wide/

f64x4_.rs

use super::*;

pick! {
  if #[cfg(target_feature="avx")] {
    #[derive(Default, Clone, Copy, PartialEq)]
    #[repr(C, align(32))]
    pub struct f64x4 { pub(crate) avx: m256d }
  } else {
    #[derive(Default, Clone, Copy, PartialEq)]
    #[repr(C, align(32))]
    pub struct f64x4 { pub(crate) a: f64x2, pub(crate) b: f64x2 }
  }
}

macro_rules! const_f64_as_f64x4 {
  ($i:ident, $f:expr) => {
    #[allow(non_upper_case_globals)]
    pub const $i: f64x4 = f64x4::new([$f; 4]);
  };
}

impl f64x4 {
  const_f64_as_f64x4!(ONE, 1.0);
  const_f64_as_f64x4!(ZERO, 0.0);
  const_f64_as_f64x4!(HALF, 0.5);
  const_f64_as_f64x4!(EPSILON, f64::EPSILON);
  const_f64_as_f64x4!(MIN, f64::MIN);
  const_f64_as_f64x4!(MIN_POSITIVE, f64::MIN_POSITIVE);
  const_f64_as_f64x4!(MAX, f64::MAX);
  const_f64_as_f64x4!(NAN, f64::NAN);
  const_f64_as_f64x4!(INFINITY, f64::INFINITY);
  const_f64_as_f64x4!(NEG_INFINITY, f64::NEG_INFINITY);
  const_f64_as_f64x4!(E, core::f64::consts::E);
  const_f64_as_f64x4!(FRAC_1_PI, core::f64::consts::FRAC_1_PI);
  const_f64_as_f64x4!(FRAC_2_PI, core::f64::consts::FRAC_2_PI);
  const_f64_as_f64x4!(FRAC_2_SQRT_PI, core::f64::consts::FRAC_2_SQRT_PI);
  const_f64_as_f64x4!(FRAC_1_SQRT_2, core::f64::consts::FRAC_1_SQRT_2);
  const_f64_as_f64x4!(FRAC_PI_2, core::f64::consts::FRAC_PI_2);
  const_f64_as_f64x4!(FRAC_PI_3, core::f64::consts::FRAC_PI_3);
  const_f64_as_f64x4!(FRAC_PI_4, core::f64::consts::FRAC_PI_4);
  const_f64_as_f64x4!(FRAC_PI_6, core::f64::consts::FRAC_PI_6);
  const_f64_as_f64x4!(FRAC_PI_8, core::f64::consts::FRAC_PI_8);
  const_f64_as_f64x4!(LN_2, core::f64::consts::LN_2);
  const_f64_as_f64x4!(LN_10, core::f64::consts::LN_10);
  const_f64_as_f64x4!(LOG2_E, core::f64::consts::LOG2_E);
  const_f64_as_f64x4!(LOG10_E, core::f64::consts::LOG10_E);
  const_f64_as_f64x4!(LOG10_2, core::f64::consts::LOG10_2);
  const_f64_as_f64x4!(LOG2_10, core::f64::consts::LOG2_10);
  const_f64_as_f64x4!(PI, core::f64::consts::PI);
  const_f64_as_f64x4!(SQRT_2, core::f64::consts::SQRT_2);
  const_f64_as_f64x4!(TAU, core::f64::consts::TAU);
}

unsafe impl Zeroable for f64x4 {}
unsafe impl Pod for f64x4 {}

impl AlignTo for f64x4 {
  type Elem = f64;
}

impl Add for f64x4 {
  type Output = Self;
  #[inline]
  fn add(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: add_m256d(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.add(rhs.a),
          b : self.b.add(rhs.b),
        }
      }
    }
  }
}

impl Sub for f64x4 {
  type Output = Self;
  #[inline]
  fn sub(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: sub_m256d(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.sub(rhs.a),
          b : self.b.sub(rhs.b),
        }
      }
    }
  }
}

impl Mul for f64x4 {
  type Output = Self;
  #[inline]
  fn mul(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: mul_m256d(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.mul(rhs.a),
          b : self.b.mul(rhs.b),
        }
      }
    }
  }
}

impl Div for f64x4 {
  type Output = Self;
  #[inline]
  fn div(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: div_m256d(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.div(rhs.a),
          b : self.b.div(rhs.b),
        }
      }
    }
  }
}

impl Rem for f64x4 {
  type Output = Self;
  #[inline]
  fn rem(self, rhs: Self) -> Self::Output {
    Self::new([
      self.to_array()[0] % rhs.to_array()[0],
      self.to_array()[1] % rhs.to_array()[1],
      self.to_array()[2] % rhs.to_array()[2],
      self.to_array()[3] % rhs.to_array()[3],
    ])
  }
}

impl Neg for f64x4 {
  type Output = Self;
  #[inline]
  fn neg(self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: bitxor_m256d(self.avx, Self::splat(-0.0).avx) }
      } else {
        Self {
          a : self.a.neg(),
          b : self.b.neg(),
        }
      }
    }
  }
}

impl Add<f64> for f64x4 {
  type Output = Self;
  #[inline]
  fn add(self, rhs: f64) -> Self::Output {
    self.add(Self::splat(rhs))
  }
}

impl Sub<f64> for f64x4 {
  type Output = Self;
  #[inline]
  fn sub(self, rhs: f64) -> Self::Output {
    self.sub(Self::splat(rhs))
  }
}

impl Mul<f64> for f64x4 {
  type Output = Self;
  #[inline]
  fn mul(self, rhs: f64) -> Self::Output {
    self.mul(Self::splat(rhs))
  }
}

impl Div<f64> for f64x4 {
  type Output = Self;
  #[inline]
  fn div(self, rhs: f64) -> Self::Output {
    self.div(Self::splat(rhs))
  }
}

impl Rem<f64> for f64x4 {
  type Output = Self;
  #[inline]
  fn rem(self, rhs: f64) -> Self::Output {
    self.rem(Self::splat(rhs))
  }
}

impl Add<f64x4> for f64 {
  type Output = f64x4;
  #[inline]
  fn add(self, rhs: f64x4) -> Self::Output {
    f64x4::splat(self).add(rhs)
  }
}

impl Sub<f64x4> for f64 {
  type Output = f64x4;
  #[inline]
  fn sub(self, rhs: f64x4) -> Self::Output {
    f64x4::splat(self).sub(rhs)
  }
}

impl Mul<f64x4> for f64 {
  type Output = f64x4;
  #[inline]
  fn mul(self, rhs: f64x4) -> Self::Output {
    f64x4::splat(self).mul(rhs)
  }
}

impl Div<f64x4> for f64 {
  type Output = f64x4;
  #[inline]
  fn div(self, rhs: f64x4) -> Self::Output {
    f64x4::splat(self).div(rhs)
  }
}

impl Rem<f64x4> for f64 {
  type Output = f64x4;
  #[inline]
  fn rem(self, rhs: f64x4) -> Self::Output {
    f64x4::splat(self).rem(rhs)
  }
}

impl BitAnd for f64x4 {
  type Output = Self;
  #[inline]
  fn bitand(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: bitand_m256d(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.bitand(rhs.a),
          b : self.b.bitand(rhs.b),
        }
      }
    }
  }
}

impl BitOr for f64x4 {
  type Output = Self;
  #[inline]
  fn bitor(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: bitor_m256d(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.bitor(rhs.a),
          b : self.b.bitor(rhs.b),
        }
      }
    }
  }
}

impl BitXor for f64x4 {
  type Output = Self;
  #[inline]
  fn bitxor(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: bitxor_m256d(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.bitxor(rhs.a),
          b : self.b.bitxor(rhs.b),
        }
      }
    }
  }
}

#[expect(deprecated)]
impl CmpEq for f64x4 {
  type Output = Self;
  #[inline]
  fn simd_eq(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")]{
        Self { avx: cmp_op_mask_m256d::<{cmp_op!(EqualOrdered)}>(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.simd_eq(rhs.a),
          b : self.b.simd_eq(rhs.b),
        }
      }
    }
  }
}

#[expect(deprecated)]
impl CmpGe for f64x4 {
  type Output = Self;
  #[inline]
  fn simd_ge(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")]{
        Self { avx: cmp_op_mask_m256d::<{cmp_op!(GreaterEqualOrdered)}>(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.simd_ge(rhs.a),
          b : self.b.simd_ge(rhs.b),
        }
      }
    }
  }
}

#[expect(deprecated)]
impl CmpGt for f64x4 {
  type Output = Self;
  #[inline]
  fn simd_gt(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")]{
        Self { avx: cmp_op_mask_m256d::<{cmp_op!( GreaterThanOrdered)}>(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.simd_gt(rhs.a),
          b : self.b.simd_gt(rhs.b),
        }
      }
    }
  }
}

#[expect(deprecated)]
impl CmpNe for f64x4 {
  type Output = Self;
  #[inline]
  fn simd_ne(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")]{
        Self { avx: cmp_op_mask_m256d::<{cmp_op!(NotEqualUnordered)}>(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.simd_ne(rhs.a),
          b : self.b.simd_ne(rhs.b),
        }
      }
    }
  }
}

#[expect(deprecated)]
impl CmpLe for f64x4 {
  type Output = Self;
  #[inline]
  fn simd_le(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")]{
        Self { avx: cmp_op_mask_m256d::<{cmp_op!(LessEqualOrdered)}>(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.simd_le(rhs.a),
          b : self.b.simd_le(rhs.b),
        }
      }
    }
  }
}

#[expect(deprecated)]
impl CmpLt for f64x4 {
  type Output = Self;
  #[inline]
  fn simd_lt(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx")]{
        Self { avx: cmp_op_mask_m256d::<{cmp_op!(LessThanOrdered)}>(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.simd_lt(rhs.a),
          b : self.b.simd_lt(rhs.b),
        }
      }
    }
  }
}

impl f64x4 {
  #[inline]
  #[must_use]
  pub const fn new(array: [f64; 4]) -> Self {
    unsafe { core::mem::transmute(array) }
  }

  simd_comparison_fns!();

  #[inline]
  #[must_use]
  pub fn blend(self, t: Self, f: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: blend_varying_m256d(f.avx, t.avx, self.avx) }
      } else {
        Self {
          a : self.a.blend(t.a, f.a),
          b : self.b.blend(t.b, f.b),
        }
      }
    }
  }

  #[inline]
  #[must_use]
  pub fn abs(self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        let non_sign_bits = f64x4::from(f64::from_bits(i64::MAX as u64));
        self & non_sign_bits
      } else {
        Self {
          a : self.a.abs(),
          b : self.b.abs(),
        }
      }
    }
  }

  #[inline]
  #[must_use]
  pub fn signum(self) -> Self {
    let result = Self::ONE | self & -Self::ZERO;

    self.is_nan().blend(self, result)
  }

  #[inline]
  #[must_use]
  pub fn floor(self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: floor_m256d(self.avx) }
      } else {
        Self {
          a : self.a.floor(),
          b : self.b.floor(),
        }
      }
    }
  }
  #[inline]
  #[must_use]
  pub fn ceil(self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: ceil_m256d(self.avx) }
      } else {
        Self {
          a : self.a.ceil(),
          b : self.b.ceil(),
        }
      }
    }
  }

  /// Calculates the lanewise maximum of both vectors. This is a faster
  /// implementation than `max`, but it doesn't specify any behavior if NaNs are
  /// involved.
  #[inline]
  #[must_use]
  pub fn fast_max(self, rhs: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: max_m256d(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.fast_max(rhs.a),
          b : self.b.fast_max(rhs.b),
        }
      }
    }
  }

  /// Calculates the lanewise maximum of both vectors. If either lane is NaN,
  /// the other lane gets chosen. Use `fast_max` for a faster implementation
  /// that doesn't handle NaNs.
  #[inline]
  #[must_use]
  pub fn max(self, rhs: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        // max_m256d seems to do rhs < self ? self : rhs. So if there's any NaN
        // involved, it chooses rhs, so we need to specifically check rhs for
        // NaN.
        rhs.is_nan().blend(self, Self { avx: max_m256d(self.avx, rhs.avx) })
      } else {
        Self {
          a : self.a.max(rhs.a),
          b : self.b.max(rhs.b),
        }
      }
    }
  }

  /// Calculates the lanewise minimum of both vectors. This is a faster
  /// implementation than `min`, but it doesn't specify any behavior if NaNs are
  /// involved.
  #[inline]
  #[must_use]
  pub fn fast_min(self, rhs: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: min_m256d(self.avx, rhs.avx) }
      } else {
        Self {
          a : self.a.fast_min(rhs.a),
          b : self.b.fast_min(rhs.b),
        }
      }
    }
  }

  /// Calculates the lanewise minimum of both vectors. If either lane is NaN,
  /// the other lane gets chosen. Use `fast_min` for a faster implementation
  /// that doesn't handle NaNs.
  #[inline]
  #[must_use]
  pub fn min(self, rhs: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        // min_m256d seems to do rhs < self ? self : rhs. So if there's any NaN
        // involved, it chooses rhs, so we need to specifically check rhs for
        // NaN.
        rhs.is_nan().blend(self, Self { avx: min_m256d(self.avx, rhs.avx) })
      } else {
        Self {
          a : self.a.min(rhs.a),
          b : self.b.min(rhs.b),
        }
      }
    }
  }

  /// Restrict a value to a certain interval unless it is NaN.
  ///
  /// If `self` is NaN, or `min` is NaN, or `max` is NaN, the result is NaN.
  /// If `min > max`, the result is `min`, since `fast_max(min)` dominates.
  #[inline]
  #[must_use]
  pub fn clamp(self, min: Self, max: Self) -> Self {
    let is_nan = self.is_nan() | min.is_nan() | max.is_nan();
    let clamped = self.fast_min(max).fast_max(min);
    is_nan.blend(Self::splat(f64::NAN), clamped)
  }

  /// Restrict a value to a certain interval unless it is NaN.
  ///
  /// Avoids NaN detection; same speed as the old `clamp` prior to IEEE 754-2019
  /// compliance. Does not specify any
  /// behavior if NaNs are involved, and if `min > max` the result is
  /// unspecified.
  #[inline]
  #[must_use]
  pub fn fast_clamp(self, min: Self, max: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        // For both `min_m256d` and `max_m256d` if any input is NaN, `rhs` gets
        // chosen. For `self` to be chosen, `self` must be the second argument.
        Self { avx: min_m256d(max.avx, max_m256d(min.avx, self.avx)) }
      } else {
        Self {
          a: self.a.fast_clamp(min.a, max.a),
          b: self.b.fast_clamp(min.b, max.b),
        }
      }
    }
  }

  #[inline]
  #[must_use]
  pub fn midpoint(self, other: Self) -> Self {
    (self + other) * 0.5
  }

  #[inline]
  #[must_use]
  pub fn is_nan(self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: cmp_op_mask_m256d::<{cmp_op!(Unordered)}>(self.avx, self.avx ) }
      } else {
        Self {
          a : self.a.is_nan(),
          b : self.b.is_nan(),
        }
      }
    }
  }

  #[inline]
  #[must_use]
  pub fn is_finite(self) -> Self {
    let shifted_exp_mask = u64x4::from(0xFFE0000000000000);
    let u: u64x4 = cast(self);
    let shift_u = u << 1_u64;
    let out = !(shift_u & shifted_exp_mask).simd_eq(shifted_exp_mask);
    cast(out)
  }

  #[inline]
  #[must_use]
  pub fn is_inf(self) -> Self {
    let shifted_inf = u64x4::from(0xFFE0000000000000);
    let u: u64x4 = cast(self);
    let shift_u = u << 1_u64;
    let out = (shift_u).simd_eq(shifted_inf);
    cast(out)
  }

  #[inline]
  #[must_use]
  pub fn round(self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: round_m256d::<{round_op!(Nearest)}>(self.avx) }
      } else {
        Self {
          a : self.a.round(),
          b : self.b.round(),
        }
      }
    }
  }

  #[inline]
  #[must_use]
  pub fn fast_round_int(self) -> i64x4 {
    pick! {
      if #[cfg(all(target_feature="avx512dq", target_feature="avx512vl"))] {
        #[cfg(target_arch = "x86")]
        use core::arch::x86::_mm256_cvtpd_epi64;
        #[cfg(target_arch = "x86_64")]
        use core::arch::x86_64::_mm256_cvtpd_epi64;

        // TODO(safe_arch): Add `_mm256_cvtpd_epi64`.
        cast(m256i(unsafe { _mm256_cvtpd_epi64(self.avx.0) }))
      } else {
        let [a, b]: [f64x2; 2] = cast(self);
        cast([a.fast_round_int(), b.fast_round_int()])
      }
    }
  }

  #[inline]
  #[must_use]
  pub fn round_int(self) -> i64x4 {
    pick! {
      if #[cfg(all(target_feature="avx512dq", target_feature="avx512vl"))] {
        #[cfg(target_arch = "x86")]
        use core::arch::x86::_mm256_cvtpd_epi64;
        #[cfg(target_arch = "x86_64")]
        use core::arch::x86_64::_mm256_cvtpd_epi64;

        // Based on: https://github.com/v8/v8/blob/210987a552a2bf2a854b0baa9588a5959ff3979d/src/codegen/shared-ia32-x64/macro-assembler-shared-ia32-x64.h#L489-L504
        let non_nan_mask = self.simd_eq(self);
        let non_nan = self & non_nan_mask;
        let flip_to_max: i64x4 = cast(self.simd_ge(Self::splat(9223372036854775808.0)));

        // TODO(safe_arch): Add `_mm256_cvtpd_epi64`.
        let cast: i64x4 = cast(m256i(unsafe { _mm256_cvtpd_epi64(non_nan.avx.0) }));
        flip_to_max ^ cast
      } else {
        let [a, b]: [f64x2; 2] = cast(self);
        cast([a.round_int(), b.round_int()])
      }
    }
  }

  #[inline]
  #[must_use]
  pub fn trunc(self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: round_m256d::<{round_op!(Zero)}>(self.avx) }
      } else {
        Self {
          a : self.a.trunc(),
          b : self.b.trunc(),
        }
      }
    }
  }

  /// Truncates each lane into an integer. This is a faster implementation than
  /// `trunc_int`, but it doesn't handle out of range values or NaNs. For those
  /// values you get implementation defined behavior.
  #[inline]
  #[must_use]
  pub fn fast_trunc_int(self) -> i64x4 {
    pick! {
      if #[cfg(all(target_feature="avx512dq", target_feature="avx512vl"))] {
        #[cfg(target_arch = "x86")]
        use core::arch::x86::_mm256_cvttpd_epi64;
        #[cfg(target_arch = "x86_64")]
        use core::arch::x86_64::_mm256_cvttpd_epi64;

        // TODO(safe_arch): Add `_mm256_cvttpd_epi64`.
        cast(m256i(unsafe { _mm256_cvttpd_epi64(self.avx.0) }))
      } else {
        let [a, b]: [f64x2; 2] = cast(self);
        cast([a.fast_trunc_int(), b.fast_trunc_int()])
      }
    }
  }

  /// Truncates each lane into an integer. This saturates out of range values
  /// and turns NaNs into 0. Use `fast_trunc_int` for a faster implementation
  /// that doesn't handle out of range values or NaNs.
  #[inline]
  #[must_use]
  pub fn trunc_int(self) -> i64x4 {
    pick! {
      if #[cfg(all(target_feature="avx512dq", target_feature="avx512vl"))] {
        #[cfg(target_arch = "x86")]
        use core::arch::x86::_mm256_cvttpd_epi64;
        #[cfg(target_arch = "x86_64")]
        use core::arch::x86_64::_mm256_cvttpd_epi64;

        // Based on: https://github.com/v8/v8/blob/210987a552a2bf2a854b0baa9588a5959ff3979d/src/codegen/shared-ia32-x64/macro-assembler-shared-ia32-x64.h#L489-L504
        let non_nan_mask = self.simd_eq(self);
        let non_nan = self & non_nan_mask;
        let flip_to_max: i64x4 = cast(self.simd_ge(Self::splat(9223372036854775808.0)));

        // TODO(safe_arch): Add `_mm256_cvttpd_epi64`.
        let cast: i64x4 = cast(m256i(unsafe { _mm256_cvttpd_epi64(non_nan.avx.0) }));
        flip_to_max ^ cast
      } else {
        let [a, b]: [f64x2; 2] = cast(self);
        cast([a.trunc_int(), b.trunc_int()])
      }
    }
  }

  #[inline]
  #[must_use]
  pub fn fract(self) -> Self {
    self - self.trunc()
  }

  /// Performs a multiply-add operation: `self * m + a`
  ///
  /// When hardware FMA support is available, this computes the result with a
  /// single rounding operation. Without FMA support, it falls back to separate
  /// multiply and add operations with two roundings.
  ///
  /// # Platform-specific behavior
  /// - On `x86`/`x86_64` with AVX+FMA: Uses 256-bit `vfmadd` (single rounding,
  ///   best accuracy)
  /// - On `x86`/`x86_64` with AVX only: Uses `(self * m) + a` (two roundings)
  /// - Other platforms: Delegates to [`f64x2`] (may use NEON FMA or fallback)
  ///
  /// # Examples
  /// ```
  /// # use wide::f64x4;
  /// let a = f64x4::from([1.0, 2.0, 3.0, 4.0]);
  /// let b = f64x4::from([2.0; 4]);
  /// let c = f64x4::from([10.0; 4]);
  ///
  /// let result = a.mul_add(b, c);
  ///
  /// let expected = f64x4::from([12.0, 14.0, 16.0, 18.0]);
  /// assert_eq!(result, expected);
  /// ```
  #[inline]
  #[must_use]
  pub fn mul_add(self, m: Self, a: Self) -> Self {
    pick! {
      if #[cfg(all(target_feature="avx",target_feature="fma"))] {
        Self { avx: fused_mul_add_m256d(self.avx, m.avx, a.avx) }
      } else if #[cfg(target_feature="avx")] {
        // still want to use 256 bit ops
        (self * m) + a
      } else {
        Self {
          a : self.a.mul_add(m.a, a.a),
          b : self.b.mul_add(m.b, a.b),
        }
      }
    }
  }

  /// Performs a multiply-subtract operation: `self * m - s`
  ///
  /// When hardware FMA support is available, this computes the result with a
  /// single rounding operation. Without FMA support, it falls back to separate
  /// multiply and subtract operations with two roundings.
  ///
  /// # Platform-specific behavior
  /// - On `x86`/`x86_64` with AVX+FMA: Uses 256-bit `vfmsub` (single rounding,
  ///   best accuracy)
  /// - On `x86`/`x86_64` with AVX only: Uses `(self * m) - s` (two roundings)
  /// - Other platforms: Delegates to [`f64x2`] (may use NEON FMA or fallback)
  ///
  /// # Examples
  /// ```
  /// # use wide::f64x4;
  /// let a = f64x4::from([10.0; 4]);
  /// let b = f64x4::from([2.0; 4]);
  /// let c = f64x4::from([5.0; 4]);
  ///
  /// let result = a.mul_sub(b, c);
  ///
  /// let expected = f64x4::from([15.0; 4]);
  /// assert_eq!(result, expected);
  /// ```
  #[inline]
  #[must_use]
  pub fn mul_sub(self, m: Self, s: Self) -> Self {
    pick! {
      if #[cfg(all(target_feature="avx",target_feature="fma"))] {
        Self { avx: fused_mul_sub_m256d(self.avx, m.avx, s.avx) }
      } else if #[cfg(target_feature="avx")] {
        // still want to use 256 bit ops
        (self * m) - s
      } else {
        Self {
          a : self.a.mul_sub(m.a, s.a),
          b : self.b.mul_sub(m.b, s.b),
        }
      }
    }
  }

  /// Performs a negative multiply-add operation: `a - (self * m)`
  ///
  /// When hardware FMA support is available, this computes the result with a
  /// single rounding operation. Without FMA support, it falls back to separate
  /// operations with two roundings.
  ///
  /// # Platform-specific behavior
  /// - On `x86`/`x86_64` with AVX+FMA: Uses 256-bit `vfnmadd` (single rounding,
  ///   best accuracy)
  /// - On `x86`/`x86_64` with AVX only: Uses `a - (self * m)` (two roundings)
  /// - Other platforms: Delegates to [`f64x2`] (may use NEON FMA or fallback)
  ///
  /// # Examples
  /// ```
  /// # use wide::f64x4;
  /// let a = f64x4::from([3.0; 4]);
  /// let b = f64x4::from([2.0; 4]);
  /// let c = f64x4::from([10.0; 4]);
  ///
  /// let result = a.mul_neg_add(b, c);
  ///
  /// let expected = f64x4::from([4.0; 4]);
  /// assert_eq!(result, expected);
  /// ```
  #[inline]
  #[must_use]
  pub fn mul_neg_add(self, m: Self, a: Self) -> Self {
    pick! {
      if #[cfg(all(target_feature="avx",target_feature="fma"))] {
        Self { avx: fused_mul_neg_add_m256d(self.avx, m.avx, a.avx) }
      } else if #[cfg(target_feature="avx")] {
        // still want to use 256 bit ops
        a - (self * m)
      } else {
        Self {
          a : self.a.mul_neg_add(m.a, a.a),
          b : self.b.mul_neg_add(m.b, a.b),
        }
      }
    }
  }

  /// Performs a negative multiply-subtract operation: `-(self * m) - s`
  ///
  /// When hardware FMA support is available, this computes the result with a
  /// single rounding operation. Without FMA support, it falls back to separate
  /// operations with two roundings.
  ///
  /// # Platform-specific behavior
  /// - On `x86`/`x86_64` with AVX+FMA: Uses 256-bit `vfnmsub` (single rounding,
  ///   best accuracy)
  /// - On `x86`/`x86_64` with AVX only: Uses `-(self * m) - s` (two roundings)
  /// - Other platforms: Delegates to [`f64x2`] (may use NEON FMA or fallback)
  ///
  /// # Examples
  /// ```
  /// # use wide::f64x4;
  /// let a = f64x4::from([3.0; 4]);
  /// let b = f64x4::from([2.0; 4]);
  /// let c = f64x4::from([1.0; 4]);
  ///
  /// let result = a.mul_neg_sub(b, c);
  ///
  /// let expected = f64x4::from([-7.0; 4]);
  /// assert_eq!(result, expected);
  /// ```
  #[inline]
  #[must_use]
  pub fn mul_neg_sub(self, m: Self, s: Self) -> Self {
    pick! {
       if #[cfg(all(target_feature="avx",target_feature="fma"))] {
         Self { avx: fused_mul_neg_sub_m256d(self.avx, m.avx, s.avx) }
        } else if #[cfg(target_feature="avx")] {
          // still want to use 256 bit ops
          -(self * m) - s
        } else {
         Self {
           a : self.a.mul_neg_sub(m.a, s.a),
           b : self.b.mul_neg_sub(m.b, s.b),
         }
       }
    }
  }

  #[inline]
  #[must_use]
  pub fn div_euclid(self, rhs: Self) -> Self {
    let q = (self / rhs).trunc();
    (self % rhs)
      .simd_lt(Self::ZERO)
      .blend(rhs.simd_gt(Self::ZERO).blend(q - Self::ONE, q + Self::ONE), q)
  }

  #[inline]
  #[must_use]
  pub fn rem_euclid(self, rhs: Self) -> Self {
    let r = self % rhs;
    r.simd_lt(Self::ZERO).blend(r + rhs.abs(), r)
  }

  #[inline]
  #[must_use]
  pub fn flip_signs(self, signs: Self) -> Self {
    self ^ (signs & Self::from(-0.0))
  }

  #[inline]
  #[must_use]
  pub fn copysign(self, sign: Self) -> Self {
    let magnitude_mask = Self::from(f64::from_bits(u64::MAX >> 1));
    (self & magnitude_mask) | (sign & Self::from(-0.0))
  }

  #[inline]
  pub fn asin_acos(self) -> (Self, Self) {
    // Based on the Agner Fog "vector class library":
    // https://github.com/vectorclass/version2/blob/master/vectormath_trig.h
    const_f64_as_f64x4!(R4asin, 2.967721961301243206100E-3);
    const_f64_as_f64x4!(R3asin, -5.634242780008963776856E-1);
    const_f64_as_f64x4!(R2asin, 6.968710824104713396794E0);
    const_f64_as_f64x4!(R1asin, -2.556901049652824852289E1);
    const_f64_as_f64x4!(R0asin, 2.853665548261061424989E1);

    const_f64_as_f64x4!(S3asin, -2.194779531642920639778E1);
    const_f64_as_f64x4!(S2asin, 1.470656354026814941758E2);
    const_f64_as_f64x4!(S1asin, -3.838770957603691357202E2);
    const_f64_as_f64x4!(S0asin, 3.424398657913078477438E2);

    const_f64_as_f64x4!(P5asin, 4.253011369004428248960E-3);
    const_f64_as_f64x4!(P4asin, -6.019598008014123785661E-1);
    const_f64_as_f64x4!(P3asin, 5.444622390564711410273E0);
    const_f64_as_f64x4!(P2asin, -1.626247967210700244449E1);
    const_f64_as_f64x4!(P1asin, 1.956261983317594739197E1);
    const_f64_as_f64x4!(P0asin, -8.198089802484824371615E0);

    const_f64_as_f64x4!(Q4asin, -1.474091372988853791896E1);
    const_f64_as_f64x4!(Q3asin, 7.049610280856842141659E1);
    const_f64_as_f64x4!(Q2asin, -1.471791292232726029859E2);
    const_f64_as_f64x4!(Q1asin, 1.395105614657485689735E2);
    const_f64_as_f64x4!(Q0asin, -4.918853881490881290097E1);

    let xa = self.abs();

    let big = xa.simd_ge(f64x4::splat(0.625));

    let x1 = big.blend(f64x4::splat(1.0) - xa, xa * xa);

    let x2 = x1 * x1;
    let x3 = x2 * x1;
    let x4 = x2 * x2;
    let x5 = x4 * x1;

    let do_big = big.any();
    let do_small = !big.all();

    let mut rx = f64x4::default();
    let mut sx = f64x4::default();
    let mut px = f64x4::default();
    let mut qx = f64x4::default();

    if do_big {
      rx = x3.mul_add(R3asin, x2 * R2asin)
        + x4.mul_add(R4asin, x1.mul_add(R1asin, R0asin));
      sx =
        x3.mul_add(S3asin, x4) + x2.mul_add(S2asin, x1.mul_add(S1asin, S0asin));
    }

    if do_small {
      px = x3.mul_add(P3asin, P0asin)
        + x4.mul_add(P4asin, x1 * P1asin)
        + x5.mul_add(P5asin, x2 * P2asin);
      qx = x4.mul_add(Q4asin, x5)
        + x3.mul_add(Q3asin, x1 * Q1asin)
        + x2.mul_add(Q2asin, Q0asin);
    };

    let vx = big.blend(rx, px);
    let wx = big.blend(sx, qx);

    let y1 = vx / wx * x1;

    let mut z1 = f64x4::default();
    let mut z2 = f64x4::default();
    if do_big {
      let xb = (x1 + x1).sqrt();
      z1 = xb.mul_add(y1, xb);
    }

    if do_small {
      z2 = xa.mul_add(y1, xa);
    }

    // asin
    let z3 = f64x4::FRAC_PI_2 - z1;
    let asin = big.blend(z3, z2);
    let asin = asin.flip_signs(self);

    // acos
    let z3 = self.simd_lt(f64x4::ZERO).blend(f64x4::PI - z1, z1);
    let z4 = f64x4::FRAC_PI_2 - z2.flip_signs(self);
    let acos = big.blend(z3, z4);

    (asin, acos)
  }

  #[inline]
  pub fn acos(self) -> Self {
    // Based on the Agner Fog "vector class library":
    // https://github.com/vectorclass/version2/blob/master/vectormath_trig.h
    const_f64_as_f64x4!(R4asin, 2.967721961301243206100E-3);
    const_f64_as_f64x4!(R3asin, -5.634242780008963776856E-1);
    const_f64_as_f64x4!(R2asin, 6.968710824104713396794E0);
    const_f64_as_f64x4!(R1asin, -2.556901049652824852289E1);
    const_f64_as_f64x4!(R0asin, 2.853665548261061424989E1);

    const_f64_as_f64x4!(S3asin, -2.194779531642920639778E1);
    const_f64_as_f64x4!(S2asin, 1.470656354026814941758E2);
    const_f64_as_f64x4!(S1asin, -3.838770957603691357202E2);
    const_f64_as_f64x4!(S0asin, 3.424398657913078477438E2);

    const_f64_as_f64x4!(P5asin, 4.253011369004428248960E-3);
    const_f64_as_f64x4!(P4asin, -6.019598008014123785661E-1);
    const_f64_as_f64x4!(P3asin, 5.444622390564711410273E0);
    const_f64_as_f64x4!(P2asin, -1.626247967210700244449E1);
    const_f64_as_f64x4!(P1asin, 1.956261983317594739197E1);
    const_f64_as_f64x4!(P0asin, -8.198089802484824371615E0);

    const_f64_as_f64x4!(Q4asin, -1.474091372988853791896E1);
    const_f64_as_f64x4!(Q3asin, 7.049610280856842141659E1);
    const_f64_as_f64x4!(Q2asin, -1.471791292232726029859E2);
    const_f64_as_f64x4!(Q1asin, 1.395105614657485689735E2);
    const_f64_as_f64x4!(Q0asin, -4.918853881490881290097E1);

    let xa = self.abs();

    let big = xa.simd_ge(f64x4::splat(0.625));

    let x1 = big.blend(f64x4::splat(1.0) - xa, xa * xa);

    let x2 = x1 * x1;
    let x3 = x2 * x1;
    let x4 = x2 * x2;
    let x5 = x4 * x1;

    let do_big = big.any();
    let do_small = !big.all();

    let mut rx = f64x4::default();
    let mut sx = f64x4::default();
    let mut px = f64x4::default();
    let mut qx = f64x4::default();

    if do_big {
      rx = x3.mul_add(R3asin, x2 * R2asin)
        + x4.mul_add(R4asin, x1.mul_add(R1asin, R0asin));
      sx =
        x3.mul_add(S3asin, x4) + x2.mul_add(S2asin, x1.mul_add(S1asin, S0asin));
    }
    if do_small {
      px = x3.mul_add(P3asin, P0asin)
        + x4.mul_add(P4asin, x1 * P1asin)
        + x5.mul_add(P5asin, x2 * P2asin);
      qx = x4.mul_add(Q4asin, x5)
        + x3.mul_add(Q3asin, x1 * Q1asin)
        + x2.mul_add(Q2asin, Q0asin);
    };

    let vx = big.blend(rx, px);
    let wx = big.blend(sx, qx);

    let y1 = vx / wx * x1;

    let mut z1 = f64x4::default();
    let mut z2 = f64x4::default();
    if do_big {
      let xb = (x1 + x1).sqrt();
      z1 = xb.mul_add(y1, xb);
    }

    if do_small {
      z2 = xa.mul_add(y1, xa);
    }

    // acos
    let z3 = self.simd_lt(f64x4::ZERO).blend(f64x4::PI - z1, z1);
    let z4 = f64x4::FRAC_PI_2 - z2.flip_signs(self);
    let acos = big.blend(z3, z4);

    acos
  }
  #[inline]
  #[must_use]
  pub fn asin(self) -> Self {
    // Based on the Agner Fog "vector class library":
    // https://github.com/vectorclass/version2/blob/master/vectormath_trig.h
    const_f64_as_f64x4!(R4asin, 2.967721961301243206100E-3);
    const_f64_as_f64x4!(R3asin, -5.634242780008963776856E-1);
    const_f64_as_f64x4!(R2asin, 6.968710824104713396794E0);
    const_f64_as_f64x4!(R1asin, -2.556901049652824852289E1);
    const_f64_as_f64x4!(R0asin, 2.853665548261061424989E1);

    const_f64_as_f64x4!(S3asin, -2.194779531642920639778E1);
    const_f64_as_f64x4!(S2asin, 1.470656354026814941758E2);
    const_f64_as_f64x4!(S1asin, -3.838770957603691357202E2);
    const_f64_as_f64x4!(S0asin, 3.424398657913078477438E2);

    const_f64_as_f64x4!(P5asin, 4.253011369004428248960E-3);
    const_f64_as_f64x4!(P4asin, -6.019598008014123785661E-1);
    const_f64_as_f64x4!(P3asin, 5.444622390564711410273E0);
    const_f64_as_f64x4!(P2asin, -1.626247967210700244449E1);
    const_f64_as_f64x4!(P1asin, 1.956261983317594739197E1);
    const_f64_as_f64x4!(P0asin, -8.198089802484824371615E0);

    const_f64_as_f64x4!(Q4asin, -1.474091372988853791896E1);
    const_f64_as_f64x4!(Q3asin, 7.049610280856842141659E1);
    const_f64_as_f64x4!(Q2asin, -1.471791292232726029859E2);
    const_f64_as_f64x4!(Q1asin, 1.395105614657485689735E2);
    const_f64_as_f64x4!(Q0asin, -4.918853881490881290097E1);

    let xa = self.abs();

    let big = xa.simd_ge(f64x4::splat(0.625));

    let x1 = big.blend(f64x4::splat(1.0) - xa, xa * xa);

    let x2 = x1 * x1;
    let x3 = x2 * x1;
    let x4 = x2 * x2;
    let x5 = x4 * x1;

    let do_big = big.any();
    let do_small = !big.all();

    let mut rx = f64x4::default();
    let mut sx = f64x4::default();
    let mut px = f64x4::default();
    let mut qx = f64x4::default();

    if do_big {
      rx = x3.mul_add(R3asin, x2 * R2asin)
        + x4.mul_add(R4asin, x1.mul_add(R1asin, R0asin));
      sx =
        x3.mul_add(S3asin, x4) + x2.mul_add(S2asin, x1.mul_add(S1asin, S0asin));
    }
    if do_small {
      px = x3.mul_add(P3asin, P0asin)
        + x4.mul_add(P4asin, x1 * P1asin)
        + x5.mul_add(P5asin, x2 * P2asin);
      qx = x4.mul_add(Q4asin, x5)
        + x3.mul_add(Q3asin, x1 * Q1asin)
        + x2.mul_add(Q2asin, Q0asin);
    };

    let vx = big.blend(rx, px);
    let wx = big.blend(sx, qx);

    let y1 = vx / wx * x1;

    let mut z1 = f64x4::default();
    let mut z2 = f64x4::default();
    if do_big {
      let xb = (x1 + x1).sqrt();
      z1 = xb.mul_add(y1, xb);
    }

    if do_small {
      z2 = xa.mul_add(y1, xa);
    }

    // asin
    let z3 = f64x4::FRAC_PI_2 - z1;
    let asin = big.blend(z3, z2);
    let asin = asin.flip_signs(self);

    asin
  }

  #[inline]
  pub fn atan(self) -> Self {
    // Based on the Agner Fog "vector class library":
    // https://github.com/vectorclass/version2/blob/master/vectormath_trig.h
    const_f64_as_f64x4!(MORE_BITS, 6.123233995736765886130E-17);
    const_f64_as_f64x4!(MORE_BITS_O2, 6.123233995736765886130E-17 * 0.5);
    const_f64_as_f64x4!(T3PO8, core::f64::consts::SQRT_2 + 1.0);

    const_f64_as_f64x4!(P4atan, -8.750608600031904122785E-1);
    const_f64_as_f64x4!(P3atan, -1.615753718733365076637E1);
    const_f64_as_f64x4!(P2atan, -7.500855792314704667340E1);
    const_f64_as_f64x4!(P1atan, -1.228866684490136173410E2);
    const_f64_as_f64x4!(P0atan, -6.485021904942025371773E1);

    const_f64_as_f64x4!(Q4atan, 2.485846490142306297962E1);
    const_f64_as_f64x4!(Q3atan, 1.650270098316988542046E2);
    const_f64_as_f64x4!(Q2atan, 4.328810604912902668951E2);
    const_f64_as_f64x4!(Q1atan, 4.853903996359136964868E2);
    const_f64_as_f64x4!(Q0atan, 1.945506571482613964425E2);

    let t = self.abs();

    // small:  t < 0.66
    // medium: t <= t <= 2.4142 (1+sqrt(2))
    // big:    t > 2.4142
    let notbig = t.simd_le(T3PO8);
    let notsmal = t.simd_ge(Self::splat(0.66));

    let mut s = notbig.blend(Self::FRAC_PI_4, Self::FRAC_PI_2);
    s = notsmal & s;
    let mut fac = notbig.blend(MORE_BITS_O2, MORE_BITS);
    fac = notsmal & fac;

    // small:  z = t / 1.0;
    // medium: z = (t-1.0) / (t+1.0);
    // big:    z = -1.0 / t;
    let mut a = notbig & t;
    a = notsmal.blend(a - Self::ONE, a);
    let mut b = notbig & Self::ONE;
    b = notsmal.blend(b + t, b);
    let z = a / b;

    let zz = z * z;

    let px = polynomial_4!(zz, P0atan, P1atan, P2atan, P3atan, P4atan);
    let qx = polynomial_5n!(zz, Q0atan, Q1atan, Q2atan, Q3atan, Q4atan);

    let mut re = (px / qx).mul_add(z * zz, z);
    re += s + fac;

    // get sign bit
    re = (self.is_sign_negative()).blend(-re, re);

    re
  }

  #[inline]
  pub fn atan2(self, x: Self) -> Self {
    // Based on the Agner Fog "vector class library":
    // https://github.com/vectorclass/version2/blob/master/vectormath_trig.h
    const_f64_as_f64x4!(MORE_BITS, 6.123233995736765886130E-17);
    const_f64_as_f64x4!(MORE_BITS_O2, 6.123233995736765886130E-17 * 0.5);
    const_f64_as_f64x4!(T3PO8, core::f64::consts::SQRT_2 + 1.0);

    const_f64_as_f64x4!(P4atan, -8.750608600031904122785E-1);
    const_f64_as_f64x4!(P3atan, -1.615753718733365076637E1);
    const_f64_as_f64x4!(P2atan, -7.500855792314704667340E1);
    const_f64_as_f64x4!(P1atan, -1.228866684490136173410E2);
    const_f64_as_f64x4!(P0atan, -6.485021904942025371773E1);

    const_f64_as_f64x4!(Q4atan, 2.485846490142306297962E1);
    const_f64_as_f64x4!(Q3atan, 1.650270098316988542046E2);
    const_f64_as_f64x4!(Q2atan, 4.328810604912902668951E2);
    const_f64_as_f64x4!(Q1atan, 4.853903996359136964868E2);
    const_f64_as_f64x4!(Q0atan, 1.945506571482613964425E2);

    let y = self;

    // move in first octant
    let x1 = x.abs();
    let y1 = y.abs();
    let swapxy = y1.simd_gt(x1);
    // swap x and y if y1 > x1
    let mut x2 = swapxy.blend(y1, x1);
    let mut y2 = swapxy.blend(x1, y1);

    // check for special case: x and y are both +/- INF
    let both_infinite = x.is_inf() & y.is_inf();
    if both_infinite.any() {
      let minus_one = -Self::ONE;
      x2 = both_infinite.blend(x2 & minus_one, x2);
      y2 = both_infinite.blend(y2 & minus_one, y2);
    }

    // x = y = 0 gives NAN here
    let t = y2 / x2;

    // small:  t < 0.66
    // medium: t <= t <= 2.4142 (1+sqrt(2))
    // big:    t > 2.4142
    let notbig = t.simd_le(T3PO8);
    let notsmal = t.simd_ge(Self::splat(0.66));

    let mut s = notbig.blend(Self::FRAC_PI_4, Self::FRAC_PI_2);
    s = notsmal & s;
    let mut fac = notbig.blend(MORE_BITS_O2, MORE_BITS);
    fac = notsmal & fac;

    // small:  z = t / 1.0;
    // medium: z = (t-1.0) / (t+1.0);
    // big:    z = -1.0 / t;
    let mut a = notbig & t;
    a = notsmal.blend(a - Self::ONE, a);
    let mut b = notbig & Self::ONE;
    b = notsmal.blend(b + t, b);
    let z = a / b;

    let zz = z * z;

    let px = polynomial_4!(zz, P0atan, P1atan, P2atan, P3atan, P4atan);
    let qx = polynomial_5n!(zz, Q0atan, Q1atan, Q2atan, Q3atan, Q4atan);

    let mut re = (px / qx).mul_add(z * zz, z);
    re += s + fac;

    // move back in place
    re = swapxy.blend(Self::FRAC_PI_2 - re, re);
    re = ((x | y).simd_eq(Self::ZERO)).blend(Self::ZERO, re);
    re = (x.is_sign_negative()).blend(Self::PI - re, re);

    // get sign bit
    re = (y.is_sign_negative()).blend(-re, re);

    re
  }

  #[inline]
  #[must_use]
  pub fn sin_cos(self) -> (Self, Self) {
    // Based on the Agner Fog "vector class library":
    // https://github.com/vectorclass/version2/blob/master/vectormath_trig.h

    const_f64_as_f64x4!(P0sin, -1.66666666666666307295E-1);
    const_f64_as_f64x4!(P1sin, 8.33333333332211858878E-3);
    const_f64_as_f64x4!(P2sin, -1.98412698295895385996E-4);
    const_f64_as_f64x4!(P3sin, 2.75573136213857245213E-6);
    const_f64_as_f64x4!(P4sin, -2.50507477628578072866E-8);
    const_f64_as_f64x4!(P5sin, 1.58962301576546568060E-10);

    const_f64_as_f64x4!(P0cos, 4.16666666666665929218E-2);
    const_f64_as_f64x4!(P1cos, -1.38888888888730564116E-3);
    const_f64_as_f64x4!(P2cos, 2.48015872888517045348E-5);
    const_f64_as_f64x4!(P3cos, -2.75573141792967388112E-7);
    const_f64_as_f64x4!(P4cos, 2.08757008419747316778E-9);
    const_f64_as_f64x4!(P5cos, -1.13585365213876817300E-11);

    const_f64_as_f64x4!(DP1, 7.853981554508209228515625E-1 * 2.);
    const_f64_as_f64x4!(DP2, 7.94662735614792836714E-9 * 2.);
    const_f64_as_f64x4!(DP3, 3.06161699786838294307E-17 * 2.);

    const_f64_as_f64x4!(TWO_OVER_PI, 2.0 / core::f64::consts::PI);

    let xa = self.abs();

    let y = (xa * TWO_OVER_PI).round();
    let q = y.round_int();

    let x = y.mul_neg_add(DP3, y.mul_neg_add(DP2, y.mul_neg_add(DP1, xa)));

    let x2 = x * x;
    let mut s = polynomial_5!(x2, P0sin, P1sin, P2sin, P3sin, P4sin, P5sin);
    let mut c = polynomial_5!(x2, P0cos, P1cos, P2cos, P3cos, P4cos, P5cos);
    s = (x * x2).mul_add(s, x);
    c =
      (x2 * x2).mul_add(c, x2.mul_neg_add(f64x4::from(0.5), f64x4::from(1.0)));

    let swap = !((q & i64x4::from(1)).simd_eq(i64x4::from(0)));

    let mut overflow: f64x4 = cast(q.simd_gt(i64x4::from(0x80000000000000)));
    overflow &= xa.is_finite();
    s = overflow.blend(f64x4::from(0.0), s);
    c = overflow.blend(f64x4::from(1.0), c);

    // calc sin
    let mut sin1 = cast::<_, f64x4>(swap).blend(c, s);
    let sign_sin: i64x4 = (q << 62) ^ cast::<_, i64x4>(self);
    sin1 = sin1.flip_signs(cast(sign_sin));

    // calc cos
    let mut cos1 = cast::<_, f64x4>(swap).blend(s, c);
    let sign_cos: i64x4 = ((q + i64x4::from(1)) & i64x4::from(2)) << 62;
    cos1 ^= cast::<_, f64x4>(sign_cos);

    // IEEE 754: sin/cos(±∞) = NaN, sin/cos(NaN) = NaN
    let finite = self.is_finite();
    let nan = Self::splat(f64::NAN);
    let sin_final = finite.blend(sin1, nan);
    let cos_final = finite.blend(cos1, nan);

    (sin_final, cos_final)
  }
  #[inline]
  #[must_use]
  pub fn sin(self) -> Self {
    let (s, _) = self.sin_cos();
    s
  }
  #[inline]
  #[must_use]
  pub fn cos(self) -> Self {
    let (_, c) = self.sin_cos();
    c
  }
  #[inline]
  #[must_use]
  pub fn tan(self) -> Self {
    let (s, c) = self.sin_cos();
    s / c
  }

  /// Calculates hyperbolic sine: `(e^self - e^(-self))/2`.
  #[inline]
  #[must_use]
  pub fn sinh(self) -> Self {
    const_f64_as_f64x4!(P0, 1.0);
    const_f64_as_f64x4!(P1, 1.0 / 6.0);
    const_f64_as_f64x4!(P2, 1.0 / 120.0);
    const_f64_as_f64x4!(P3, 1.0 / 5040.0);
    const_f64_as_f64x4!(P4, 1.0 / 362880.0);
    const_f64_as_f64x4!(P5, 1.0 / 39916800.0);
    const_f64_as_f64x4!(P6, 1.0 / 6227020800.0);
    let a = self.abs();
    // |x| < 0.5: Taylor poly; last truncation term < 1 ULP at x=0.5 for both types
    let small = a.simd_lt(f64x4::from(0.5));
    let t = a * a;
    let poly = a * polynomial_6!(t, P0, P1, P2, P3, P4, P5, P6);
    let exp_based = {
      let e = a.exp();
      (e - Self::ONE / e) * Self::HALF
    };
    let result = small.blend(poly, exp_based);
    result.flip_signs(self)
  }

  /// Calculates hyperbolic cosine: `(e^self + e^(-self))/2`.
  #[inline]
  #[must_use]
  pub fn cosh(self) -> Self {
    const_f64_as_f64x4!(P0, 1.0);
    const_f64_as_f64x4!(P1, 1.0 / 2.0);
    const_f64_as_f64x4!(P2, 1.0 / 24.0);
    const_f64_as_f64x4!(P3, 1.0 / 720.0);
    const_f64_as_f64x4!(P4, 1.0 / 40320.0);
    const_f64_as_f64x4!(P5, 1.0 / 3628800.0);
    const_f64_as_f64x4!(P6, 1.0 / 479001600.0);
    const_f64_as_f64x4!(P7, 1.0 / 87178291200.0);
    let a = self.abs();
    // |x| < 0.5: Taylor poly; last truncation term < 1 ULP at x=0.5 for both types
    let small = a.simd_lt(f64x4::from(0.5));
    let t = a * a;
    let poly = polynomial_7!(t, P0, P1, P2, P3, P4, P5, P6, P7);
    let exp_based = {
      let e = a.exp();
      (e + Self::ONE / e) * Self::HALF
    };
    small.blend(poly, exp_based)
  }

  /// Calculates hyperbolic tangent: `sinh(self)/cosh(self)`.
  #[inline]
  #[must_use]
  pub fn tanh(self) -> Self {
    // |x| < 5e-8: tanh(x) ≈ x, error x³/3 < 16·ULP(x)
    // bound: x² < 48·2⁻⁵² → x < 1.03e-7; 5e-8 has 2× margin
    // |x| > 19.062: tanh(x) = ±1 to f64 precision (e⁻²ˣ < 2⁻⁵⁴)
    let a = self.abs();
    let large = a.simd_gt(f64x4::from(19.062));
    if large.all() {
      return Self::ONE.flip_signs(self);
    }
    let small = a.simd_lt(f64x4::from(5e-8));
    let exp_based = {
      let t = (Self::from(-2.0) * a).exp_m1();
      let pos = -t / (t + Self::from(2.0));
      pos.flip_signs(self)
    };
    let result = small.blend(self, exp_based);
    large.blend(Self::ONE.flip_signs(self), result)
  }

  /// Calculates the cube root: `self^(1/3)`.
  #[inline]
  #[must_use]
  pub fn cbrt(self) -> Self {
    let a = self.abs();
    let zero = a.simd_eq(Self::ZERO);
    if zero.all() {
      return self; // preserves -0.0
    }
    let inf = a.is_inf();
    let nan = self.is_nan();

    const SUBN_SCALE: f64 = 1.8014398509481984e16;
    const SUBN_CBRT: f64 = 262144.0;
    let tiny = a.simd_lt(Self::from(f64::MIN_POSITIVE));
    let a = tiny.blend(a * Self::from(SUBN_SCALE), a);

    let e = Self::exponent(a) + Self::ONE;
    let d = Self::fraction_2(a);

    // C0..C5 from SLEEF's minimax polynomial for 1/cbrt(d) on [0.5, 1.0)
    // Naoki Shibata et al., "SLEEF: A Portable Vectorized Library of C99
    // Mathematical Functions", https://sleef.org / https://github.com/shibatch/sleef
    // Licensed under the Boost Software License 1.0.
    const_f64_as_f64x4!(C0, 2.2307275302496609725722);
    const_f64_as_f64x4!(C1, -3.85841935510444988821632);
    const_f64_as_f64x4!(C2, 6.03990368989458747961407);
    const_f64_as_f64x4!(C3, -5.73353060922947843636166);
    const_f64_as_f64x4!(C4, 2.96155103020039511818595);
    const_f64_as_f64x4!(C5, -0.640245898480692909870982);
    let mut x = polynomial_5!(d, C0, C1, C2, C3, C4, C5);

    // Newton for 1/cbrt: x = x - (d * x^4 - x) / 3.
    let x2 = x * x;
    let x4 = x2 * x2;
    x = x - d.mul_add(x4, -x) * Self::from(1.0 / 3.0);

    // cbrt(d) = d * x^2, then polish
    let mut y = (d * x) * x;
    let yx = y * x;
    let t = Self::from(2.0 / 3.0);
    y = y - t * y * (yx - Self::ONE);

    // Scale by 2^(e/3) = 2^k * 2^(r/3)
    let three = Self::from(3.0);
    let two = Self::from(2.0);
    let neg = e.simd_lt(Self::ZERO);
    let e_adj = neg.blend(e - two, e);
    let k = (e_adj / three).trunc();
    let r = e - three * k;
    const_f64_as_f64x4!(CBRT2, 1.2599210498948732);
    const_f64_as_f64x4!(CBRT4, 1.5874010519681994);
    y = r.simd_eq(Self::ONE).blend(y * CBRT2, y);
    y = r.simd_eq(two).blend(y * CBRT4, y);
    y *= Self::vm_pow2n(k);
    y = tiny.blend(y / Self::from(SUBN_CBRT), y);

    let result = y.flip_signs(self);
    let result = nan.blend(self, result);
    let result = zero.blend(self, result);
    let result = inf.blend(self, result);
    result
  }

  #[inline]
  #[must_use]
  pub fn to_degrees(self) -> Self {
    const_f64_as_f64x4!(RAD_TO_DEG_RATIO, 180.0_f64 / core::f64::consts::PI);
    self * RAD_TO_DEG_RATIO
  }
  #[inline]
  #[must_use]
  pub fn to_radians(self) -> Self {
    const_f64_as_f64x4!(DEG_TO_RAD_RATIO, core::f64::consts::PI / 180.0_f64);
    self * DEG_TO_RAD_RATIO
  }

  #[inline]
  #[must_use]
  pub fn recip(self) -> Self {
    // There does not seem to be a `recip` intrinsic for any architecture. The
    // closest is `_mm256_rcp14_pd` which has relative error.
    Self::ONE / self
  }

  #[inline]
  #[must_use]
  pub fn recip_sqrt(self) -> Self {
    // There does not seem to be a `recip_sqrt` intrinsic for any architecture.
    // The closest is `_mm256_rsqrt14_pd` which has relative error.
    Self::ONE / self.sqrt()
  }

  #[inline]
  #[must_use]
  pub fn sqrt(self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: sqrt_m256d(self.avx) }
      } else {
        Self {
          a : self.a.sqrt(),
          b : self.b.sqrt(),
        }
      }
    }
  }
  #[inline]
  #[must_use]
  #[doc(alias("movemask", "move_mask"))]
  pub fn to_bitmask(self) -> u32 {
    pick! {
      if #[cfg(target_feature="avx")] {
        move_mask_m256d(self.avx) as u32
      } else {
        (self.b.to_bitmask() << 2) | self.a.to_bitmask()
      }
    }
  }
  #[inline]
  #[must_use]
  pub fn any(self) -> bool {
    pick! {
      if #[cfg(target_feature="avx")] {
        move_mask_m256d(self.avx) != 0
      } else {
        self.a.any() || self.b.any()
      }
    }
  }
  #[inline]
  #[must_use]
  pub fn all(self) -> bool {
    pick! {
      if #[cfg(target_feature="avx")] {
        move_mask_m256d(self.avx) == 0b1111
      } else {
        self.a.all() && self.b.all()
      }
    }
  }
  #[inline]
  #[must_use]
  pub fn none(self) -> bool {
    !self.any()
  }

  #[inline]
  fn vm_pow2n(self) -> Self {
    const_f64_as_f64x4!(pow2_52, 4503599627370496.0);
    const_f64_as_f64x4!(bias, 1023.0);
    let a = self + (bias + pow2_52);
    let c = cast::<_, i64x4>(a) << 52;
    let std_result = cast::<_, f64x4>(c);

    let min_exp = f64x4::from(-1022.0);
    let is_sub = self.simd_lt(min_exp);
    if is_sub.any() {
      let valid = self.simd_ge(f64x4::from(-1074.0));
      let shift_f = self + f64x4::from(1074.0);
      let mut shift_i = shift_f.trunc_int();
      shift_i = cast::<_, i64x4>(valid).blend(shift_i, i64x4::ZERO);
      let mantissa = i64x4::ONE << shift_i;
      let sub_result = cast::<_, f64x4>(mantissa);
      let sub_result = valid.blend(sub_result, f64x4::ZERO);
      is_sub.blend(sub_result, std_result)
    } else {
      std_result
    }
  }

  /// Calculate the exponent of a packed `f64x4`
  #[inline]
  #[must_use]
  pub fn exp(self) -> Self {
    const_f64_as_f64x4!(P2, 1.0 / 2.0);
    const_f64_as_f64x4!(P3, 1.0 / 6.0);
    const_f64_as_f64x4!(P4, 1.0 / 24.0);
    const_f64_as_f64x4!(P5, 1.0 / 120.0);
    const_f64_as_f64x4!(P6, 1.0 / 720.0);
    const_f64_as_f64x4!(P7, 1.0 / 5040.0);
    const_f64_as_f64x4!(P8, 1.0 / 40320.0);
    const_f64_as_f64x4!(P9, 1.0 / 362880.0);
    const_f64_as_f64x4!(P10, 1.0 / 3628800.0);
    const_f64_as_f64x4!(P11, 1.0 / 39916800.0);
    const_f64_as_f64x4!(P12, 1.0 / 479001600.0);
    const_f64_as_f64x4!(P13, 1.0 / 6227020800.0);
    // LN2D_HI/LO: double-double decomposition of ln(2) for exp range reduction,
    // following fdlibm's approach (Sun Microsystems, https://www.netlib.org/fdlibm/ e_exp.c).
    // Values chosen so LN2D_HI + LN2D_LO = ln(2) to full f64 precision.
    const_f64_as_f64x4!(LN2D_HI, 0.693145751953125);
    const_f64_as_f64x4!(LN2D_LO, 1.42860682030941723212E-6);
    let max_x = f64x4::from(709.783);
    let min_x = f64x4::from(-744.79);
    let finite = self.is_finite();
    // x < min_x: e^x underflows to 0 -- skip the entire pipeline
    let neg_underflow = self.simd_lt(min_x) & finite;
    if neg_underflow.all() {
      return Self::ZERO;
    }
    let max_r = f64x4::from(1023.0);
    let r = (self * Self::LOG2_E).round();
    let big = r.simd_gt(max_r);
    let r_safe = big.blend(max_r, r);
    let excess = r - max_r;
    let excess = big.blend(excess, Self::ZERO);
    let scale = Self::vm_pow2n(excess);
    let x = r.mul_neg_add(LN2D_HI, self);
    let x = r.mul_neg_add(LN2D_LO, x);
    let z =
      polynomial_13!(x, P2, P3, P4, P5, P6, P7, P8, P9, P10, P11, P12, P13);
    let n2 = Self::vm_pow2n(r_safe);
    let z = (z + Self::ONE) * scale * n2;
    let nan_mask = self.is_nan();
    let mut result = nan_mask.blend(Self::nan_pow(), z);
    let pos_overflow = self.simd_gt(max_x) & finite;
    result = pos_overflow.blend(Self::infinity(), result);
    result = neg_underflow.blend(Self::ZERO, result);
    let pos_inf = !finite & !self.is_sign_negative() & !nan_mask;
    result = pos_inf.blend(Self::infinity(), result);
    let neg_inf = !finite & self.is_sign_negative() & !nan_mask;
    result = neg_inf.blend(Self::ZERO, result);
    result
  }

  /// Calculate `e^self - 1` for each lane.
  /// Accurate even for very small values.
  #[inline]
  #[must_use]
  pub fn exp_m1(self) -> Self {
    const_f64_as_f64x4!(P2, 1.0 / 2.0);
    const_f64_as_f64x4!(P3, 1.0 / 6.0);
    const_f64_as_f64x4!(P4, 1.0 / 24.0);
    const_f64_as_f64x4!(P5, 1.0 / 120.0);
    const_f64_as_f64x4!(P6, 1.0 / 720.0);
    const_f64_as_f64x4!(P7, 1.0 / 5040.0);
    const_f64_as_f64x4!(P8, 1.0 / 40320.0);
    const_f64_as_f64x4!(P9, 1.0 / 362880.0);
    const_f64_as_f64x4!(P10, 1.0 / 3628800.0);
    const_f64_as_f64x4!(P11, 1.0 / 39916800.0);
    const_f64_as_f64x4!(P12, 1.0 / 479001600.0);
    const_f64_as_f64x4!(P13, 1.0 / 6227020800.0);
    // LN2D_HI/LO: double-double decomposition of ln(2) for exp range reduction,
    // following fdlibm's approach (Sun Microsystems, https://www.netlib.org/fdlibm/ e_exp.c).
    const_f64_as_f64x4!(LN2D_HI, 0.693145751953125);
    const_f64_as_f64x4!(LN2D_LO, 1.42860682030941723212E-6);
    // x < -37.429: e^x < 2⁻⁵⁴, exp_m1(x) = -1.0 exactly (mantissa exhaustion)
    // IEEE simd_lt returns false for NaN, so NaN lanes can't reach here.
    // -inf is < -37.429, and exp_m1(-inf) = -1.0, also correct.
    if self.simd_lt(f64x4::from(-37.429)).all() {
      return f64x4::from(-1.0);
    }
    // max_x = ln(f64::MAX) ≈ 709.7827129, max_r = 1023 (IEEE max normal
    // exponent) min_x = -1074.5 ln(2) ≈ -744.79: min r for vm_pow2n to
    // construct subnormal
    let max_x = f64x4::from(709.783);
    let min_x = f64x4::from(-744.79);
    let max_r = f64x4::from(1023.0);
    let r = (self * Self::LOG2_E).round();
    let big = r.simd_gt(max_r);
    let r_safe = big.blend(max_r, r);
    let excess = r - max_r;
    let excess = big.blend(excess, Self::ZERO);
    let scale = Self::vm_pow2n(excess);
    let x = r.mul_neg_add(LN2D_HI, self);
    let x = r.mul_neg_add(LN2D_LO, x);
    let z =
      polynomial_13!(x, P2, P3, P4, P5, P6, P7, P8, P9, P10, P11, P12, P13);
    let n2 = Self::vm_pow2n(r_safe);
    let exp_val = (z + Self::ONE) * scale * n2;
    // When r == 0, z is already e^x - 1 from the Taylor poly.
    // Computing (z+1) - 1 would lose low bits for small x (catastrophic
    // cancellation at z ~ 0), so keep z directly.
    let r_is_zero = r.simd_eq(Self::ZERO);
    let z = r_is_zero.blend(z, exp_val - Self::ONE);
    let nan_mask = self.is_nan();
    let finite = self.is_finite();
    let mut result = nan_mask.blend(Self::nan_pow(), z);
    let pos_overflow = self.simd_gt(max_x) & finite;
    result = pos_overflow.blend(Self::infinity(), result);
    let neg_underflow = self.simd_lt(min_x) & finite;
    result = neg_underflow.blend(-Self::ONE, result);
    let pos_inf = !finite & !self.is_sign_negative() & !nan_mask;
    result = pos_inf.blend(Self::infinity(), result);
    let neg_inf = !finite & self.is_sign_negative() & !nan_mask;
    result = neg_inf.blend(-Self::ONE, result);
    let is_zero = self.simd_eq(Self::ZERO);
    result = is_zero.blend(self, result);
    result
  }

  /// Returns `2^self`.
  #[inline]
  #[must_use]
  pub fn exp2(self) -> Self {
    const_f64_as_f64x4!(P2, 1.0 / 2.0);
    const_f64_as_f64x4!(P3, 1.0 / 6.0);
    const_f64_as_f64x4!(P4, 1.0 / 24.0);
    const_f64_as_f64x4!(P5, 1.0 / 120.0);
    const_f64_as_f64x4!(P6, 1.0 / 720.0);
    const_f64_as_f64x4!(P7, 1.0 / 5040.0);
    const_f64_as_f64x4!(P8, 1.0 / 40320.0);
    const_f64_as_f64x4!(P9, 1.0 / 362880.0);
    const_f64_as_f64x4!(P10, 1.0 / 3628800.0);

    // max_x = log2(f64::MAX) ≈ 1023.9999999999999
    // min_x = log2(f64::MIN_POSITIVE) - 52 ≈ -1022 - 52 = -1074
    let max_x = f64x4::from(1023.9999999999999);
    let min_x = f64x4::from(-1074.5);
    let finite = self.is_finite();
    let neg_underflow = self.simd_lt(min_x) & finite;
    if neg_underflow.all() {
      return Self::ZERO;
    }

    let round = self.round();
    let max_r = f64x4::from(1023.0);
    let big = round.simd_gt(max_r);
    let r_safe = big.blend(max_r, round);
    let excess = round - max_r;
    let excess = big.blend(excess, Self::ZERO);
    let scale = Self::vm_pow2n(excess);

    let fract = (self - round) * Self::LN_2;
    let fract_partial_exp2 =
      polynomial_8!(fract, P2, P3, P4, P5, P6, P7, P8, P9, P10);
    let fract2 = fract * fract;
    let fract_exp2 = fract_partial_exp2.mul_add(fract2, fract) + Self::ONE;

    let n2 = Self::vm_pow2n(r_safe);
    let result = fract_exp2 * scale * n2;

    let nan_mask = self.is_nan();
    let mut result = nan_mask.blend(Self::nan_pow(), result);
    let pos_overflow = self.simd_gt(max_x) & finite;
    result = pos_overflow.blend(Self::infinity(), result);
    result = neg_underflow.blend(Self::ZERO, result);
    let pos_inf = !finite & !self.is_sign_negative() & !nan_mask;
    result = pos_inf.blend(Self::infinity(), result);
    let neg_inf = !finite & self.is_sign_negative() & !nan_mask;
    result = neg_inf.blend(Self::ZERO, result);
    result
  }

  #[inline]
  fn exponent(self) -> f64x4 {
    const_f64_as_f64x4!(pow2_52, 4503599627370496.0);
    const_f64_as_f64x4!(bias, 1023.0);
    let a = cast::<_, u64x4>(self);
    let b = a >> 52;
    let c = b | cast::<_, u64x4>(pow2_52);
    let d = cast::<_, f64x4>(c);
    let e = d - (pow2_52 + bias);
    e
  }

  #[inline]
  fn fraction_2(self) -> Self {
    let t1 = cast::<_, u64x4>(self);
    let t2 = cast::<_, u64x4>(
      (t1 & u64x4::from(0x000FFFFFFFFFFFFF)) | u64x4::from(0x3FE0000000000000),
    );
    cast::<_, f64x4>(t2)
  }
  #[inline]
  fn is_zero_or_subnormal(self) -> Self {
    let t = cast::<_, i64x4>(self);
    let t = t & i64x4::splat(0x7FF0000000000000);
    let mask = t.simd_eq(i64x4::splat(0));
    cast::<_, f64x4>(mask)
  }
  #[inline]
  fn infinity() -> Self {
    cast::<_, f64x4>(i64x4::splat(0x7FF0000000000000))
  }
  #[inline]
  fn nan_log() -> Self {
    cast::<_, f64x4>(i64x4::splat(0x7FF8000000000000 | 0x101 << 29))
  }
  #[inline]
  fn nan_pow() -> Self {
    cast::<_, f64x4>(i64x4::splat(0x7FF8000000000000 | 0x101 << 29))
  }

  /// Returns true for each element if it has a positive sign, including `+0.0`,
  /// `NaN`s with positive sign bit and positive infinity.
  #[inline]
  #[must_use]
  pub fn is_sign_positive(self) -> Self {
    const SIGN_MASK: u64x4 = u64x4::splat((-0.0_f64).to_bits());

    let bits = cast::<f64x4, u64x4>(self);
    let sign = bits & SIGN_MASK;
    let result = sign.simd_eq(u64x4::ZERO);
    cast::<u64x4, f64x4>(result)
  }

  /// Returns true for each element if it has a negative sign, including `-0.0`,
  /// `NaN`s with negative sign bit and negative infinity.
  #[inline]
  #[must_use]
  pub fn is_sign_negative(self) -> Self {
    const SIGN_MASK: u64x4 = u64x4::splat((-0.0_f64).to_bits());

    let bits = cast::<f64x4, u64x4>(self);
    let sign = bits & SIGN_MASK;
    let result = sign.simd_eq(SIGN_MASK);
    cast::<u64x4, f64x4>(result)
  }

  /// horizontal add of all the elements of the vector
  #[inline]
  pub fn reduce_add(self) -> f64 {
    pick! {
      if #[cfg(target_feature="avx")] {
        // From https://stackoverflow.com/questions/49941645/get-sum-of-values-stored-in-m256d-with-sse-avx
        let lo = cast_to_m128d_from_m256d(self.avx);
        let hi = extract_m128d_from_m256d::<1>(self.avx);
        let lo = add_m128d(lo,hi);
        let hi64 = unpack_high_m128d(lo,lo);
        let sum = add_m128d_s(lo,hi64);
        get_f64_from_m128d_s(sum)
      } else {
        self.a.reduce_add() + self.b.reduce_add()
      }
    }
  }

  /// horizontal multiplication of all the elements of the vector
  #[inline]
  #[must_use]
  pub fn reduce_mul(self) -> f64 {
    pick! {
      if #[cfg(target_feature="avx")] {
        // From https://stackoverflow.com/questions/49941645/get-sum-of-values-stored-in-m256d-with-sse-avx
        let lo = cast_to_m128d_from_m256d(self.avx);
        let hi = extract_m128d_from_m256d::<1>(self.avx);
        let lo = mul_m128d(lo,hi);
        let hi64 = unpack_high_m128d(lo,lo);
        let product = mul_m128d_s(lo,hi64);
        get_f64_from_m128d_s(product)
      } else {
        self.a.reduce_mul() * self.b.reduce_mul()
      }
    }
  }

  /// Natural log (ln(x))
  #[inline]
  #[must_use]
  pub fn ln(self) -> Self {
    const_f64_as_f64x4!(HALF, 0.5);
    const_f64_as_f64x4!(P0, 7.70838733755885391666E0);
    const_f64_as_f64x4!(P1, 1.79368678507819816313E1);
    const_f64_as_f64x4!(P2, 1.44989225341610930846E1);
    const_f64_as_f64x4!(P3, 4.70579119878881725854E0);
    const_f64_as_f64x4!(P4, 4.97494994976747001425E-1);
    const_f64_as_f64x4!(P5, 1.01875663804580931796E-4);

    const_f64_as_f64x4!(Q0, 2.31251620126765340583E1);
    const_f64_as_f64x4!(Q1, 7.11544750618563894466E1);
    const_f64_as_f64x4!(Q2, 8.29875266912776603211E1);
    const_f64_as_f64x4!(Q3, 4.52279145837532221105E1);
    const_f64_as_f64x4!(Q4, 1.12873587189167450590E1);
    // LN2F_HI/LO from fdlibm (Freely Distributable LIBM)
    // Sun Microsystems, Inc. https://www.netlib.org/fdlibm/
    // e_log.c: bit-exact double-double decomposition of ln(2) for f64.
    // Replaced the original f32-literals (0.693359375, -2.12194440e-4)
    // which had ~10 significant digits, causing ~630 ULP error in f64 ln.
    const_f64_as_f64x4!(LN2F_HI, f64::from_bits(0x3FE62E42FEE00000));
    const_f64_as_f64x4!(LN2F_LO, f64::from_bits(0x3DEA39EF35793C76));
    const_f64_as_f64x4!(VM_SQRT2, 1.414213562373095048801);
    const_f64_as_f64x4!(VM_SMALLEST_NORMAL, 2.2250738585072014E-308);

    let x1 = self;
    let x = Self::fraction_2(x1);
    let e = Self::exponent(x1);
    let mask = x.simd_gt(VM_SQRT2 * HALF);
    let x = (!mask).blend(x + x, x);
    let fe = mask.blend(e + Self::ONE, e);
    let x = x - Self::ONE;
    let px = polynomial_5!(x, P0, P1, P2, P3, P4, P5);
    let x2 = x * x;
    let px = x2 * x * px;
    let qx = polynomial_5n!(x, Q0, Q1, Q2, Q3, Q4);
    let res = px / qx;
    let res = fe.mul_add(LN2F_LO, res);
    let res = res + x2.mul_neg_add(HALF, x);
    let res = fe.mul_add(LN2F_HI, res);
    let overflow = !self.is_finite();
    let underflow = x1.simd_lt(VM_SMALLEST_NORMAL);
    let mask = overflow | underflow;
    if !mask.any() {
      res
    } else {
      let is_zero = self.is_zero_or_subnormal();
      let res = underflow.blend(Self::nan_log(), res);
      // Note: is_zero_or_subnormal() lumps subnormals (exponent==0) with zero.
      // Both get -Inf here. True subnormal inputs (~5e-324..2.225e-308) should
      // produce a finite negative result, but are vanishingly rare in
      // practice.
      let res = is_zero.blend(-Self::infinity(), res);
      let res = overflow.blend(self, res);
      // This must come *after* overflow.blend to overwrite ln(-∞) = -∞ to NaN
      let res = (!self.is_finite() & self.is_sign_negative())
        .blend(Self::nan_log(), res);
      res
    }
  }

  /// Calculate `ln(1 + self)` for each lane.
  /// Accurate even for very small values.
  #[inline]
  #[must_use]
  pub fn ln_1p(self) -> Self {
    // Based on the identity ln(1+x) = x·ln(1+x)/((1+x)-1), i.e. x·ln(u)/(u-1)
    // where u = 1+x. From MUSL libc (Rich Felker et al., https://musl.libc.org) src/math/log1p.c
    // and fdlibm (Sun Microsystems, https://www.netlib.org/fdlibm/) s_log1p.c.
    // When 1+x rounds to 1 exactly (subnormal x), return x directly.
    // When 1+x overflows (+inf), return ln(u) without correction.
    // Mathematically exact: compensates for the rounding loss in 1+x without
    // needing a series threshold.
    let u = self + Self::ONE;
    let eq = u.simd_eq(Self::ONE);
    let ln_u = Self::ln(u);
    let correction = self * (ln_u / (u - Self::ONE));
    let result = eq.blend(self, correction);
    let over = u.is_inf();
    over.blend(ln_u, result)
  }

  #[inline]
  #[must_use]
  pub fn log2(self) -> Self {
    Self::ln(self) * Self::LOG2_E
  }
  #[inline]
  #[must_use]
  pub fn log10(self) -> Self {
    Self::ln(self) * Self::LOG10_E
  }

  #[inline]
  #[must_use]
  pub fn pow_f64x4(self, y: Self) -> Self {
    const_f64_as_f64x4!(ln2d_hi, 0.693145751953125);
    const_f64_as_f64x4!(ln2d_lo, 1.42860682030941723212E-6);
    const_f64_as_f64x4!(P0log, 2.0039553499201281259648E1);
    const_f64_as_f64x4!(P1log, 5.7112963590585538103336E1);
    const_f64_as_f64x4!(P2log, 6.0949667980987787057556E1);
    const_f64_as_f64x4!(P3log, 2.9911919328553073277375E1);
    const_f64_as_f64x4!(P4log, 6.5787325942061044846969E0);
    const_f64_as_f64x4!(P5log, 4.9854102823193375972212E-1);
    const_f64_as_f64x4!(P6log, 4.5270000862445199635215E-5);
    const_f64_as_f64x4!(Q0log, 6.0118660497603843919306E1);
    const_f64_as_f64x4!(Q1log, 2.1642788614495947685003E2);
    const_f64_as_f64x4!(Q2log, 3.0909872225312059774938E2);
    const_f64_as_f64x4!(Q3log, 2.2176239823732856465394E2);
    const_f64_as_f64x4!(Q4log, 8.3047565967967209469434E1);
    const_f64_as_f64x4!(Q5log, 1.5062909083469192043167E1);

    // Taylor expansion constants
    const_f64_as_f64x4!(p2, 1.0 / 2.0); // coefficients for Taylor expansion of exp
    const_f64_as_f64x4!(p3, 1.0 / 6.0);
    const_f64_as_f64x4!(p4, 1.0 / 24.0);
    const_f64_as_f64x4!(p5, 1.0 / 120.0);
    const_f64_as_f64x4!(p6, 1.0 / 720.0);
    const_f64_as_f64x4!(p7, 1.0 / 5040.0);
    const_f64_as_f64x4!(p8, 1.0 / 40320.0);
    const_f64_as_f64x4!(p9, 1.0 / 362880.0);
    const_f64_as_f64x4!(p10, 1.0 / 3628800.0);
    const_f64_as_f64x4!(p11, 1.0 / 39916800.0);
    const_f64_as_f64x4!(p12, 1.0 / 479001600.0);
    const_f64_as_f64x4!(p13, 1.0 / 6227020800.0);

    let x1 = self.abs();
    let x = x1.fraction_2();
    let mask = x.simd_gt(f64x4::SQRT_2 * f64x4::HALF);
    let x = (!mask).blend(x + x, x);
    let x = x - f64x4::ONE;
    let x2 = x * x;
    let px = polynomial_6!(x, P0log, P1log, P2log, P3log, P4log, P5log, P6log);
    let px = px * x * x2;
    let qx = polynomial_6n!(x, Q0log, Q1log, Q2log, Q3log, Q4log, Q5log);
    let lg1 = px / qx;

    let ef = x1.exponent();
    let ef = mask.blend(ef + f64x4::ONE, ef);
    let e1 = (ef * y).round();
    let yr = ef.mul_sub(y, e1);

    let lg = f64x4::HALF.mul_neg_add(x2, x) + lg1;
    let x2err = (f64x4::HALF * x).mul_sub(x, f64x4::HALF * x2);
    let lg_err = f64x4::HALF.mul_add(x2, lg - x) - lg1;

    let e2 = (lg * y * f64x4::LOG2_E).round();
    let v = lg.mul_sub(y, e2 * ln2d_hi);
    let v = e2.mul_neg_add(ln2d_lo, v);
    let v = v - (lg_err + x2err).mul_sub(y, yr * f64x4::LN_2);

    let x = v;
    let e3 = (x * f64x4::LOG2_E).round();
    let x = e3.mul_neg_add(f64x4::LN_2, x);
    let z =
      polynomial_13!(x, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11, p12, p13)
        + f64x4::ONE;
    let ee = e1 + e2 + e3;
    let ei = cast::<_, i64x4>(ee.round_int());
    let ej = cast::<_, i64x4>(ei + (cast::<_, i64x4>(z) >> 52));

    let overflow = cast::<_, f64x4>(!ej.simd_lt(i64x4::splat(0x07FF)))
      | ee.simd_gt(f64x4::splat(3000.0));
    let underflow = cast::<_, f64x4>(!ej.simd_gt(i64x4::splat(0x000)))
      | ee.simd_lt(f64x4::splat(-3000.0));

    // Add exponent by integer addition
    let z = cast::<_, f64x4>(cast::<_, i64x4>(z) + (ei << 52));

    // Check for overflow/underflow
    let z = if (overflow | underflow).any() {
      let z = underflow.blend(f64x4::ZERO, z);
      overflow.blend(Self::infinity(), z)
    } else {
      z
    };

    // Check for self == 0
    let x_zero = self.is_zero_or_subnormal();
    let z = x_zero.blend(
      y.simd_lt(f64x4::ZERO).blend(
        Self::infinity(),
        y.simd_eq(f64x4::ZERO).blend(f64x4::ONE, f64x4::ZERO),
      ),
      z,
    );

    let x_sign = self.is_sign_negative();

    let z = if x_sign.any() {
      // Y into an integer
      let yi = y.simd_eq(y.round());
      // Is y odd?
      let y_odd = cast::<_, i64x4>(y.round_int() << 63).round_float();
      let z1 =
        yi.blend(z | y_odd, self.simd_eq(Self::ZERO).blend(z, Self::nan_pow()));
      x_sign.blend(z1, z)
    } else {
      z
    };

    let x_finite = self.is_finite();
    let y_finite = y.is_finite();
    let e_finite = ee.is_finite();

    if (x_finite & y_finite & (e_finite | x_zero)).all() {
      return z;
    }

    (self.is_nan() | y.is_nan()).blend(self + y, z)
  }
  #[inline]
  pub fn powf(self, y: f64) -> Self {
    Self::pow_f64x4(self, f64x4::splat(y))
  }

  // Sometimes used for `transpose`.
  #[must_use]
  #[inline]
  #[allow(dead_code)]
  pub(crate) fn unpack_lo(self, b: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        // `unpack_lo_m256d` cannot be used because it actually performs a
        // different operation.
        let [aa, _]: [f64x2; 2] = cast(self);
        let [ba, _]: [f64x2; 2] = cast(b);
        cast([aa.unpack_lo(ba), aa.unpack_hi(ba)])
      } else {
        Self { a: self.a.unpack_lo(b.a), b: self.a.unpack_hi(b.a) }
      }
    }
  }

  // Sometimes used for `transpose`.
  #[must_use]
  #[inline]
  #[allow(dead_code)]
  pub(crate) fn unpack_hi(self, b: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        // `unpack_hi_m256d` cannot be used because it actually performs a
        // different operation.
        let [_, ab]: [f64x2; 2] = cast(self);
        let [_, bb]: [f64x2; 2] = cast(b);
        cast([ab.unpack_lo(bb), ab.unpack_hi(bb)])
      } else {
        Self { a: self.b.unpack_lo(b.b), b: self.b.unpack_hi(b.b) }
      }
    }
  }

  /// Transpose matrix of 4x4 `f64` matrix.
  #[must_use]
  #[inline]
  pub fn transpose(data: [f64x4; 4]) -> [f64x4; 4] {
    pick! {
      if #[cfg(target_feature="avx")] {
        // Can this be optimized?
        let a = data[0].unpack_lo(data[2]);
        let b = data[1].unpack_lo(data[3]);
        let c = data[0].unpack_hi(data[2]);
        let d = data[1].unpack_hi(data[3]);
        [
          a.unpack_lo(b),
          a.unpack_hi(b),
          c.unpack_lo(d),
          c.unpack_hi(d),
        ]
      } else {
        #[inline(always)]
        fn transpose_column(data: &[f64x4; 4], index: usize) -> f64x4 {
          f64x4::new([
            data[0].as_array()[index],
            data[1].as_array()[index],
            data[2].as_array()[index],
            data[3].as_array()[index],
          ])
        }

        [
          transpose_column(&data, 0),
          transpose_column(&data, 1),
          transpose_column(&data, 2),
          transpose_column(&data, 3),
        ]
      }
    }
  }

  #[inline]
  pub fn to_array(self) -> [f64; 4] {
    cast(self)
  }

  #[inline]
  pub fn as_array(&self) -> &[f64; 4] {
    cast_ref(self)
  }

  #[inline]
  pub fn as_mut_array(&mut self) -> &mut [f64; 4] {
    cast_mut(self)
  }

  #[inline]
  pub fn from_i32x4(v: i32x4) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: convert_to_m256d_from_i32_m128i(v.sse) }
      } else {
        Self::new([
          v.as_array()[0] as f64,
          v.as_array()[1] as f64,
          v.as_array()[2] as f64,
          v.as_array()[3] as f64,
        ])
      }
    }
  }
}

impl From<i32x4> for f64x4 {
  #[inline]
  fn from(v: i32x4) -> Self {
    Self::from_i32x4(v)
  }
}

impl Not for f64x4 {
  type Output = Self;
  #[inline]
  fn not(self) -> Self {
    pick! {
      if #[cfg(target_feature="avx")] {
        Self { avx: self.avx.not()  }
      } else {
        Self {
          a : self.a.not(),
          b : self.b.not(),
        }
      }
    }
  }
}