use super::v128;
use crate::core_arch::simd;

#[cfg(test)]
use stdarch_test::assert_instr;

#[allow(improper_ctypes)]
extern "C" {
    #[link_name = "llvm.wasm.relaxed.swizzle"]
    fn llvm_relaxed_swizzle(a: simd::i8x16, b: simd::i8x16) -> simd::i8x16;
    #[link_name = "llvm.wasm.relaxed.trunc.signed"]
    fn llvm_relaxed_trunc_signed(a: simd::f32x4) -> simd::i32x4;
    #[link_name = "llvm.wasm.relaxed.trunc.unsigned"]
    fn llvm_relaxed_trunc_unsigned(a: simd::f32x4) -> simd::i32x4;
    #[link_name = "llvm.wasm.relaxed.trunc.signed.zero"]
    fn llvm_relaxed_trunc_signed_zero(a: simd::f64x2) -> simd::i32x4;
    #[link_name = "llvm.wasm.relaxed.trunc.unsigned.zero"]
    fn llvm_relaxed_trunc_unsigned_zero(a: simd::f64x2) -> simd::i32x4;

    #[link_name = "llvm.wasm.fma.v4f32"]
    fn llvm_f32x4_fma(a: simd::f32x4, b: simd::f32x4, c: simd::f32x4) -> simd::f32x4;
    #[link_name = "llvm.wasm.fms.v4f32"]
    fn llvm_f32x4_fms(a: simd::f32x4, b: simd::f32x4, c: simd::f32x4) -> simd::f32x4;
    #[link_name = "llvm.wasm.fma.v2f64"]
    fn llvm_f64x2_fma(a: simd::f64x2, b: simd::f64x2, c: simd::f64x2) -> simd::f64x2;
    #[link_name = "llvm.wasm.fms.v2f64"]
    fn llvm_f64x2_fms(a: simd::f64x2, b: simd::f64x2, c: simd::f64x2) -> simd::f64x2;

    #[link_name = "llvm.wasm.laneselect.v16i8"]
    fn llvm_i8x16_laneselect(a: simd::i8x16, b: simd::i8x16, c: simd::i8x16) -> simd::i8x16;
    #[link_name = "llvm.wasm.laneselect.v8i16"]
    fn llvm_i16x8_laneselect(a: simd::i16x8, b: simd::i16x8, c: simd::i16x8) -> simd::i16x8;
    #[link_name = "llvm.wasm.laneselect.v4i32"]
    fn llvm_i32x4_laneselect(a: simd::i32x4, b: simd::i32x4, c: simd::i32x4) -> simd::i32x4;
    #[link_name = "llvm.wasm.laneselect.v2i64"]
    fn llvm_i64x2_laneselect(a: simd::i64x2, b: simd::i64x2, c: simd::i64x2) -> simd::i64x2;

    #[link_name = "llvm.wasm.relaxed.min.v4f32"]
    fn llvm_f32x4_relaxed_min(a: simd::f32x4, b: simd::f32x4) -> simd::f32x4;
    #[link_name = "llvm.wasm.relaxed.min.v2f64"]
    fn llvm_f64x2_relaxed_min(a: simd::f64x2, b: simd::f64x2) -> simd::f64x2;
    #[link_name = "llvm.wasm.relaxed.max.v4f32"]
    fn llvm_f32x4_relaxed_max(a: simd::f32x4, b: simd::f32x4) -> simd::f32x4;
    #[link_name = "llvm.wasm.relaxed.max.v2f64"]
    fn llvm_f64x2_relaxed_max(a: simd::f64x2, b: simd::f64x2) -> simd::f64x2;

    #[link_name = "llvm.wasm.relaxed.q15mulr.signed"]
    fn llvm_relaxed_q15mulr_signed(a: simd::i16x8, b: simd::i16x8) -> simd::i16x8;
    #[link_name = "llvm.wasm.dot.i8x16.i7x16.signed"]
    fn llvm_i16x8_relaxed_dot_i8x16_i7x16_s(a: simd::i8x16, b: simd::i8x16) -> simd::i16x8;
    #[link_name = "llvm.wasm.dot.i8x16.i7x16.add.signed"]
    fn llvm_i32x4_relaxed_dot_i8x16_i7x16_add_s(
        a: simd::i8x16,
        b: simd::i8x16,
        c: simd::i32x4,
    ) -> simd::i32x4;
}

/// A relaxed version of `i8x16_swizzle(a, s)` which selects lanes from `a`
/// using indices in `s`.
///
/// Indices in the range `[0,15]` will select the `i`-th element of `a`.
/// If the high bit of any element of `s` is set (meaning 128 or greater) then
/// the corresponding output lane is guaranteed to be zero. Otherwise if the
/// element of `s` is within the range `[16,128)` then the output lane is either
/// 0 or `a[s[i] % 16]` depending on the implementation.
#[inline]
#[cfg_attr(test, assert_instr(i8x16.relaxed_swizzle))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i8x16.relaxed_swizzle"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn i8x16_relaxed_swizzle(a: v128, s: v128) -> v128 {
    unsafe { llvm_relaxed_swizzle(a.as_i8x16(), s.as_i8x16()).v128() }
}

/// A relaxed version of `i32x4_trunc_sat_f32x4(a)` converts the `f32` lanes
/// of `a` to signed 32-bit integers.
///
/// Values which don't fit in 32-bit integers or are NaN may have the same
/// result as `i32x4_trunc_sat_f32x4` or may return `i32::MIN`.
#[inline]
#[cfg_attr(test, assert_instr(i32x4.relaxed_trunc_f32x4_s))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i32x4.relaxed_trunc_f32x4_s"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn i32x4_relaxed_trunc_f32x4(a: v128) -> v128 {
    unsafe { llvm_relaxed_trunc_signed(a.as_f32x4()).v128() }
}

/// A relaxed version of `u32x4_trunc_sat_f32x4(a)` converts the `f32` lanes
/// of `a` to unsigned 32-bit integers.
///
/// Values which don't fit in 32-bit unsigned integers or are NaN may have the
/// same result as `u32x4_trunc_sat_f32x4` or may return `u32::MAX`.
#[inline]
#[cfg_attr(test, assert_instr(i32x4.relaxed_trunc_f32x4_u))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i32x4.relaxed_trunc_f32x4_u"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn u32x4_relaxed_trunc_f32x4(a: v128) -> v128 {
    unsafe { llvm_relaxed_trunc_unsigned(a.as_f32x4()).v128() }
}

/// A relaxed version of `i32x4_trunc_sat_f64x2_zero(a)` converts the `f64`
/// lanes of `a` to signed 32-bit integers and the upper two lanes are zero.
///
/// Values which don't fit in 32-bit integers or are NaN may have the same
/// result as `i32x4_trunc_sat_f32x4` or may return `i32::MIN`.
#[inline]
#[cfg_attr(test, assert_instr(i32x4.relaxed_trunc_f64x2_s_zero))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i32x4.relaxed_trunc_f64x2_s_zero"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn i32x4_relaxed_trunc_f64x2_zero(a: v128) -> v128 {
    unsafe { llvm_relaxed_trunc_signed_zero(a.as_f64x2()).v128() }
}

/// A relaxed version of `u32x4_trunc_sat_f64x2_zero(a)` converts the `f64`
/// lanes of `a` to unsigned 32-bit integers and the upper two lanes are zero.
///
/// Values which don't fit in 32-bit unsigned integers or are NaN may have the
/// same result as `u32x4_trunc_sat_f32x4` or may return `u32::MAX`.
#[inline]
#[cfg_attr(test, assert_instr(i32x4.relaxed_trunc_f64x2_u_zero))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i32x4.relaxed_trunc_f64x2_u_zero"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn u32x4_relaxed_trunc_f64x2_zero(a: v128) -> v128 {
    unsafe { llvm_relaxed_trunc_unsigned_zero(a.as_f64x2()).v128() }
}

/// Computes `a * b + c` with either one rounding or two roundings.
#[inline]
#[cfg_attr(test, assert_instr(f32x4.relaxed_madd))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("f32x4.relaxed_madd"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn f32x4_relaxed_madd(a: v128, b: v128, c: v128) -> v128 {
    unsafe { llvm_f32x4_fma(a.as_f32x4(), b.as_f32x4(), c.as_f32x4()).v128() }
}

/// Computes `-a * b + c` with either one rounding or two roundings.
#[inline]
#[cfg_attr(test, assert_instr(f32x4.relaxed_nmadd))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("f32x4.relaxed_nmadd"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn f32x4_relaxed_nmadd(a: v128, b: v128, c: v128) -> v128 {
    unsafe { llvm_f32x4_fms(a.as_f32x4(), b.as_f32x4(), c.as_f32x4()).v128() }
}

/// Computes `a * b + c` with either one rounding or two roundings.
#[inline]
#[cfg_attr(test, assert_instr(f64x2.relaxed_madd))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("f64x2.relaxed_madd"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn f64x2_relaxed_madd(a: v128, b: v128, c: v128) -> v128 {
    unsafe { llvm_f64x2_fma(a.as_f64x2(), b.as_f64x2(), c.as_f64x2()).v128() }
}

/// Computes `-a * b + c` with either one rounding or two roundings.
#[inline]
#[cfg_attr(test, assert_instr(f64x2.relaxed_nmadd))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("f64x2.relaxed_nmadd"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn f64x2_relaxed_nmadd(a: v128, b: v128, c: v128) -> v128 {
    unsafe { llvm_f64x2_fms(a.as_f64x2(), b.as_f64x2(), c.as_f64x2()).v128() }
}

/// A relaxed version of `v128_bitselect` where this either behaves the same as
/// `v128_bitselect` or the high bit of each lane `m` is inspected and the
/// corresponding lane of `a` is chosen if the bit is 1 or the lane of `b` is
/// chosen if it's zero.
///
/// If the `m` mask's lanes are either all-one or all-zero then this instruction
/// is the same as `v128_bitselect`.
#[inline]
#[cfg_attr(test, assert_instr(i8x16.relaxed_laneselect))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i8x16.relaxed_laneselect"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn i8x16_relaxed_laneselect(a: v128, b: v128, m: v128) -> v128 {
    unsafe { llvm_i8x16_laneselect(a.as_i8x16(), b.as_i8x16(), m.as_i8x16()).v128() }
}

/// A relaxed version of `v128_bitselect` where this either behaves the same as
/// `v128_bitselect` or the high bit of each lane `m` is inspected and the
/// corresponding lane of `a` is chosen if the bit is 1 or the lane of `b` is
/// chosen if it's zero.
///
/// If the `m` mask's lanes are either all-one or all-zero then this instruction
/// is the same as `v128_bitselect`.
#[inline]
#[cfg_attr(test, assert_instr(i16x8.relaxed_laneselect))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i16x8.relaxed_laneselect"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn i16x8_relaxed_laneselect(a: v128, b: v128, m: v128) -> v128 {
    unsafe { llvm_i16x8_laneselect(a.as_i16x8(), b.as_i16x8(), m.as_i16x8()).v128() }
}

/// A relaxed version of `v128_bitselect` where this either behaves the same as
/// `v128_bitselect` or the high bit of each lane `m` is inspected and the
/// corresponding lane of `a` is chosen if the bit is 1 or the lane of `b` is
/// chosen if it's zero.
///
/// If the `m` mask's lanes are either all-one or all-zero then this instruction
/// is the same as `v128_bitselect`.
#[inline]
#[cfg_attr(test, assert_instr(i32x4.relaxed_laneselect))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i32x4.relaxed_laneselect"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn i32x4_relaxed_laneselect(a: v128, b: v128, m: v128) -> v128 {
    unsafe { llvm_i32x4_laneselect(a.as_i32x4(), b.as_i32x4(), m.as_i32x4()).v128() }
}

/// A relaxed version of `v128_bitselect` where this either behaves the same as
/// `v128_bitselect` or the high bit of each lane `m` is inspected and the
/// corresponding lane of `a` is chosen if the bit is 1 or the lane of `b` is
/// chosen if it's zero.
///
/// If the `m` mask's lanes are either all-one or all-zero then this instruction
/// is the same as `v128_bitselect`.
#[inline]
#[cfg_attr(test, assert_instr(i64x2.relaxed_laneselect))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i64x2.relaxed_laneselect"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn i64x2_relaxed_laneselect(a: v128, b: v128, m: v128) -> v128 {
    unsafe { llvm_i64x2_laneselect(a.as_i64x2(), b.as_i64x2(), m.as_i64x2()).v128() }
}

/// A relaxed version of `f32x4_min` which is either `f32x4_min` or
/// `f32x4_pmin`.
#[inline]
#[cfg_attr(test, assert_instr(f32x4.relaxed_min))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("f32x4.relaxed_min"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn f32x4_relaxed_min(a: v128, b: v128) -> v128 {
    unsafe { llvm_f32x4_relaxed_min(a.as_f32x4(), b.as_f32x4()).v128() }
}

/// A relaxed version of `f32x4_max` which is either `f32x4_max` or
/// `f32x4_pmax`.
#[inline]
#[cfg_attr(test, assert_instr(f32x4.relaxed_max))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("f32x4.relaxed_max"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn f32x4_relaxed_max(a: v128, b: v128) -> v128 {
    unsafe { llvm_f32x4_relaxed_max(a.as_f32x4(), b.as_f32x4()).v128() }
}

/// A relaxed version of `f64x2_min` which is either `f64x2_min` or
/// `f64x2_pmin`.
#[inline]
#[cfg_attr(test, assert_instr(f64x2.relaxed_min))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("f64x2.relaxed_min"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn f64x2_relaxed_min(a: v128, b: v128) -> v128 {
    unsafe { llvm_f64x2_relaxed_min(a.as_f64x2(), b.as_f64x2()).v128() }
}

/// A relaxed version of `f64x2_max` which is either `f64x2_max` or
/// `f64x2_pmax`.
#[inline]
#[cfg_attr(test, assert_instr(f64x2.relaxed_max))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("f64x2.relaxed_max"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn f64x2_relaxed_max(a: v128, b: v128) -> v128 {
    unsafe { llvm_f64x2_relaxed_max(a.as_f64x2(), b.as_f64x2()).v128() }
}

/// A relaxed version of `i16x8_relaxed_q15mulr` where if both lanes are
/// `i16::MIN` then the result is either `i16::MIN` or `i16::MAX`.
#[inline]
#[cfg_attr(test, assert_instr(i16x8.relaxed_q15mulr_s))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i16x8.relaxed_q15mulr_s"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn i16x8_relaxed_q15mulr(a: v128, b: v128) -> v128 {
    unsafe { llvm_relaxed_q15mulr_signed(a.as_i16x8(), b.as_i16x8()).v128() }
}

/// A relaxed dot-product instruction.
///
/// This instruction will perform pairwise products of the 8-bit values in `a`
/// and `b` and then accumulate adjacent pairs into 16-bit results producing a
/// final `i16x8` vector. The bytes of `a` are always interpreted as signed and
/// the bytes in `b` may be interpreted as signed or unsigned. If the top bit in
/// `b` isn't set then the value is the same regardless of whether it's signed
/// or unsigned.
///
/// The accumulation into 16-bit values may be saturated on some platforms, and
/// on other platforms it may wrap-around on overflow.
#[inline]
#[cfg_attr(test, assert_instr(i16x8.relaxed_dot_i8x16_i7x16_s))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i16x8.relaxed_dot_i8x16_i7x16_s"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn i16x8_relaxed_dot_i8x16_i7x16(a: v128, b: v128) -> v128 {
    unsafe { llvm_i16x8_relaxed_dot_i8x16_i7x16_s(a.as_i8x16(), b.as_i8x16()).v128() }
}

/// Similar to [`i16x8_relaxed_dot_i8x16_i7x16`] except that the intermediate
/// `i16x8` result is fed into `i32x4_extadd_pairwise_i16x8` followed by
/// `i32x4_add` to add the value `c` to the result.
#[inline]
#[cfg_attr(test, assert_instr(i32x4.relaxed_dot_i8x16_i7x16_add_s))]
#[target_feature(enable = "relaxed-simd")]
#[doc(alias("i32x4.relaxed_dot_i8x16_i7x16_add_s"))]
#[unstable(feature = "stdarch_wasm_relaxed_simd", issue = "111196")]
pub fn i32x4_relaxed_dot_i8x16_i7x16_add(a: v128, b: v128, c: v128) -> v128 {
    unsafe {
        llvm_i32x4_relaxed_dot_i8x16_i7x16_add_s(a.as_i8x16(), b.as_i8x16(), c.as_i32x4()).v128()
    }
}

#[cfg(test)]
mod tests {
    use super::super::simd128::*;
    use super::*;
    use core::ops::{Add, Div, Mul, Neg, Sub};

    use std::fmt::Debug;
    use std::mem::transmute;
    use std::num::Wrapping;
    use std::prelude::v1::*;

    fn compare_bytes(a: v128, b: &[v128]) {
        let a: [u8; 16] = unsafe { transmute(a) };
        if b.iter().any(|b| {
            let b: [u8; 16] = unsafe { transmute(*b) };
            a == b
        }) {
            return;
        }
        eprintln!("input vector {a:?}");
        eprintln!("did not match any output:");
        for b in b {
            eprintln!("  {b:?}");
        }
    }

    #[test]
    fn test_relaxed_swizzle() {
        compare_bytes(
            i8x16_relaxed_swizzle(
                i8x16(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15),
                i8x16(2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0, 1),
            ),
            &[i8x16(2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0, 1)],
        );
        compare_bytes(
            i8x16_relaxed_swizzle(
                i8x16(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15),
                u8x16(0x80, 0xff, 16, 17, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0),
            ),
            &[
                i8x16(0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0),
                i8x16(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0),
            ],
        );
    }

    #[test]
    fn test_relaxed_trunc() {
        compare_bytes(
            i32x4_relaxed_trunc_f32x4(f32x4(1.0, 2.0, -1., -4.)),
            &[i32x4(1, 2, -1, -4)],
        );
        compare_bytes(
            i32x4_relaxed_trunc_f32x4(f32x4(f32::NEG_INFINITY, f32::NAN, -0.0, f32::INFINITY)),
            &[
                i32x4(i32::MIN, 0, 0, i32::MAX),
                i32x4(i32::MIN, i32::MIN, 0, i32::MIN),
            ],
        );
        compare_bytes(
            i32x4_relaxed_trunc_f64x2_zero(f64x2(1.0, -3.0)),
            &[i32x4(1, -3, 0, 0)],
        );
        compare_bytes(
            i32x4_relaxed_trunc_f64x2_zero(f64x2(f64::INFINITY, f64::NAN)),
            &[i32x4(i32::MAX, 0, 0, 0), i32x4(i32::MIN, i32::MIN, 0, 0)],
        );

        compare_bytes(
            u32x4_relaxed_trunc_f32x4(f32x4(1.0, 2.0, 5., 100.)),
            &[i32x4(1, 2, 5, 100)],
        );
        compare_bytes(
            u32x4_relaxed_trunc_f32x4(f32x4(f32::NEG_INFINITY, f32::NAN, -0.0, f32::INFINITY)),
            &[
                u32x4(u32::MAX, 0, 0, u32::MAX),
                u32x4(u32::MAX, u32::MAX, 0, u32::MAX),
            ],
        );
        compare_bytes(
            u32x4_relaxed_trunc_f64x2_zero(f64x2(1.0, 3.0)),
            &[u32x4(1, 3, 0, 0)],
        );
        compare_bytes(
            u32x4_relaxed_trunc_f64x2_zero(f64x2(f64::INFINITY, f64::NAN)),
            &[i32x4(i32::MAX, 0, 0, 0), i32x4(i32::MIN, i32::MIN, 0, 0)],
        );
    }

    #[test]
    fn test_madd() {
        let floats = [
            f32::NAN,
            f32::NEG_INFINITY,
            f32::INFINITY,
            1.0,
            2.0,
            -1.0,
            0.0,
            100.3,
            7.8,
            9.4,
        ];
        for &a in floats.iter() {
            for &b in floats.iter() {
                for &c in floats.iter() {
                    let f1 = a * b + c;
                    let f2 = a.mul_add(b, c);
                    compare_bytes(
                        f32x4_relaxed_madd(f32x4(a, a, a, a), f32x4(b, b, b, b), f32x4(c, c, c, c)),
                        &[f32x4(f1, f1, f1, f1), f32x4(f2, f2, f2, f2)],
                    );

                    let f1 = -a * b + c;
                    let f2 = (-a).mul_add(b, c);
                    compare_bytes(
                        f32x4_relaxed_nmadd(
                            f32x4(a, a, a, a),
                            f32x4(b, b, b, b),
                            f32x4(c, c, c, c),
                        ),
                        &[f32x4(f1, f1, f1, f1), f32x4(f2, f2, f2, f2)],
                    );

                    let a = f64::from(a);
                    let b = f64::from(b);
                    let c = f64::from(c);
                    let f1 = a * b + c;
                    let f2 = a.mul_add(b, c);
                    compare_bytes(
                        f64x2_relaxed_madd(f64x2(a, a), f64x2(b, b), f64x2(c, c)),
                        &[f64x2(f1, f1), f64x2(f2, f2)],
                    );
                    let f1 = -a * b + c;
                    let f2 = (-a).mul_add(b, c);
                    compare_bytes(
                        f64x2_relaxed_nmadd(f64x2(a, a), f64x2(b, b), f64x2(c, c)),
                        &[f64x2(f1, f1), f64x2(f2, f2)],
                    );
                }
            }
        }
    }
}