thermite 0.2.1

use generic_array::{
    GenericArray,
    sequence::GenericSequence,
    typenum::{self, Unsigned},
};

use crate::{
    isa::InstructionSet,
    register::{
        BitshiftRegister, BitwiseRegister, CastRegister, CoreRegister, ExtendRegister, IntegerRegister,
        InterleaveRegister, MaskElement, MaskRegister, NumericRegister, PartialOrdRegister, Register, Storage,
        UnsignedIntegerRegister, array::ArrayRegister, empty_reg, reg,
    },
};

use super::arch;

#[cfg_attr(not(feature = "document_registers"), doc(hidden))]
#[derive(Debug, Clone, Copy, Hash)]
pub struct U8x16V2;

#[thermite_macros::inline_always]
impl CoreRegister for U8x16V2 {
    type NativeIsa = crate::backend::x86_v2::X86V2;
    type Lanes = typenum::U16;
    type Storage = arch::__m128i;
    type Mask = Self;

    const IS_EMULATED: bool = false;

    const EMPTY: Storage<Self> = empty_reg::<Self>();

    const HAS_EQUAL_SIZE_MASK: bool = true;

    fn blendv(mask: Storage<Self::Mask>, lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_blendv_epi8(lhs, rhs, mask) }
    }

    fn zz(mask: Storage<Self::Mask>, value: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_and_si128(value, mask) }
    }

    fn nz(mask: Storage<Self::Mask>, value: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_andnot_si128(mask, value) }
    }

    fn zeroupper_z<Z: crate::register::ZeroUpper>(value: Storage<Self>) -> Storage<Self> {
        super::I8x16V2::zeroupper_z::<Z>(value) // same mask layout
    }

    fn from_mask(mask: Storage<Self::Mask>) -> Storage<Self> {
        mask
    }
}

#[thermite_macros::inline_always]
impl InterleaveRegister for U8x16V2 {
    fn interleave(a: Storage<Self>, b: Storage<Self>) -> (Storage<Self>, Storage<Self>) {
        super::I8x16V2::interleave(a, b)
    }

    fn deinterleave(a: Storage<Self>, b: Storage<Self>) -> (Storage<Self>, Storage<Self>) {
        super::I8x16V2::deinterleave(a, b)
    }
}

#[thermite_macros::inline_always]
impl MaskRegister for U8x16V2 {
    fn set(mut mask: Storage<Self::Mask>, lane: usize, value: bool) -> Storage<Self> {
        Self::as_mut_slice(&mut mask)[lane] = if value { MaskElement::TRUTHY } else { MaskElement::FALSY };
        mask
    }

    fn test(mask: Storage<Self::Mask>, lane: usize) -> bool {
        Self::as_slice(&mask)[lane].to_bool()
    }

    const FALSY: Storage<Self> = reg::<Self, 16>([0; 16]);
    const TRUTHY: Storage<Self> = reg::<Self, 16>([!0; 16]);

    fn all(value: Storage<Self>) -> bool {
        unsafe { arch::_mm_movemask_epi8(value) as u32 == 0xFFFF }
    }

    fn any(value: Storage<Self>) -> bool {
        unsafe { arch::_mm_movemask_epi8(value) != 0 }
    }

    fn none(value: Storage<Self>) -> bool {
        unsafe { arch::_mm_movemask_epi8(value) == 0 }
    }

    fn from_native_bitmask(bitmask: u64) -> Storage<Self> {
        unsafe { arch::_mm_movm_epi8x_v2(bitmask) }
    }

    fn native_bitmask(value: Storage<Self>) -> Option<u64> {
        Some(unsafe { (arch::_mm_movemask_epi8(value) as u32) as u64 })
    }

    #[cfg(feature = "bitvec")]
    fn fill_bitmask(value: Storage<Self>, view: &mut bitvec::slice::BitSlice<u32>) {
        let mask = Self::native_bitmask(value).unwrap() as u32;
        let mask = bitvec::slice::BitSlice::from_slice(core::slice::from_ref(&mask));
        view.copy_from_bitslice(&mask[..Self::Lanes::USIZE]);
    }
}

#[rustfmt::skip] #[thermite_macros::inline_always]
impl BitwiseRegister for U8x16V2 {
    fn bitxor(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_xor_si128(lhs, rhs) }
    }

    fn bitand(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_and_si128(lhs, rhs) }
    }

    fn bitandnot(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_andnot_si128(lhs, rhs) }
    }

    fn bitor(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_or_si128(lhs, rhs) }
    }

    fn not(value: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_xor_si128(value, arch::_mm_set1_epi8(-1)) }
    }
}

#[thermite_macros::inline_always]
impl ExtendRegister<u8> for U8x16V2 {
    fn extend(value: Storage<u8>) -> Storage<Self> {
        unsafe { arch::_mm_setr_epi8(value as i8, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) }
    }

    fn narrow(value: Storage<Self>) -> Storage<u8> {
        unsafe { arch::_mm_extract_epi8::<0>(value) as u8 }
    }
}

#[thermite_macros::inline_always]
impl Register for U8x16V2 {
    type Element = u8;

    type Signed = super::I8x16V2;
    type Unsigned = super::U8x16V2;

    fn into_mask(value: Storage<Self>) -> Storage<Self::Mask> {
        Self::ne(value, Self::ZERO)
    }

    fn into_mask_unchecked(value: Storage<Self>) -> Storage<Self::Mask> {
        value
    }

    fn msb_to_mask(value: Storage<Self>) -> Storage<Self::Mask> {
        // Smear bit 7 of each byte (same as the signed register).
        unsafe { arch::_mm_cmpgt_epi8(arch::_mm_setzero_si128(), value) }
    }

    fn new(value: GenericArray<Self::Element, Self::Lanes>) -> Storage<Self> {
        unsafe { arch::_mm_loadu_si128(value.as_ptr() as *const _) }
    }

    fn single(value: Self::Element) -> Storage<Self> {
        unsafe { arch::_mm_setr_epi8(value as i8, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) }
    }

    impl_native_extract!(@epi8x16);

    fn splat(value: Self::Element) -> Storage<Self> {
        unsafe { arch::_mm_set1_epi8(value as i8) }
    }

    unsafe fn load(ptr: *const Self::Element) -> Storage<Self> {
        unsafe { arch::_mm_load_si128(ptr as *const _) }
    }

    unsafe fn load_unaligned(ptr: *const Self::Element) -> Storage<Self> {
        unsafe { arch::_mm_loadu_si128(ptr as *const _) }
    }

    unsafe fn store(ptr: *mut Self::Element, value: Storage<Self>) {
        unsafe { arch::_mm_store_si128(ptr as *mut _, value) }
    }

    unsafe fn store_unaligned(ptr: *mut Self::Element, value: Storage<Self>) {
        unsafe { arch::_mm_storeu_si128(ptr as *mut _, value) }
    }

    unsafe fn load_stream(ptr: *const Self::Element) -> Storage<Self> {
        unsafe { arch::_mm_stream_load_si128(ptr as _) }
    }

    unsafe fn store_stream(ptr: *mut Self::Element, value: Storage<Self>) {
        unsafe { arch::_mm_stream_si128(ptr as _, value) }
    }

    fn reverse(value: Storage<Self>) -> Storage<Self> {
        unsafe {
            arch::_mm_shuffle_epi8(
                value,
                arch::_mm_setr_epi8(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0),
            )
        }
    }

    fn swap_bytes(value: Storage<Self>) -> Storage<Self> {
        value
    }

    const HAS_PERMUTEV: bool = true;

    fn permutev(value: Storage<Self>, idxs: GenericArray<u32, Self::Lanes>) -> Storage<Self> {
        // Lane permutation is bit-pattern identical for signed/unsigned (same storage).
        super::I8x16V2::permutev(value, idxs)
    }

    compress_via_wide!();

    // Byte-identical to the signed register (both are raw `__m128i`); reuse it.
    // hand-written body below, so the flag is set here rather than by an
    // `impl_*_align*!` macro
    const HAS_NATIVE_ALIGN: bool = true;

    fn align<const OFFSET: usize>(a: Storage<Self>, b: Storage<Self>) -> Storage<Self> {
        super::I8x16V2::align::<OFFSET>(a, b)
    }
}

#[rustfmt::skip] #[thermite_macros::inline_always]
impl BitshiftRegister for U8x16V2 {
    const HAS_TRUE_SHIFTV: bool = false;
    const HAS_WIDE_BYTE_SHIFTS: bool = true;

    fn bshli<const IMM8: i32>(value: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_bslli_si128(value, IMM8) }
    }

    fn bshri<const IMM8: i32>(value: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_bsrli_si128(value, IMM8) }
    }

    fn shl(value: Storage<Self>, shift: u32) -> Storage<Self> {
        unsafe { arch::_mm_sll_epi8x_v1(value, shift) }
    }

    fn shr(value: Storage<Self>, shift: u32) -> Storage<Self> {
        unsafe { arch::_mm_srl_epi8x_v1(value, shift) }
    }

    fn shli<const IMM8: i32>(value: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_slli_epi8x_v1::<IMM8>(value) }
    }

    fn shri<const IMM8: i32>(value: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_srli_epi8x_v1::<IMM8>(value) }
    }
}

#[thermite_macros::inline_always]
impl PartialOrdRegister for U8x16V2 {
    fn gt(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_cmpgt_epu8x_v1(lhs, rhs) }
    }

    fn lt(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_cmpgt_epu8x_v1(rhs, lhs) }
    }

    fn eq(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_cmpeq_epi8(lhs, rhs) }
    }
}

#[thermite_macros::inline_always]
impl NumericRegister for U8x16V2 {
    const ZERO: Storage<Self> = reg::<Self, 16>([0; 16]);
    const ONE: Storage<Self> = reg::<Self, 16>([1; 16]);
    const TWO: Storage<Self> = reg::<Self, 16>([2; 16]);

    const MIN: Storage<Self> = reg::<Self, 16>([u8::MIN; 16]);
    const MAX: Storage<Self> = reg::<Self, 16>([u8::MAX; 16]);

    fn min_element(value: Storage<Self>) -> Self::Element {
        _mm_reduce_epi8_v2!(value; _mm_min_epu8) as u8
    }

    fn max_element(value: Storage<Self>) -> Self::Element {
        _mm_reduce_epi8_v2!(value; _mm_max_epu8) as u8
    }

    fn sum_elements(value: Storage<Self>) -> Self::Element {
        _mm_reduce_epi8_v2!(value; _mm_add_epi8) as u8
    }

    fn prod_elements(value: Storage<Self>) -> Self::Element {
        Self::as_slice(&value).iter().copied().fold(1u8, u8::wrapping_mul)
    }

    fn offset() -> Storage<Self> {
        Self::splat(<Self::Lanes as typenum::Unsigned>::U8)
    }

    fn indexed() -> Storage<Self> {
        Self::new(GenericArray::generate(|i| i as u8))
    }

    fn add(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_add_epi8(lhs, rhs) }
    }

    fn sub(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_sub_epi8(lhs, rhs) }
    }

    fn add_c(mask: Storage<Self::Mask>, lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_add_epi8(lhs, arch::_mm_and_si128(rhs, mask)) }
    }

    fn sub_c(mask: Storage<Self::Mask>, lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_sub_epi8(lhs, arch::_mm_and_si128(rhs, mask)) }
    }

    fn mul(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_mullo_epi8x_v1(lhs, rhs) }
    }

    fn div(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        Self::zip(lhs, rhs, |a, b| if b == 0 { 0 } else { a / b })
    }

    fn rem(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        Self::zip(lhs, rhs, |a, b| if b == 0 { 0 } else { a % b })
    }

    fn min(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_min_epu8(lhs, rhs) }
    }

    fn max(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_max_epu8(lhs, rhs) }
    }
}

#[thermite_macros::inline_always]
impl IntegerRegister for U8x16V2 {
    fn mulhi(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_mulhi_epu8x_v1(lhs, rhs) }
    }

    fn mullo(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_mullo_epi8x_v1(lhs, rhs) }
    }

    fn saturating_add(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_adds_epu8(lhs, rhs) }
    }

    fn saturating_sub(lhs: Storage<Self>, rhs: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_subs_epu8(lhs, rhs) }
    }

    fn div_branched(value: Storage<Self>, divider: crate::Divider<Self::Element>) -> Storage<Self> {
        arch::div_epu::<Self>(value, divider.multiplier(), divider.shift())
    }

    fn div_branchfree(value: Storage<Self>, divider: crate::BranchfreeDivider<Self::Element>) -> Storage<Self> {
        arch::div_epu_bf::<Self>(value, divider.multiplier(), divider.shift())
    }

    fn divv_branchfree(value: Storage<Self>, dividers: crate::divider::vector::VectorDivider<Self>) -> Storage<Self> {
        arch::divv_epu_bf::<Self>(value, dividers.multipliers.0, dividers.shifts.0)
    }

    const HAS_HARDWARE_POPCNT: bool = false;

    fn count_ones(value: Storage<Self>) -> Storage<Self> {
        unsafe { arch::_mm_popcnt_epi8x_v2(value) }
    }

    fn leading_zeros(value: Storage<Self>) -> Storage<Self> {
        Self::sub(Self::splat(8), Self::ilog2p1(value))
    }

    fn trailing_zeros(value: Storage<Self>) -> Storage<Self> {
        super::I8x16V2::trailing_zeros(value)
    }
}

#[thermite_macros::inline_always]
impl UnsignedIntegerRegister for U8x16V2 {}

// Saturating narrow u16x16 -> u8x16: clamp each half (`pminuw`) then two-source `packuswb`.
#[thermite_macros::inline_always]
impl CastRegister<ArrayRegister<super::U16x8V2, 2>> for U8x16V2 {
    fn saturating_cast_from(value: Storage<ArrayRegister<super::U16x8V2, 2>>) -> Storage<Self> {
        unsafe {
            let max = arch::_mm_set1_epi16(0xFF);
            let lo = arch::_mm_min_epu16(value.0[0], max);
            let hi = arch::_mm_min_epu16(value.0[1], max);
            arch::_mm_packus_epi16(lo, hi)
        }
    }

    fn cast_from(value: Storage<ArrayRegister<super::U16x8V2, 2>>) -> Storage<Self> {
        unsafe { arch::_mm_cvt2epi16_epi8x_v2(value.0) }
    }
}

// Saturating narrow u32x16 -> u8x16: clamp + `packusdw` to u16, then clamp + `packuswb` to u8.
#[thermite_macros::inline_always]
impl CastRegister<ArrayRegister<super::U32x4V2, 4>> for U8x16V2 {
    fn saturating_cast_from(value: Storage<ArrayRegister<super::U32x4V2, 4>>) -> Storage<Self> {
        unsafe {
            let m32 = arch::_mm_set1_epi32(0xFFFF);
            let w0 = arch::_mm_packus_epi32(
                arch::_mm_min_epu32(value.0[0], m32),
                arch::_mm_min_epu32(value.0[1], m32),
            );
            let w1 = arch::_mm_packus_epi32(
                arch::_mm_min_epu32(value.0[2], m32),
                arch::_mm_min_epu32(value.0[3], m32),
            );
            let m16 = arch::_mm_set1_epi16(0xFF);
            arch::_mm_packus_epi16(arch::_mm_min_epu16(w0, m16), arch::_mm_min_epu16(w1, m16))
        }
    }

    fn cast_from(value: Storage<ArrayRegister<super::U32x4V2, 4>>) -> Storage<Self> {
        unsafe { arch::_mm_cvt4epi32_epi8x_v2(value.0) }
    }
}

// Saturating narrow u64x16 -> u8x16: clamp the high end (polyfilled 64-bit min) + truncating narrow.
#[thermite_macros::inline_always]
impl CastRegister<ArrayRegister<super::U64x2V2, 8>> for U8x16V2 {
    fn saturating_cast_from(value: Storage<ArrayRegister<super::U64x2V2, 8>>) -> Storage<Self> {
        type Src = ArrayRegister<super::U64x2V2, 8>;
        let hi = <Src as Register>::splat(u8::MAX as u64);
        let clamped = <Src as NumericRegister>::min(value, hi);
        <Self as CastRegister<Src>>::cast_from(clamped)
    }

    fn cast_from(value: Storage<ArrayRegister<super::U64x2V2, 8>>) -> Storage<Self> {
        unsafe { arch::_mm_cvt8epi64_epi8x_v2(value.0) }
    }
}