ic-ec 0.3.2

//! P-384's field operations and point formulas, each one `asm!` block over
//! `mulx`/`adcx`/`adox`.
//!
//! GENERATED by `scripts/gen_p384_x86_point.py`; edit that, not this.
//!
//! The field operations are specialised to p = 2^384 - 2^128 - 2^96 + 2^32 -
//! 1: Montgomery reduction takes two `mulx` per word, by constants, where a
//! generic modulus takes six; squaring computes each cross product once; and
//! every final correction is a conditional move or a masked addition, in
//! registers. The point formulas compute the values of `point.rs` --
//! doubling, addition and mixed addition -- over a buffer of 48-byte slots,
//! their operations ordered for the core to overlap them; the doubling also
//! in place, for the scalar multiplication. Nothing branches and nothing is
//! indexed by data.

use core::arch::asm;
use core::mem::MaybeUninit;

/// `a * b * 2^-384 mod p`, canonical.
///
/// Slots, 48 bytes each, in order: `a`, `b`, `out`, `scratch`.
///
/// # Safety
///
/// The CPU must support BMI2 and ADX; see `super::p256_x86::has_adx`.
/// Every input slot must hold a value below p.
#[inline]
pub unsafe fn fmul_adx(buf: &mut [MaybeUninit<u64>; 24]) {
    // SAFETY: the caller guarantees the instructions exist. The block reads
    // and writes only the buffer behind `{s}`, a live, exclusive reference
    // of 192 bytes, at offsets below that; it writes only the
    // registers it declares, and does not touch the stack. It reads no slot
    // before writing it but the inputs, which the caller has written.
    unsafe {
        asm!(
            ".p2align 6",
            // out = mul(a, b)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 48]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 56]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 64]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 72]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 80]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 88]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 136], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 136]",
            "mov qword ptr [{s} + 96], {w6}",
            "mov qword ptr [{s} + 104], {w7}",
            "mov qword ptr [{s} + 112], {w0}",
            "mov qword ptr [{s} + 120], {w1}",
            "mov qword ptr [{s} + 128], {w2}",
            "mov qword ptr [{s} + 136], {w3}",
            s = in(reg) buf.as_mut_ptr(),
            w0 = out(reg) _,
            w1 = out(reg) _,
            w2 = out(reg) _,
            w3 = out(reg) _,
            w4 = out(reg) _,
            w5 = out(reg) _,
            w6 = out(reg) _,
            w7 = out(reg) _,
            c = out(reg) _,
            k = out(reg) _,
            out("rax") _,
            out("rdx") _,
            options(nostack),
        );
    }
}
#[allow(dead_code)]
pub const FMUL_OUT: usize = 2;

/// `a^2 * 2^-384 mod p`, canonical.
///
/// Slots, 48 bytes each, in order: `a`, `out`, `scratch`.
///
/// # Safety
///
/// The CPU must support BMI2 and ADX; see `super::p256_x86::has_adx`.
/// Every input slot must hold a value below p.
#[inline]
pub unsafe fn fsqr_adx(buf: &mut [MaybeUninit<u64>; 18]) {
    // SAFETY: the caller guarantees the instructions exist. The block reads
    // and writes only the buffer behind `{s}`, a live, exclusive reference
    // of 144 bytes, at offsets below that; it writes only the
    // registers it declares, and does not touch the stack. It reads no slot
    // before writing it but the inputs, which the caller has written.
    unsafe {
        asm!(
            ".p2align 6",
            // out = sqr(a)
            "mov rdx, qword ptr [{s}]",
            "mulx {w2}, {w1}, qword ptr [{s} + 8]",
            "mulx {w3}, rax, qword ptr [{s} + 16]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 24]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 32]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 40]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 8]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 16]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 24]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 32]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 40]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 16]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 24]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 32]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 40]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 24]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 32]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 40]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 32]",
            "mulx {k}, rax, qword ptr [{s} + 40]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 128], {k}",
            "mov qword ptr [{s} + 136], rax",
            "mov rdx, qword ptr [{s}]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 8]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 16]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 24]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 32]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 40]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 128], rax",
            "adc qword ptr [{s} + 136], rdx",
            "mov qword ptr [{s} + 96], {w6}",
            "mov qword ptr [{s} + 104], {w7}",
            "mov qword ptr [{s} + 112], {c}",
            "mov qword ptr [{s} + 120], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 96]",
            "adc {w1}, qword ptr [{s} + 104]",
            "adc {w2}, qword ptr [{s} + 112]",
            "adc {w3}, qword ptr [{s} + 120]",
            "adc {w4}, qword ptr [{s} + 128]",
            "adc {w5}, qword ptr [{s} + 136]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 88], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 88]",
            "mov qword ptr [{s} + 48], {k}",
            "mov qword ptr [{s} + 56], {w1}",
            "mov qword ptr [{s} + 64], {w2}",
            "mov qword ptr [{s} + 72], {w3}",
            "mov qword ptr [{s} + 80], {w4}",
            "mov qword ptr [{s} + 88], {w5}",
            s = in(reg) buf.as_mut_ptr(),
            w0 = out(reg) _,
            w1 = out(reg) _,
            w2 = out(reg) _,
            w3 = out(reg) _,
            w4 = out(reg) _,
            w5 = out(reg) _,
            w6 = out(reg) _,
            w7 = out(reg) _,
            c = out(reg) _,
            k = out(reg) _,
            out("rax") _,
            out("rdx") _,
            options(nostack),
        );
    }
}
#[allow(dead_code)]
pub const FSQR_OUT: usize = 1;

/// `a + b mod p`; alone, only for the tests.
///
/// Slots, 48 bytes each, in order: `a`, `b`, `out`, `scratch`.
///
/// # Safety
///
/// The CPU must support BMI2 and ADX; see `super::p256_x86::has_adx`.
/// Every input slot must hold a value below p.
#[cfg(test)]
#[inline(never)]
pub unsafe fn fadd_adx(buf: &mut [MaybeUninit<u64>; 24]) {
    // SAFETY: the caller guarantees the instructions exist. The block reads
    // and writes only the buffer behind `{s}`, a live, exclusive reference
    // of 192 bytes, at offsets below that; it writes only the
    // registers it declares, and does not touch the stack. It reads no slot
    // before writing it but the inputs, which the caller has written.
    unsafe {
        asm!(
            ".p2align 6",
            // out = add(a, b)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s}]",
            "mov {w1}, qword ptr [{s} + 8]",
            "mov {w2}, qword ptr [{s} + 16]",
            "mov {w3}, qword ptr [{s} + 24]",
            "mov {w4}, qword ptr [{s} + 32]",
            "mov {w5}, qword ptr [{s} + 40]",
            "add {w0}, qword ptr [{s} + 48]",
            "adc {w1}, qword ptr [{s} + 56]",
            "adc {w2}, qword ptr [{s} + 64]",
            "adc {w3}, qword ptr [{s} + 72]",
            "adc {w4}, qword ptr [{s} + 80]",
            "adc {w5}, qword ptr [{s} + 88]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 136], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 136]",
            "mov qword ptr [{s} + 96], {w0}",
            "mov qword ptr [{s} + 104], {w1}",
            "mov qword ptr [{s} + 112], {w2}",
            "mov qword ptr [{s} + 120], {w3}",
            "mov qword ptr [{s} + 128], {w4}",
            "mov qword ptr [{s} + 136], {w5}",
            s = in(reg) buf.as_mut_ptr(),
            w0 = out(reg) _,
            w1 = out(reg) _,
            w2 = out(reg) _,
            w3 = out(reg) _,
            w4 = out(reg) _,
            w5 = out(reg) _,
            w6 = out(reg) _,
            w7 = out(reg) _,
            c = out(reg) _,
            k = out(reg) _,
            out("rax") _,
            out("rdx") _,
            options(nostack),
        );
    }
}
#[cfg(test)]
#[allow(dead_code)]
pub const FADD_OUT: usize = 2;

/// `a - b mod p`; alone, only for the tests.
///
/// Slots, 48 bytes each, in order: `a`, `b`, `out`, `scratch`.
///
/// # Safety
///
/// The CPU must support BMI2 and ADX; see `super::p256_x86::has_adx`.
/// Every input slot must hold a value below p.
#[cfg(test)]
#[inline(never)]
pub unsafe fn fsub_adx(buf: &mut [MaybeUninit<u64>; 24]) {
    // SAFETY: the caller guarantees the instructions exist. The block reads
    // and writes only the buffer behind `{s}`, a live, exclusive reference
    // of 192 bytes, at offsets below that; it writes only the
    // registers it declares, and does not touch the stack. It reads no slot
    // before writing it but the inputs, which the caller has written.
    unsafe {
        asm!(
            ".p2align 6",
            // out = sub(a, b)
            "mov {w0}, qword ptr [{s}]",
            "mov {w1}, qword ptr [{s} + 8]",
            "mov {w2}, qword ptr [{s} + 16]",
            "mov {w3}, qword ptr [{s} + 24]",
            "mov {w4}, qword ptr [{s} + 32]",
            "mov {w5}, qword ptr [{s} + 40]",
            "sub {w0}, qword ptr [{s} + 48]",
            "sbb {w1}, qword ptr [{s} + 56]",
            "sbb {w2}, qword ptr [{s} + 64]",
            "sbb {w3}, qword ptr [{s} + 72]",
            "sbb {w4}, qword ptr [{s} + 80]",
            "sbb {w5}, qword ptr [{s} + 88]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 96], {w0}",
            "mov qword ptr [{s} + 104], {w1}",
            "mov qword ptr [{s} + 112], {w2}",
            "mov qword ptr [{s} + 120], {w3}",
            "mov qword ptr [{s} + 128], {w4}",
            "mov qword ptr [{s} + 136], {w5}",
            s = in(reg) buf.as_mut_ptr(),
            w0 = out(reg) _,
            w1 = out(reg) _,
            w2 = out(reg) _,
            w3 = out(reg) _,
            w4 = out(reg) _,
            w5 = out(reg) _,
            w6 = out(reg) _,
            c = out(reg) _,
            out("rax") _,
            out("rdx") _,
            options(nostack),
        );
    }
}
#[cfg(test)]
#[allow(dead_code)]
pub const FSUB_OUT: usize = 2;

/// `3 a - 8 b mod p`; alone, only for the tests.
///
/// Slots, 48 bytes each, in order: `a`, `b`, `out`, `scratch`.
///
/// # Safety
///
/// The CPU must support BMI2 and ADX; see `super::p256_x86::has_adx`.
/// Every input slot must hold a value below p.
#[cfg(test)]
#[inline(never)]
pub unsafe fn fcms_3_8_adx(buf: &mut [MaybeUninit<u64>; 24]) {
    // SAFETY: the caller guarantees the instructions exist. The block reads
    // and writes only the buffer behind `{s}`, a live, exclusive reference
    // of 192 bytes, at offsets below that; it writes only the
    // registers it declares, and does not touch the stack. It reads no slot
    // before writing it but the inputs, which the caller has written.
    unsafe {
        asm!(
            ".p2align 6",
            // out = 3 * a - 8 * b
            "mov {w0:e}, 0xffffffff",
            "sub {w0}, qword ptr [{s} + 48]",
            "mov {w1}, 0xffffffff00000000",
            "sbb {w1}, qword ptr [{s} + 56]",
            "mov {w2}, -2",
            "sbb {w2}, qword ptr [{s} + 64]",
            "mov {w3}, -1",
            "sbb {w3}, qword ptr [{s} + 72]",
            "mov {w4}, -1",
            "sbb {w4}, qword ptr [{s} + 80]",
            "mov {w5}, -1",
            "sbb {w5}, qword ptr [{s} + 88]",
            "mov {w6}, {w5}",
            "shr {w6}, 61",
            "shld {w5}, {w4}, 3",
            "shld {w4}, {w3}, 3",
            "shld {w3}, {w2}, 3",
            "shld {w2}, {w1}, 3",
            "shld {w1}, {w0}, 3",
            "shl {w0}, 3",
            "mov edx, 3",
            "xor {c:e}, {c:e}",
            "mulx {w7}, rax, qword ptr [{s}]",
            "adcx {w0}, rax",
            "adox {w1}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 8]",
            "adcx {w1}, rax",
            "adox {w2}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 16]",
            "adcx {w2}, rax",
            "adox {w3}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 24]",
            "adcx {w3}, rax",
            "adox {w4}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 32]",
            "adcx {w4}, rax",
            "adox {w5}, {w7}",
            "mulx rdx, rax, qword ptr [{s} + 40]",
            "adcx {w5}, rax",
            "adox rdx, {w6}",
            "adcx rdx, {c}",
            "add rdx, 1",
            "mov rax, 0xffffffff00000001",
            "mulx {w7}, rax, rax",
            "mov {c}, rdx",
            "shl {c}, 32",
            "sub {c}, rdx",
            "add {c}, {w7}",
            "add {w0}, rax",
            "adc {w1}, {c}",
            "adc {w2}, rdx",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "adc {w5}, 0",
            "cmc",
            "sbb rax, rax",
            "mov rdx, 0xffffffff00000001",
            "and rdx, rax",
            "mov {c:e}, 0xffffffff",
            "and {c}, rax",
            "and eax, 1",
            "sub {w0}, rdx",
            "sbb {w1}, {c}",
            "sbb {w2}, rax",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov qword ptr [{s} + 96], {w0}",
            "mov qword ptr [{s} + 104], {w1}",
            "mov qword ptr [{s} + 112], {w2}",
            "mov qword ptr [{s} + 120], {w3}",
            "mov qword ptr [{s} + 128], {w4}",
            "mov qword ptr [{s} + 136], {w5}",
            s = in(reg) buf.as_mut_ptr(),
            w0 = out(reg) _,
            w1 = out(reg) _,
            w2 = out(reg) _,
            w3 = out(reg) _,
            w4 = out(reg) _,
            w5 = out(reg) _,
            w6 = out(reg) _,
            w7 = out(reg) _,
            c = out(reg) _,
            out("rax") _,
            out("rdx") _,
            options(nostack),
        );
    }
}
#[cfg(test)]
#[allow(dead_code)]
pub const FCMS_3_8_OUT: usize = 2;

/// `4 a - 1 b mod p`; alone, only for the tests.
///
/// Slots, 48 bytes each, in order: `a`, `b`, `out`, `scratch`.
///
/// # Safety
///
/// The CPU must support BMI2 and ADX; see `super::p256_x86::has_adx`.
/// Every input slot must hold a value below p.
#[cfg(test)]
#[inline(never)]
pub unsafe fn fcms_4_1_adx(buf: &mut [MaybeUninit<u64>; 24]) {
    // SAFETY: the caller guarantees the instructions exist. The block reads
    // and writes only the buffer behind `{s}`, a live, exclusive reference
    // of 192 bytes, at offsets below that; it writes only the
    // registers it declares, and does not touch the stack. It reads no slot
    // before writing it but the inputs, which the caller has written.
    unsafe {
        asm!(
            ".p2align 6",
            // out = 4 * a - 1 * b
            "mov {w0:e}, 0xffffffff",
            "sub {w0}, qword ptr [{s} + 48]",
            "mov {w1}, 0xffffffff00000000",
            "sbb {w1}, qword ptr [{s} + 56]",
            "mov {w2}, -2",
            "sbb {w2}, qword ptr [{s} + 64]",
            "mov {w3}, -1",
            "sbb {w3}, qword ptr [{s} + 72]",
            "mov {w4}, -1",
            "sbb {w4}, qword ptr [{s} + 80]",
            "mov {w5}, -1",
            "sbb {w5}, qword ptr [{s} + 88]",
            "xor {w6:e}, {w6:e}",
            "mov edx, 4",
            "xor {c:e}, {c:e}",
            "mulx {w7}, rax, qword ptr [{s}]",
            "adcx {w0}, rax",
            "adox {w1}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 8]",
            "adcx {w1}, rax",
            "adox {w2}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 16]",
            "adcx {w2}, rax",
            "adox {w3}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 24]",
            "adcx {w3}, rax",
            "adox {w4}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 32]",
            "adcx {w4}, rax",
            "adox {w5}, {w7}",
            "mulx rdx, rax, qword ptr [{s} + 40]",
            "adcx {w5}, rax",
            "adox rdx, {w6}",
            "adcx rdx, {c}",
            "add rdx, 1",
            "mov rax, 0xffffffff00000001",
            "mulx {w7}, rax, rax",
            "mov {c}, rdx",
            "shl {c}, 32",
            "sub {c}, rdx",
            "add {c}, {w7}",
            "add {w0}, rax",
            "adc {w1}, {c}",
            "adc {w2}, rdx",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "adc {w5}, 0",
            "cmc",
            "sbb rax, rax",
            "mov rdx, 0xffffffff00000001",
            "and rdx, rax",
            "mov {c:e}, 0xffffffff",
            "and {c}, rax",
            "and eax, 1",
            "sub {w0}, rdx",
            "sbb {w1}, {c}",
            "sbb {w2}, rax",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov qword ptr [{s} + 96], {w0}",
            "mov qword ptr [{s} + 104], {w1}",
            "mov qword ptr [{s} + 112], {w2}",
            "mov qword ptr [{s} + 120], {w3}",
            "mov qword ptr [{s} + 128], {w4}",
            "mov qword ptr [{s} + 136], {w5}",
            s = in(reg) buf.as_mut_ptr(),
            w0 = out(reg) _,
            w1 = out(reg) _,
            w2 = out(reg) _,
            w3 = out(reg) _,
            w4 = out(reg) _,
            w5 = out(reg) _,
            w6 = out(reg) _,
            w7 = out(reg) _,
            c = out(reg) _,
            out("rax") _,
            out("rdx") _,
            options(nostack),
        );
    }
}
#[cfg(test)]
#[allow(dead_code)]
pub const FCMS_4_1_OUT: usize = 2;

/// `12 a - 9 b mod p`; alone, only for the tests.
///
/// Slots, 48 bytes each, in order: `a`, `b`, `out`, `scratch`.
///
/// # Safety
///
/// The CPU must support BMI2 and ADX; see `super::p256_x86::has_adx`.
/// Every input slot must hold a value below p.
#[cfg(test)]
#[inline(never)]
pub unsafe fn fcms_12_9_adx(buf: &mut [MaybeUninit<u64>; 24]) {
    // SAFETY: the caller guarantees the instructions exist. The block reads
    // and writes only the buffer behind `{s}`, a live, exclusive reference
    // of 192 bytes, at offsets below that; it writes only the
    // registers it declares, and does not touch the stack. It reads no slot
    // before writing it but the inputs, which the caller has written.
    unsafe {
        asm!(
            ".p2align 6",
            // out = 12 * a - 9 * b
            "mov {w0:e}, 0xffffffff",
            "sub {w0}, qword ptr [{s} + 48]",
            "mov {w1}, 0xffffffff00000000",
            "sbb {w1}, qword ptr [{s} + 56]",
            "mov {w2}, -2",
            "sbb {w2}, qword ptr [{s} + 64]",
            "mov {w3}, -1",
            "sbb {w3}, qword ptr [{s} + 72]",
            "mov {w4}, -1",
            "sbb {w4}, qword ptr [{s} + 80]",
            "mov {w5}, -1",
            "sbb {w5}, qword ptr [{s} + 88]",
            "mov edx, 9",
            "mulx {w7}, {w0}, {w0}",
            "mulx {c}, {w1}, {w1}",
            "add {w1}, {w7}",
            "mulx {w7}, {w2}, {w2}",
            "adc {w2}, {c}",
            "mulx {c}, {w3}, {w3}",
            "adc {w3}, {w7}",
            "mulx {w7}, {w4}, {w4}",
            "adc {w4}, {c}",
            "mulx {w6}, {w5}, {w5}",
            "adc {w5}, {w7}",
            "adc {w6}, 0",
            "mov edx, 12",
            "xor {c:e}, {c:e}",
            "mulx {w7}, rax, qword ptr [{s}]",
            "adcx {w0}, rax",
            "adox {w1}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 8]",
            "adcx {w1}, rax",
            "adox {w2}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 16]",
            "adcx {w2}, rax",
            "adox {w3}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 24]",
            "adcx {w3}, rax",
            "adox {w4}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 32]",
            "adcx {w4}, rax",
            "adox {w5}, {w7}",
            "mulx rdx, rax, qword ptr [{s} + 40]",
            "adcx {w5}, rax",
            "adox rdx, {w6}",
            "adcx rdx, {c}",
            "add rdx, 1",
            "mov rax, 0xffffffff00000001",
            "mulx {w7}, rax, rax",
            "mov {c}, rdx",
            "shl {c}, 32",
            "sub {c}, rdx",
            "add {c}, {w7}",
            "add {w0}, rax",
            "adc {w1}, {c}",
            "adc {w2}, rdx",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "adc {w5}, 0",
            "cmc",
            "sbb rax, rax",
            "mov rdx, 0xffffffff00000001",
            "and rdx, rax",
            "mov {c:e}, 0xffffffff",
            "and {c}, rax",
            "and eax, 1",
            "sub {w0}, rdx",
            "sbb {w1}, {c}",
            "sbb {w2}, rax",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov qword ptr [{s} + 96], {w0}",
            "mov qword ptr [{s} + 104], {w1}",
            "mov qword ptr [{s} + 112], {w2}",
            "mov qword ptr [{s} + 120], {w3}",
            "mov qword ptr [{s} + 128], {w4}",
            "mov qword ptr [{s} + 136], {w5}",
            s = in(reg) buf.as_mut_ptr(),
            w0 = out(reg) _,
            w1 = out(reg) _,
            w2 = out(reg) _,
            w3 = out(reg) _,
            w4 = out(reg) _,
            w5 = out(reg) _,
            w6 = out(reg) _,
            w7 = out(reg) _,
            c = out(reg) _,
            out("rax") _,
            out("rdx") _,
            options(nostack),
        );
    }
}
#[cfg(test)]
#[allow(dead_code)]
pub const FCMS_12_9_OUT: usize = 2;

/// Point doubling, `dbl-2001-b` for `a = -3`, as `Point::double`.
///
/// Slots, 48 bytes each, in order: `X`, `Y`, `Z`, `X3`, `Y3`, `Z3`, `delta`, `gamma`, `beta`, `m`, `u`, `m2`, `d`, `w`, `g2`, `scratch`.
///
/// # Safety
///
/// The CPU must support BMI2 and ADX; see `super::p256_x86::has_adx`.
/// Every input slot must hold a value below p.
#[inline(never)]
pub unsafe fn double_adx(buf: &mut [MaybeUninit<u64>; 96]) {
    // SAFETY: the caller guarantees the instructions exist. The block reads
    // and writes only the buffer behind `{s}`, a live, exclusive reference
    // of 768 bytes, at offsets below that; it writes only the
    // registers it declares, and does not touch the stack. It reads no slot
    // before writing it but the inputs, which the caller has written.
    unsafe {
        asm!(
            ".p2align 6",
            // delta = sqr(Z)
            "mov rdx, qword ptr [{s} + 96]",
            "mulx {w2}, {w1}, qword ptr [{s} + 104]",
            "mulx {w3}, rax, qword ptr [{s} + 112]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 120]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 128]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 136]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 104]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 112]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 120]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 112]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 120]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 120]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 128]",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 752], {k}",
            "mov qword ptr [{s} + 760], rax",
            "mov rdx, qword ptr [{s} + 96]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 104]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 112]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 120]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 128]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 136]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 752], rax",
            "adc qword ptr [{s} + 760], rdx",
            "mov qword ptr [{s} + 720], {w6}",
            "mov qword ptr [{s} + 728], {w7}",
            "mov qword ptr [{s} + 736], {c}",
            "mov qword ptr [{s} + 744], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 720]",
            "adc {w1}, qword ptr [{s} + 728]",
            "adc {w2}, qword ptr [{s} + 736]",
            "adc {w3}, qword ptr [{s} + 744]",
            "adc {w4}, qword ptr [{s} + 752]",
            "adc {w5}, qword ptr [{s} + 760]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 328], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 328]",
            "mov qword ptr [{s} + 288], {k}",
            "mov qword ptr [{s} + 296], {w1}",
            "mov qword ptr [{s} + 304], {w2}",
            "mov qword ptr [{s} + 312], {w3}",
            "mov qword ptr [{s} + 320], {w4}",
            "mov qword ptr [{s} + 328], {w5}",
            // gamma = sqr(Y)
            "mov rdx, qword ptr [{s} + 48]",
            "mulx {w2}, {w1}, qword ptr [{s} + 56]",
            "mulx {w3}, rax, qword ptr [{s} + 64]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 72]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 80]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 88]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 56]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 64]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 72]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 80]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 88]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 64]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 72]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 80]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 88]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 72]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 80]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 88]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 80]",
            "mulx {k}, rax, qword ptr [{s} + 88]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 752], {k}",
            "mov qword ptr [{s} + 760], rax",
            "mov rdx, qword ptr [{s} + 48]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 56]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 64]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 72]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 80]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 88]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 752], rax",
            "adc qword ptr [{s} + 760], rdx",
            "mov qword ptr [{s} + 720], {w6}",
            "mov qword ptr [{s} + 728], {w7}",
            "mov qword ptr [{s} + 736], {c}",
            "mov qword ptr [{s} + 744], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 720]",
            "adc {w1}, qword ptr [{s} + 728]",
            "adc {w2}, qword ptr [{s} + 736]",
            "adc {w3}, qword ptr [{s} + 744]",
            "adc {w4}, qword ptr [{s} + 752]",
            "adc {w5}, qword ptr [{s} + 760]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 376], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 376]",
            "mov qword ptr [{s} + 336], {k}",
            "mov qword ptr [{s} + 344], {w1}",
            "mov qword ptr [{s} + 352], {w2}",
            "mov qword ptr [{s} + 360], {w3}",
            "mov qword ptr [{s} + 368], {w4}",
            "mov qword ptr [{s} + 376], {w5}",
            // m = sub(X, delta)
            "mov {w0}, qword ptr [{s}]",
            "mov {w1}, qword ptr [{s} + 8]",
            "mov {w2}, qword ptr [{s} + 16]",
            "mov {w3}, qword ptr [{s} + 24]",
            "mov {w4}, qword ptr [{s} + 32]",
            "mov {w5}, qword ptr [{s} + 40]",
            "sub {w0}, qword ptr [{s} + 288]",
            "sbb {w1}, qword ptr [{s} + 296]",
            "sbb {w2}, qword ptr [{s} + 304]",
            "sbb {w3}, qword ptr [{s} + 312]",
            "sbb {w4}, qword ptr [{s} + 320]",
            "sbb {w5}, qword ptr [{s} + 328]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 432], {w0}",
            "mov qword ptr [{s} + 440], {w1}",
            "mov qword ptr [{s} + 448], {w2}",
            "mov qword ptr [{s} + 456], {w3}",
            "mov qword ptr [{s} + 464], {w4}",
            "mov qword ptr [{s} + 472], {w5}",
            // u = add(X, delta)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s}]",
            "mov {w1}, qword ptr [{s} + 8]",
            "mov {w2}, qword ptr [{s} + 16]",
            "mov {w3}, qword ptr [{s} + 24]",
            "mov {w4}, qword ptr [{s} + 32]",
            "mov {w5}, qword ptr [{s} + 40]",
            "add {w0}, qword ptr [{s} + 288]",
            "adc {w1}, qword ptr [{s} + 296]",
            "adc {w2}, qword ptr [{s} + 304]",
            "adc {w3}, qword ptr [{s} + 312]",
            "adc {w4}, qword ptr [{s} + 320]",
            "adc {w5}, qword ptr [{s} + 328]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 520], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 520]",
            "mov qword ptr [{s} + 480], {w0}",
            "mov qword ptr [{s} + 488], {w1}",
            "mov qword ptr [{s} + 496], {w2}",
            "mov qword ptr [{s} + 504], {w3}",
            "mov qword ptr [{s} + 512], {w4}",
            "mov qword ptr [{s} + 520], {w5}",
            // m = mul(m, u)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 480]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 488]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 496]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 504]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 512]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 520]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 472], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 472]",
            "mov qword ptr [{s} + 432], {w6}",
            "mov qword ptr [{s} + 440], {w7}",
            "mov qword ptr [{s} + 448], {w0}",
            "mov qword ptr [{s} + 456], {w1}",
            "mov qword ptr [{s} + 464], {w2}",
            "mov qword ptr [{s} + 472], {w3}",
            // beta = mul(X, gamma)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 336]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 344]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 352]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 360]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 368]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 376]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 424], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 424]",
            "mov qword ptr [{s} + 384], {w6}",
            "mov qword ptr [{s} + 392], {w7}",
            "mov qword ptr [{s} + 400], {w0}",
            "mov qword ptr [{s} + 408], {w1}",
            "mov qword ptr [{s} + 416], {w2}",
            "mov qword ptr [{s} + 424], {w3}",
            // m2 = sqr(m)
            "mov rdx, qword ptr [{s} + 432]",
            "mulx {w2}, {w1}, qword ptr [{s} + 440]",
            "mulx {w3}, rax, qword ptr [{s} + 448]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 456]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 464]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 472]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 440]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 448]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 456]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 464]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 472]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 448]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 456]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 464]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 472]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 456]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 464]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 472]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 464]",
            "mulx {k}, rax, qword ptr [{s} + 472]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 752], {k}",
            "mov qword ptr [{s} + 760], rax",
            "mov rdx, qword ptr [{s} + 432]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 440]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 448]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 456]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 464]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 472]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 752], rax",
            "adc qword ptr [{s} + 760], rdx",
            "mov qword ptr [{s} + 720], {w6}",
            "mov qword ptr [{s} + 728], {w7}",
            "mov qword ptr [{s} + 736], {c}",
            "mov qword ptr [{s} + 744], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 720]",
            "adc {w1}, qword ptr [{s} + 728]",
            "adc {w2}, qword ptr [{s} + 736]",
            "adc {w3}, qword ptr [{s} + 744]",
            "adc {w4}, qword ptr [{s} + 752]",
            "adc {w5}, qword ptr [{s} + 760]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 568], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 568]",
            "mov qword ptr [{s} + 528], {k}",
            "mov qword ptr [{s} + 536], {w1}",
            "mov qword ptr [{s} + 544], {w2}",
            "mov qword ptr [{s} + 552], {w3}",
            "mov qword ptr [{s} + 560], {w4}",
            "mov qword ptr [{s} + 568], {w5}",
            // w = add(Y, Z)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 48]",
            "mov {w1}, qword ptr [{s} + 56]",
            "mov {w2}, qword ptr [{s} + 64]",
            "mov {w3}, qword ptr [{s} + 72]",
            "mov {w4}, qword ptr [{s} + 80]",
            "mov {w5}, qword ptr [{s} + 88]",
            "add {w0}, qword ptr [{s} + 96]",
            "adc {w1}, qword ptr [{s} + 104]",
            "adc {w2}, qword ptr [{s} + 112]",
            "adc {w3}, qword ptr [{s} + 120]",
            "adc {w4}, qword ptr [{s} + 128]",
            "adc {w5}, qword ptr [{s} + 136]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 664], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 664]",
            "mov qword ptr [{s} + 624], {w0}",
            "mov qword ptr [{s} + 632], {w1}",
            "mov qword ptr [{s} + 640], {w2}",
            "mov qword ptr [{s} + 648], {w3}",
            "mov qword ptr [{s} + 656], {w4}",
            "mov qword ptr [{s} + 664], {w5}",
            // w = sqr(w)
            "mov rdx, qword ptr [{s} + 624]",
            "mulx {w2}, {w1}, qword ptr [{s} + 632]",
            "mulx {w3}, rax, qword ptr [{s} + 640]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 648]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 656]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 664]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 632]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 640]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 648]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 656]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 664]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 640]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 648]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 656]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 664]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 648]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 656]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 664]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 656]",
            "mulx {k}, rax, qword ptr [{s} + 664]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 752], {k}",
            "mov qword ptr [{s} + 760], rax",
            "mov rdx, qword ptr [{s} + 624]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 632]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 640]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 648]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 656]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 664]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 752], rax",
            "adc qword ptr [{s} + 760], rdx",
            "mov qword ptr [{s} + 720], {w6}",
            "mov qword ptr [{s} + 728], {w7}",
            "mov qword ptr [{s} + 736], {c}",
            "mov qword ptr [{s} + 744], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 720]",
            "adc {w1}, qword ptr [{s} + 728]",
            "adc {w2}, qword ptr [{s} + 736]",
            "adc {w3}, qword ptr [{s} + 744]",
            "adc {w4}, qword ptr [{s} + 752]",
            "adc {w5}, qword ptr [{s} + 760]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 664], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 664]",
            "mov qword ptr [{s} + 624], {k}",
            "mov qword ptr [{s} + 632], {w1}",
            "mov qword ptr [{s} + 640], {w2}",
            "mov qword ptr [{s} + 648], {w3}",
            "mov qword ptr [{s} + 656], {w4}",
            "mov qword ptr [{s} + 664], {w5}",
            // d = 12 * beta - 9 * m2
            "mov {w0:e}, 0xffffffff",
            "sub {w0}, qword ptr [{s} + 528]",
            "mov {w1}, 0xffffffff00000000",
            "sbb {w1}, qword ptr [{s} + 536]",
            "mov {w2}, -2",
            "sbb {w2}, qword ptr [{s} + 544]",
            "mov {w3}, -1",
            "sbb {w3}, qword ptr [{s} + 552]",
            "mov {w4}, -1",
            "sbb {w4}, qword ptr [{s} + 560]",
            "mov {w5}, -1",
            "sbb {w5}, qword ptr [{s} + 568]",
            "mov edx, 9",
            "mulx {w7}, {w0}, {w0}",
            "mulx {c}, {w1}, {w1}",
            "add {w1}, {w7}",
            "mulx {w7}, {w2}, {w2}",
            "adc {w2}, {c}",
            "mulx {c}, {w3}, {w3}",
            "adc {w3}, {w7}",
            "mulx {w7}, {w4}, {w4}",
            "adc {w4}, {c}",
            "mulx {w6}, {w5}, {w5}",
            "adc {w5}, {w7}",
            "adc {w6}, 0",
            "mov edx, 12",
            "xor {c:e}, {c:e}",
            "mulx {w7}, rax, qword ptr [{s} + 384]",
            "adcx {w0}, rax",
            "adox {w1}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 392]",
            "adcx {w1}, rax",
            "adox {w2}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 400]",
            "adcx {w2}, rax",
            "adox {w3}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 408]",
            "adcx {w3}, rax",
            "adox {w4}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 416]",
            "adcx {w4}, rax",
            "adox {w5}, {w7}",
            "mulx rdx, rax, qword ptr [{s} + 424]",
            "adcx {w5}, rax",
            "adox rdx, {w6}",
            "adcx rdx, {c}",
            "add rdx, 1",
            "mov rax, 0xffffffff00000001",
            "mulx {w7}, rax, rax",
            "mov {c}, rdx",
            "shl {c}, 32",
            "sub {c}, rdx",
            "add {c}, {w7}",
            "add {w0}, rax",
            "adc {w1}, {c}",
            "adc {w2}, rdx",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "adc {w5}, 0",
            "cmc",
            "sbb rax, rax",
            "mov rdx, 0xffffffff00000001",
            "and rdx, rax",
            "mov {c:e}, 0xffffffff",
            "and {c}, rax",
            "and eax, 1",
            "sub {w0}, rdx",
            "sbb {w1}, {c}",
            "sbb {w2}, rax",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov qword ptr [{s} + 576], {w0}",
            "mov qword ptr [{s} + 584], {w1}",
            "mov qword ptr [{s} + 592], {w2}",
            "mov qword ptr [{s} + 600], {w3}",
            "mov qword ptr [{s} + 608], {w4}",
            "mov qword ptr [{s} + 616], {w5}",
            // g2 = sqr(gamma)
            "mov rdx, qword ptr [{s} + 336]",
            "mulx {w2}, {w1}, qword ptr [{s} + 344]",
            "mulx {w3}, rax, qword ptr [{s} + 352]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 360]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 368]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 376]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 344]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 352]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 360]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 368]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 376]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 352]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 360]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 368]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 376]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 360]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 368]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 376]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 368]",
            "mulx {k}, rax, qword ptr [{s} + 376]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 752], {k}",
            "mov qword ptr [{s} + 760], rax",
            "mov rdx, qword ptr [{s} + 336]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 344]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 352]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 360]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 368]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 376]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 752], rax",
            "adc qword ptr [{s} + 760], rdx",
            "mov qword ptr [{s} + 720], {w6}",
            "mov qword ptr [{s} + 728], {w7}",
            "mov qword ptr [{s} + 736], {c}",
            "mov qword ptr [{s} + 744], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 720]",
            "adc {w1}, qword ptr [{s} + 728]",
            "adc {w2}, qword ptr [{s} + 736]",
            "adc {w3}, qword ptr [{s} + 744]",
            "adc {w4}, qword ptr [{s} + 752]",
            "adc {w5}, qword ptr [{s} + 760]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 712], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 712]",
            "mov qword ptr [{s} + 672], {k}",
            "mov qword ptr [{s} + 680], {w1}",
            "mov qword ptr [{s} + 688], {w2}",
            "mov qword ptr [{s} + 696], {w3}",
            "mov qword ptr [{s} + 704], {w4}",
            "mov qword ptr [{s} + 712], {w5}",
            // u = mul(d, m)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 432]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 576]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 584]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 592]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 600]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 608]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 616]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 440]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 576]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 584]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 592]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 600]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 608]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 616]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 448]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 576]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 584]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 592]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 600]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 608]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 616]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 456]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 576]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 584]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 592]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 600]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 608]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 616]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 464]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 576]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 584]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 592]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 600]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 608]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 616]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 472]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 576]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 584]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 592]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 600]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 608]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 616]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 520], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 520]",
            "mov qword ptr [{s} + 480], {w6}",
            "mov qword ptr [{s} + 488], {w7}",
            "mov qword ptr [{s} + 496], {w0}",
            "mov qword ptr [{s} + 504], {w1}",
            "mov qword ptr [{s} + 512], {w2}",
            "mov qword ptr [{s} + 520], {w3}",
            // X3 = 4 * beta - 1 * d
            "mov {w0:e}, 0xffffffff",
            "sub {w0}, qword ptr [{s} + 576]",
            "mov {w1}, 0xffffffff00000000",
            "sbb {w1}, qword ptr [{s} + 584]",
            "mov {w2}, -2",
            "sbb {w2}, qword ptr [{s} + 592]",
            "mov {w3}, -1",
            "sbb {w3}, qword ptr [{s} + 600]",
            "mov {w4}, -1",
            "sbb {w4}, qword ptr [{s} + 608]",
            "mov {w5}, -1",
            "sbb {w5}, qword ptr [{s} + 616]",
            "xor {w6:e}, {w6:e}",
            "mov edx, 4",
            "xor {c:e}, {c:e}",
            "mulx {w7}, rax, qword ptr [{s} + 384]",
            "adcx {w0}, rax",
            "adox {w1}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 392]",
            "adcx {w1}, rax",
            "adox {w2}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 400]",
            "adcx {w2}, rax",
            "adox {w3}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 408]",
            "adcx {w3}, rax",
            "adox {w4}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 416]",
            "adcx {w4}, rax",
            "adox {w5}, {w7}",
            "mulx rdx, rax, qword ptr [{s} + 424]",
            "adcx {w5}, rax",
            "adox rdx, {w6}",
            "adcx rdx, {c}",
            "add rdx, 1",
            "mov rax, 0xffffffff00000001",
            "mulx {w7}, rax, rax",
            "mov {c}, rdx",
            "shl {c}, 32",
            "sub {c}, rdx",
            "add {c}, {w7}",
            "add {w0}, rax",
            "adc {w1}, {c}",
            "adc {w2}, rdx",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "adc {w5}, 0",
            "cmc",
            "sbb rax, rax",
            "mov rdx, 0xffffffff00000001",
            "and rdx, rax",
            "mov {c:e}, 0xffffffff",
            "and {c}, rax",
            "and eax, 1",
            "sub {w0}, rdx",
            "sbb {w1}, {c}",
            "sbb {w2}, rax",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov qword ptr [{s} + 144], {w0}",
            "mov qword ptr [{s} + 152], {w1}",
            "mov qword ptr [{s} + 160], {w2}",
            "mov qword ptr [{s} + 168], {w3}",
            "mov qword ptr [{s} + 176], {w4}",
            "mov qword ptr [{s} + 184], {w5}",
            // w = sub(w, gamma)
            "mov {w0}, qword ptr [{s} + 624]",
            "mov {w1}, qword ptr [{s} + 632]",
            "mov {w2}, qword ptr [{s} + 640]",
            "mov {w3}, qword ptr [{s} + 648]",
            "mov {w4}, qword ptr [{s} + 656]",
            "mov {w5}, qword ptr [{s} + 664]",
            "sub {w0}, qword ptr [{s} + 336]",
            "sbb {w1}, qword ptr [{s} + 344]",
            "sbb {w2}, qword ptr [{s} + 352]",
            "sbb {w3}, qword ptr [{s} + 360]",
            "sbb {w4}, qword ptr [{s} + 368]",
            "sbb {w5}, qword ptr [{s} + 376]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 624], {w0}",
            "mov qword ptr [{s} + 632], {w1}",
            "mov qword ptr [{s} + 640], {w2}",
            "mov qword ptr [{s} + 648], {w3}",
            "mov qword ptr [{s} + 656], {w4}",
            "mov qword ptr [{s} + 664], {w5}",
            // Z3 = sub(w, delta)
            "mov {w0}, qword ptr [{s} + 624]",
            "mov {w1}, qword ptr [{s} + 632]",
            "mov {w2}, qword ptr [{s} + 640]",
            "mov {w3}, qword ptr [{s} + 648]",
            "mov {w4}, qword ptr [{s} + 656]",
            "mov {w5}, qword ptr [{s} + 664]",
            "sub {w0}, qword ptr [{s} + 288]",
            "sbb {w1}, qword ptr [{s} + 296]",
            "sbb {w2}, qword ptr [{s} + 304]",
            "sbb {w3}, qword ptr [{s} + 312]",
            "sbb {w4}, qword ptr [{s} + 320]",
            "sbb {w5}, qword ptr [{s} + 328]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 240], {w0}",
            "mov qword ptr [{s} + 248], {w1}",
            "mov qword ptr [{s} + 256], {w2}",
            "mov qword ptr [{s} + 264], {w3}",
            "mov qword ptr [{s} + 272], {w4}",
            "mov qword ptr [{s} + 280], {w5}",
            // Y3 = 3 * u - 8 * g2
            "mov {w0:e}, 0xffffffff",
            "sub {w0}, qword ptr [{s} + 672]",
            "mov {w1}, 0xffffffff00000000",
            "sbb {w1}, qword ptr [{s} + 680]",
            "mov {w2}, -2",
            "sbb {w2}, qword ptr [{s} + 688]",
            "mov {w3}, -1",
            "sbb {w3}, qword ptr [{s} + 696]",
            "mov {w4}, -1",
            "sbb {w4}, qword ptr [{s} + 704]",
            "mov {w5}, -1",
            "sbb {w5}, qword ptr [{s} + 712]",
            "mov {w6}, {w5}",
            "shr {w6}, 61",
            "shld {w5}, {w4}, 3",
            "shld {w4}, {w3}, 3",
            "shld {w3}, {w2}, 3",
            "shld {w2}, {w1}, 3",
            "shld {w1}, {w0}, 3",
            "shl {w0}, 3",
            "mov edx, 3",
            "xor {c:e}, {c:e}",
            "mulx {w7}, rax, qword ptr [{s} + 480]",
            "adcx {w0}, rax",
            "adox {w1}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 488]",
            "adcx {w1}, rax",
            "adox {w2}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 496]",
            "adcx {w2}, rax",
            "adox {w3}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 504]",
            "adcx {w3}, rax",
            "adox {w4}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 512]",
            "adcx {w4}, rax",
            "adox {w5}, {w7}",
            "mulx rdx, rax, qword ptr [{s} + 520]",
            "adcx {w5}, rax",
            "adox rdx, {w6}",
            "adcx rdx, {c}",
            "add rdx, 1",
            "mov rax, 0xffffffff00000001",
            "mulx {w7}, rax, rax",
            "mov {c}, rdx",
            "shl {c}, 32",
            "sub {c}, rdx",
            "add {c}, {w7}",
            "add {w0}, rax",
            "adc {w1}, {c}",
            "adc {w2}, rdx",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "adc {w5}, 0",
            "cmc",
            "sbb rax, rax",
            "mov rdx, 0xffffffff00000001",
            "and rdx, rax",
            "mov {c:e}, 0xffffffff",
            "and {c}, rax",
            "and eax, 1",
            "sub {w0}, rdx",
            "sbb {w1}, {c}",
            "sbb {w2}, rax",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov qword ptr [{s} + 192], {w0}",
            "mov qword ptr [{s} + 200], {w1}",
            "mov qword ptr [{s} + 208], {w2}",
            "mov qword ptr [{s} + 216], {w3}",
            "mov qword ptr [{s} + 224], {w4}",
            "mov qword ptr [{s} + 232], {w5}",
            s = in(reg) buf.as_mut_ptr(),
            w0 = out(reg) _,
            w1 = out(reg) _,
            w2 = out(reg) _,
            w3 = out(reg) _,
            w4 = out(reg) _,
            w5 = out(reg) _,
            w6 = out(reg) _,
            w7 = out(reg) _,
            c = out(reg) _,
            k = out(reg) _,
            out("rax") _,
            out("rdx") _,
            options(nostack),
        );
    }
}
#[allow(dead_code)]
pub const DOUBLE_X3: usize = 3;
#[allow(dead_code)]
pub const DOUBLE_Y3: usize = 4;
#[allow(dead_code)]
pub const DOUBLE_Z3: usize = 5;

/// `double_adx` with X3, Y3 and Z3 written over X, Y and Z: the scalar multiplication's doubling, on a buffer whose first slots are its accumulator.
///
/// Slots, 48 bytes each, in order: `X`, `Y`, `Z`, `delta`, `gamma`, `beta`, `m`, `u`, `m2`, `d`, `w`, `g2`, `scratch`.
///
/// # Safety
///
/// The CPU must support BMI2 and ADX; see `super::p256_x86::has_adx`.
/// Every input slot must hold a value below p.
#[inline(never)]
pub unsafe fn double_in_place_adx(buf: &mut [MaybeUninit<u64>; 78]) {
    // SAFETY: the caller guarantees the instructions exist. The block reads
    // and writes only the buffer behind `{s}`, a live, exclusive reference
    // of 624 bytes, at offsets below that; it writes only the
    // registers it declares, and does not touch the stack. It reads no slot
    // before writing it but the inputs, which the caller has written.
    unsafe {
        asm!(
            ".p2align 6",
            // delta = sqr(Z)
            "mov rdx, qword ptr [{s} + 96]",
            "mulx {w2}, {w1}, qword ptr [{s} + 104]",
            "mulx {w3}, rax, qword ptr [{s} + 112]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 120]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 128]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 136]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 104]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 112]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 120]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 112]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 120]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 120]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 128]",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 608], {k}",
            "mov qword ptr [{s} + 616], rax",
            "mov rdx, qword ptr [{s} + 96]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 104]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 112]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 120]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 128]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 136]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 608], rax",
            "adc qword ptr [{s} + 616], rdx",
            "mov qword ptr [{s} + 576], {w6}",
            "mov qword ptr [{s} + 584], {w7}",
            "mov qword ptr [{s} + 592], {c}",
            "mov qword ptr [{s} + 600], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 576]",
            "adc {w1}, qword ptr [{s} + 584]",
            "adc {w2}, qword ptr [{s} + 592]",
            "adc {w3}, qword ptr [{s} + 600]",
            "adc {w4}, qword ptr [{s} + 608]",
            "adc {w5}, qword ptr [{s} + 616]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 184], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 184]",
            "mov qword ptr [{s} + 144], {k}",
            "mov qword ptr [{s} + 152], {w1}",
            "mov qword ptr [{s} + 160], {w2}",
            "mov qword ptr [{s} + 168], {w3}",
            "mov qword ptr [{s} + 176], {w4}",
            "mov qword ptr [{s} + 184], {w5}",
            // gamma = sqr(Y)
            "mov rdx, qword ptr [{s} + 48]",
            "mulx {w2}, {w1}, qword ptr [{s} + 56]",
            "mulx {w3}, rax, qword ptr [{s} + 64]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 72]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 80]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 88]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 56]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 64]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 72]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 80]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 88]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 64]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 72]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 80]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 88]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 72]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 80]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 88]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 80]",
            "mulx {k}, rax, qword ptr [{s} + 88]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 608], {k}",
            "mov qword ptr [{s} + 616], rax",
            "mov rdx, qword ptr [{s} + 48]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 56]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 64]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 72]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 80]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 88]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 608], rax",
            "adc qword ptr [{s} + 616], rdx",
            "mov qword ptr [{s} + 576], {w6}",
            "mov qword ptr [{s} + 584], {w7}",
            "mov qword ptr [{s} + 592], {c}",
            "mov qword ptr [{s} + 600], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 576]",
            "adc {w1}, qword ptr [{s} + 584]",
            "adc {w2}, qword ptr [{s} + 592]",
            "adc {w3}, qword ptr [{s} + 600]",
            "adc {w4}, qword ptr [{s} + 608]",
            "adc {w5}, qword ptr [{s} + 616]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 232], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 232]",
            "mov qword ptr [{s} + 192], {k}",
            "mov qword ptr [{s} + 200], {w1}",
            "mov qword ptr [{s} + 208], {w2}",
            "mov qword ptr [{s} + 216], {w3}",
            "mov qword ptr [{s} + 224], {w4}",
            "mov qword ptr [{s} + 232], {w5}",
            // m = sub(X, delta)
            "mov {w0}, qword ptr [{s}]",
            "mov {w1}, qword ptr [{s} + 8]",
            "mov {w2}, qword ptr [{s} + 16]",
            "mov {w3}, qword ptr [{s} + 24]",
            "mov {w4}, qword ptr [{s} + 32]",
            "mov {w5}, qword ptr [{s} + 40]",
            "sub {w0}, qword ptr [{s} + 144]",
            "sbb {w1}, qword ptr [{s} + 152]",
            "sbb {w2}, qword ptr [{s} + 160]",
            "sbb {w3}, qword ptr [{s} + 168]",
            "sbb {w4}, qword ptr [{s} + 176]",
            "sbb {w5}, qword ptr [{s} + 184]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 288], {w0}",
            "mov qword ptr [{s} + 296], {w1}",
            "mov qword ptr [{s} + 304], {w2}",
            "mov qword ptr [{s} + 312], {w3}",
            "mov qword ptr [{s} + 320], {w4}",
            "mov qword ptr [{s} + 328], {w5}",
            // u = add(X, delta)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s}]",
            "mov {w1}, qword ptr [{s} + 8]",
            "mov {w2}, qword ptr [{s} + 16]",
            "mov {w3}, qword ptr [{s} + 24]",
            "mov {w4}, qword ptr [{s} + 32]",
            "mov {w5}, qword ptr [{s} + 40]",
            "add {w0}, qword ptr [{s} + 144]",
            "adc {w1}, qword ptr [{s} + 152]",
            "adc {w2}, qword ptr [{s} + 160]",
            "adc {w3}, qword ptr [{s} + 168]",
            "adc {w4}, qword ptr [{s} + 176]",
            "adc {w5}, qword ptr [{s} + 184]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 376], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 376]",
            "mov qword ptr [{s} + 336], {w0}",
            "mov qword ptr [{s} + 344], {w1}",
            "mov qword ptr [{s} + 352], {w2}",
            "mov qword ptr [{s} + 360], {w3}",
            "mov qword ptr [{s} + 368], {w4}",
            "mov qword ptr [{s} + 376], {w5}",
            // m = mul(m, u)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 336]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 288]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 296]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 304]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 312]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 320]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 328]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 344]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 288]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 296]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 304]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 312]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 320]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 328]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 352]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 288]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 296]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 304]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 312]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 320]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 328]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 360]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 288]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 296]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 304]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 312]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 320]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 328]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 368]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 288]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 296]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 304]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 312]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 320]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 328]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 376]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 288]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 296]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 304]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 312]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 320]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 328]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 328], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 328]",
            "mov qword ptr [{s} + 288], {w6}",
            "mov qword ptr [{s} + 296], {w7}",
            "mov qword ptr [{s} + 304], {w0}",
            "mov qword ptr [{s} + 312], {w1}",
            "mov qword ptr [{s} + 320], {w2}",
            "mov qword ptr [{s} + 328], {w3}",
            // beta = mul(X, gamma)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 192]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 200]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 208]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 216]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 224]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 232]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 280], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 280]",
            "mov qword ptr [{s} + 240], {w6}",
            "mov qword ptr [{s} + 248], {w7}",
            "mov qword ptr [{s} + 256], {w0}",
            "mov qword ptr [{s} + 264], {w1}",
            "mov qword ptr [{s} + 272], {w2}",
            "mov qword ptr [{s} + 280], {w3}",
            // m2 = sqr(m)
            "mov rdx, qword ptr [{s} + 288]",
            "mulx {w2}, {w1}, qword ptr [{s} + 296]",
            "mulx {w3}, rax, qword ptr [{s} + 304]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 312]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 320]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 328]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 296]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 304]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 312]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 320]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 328]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 304]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 312]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 320]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 328]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 312]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 320]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 328]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 320]",
            "mulx {k}, rax, qword ptr [{s} + 328]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 608], {k}",
            "mov qword ptr [{s} + 616], rax",
            "mov rdx, qword ptr [{s} + 288]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 296]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 304]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 312]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 320]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 328]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 608], rax",
            "adc qword ptr [{s} + 616], rdx",
            "mov qword ptr [{s} + 576], {w6}",
            "mov qword ptr [{s} + 584], {w7}",
            "mov qword ptr [{s} + 592], {c}",
            "mov qword ptr [{s} + 600], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 576]",
            "adc {w1}, qword ptr [{s} + 584]",
            "adc {w2}, qword ptr [{s} + 592]",
            "adc {w3}, qword ptr [{s} + 600]",
            "adc {w4}, qword ptr [{s} + 608]",
            "adc {w5}, qword ptr [{s} + 616]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 424], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 424]",
            "mov qword ptr [{s} + 384], {k}",
            "mov qword ptr [{s} + 392], {w1}",
            "mov qword ptr [{s} + 400], {w2}",
            "mov qword ptr [{s} + 408], {w3}",
            "mov qword ptr [{s} + 416], {w4}",
            "mov qword ptr [{s} + 424], {w5}",
            // w = add(Y, Z)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 48]",
            "mov {w1}, qword ptr [{s} + 56]",
            "mov {w2}, qword ptr [{s} + 64]",
            "mov {w3}, qword ptr [{s} + 72]",
            "mov {w4}, qword ptr [{s} + 80]",
            "mov {w5}, qword ptr [{s} + 88]",
            "add {w0}, qword ptr [{s} + 96]",
            "adc {w1}, qword ptr [{s} + 104]",
            "adc {w2}, qword ptr [{s} + 112]",
            "adc {w3}, qword ptr [{s} + 120]",
            "adc {w4}, qword ptr [{s} + 128]",
            "adc {w5}, qword ptr [{s} + 136]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 520], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 520]",
            "mov qword ptr [{s} + 480], {w0}",
            "mov qword ptr [{s} + 488], {w1}",
            "mov qword ptr [{s} + 496], {w2}",
            "mov qword ptr [{s} + 504], {w3}",
            "mov qword ptr [{s} + 512], {w4}",
            "mov qword ptr [{s} + 520], {w5}",
            // w = sqr(w)
            "mov rdx, qword ptr [{s} + 480]",
            "mulx {w2}, {w1}, qword ptr [{s} + 488]",
            "mulx {w3}, rax, qword ptr [{s} + 496]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 504]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 512]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 520]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 488]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 496]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 504]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 512]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 520]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 496]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 504]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 512]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 520]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 504]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 512]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 520]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 512]",
            "mulx {k}, rax, qword ptr [{s} + 520]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 608], {k}",
            "mov qword ptr [{s} + 616], rax",
            "mov rdx, qword ptr [{s} + 480]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 488]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 496]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 504]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 512]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 520]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 608], rax",
            "adc qword ptr [{s} + 616], rdx",
            "mov qword ptr [{s} + 576], {w6}",
            "mov qword ptr [{s} + 584], {w7}",
            "mov qword ptr [{s} + 592], {c}",
            "mov qword ptr [{s} + 600], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 576]",
            "adc {w1}, qword ptr [{s} + 584]",
            "adc {w2}, qword ptr [{s} + 592]",
            "adc {w3}, qword ptr [{s} + 600]",
            "adc {w4}, qword ptr [{s} + 608]",
            "adc {w5}, qword ptr [{s} + 616]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 520], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 520]",
            "mov qword ptr [{s} + 480], {k}",
            "mov qword ptr [{s} + 488], {w1}",
            "mov qword ptr [{s} + 496], {w2}",
            "mov qword ptr [{s} + 504], {w3}",
            "mov qword ptr [{s} + 512], {w4}",
            "mov qword ptr [{s} + 520], {w5}",
            // d = 12 * beta - 9 * m2
            "mov {w0:e}, 0xffffffff",
            "sub {w0}, qword ptr [{s} + 384]",
            "mov {w1}, 0xffffffff00000000",
            "sbb {w1}, qword ptr [{s} + 392]",
            "mov {w2}, -2",
            "sbb {w2}, qword ptr [{s} + 400]",
            "mov {w3}, -1",
            "sbb {w3}, qword ptr [{s} + 408]",
            "mov {w4}, -1",
            "sbb {w4}, qword ptr [{s} + 416]",
            "mov {w5}, -1",
            "sbb {w5}, qword ptr [{s} + 424]",
            "mov edx, 9",
            "mulx {w7}, {w0}, {w0}",
            "mulx {c}, {w1}, {w1}",
            "add {w1}, {w7}",
            "mulx {w7}, {w2}, {w2}",
            "adc {w2}, {c}",
            "mulx {c}, {w3}, {w3}",
            "adc {w3}, {w7}",
            "mulx {w7}, {w4}, {w4}",
            "adc {w4}, {c}",
            "mulx {w6}, {w5}, {w5}",
            "adc {w5}, {w7}",
            "adc {w6}, 0",
            "mov edx, 12",
            "xor {c:e}, {c:e}",
            "mulx {w7}, rax, qword ptr [{s} + 240]",
            "adcx {w0}, rax",
            "adox {w1}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 248]",
            "adcx {w1}, rax",
            "adox {w2}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 256]",
            "adcx {w2}, rax",
            "adox {w3}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 264]",
            "adcx {w3}, rax",
            "adox {w4}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 272]",
            "adcx {w4}, rax",
            "adox {w5}, {w7}",
            "mulx rdx, rax, qword ptr [{s} + 280]",
            "adcx {w5}, rax",
            "adox rdx, {w6}",
            "adcx rdx, {c}",
            "add rdx, 1",
            "mov rax, 0xffffffff00000001",
            "mulx {w7}, rax, rax",
            "mov {c}, rdx",
            "shl {c}, 32",
            "sub {c}, rdx",
            "add {c}, {w7}",
            "add {w0}, rax",
            "adc {w1}, {c}",
            "adc {w2}, rdx",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "adc {w5}, 0",
            "cmc",
            "sbb rax, rax",
            "mov rdx, 0xffffffff00000001",
            "and rdx, rax",
            "mov {c:e}, 0xffffffff",
            "and {c}, rax",
            "and eax, 1",
            "sub {w0}, rdx",
            "sbb {w1}, {c}",
            "sbb {w2}, rax",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov qword ptr [{s} + 432], {w0}",
            "mov qword ptr [{s} + 440], {w1}",
            "mov qword ptr [{s} + 448], {w2}",
            "mov qword ptr [{s} + 456], {w3}",
            "mov qword ptr [{s} + 464], {w4}",
            "mov qword ptr [{s} + 472], {w5}",
            // g2 = sqr(gamma)
            "mov rdx, qword ptr [{s} + 192]",
            "mulx {w2}, {w1}, qword ptr [{s} + 200]",
            "mulx {w3}, rax, qword ptr [{s} + 208]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 216]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 224]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 232]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 200]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 208]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 216]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 224]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 232]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 208]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 216]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 224]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 232]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 216]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 224]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 232]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 224]",
            "mulx {k}, rax, qword ptr [{s} + 232]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 608], {k}",
            "mov qword ptr [{s} + 616], rax",
            "mov rdx, qword ptr [{s} + 192]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 200]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 208]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 216]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 224]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 232]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 608], rax",
            "adc qword ptr [{s} + 616], rdx",
            "mov qword ptr [{s} + 576], {w6}",
            "mov qword ptr [{s} + 584], {w7}",
            "mov qword ptr [{s} + 592], {c}",
            "mov qword ptr [{s} + 600], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 576]",
            "adc {w1}, qword ptr [{s} + 584]",
            "adc {w2}, qword ptr [{s} + 592]",
            "adc {w3}, qword ptr [{s} + 600]",
            "adc {w4}, qword ptr [{s} + 608]",
            "adc {w5}, qword ptr [{s} + 616]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 568], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 568]",
            "mov qword ptr [{s} + 528], {k}",
            "mov qword ptr [{s} + 536], {w1}",
            "mov qword ptr [{s} + 544], {w2}",
            "mov qword ptr [{s} + 552], {w3}",
            "mov qword ptr [{s} + 560], {w4}",
            "mov qword ptr [{s} + 568], {w5}",
            // u = mul(d, m)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 288]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 296]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 304]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 312]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 320]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 328]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 376], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 376]",
            "mov qword ptr [{s} + 336], {w6}",
            "mov qword ptr [{s} + 344], {w7}",
            "mov qword ptr [{s} + 352], {w0}",
            "mov qword ptr [{s} + 360], {w1}",
            "mov qword ptr [{s} + 368], {w2}",
            "mov qword ptr [{s} + 376], {w3}",
            // X = 4 * beta - 1 * d
            "mov {w0:e}, 0xffffffff",
            "sub {w0}, qword ptr [{s} + 432]",
            "mov {w1}, 0xffffffff00000000",
            "sbb {w1}, qword ptr [{s} + 440]",
            "mov {w2}, -2",
            "sbb {w2}, qword ptr [{s} + 448]",
            "mov {w3}, -1",
            "sbb {w3}, qword ptr [{s} + 456]",
            "mov {w4}, -1",
            "sbb {w4}, qword ptr [{s} + 464]",
            "mov {w5}, -1",
            "sbb {w5}, qword ptr [{s} + 472]",
            "xor {w6:e}, {w6:e}",
            "mov edx, 4",
            "xor {c:e}, {c:e}",
            "mulx {w7}, rax, qword ptr [{s} + 240]",
            "adcx {w0}, rax",
            "adox {w1}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 248]",
            "adcx {w1}, rax",
            "adox {w2}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 256]",
            "adcx {w2}, rax",
            "adox {w3}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 264]",
            "adcx {w3}, rax",
            "adox {w4}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 272]",
            "adcx {w4}, rax",
            "adox {w5}, {w7}",
            "mulx rdx, rax, qword ptr [{s} + 280]",
            "adcx {w5}, rax",
            "adox rdx, {w6}",
            "adcx rdx, {c}",
            "add rdx, 1",
            "mov rax, 0xffffffff00000001",
            "mulx {w7}, rax, rax",
            "mov {c}, rdx",
            "shl {c}, 32",
            "sub {c}, rdx",
            "add {c}, {w7}",
            "add {w0}, rax",
            "adc {w1}, {c}",
            "adc {w2}, rdx",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "adc {w5}, 0",
            "cmc",
            "sbb rax, rax",
            "mov rdx, 0xffffffff00000001",
            "and rdx, rax",
            "mov {c:e}, 0xffffffff",
            "and {c}, rax",
            "and eax, 1",
            "sub {w0}, rdx",
            "sbb {w1}, {c}",
            "sbb {w2}, rax",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov qword ptr [{s}], {w0}",
            "mov qword ptr [{s} + 8], {w1}",
            "mov qword ptr [{s} + 16], {w2}",
            "mov qword ptr [{s} + 24], {w3}",
            "mov qword ptr [{s} + 32], {w4}",
            "mov qword ptr [{s} + 40], {w5}",
            // w = sub(w, gamma)
            "mov {w0}, qword ptr [{s} + 480]",
            "mov {w1}, qword ptr [{s} + 488]",
            "mov {w2}, qword ptr [{s} + 496]",
            "mov {w3}, qword ptr [{s} + 504]",
            "mov {w4}, qword ptr [{s} + 512]",
            "mov {w5}, qword ptr [{s} + 520]",
            "sub {w0}, qword ptr [{s} + 192]",
            "sbb {w1}, qword ptr [{s} + 200]",
            "sbb {w2}, qword ptr [{s} + 208]",
            "sbb {w3}, qword ptr [{s} + 216]",
            "sbb {w4}, qword ptr [{s} + 224]",
            "sbb {w5}, qword ptr [{s} + 232]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 480], {w0}",
            "mov qword ptr [{s} + 488], {w1}",
            "mov qword ptr [{s} + 496], {w2}",
            "mov qword ptr [{s} + 504], {w3}",
            "mov qword ptr [{s} + 512], {w4}",
            "mov qword ptr [{s} + 520], {w5}",
            // Z = sub(w, delta)
            "mov {w0}, qword ptr [{s} + 480]",
            "mov {w1}, qword ptr [{s} + 488]",
            "mov {w2}, qword ptr [{s} + 496]",
            "mov {w3}, qword ptr [{s} + 504]",
            "mov {w4}, qword ptr [{s} + 512]",
            "mov {w5}, qword ptr [{s} + 520]",
            "sub {w0}, qword ptr [{s} + 144]",
            "sbb {w1}, qword ptr [{s} + 152]",
            "sbb {w2}, qword ptr [{s} + 160]",
            "sbb {w3}, qword ptr [{s} + 168]",
            "sbb {w4}, qword ptr [{s} + 176]",
            "sbb {w5}, qword ptr [{s} + 184]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 96], {w0}",
            "mov qword ptr [{s} + 104], {w1}",
            "mov qword ptr [{s} + 112], {w2}",
            "mov qword ptr [{s} + 120], {w3}",
            "mov qword ptr [{s} + 128], {w4}",
            "mov qword ptr [{s} + 136], {w5}",
            // Y = 3 * u - 8 * g2
            "mov {w0:e}, 0xffffffff",
            "sub {w0}, qword ptr [{s} + 528]",
            "mov {w1}, 0xffffffff00000000",
            "sbb {w1}, qword ptr [{s} + 536]",
            "mov {w2}, -2",
            "sbb {w2}, qword ptr [{s} + 544]",
            "mov {w3}, -1",
            "sbb {w3}, qword ptr [{s} + 552]",
            "mov {w4}, -1",
            "sbb {w4}, qword ptr [{s} + 560]",
            "mov {w5}, -1",
            "sbb {w5}, qword ptr [{s} + 568]",
            "mov {w6}, {w5}",
            "shr {w6}, 61",
            "shld {w5}, {w4}, 3",
            "shld {w4}, {w3}, 3",
            "shld {w3}, {w2}, 3",
            "shld {w2}, {w1}, 3",
            "shld {w1}, {w0}, 3",
            "shl {w0}, 3",
            "mov edx, 3",
            "xor {c:e}, {c:e}",
            "mulx {w7}, rax, qword ptr [{s} + 336]",
            "adcx {w0}, rax",
            "adox {w1}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 344]",
            "adcx {w1}, rax",
            "adox {w2}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 352]",
            "adcx {w2}, rax",
            "adox {w3}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 360]",
            "adcx {w3}, rax",
            "adox {w4}, {w7}",
            "mulx {w7}, rax, qword ptr [{s} + 368]",
            "adcx {w4}, rax",
            "adox {w5}, {w7}",
            "mulx rdx, rax, qword ptr [{s} + 376]",
            "adcx {w5}, rax",
            "adox rdx, {w6}",
            "adcx rdx, {c}",
            "add rdx, 1",
            "mov rax, 0xffffffff00000001",
            "mulx {w7}, rax, rax",
            "mov {c}, rdx",
            "shl {c}, 32",
            "sub {c}, rdx",
            "add {c}, {w7}",
            "add {w0}, rax",
            "adc {w1}, {c}",
            "adc {w2}, rdx",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "adc {w5}, 0",
            "cmc",
            "sbb rax, rax",
            "mov rdx, 0xffffffff00000001",
            "and rdx, rax",
            "mov {c:e}, 0xffffffff",
            "and {c}, rax",
            "and eax, 1",
            "sub {w0}, rdx",
            "sbb {w1}, {c}",
            "sbb {w2}, rax",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov qword ptr [{s} + 48], {w0}",
            "mov qword ptr [{s} + 56], {w1}",
            "mov qword ptr [{s} + 64], {w2}",
            "mov qword ptr [{s} + 72], {w3}",
            "mov qword ptr [{s} + 80], {w4}",
            "mov qword ptr [{s} + 88], {w5}",
            s = in(reg) buf.as_mut_ptr(),
            w0 = out(reg) _,
            w1 = out(reg) _,
            w2 = out(reg) _,
            w3 = out(reg) _,
            w4 = out(reg) _,
            w5 = out(reg) _,
            w6 = out(reg) _,
            w7 = out(reg) _,
            c = out(reg) _,
            k = out(reg) _,
            out("rax") _,
            out("rdx") _,
            options(nostack),
        );
    }
}
#[allow(dead_code)]
pub const DOUBLE_IN_PLACE_X: usize = 0;
#[allow(dead_code)]
pub const DOUBLE_IN_PLACE_Y: usize = 1;
#[allow(dead_code)]
pub const DOUBLE_IN_PLACE_Z: usize = 2;

/// Jacobian addition, `add-2007-bl`, as `Point::add_formula`; also leaves `H` and `R`.
///
/// Slots, 48 bytes each, in order: `X1`, `Y1`, `Z1`, `X2`, `Y2`, `Z2`, `X3`, `Y3`, `Z3`, `H`, `R`, `z1z1`, `z2z2`, `u1`, `u2`, `s1`, `s2`, `i`, `j`, `v`, `t1`, `t2`, `t3`, `t4`, `scratch`.
///
/// # Safety
///
/// The CPU must support BMI2 and ADX; see `super::p256_x86::has_adx`.
/// Every input slot must hold a value below p.
#[inline(never)]
pub unsafe fn add_adx(buf: &mut [MaybeUninit<u64>; 150]) {
    // SAFETY: the caller guarantees the instructions exist. The block reads
    // and writes only the buffer behind `{s}`, a live, exclusive reference
    // of 1200 bytes, at offsets below that; it writes only the
    // registers it declares, and does not touch the stack. It reads no slot
    // before writing it but the inputs, which the caller has written.
    unsafe {
        asm!(
            ".p2align 6",
            // z1z1 = sqr(Z1)
            "mov rdx, qword ptr [{s} + 96]",
            "mulx {w2}, {w1}, qword ptr [{s} + 104]",
            "mulx {w3}, rax, qword ptr [{s} + 112]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 120]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 128]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 136]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 104]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 112]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 120]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 112]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 120]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 120]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 128]",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 1184], {k}",
            "mov qword ptr [{s} + 1192], rax",
            "mov rdx, qword ptr [{s} + 96]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 104]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 112]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 120]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 128]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 136]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 1184], rax",
            "adc qword ptr [{s} + 1192], rdx",
            "mov qword ptr [{s} + 1152], {w6}",
            "mov qword ptr [{s} + 1160], {w7}",
            "mov qword ptr [{s} + 1168], {c}",
            "mov qword ptr [{s} + 1176], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 1152]",
            "adc {w1}, qword ptr [{s} + 1160]",
            "adc {w2}, qword ptr [{s} + 1168]",
            "adc {w3}, qword ptr [{s} + 1176]",
            "adc {w4}, qword ptr [{s} + 1184]",
            "adc {w5}, qword ptr [{s} + 1192]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 568], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 568]",
            "mov qword ptr [{s} + 528], {k}",
            "mov qword ptr [{s} + 536], {w1}",
            "mov qword ptr [{s} + 544], {w2}",
            "mov qword ptr [{s} + 552], {w3}",
            "mov qword ptr [{s} + 560], {w4}",
            "mov qword ptr [{s} + 568], {w5}",
            // z2z2 = sqr(Z2)
            "mov rdx, qword ptr [{s} + 240]",
            "mulx {w2}, {w1}, qword ptr [{s} + 248]",
            "mulx {w3}, rax, qword ptr [{s} + 256]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 264]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 272]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 280]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 248]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 256]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 264]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 272]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 280]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 256]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 264]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 272]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 280]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 264]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 272]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 280]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 272]",
            "mulx {k}, rax, qword ptr [{s} + 280]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 1184], {k}",
            "mov qword ptr [{s} + 1192], rax",
            "mov rdx, qword ptr [{s} + 240]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 248]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 256]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 264]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 272]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 280]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 1184], rax",
            "adc qword ptr [{s} + 1192], rdx",
            "mov qword ptr [{s} + 1152], {w6}",
            "mov qword ptr [{s} + 1160], {w7}",
            "mov qword ptr [{s} + 1168], {c}",
            "mov qword ptr [{s} + 1176], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 1152]",
            "adc {w1}, qword ptr [{s} + 1160]",
            "adc {w2}, qword ptr [{s} + 1168]",
            "adc {w3}, qword ptr [{s} + 1176]",
            "adc {w4}, qword ptr [{s} + 1184]",
            "adc {w5}, qword ptr [{s} + 1192]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 616], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 616]",
            "mov qword ptr [{s} + 576], {k}",
            "mov qword ptr [{s} + 584], {w1}",
            "mov qword ptr [{s} + 592], {w2}",
            "mov qword ptr [{s} + 600], {w3}",
            "mov qword ptr [{s} + 608], {w4}",
            "mov qword ptr [{s} + 616], {w5}",
            // t1 = mul(Y1, Z2)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 240]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 248]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 256]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 264]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 272]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 280]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 1000], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 1000]",
            "mov qword ptr [{s} + 960], {w6}",
            "mov qword ptr [{s} + 968], {w7}",
            "mov qword ptr [{s} + 976], {w0}",
            "mov qword ptr [{s} + 984], {w1}",
            "mov qword ptr [{s} + 992], {w2}",
            "mov qword ptr [{s} + 1000], {w3}",
            // u2 = mul(X2, z1z1)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 528]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 536]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 544]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 552]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 560]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 568]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 712], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 712]",
            "mov qword ptr [{s} + 672], {w6}",
            "mov qword ptr [{s} + 680], {w7}",
            "mov qword ptr [{s} + 688], {w0}",
            "mov qword ptr [{s} + 696], {w1}",
            "mov qword ptr [{s} + 704], {w2}",
            "mov qword ptr [{s} + 712], {w3}",
            // u1 = mul(X1, z2z2)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 576]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 584]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 592]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 600]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 608]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 616]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 664], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 664]",
            "mov qword ptr [{s} + 624], {w6}",
            "mov qword ptr [{s} + 632], {w7}",
            "mov qword ptr [{s} + 640], {w0}",
            "mov qword ptr [{s} + 648], {w1}",
            "mov qword ptr [{s} + 656], {w2}",
            "mov qword ptr [{s} + 664], {w3}",
            // t2 = mul(Y2, Z1)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 96]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 104]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 112]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 120]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 128]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 136]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 1048], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 1048]",
            "mov qword ptr [{s} + 1008], {w6}",
            "mov qword ptr [{s} + 1016], {w7}",
            "mov qword ptr [{s} + 1024], {w0}",
            "mov qword ptr [{s} + 1032], {w1}",
            "mov qword ptr [{s} + 1040], {w2}",
            "mov qword ptr [{s} + 1048], {w3}",
            // H = sub(u2, u1)
            "mov {w0}, qword ptr [{s} + 672]",
            "mov {w1}, qword ptr [{s} + 680]",
            "mov {w2}, qword ptr [{s} + 688]",
            "mov {w3}, qword ptr [{s} + 696]",
            "mov {w4}, qword ptr [{s} + 704]",
            "mov {w5}, qword ptr [{s} + 712]",
            "sub {w0}, qword ptr [{s} + 624]",
            "sbb {w1}, qword ptr [{s} + 632]",
            "sbb {w2}, qword ptr [{s} + 640]",
            "sbb {w3}, qword ptr [{s} + 648]",
            "sbb {w4}, qword ptr [{s} + 656]",
            "sbb {w5}, qword ptr [{s} + 664]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 432], {w0}",
            "mov qword ptr [{s} + 440], {w1}",
            "mov qword ptr [{s} + 448], {w2}",
            "mov qword ptr [{s} + 456], {w3}",
            "mov qword ptr [{s} + 464], {w4}",
            "mov qword ptr [{s} + 472], {w5}",
            // i = add(H, H)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 432]",
            "mov {w1}, qword ptr [{s} + 440]",
            "mov {w2}, qword ptr [{s} + 448]",
            "mov {w3}, qword ptr [{s} + 456]",
            "mov {w4}, qword ptr [{s} + 464]",
            "mov {w5}, qword ptr [{s} + 472]",
            "add {w0}, qword ptr [{s} + 432]",
            "adc {w1}, qword ptr [{s} + 440]",
            "adc {w2}, qword ptr [{s} + 448]",
            "adc {w3}, qword ptr [{s} + 456]",
            "adc {w4}, qword ptr [{s} + 464]",
            "adc {w5}, qword ptr [{s} + 472]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 856], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 856]",
            "mov qword ptr [{s} + 816], {w0}",
            "mov qword ptr [{s} + 824], {w1}",
            "mov qword ptr [{s} + 832], {w2}",
            "mov qword ptr [{s} + 840], {w3}",
            "mov qword ptr [{s} + 848], {w4}",
            "mov qword ptr [{s} + 856], {w5}",
            // i = sqr(i)
            "mov rdx, qword ptr [{s} + 816]",
            "mulx {w2}, {w1}, qword ptr [{s} + 824]",
            "mulx {w3}, rax, qword ptr [{s} + 832]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 840]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 848]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 856]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 824]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 832]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 840]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 848]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 856]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 832]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 840]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 848]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 856]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 840]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 848]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 856]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 848]",
            "mulx {k}, rax, qword ptr [{s} + 856]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 1184], {k}",
            "mov qword ptr [{s} + 1192], rax",
            "mov rdx, qword ptr [{s} + 816]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 824]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 832]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 840]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 848]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 856]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 1184], rax",
            "adc qword ptr [{s} + 1192], rdx",
            "mov qword ptr [{s} + 1152], {w6}",
            "mov qword ptr [{s} + 1160], {w7}",
            "mov qword ptr [{s} + 1168], {c}",
            "mov qword ptr [{s} + 1176], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 1152]",
            "adc {w1}, qword ptr [{s} + 1160]",
            "adc {w2}, qword ptr [{s} + 1168]",
            "adc {w3}, qword ptr [{s} + 1176]",
            "adc {w4}, qword ptr [{s} + 1184]",
            "adc {w5}, qword ptr [{s} + 1192]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 856], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 856]",
            "mov qword ptr [{s} + 816], {k}",
            "mov qword ptr [{s} + 824], {w1}",
            "mov qword ptr [{s} + 832], {w2}",
            "mov qword ptr [{s} + 840], {w3}",
            "mov qword ptr [{s} + 848], {w4}",
            "mov qword ptr [{s} + 856], {w5}",
            // s1 = mul(t1, z2z2)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 576]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 960]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 968]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 976]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 984]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 992]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1000]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 584]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 960]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 968]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 976]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 984]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 992]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1000]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 592]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 960]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 968]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 976]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 984]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 992]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1000]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 600]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 960]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 968]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 976]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 984]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 992]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1000]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 608]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 960]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 968]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 976]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 984]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 992]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1000]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 616]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 960]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 968]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 976]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 984]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 992]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1000]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 760], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 760]",
            "mov qword ptr [{s} + 720], {w6}",
            "mov qword ptr [{s} + 728], {w7}",
            "mov qword ptr [{s} + 736], {w0}",
            "mov qword ptr [{s} + 744], {w1}",
            "mov qword ptr [{s} + 752], {w2}",
            "mov qword ptr [{s} + 760], {w3}",
            // v = mul(u1, i)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 816]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 624]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 632]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 640]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 648]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 656]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 664]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 824]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 624]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 632]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 640]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 648]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 656]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 664]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 832]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 624]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 632]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 640]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 648]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 656]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 664]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 840]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 624]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 632]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 640]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 648]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 656]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 664]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 848]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 624]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 632]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 640]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 648]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 656]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 664]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 856]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 624]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 632]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 640]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 648]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 656]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 664]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 952], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 952]",
            "mov qword ptr [{s} + 912], {w6}",
            "mov qword ptr [{s} + 920], {w7}",
            "mov qword ptr [{s} + 928], {w0}",
            "mov qword ptr [{s} + 936], {w1}",
            "mov qword ptr [{s} + 944], {w2}",
            "mov qword ptr [{s} + 952], {w3}",
            // s2 = mul(t2, z1z1)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 528]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 1008]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1016]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1024]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1032]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1040]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1048]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 536]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 1008]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1016]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1024]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1032]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1040]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1048]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 544]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 1008]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1016]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1024]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1032]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1040]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1048]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 552]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 1008]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1016]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1024]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1032]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1040]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1048]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 560]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 1008]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1016]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1024]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1032]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1040]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1048]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 568]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 1008]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1016]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1024]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1032]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1040]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 1048]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 808], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 808]",
            "mov qword ptr [{s} + 768], {w6}",
            "mov qword ptr [{s} + 776], {w7}",
            "mov qword ptr [{s} + 784], {w0}",
            "mov qword ptr [{s} + 792], {w1}",
            "mov qword ptr [{s} + 800], {w2}",
            "mov qword ptr [{s} + 808], {w3}",
            // j = mul(H, i)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 816]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 824]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 832]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 840]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 848]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 856]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 904], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 904]",
            "mov qword ptr [{s} + 864], {w6}",
            "mov qword ptr [{s} + 872], {w7}",
            "mov qword ptr [{s} + 880], {w0}",
            "mov qword ptr [{s} + 888], {w1}",
            "mov qword ptr [{s} + 896], {w2}",
            "mov qword ptr [{s} + 904], {w3}",
            // R = sub(s2, s1)
            "mov {w0}, qword ptr [{s} + 768]",
            "mov {w1}, qword ptr [{s} + 776]",
            "mov {w2}, qword ptr [{s} + 784]",
            "mov {w3}, qword ptr [{s} + 792]",
            "mov {w4}, qword ptr [{s} + 800]",
            "mov {w5}, qword ptr [{s} + 808]",
            "sub {w0}, qword ptr [{s} + 720]",
            "sbb {w1}, qword ptr [{s} + 728]",
            "sbb {w2}, qword ptr [{s} + 736]",
            "sbb {w3}, qword ptr [{s} + 744]",
            "sbb {w4}, qword ptr [{s} + 752]",
            "sbb {w5}, qword ptr [{s} + 760]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 480], {w0}",
            "mov qword ptr [{s} + 488], {w1}",
            "mov qword ptr [{s} + 496], {w2}",
            "mov qword ptr [{s} + 504], {w3}",
            "mov qword ptr [{s} + 512], {w4}",
            "mov qword ptr [{s} + 520], {w5}",
            // R = add(R, R)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 480]",
            "mov {w1}, qword ptr [{s} + 488]",
            "mov {w2}, qword ptr [{s} + 496]",
            "mov {w3}, qword ptr [{s} + 504]",
            "mov {w4}, qword ptr [{s} + 512]",
            "mov {w5}, qword ptr [{s} + 520]",
            "add {w0}, qword ptr [{s} + 480]",
            "adc {w1}, qword ptr [{s} + 488]",
            "adc {w2}, qword ptr [{s} + 496]",
            "adc {w3}, qword ptr [{s} + 504]",
            "adc {w4}, qword ptr [{s} + 512]",
            "adc {w5}, qword ptr [{s} + 520]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 520], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 520]",
            "mov qword ptr [{s} + 480], {w0}",
            "mov qword ptr [{s} + 488], {w1}",
            "mov qword ptr [{s} + 496], {w2}",
            "mov qword ptr [{s} + 504], {w3}",
            "mov qword ptr [{s} + 512], {w4}",
            "mov qword ptr [{s} + 520], {w5}",
            // X3 = sqr(R)
            "mov rdx, qword ptr [{s} + 480]",
            "mulx {w2}, {w1}, qword ptr [{s} + 488]",
            "mulx {w3}, rax, qword ptr [{s} + 496]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 504]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 512]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 520]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 488]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 496]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 504]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 512]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 520]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 496]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 504]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 512]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 520]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 504]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 512]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 520]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 512]",
            "mulx {k}, rax, qword ptr [{s} + 520]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 1184], {k}",
            "mov qword ptr [{s} + 1192], rax",
            "mov rdx, qword ptr [{s} + 480]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 488]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 496]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 504]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 512]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 520]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 1184], rax",
            "adc qword ptr [{s} + 1192], rdx",
            "mov qword ptr [{s} + 1152], {w6}",
            "mov qword ptr [{s} + 1160], {w7}",
            "mov qword ptr [{s} + 1168], {c}",
            "mov qword ptr [{s} + 1176], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 1152]",
            "adc {w1}, qword ptr [{s} + 1160]",
            "adc {w2}, qword ptr [{s} + 1168]",
            "adc {w3}, qword ptr [{s} + 1176]",
            "adc {w4}, qword ptr [{s} + 1184]",
            "adc {w5}, qword ptr [{s} + 1192]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 328], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 328]",
            "mov qword ptr [{s} + 288], {k}",
            "mov qword ptr [{s} + 296], {w1}",
            "mov qword ptr [{s} + 304], {w2}",
            "mov qword ptr [{s} + 312], {w3}",
            "mov qword ptr [{s} + 320], {w4}",
            "mov qword ptr [{s} + 328], {w5}",
            // Z3 = add(Z1, Z2)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 96]",
            "mov {w1}, qword ptr [{s} + 104]",
            "mov {w2}, qword ptr [{s} + 112]",
            "mov {w3}, qword ptr [{s} + 120]",
            "mov {w4}, qword ptr [{s} + 128]",
            "mov {w5}, qword ptr [{s} + 136]",
            "add {w0}, qword ptr [{s} + 240]",
            "adc {w1}, qword ptr [{s} + 248]",
            "adc {w2}, qword ptr [{s} + 256]",
            "adc {w3}, qword ptr [{s} + 264]",
            "adc {w4}, qword ptr [{s} + 272]",
            "adc {w5}, qword ptr [{s} + 280]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 424], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 424]",
            "mov qword ptr [{s} + 384], {w0}",
            "mov qword ptr [{s} + 392], {w1}",
            "mov qword ptr [{s} + 400], {w2}",
            "mov qword ptr [{s} + 408], {w3}",
            "mov qword ptr [{s} + 416], {w4}",
            "mov qword ptr [{s} + 424], {w5}",
            // Z3 = sqr(Z3)
            "mov rdx, qword ptr [{s} + 384]",
            "mulx {w2}, {w1}, qword ptr [{s} + 392]",
            "mulx {w3}, rax, qword ptr [{s} + 400]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 408]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 416]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 424]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 392]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 400]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 408]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 416]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 424]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 400]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 408]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 416]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 424]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 408]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 416]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 424]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 416]",
            "mulx {k}, rax, qword ptr [{s} + 424]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 1184], {k}",
            "mov qword ptr [{s} + 1192], rax",
            "mov rdx, qword ptr [{s} + 384]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 392]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 400]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 408]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 416]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 424]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 1184], rax",
            "adc qword ptr [{s} + 1192], rdx",
            "mov qword ptr [{s} + 1152], {w6}",
            "mov qword ptr [{s} + 1160], {w7}",
            "mov qword ptr [{s} + 1168], {c}",
            "mov qword ptr [{s} + 1176], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 1152]",
            "adc {w1}, qword ptr [{s} + 1160]",
            "adc {w2}, qword ptr [{s} + 1168]",
            "adc {w3}, qword ptr [{s} + 1176]",
            "adc {w4}, qword ptr [{s} + 1184]",
            "adc {w5}, qword ptr [{s} + 1192]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 424], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 424]",
            "mov qword ptr [{s} + 384], {k}",
            "mov qword ptr [{s} + 392], {w1}",
            "mov qword ptr [{s} + 400], {w2}",
            "mov qword ptr [{s} + 408], {w3}",
            "mov qword ptr [{s} + 416], {w4}",
            "mov qword ptr [{s} + 424], {w5}",
            // t3 = mul(s1, j)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 864]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 720]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 728]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 736]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 744]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 752]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 760]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 872]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 720]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 728]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 736]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 744]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 752]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 760]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 880]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 720]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 728]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 736]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 744]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 752]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 760]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 888]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 720]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 728]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 736]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 744]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 752]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 760]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 896]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 720]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 728]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 736]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 744]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 752]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 760]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 904]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 720]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 728]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 736]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 744]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 752]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 760]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 1096], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 1096]",
            "mov qword ptr [{s} + 1056], {w6}",
            "mov qword ptr [{s} + 1064], {w7}",
            "mov qword ptr [{s} + 1072], {w0}",
            "mov qword ptr [{s} + 1080], {w1}",
            "mov qword ptr [{s} + 1088], {w2}",
            "mov qword ptr [{s} + 1096], {w3}",
            // t1 = add(v, v)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 912]",
            "mov {w1}, qword ptr [{s} + 920]",
            "mov {w2}, qword ptr [{s} + 928]",
            "mov {w3}, qword ptr [{s} + 936]",
            "mov {w4}, qword ptr [{s} + 944]",
            "mov {w5}, qword ptr [{s} + 952]",
            "add {w0}, qword ptr [{s} + 912]",
            "adc {w1}, qword ptr [{s} + 920]",
            "adc {w2}, qword ptr [{s} + 928]",
            "adc {w3}, qword ptr [{s} + 936]",
            "adc {w4}, qword ptr [{s} + 944]",
            "adc {w5}, qword ptr [{s} + 952]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 1000], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 1000]",
            "mov qword ptr [{s} + 960], {w0}",
            "mov qword ptr [{s} + 968], {w1}",
            "mov qword ptr [{s} + 976], {w2}",
            "mov qword ptr [{s} + 984], {w3}",
            "mov qword ptr [{s} + 992], {w4}",
            "mov qword ptr [{s} + 1000], {w5}",
            // X3 = sub(X3, j)
            "mov {w0}, qword ptr [{s} + 288]",
            "mov {w1}, qword ptr [{s} + 296]",
            "mov {w2}, qword ptr [{s} + 304]",
            "mov {w3}, qword ptr [{s} + 312]",
            "mov {w4}, qword ptr [{s} + 320]",
            "mov {w5}, qword ptr [{s} + 328]",
            "sub {w0}, qword ptr [{s} + 864]",
            "sbb {w1}, qword ptr [{s} + 872]",
            "sbb {w2}, qword ptr [{s} + 880]",
            "sbb {w3}, qword ptr [{s} + 888]",
            "sbb {w4}, qword ptr [{s} + 896]",
            "sbb {w5}, qword ptr [{s} + 904]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 288], {w0}",
            "mov qword ptr [{s} + 296], {w1}",
            "mov qword ptr [{s} + 304], {w2}",
            "mov qword ptr [{s} + 312], {w3}",
            "mov qword ptr [{s} + 320], {w4}",
            "mov qword ptr [{s} + 328], {w5}",
            // X3 = sub(X3, t1)
            "mov {w0}, qword ptr [{s} + 288]",
            "mov {w1}, qword ptr [{s} + 296]",
            "mov {w2}, qword ptr [{s} + 304]",
            "mov {w3}, qword ptr [{s} + 312]",
            "mov {w4}, qword ptr [{s} + 320]",
            "mov {w5}, qword ptr [{s} + 328]",
            "sub {w0}, qword ptr [{s} + 960]",
            "sbb {w1}, qword ptr [{s} + 968]",
            "sbb {w2}, qword ptr [{s} + 976]",
            "sbb {w3}, qword ptr [{s} + 984]",
            "sbb {w4}, qword ptr [{s} + 992]",
            "sbb {w5}, qword ptr [{s} + 1000]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 288], {w0}",
            "mov qword ptr [{s} + 296], {w1}",
            "mov qword ptr [{s} + 304], {w2}",
            "mov qword ptr [{s} + 312], {w3}",
            "mov qword ptr [{s} + 320], {w4}",
            "mov qword ptr [{s} + 328], {w5}",
            // t4 = sub(v, X3)
            "mov {w0}, qword ptr [{s} + 912]",
            "mov {w1}, qword ptr [{s} + 920]",
            "mov {w2}, qword ptr [{s} + 928]",
            "mov {w3}, qword ptr [{s} + 936]",
            "mov {w4}, qword ptr [{s} + 944]",
            "mov {w5}, qword ptr [{s} + 952]",
            "sub {w0}, qword ptr [{s} + 288]",
            "sbb {w1}, qword ptr [{s} + 296]",
            "sbb {w2}, qword ptr [{s} + 304]",
            "sbb {w3}, qword ptr [{s} + 312]",
            "sbb {w4}, qword ptr [{s} + 320]",
            "sbb {w5}, qword ptr [{s} + 328]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 1104], {w0}",
            "mov qword ptr [{s} + 1112], {w1}",
            "mov qword ptr [{s} + 1120], {w2}",
            "mov qword ptr [{s} + 1128], {w3}",
            "mov qword ptr [{s} + 1136], {w4}",
            "mov qword ptr [{s} + 1144], {w5}",
            // Z3 = sub(Z3, z1z1)
            "mov {w0}, qword ptr [{s} + 384]",
            "mov {w1}, qword ptr [{s} + 392]",
            "mov {w2}, qword ptr [{s} + 400]",
            "mov {w3}, qword ptr [{s} + 408]",
            "mov {w4}, qword ptr [{s} + 416]",
            "mov {w5}, qword ptr [{s} + 424]",
            "sub {w0}, qword ptr [{s} + 528]",
            "sbb {w1}, qword ptr [{s} + 536]",
            "sbb {w2}, qword ptr [{s} + 544]",
            "sbb {w3}, qword ptr [{s} + 552]",
            "sbb {w4}, qword ptr [{s} + 560]",
            "sbb {w5}, qword ptr [{s} + 568]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 384], {w0}",
            "mov qword ptr [{s} + 392], {w1}",
            "mov qword ptr [{s} + 400], {w2}",
            "mov qword ptr [{s} + 408], {w3}",
            "mov qword ptr [{s} + 416], {w4}",
            "mov qword ptr [{s} + 424], {w5}",
            // Y3 = mul(R, t4)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 1104]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 480]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 488]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 496]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 504]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 512]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 520]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 1112]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 480]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 488]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 496]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 504]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 512]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 520]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 1120]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 480]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 488]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 496]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 504]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 512]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 520]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 1128]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 480]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 488]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 496]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 504]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 512]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 520]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 1136]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 480]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 488]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 496]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 504]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 512]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 520]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 1144]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 480]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 488]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 496]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 504]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 512]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 520]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 376], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 376]",
            "mov qword ptr [{s} + 336], {w6}",
            "mov qword ptr [{s} + 344], {w7}",
            "mov qword ptr [{s} + 352], {w0}",
            "mov qword ptr [{s} + 360], {w1}",
            "mov qword ptr [{s} + 368], {w2}",
            "mov qword ptr [{s} + 376], {w3}",
            // Z3 = sub(Z3, z2z2)
            "mov {w0}, qword ptr [{s} + 384]",
            "mov {w1}, qword ptr [{s} + 392]",
            "mov {w2}, qword ptr [{s} + 400]",
            "mov {w3}, qword ptr [{s} + 408]",
            "mov {w4}, qword ptr [{s} + 416]",
            "mov {w5}, qword ptr [{s} + 424]",
            "sub {w0}, qword ptr [{s} + 576]",
            "sbb {w1}, qword ptr [{s} + 584]",
            "sbb {w2}, qword ptr [{s} + 592]",
            "sbb {w3}, qword ptr [{s} + 600]",
            "sbb {w4}, qword ptr [{s} + 608]",
            "sbb {w5}, qword ptr [{s} + 616]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 384], {w0}",
            "mov qword ptr [{s} + 392], {w1}",
            "mov qword ptr [{s} + 400], {w2}",
            "mov qword ptr [{s} + 408], {w3}",
            "mov qword ptr [{s} + 416], {w4}",
            "mov qword ptr [{s} + 424], {w5}",
            // Z3 = mul(Z3, H)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 432]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 440]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 448]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 456]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 464]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 472]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 424], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 424]",
            "mov qword ptr [{s} + 384], {w6}",
            "mov qword ptr [{s} + 392], {w7}",
            "mov qword ptr [{s} + 400], {w0}",
            "mov qword ptr [{s} + 408], {w1}",
            "mov qword ptr [{s} + 416], {w2}",
            "mov qword ptr [{s} + 424], {w3}",
            // t3 = add(t3, t3)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 1056]",
            "mov {w1}, qword ptr [{s} + 1064]",
            "mov {w2}, qword ptr [{s} + 1072]",
            "mov {w3}, qword ptr [{s} + 1080]",
            "mov {w4}, qword ptr [{s} + 1088]",
            "mov {w5}, qword ptr [{s} + 1096]",
            "add {w0}, qword ptr [{s} + 1056]",
            "adc {w1}, qword ptr [{s} + 1064]",
            "adc {w2}, qword ptr [{s} + 1072]",
            "adc {w3}, qword ptr [{s} + 1080]",
            "adc {w4}, qword ptr [{s} + 1088]",
            "adc {w5}, qword ptr [{s} + 1096]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 1096], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 1096]",
            "mov qword ptr [{s} + 1056], {w0}",
            "mov qword ptr [{s} + 1064], {w1}",
            "mov qword ptr [{s} + 1072], {w2}",
            "mov qword ptr [{s} + 1080], {w3}",
            "mov qword ptr [{s} + 1088], {w4}",
            "mov qword ptr [{s} + 1096], {w5}",
            // Y3 = sub(Y3, t3)
            "mov {w0}, qword ptr [{s} + 336]",
            "mov {w1}, qword ptr [{s} + 344]",
            "mov {w2}, qword ptr [{s} + 352]",
            "mov {w3}, qword ptr [{s} + 360]",
            "mov {w4}, qword ptr [{s} + 368]",
            "mov {w5}, qword ptr [{s} + 376]",
            "sub {w0}, qword ptr [{s} + 1056]",
            "sbb {w1}, qword ptr [{s} + 1064]",
            "sbb {w2}, qword ptr [{s} + 1072]",
            "sbb {w3}, qword ptr [{s} + 1080]",
            "sbb {w4}, qword ptr [{s} + 1088]",
            "sbb {w5}, qword ptr [{s} + 1096]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 336], {w0}",
            "mov qword ptr [{s} + 344], {w1}",
            "mov qword ptr [{s} + 352], {w2}",
            "mov qword ptr [{s} + 360], {w3}",
            "mov qword ptr [{s} + 368], {w4}",
            "mov qword ptr [{s} + 376], {w5}",
            s = in(reg) buf.as_mut_ptr(),
            w0 = out(reg) _,
            w1 = out(reg) _,
            w2 = out(reg) _,
            w3 = out(reg) _,
            w4 = out(reg) _,
            w5 = out(reg) _,
            w6 = out(reg) _,
            w7 = out(reg) _,
            c = out(reg) _,
            k = out(reg) _,
            out("rax") _,
            out("rdx") _,
            options(nostack),
        );
    }
}
#[allow(dead_code)]
pub const ADD_X3: usize = 6;
#[allow(dead_code)]
pub const ADD_Y3: usize = 7;
#[allow(dead_code)]
pub const ADD_Z3: usize = 8;
#[allow(dead_code)]
pub const ADD_H: usize = 9;
#[allow(dead_code)]
pub const ADD_R: usize = 10;

/// Mixed addition, `madd-2007-bl`, as `Point::add_affine_formula`; also leaves `H` and `R`.
///
/// Slots, 48 bytes each, in order: `X1`, `Y1`, `Z1`, `x2`, `y2`, `X3`, `Y3`, `Z3`, `H`, `R`, `z1z1`, `u2`, `s2`, `hh`, `i`, `j`, `v`, `t`, `t2`, `t3`, `scratch`.
///
/// # Safety
///
/// The CPU must support BMI2 and ADX; see `super::p256_x86::has_adx`.
/// Every input slot must hold a value below p.
#[inline(never)]
pub unsafe fn add_affine_adx(buf: &mut [MaybeUninit<u64>; 126]) {
    // SAFETY: the caller guarantees the instructions exist. The block reads
    // and writes only the buffer behind `{s}`, a live, exclusive reference
    // of 1008 bytes, at offsets below that; it writes only the
    // registers it declares, and does not touch the stack. It reads no slot
    // before writing it but the inputs, which the caller has written.
    unsafe {
        asm!(
            ".p2align 6",
            // z1z1 = sqr(Z1)
            "mov rdx, qword ptr [{s} + 96]",
            "mulx {w2}, {w1}, qword ptr [{s} + 104]",
            "mulx {w3}, rax, qword ptr [{s} + 112]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 120]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 128]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 136]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 104]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 112]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 120]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 112]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 120]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 120]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 128]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 128]",
            "mulx {k}, rax, qword ptr [{s} + 136]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 992], {k}",
            "mov qword ptr [{s} + 1000], rax",
            "mov rdx, qword ptr [{s} + 96]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 104]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 112]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 120]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 128]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 136]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 992], rax",
            "adc qword ptr [{s} + 1000], rdx",
            "mov qword ptr [{s} + 960], {w6}",
            "mov qword ptr [{s} + 968], {w7}",
            "mov qword ptr [{s} + 976], {c}",
            "mov qword ptr [{s} + 984], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 960]",
            "adc {w1}, qword ptr [{s} + 968]",
            "adc {w2}, qword ptr [{s} + 976]",
            "adc {w3}, qword ptr [{s} + 984]",
            "adc {w4}, qword ptr [{s} + 992]",
            "adc {w5}, qword ptr [{s} + 1000]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 520], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 520]",
            "mov qword ptr [{s} + 480], {k}",
            "mov qword ptr [{s} + 488], {w1}",
            "mov qword ptr [{s} + 496], {w2}",
            "mov qword ptr [{s} + 504], {w3}",
            "mov qword ptr [{s} + 512], {w4}",
            "mov qword ptr [{s} + 520], {w5}",
            // t = mul(y2, Z1)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 96]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 104]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 112]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 120]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 128]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 136]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 192]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 200]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 208]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 216]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 224]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 232]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 856], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 856]",
            "mov qword ptr [{s} + 816], {w6}",
            "mov qword ptr [{s} + 824], {w7}",
            "mov qword ptr [{s} + 832], {w0}",
            "mov qword ptr [{s} + 840], {w1}",
            "mov qword ptr [{s} + 848], {w2}",
            "mov qword ptr [{s} + 856], {w3}",
            // u2 = mul(x2, z1z1)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 480]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 488]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 496]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 504]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 512]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 520]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 144]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 152]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 160]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 168]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 176]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 184]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 568], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 568]",
            "mov qword ptr [{s} + 528], {w6}",
            "mov qword ptr [{s} + 536], {w7}",
            "mov qword ptr [{s} + 544], {w0}",
            "mov qword ptr [{s} + 552], {w1}",
            "mov qword ptr [{s} + 560], {w2}",
            "mov qword ptr [{s} + 568], {w3}",
            // s2 = mul(t, z1z1)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 480]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 816]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 824]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 832]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 840]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 848]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 856]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 488]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 816]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 824]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 832]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 840]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 848]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 856]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 496]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 816]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 824]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 832]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 840]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 848]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 856]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 504]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 816]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 824]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 832]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 840]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 848]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 856]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 512]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 816]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 824]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 832]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 840]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 848]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 856]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 520]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 816]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 824]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 832]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 840]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 848]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 856]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 616], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 616]",
            "mov qword ptr [{s} + 576], {w6}",
            "mov qword ptr [{s} + 584], {w7}",
            "mov qword ptr [{s} + 592], {w0}",
            "mov qword ptr [{s} + 600], {w1}",
            "mov qword ptr [{s} + 608], {w2}",
            "mov qword ptr [{s} + 616], {w3}",
            // H = sub(u2, X1)
            "mov {w0}, qword ptr [{s} + 528]",
            "mov {w1}, qword ptr [{s} + 536]",
            "mov {w2}, qword ptr [{s} + 544]",
            "mov {w3}, qword ptr [{s} + 552]",
            "mov {w4}, qword ptr [{s} + 560]",
            "mov {w5}, qword ptr [{s} + 568]",
            "sub {w0}, qword ptr [{s}]",
            "sbb {w1}, qword ptr [{s} + 8]",
            "sbb {w2}, qword ptr [{s} + 16]",
            "sbb {w3}, qword ptr [{s} + 24]",
            "sbb {w4}, qword ptr [{s} + 32]",
            "sbb {w5}, qword ptr [{s} + 40]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 384], {w0}",
            "mov qword ptr [{s} + 392], {w1}",
            "mov qword ptr [{s} + 400], {w2}",
            "mov qword ptr [{s} + 408], {w3}",
            "mov qword ptr [{s} + 416], {w4}",
            "mov qword ptr [{s} + 424], {w5}",
            // hh = sqr(H)
            "mov rdx, qword ptr [{s} + 384]",
            "mulx {w2}, {w1}, qword ptr [{s} + 392]",
            "mulx {w3}, rax, qword ptr [{s} + 400]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 408]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 416]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 424]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 392]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 400]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 408]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 416]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 424]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 400]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 408]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 416]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 424]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 408]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 416]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 424]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 416]",
            "mulx {k}, rax, qword ptr [{s} + 424]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 992], {k}",
            "mov qword ptr [{s} + 1000], rax",
            "mov rdx, qword ptr [{s} + 384]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 392]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 400]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 408]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 416]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 424]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 992], rax",
            "adc qword ptr [{s} + 1000], rdx",
            "mov qword ptr [{s} + 960], {w6}",
            "mov qword ptr [{s} + 968], {w7}",
            "mov qword ptr [{s} + 976], {c}",
            "mov qword ptr [{s} + 984], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 960]",
            "adc {w1}, qword ptr [{s} + 968]",
            "adc {w2}, qword ptr [{s} + 976]",
            "adc {w3}, qword ptr [{s} + 984]",
            "adc {w4}, qword ptr [{s} + 992]",
            "adc {w5}, qword ptr [{s} + 1000]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 664], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 664]",
            "mov qword ptr [{s} + 624], {k}",
            "mov qword ptr [{s} + 632], {w1}",
            "mov qword ptr [{s} + 640], {w2}",
            "mov qword ptr [{s} + 648], {w3}",
            "mov qword ptr [{s} + 656], {w4}",
            "mov qword ptr [{s} + 664], {w5}",
            // R = sub(s2, Y1)
            "mov {w0}, qword ptr [{s} + 576]",
            "mov {w1}, qword ptr [{s} + 584]",
            "mov {w2}, qword ptr [{s} + 592]",
            "mov {w3}, qword ptr [{s} + 600]",
            "mov {w4}, qword ptr [{s} + 608]",
            "mov {w5}, qword ptr [{s} + 616]",
            "sub {w0}, qword ptr [{s} + 48]",
            "sbb {w1}, qword ptr [{s} + 56]",
            "sbb {w2}, qword ptr [{s} + 64]",
            "sbb {w3}, qword ptr [{s} + 72]",
            "sbb {w4}, qword ptr [{s} + 80]",
            "sbb {w5}, qword ptr [{s} + 88]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 432], {w0}",
            "mov qword ptr [{s} + 440], {w1}",
            "mov qword ptr [{s} + 448], {w2}",
            "mov qword ptr [{s} + 456], {w3}",
            "mov qword ptr [{s} + 464], {w4}",
            "mov qword ptr [{s} + 472], {w5}",
            // R = add(R, R)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 432]",
            "mov {w1}, qword ptr [{s} + 440]",
            "mov {w2}, qword ptr [{s} + 448]",
            "mov {w3}, qword ptr [{s} + 456]",
            "mov {w4}, qword ptr [{s} + 464]",
            "mov {w5}, qword ptr [{s} + 472]",
            "add {w0}, qword ptr [{s} + 432]",
            "adc {w1}, qword ptr [{s} + 440]",
            "adc {w2}, qword ptr [{s} + 448]",
            "adc {w3}, qword ptr [{s} + 456]",
            "adc {w4}, qword ptr [{s} + 464]",
            "adc {w5}, qword ptr [{s} + 472]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 472], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 472]",
            "mov qword ptr [{s} + 432], {w0}",
            "mov qword ptr [{s} + 440], {w1}",
            "mov qword ptr [{s} + 448], {w2}",
            "mov qword ptr [{s} + 456], {w3}",
            "mov qword ptr [{s} + 464], {w4}",
            "mov qword ptr [{s} + 472], {w5}",
            // X3 = sqr(R)
            "mov rdx, qword ptr [{s} + 432]",
            "mulx {w2}, {w1}, qword ptr [{s} + 440]",
            "mulx {w3}, rax, qword ptr [{s} + 448]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 456]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 464]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 472]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 440]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 448]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 456]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 464]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 472]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 448]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 456]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 464]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 472]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 456]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 464]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 472]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 464]",
            "mulx {k}, rax, qword ptr [{s} + 472]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 992], {k}",
            "mov qword ptr [{s} + 1000], rax",
            "mov rdx, qword ptr [{s} + 432]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 440]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 448]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 456]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 464]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 472]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 992], rax",
            "adc qword ptr [{s} + 1000], rdx",
            "mov qword ptr [{s} + 960], {w6}",
            "mov qword ptr [{s} + 968], {w7}",
            "mov qword ptr [{s} + 976], {c}",
            "mov qword ptr [{s} + 984], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 960]",
            "adc {w1}, qword ptr [{s} + 968]",
            "adc {w2}, qword ptr [{s} + 976]",
            "adc {w3}, qword ptr [{s} + 984]",
            "adc {w4}, qword ptr [{s} + 992]",
            "adc {w5}, qword ptr [{s} + 1000]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 280], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 280]",
            "mov qword ptr [{s} + 240], {k}",
            "mov qword ptr [{s} + 248], {w1}",
            "mov qword ptr [{s} + 256], {w2}",
            "mov qword ptr [{s} + 264], {w3}",
            "mov qword ptr [{s} + 272], {w4}",
            "mov qword ptr [{s} + 280], {w5}",
            // i = add(hh, hh)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 624]",
            "mov {w1}, qword ptr [{s} + 632]",
            "mov {w2}, qword ptr [{s} + 640]",
            "mov {w3}, qword ptr [{s} + 648]",
            "mov {w4}, qword ptr [{s} + 656]",
            "mov {w5}, qword ptr [{s} + 664]",
            "add {w0}, qword ptr [{s} + 624]",
            "adc {w1}, qword ptr [{s} + 632]",
            "adc {w2}, qword ptr [{s} + 640]",
            "adc {w3}, qword ptr [{s} + 648]",
            "adc {w4}, qword ptr [{s} + 656]",
            "adc {w5}, qword ptr [{s} + 664]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 712], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 712]",
            "mov qword ptr [{s} + 672], {w0}",
            "mov qword ptr [{s} + 680], {w1}",
            "mov qword ptr [{s} + 688], {w2}",
            "mov qword ptr [{s} + 696], {w3}",
            "mov qword ptr [{s} + 704], {w4}",
            "mov qword ptr [{s} + 712], {w5}",
            // i = add(i, i)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 672]",
            "mov {w1}, qword ptr [{s} + 680]",
            "mov {w2}, qword ptr [{s} + 688]",
            "mov {w3}, qword ptr [{s} + 696]",
            "mov {w4}, qword ptr [{s} + 704]",
            "mov {w5}, qword ptr [{s} + 712]",
            "add {w0}, qword ptr [{s} + 672]",
            "adc {w1}, qword ptr [{s} + 680]",
            "adc {w2}, qword ptr [{s} + 688]",
            "adc {w3}, qword ptr [{s} + 696]",
            "adc {w4}, qword ptr [{s} + 704]",
            "adc {w5}, qword ptr [{s} + 712]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 712], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 712]",
            "mov qword ptr [{s} + 672], {w0}",
            "mov qword ptr [{s} + 680], {w1}",
            "mov qword ptr [{s} + 688], {w2}",
            "mov qword ptr [{s} + 696], {w3}",
            "mov qword ptr [{s} + 704], {w4}",
            "mov qword ptr [{s} + 712], {w5}",
            // v = mul(X1, i)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 672]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 680]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 688]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 696]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 704]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 712]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s}]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 8]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 16]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 24]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 32]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 40]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 808], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 808]",
            "mov qword ptr [{s} + 768], {w6}",
            "mov qword ptr [{s} + 776], {w7}",
            "mov qword ptr [{s} + 784], {w0}",
            "mov qword ptr [{s} + 792], {w1}",
            "mov qword ptr [{s} + 800], {w2}",
            "mov qword ptr [{s} + 808], {w3}",
            // j = mul(H, i)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 672]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 680]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 688]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 696]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 704]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 712]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 384]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 392]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 400]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 408]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 416]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 424]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 760], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 760]",
            "mov qword ptr [{s} + 720], {w6}",
            "mov qword ptr [{s} + 728], {w7}",
            "mov qword ptr [{s} + 736], {w0}",
            "mov qword ptr [{s} + 744], {w1}",
            "mov qword ptr [{s} + 752], {w2}",
            "mov qword ptr [{s} + 760], {w3}",
            // Z3 = add(Z1, H)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 96]",
            "mov {w1}, qword ptr [{s} + 104]",
            "mov {w2}, qword ptr [{s} + 112]",
            "mov {w3}, qword ptr [{s} + 120]",
            "mov {w4}, qword ptr [{s} + 128]",
            "mov {w5}, qword ptr [{s} + 136]",
            "add {w0}, qword ptr [{s} + 384]",
            "adc {w1}, qword ptr [{s} + 392]",
            "adc {w2}, qword ptr [{s} + 400]",
            "adc {w3}, qword ptr [{s} + 408]",
            "adc {w4}, qword ptr [{s} + 416]",
            "adc {w5}, qword ptr [{s} + 424]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 376], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 376]",
            "mov qword ptr [{s} + 336], {w0}",
            "mov qword ptr [{s} + 344], {w1}",
            "mov qword ptr [{s} + 352], {w2}",
            "mov qword ptr [{s} + 360], {w3}",
            "mov qword ptr [{s} + 368], {w4}",
            "mov qword ptr [{s} + 376], {w5}",
            // Z3 = sqr(Z3)
            "mov rdx, qword ptr [{s} + 336]",
            "mulx {w2}, {w1}, qword ptr [{s} + 344]",
            "mulx {w3}, rax, qword ptr [{s} + 352]",
            "add {w2}, rax",
            "mulx {w4}, rax, qword ptr [{s} + 360]",
            "adc {w3}, rax",
            "mulx {w5}, rax, qword ptr [{s} + 368]",
            "adc {w4}, rax",
            "mulx {w6}, rax, qword ptr [{s} + 376]",
            "adc {w5}, rax",
            "adc {w6}, 0",
            "mov rdx, qword ptr [{s} + 344]",
            "xor {w7:e}, {w7:e}",
            "mulx {k}, rax, qword ptr [{s} + 352]",
            "adcx {w3}, rax",
            "adox {w4}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 360]",
            "adcx {w4}, rax",
            "adox {w5}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 368]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 376]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "mov rdx, qword ptr [{s} + 352]",
            "xor {c:e}, {c:e}",
            "mulx {k}, rax, qword ptr [{s} + 360]",
            "adcx {w5}, rax",
            "adox {w6}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 368]",
            "adcx {w6}, rax",
            "adox {w7}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 376]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mov eax, 0",
            "adcx {c}, rax",
            "mov rdx, qword ptr [{s} + 360]",
            "xor {w0:e}, {w0:e}",
            "mulx {k}, rax, qword ptr [{s} + 368]",
            "adcx {w7}, rax",
            "adox {c}, {k}",
            "mulx {k}, rax, qword ptr [{s} + 376]",
            "adcx {c}, rax",
            "adox {w0}, {k}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "mov rdx, qword ptr [{s} + 368]",
            "mulx {k}, rax, qword ptr [{s} + 376]",
            "add {w0}, rax",
            "adc {k}, 0",
            "add {w1}, {w1}",
            "adc {w2}, {w2}",
            "adc {w3}, {w3}",
            "adc {w4}, {w4}",
            "adc {w5}, {w5}",
            "adc {w6}, {w6}",
            "adc {w7}, {w7}",
            "adc {c}, {c}",
            "adc {w0}, {w0}",
            "adc {k}, {k}",
            "mov eax, 0",
            "adc eax, 0",
            "mov qword ptr [{s} + 992], {k}",
            "mov qword ptr [{s} + 1000], rax",
            "mov rdx, qword ptr [{s} + 336]",
            "mulx rdx, {k}, rdx",
            "add {w1}, rdx",
            "mov rdx, qword ptr [{s} + 344]",
            "mulx rdx, rax, rdx",
            "adc {w2}, rax",
            "adc {w3}, rdx",
            "mov rdx, qword ptr [{s} + 352]",
            "mulx rdx, rax, rdx",
            "adc {w4}, rax",
            "adc {w5}, rdx",
            "mov rdx, qword ptr [{s} + 360]",
            "mulx rdx, rax, rdx",
            "adc {w6}, rax",
            "adc {w7}, rdx",
            "mov rdx, qword ptr [{s} + 368]",
            "mulx rdx, rax, rdx",
            "adc {c}, rax",
            "adc {w0}, rdx",
            "mov rdx, qword ptr [{s} + 376]",
            "mulx rdx, rax, rdx",
            "adc qword ptr [{s} + 992], rax",
            "adc qword ptr [{s} + 1000], rdx",
            "mov qword ptr [{s} + 960], {w6}",
            "mov qword ptr [{s} + 968], {w7}",
            "mov qword ptr [{s} + 976], {c}",
            "mov qword ptr [{s} + 984], {w0}",
            "mov {w7}, 0xffffffff00000001",
            "mov {c:e}, 0xffffffff",
            "mov rdx, {k}",
            "shl rdx, 32",
            "add rdx, {k}",
            "mulx rax, {k}, {w7}",
            "mulx {k}, {w6}, {c}",
            "add rax, {w6}",
            "adc {k}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {k}",
            "sbb {w3}, {w6}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "mov {k}, rdx",
            "sbb {k}, 0",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {w7}",
            "mulx {w1}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w1}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {w6}",
            "sbb {w5}, 0",
            "sbb {k}, 0",
            "mov {w1}, rdx",
            "sbb {w1}, 0",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {w7}",
            "mulx {w2}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w2}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {w6}",
            "sbb {k}, 0",
            "sbb {w1}, 0",
            "mov {w2}, rdx",
            "sbb {w2}, 0",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {w7}",
            "mulx {w3}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w3}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {k}, {w6}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "mov {w3}, rdx",
            "sbb {w3}, 0",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {w7}",
            "mulx {w4}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w4}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {w5}, rax",
            "sbb {k}, {w4}",
            "sbb {w1}, {w6}",
            "sbb {w2}, 0",
            "sbb {w3}, 0",
            "mov {w4}, rdx",
            "sbb {w4}, 0",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {w7}",
            "mulx {w5}, {w6}, {c}",
            "add rax, {w6}",
            "adc {w5}, rdx",
            "mov {w6:e}, 0",
            "adc {w6}, 0",
            "sub {k}, rax",
            "sbb {w1}, {w5}",
            "sbb {w2}, {w6}",
            "sbb {w3}, 0",
            "sbb {w4}, 0",
            "mov {w5}, rdx",
            "sbb {w5}, 0",
            "xor {w7:e}, {w7:e}",
            "add {k}, qword ptr [{s} + 960]",
            "adc {w1}, qword ptr [{s} + 968]",
            "adc {w2}, qword ptr [{s} + 976]",
            "adc {w3}, qword ptr [{s} + 984]",
            "adc {w4}, qword ptr [{s} + 992]",
            "adc {w5}, qword ptr [{s} + 1000]",
            "adc {w7}, 0",
            "mov qword ptr [{s} + 376], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w0:e}, 1",
            "mov {w6}, {w3}",
            "mov {c}, {w4}",
            "add rax, {k}",
            "adc rdx, {w1}",
            "adc {w0}, {w2}",
            "adc {w6}, 0",
            "adc {c}, 0",
            "adc {w5}, 0",
            "adc {w7}, 0",
            "cmovnz {k}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w0}",
            "cmovnz {w3}, {w6}",
            "cmovnz {w4}, {c}",
            "cmovz {w5}, qword ptr [{s} + 376]",
            "mov qword ptr [{s} + 336], {k}",
            "mov qword ptr [{s} + 344], {w1}",
            "mov qword ptr [{s} + 352], {w2}",
            "mov qword ptr [{s} + 360], {w3}",
            "mov qword ptr [{s} + 368], {w4}",
            "mov qword ptr [{s} + 376], {w5}",
            // t2 = mul(Y1, j)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 720]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 728]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 736]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 744]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 752]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 760]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 48]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 56]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 64]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 72]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 80]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 88]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 904], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 904]",
            "mov qword ptr [{s} + 864], {w6}",
            "mov qword ptr [{s} + 872], {w7}",
            "mov qword ptr [{s} + 880], {w0}",
            "mov qword ptr [{s} + 888], {w1}",
            "mov qword ptr [{s} + 896], {w2}",
            "mov qword ptr [{s} + 904], {w3}",
            // t = add(v, v)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 768]",
            "mov {w1}, qword ptr [{s} + 776]",
            "mov {w2}, qword ptr [{s} + 784]",
            "mov {w3}, qword ptr [{s} + 792]",
            "mov {w4}, qword ptr [{s} + 800]",
            "mov {w5}, qword ptr [{s} + 808]",
            "add {w0}, qword ptr [{s} + 768]",
            "adc {w1}, qword ptr [{s} + 776]",
            "adc {w2}, qword ptr [{s} + 784]",
            "adc {w3}, qword ptr [{s} + 792]",
            "adc {w4}, qword ptr [{s} + 800]",
            "adc {w5}, qword ptr [{s} + 808]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 856], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 856]",
            "mov qword ptr [{s} + 816], {w0}",
            "mov qword ptr [{s} + 824], {w1}",
            "mov qword ptr [{s} + 832], {w2}",
            "mov qword ptr [{s} + 840], {w3}",
            "mov qword ptr [{s} + 848], {w4}",
            "mov qword ptr [{s} + 856], {w5}",
            // X3 = sub(X3, j)
            "mov {w0}, qword ptr [{s} + 240]",
            "mov {w1}, qword ptr [{s} + 248]",
            "mov {w2}, qword ptr [{s} + 256]",
            "mov {w3}, qword ptr [{s} + 264]",
            "mov {w4}, qword ptr [{s} + 272]",
            "mov {w5}, qword ptr [{s} + 280]",
            "sub {w0}, qword ptr [{s} + 720]",
            "sbb {w1}, qword ptr [{s} + 728]",
            "sbb {w2}, qword ptr [{s} + 736]",
            "sbb {w3}, qword ptr [{s} + 744]",
            "sbb {w4}, qword ptr [{s} + 752]",
            "sbb {w5}, qword ptr [{s} + 760]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 240], {w0}",
            "mov qword ptr [{s} + 248], {w1}",
            "mov qword ptr [{s} + 256], {w2}",
            "mov qword ptr [{s} + 264], {w3}",
            "mov qword ptr [{s} + 272], {w4}",
            "mov qword ptr [{s} + 280], {w5}",
            // X3 = sub(X3, t)
            "mov {w0}, qword ptr [{s} + 240]",
            "mov {w1}, qword ptr [{s} + 248]",
            "mov {w2}, qword ptr [{s} + 256]",
            "mov {w3}, qword ptr [{s} + 264]",
            "mov {w4}, qword ptr [{s} + 272]",
            "mov {w5}, qword ptr [{s} + 280]",
            "sub {w0}, qword ptr [{s} + 816]",
            "sbb {w1}, qword ptr [{s} + 824]",
            "sbb {w2}, qword ptr [{s} + 832]",
            "sbb {w3}, qword ptr [{s} + 840]",
            "sbb {w4}, qword ptr [{s} + 848]",
            "sbb {w5}, qword ptr [{s} + 856]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 240], {w0}",
            "mov qword ptr [{s} + 248], {w1}",
            "mov qword ptr [{s} + 256], {w2}",
            "mov qword ptr [{s} + 264], {w3}",
            "mov qword ptr [{s} + 272], {w4}",
            "mov qword ptr [{s} + 280], {w5}",
            // t3 = sub(v, X3)
            "mov {w0}, qword ptr [{s} + 768]",
            "mov {w1}, qword ptr [{s} + 776]",
            "mov {w2}, qword ptr [{s} + 784]",
            "mov {w3}, qword ptr [{s} + 792]",
            "mov {w4}, qword ptr [{s} + 800]",
            "mov {w5}, qword ptr [{s} + 808]",
            "sub {w0}, qword ptr [{s} + 240]",
            "sbb {w1}, qword ptr [{s} + 248]",
            "sbb {w2}, qword ptr [{s} + 256]",
            "sbb {w3}, qword ptr [{s} + 264]",
            "sbb {w4}, qword ptr [{s} + 272]",
            "sbb {w5}, qword ptr [{s} + 280]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 912], {w0}",
            "mov qword ptr [{s} + 920], {w1}",
            "mov qword ptr [{s} + 928], {w2}",
            "mov qword ptr [{s} + 936], {w3}",
            "mov qword ptr [{s} + 944], {w4}",
            "mov qword ptr [{s} + 952], {w5}",
            // Z3 = sub(Z3, z1z1)
            "mov {w0}, qword ptr [{s} + 336]",
            "mov {w1}, qword ptr [{s} + 344]",
            "mov {w2}, qword ptr [{s} + 352]",
            "mov {w3}, qword ptr [{s} + 360]",
            "mov {w4}, qword ptr [{s} + 368]",
            "mov {w5}, qword ptr [{s} + 376]",
            "sub {w0}, qword ptr [{s} + 480]",
            "sbb {w1}, qword ptr [{s} + 488]",
            "sbb {w2}, qword ptr [{s} + 496]",
            "sbb {w3}, qword ptr [{s} + 504]",
            "sbb {w4}, qword ptr [{s} + 512]",
            "sbb {w5}, qword ptr [{s} + 520]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 336], {w0}",
            "mov qword ptr [{s} + 344], {w1}",
            "mov qword ptr [{s} + 352], {w2}",
            "mov qword ptr [{s} + 360], {w3}",
            "mov qword ptr [{s} + 368], {w4}",
            "mov qword ptr [{s} + 376], {w5}",
            // Y3 = mul(R, t3)
            "xor {w0:e}, {w0:e}",
            "xor {w1:e}, {w1:e}",
            "xor {w2:e}, {w2:e}",
            "xor {w3:e}, {w3:e}",
            "xor {w4:e}, {w4:e}",
            "xor {w5:e}, {w5:e}",
            "xor {w6:e}, {w6:e}",
            "mov {k}, 0xffffffff00000001",
            "mov rdx, qword ptr [{s} + 912]",
            "xor {w7:e}, {w7:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mov eax, 0",
            "adcx {w6}, rax",
            "adox {w7}, rax",
            "adcx {w7}, rax",
            "mov rdx, {w0}",
            "shl rdx, 32",
            "add rdx, {w0}",
            "mulx rax, {w0}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w0}, {c}, {c}",
            "add rax, {c}",
            "adc {w0}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w1}, rax",
            "sbb {w2}, {w0}",
            "sbb {w3}, {c}",
            "sbb {w4}, 0",
            "sbb {w5}, 0",
            "sbb rdx, 0",
            "add {w6}, rdx",
            "adc {w7}, 0",
            "mov rdx, qword ptr [{s} + 920]",
            "xor {w0:e}, {w0:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mov eax, 0",
            "adcx {w7}, rax",
            "adox {w0}, rax",
            "adcx {w0}, rax",
            "mov rdx, {w1}",
            "shl rdx, 32",
            "add rdx, {w1}",
            "mulx rax, {w1}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w1}, {c}, {c}",
            "add rax, {c}",
            "adc {w1}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w2}, rax",
            "sbb {w3}, {w1}",
            "sbb {w4}, {c}",
            "sbb {w5}, 0",
            "sbb {w6}, 0",
            "sbb rdx, 0",
            "add {w7}, rdx",
            "adc {w0}, 0",
            "mov rdx, qword ptr [{s} + 928]",
            "xor {w1:e}, {w1:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mov eax, 0",
            "adcx {w0}, rax",
            "adox {w1}, rax",
            "adcx {w1}, rax",
            "mov rdx, {w2}",
            "shl rdx, 32",
            "add rdx, {w2}",
            "mulx rax, {w2}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w2}, {c}, {c}",
            "add rax, {c}",
            "adc {w2}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w3}, rax",
            "sbb {w4}, {w2}",
            "sbb {w5}, {c}",
            "sbb {w6}, 0",
            "sbb {w7}, 0",
            "sbb rdx, 0",
            "add {w0}, rdx",
            "adc {w1}, 0",
            "mov rdx, qword ptr [{s} + 936]",
            "xor {w2:e}, {w2:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w3}, rax",
            "adox {w4}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mov eax, 0",
            "adcx {w1}, rax",
            "adox {w2}, rax",
            "adcx {w2}, rax",
            "mov rdx, {w3}",
            "shl rdx, 32",
            "add rdx, {w3}",
            "mulx rax, {w3}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w3}, {c}, {c}",
            "add rax, {c}",
            "adc {w3}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w4}, rax",
            "sbb {w5}, {w3}",
            "sbb {w6}, {c}",
            "sbb {w7}, 0",
            "sbb {w0}, 0",
            "sbb rdx, 0",
            "add {w1}, rdx",
            "adc {w2}, 0",
            "mov rdx, qword ptr [{s} + 944]",
            "xor {w3:e}, {w3:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w4}, rax",
            "adox {w5}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mov eax, 0",
            "adcx {w2}, rax",
            "adox {w3}, rax",
            "adcx {w3}, rax",
            "mov rdx, {w4}",
            "shl rdx, 32",
            "add rdx, {w4}",
            "mulx rax, {w4}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w4}, {c}, {c}",
            "add rax, {c}",
            "adc {w4}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w5}, rax",
            "sbb {w6}, {w4}",
            "sbb {w7}, {c}",
            "sbb {w0}, 0",
            "sbb {w1}, 0",
            "sbb rdx, 0",
            "add {w2}, rdx",
            "adc {w3}, 0",
            "mov rdx, qword ptr [{s} + 952]",
            "xor {w4:e}, {w4:e}",
            "mulx {c}, rax, qword ptr [{s} + 432]",
            "adcx {w5}, rax",
            "adox {w6}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 440]",
            "adcx {w6}, rax",
            "adox {w7}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 448]",
            "adcx {w7}, rax",
            "adox {w0}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 456]",
            "adcx {w0}, rax",
            "adox {w1}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 464]",
            "adcx {w1}, rax",
            "adox {w2}, {c}",
            "mulx {c}, rax, qword ptr [{s} + 472]",
            "adcx {w2}, rax",
            "adox {w3}, {c}",
            "mov eax, 0",
            "adcx {w3}, rax",
            "adox {w4}, rax",
            "adcx {w4}, rax",
            "mov rdx, {w5}",
            "shl rdx, 32",
            "add rdx, {w5}",
            "mulx rax, {w5}, {k}",
            "mov {c:e}, 0xffffffff",
            "mulx {w5}, {c}, {c}",
            "add rax, {c}",
            "adc {w5}, rdx",
            "mov {c:e}, 0",
            "adc {c}, 0",
            "sub {w6}, rax",
            "sbb {w7}, {w5}",
            "sbb {w0}, {c}",
            "sbb {w1}, 0",
            "sbb {w2}, 0",
            "sbb rdx, 0",
            "add {w3}, rdx",
            "adc {w4}, 0",
            "mov qword ptr [{s} + 328], {w3}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w5:e}, 1",
            "mov {c}, {w1}",
            "mov {k}, {w2}",
            "add rax, {w6}",
            "adc rdx, {w7}",
            "adc {w5}, {w0}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w3}, 0",
            "adc {w4}, 0",
            "cmovnz {w6}, rax",
            "cmovnz {w7}, rdx",
            "cmovnz {w0}, {w5}",
            "cmovnz {w1}, {c}",
            "cmovnz {w2}, {k}",
            "cmovz {w3}, qword ptr [{s} + 328]",
            "mov qword ptr [{s} + 288], {w6}",
            "mov qword ptr [{s} + 296], {w7}",
            "mov qword ptr [{s} + 304], {w0}",
            "mov qword ptr [{s} + 312], {w1}",
            "mov qword ptr [{s} + 320], {w2}",
            "mov qword ptr [{s} + 328], {w3}",
            // Z3 = sub(Z3, hh)
            "mov {w0}, qword ptr [{s} + 336]",
            "mov {w1}, qword ptr [{s} + 344]",
            "mov {w2}, qword ptr [{s} + 352]",
            "mov {w3}, qword ptr [{s} + 360]",
            "mov {w4}, qword ptr [{s} + 368]",
            "mov {w5}, qword ptr [{s} + 376]",
            "sub {w0}, qword ptr [{s} + 624]",
            "sbb {w1}, qword ptr [{s} + 632]",
            "sbb {w2}, qword ptr [{s} + 640]",
            "sbb {w3}, qword ptr [{s} + 648]",
            "sbb {w4}, qword ptr [{s} + 656]",
            "sbb {w5}, qword ptr [{s} + 664]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 336], {w0}",
            "mov qword ptr [{s} + 344], {w1}",
            "mov qword ptr [{s} + 352], {w2}",
            "mov qword ptr [{s} + 360], {w3}",
            "mov qword ptr [{s} + 368], {w4}",
            "mov qword ptr [{s} + 376], {w5}",
            // t2 = add(t2, t2)
            "xor {w6:e}, {w6:e}",
            "mov {w0}, qword ptr [{s} + 864]",
            "mov {w1}, qword ptr [{s} + 872]",
            "mov {w2}, qword ptr [{s} + 880]",
            "mov {w3}, qword ptr [{s} + 888]",
            "mov {w4}, qword ptr [{s} + 896]",
            "mov {w5}, qword ptr [{s} + 904]",
            "add {w0}, qword ptr [{s} + 864]",
            "adc {w1}, qword ptr [{s} + 872]",
            "adc {w2}, qword ptr [{s} + 880]",
            "adc {w3}, qword ptr [{s} + 888]",
            "adc {w4}, qword ptr [{s} + 896]",
            "adc {w5}, qword ptr [{s} + 904]",
            "adc {w6}, 0",
            "mov qword ptr [{s} + 904], {w5}",
            "mov rax, 0xffffffff00000001",
            "mov edx, 0xffffffff",
            "mov {w7:e}, 1",
            "mov {c}, {w3}",
            "mov {k}, {w4}",
            "add rax, {w0}",
            "adc rdx, {w1}",
            "adc {w7}, {w2}",
            "adc {c}, 0",
            "adc {k}, 0",
            "adc {w5}, 0",
            "adc {w6}, 0",
            "cmovnz {w0}, rax",
            "cmovnz {w1}, rdx",
            "cmovnz {w2}, {w7}",
            "cmovnz {w3}, {c}",
            "cmovnz {w4}, {k}",
            "cmovz {w5}, qword ptr [{s} + 904]",
            "mov qword ptr [{s} + 864], {w0}",
            "mov qword ptr [{s} + 872], {w1}",
            "mov qword ptr [{s} + 880], {w2}",
            "mov qword ptr [{s} + 888], {w3}",
            "mov qword ptr [{s} + 896], {w4}",
            "mov qword ptr [{s} + 904], {w5}",
            // Y3 = sub(Y3, t2)
            "mov {w0}, qword ptr [{s} + 288]",
            "mov {w1}, qword ptr [{s} + 296]",
            "mov {w2}, qword ptr [{s} + 304]",
            "mov {w3}, qword ptr [{s} + 312]",
            "mov {w4}, qword ptr [{s} + 320]",
            "mov {w5}, qword ptr [{s} + 328]",
            "sub {w0}, qword ptr [{s} + 864]",
            "sbb {w1}, qword ptr [{s} + 872]",
            "sbb {w2}, qword ptr [{s} + 880]",
            "sbb {w3}, qword ptr [{s} + 888]",
            "sbb {w4}, qword ptr [{s} + 896]",
            "sbb {w5}, qword ptr [{s} + 904]",
            "sbb {c}, {c}",
            "mov eax, 0xffffffff",
            "and rax, {c}",
            "mov rdx, {c}",
            "shl rdx, 32",
            "mov {w6}, {c}",
            "and {w6}, -2",
            "add {w0}, rax",
            "adc {w1}, rdx",
            "adc {w2}, {w6}",
            "adc {w3}, {c}",
            "adc {w4}, {c}",
            "adc {w5}, {c}",
            "mov qword ptr [{s} + 288], {w0}",
            "mov qword ptr [{s} + 296], {w1}",
            "mov qword ptr [{s} + 304], {w2}",
            "mov qword ptr [{s} + 312], {w3}",
            "mov qword ptr [{s} + 320], {w4}",
            "mov qword ptr [{s} + 328], {w5}",
            s = in(reg) buf.as_mut_ptr(),
            w0 = out(reg) _,
            w1 = out(reg) _,
            w2 = out(reg) _,
            w3 = out(reg) _,
            w4 = out(reg) _,
            w5 = out(reg) _,
            w6 = out(reg) _,
            w7 = out(reg) _,
            c = out(reg) _,
            k = out(reg) _,
            out("rax") _,
            out("rdx") _,
            options(nostack),
        );
    }
}
#[allow(dead_code)]
pub const ADD_AFFINE_X1: usize = 0;
#[allow(dead_code)]
pub const ADD_AFFINE_Y1: usize = 1;
#[allow(dead_code)]
pub const ADD_AFFINE_Z1: usize = 2;
#[allow(dead_code)]
pub const ADD_AFFINE_X2: usize = 3;
#[allow(dead_code)]
pub const ADD_AFFINE_Y2: usize = 4;
#[allow(dead_code)]
pub const ADD_AFFINE_X3: usize = 5;
#[allow(dead_code)]
pub const ADD_AFFINE_Y3: usize = 6;
#[allow(dead_code)]
pub const ADD_AFFINE_Z3: usize = 7;
#[allow(dead_code)]
pub const ADD_AFFINE_H: usize = 8;
#[allow(dead_code)]
pub const ADD_AFFINE_R: usize = 9;