rind 0.1.3

// generated source. do not edit.
#![allow(non_upper_case_globals, unused_macros, unused_imports)]
use crate::low::macros::*;

// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
// SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT-0

// ----------------------------------------------------------------------------
// Scalar multiplication for precomputed point on NIST curve P-256
// Input scalar[4], blocksize, table[]; output res[8]
//
// extern void p256_scalarmulbase_alt
//   (uint64_t res[static 8],
//    const uint64_t scalar[static 4],
//    uint64_t blocksize,
//    const uint64_t *table);
//
// Given scalar = n and point = P, assumed to be on the NIST elliptic
// curve P-256, the input argument "table" is expected to be a table of
// multiples of the point P in Montgomery-affine form, with each block
// corresponding to "blocksize" bits of the scalar as follows, where
// B = 2^{blocksize-1} (e.g. B = 8 for blocksize = 4):
//
// For each i,j with blocksize * i <= 256 and 1 <= j <= B
// the multiple 2^{blocksize * i} * j * P is stored at
// tab[8 * (B * i + (j - 1))], considered as uint64_t pointers
// or tab + 64 * (B * i + (j - 1)) as byte pointers.
//
// Standard ARM ABI: X0 = res, X1 = scalar, X2 = blocksize, X3 = table
// ----------------------------------------------------------------------------

// Size of individual field elements

macro_rules! NUMSIZE {
    () => {
        "32"
    };
}

// Safe copies of inputs and additional variables, with some aliasing

macro_rules! res {
    () => {
        "x19"
    };
}
macro_rules! blocksize {
    () => {
        "x20"
    };
}
macro_rules! table {
    () => {
        "x21"
    };
}
macro_rules! i {
    () => {
        "x22"
    };
}
macro_rules! bf {
    () => {
        "x23"
    };
}
macro_rules! cf {
    () => {
        "x24"
    };
}
macro_rules! j {
    () => {
        "x25"
    };
}

// Intermediate variables on the stack. The last z2, z3 values can
// safely be overlaid on "nacc", which is no longer needed at the end.
// Uppercase syntactic variants make x86_att version simpler to generate

macro_rules! rscalar { () => { Q!("sp, # (0 * " NUMSIZE!() ")") } }
macro_rules! acc { () => { Q!("sp, # (1 * " NUMSIZE!() ")") } }
macro_rules! nacc { () => { Q!("sp, # (4 * " NUMSIZE!() ")") } }
macro_rules! tabent { () => { Q!("sp, # (7 * " NUMSIZE!() ")") } }

macro_rules! z2 { () => { Q!("sp, # (4 * " NUMSIZE!() ")") } }
macro_rules! z3 { () => { Q!("sp, # (5 * " NUMSIZE!() ")") } }

macro_rules! NSPACE { () => { Q!("9 * " NUMSIZE!()) } }

// Loading large constants

macro_rules! movbig {
    ($nn:expr, $n3:expr, $n2:expr, $n1:expr, $n0:expr) => { Q!(
        "movz " $nn ", " $n0 ";\n"
        "movk " $nn ", " $n1 ", lsl #16;\n"
        "movk " $nn ", " $n2 ", lsl #32;\n"
        "movk " $nn ", " $n3 ", lsl #48"
    )}
}

/// Scalar multiplication for precomputed point on NIST curve P-256
///
/// Input scalar[4], blocksize, table[]; output res[8]
///
/// Given scalar = n and point = P, assumed to be on the NIST elliptic
/// curve P-256, the input argument "table" is expected to be a table of
/// multiples of the point P in Montgomery-affine form, with each block
/// corresponding to "blocksize" bits of the scalar as follows, where
/// B = 2^{blocksize-1} (e.g. B = 8 for blocksize = 4):
///
/// For each i,j with blocksize * i <= 256 and 1 <= j <= B
/// the multiple 2^{blocksize * i} * j * P is stored at
/// tab[8 * (B * i + (j - 1))], considered as uint64_t pointers
/// or tab + 64 * (B * i + (j - 1)) as byte pointers.
pub(crate) fn p256_scalarmulbase(res: &mut [u64; 8], scalar: &[u64; 4], table: &[[u64; 256]; 43]) {
    // SAFETY: inline assembly. see [crate::low::inline_assembly_safety] for safety info.
    unsafe {
        core::arch::asm!(

        Q!("    stp             " "x19, x20, [sp, #-16] !"),
        Q!("    stp             " "x21, x22, [sp, #-16] !"),
        Q!("    stp             " "x23, x24, [sp, #-16] !"),
        Q!("    stp             " "x25, x30, [sp, #-16] !"),
        Q!("    sub             " "sp, sp, # (" NSPACE!() "+ 0)"),

        // Preserve the input arguments except the scalar, since that gets absorbed
        // immediately. The "table" value subsequently gets shifted up each iteration
        // of the loop, while "res" and "blocksize" are static throughout.

        Q!("    mov             " res!() ", x0"),
        Q!("    mov             " blocksize!() ", x2"),
        Q!("    mov             " table!() ", x3"),

        // Load the digits of group order n_256 = [x15;x14;x13;x12]

        movbig!("x12", "#0xf3b9", "#0xcac2", "#0xfc63", "#0x2551"),
        movbig!("x13", "#0xbce6", "#0xfaad", "#0xa717", "#0x9e84"),
        Q!("    mov             " "x14, #0xffffffffffffffff"),
        Q!("    mov             " "x15, #0xffffffff00000000"),

        // First, reduce the input scalar mod n_256, i.e. conditionally subtract n_256
        // Store it to "rscalar" (reduced scalar)

        Q!("    ldp             " "x2, x3, [x1]"),
        Q!("    ldp             " "x4, x5, [x1, #16]"),

        Q!("    subs            " "x6, x2, x12"),
        Q!("    sbcs            " "x7, x3, x13"),
        Q!("    sbcs            " "x8, x4, x14"),
        Q!("    sbcs            " "x9, x5, x15"),

        Q!("    csel            " "x2, x2, x6, cc"),
        Q!("    csel            " "x3, x3, x7, cc"),
        Q!("    csel            " "x4, x4, x8, cc"),
        Q!("    csel            " "x5, x5, x9, cc"),

        Q!("    stp             " "x2, x3, [" rscalar!() "]"),
        Q!("    stp             " "x4, x5, [" rscalar!() "+ 16]"),

        // Initialize the accumulator to all zeros and the "carry flag" cf to 0

        Q!("    stp             " "xzr, xzr, [" acc!() "]"),
        Q!("    stp             " "xzr, xzr, [" acc!() "+ 16]"),
        Q!("    stp             " "xzr, xzr, [" acc!() "+ 32]"),
        Q!("    stp             " "xzr, xzr, [" acc!() "+ 48]"),
        Q!("    stp             " "xzr, xzr, [" acc!() "+ 64]"),
        Q!("    stp             " "xzr, xzr, [" acc!() "+ 80]"),
        Q!("    mov             " cf!() ", xzr"),

        // Main loop over {i >= 0 | blocksize * i <= 256}. Note the non-strict
        // inequality, to allow top carry for any choices of blocksize.

        Q!("    mov             " i!() ", xzr"),

        Q!(Label!("Lp256_scalarmulbase_alt_loop", 2) ":"),

        // The next raw bitfield is bf = bitfield(blocksize * i,blocksize) + cf,
        // adding in the deferred carry cf. We then shift the whole scalar right
        // by blocksize so we can keep picking bitfield(0,blocksize).

        Q!("    ldp             " "x0, x1, [" rscalar!() "]"),
        Q!("    ldp             " "x2, x3, [" rscalar!() "+ 16]"),

        Q!("    mov             " "x4, #1"),
        Q!("    lsl             " "x4, x4, " blocksize!()),
        Q!("    sub             " "x4, x4, #1"),
        Q!("    and             " "x4, x4, x0"),
        Q!("    add             " bf!() ", x4, " cf!()),

        Q!("    neg             " "x8, " blocksize!()),

        Q!("    lsl             " "x5, x1, x8"),

        Q!("    lsr             " "x0, x0, " blocksize!()),
        Q!("    orr             " "x0, x0, x5"),

        Q!("    lsl             " "x6, x2, x8"),
        Q!("    lsr             " "x1, x1, " blocksize!()),
        Q!("    orr             " "x1, x1, x6"),

        Q!("    lsl             " "x7, x3, x8"),
        Q!("    lsr             " "x2, x2, " blocksize!()),
        Q!("    orr             " "x2, x2, x7"),

        Q!("    lsr             " "x3, x3, " blocksize!()),

        Q!("    stp             " "x0, x1, [" rscalar!() "]"),
        Q!("    stp             " "x2, x3, [" rscalar!() "+ 16]"),

        // Now if bf <= B we just select entry j, unnegated and set cf = 0.
        // If bf > B we set j = 2 * B - bf and negate the j'th entry, setting cf = 1.
        // In either case we ultimately add bf, in the latter case with deferred
        // carry as 2 * B - (2 * B - bf) = bf.

        Q!("    mov             " "x0, #1"),
        Q!("    lsl             " "x1, x0, " blocksize!()),
        Q!("    lsr             " "x0, x1, #1"),

        Q!("    sub             " "x2, x1, " bf!()),

        Q!("    cmp             " "x0, " bf!()),
        Q!("    cset            " cf!() ", cc"),
        Q!("    csel            " j!() ", x2, " bf!() ", cc"),

        // Load table entry j - 1 for nonzero j in constant-time style.

        Q!("    mov             " "x16, #1"),
        Q!("    lsl             " "x16, x16, " blocksize!()),
        Q!("    lsr             " "x16, x16, #1"),
        Q!("    mov             " "x17, " j!()),

        Q!(Label!("Lp256_scalarmulbase_alt_tabloop", 3) ":"),
        Q!("    ldp             " "x8, x9, [" table!() "]"),
        Q!("    ldp             " "x10, x11, [" table!() ", #16]"),
        Q!("    ldp             " "x12, x13, [" table!() ", #32]"),
        Q!("    ldp             " "x14, x15, [" table!() ", #48]"),

        Q!("    subs            " "x17, x17, #1"),
        Q!("    csel            " "x0, x8, x0, eq"),
        Q!("    csel            " "x1, x9, x1, eq"),
        Q!("    csel            " "x2, x10, x2, eq"),
        Q!("    csel            " "x3, x11, x3, eq"),
        Q!("    csel            " "x4, x12, x4, eq"),
        Q!("    csel            " "x5, x13, x5, eq"),
        Q!("    csel            " "x6, x14, x6, eq"),
        Q!("    csel            " "x7, x15, x7, eq"),

        Q!("    add             " table!() ", " table!() ", #64"),

        Q!("    sub             " "x16, x16, #1"),
        Q!("    cbnz            " "x16, " Label!("Lp256_scalarmulbase_alt_tabloop", 3, Before)),

        // Before storing back, optionally negate the y coordinate of the table entry

        Q!("    stp             " "x0, x1, [" tabent!() "]"),
        Q!("    stp             " "x2, x3, [" tabent!() "+ 16]"),

        Q!("    mov             " "x0, 0xffffffffffffffff"),
        Q!("    subs            " "x0, x0, x4"),
        Q!("    mov             " "x1, 0x00000000ffffffff"),
        Q!("    sbcs            " "x1, x1, x5"),
        Q!("    mov             " "x3, 0xffffffff00000001"),
        Q!("    sbcs            " "x2, xzr, x6"),
        Q!("    sbc             " "x3, x3, x7"),

        Q!("    cmp             " cf!() ", xzr"),
        Q!("    csel            " "x4, x0, x4, ne"),
        Q!("    csel            " "x5, x1, x5, ne"),
        Q!("    csel            " "x6, x2, x6, ne"),
        Q!("    csel            " "x7, x3, x7, ne"),

        Q!("    stp             " "x4, x5, [" tabent!() "+ 32]"),
        Q!("    stp             " "x6, x7, [" tabent!() "+ 48]"),

        // Add the adjusted table point to the accumulator

        Q!("    add             " "x0, " nacc!()),
        Q!("    add             " "x1, " acc!()),
        Q!("    add             " "x2, " tabent!()),
        Q!("    bl              " Label!("Lp256_scalarmulbase_alt_local_p256_montjmixadd", 4, After)),

        // However, only commit that update to the accumulator if j is nonzero,
        // because the mixed addition function does not handle this case directly,
        // and in any case we didn't choose the table entry appropriately.

        Q!("    cmp             " j!() ", xzr"),
        Q!("    ldp             " "x0, x1, [" acc!() "]"),
        Q!("    ldp             " "x12, x13, [" nacc!() "]"),
        Q!("    csel            " "x0, x12, x0, ne"),
        Q!("    csel            " "x1, x13, x1, ne"),

        Q!("    ldp             " "x2, x3, [" acc!() "+ 16]"),
        Q!("    ldp             " "x12, x13, [" nacc!() "+ 16]"),
        Q!("    csel            " "x2, x12, x2, ne"),
        Q!("    csel            " "x3, x13, x3, ne"),

        Q!("    ldp             " "x4, x5, [" acc!() "+ 32]"),
        Q!("    ldp             " "x12, x13, [" nacc!() "+ 32]"),
        Q!("    csel            " "x4, x12, x4, ne"),
        Q!("    csel            " "x5, x13, x5, ne"),

        Q!("    ldp             " "x6, x7, [" acc!() "+ 48]"),
        Q!("    ldp             " "x12, x13, [" nacc!() "+ 48]"),
        Q!("    csel            " "x6, x12, x6, ne"),
        Q!("    csel            " "x7, x13, x7, ne"),

        Q!("    ldp             " "x8, x9, [" acc!() "+ 64]"),
        Q!("    ldp             " "x12, x13, [" nacc!() "+ 64]"),
        Q!("    csel            " "x8, x12, x8, ne"),
        Q!("    csel            " "x9, x13, x9, ne"),

        Q!("    ldp             " "x10, x11, [" acc!() "+ 80]"),
        Q!("    ldp             " "x12, x13, [" nacc!() "+ 80]"),
        Q!("    csel            " "x10, x12, x10, ne"),
        Q!("    csel            " "x11, x13, x11, ne"),

        Q!("    stp             " "x0, x1, [" acc!() "]"),
        Q!("    stp             " "x2, x3, [" acc!() "+ 16]"),
        Q!("    stp             " "x4, x5, [" acc!() "+ 32]"),
        Q!("    stp             " "x6, x7, [" acc!() "+ 48]"),
        Q!("    stp             " "x8, x9, [" acc!() "+ 64]"),
        Q!("    stp             " "x10, x11, [" acc!() "+ 80]"),

        // Loop while blocksize * i <= 256

        Q!("    add             " i!() ", " i!() ", #1"),
        Q!("    mul             " "x0, " blocksize!() ", " i!()),
        Q!("    cmp             " "x0, #257"),
        Q!("    bcc             " Label!("Lp256_scalarmulbase_alt_loop", 2, Before)),

        // That's the end of the main loop, and we just need to translate
        // back from the Jacobian representation to affine. First of all,
        // let z2 = 1/z^2 and z3 = 1/z^3, both without Montgomery form

        Q!("    add             " "x0, " z2!()),
        Q!("    add             " "x1, " acc!() "+ 64"),
        Q!("    bl              " Label!("Lp256_scalarmulbase_alt_local_montsqr_p256", 5, After)),

        Q!("    add             " "x0, " z3!()),
        Q!("    add             " "x1, " acc!() "+ 64"),
        Q!("    add             " "x2, " z2!()),
        Q!("    bl              " Label!("Lp256_scalarmulbase_alt_local_montmul_p256", 6, After)),

        Q!("    add             " "x0, " z2!()),
        Q!("    add             " "x1, " z3!()),
        Q!("    bl              " Label!("Lp256_scalarmulbase_alt_local_demont_p256", 7, After)),

        Q!("    add             " "x0, " z3!()),
        Q!("    add             " "x1, " z2!()),
        Q!("    bl              " Label!("Lp256_scalarmulbase_alt_local_inv_p256", 8, After)),

        Q!("    add             " "x0, " z2!()),
        Q!("    add             " "x1, " acc!() "+ 64"),
        Q!("    add             " "x2, " z3!()),
        Q!("    bl              " Label!("Lp256_scalarmulbase_alt_local_montmul_p256", 6, After)),

        // Convert back from Jacobian (X,Y,Z) |-> (X/Z^2, Y/Z^3)

        Q!("    mov             " "x0, " res!()),
        Q!("    add             " "x1, " acc!()),
        Q!("    add             " "x2, " z2!()),
        Q!("    bl              " Label!("Lp256_scalarmulbase_alt_local_montmul_p256", 6, After)),

        Q!("    add             " "x0, " res!() ", #32"),
        Q!("    add             " "x1, " acc!() "+ 32"),
        Q!("    add             " "x2, " z3!()),
        Q!("    bl              " Label!("Lp256_scalarmulbase_alt_local_montmul_p256", 6, After)),

        // Restore stack and registers and return

        Q!("    add             " "sp, sp, # (" NSPACE!() "+ 0)"),
        Q!("    ldp             " "x25, x30, [sp], #16"),
        Q!("    ldp             " "x23, x24, [sp], #16"),
        Q!("    ldp             " "x21, x22, [sp], #16"),
        Q!("    ldp             " "x19, x20, [sp], #16"),
        // proc hoisting in -> ret after Lp256_scalarmulbase_alt_local_p256_montjmixadd
        Q!("    b               " Label!("hoist_finish", 9, After)),

        // Local copies of subroutines, complete clones at the moment

        Q!(Label!("Lp256_scalarmulbase_alt_local_demont_p256", 7) ":"),

        Q!("    ldp             " "x2, x3, [x1]"),
        Q!("    ldp             " "x4, x5, [x1, #16]"),
        Q!("    lsl             " "x7, x2, #32"),
        Q!("    subs            " "x8, x2, x7"),
        Q!("    lsr             " "x6, x2, #32"),
        Q!("    sbc             " "x2, x2, x6"),
        Q!("    adds            " "x3, x3, x7"),
        Q!("    adcs            " "x4, x4, x6"),
        Q!("    adcs            " "x5, x5, x8"),
        Q!("    adc             " "x2, x2, xzr"),
        Q!("    lsl             " "x7, x3, #32"),
        Q!("    subs            " "x8, x3, x7"),
        Q!("    lsr             " "x6, x3, #32"),
        Q!("    sbc             " "x3, x3, x6"),
        Q!("    adds            " "x4, x4, x7"),
        Q!("    adcs            " "x5, x5, x6"),
        Q!("    adcs            " "x2, x2, x8"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    lsl             " "x7, x4, #32"),
        Q!("    subs            " "x8, x4, x7"),
        Q!("    lsr             " "x6, x4, #32"),
        Q!("    sbc             " "x4, x4, x6"),
        Q!("    adds            " "x5, x5, x7"),
        Q!("    adcs            " "x2, x2, x6"),
        Q!("    adcs            " "x3, x3, x8"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    lsl             " "x7, x5, #32"),
        Q!("    subs            " "x8, x5, x7"),
        Q!("    lsr             " "x6, x5, #32"),
        Q!("    sbc             " "x5, x5, x6"),
        Q!("    adds            " "x2, x2, x7"),
        Q!("    adcs            " "x3, x3, x6"),
        Q!("    adcs            " "x4, x4, x8"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    stp             " "x2, x3, [x0]"),
        Q!("    stp             " "x4, x5, [x0, #16]"),
        Q!("    ret             " ),

        Q!(Label!("Lp256_scalarmulbase_alt_local_inv_p256", 8) ":"),

        Q!("    stp             " "x19, x20, [sp, #-16] !"),
        Q!("    stp             " "x21, x22, [sp, #-16] !"),
        Q!("    stp             " "x23, x24, [sp, #-16] !"),
        Q!("    sub             " "sp, sp, # (160 + 0)"),
        Q!("    mov             " "x20, x0"),
        Q!("    mov             " "x10, #0xffffffffffffffff"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    mov             " "x13, #0xffffffff00000001"),
        Q!("    stp             " "x10, x11, [sp]"),
        Q!("    stp             " "xzr, x13, [sp, #16]"),
        Q!("    str             " "xzr, [sp, #32]"),
        Q!("    ldp             " "x2, x3, [x1]"),
        Q!("    subs            " "x10, x2, x10"),
        Q!("    sbcs            " "x11, x3, x11"),
        Q!("    ldp             " "x4, x5, [x1, #16]"),
        Q!("    sbcs            " "x12, x4, xzr"),
        Q!("    sbcs            " "x13, x5, x13"),
        Q!("    csel            " "x2, x2, x10, cc"),
        Q!("    csel            " "x3, x3, x11, cc"),
        Q!("    csel            " "x4, x4, x12, cc"),
        Q!("    csel            " "x5, x5, x13, cc"),
        Q!("    stp             " "x2, x3, [sp, #48]"),
        Q!("    stp             " "x4, x5, [sp, #64]"),
        Q!("    str             " "xzr, [sp, #80]"),
        Q!("    stp             " "xzr, xzr, [sp, #96]"),
        Q!("    stp             " "xzr, xzr, [sp, #112]"),
        Q!("    mov             " "x10, #0x4000000000000"),
        Q!("    stp             " "x10, xzr, [sp, #128]"),
        Q!("    stp             " "xzr, xzr, [sp, #144]"),
        Q!("    mov             " "x21, #0xa"),
        Q!("    mov             " "x22, #0x1"),
        Q!("    b               " Label!("Lp256_scalarmulbase_alt_inv_midloop", 12, After)),
        Q!(Label!("Lp256_scalarmulbase_alt_inv_loop", 13) ":"),
        Q!("    cmp             " "x10, xzr"),
        Q!("    csetm           " "x14, mi"),
        Q!("    cneg            " "x10, x10, mi"),
        Q!("    cmp             " "x11, xzr"),
        Q!("    csetm           " "x15, mi"),
        Q!("    cneg            " "x11, x11, mi"),
        Q!("    cmp             " "x12, xzr"),
        Q!("    csetm           " "x16, mi"),
        Q!("    cneg            " "x12, x12, mi"),
        Q!("    cmp             " "x13, xzr"),
        Q!("    csetm           " "x17, mi"),
        Q!("    cneg            " "x13, x13, mi"),
        Q!("    and             " "x0, x10, x14"),
        Q!("    and             " "x1, x11, x15"),
        Q!("    add             " "x9, x0, x1"),
        Q!("    and             " "x0, x12, x16"),
        Q!("    and             " "x1, x13, x17"),
        Q!("    add             " "x19, x0, x1"),
        Q!("    ldr             " "x7, [sp]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x4, x9, x0"),
        Q!("    adc             " "x2, xzr, x1"),
        Q!("    ldr             " "x8, [sp, #48]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x4, x4, x0"),
        Q!("    adc             " "x2, x2, x1"),
        Q!("    eor             " "x1, x7, x16"),
        Q!("    mul             " "x0, x1, x12"),
        Q!("    umulh           " "x1, x1, x12"),
        Q!("    adds            " "x5, x19, x0"),
        Q!("    adc             " "x3, xzr, x1"),
        Q!("    eor             " "x1, x8, x17"),
        Q!("    mul             " "x0, x1, x13"),
        Q!("    umulh           " "x1, x1, x13"),
        Q!("    adds            " "x5, x5, x0"),
        Q!("    adc             " "x3, x3, x1"),
        Q!("    ldr             " "x7, [sp, #8]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x2, x2, x0"),
        Q!("    adc             " "x6, xzr, x1"),
        Q!("    ldr             " "x8, [sp, #56]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x2, x2, x0"),
        Q!("    adc             " "x6, x6, x1"),
        Q!("    extr            " "x4, x2, x4, #59"),
        Q!("    str             " "x4, [sp]"),
        Q!("    eor             " "x1, x7, x16"),
        Q!("    mul             " "x0, x1, x12"),
        Q!("    umulh           " "x1, x1, x12"),
        Q!("    adds            " "x3, x3, x0"),
        Q!("    adc             " "x4, xzr, x1"),
        Q!("    eor             " "x1, x8, x17"),
        Q!("    mul             " "x0, x1, x13"),
        Q!("    umulh           " "x1, x1, x13"),
        Q!("    adds            " "x3, x3, x0"),
        Q!("    adc             " "x4, x4, x1"),
        Q!("    extr            " "x5, x3, x5, #59"),
        Q!("    str             " "x5, [sp, #48]"),
        Q!("    ldr             " "x7, [sp, #16]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x6, x6, x0"),
        Q!("    adc             " "x5, xzr, x1"),
        Q!("    ldr             " "x8, [sp, #64]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x6, x6, x0"),
        Q!("    adc             " "x5, x5, x1"),
        Q!("    extr            " "x2, x6, x2, #59"),
        Q!("    str             " "x2, [sp, #8]"),
        Q!("    eor             " "x1, x7, x16"),
        Q!("    mul             " "x0, x1, x12"),
        Q!("    umulh           " "x1, x1, x12"),
        Q!("    adds            " "x4, x4, x0"),
        Q!("    adc             " "x2, xzr, x1"),
        Q!("    eor             " "x1, x8, x17"),
        Q!("    mul             " "x0, x1, x13"),
        Q!("    umulh           " "x1, x1, x13"),
        Q!("    adds            " "x4, x4, x0"),
        Q!("    adc             " "x2, x2, x1"),
        Q!("    extr            " "x3, x4, x3, #59"),
        Q!("    str             " "x3, [sp, #56]"),
        Q!("    ldr             " "x7, [sp, #24]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    ldr             " "x23, [sp, #32]"),
        Q!("    eor             " "x3, x23, x14"),
        Q!("    and             " "x3, x3, x10"),
        Q!("    neg             " "x3, x3"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x5, x5, x0"),
        Q!("    adc             " "x3, x3, x1"),
        Q!("    ldr             " "x8, [sp, #72]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    ldr             " "x24, [sp, #80]"),
        Q!("    eor             " "x0, x24, x15"),
        Q!("    and             " "x0, x0, x11"),
        Q!("    sub             " "x3, x3, x0"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x5, x5, x0"),
        Q!("    adc             " "x3, x3, x1"),
        Q!("    extr            " "x6, x5, x6, #59"),
        Q!("    str             " "x6, [sp, #16]"),
        Q!("    extr            " "x5, x3, x5, #59"),
        Q!("    str             " "x5, [sp, #24]"),
        Q!("    asr             " "x3, x3, #59"),
        Q!("    str             " "x3, [sp, #32]"),
        Q!("    eor             " "x1, x7, x16"),
        Q!("    eor             " "x5, x23, x16"),
        Q!("    and             " "x5, x5, x12"),
        Q!("    neg             " "x5, x5"),
        Q!("    mul             " "x0, x1, x12"),
        Q!("    umulh           " "x1, x1, x12"),
        Q!("    adds            " "x2, x2, x0"),
        Q!("    adc             " "x5, x5, x1"),
        Q!("    eor             " "x1, x8, x17"),
        Q!("    eor             " "x0, x24, x17"),
        Q!("    and             " "x0, x0, x13"),
        Q!("    sub             " "x5, x5, x0"),
        Q!("    mul             " "x0, x1, x13"),
        Q!("    umulh           " "x1, x1, x13"),
        Q!("    adds            " "x2, x2, x0"),
        Q!("    adc             " "x5, x5, x1"),
        Q!("    extr            " "x4, x2, x4, #59"),
        Q!("    str             " "x4, [sp, #64]"),
        Q!("    extr            " "x2, x5, x2, #59"),
        Q!("    str             " "x2, [sp, #72]"),
        Q!("    asr             " "x5, x5, #59"),
        Q!("    str             " "x5, [sp, #80]"),
        Q!("    ldr             " "x7, [sp, #96]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x4, x9, x0"),
        Q!("    adc             " "x2, xzr, x1"),
        Q!("    ldr             " "x8, [sp, #128]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x4, x4, x0"),
        Q!("    str             " "x4, [sp, #96]"),
        Q!("    adc             " "x2, x2, x1"),
        Q!("    eor             " "x1, x7, x16"),
        Q!("    mul             " "x0, x1, x12"),
        Q!("    umulh           " "x1, x1, x12"),
        Q!("    adds            " "x5, x19, x0"),
        Q!("    adc             " "x3, xzr, x1"),
        Q!("    eor             " "x1, x8, x17"),
        Q!("    mul             " "x0, x1, x13"),
        Q!("    umulh           " "x1, x1, x13"),
        Q!("    adds            " "x5, x5, x0"),
        Q!("    str             " "x5, [sp, #128]"),
        Q!("    adc             " "x3, x3, x1"),
        Q!("    ldr             " "x7, [sp, #104]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x2, x2, x0"),
        Q!("    adc             " "x6, xzr, x1"),
        Q!("    ldr             " "x8, [sp, #136]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x2, x2, x0"),
        Q!("    str             " "x2, [sp, #104]"),
        Q!("    adc             " "x6, x6, x1"),
        Q!("    eor             " "x1, x7, x16"),
        Q!("    mul             " "x0, x1, x12"),
        Q!("    umulh           " "x1, x1, x12"),
        Q!("    adds            " "x3, x3, x0"),
        Q!("    adc             " "x4, xzr, x1"),
        Q!("    eor             " "x1, x8, x17"),
        Q!("    mul             " "x0, x1, x13"),
        Q!("    umulh           " "x1, x1, x13"),
        Q!("    adds            " "x3, x3, x0"),
        Q!("    str             " "x3, [sp, #136]"),
        Q!("    adc             " "x4, x4, x1"),
        Q!("    ldr             " "x7, [sp, #112]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x6, x6, x0"),
        Q!("    adc             " "x5, xzr, x1"),
        Q!("    ldr             " "x8, [sp, #144]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x6, x6, x0"),
        Q!("    str             " "x6, [sp, #112]"),
        Q!("    adc             " "x5, x5, x1"),
        Q!("    eor             " "x1, x7, x16"),
        Q!("    mul             " "x0, x1, x12"),
        Q!("    umulh           " "x1, x1, x12"),
        Q!("    adds            " "x4, x4, x0"),
        Q!("    adc             " "x2, xzr, x1"),
        Q!("    eor             " "x1, x8, x17"),
        Q!("    mul             " "x0, x1, x13"),
        Q!("    umulh           " "x1, x1, x13"),
        Q!("    adds            " "x4, x4, x0"),
        Q!("    str             " "x4, [sp, #144]"),
        Q!("    adc             " "x2, x2, x1"),
        Q!("    ldr             " "x7, [sp, #120]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    and             " "x3, x14, x10"),
        Q!("    neg             " "x3, x3"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x5, x5, x0"),
        Q!("    adc             " "x3, x3, x1"),
        Q!("    ldr             " "x8, [sp, #152]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    and             " "x0, x15, x11"),
        Q!("    sub             " "x3, x3, x0"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x5, x5, x0"),
        Q!("    adc             " "x3, x3, x1"),
        Q!("    ldp             " "x0, x1, [sp, #96]"),
        Q!("    ldr             " "x6, [sp, #112]"),
        Q!("    mov             " "x14, #0xe000000000000000"),
        Q!("    adds            " "x0, x0, x14"),
        Q!("    sbcs            " "x1, x1, xzr"),
        Q!("    mov             " "x11, #0x1fffffff"),
        Q!("    adcs            " "x6, x6, x11"),
        Q!("    mov             " "x10, #0x2000000000000000"),
        Q!("    adcs            " "x5, x5, x10"),
        Q!("    mov             " "x14, #0x1fffffffe0000000"),
        Q!("    adc             " "x3, x3, x14"),
        Q!("    lsl             " "x11, x0, #32"),
        Q!("    subs            " "x14, x0, x11"),
        Q!("    lsr             " "x10, x0, #32"),
        Q!("    sbc             " "x0, x0, x10"),
        Q!("    adds            " "x1, x1, x11"),
        Q!("    adcs            " "x6, x6, x10"),
        Q!("    adcs            " "x5, x5, x14"),
        Q!("    adcs            " "x3, x3, x0"),
        Q!("    mov             " "x14, #0xffffffffffffffff"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    csel            " "x14, x14, xzr, cs"),
        Q!("    csel            " "x11, x11, xzr, cs"),
        Q!("    csel            " "x10, x10, xzr, cs"),
        Q!("    subs            " "x1, x1, x14"),
        Q!("    sbcs            " "x6, x6, x11"),
        Q!("    sbcs            " "x5, x5, xzr"),
        Q!("    sbc             " "x3, x3, x10"),
        Q!("    stp             " "x1, x6, [sp, #96]"),
        Q!("    stp             " "x5, x3, [sp, #112]"),
        Q!("    eor             " "x1, x7, x16"),
        Q!("    and             " "x5, x16, x12"),
        Q!("    neg             " "x5, x5"),
        Q!("    mul             " "x0, x1, x12"),
        Q!("    umulh           " "x1, x1, x12"),
        Q!("    adds            " "x2, x2, x0"),
        Q!("    adc             " "x5, x5, x1"),
        Q!("    eor             " "x1, x8, x17"),
        Q!("    and             " "x0, x17, x13"),
        Q!("    sub             " "x5, x5, x0"),
        Q!("    mul             " "x0, x1, x13"),
        Q!("    umulh           " "x1, x1, x13"),
        Q!("    adds            " "x2, x2, x0"),
        Q!("    adc             " "x5, x5, x1"),
        Q!("    ldp             " "x0, x1, [sp, #128]"),
        Q!("    ldr             " "x3, [sp, #144]"),
        Q!("    mov             " "x14, #0xe000000000000000"),
        Q!("    adds            " "x0, x0, x14"),
        Q!("    sbcs            " "x1, x1, xzr"),
        Q!("    mov             " "x11, #0x1fffffff"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    mov             " "x10, #0x2000000000000000"),
        Q!("    adcs            " "x2, x2, x10"),
        Q!("    mov             " "x14, #0x1fffffffe0000000"),
        Q!("    adc             " "x5, x5, x14"),
        Q!("    lsl             " "x11, x0, #32"),
        Q!("    subs            " "x14, x0, x11"),
        Q!("    lsr             " "x10, x0, #32"),
        Q!("    sbc             " "x0, x0, x10"),
        Q!("    adds            " "x1, x1, x11"),
        Q!("    adcs            " "x3, x3, x10"),
        Q!("    adcs            " "x2, x2, x14"),
        Q!("    adcs            " "x5, x5, x0"),
        Q!("    mov             " "x14, #0xffffffffffffffff"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    csel            " "x14, x14, xzr, cs"),
        Q!("    csel            " "x11, x11, xzr, cs"),
        Q!("    csel            " "x10, x10, xzr, cs"),
        Q!("    subs            " "x1, x1, x14"),
        Q!("    sbcs            " "x3, x3, x11"),
        Q!("    sbcs            " "x2, x2, xzr"),
        Q!("    sbc             " "x5, x5, x10"),
        Q!("    stp             " "x1, x3, [sp, #128]"),
        Q!("    stp             " "x2, x5, [sp, #144]"),
        Q!(Label!("Lp256_scalarmulbase_alt_inv_midloop", 12) ":"),
        Q!("    mov             " "x1, x22"),
        Q!("    ldr             " "x2, [sp]"),
        Q!("    ldr             " "x3, [sp, #48]"),
        Q!("    and             " "x4, x2, #0xfffff"),
        Q!("    orr             " "x4, x4, #0xfffffe0000000000"),
        Q!("    and             " "x5, x3, #0xfffff"),
        Q!("    orr             " "x5, x5, #0xc000000000000000"),
        Q!("    tst             " "x5, #0x1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    add             " "x8, x4, #0x100, lsl #12"),
        Q!("    sbfx            " "x8, x8, #21, #21"),
        Q!("    mov             " "x11, #0x100000"),
        Q!("    add             " "x11, x11, x11, lsl #21"),
        Q!("    add             " "x9, x4, x11"),
        Q!("    asr             " "x9, x9, #42"),
        Q!("    add             " "x10, x5, #0x100, lsl #12"),
        Q!("    sbfx            " "x10, x10, #21, #21"),
        Q!("    add             " "x11, x5, x11"),
        Q!("    asr             " "x11, x11, #42"),
        Q!("    mul             " "x6, x8, x2"),
        Q!("    mul             " "x7, x9, x3"),
        Q!("    mul             " "x2, x10, x2"),
        Q!("    mul             " "x3, x11, x3"),
        Q!("    add             " "x4, x6, x7"),
        Q!("    add             " "x5, x2, x3"),
        Q!("    asr             " "x2, x4, #20"),
        Q!("    asr             " "x3, x5, #20"),
        Q!("    and             " "x4, x2, #0xfffff"),
        Q!("    orr             " "x4, x4, #0xfffffe0000000000"),
        Q!("    and             " "x5, x3, #0xfffff"),
        Q!("    orr             " "x5, x5, #0xc000000000000000"),
        Q!("    tst             " "x5, #0x1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    add             " "x12, x4, #0x100, lsl #12"),
        Q!("    sbfx            " "x12, x12, #21, #21"),
        Q!("    mov             " "x15, #0x100000"),
        Q!("    add             " "x15, x15, x15, lsl #21"),
        Q!("    add             " "x13, x4, x15"),
        Q!("    asr             " "x13, x13, #42"),
        Q!("    add             " "x14, x5, #0x100, lsl #12"),
        Q!("    sbfx            " "x14, x14, #21, #21"),
        Q!("    add             " "x15, x5, x15"),
        Q!("    asr             " "x15, x15, #42"),
        Q!("    mul             " "x6, x12, x2"),
        Q!("    mul             " "x7, x13, x3"),
        Q!("    mul             " "x2, x14, x2"),
        Q!("    mul             " "x3, x15, x3"),
        Q!("    add             " "x4, x6, x7"),
        Q!("    add             " "x5, x2, x3"),
        Q!("    asr             " "x2, x4, #20"),
        Q!("    asr             " "x3, x5, #20"),
        Q!("    and             " "x4, x2, #0xfffff"),
        Q!("    orr             " "x4, x4, #0xfffffe0000000000"),
        Q!("    and             " "x5, x3, #0xfffff"),
        Q!("    orr             " "x5, x5, #0xc000000000000000"),
        Q!("    tst             " "x5, #0x1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    mul             " "x2, x12, x8"),
        Q!("    mul             " "x3, x12, x9"),
        Q!("    mul             " "x6, x14, x8"),
        Q!("    mul             " "x7, x14, x9"),
        Q!("    madd            " "x8, x13, x10, x2"),
        Q!("    madd            " "x9, x13, x11, x3"),
        Q!("    madd            " "x16, x15, x10, x6"),
        Q!("    madd            " "x17, x15, x11, x7"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    tst             " "x5, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    csel            " "x6, x4, xzr, ne"),
        Q!("    ccmp            " "x1, xzr, #0x8, ne"),
        Q!("    cneg            " "x1, x1, ge"),
        Q!("    cneg            " "x6, x6, ge"),
        Q!("    csel            " "x4, x5, x4, ge"),
        Q!("    add             " "x5, x5, x6"),
        Q!("    add             " "x1, x1, #0x2"),
        Q!("    asr             " "x5, x5, #1"),
        Q!("    add             " "x12, x4, #0x100, lsl #12"),
        Q!("    sbfx            " "x12, x12, #22, #21"),
        Q!("    mov             " "x15, #0x100000"),
        Q!("    add             " "x15, x15, x15, lsl #21"),
        Q!("    add             " "x13, x4, x15"),
        Q!("    asr             " "x13, x13, #43"),
        Q!("    add             " "x14, x5, #0x100, lsl #12"),
        Q!("    sbfx            " "x14, x14, #22, #21"),
        Q!("    add             " "x15, x5, x15"),
        Q!("    asr             " "x15, x15, #43"),
        Q!("    mneg            " "x2, x12, x8"),
        Q!("    mneg            " "x3, x12, x9"),
        Q!("    mneg            " "x4, x14, x8"),
        Q!("    mneg            " "x5, x14, x9"),
        Q!("    msub            " "x10, x13, x16, x2"),
        Q!("    msub            " "x11, x13, x17, x3"),
        Q!("    msub            " "x12, x15, x16, x4"),
        Q!("    msub            " "x13, x15, x17, x5"),
        Q!("    mov             " "x22, x1"),
        Q!("    subs            " "x21, x21, #0x1"),
        Q!("    bne             " Label!("Lp256_scalarmulbase_alt_inv_loop", 13, Before)),
        Q!("    ldr             " "x0, [sp]"),
        Q!("    ldr             " "x1, [sp, #48]"),
        Q!("    mul             " "x0, x0, x10"),
        Q!("    madd            " "x1, x1, x11, x0"),
        Q!("    asr             " "x0, x1, #63"),
        Q!("    cmp             " "x10, xzr"),
        Q!("    csetm           " "x14, mi"),
        Q!("    cneg            " "x10, x10, mi"),
        Q!("    eor             " "x14, x14, x0"),
        Q!("    cmp             " "x11, xzr"),
        Q!("    csetm           " "x15, mi"),
        Q!("    cneg            " "x11, x11, mi"),
        Q!("    eor             " "x15, x15, x0"),
        Q!("    cmp             " "x12, xzr"),
        Q!("    csetm           " "x16, mi"),
        Q!("    cneg            " "x12, x12, mi"),
        Q!("    eor             " "x16, x16, x0"),
        Q!("    cmp             " "x13, xzr"),
        Q!("    csetm           " "x17, mi"),
        Q!("    cneg            " "x13, x13, mi"),
        Q!("    eor             " "x17, x17, x0"),
        Q!("    and             " "x0, x10, x14"),
        Q!("    and             " "x1, x11, x15"),
        Q!("    add             " "x9, x0, x1"),
        Q!("    ldr             " "x7, [sp, #96]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x4, x9, x0"),
        Q!("    adc             " "x2, xzr, x1"),
        Q!("    ldr             " "x8, [sp, #128]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x4, x4, x0"),
        Q!("    str             " "x4, [sp, #96]"),
        Q!("    adc             " "x2, x2, x1"),
        Q!("    ldr             " "x7, [sp, #104]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x2, x2, x0"),
        Q!("    adc             " "x6, xzr, x1"),
        Q!("    ldr             " "x8, [sp, #136]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x2, x2, x0"),
        Q!("    str             " "x2, [sp, #104]"),
        Q!("    adc             " "x6, x6, x1"),
        Q!("    ldr             " "x7, [sp, #112]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x6, x6, x0"),
        Q!("    adc             " "x5, xzr, x1"),
        Q!("    ldr             " "x8, [sp, #144]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x6, x6, x0"),
        Q!("    str             " "x6, [sp, #112]"),
        Q!("    adc             " "x5, x5, x1"),
        Q!("    ldr             " "x7, [sp, #120]"),
        Q!("    eor             " "x1, x7, x14"),
        Q!("    and             " "x3, x14, x10"),
        Q!("    neg             " "x3, x3"),
        Q!("    mul             " "x0, x1, x10"),
        Q!("    umulh           " "x1, x1, x10"),
        Q!("    adds            " "x5, x5, x0"),
        Q!("    adc             " "x3, x3, x1"),
        Q!("    ldr             " "x8, [sp, #152]"),
        Q!("    eor             " "x1, x8, x15"),
        Q!("    and             " "x0, x15, x11"),
        Q!("    sub             " "x3, x3, x0"),
        Q!("    mul             " "x0, x1, x11"),
        Q!("    umulh           " "x1, x1, x11"),
        Q!("    adds            " "x5, x5, x0"),
        Q!("    adc             " "x3, x3, x1"),
        Q!("    ldp             " "x0, x1, [sp, #96]"),
        Q!("    ldr             " "x2, [sp, #112]"),
        Q!("    mov             " "x14, #0xe000000000000000"),
        Q!("    adds            " "x0, x0, x14"),
        Q!("    sbcs            " "x1, x1, xzr"),
        Q!("    mov             " "x11, #0x1fffffff"),
        Q!("    adcs            " "x2, x2, x11"),
        Q!("    mov             " "x10, #0x2000000000000000"),
        Q!("    adcs            " "x5, x5, x10"),
        Q!("    mov             " "x14, #0x1fffffffe0000000"),
        Q!("    adc             " "x3, x3, x14"),
        Q!("    lsl             " "x11, x0, #32"),
        Q!("    subs            " "x14, x0, x11"),
        Q!("    lsr             " "x10, x0, #32"),
        Q!("    sbc             " "x0, x0, x10"),
        Q!("    adds            " "x1, x1, x11"),
        Q!("    adcs            " "x2, x2, x10"),
        Q!("    adcs            " "x5, x5, x14"),
        Q!("    adcs            " "x3, x3, x0"),
        Q!("    mov             " "x14, #0xffffffffffffffff"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    csel            " "x14, x14, xzr, cs"),
        Q!("    csel            " "x11, x11, xzr, cs"),
        Q!("    csel            " "x10, x10, xzr, cs"),
        Q!("    subs            " "x1, x1, x14"),
        Q!("    sbcs            " "x2, x2, x11"),
        Q!("    sbcs            " "x5, x5, xzr"),
        Q!("    sbc             " "x3, x3, x10"),
        Q!("    mov             " "x10, #0xffffffffffffffff"),
        Q!("    subs            " "x10, x1, x10"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    sbcs            " "x11, x2, x11"),
        Q!("    mov             " "x13, #0xffffffff00000001"),
        Q!("    sbcs            " "x12, x5, xzr"),
        Q!("    sbcs            " "x13, x3, x13"),
        Q!("    csel            " "x10, x1, x10, cc"),
        Q!("    csel            " "x11, x2, x11, cc"),
        Q!("    csel            " "x12, x5, x12, cc"),
        Q!("    csel            " "x13, x3, x13, cc"),
        Q!("    stp             " "x10, x11, [x20]"),
        Q!("    stp             " "x12, x13, [x20, #16]"),
        Q!("    add             " "sp, sp, # (160 + 0)"),
        Q!("    ldp             " "x23, x24, [sp], #16"),
        Q!("    ldp             " "x21, x22, [sp], #16"),
        Q!("    ldp             " "x19, x20, [sp], #16"),
        Q!("    ret             " ),

        Q!(Label!("Lp256_scalarmulbase_alt_local_montmul_p256", 6) ":"),

        Q!("    ldp             " "x3, x4, [x1]"),
        Q!("    ldp             " "x7, x8, [x2]"),
        Q!("    mul             " "x12, x3, x7"),
        Q!("    umulh           " "x13, x3, x7"),
        Q!("    mul             " "x11, x3, x8"),
        Q!("    umulh           " "x14, x3, x8"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    ldp             " "x9, x10, [x2, #16]"),
        Q!("    mul             " "x11, x3, x9"),
        Q!("    umulh           " "x15, x3, x9"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x3, x10"),
        Q!("    umulh           " "x16, x3, x10"),
        Q!("    adcs            " "x15, x15, x11"),
        Q!("    adc             " "x16, x16, xzr"),
        Q!("    ldp             " "x5, x6, [x1, #16]"),
        Q!("    mul             " "x11, x4, x7"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    mul             " "x11, x4, x8"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x4, x9"),
        Q!("    adcs            " "x15, x15, x11"),
        Q!("    mul             " "x11, x4, x10"),
        Q!("    adcs            " "x16, x16, x11"),
        Q!("    umulh           " "x3, x4, x10"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    umulh           " "x11, x4, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    umulh           " "x11, x4, x8"),
        Q!("    adcs            " "x15, x15, x11"),
        Q!("    umulh           " "x11, x4, x9"),
        Q!("    adcs            " "x16, x16, x11"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    mul             " "x11, x5, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    mul             " "x11, x5, x8"),
        Q!("    adcs            " "x15, x15, x11"),
        Q!("    mul             " "x11, x5, x9"),
        Q!("    adcs            " "x16, x16, x11"),
        Q!("    mul             " "x11, x5, x10"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x4, x5, x10"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    umulh           " "x11, x5, x7"),
        Q!("    adds            " "x15, x15, x11"),
        Q!("    umulh           " "x11, x5, x8"),
        Q!("    adcs            " "x16, x16, x11"),
        Q!("    umulh           " "x11, x5, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    mul             " "x11, x6, x7"),
        Q!("    adds            " "x15, x15, x11"),
        Q!("    mul             " "x11, x6, x8"),
        Q!("    adcs            " "x16, x16, x11"),
        Q!("    mul             " "x11, x6, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    mul             " "x11, x6, x10"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    umulh           " "x5, x6, x10"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    adds            " "x13, x13, x12, lsl #32"),
        Q!("    lsr             " "x11, x12, #32"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x12, x10"),
        Q!("    umulh           " "x12, x12, x10"),
        Q!("    adcs            " "x15, x15, x11"),
        Q!("    adc             " "x12, x12, xzr"),
        Q!("    umulh           " "x11, x6, x7"),
        Q!("    adds            " "x16, x16, x11"),
        Q!("    umulh           " "x11, x6, x8"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x11, x6, x9"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    adds            " "x14, x14, x13, lsl #32"),
        Q!("    lsr             " "x11, x13, #32"),
        Q!("    adcs            " "x15, x15, x11"),
        Q!("    mul             " "x11, x13, x10"),
        Q!("    umulh           " "x13, x13, x10"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    adc             " "x13, x13, xzr"),
        Q!("    adds            " "x15, x15, x14, lsl #32"),
        Q!("    lsr             " "x11, x14, #32"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    mul             " "x11, x14, x10"),
        Q!("    umulh           " "x14, x14, x10"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x12, x12, x15, lsl #32"),
        Q!("    lsr             " "x11, x15, #32"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    mul             " "x11, x15, x10"),
        Q!("    umulh           " "x15, x15, x10"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    adc             " "x15, x15, xzr"),
        Q!("    adds            " "x12, x12, x16"),
        Q!("    adcs            " "x13, x13, x3"),
        Q!("    adcs            " "x14, x14, x4"),
        Q!("    adcs            " "x15, x15, x5"),
        Q!("    cset            " "x8, cs"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    adds            " "x16, x12, #0x1"),
        Q!("    sbcs            " "x3, x13, x11"),
        Q!("    sbcs            " "x4, x14, xzr"),
        Q!("    sbcs            " "x5, x15, x10"),
        Q!("    sbcs            " "xzr, x8, xzr"),
        Q!("    csel            " "x12, x12, x16, cc"),
        Q!("    csel            " "x13, x13, x3, cc"),
        Q!("    csel            " "x14, x14, x4, cc"),
        Q!("    csel            " "x15, x15, x5, cc"),
        Q!("    stp             " "x12, x13, [x0]"),
        Q!("    stp             " "x14, x15, [x0, #16]"),
        Q!("    ret             " ),

        Q!(Label!("Lp256_scalarmulbase_alt_local_montsqr_p256", 5) ":"),

        Q!("    ldp             " "x2, x3, [x1]"),
        Q!("    mul             " "x9, x2, x3"),
        Q!("    umulh           " "x10, x2, x3"),
        Q!("    ldp             " "x4, x5, [x1, #16]"),
        Q!("    mul             " "x11, x2, x5"),
        Q!("    umulh           " "x12, x2, x5"),
        Q!("    mul             " "x6, x2, x4"),
        Q!("    umulh           " "x7, x2, x4"),
        Q!("    adds            " "x10, x10, x6"),
        Q!("    adcs            " "x11, x11, x7"),
        Q!("    mul             " "x6, x3, x4"),
        Q!("    umulh           " "x7, x3, x4"),
        Q!("    adc             " "x7, x7, xzr"),
        Q!("    adds            " "x11, x11, x6"),
        Q!("    mul             " "x13, x4, x5"),
        Q!("    umulh           " "x14, x4, x5"),
        Q!("    adcs            " "x12, x12, x7"),
        Q!("    mul             " "x6, x3, x5"),
        Q!("    umulh           " "x7, x3, x5"),
        Q!("    adc             " "x7, x7, xzr"),
        Q!("    adds            " "x12, x12, x6"),
        Q!("    adcs            " "x13, x13, x7"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x9, x9, x9"),
        Q!("    adcs            " "x10, x10, x10"),
        Q!("    adcs            " "x11, x11, x11"),
        Q!("    adcs            " "x12, x12, x12"),
        Q!("    adcs            " "x13, x13, x13"),
        Q!("    adcs            " "x14, x14, x14"),
        Q!("    cset            " "x7, cs"),
        Q!("    umulh           " "x6, x2, x2"),
        Q!("    mul             " "x8, x2, x2"),
        Q!("    adds            " "x9, x9, x6"),
        Q!("    mul             " "x6, x3, x3"),
        Q!("    adcs            " "x10, x10, x6"),
        Q!("    umulh           " "x6, x3, x3"),
        Q!("    adcs            " "x11, x11, x6"),
        Q!("    mul             " "x6, x4, x4"),
        Q!("    adcs            " "x12, x12, x6"),
        Q!("    umulh           " "x6, x4, x4"),
        Q!("    adcs            " "x13, x13, x6"),
        Q!("    mul             " "x6, x5, x5"),
        Q!("    adcs            " "x14, x14, x6"),
        Q!("    umulh           " "x6, x5, x5"),
        Q!("    adc             " "x7, x7, x6"),
        Q!("    mov             " "x5, #0xffffffff00000001"),
        Q!("    adds            " "x9, x9, x8, lsl #32"),
        Q!("    lsr             " "x2, x8, #32"),
        Q!("    adcs            " "x10, x10, x2"),
        Q!("    mul             " "x2, x8, x5"),
        Q!("    umulh           " "x8, x8, x5"),
        Q!("    adcs            " "x11, x11, x2"),
        Q!("    adc             " "x8, x8, xzr"),
        Q!("    adds            " "x10, x10, x9, lsl #32"),
        Q!("    lsr             " "x2, x9, #32"),
        Q!("    adcs            " "x11, x11, x2"),
        Q!("    mul             " "x2, x9, x5"),
        Q!("    umulh           " "x9, x9, x5"),
        Q!("    adcs            " "x8, x8, x2"),
        Q!("    adc             " "x9, x9, xzr"),
        Q!("    adds            " "x11, x11, x10, lsl #32"),
        Q!("    lsr             " "x2, x10, #32"),
        Q!("    adcs            " "x8, x8, x2"),
        Q!("    mul             " "x2, x10, x5"),
        Q!("    umulh           " "x10, x10, x5"),
        Q!("    adcs            " "x9, x9, x2"),
        Q!("    adc             " "x10, x10, xzr"),
        Q!("    adds            " "x8, x8, x11, lsl #32"),
        Q!("    lsr             " "x2, x11, #32"),
        Q!("    adcs            " "x9, x9, x2"),
        Q!("    mul             " "x2, x11, x5"),
        Q!("    umulh           " "x11, x11, x5"),
        Q!("    adcs            " "x10, x10, x2"),
        Q!("    adc             " "x11, x11, xzr"),
        Q!("    adds            " "x8, x8, x12"),
        Q!("    adcs            " "x9, x9, x13"),
        Q!("    adcs            " "x10, x10, x14"),
        Q!("    adcs            " "x11, x11, x7"),
        Q!("    cset            " "x2, cs"),
        Q!("    mov             " "x3, #0xffffffff"),
        Q!("    adds            " "x12, x8, #0x1"),
        Q!("    sbcs            " "x13, x9, x3"),
        Q!("    sbcs            " "x14, x10, xzr"),
        Q!("    sbcs            " "x7, x11, x5"),
        Q!("    sbcs            " "xzr, x2, xzr"),
        Q!("    csel            " "x8, x8, x12, cc"),
        Q!("    csel            " "x9, x9, x13, cc"),
        Q!("    csel            " "x10, x10, x14, cc"),
        Q!("    csel            " "x11, x11, x7, cc"),
        Q!("    stp             " "x8, x9, [x0]"),
        Q!("    stp             " "x10, x11, [x0, #16]"),
        Q!("    ret             " ),

        Q!(Label!("Lp256_scalarmulbase_alt_local_p256_montjmixadd", 4) ":"),

        Q!("    sub             " "sp, sp, # (192 + 0)"),
        Q!("    mov             " "x15, x0"),
        Q!("    mov             " "x16, x1"),
        Q!("    mov             " "x17, x2"),
        Q!("    ldp             " "x2, x3, [x16, #64]"),
        Q!("    mul             " "x9, x2, x3"),
        Q!("    umulh           " "x10, x2, x3"),
        Q!("    ldp             " "x4, x5, [x16, #80]"),
        Q!("    mul             " "x11, x2, x5"),
        Q!("    umulh           " "x12, x2, x5"),
        Q!("    mul             " "x6, x2, x4"),
        Q!("    umulh           " "x7, x2, x4"),
        Q!("    adds            " "x10, x10, x6"),
        Q!("    adcs            " "x11, x11, x7"),
        Q!("    mul             " "x6, x3, x4"),
        Q!("    umulh           " "x7, x3, x4"),
        Q!("    adc             " "x7, x7, xzr"),
        Q!("    adds            " "x11, x11, x6"),
        Q!("    mul             " "x13, x4, x5"),
        Q!("    umulh           " "x14, x4, x5"),
        Q!("    adcs            " "x12, x12, x7"),
        Q!("    mul             " "x6, x3, x5"),
        Q!("    umulh           " "x7, x3, x5"),
        Q!("    adc             " "x7, x7, xzr"),
        Q!("    adds            " "x12, x12, x6"),
        Q!("    adcs            " "x13, x13, x7"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x9, x9, x9"),
        Q!("    adcs            " "x10, x10, x10"),
        Q!("    adcs            " "x11, x11, x11"),
        Q!("    adcs            " "x12, x12, x12"),
        Q!("    adcs            " "x13, x13, x13"),
        Q!("    adcs            " "x14, x14, x14"),
        Q!("    cset            " "x7, cs"),
        Q!("    umulh           " "x6, x2, x2"),
        Q!("    mul             " "x8, x2, x2"),
        Q!("    adds            " "x9, x9, x6"),
        Q!("    mul             " "x6, x3, x3"),
        Q!("    adcs            " "x10, x10, x6"),
        Q!("    umulh           " "x6, x3, x3"),
        Q!("    adcs            " "x11, x11, x6"),
        Q!("    mul             " "x6, x4, x4"),
        Q!("    adcs            " "x12, x12, x6"),
        Q!("    umulh           " "x6, x4, x4"),
        Q!("    adcs            " "x13, x13, x6"),
        Q!("    mul             " "x6, x5, x5"),
        Q!("    adcs            " "x14, x14, x6"),
        Q!("    umulh           " "x6, x5, x5"),
        Q!("    adc             " "x7, x7, x6"),
        Q!("    adds            " "x9, x9, x8, lsl #32"),
        Q!("    lsr             " "x3, x8, #32"),
        Q!("    adcs            " "x10, x10, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x8, x3"),
        Q!("    umulh           " "x8, x8, x3"),
        Q!("    adcs            " "x11, x11, x2"),
        Q!("    adc             " "x8, x8, xzr"),
        Q!("    adds            " "x10, x10, x9, lsl #32"),
        Q!("    lsr             " "x3, x9, #32"),
        Q!("    adcs            " "x11, x11, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x9, x3"),
        Q!("    umulh           " "x9, x9, x3"),
        Q!("    adcs            " "x8, x8, x2"),
        Q!("    adc             " "x9, x9, xzr"),
        Q!("    adds            " "x11, x11, x10, lsl #32"),
        Q!("    lsr             " "x3, x10, #32"),
        Q!("    adcs            " "x8, x8, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x10, x3"),
        Q!("    umulh           " "x10, x10, x3"),
        Q!("    adcs            " "x9, x9, x2"),
        Q!("    adc             " "x10, x10, xzr"),
        Q!("    adds            " "x8, x8, x11, lsl #32"),
        Q!("    lsr             " "x3, x11, #32"),
        Q!("    adcs            " "x9, x9, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x11, x3"),
        Q!("    umulh           " "x11, x11, x3"),
        Q!("    adcs            " "x10, x10, x2"),
        Q!("    adc             " "x11, x11, xzr"),
        Q!("    adds            " "x8, x8, x12"),
        Q!("    adcs            " "x9, x9, x13"),
        Q!("    adcs            " "x10, x10, x14"),
        Q!("    adcs            " "x11, x11, x7"),
        Q!("    mov             " "x2, #0xffffffffffffffff"),
        Q!("    csel            " "x2, xzr, x2, cc"),
        Q!("    mov             " "x3, #0xffffffff"),
        Q!("    csel            " "x3, xzr, x3, cc"),
        Q!("    mov             " "x5, #0xffffffff00000001"),
        Q!("    csel            " "x5, xzr, x5, cc"),
        Q!("    subs            " "x8, x8, x2"),
        Q!("    sbcs            " "x9, x9, x3"),
        Q!("    sbcs            " "x10, x10, xzr"),
        Q!("    sbc             " "x11, x11, x5"),
        Q!("    stp             " "x8, x9, [sp]"),
        Q!("    stp             " "x10, x11, [sp, #16]"),
        Q!("    ldp             " "x3, x4, [x16, #64]"),
        Q!("    ldp             " "x7, x8, [x17, #32]"),
        Q!("    mul             " "x12, x3, x7"),
        Q!("    umulh           " "x13, x3, x7"),
        Q!("    mul             " "x11, x3, x8"),
        Q!("    umulh           " "x14, x3, x8"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    ldp             " "x9, x10, [x17, #48]"),
        Q!("    mul             " "x11, x3, x9"),
        Q!("    umulh           " "x0, x3, x9"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x3, x10"),
        Q!("    umulh           " "x1, x3, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x1, x1, xzr"),
        Q!("    ldp             " "x5, x6, [x16, #80]"),
        Q!("    mul             " "x11, x4, x7"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    mul             " "x11, x4, x8"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x4, x9"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x4, x10"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x3, x4, x10"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    umulh           " "x11, x4, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    umulh           " "x11, x4, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x4, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    mul             " "x11, x5, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    mul             " "x11, x5, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x5, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x5, x10"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x4, x5, x10"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    umulh           " "x11, x5, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x5, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x5, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    mul             " "x11, x6, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    mul             " "x11, x6, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x6, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    mul             " "x11, x6, x10"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    umulh           " "x5, x6, x10"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    adds            " "x13, x13, x12, lsl #32"),
        Q!("    lsr             " "x11, x12, #32"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x12, x10"),
        Q!("    umulh           " "x12, x12, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x12, x12, xzr"),
        Q!("    umulh           " "x11, x6, x7"),
        Q!("    adds            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x6, x8"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x11, x6, x9"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    adds            " "x14, x14, x13, lsl #32"),
        Q!("    lsr             " "x11, x13, #32"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x13, x10"),
        Q!("    umulh           " "x13, x13, x10"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    adc             " "x13, x13, xzr"),
        Q!("    adds            " "x0, x0, x14, lsl #32"),
        Q!("    lsr             " "x11, x14, #32"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    mul             " "x11, x14, x10"),
        Q!("    umulh           " "x14, x14, x10"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x12, x12, x0, lsl #32"),
        Q!("    lsr             " "x11, x0, #32"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    mul             " "x11, x0, x10"),
        Q!("    umulh           " "x0, x0, x10"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    adc             " "x0, x0, xzr"),
        Q!("    adds            " "x12, x12, x1"),
        Q!("    adcs            " "x13, x13, x3"),
        Q!("    adcs            " "x14, x14, x4"),
        Q!("    adcs            " "x0, x0, x5"),
        Q!("    cset            " "x8, cs"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    adds            " "x1, x12, #0x1"),
        Q!("    sbcs            " "x3, x13, x11"),
        Q!("    sbcs            " "x4, x14, xzr"),
        Q!("    sbcs            " "x5, x0, x10"),
        Q!("    sbcs            " "xzr, x8, xzr"),
        Q!("    csel            " "x12, x12, x1, cc"),
        Q!("    csel            " "x13, x13, x3, cc"),
        Q!("    csel            " "x14, x14, x4, cc"),
        Q!("    csel            " "x0, x0, x5, cc"),
        Q!("    stp             " "x12, x13, [sp, #32]"),
        Q!("    stp             " "x14, x0, [sp, #48]"),
        Q!("    ldp             " "x3, x4, [sp]"),
        Q!("    ldp             " "x7, x8, [x17]"),
        Q!("    mul             " "x12, x3, x7"),
        Q!("    umulh           " "x13, x3, x7"),
        Q!("    mul             " "x11, x3, x8"),
        Q!("    umulh           " "x14, x3, x8"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    ldp             " "x9, x10, [x17, #16]"),
        Q!("    mul             " "x11, x3, x9"),
        Q!("    umulh           " "x0, x3, x9"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x3, x10"),
        Q!("    umulh           " "x1, x3, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x1, x1, xzr"),
        Q!("    ldp             " "x5, x6, [sp, #16]"),
        Q!("    mul             " "x11, x4, x7"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    mul             " "x11, x4, x8"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x4, x9"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x4, x10"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x3, x4, x10"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    umulh           " "x11, x4, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    umulh           " "x11, x4, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x4, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    mul             " "x11, x5, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    mul             " "x11, x5, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x5, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x5, x10"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x4, x5, x10"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    umulh           " "x11, x5, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x5, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x5, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    mul             " "x11, x6, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    mul             " "x11, x6, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x6, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    mul             " "x11, x6, x10"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    umulh           " "x5, x6, x10"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    adds            " "x13, x13, x12, lsl #32"),
        Q!("    lsr             " "x11, x12, #32"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x12, x10"),
        Q!("    umulh           " "x12, x12, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x12, x12, xzr"),
        Q!("    umulh           " "x11, x6, x7"),
        Q!("    adds            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x6, x8"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x11, x6, x9"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    adds            " "x14, x14, x13, lsl #32"),
        Q!("    lsr             " "x11, x13, #32"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x13, x10"),
        Q!("    umulh           " "x13, x13, x10"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    adc             " "x13, x13, xzr"),
        Q!("    adds            " "x0, x0, x14, lsl #32"),
        Q!("    lsr             " "x11, x14, #32"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    mul             " "x11, x14, x10"),
        Q!("    umulh           " "x14, x14, x10"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x12, x12, x0, lsl #32"),
        Q!("    lsr             " "x11, x0, #32"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    mul             " "x11, x0, x10"),
        Q!("    umulh           " "x0, x0, x10"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    adc             " "x0, x0, xzr"),
        Q!("    adds            " "x12, x12, x1"),
        Q!("    adcs            " "x13, x13, x3"),
        Q!("    adcs            " "x14, x14, x4"),
        Q!("    adcs            " "x0, x0, x5"),
        Q!("    cset            " "x8, cs"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    adds            " "x1, x12, #0x1"),
        Q!("    sbcs            " "x3, x13, x11"),
        Q!("    sbcs            " "x4, x14, xzr"),
        Q!("    sbcs            " "x5, x0, x10"),
        Q!("    sbcs            " "xzr, x8, xzr"),
        Q!("    csel            " "x12, x12, x1, cc"),
        Q!("    csel            " "x13, x13, x3, cc"),
        Q!("    csel            " "x14, x14, x4, cc"),
        Q!("    csel            " "x0, x0, x5, cc"),
        Q!("    stp             " "x12, x13, [sp, #64]"),
        Q!("    stp             " "x14, x0, [sp, #80]"),
        Q!("    ldp             " "x3, x4, [sp]"),
        Q!("    ldp             " "x7, x8, [sp, #32]"),
        Q!("    mul             " "x12, x3, x7"),
        Q!("    umulh           " "x13, x3, x7"),
        Q!("    mul             " "x11, x3, x8"),
        Q!("    umulh           " "x14, x3, x8"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    ldp             " "x9, x10, [sp, #48]"),
        Q!("    mul             " "x11, x3, x9"),
        Q!("    umulh           " "x0, x3, x9"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x3, x10"),
        Q!("    umulh           " "x1, x3, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x1, x1, xzr"),
        Q!("    ldp             " "x5, x6, [sp, #16]"),
        Q!("    mul             " "x11, x4, x7"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    mul             " "x11, x4, x8"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x4, x9"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x4, x10"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x3, x4, x10"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    umulh           " "x11, x4, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    umulh           " "x11, x4, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x4, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    mul             " "x11, x5, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    mul             " "x11, x5, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x5, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x5, x10"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x4, x5, x10"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    umulh           " "x11, x5, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x5, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x5, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    mul             " "x11, x6, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    mul             " "x11, x6, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x6, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    mul             " "x11, x6, x10"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    umulh           " "x5, x6, x10"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    adds            " "x13, x13, x12, lsl #32"),
        Q!("    lsr             " "x11, x12, #32"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x12, x10"),
        Q!("    umulh           " "x12, x12, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x12, x12, xzr"),
        Q!("    umulh           " "x11, x6, x7"),
        Q!("    adds            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x6, x8"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x11, x6, x9"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    adds            " "x14, x14, x13, lsl #32"),
        Q!("    lsr             " "x11, x13, #32"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x13, x10"),
        Q!("    umulh           " "x13, x13, x10"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    adc             " "x13, x13, xzr"),
        Q!("    adds            " "x0, x0, x14, lsl #32"),
        Q!("    lsr             " "x11, x14, #32"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    mul             " "x11, x14, x10"),
        Q!("    umulh           " "x14, x14, x10"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x12, x12, x0, lsl #32"),
        Q!("    lsr             " "x11, x0, #32"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    mul             " "x11, x0, x10"),
        Q!("    umulh           " "x0, x0, x10"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    adc             " "x0, x0, xzr"),
        Q!("    adds            " "x12, x12, x1"),
        Q!("    adcs            " "x13, x13, x3"),
        Q!("    adcs            " "x14, x14, x4"),
        Q!("    adcs            " "x0, x0, x5"),
        Q!("    cset            " "x8, cs"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    adds            " "x1, x12, #0x1"),
        Q!("    sbcs            " "x3, x13, x11"),
        Q!("    sbcs            " "x4, x14, xzr"),
        Q!("    sbcs            " "x5, x0, x10"),
        Q!("    sbcs            " "xzr, x8, xzr"),
        Q!("    csel            " "x12, x12, x1, cc"),
        Q!("    csel            " "x13, x13, x3, cc"),
        Q!("    csel            " "x14, x14, x4, cc"),
        Q!("    csel            " "x0, x0, x5, cc"),
        Q!("    stp             " "x12, x13, [sp, #32]"),
        Q!("    stp             " "x14, x0, [sp, #48]"),
        Q!("    ldp             " "x5, x6, [sp, #64]"),
        Q!("    ldp             " "x4, x3, [x16]"),
        Q!("    subs            " "x5, x5, x4"),
        Q!("    sbcs            " "x6, x6, x3"),
        Q!("    ldp             " "x7, x8, [sp, #80]"),
        Q!("    ldp             " "x4, x3, [x16, #16]"),
        Q!("    sbcs            " "x7, x7, x4"),
        Q!("    sbcs            " "x8, x8, x3"),
        Q!("    csetm           " "x3, cc"),
        Q!("    adds            " "x5, x5, x3"),
        Q!("    mov             " "x4, #0xffffffff"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adcs            " "x6, x6, x4"),
        Q!("    adcs            " "x7, x7, xzr"),
        Q!("    mov             " "x4, #0xffffffff00000001"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adc             " "x8, x8, x4"),
        Q!("    stp             " "x5, x6, [sp, #160]"),
        Q!("    stp             " "x7, x8, [sp, #176]"),
        Q!("    ldp             " "x5, x6, [sp, #32]"),
        Q!("    ldp             " "x4, x3, [x16, #32]"),
        Q!("    subs            " "x5, x5, x4"),
        Q!("    sbcs            " "x6, x6, x3"),
        Q!("    ldp             " "x7, x8, [sp, #48]"),
        Q!("    ldp             " "x4, x3, [x16, #48]"),
        Q!("    sbcs            " "x7, x7, x4"),
        Q!("    sbcs            " "x8, x8, x3"),
        Q!("    csetm           " "x3, cc"),
        Q!("    adds            " "x5, x5, x3"),
        Q!("    mov             " "x4, #0xffffffff"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adcs            " "x6, x6, x4"),
        Q!("    adcs            " "x7, x7, xzr"),
        Q!("    mov             " "x4, #0xffffffff00000001"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adc             " "x8, x8, x4"),
        Q!("    stp             " "x5, x6, [sp, #32]"),
        Q!("    stp             " "x7, x8, [sp, #48]"),
        Q!("    ldp             " "x2, x3, [sp, #160]"),
        Q!("    mul             " "x9, x2, x3"),
        Q!("    umulh           " "x10, x2, x3"),
        Q!("    ldp             " "x4, x5, [sp, #176]"),
        Q!("    mul             " "x11, x2, x5"),
        Q!("    umulh           " "x12, x2, x5"),
        Q!("    mul             " "x6, x2, x4"),
        Q!("    umulh           " "x7, x2, x4"),
        Q!("    adds            " "x10, x10, x6"),
        Q!("    adcs            " "x11, x11, x7"),
        Q!("    mul             " "x6, x3, x4"),
        Q!("    umulh           " "x7, x3, x4"),
        Q!("    adc             " "x7, x7, xzr"),
        Q!("    adds            " "x11, x11, x6"),
        Q!("    mul             " "x13, x4, x5"),
        Q!("    umulh           " "x14, x4, x5"),
        Q!("    adcs            " "x12, x12, x7"),
        Q!("    mul             " "x6, x3, x5"),
        Q!("    umulh           " "x7, x3, x5"),
        Q!("    adc             " "x7, x7, xzr"),
        Q!("    adds            " "x12, x12, x6"),
        Q!("    adcs            " "x13, x13, x7"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x9, x9, x9"),
        Q!("    adcs            " "x10, x10, x10"),
        Q!("    adcs            " "x11, x11, x11"),
        Q!("    adcs            " "x12, x12, x12"),
        Q!("    adcs            " "x13, x13, x13"),
        Q!("    adcs            " "x14, x14, x14"),
        Q!("    cset            " "x7, cs"),
        Q!("    umulh           " "x6, x2, x2"),
        Q!("    mul             " "x8, x2, x2"),
        Q!("    adds            " "x9, x9, x6"),
        Q!("    mul             " "x6, x3, x3"),
        Q!("    adcs            " "x10, x10, x6"),
        Q!("    umulh           " "x6, x3, x3"),
        Q!("    adcs            " "x11, x11, x6"),
        Q!("    mul             " "x6, x4, x4"),
        Q!("    adcs            " "x12, x12, x6"),
        Q!("    umulh           " "x6, x4, x4"),
        Q!("    adcs            " "x13, x13, x6"),
        Q!("    mul             " "x6, x5, x5"),
        Q!("    adcs            " "x14, x14, x6"),
        Q!("    umulh           " "x6, x5, x5"),
        Q!("    adc             " "x7, x7, x6"),
        Q!("    adds            " "x9, x9, x8, lsl #32"),
        Q!("    lsr             " "x3, x8, #32"),
        Q!("    adcs            " "x10, x10, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x8, x3"),
        Q!("    umulh           " "x8, x8, x3"),
        Q!("    adcs            " "x11, x11, x2"),
        Q!("    adc             " "x8, x8, xzr"),
        Q!("    adds            " "x10, x10, x9, lsl #32"),
        Q!("    lsr             " "x3, x9, #32"),
        Q!("    adcs            " "x11, x11, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x9, x3"),
        Q!("    umulh           " "x9, x9, x3"),
        Q!("    adcs            " "x8, x8, x2"),
        Q!("    adc             " "x9, x9, xzr"),
        Q!("    adds            " "x11, x11, x10, lsl #32"),
        Q!("    lsr             " "x3, x10, #32"),
        Q!("    adcs            " "x8, x8, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x10, x3"),
        Q!("    umulh           " "x10, x10, x3"),
        Q!("    adcs            " "x9, x9, x2"),
        Q!("    adc             " "x10, x10, xzr"),
        Q!("    adds            " "x8, x8, x11, lsl #32"),
        Q!("    lsr             " "x3, x11, #32"),
        Q!("    adcs            " "x9, x9, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x11, x3"),
        Q!("    umulh           " "x11, x11, x3"),
        Q!("    adcs            " "x10, x10, x2"),
        Q!("    adc             " "x11, x11, xzr"),
        Q!("    adds            " "x8, x8, x12"),
        Q!("    adcs            " "x9, x9, x13"),
        Q!("    adcs            " "x10, x10, x14"),
        Q!("    adcs            " "x11, x11, x7"),
        Q!("    mov             " "x2, #0xffffffffffffffff"),
        Q!("    csel            " "x2, xzr, x2, cc"),
        Q!("    mov             " "x3, #0xffffffff"),
        Q!("    csel            " "x3, xzr, x3, cc"),
        Q!("    mov             " "x5, #0xffffffff00000001"),
        Q!("    csel            " "x5, xzr, x5, cc"),
        Q!("    subs            " "x8, x8, x2"),
        Q!("    sbcs            " "x9, x9, x3"),
        Q!("    sbcs            " "x10, x10, xzr"),
        Q!("    sbc             " "x11, x11, x5"),
        Q!("    stp             " "x8, x9, [sp, #96]"),
        Q!("    stp             " "x10, x11, [sp, #112]"),
        Q!("    ldp             " "x2, x3, [sp, #32]"),
        Q!("    mul             " "x9, x2, x3"),
        Q!("    umulh           " "x10, x2, x3"),
        Q!("    ldp             " "x4, x5, [sp, #48]"),
        Q!("    mul             " "x11, x2, x5"),
        Q!("    umulh           " "x12, x2, x5"),
        Q!("    mul             " "x6, x2, x4"),
        Q!("    umulh           " "x7, x2, x4"),
        Q!("    adds            " "x10, x10, x6"),
        Q!("    adcs            " "x11, x11, x7"),
        Q!("    mul             " "x6, x3, x4"),
        Q!("    umulh           " "x7, x3, x4"),
        Q!("    adc             " "x7, x7, xzr"),
        Q!("    adds            " "x11, x11, x6"),
        Q!("    mul             " "x13, x4, x5"),
        Q!("    umulh           " "x14, x4, x5"),
        Q!("    adcs            " "x12, x12, x7"),
        Q!("    mul             " "x6, x3, x5"),
        Q!("    umulh           " "x7, x3, x5"),
        Q!("    adc             " "x7, x7, xzr"),
        Q!("    adds            " "x12, x12, x6"),
        Q!("    adcs            " "x13, x13, x7"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x9, x9, x9"),
        Q!("    adcs            " "x10, x10, x10"),
        Q!("    adcs            " "x11, x11, x11"),
        Q!("    adcs            " "x12, x12, x12"),
        Q!("    adcs            " "x13, x13, x13"),
        Q!("    adcs            " "x14, x14, x14"),
        Q!("    cset            " "x7, cs"),
        Q!("    umulh           " "x6, x2, x2"),
        Q!("    mul             " "x8, x2, x2"),
        Q!("    adds            " "x9, x9, x6"),
        Q!("    mul             " "x6, x3, x3"),
        Q!("    adcs            " "x10, x10, x6"),
        Q!("    umulh           " "x6, x3, x3"),
        Q!("    adcs            " "x11, x11, x6"),
        Q!("    mul             " "x6, x4, x4"),
        Q!("    adcs            " "x12, x12, x6"),
        Q!("    umulh           " "x6, x4, x4"),
        Q!("    adcs            " "x13, x13, x6"),
        Q!("    mul             " "x6, x5, x5"),
        Q!("    adcs            " "x14, x14, x6"),
        Q!("    umulh           " "x6, x5, x5"),
        Q!("    adc             " "x7, x7, x6"),
        Q!("    adds            " "x9, x9, x8, lsl #32"),
        Q!("    lsr             " "x3, x8, #32"),
        Q!("    adcs            " "x10, x10, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x8, x3"),
        Q!("    umulh           " "x8, x8, x3"),
        Q!("    adcs            " "x11, x11, x2"),
        Q!("    adc             " "x8, x8, xzr"),
        Q!("    adds            " "x10, x10, x9, lsl #32"),
        Q!("    lsr             " "x3, x9, #32"),
        Q!("    adcs            " "x11, x11, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x9, x3"),
        Q!("    umulh           " "x9, x9, x3"),
        Q!("    adcs            " "x8, x8, x2"),
        Q!("    adc             " "x9, x9, xzr"),
        Q!("    adds            " "x11, x11, x10, lsl #32"),
        Q!("    lsr             " "x3, x10, #32"),
        Q!("    adcs            " "x8, x8, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x10, x3"),
        Q!("    umulh           " "x10, x10, x3"),
        Q!("    adcs            " "x9, x9, x2"),
        Q!("    adc             " "x10, x10, xzr"),
        Q!("    adds            " "x8, x8, x11, lsl #32"),
        Q!("    lsr             " "x3, x11, #32"),
        Q!("    adcs            " "x9, x9, x3"),
        Q!("    mov             " "x3, #0xffffffff00000001"),
        Q!("    mul             " "x2, x11, x3"),
        Q!("    umulh           " "x11, x11, x3"),
        Q!("    adcs            " "x10, x10, x2"),
        Q!("    adc             " "x11, x11, xzr"),
        Q!("    adds            " "x8, x8, x12"),
        Q!("    adcs            " "x9, x9, x13"),
        Q!("    adcs            " "x10, x10, x14"),
        Q!("    adcs            " "x11, x11, x7"),
        Q!("    cset            " "x2, cs"),
        Q!("    mov             " "x3, #0xffffffff"),
        Q!("    mov             " "x5, #0xffffffff00000001"),
        Q!("    adds            " "x12, x8, #0x1"),
        Q!("    sbcs            " "x13, x9, x3"),
        Q!("    sbcs            " "x14, x10, xzr"),
        Q!("    sbcs            " "x7, x11, x5"),
        Q!("    sbcs            " "xzr, x2, xzr"),
        Q!("    csel            " "x8, x8, x12, cc"),
        Q!("    csel            " "x9, x9, x13, cc"),
        Q!("    csel            " "x10, x10, x14, cc"),
        Q!("    csel            " "x11, x11, x7, cc"),
        Q!("    stp             " "x8, x9, [sp]"),
        Q!("    stp             " "x10, x11, [sp, #16]"),
        Q!("    ldp             " "x3, x4, [sp, #96]"),
        Q!("    ldp             " "x7, x8, [x16]"),
        Q!("    mul             " "x12, x3, x7"),
        Q!("    umulh           " "x13, x3, x7"),
        Q!("    mul             " "x11, x3, x8"),
        Q!("    umulh           " "x14, x3, x8"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    ldp             " "x9, x10, [x16, #16]"),
        Q!("    mul             " "x11, x3, x9"),
        Q!("    umulh           " "x0, x3, x9"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x3, x10"),
        Q!("    umulh           " "x1, x3, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x1, x1, xzr"),
        Q!("    ldp             " "x5, x6, [sp, #112]"),
        Q!("    mul             " "x11, x4, x7"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    mul             " "x11, x4, x8"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x4, x9"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x4, x10"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x3, x4, x10"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    umulh           " "x11, x4, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    umulh           " "x11, x4, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x4, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    mul             " "x11, x5, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    mul             " "x11, x5, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x5, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x5, x10"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x4, x5, x10"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    umulh           " "x11, x5, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x5, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x5, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    mul             " "x11, x6, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    mul             " "x11, x6, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x6, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    mul             " "x11, x6, x10"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    umulh           " "x5, x6, x10"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    adds            " "x13, x13, x12, lsl #32"),
        Q!("    lsr             " "x11, x12, #32"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x12, x10"),
        Q!("    umulh           " "x12, x12, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x12, x12, xzr"),
        Q!("    umulh           " "x11, x6, x7"),
        Q!("    adds            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x6, x8"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x11, x6, x9"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    adds            " "x14, x14, x13, lsl #32"),
        Q!("    lsr             " "x11, x13, #32"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x13, x10"),
        Q!("    umulh           " "x13, x13, x10"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    adc             " "x13, x13, xzr"),
        Q!("    adds            " "x0, x0, x14, lsl #32"),
        Q!("    lsr             " "x11, x14, #32"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    mul             " "x11, x14, x10"),
        Q!("    umulh           " "x14, x14, x10"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x12, x12, x0, lsl #32"),
        Q!("    lsr             " "x11, x0, #32"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    mul             " "x11, x0, x10"),
        Q!("    umulh           " "x0, x0, x10"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    adc             " "x0, x0, xzr"),
        Q!("    adds            " "x12, x12, x1"),
        Q!("    adcs            " "x13, x13, x3"),
        Q!("    adcs            " "x14, x14, x4"),
        Q!("    adcs            " "x0, x0, x5"),
        Q!("    cset            " "x8, cs"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    adds            " "x1, x12, #0x1"),
        Q!("    sbcs            " "x3, x13, x11"),
        Q!("    sbcs            " "x4, x14, xzr"),
        Q!("    sbcs            " "x5, x0, x10"),
        Q!("    sbcs            " "xzr, x8, xzr"),
        Q!("    csel            " "x12, x12, x1, cc"),
        Q!("    csel            " "x13, x13, x3, cc"),
        Q!("    csel            " "x14, x14, x4, cc"),
        Q!("    csel            " "x0, x0, x5, cc"),
        Q!("    stp             " "x12, x13, [sp, #128]"),
        Q!("    stp             " "x14, x0, [sp, #144]"),
        Q!("    ldp             " "x3, x4, [sp, #96]"),
        Q!("    ldp             " "x7, x8, [sp, #64]"),
        Q!("    mul             " "x12, x3, x7"),
        Q!("    umulh           " "x13, x3, x7"),
        Q!("    mul             " "x11, x3, x8"),
        Q!("    umulh           " "x14, x3, x8"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    ldp             " "x9, x10, [sp, #80]"),
        Q!("    mul             " "x11, x3, x9"),
        Q!("    umulh           " "x0, x3, x9"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x3, x10"),
        Q!("    umulh           " "x1, x3, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x1, x1, xzr"),
        Q!("    ldp             " "x5, x6, [sp, #112]"),
        Q!("    mul             " "x11, x4, x7"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    mul             " "x11, x4, x8"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x4, x9"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x4, x10"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x3, x4, x10"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    umulh           " "x11, x4, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    umulh           " "x11, x4, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x4, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    mul             " "x11, x5, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    mul             " "x11, x5, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x5, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x5, x10"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x4, x5, x10"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    umulh           " "x11, x5, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x5, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x5, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    mul             " "x11, x6, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    mul             " "x11, x6, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x6, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    mul             " "x11, x6, x10"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    umulh           " "x5, x6, x10"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    adds            " "x13, x13, x12, lsl #32"),
        Q!("    lsr             " "x11, x12, #32"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x12, x10"),
        Q!("    umulh           " "x12, x12, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x12, x12, xzr"),
        Q!("    umulh           " "x11, x6, x7"),
        Q!("    adds            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x6, x8"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x11, x6, x9"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    adds            " "x14, x14, x13, lsl #32"),
        Q!("    lsr             " "x11, x13, #32"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x13, x10"),
        Q!("    umulh           " "x13, x13, x10"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    adc             " "x13, x13, xzr"),
        Q!("    adds            " "x0, x0, x14, lsl #32"),
        Q!("    lsr             " "x11, x14, #32"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    mul             " "x11, x14, x10"),
        Q!("    umulh           " "x14, x14, x10"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x12, x12, x0, lsl #32"),
        Q!("    lsr             " "x11, x0, #32"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    mul             " "x11, x0, x10"),
        Q!("    umulh           " "x0, x0, x10"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    adc             " "x0, x0, xzr"),
        Q!("    adds            " "x12, x12, x1"),
        Q!("    adcs            " "x13, x13, x3"),
        Q!("    adcs            " "x14, x14, x4"),
        Q!("    adcs            " "x0, x0, x5"),
        Q!("    cset            " "x8, cs"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    adds            " "x1, x12, #0x1"),
        Q!("    sbcs            " "x3, x13, x11"),
        Q!("    sbcs            " "x4, x14, xzr"),
        Q!("    sbcs            " "x5, x0, x10"),
        Q!("    sbcs            " "xzr, x8, xzr"),
        Q!("    csel            " "x12, x12, x1, cc"),
        Q!("    csel            " "x13, x13, x3, cc"),
        Q!("    csel            " "x14, x14, x4, cc"),
        Q!("    csel            " "x0, x0, x5, cc"),
        Q!("    stp             " "x12, x13, [sp, #64]"),
        Q!("    stp             " "x14, x0, [sp, #80]"),
        Q!("    ldp             " "x5, x6, [sp]"),
        Q!("    ldp             " "x4, x3, [sp, #128]"),
        Q!("    subs            " "x5, x5, x4"),
        Q!("    sbcs            " "x6, x6, x3"),
        Q!("    ldp             " "x7, x8, [sp, #16]"),
        Q!("    ldp             " "x4, x3, [sp, #144]"),
        Q!("    sbcs            " "x7, x7, x4"),
        Q!("    sbcs            " "x8, x8, x3"),
        Q!("    csetm           " "x3, cc"),
        Q!("    adds            " "x5, x5, x3"),
        Q!("    mov             " "x4, #0xffffffff"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adcs            " "x6, x6, x4"),
        Q!("    adcs            " "x7, x7, xzr"),
        Q!("    mov             " "x4, #0xffffffff00000001"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adc             " "x8, x8, x4"),
        Q!("    stp             " "x5, x6, [sp]"),
        Q!("    stp             " "x7, x8, [sp, #16]"),
        Q!("    ldp             " "x5, x6, [sp, #64]"),
        Q!("    ldp             " "x4, x3, [sp, #128]"),
        Q!("    subs            " "x5, x5, x4"),
        Q!("    sbcs            " "x6, x6, x3"),
        Q!("    ldp             " "x7, x8, [sp, #80]"),
        Q!("    ldp             " "x4, x3, [sp, #144]"),
        Q!("    sbcs            " "x7, x7, x4"),
        Q!("    sbcs            " "x8, x8, x3"),
        Q!("    csetm           " "x3, cc"),
        Q!("    adds            " "x5, x5, x3"),
        Q!("    mov             " "x4, #0xffffffff"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adcs            " "x6, x6, x4"),
        Q!("    adcs            " "x7, x7, xzr"),
        Q!("    mov             " "x4, #0xffffffff00000001"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adc             " "x8, x8, x4"),
        Q!("    stp             " "x5, x6, [sp, #96]"),
        Q!("    stp             " "x7, x8, [sp, #112]"),
        Q!("    ldp             " "x3, x4, [sp, #160]"),
        Q!("    ldp             " "x7, x8, [x16, #64]"),
        Q!("    mul             " "x12, x3, x7"),
        Q!("    umulh           " "x13, x3, x7"),
        Q!("    mul             " "x11, x3, x8"),
        Q!("    umulh           " "x14, x3, x8"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    ldp             " "x9, x10, [x16, #80]"),
        Q!("    mul             " "x11, x3, x9"),
        Q!("    umulh           " "x0, x3, x9"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x3, x10"),
        Q!("    umulh           " "x1, x3, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x1, x1, xzr"),
        Q!("    ldp             " "x5, x6, [sp, #176]"),
        Q!("    mul             " "x11, x4, x7"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    mul             " "x11, x4, x8"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x4, x9"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x4, x10"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x3, x4, x10"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    umulh           " "x11, x4, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    umulh           " "x11, x4, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x4, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    mul             " "x11, x5, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    mul             " "x11, x5, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x5, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x5, x10"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x4, x5, x10"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    umulh           " "x11, x5, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x5, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x5, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    mul             " "x11, x6, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    mul             " "x11, x6, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x6, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    mul             " "x11, x6, x10"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    umulh           " "x5, x6, x10"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    adds            " "x13, x13, x12, lsl #32"),
        Q!("    lsr             " "x11, x12, #32"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x12, x10"),
        Q!("    umulh           " "x12, x12, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x12, x12, xzr"),
        Q!("    umulh           " "x11, x6, x7"),
        Q!("    adds            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x6, x8"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x11, x6, x9"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    adds            " "x14, x14, x13, lsl #32"),
        Q!("    lsr             " "x11, x13, #32"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x13, x10"),
        Q!("    umulh           " "x13, x13, x10"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    adc             " "x13, x13, xzr"),
        Q!("    adds            " "x0, x0, x14, lsl #32"),
        Q!("    lsr             " "x11, x14, #32"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    mul             " "x11, x14, x10"),
        Q!("    umulh           " "x14, x14, x10"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x12, x12, x0, lsl #32"),
        Q!("    lsr             " "x11, x0, #32"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    mul             " "x11, x0, x10"),
        Q!("    umulh           " "x0, x0, x10"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    adc             " "x0, x0, xzr"),
        Q!("    adds            " "x12, x12, x1"),
        Q!("    adcs            " "x13, x13, x3"),
        Q!("    adcs            " "x14, x14, x4"),
        Q!("    adcs            " "x0, x0, x5"),
        Q!("    cset            " "x8, cs"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    adds            " "x1, x12, #0x1"),
        Q!("    sbcs            " "x3, x13, x11"),
        Q!("    sbcs            " "x4, x14, xzr"),
        Q!("    sbcs            " "x5, x0, x10"),
        Q!("    sbcs            " "xzr, x8, xzr"),
        Q!("    csel            " "x12, x12, x1, cc"),
        Q!("    csel            " "x13, x13, x3, cc"),
        Q!("    csel            " "x14, x14, x4, cc"),
        Q!("    csel            " "x0, x0, x5, cc"),
        Q!("    stp             " "x12, x13, [sp, #160]"),
        Q!("    stp             " "x14, x0, [sp, #176]"),
        Q!("    ldp             " "x5, x6, [sp]"),
        Q!("    ldp             " "x4, x3, [sp, #64]"),
        Q!("    subs            " "x5, x5, x4"),
        Q!("    sbcs            " "x6, x6, x3"),
        Q!("    ldp             " "x7, x8, [sp, #16]"),
        Q!("    ldp             " "x4, x3, [sp, #80]"),
        Q!("    sbcs            " "x7, x7, x4"),
        Q!("    sbcs            " "x8, x8, x3"),
        Q!("    csetm           " "x3, cc"),
        Q!("    adds            " "x5, x5, x3"),
        Q!("    mov             " "x4, #0xffffffff"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adcs            " "x6, x6, x4"),
        Q!("    adcs            " "x7, x7, xzr"),
        Q!("    mov             " "x4, #0xffffffff00000001"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adc             " "x8, x8, x4"),
        Q!("    stp             " "x5, x6, [sp]"),
        Q!("    stp             " "x7, x8, [sp, #16]"),
        Q!("    ldp             " "x5, x6, [sp, #128]"),
        Q!("    ldp             " "x4, x3, [sp]"),
        Q!("    subs            " "x5, x5, x4"),
        Q!("    sbcs            " "x6, x6, x3"),
        Q!("    ldp             " "x7, x8, [sp, #144]"),
        Q!("    ldp             " "x4, x3, [sp, #16]"),
        Q!("    sbcs            " "x7, x7, x4"),
        Q!("    sbcs            " "x8, x8, x3"),
        Q!("    csetm           " "x3, cc"),
        Q!("    adds            " "x5, x5, x3"),
        Q!("    mov             " "x4, #0xffffffff"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adcs            " "x6, x6, x4"),
        Q!("    adcs            " "x7, x7, xzr"),
        Q!("    mov             " "x4, #0xffffffff00000001"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adc             " "x8, x8, x4"),
        Q!("    stp             " "x5, x6, [sp, #128]"),
        Q!("    stp             " "x7, x8, [sp, #144]"),
        Q!("    ldp             " "x3, x4, [sp, #96]"),
        Q!("    ldp             " "x7, x8, [x16, #32]"),
        Q!("    mul             " "x12, x3, x7"),
        Q!("    umulh           " "x13, x3, x7"),
        Q!("    mul             " "x11, x3, x8"),
        Q!("    umulh           " "x14, x3, x8"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    ldp             " "x9, x10, [x16, #48]"),
        Q!("    mul             " "x11, x3, x9"),
        Q!("    umulh           " "x0, x3, x9"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x3, x10"),
        Q!("    umulh           " "x1, x3, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x1, x1, xzr"),
        Q!("    ldp             " "x5, x6, [sp, #112]"),
        Q!("    mul             " "x11, x4, x7"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    mul             " "x11, x4, x8"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x4, x9"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x4, x10"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x3, x4, x10"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    umulh           " "x11, x4, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    umulh           " "x11, x4, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x4, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    mul             " "x11, x5, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    mul             " "x11, x5, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x5, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x5, x10"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x4, x5, x10"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    umulh           " "x11, x5, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x5, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x5, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    mul             " "x11, x6, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    mul             " "x11, x6, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x6, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    mul             " "x11, x6, x10"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    umulh           " "x5, x6, x10"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    adds            " "x13, x13, x12, lsl #32"),
        Q!("    lsr             " "x11, x12, #32"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x12, x10"),
        Q!("    umulh           " "x12, x12, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x12, x12, xzr"),
        Q!("    umulh           " "x11, x6, x7"),
        Q!("    adds            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x6, x8"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x11, x6, x9"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    adds            " "x14, x14, x13, lsl #32"),
        Q!("    lsr             " "x11, x13, #32"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x13, x10"),
        Q!("    umulh           " "x13, x13, x10"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    adc             " "x13, x13, xzr"),
        Q!("    adds            " "x0, x0, x14, lsl #32"),
        Q!("    lsr             " "x11, x14, #32"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    mul             " "x11, x14, x10"),
        Q!("    umulh           " "x14, x14, x10"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x12, x12, x0, lsl #32"),
        Q!("    lsr             " "x11, x0, #32"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    mul             " "x11, x0, x10"),
        Q!("    umulh           " "x0, x0, x10"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    adc             " "x0, x0, xzr"),
        Q!("    adds            " "x12, x12, x1"),
        Q!("    adcs            " "x13, x13, x3"),
        Q!("    adcs            " "x14, x14, x4"),
        Q!("    adcs            " "x0, x0, x5"),
        Q!("    cset            " "x8, cs"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    adds            " "x1, x12, #0x1"),
        Q!("    sbcs            " "x3, x13, x11"),
        Q!("    sbcs            " "x4, x14, xzr"),
        Q!("    sbcs            " "x5, x0, x10"),
        Q!("    sbcs            " "xzr, x8, xzr"),
        Q!("    csel            " "x12, x12, x1, cc"),
        Q!("    csel            " "x13, x13, x3, cc"),
        Q!("    csel            " "x14, x14, x4, cc"),
        Q!("    csel            " "x0, x0, x5, cc"),
        Q!("    stp             " "x12, x13, [sp, #96]"),
        Q!("    stp             " "x14, x0, [sp, #112]"),
        Q!("    ldp             " "x3, x4, [sp, #32]"),
        Q!("    ldp             " "x7, x8, [sp, #128]"),
        Q!("    mul             " "x12, x3, x7"),
        Q!("    umulh           " "x13, x3, x7"),
        Q!("    mul             " "x11, x3, x8"),
        Q!("    umulh           " "x14, x3, x8"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    ldp             " "x9, x10, [sp, #144]"),
        Q!("    mul             " "x11, x3, x9"),
        Q!("    umulh           " "x0, x3, x9"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x3, x10"),
        Q!("    umulh           " "x1, x3, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x1, x1, xzr"),
        Q!("    ldp             " "x5, x6, [sp, #48]"),
        Q!("    mul             " "x11, x4, x7"),
        Q!("    adds            " "x13, x13, x11"),
        Q!("    mul             " "x11, x4, x8"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x4, x9"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x4, x10"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x3, x4, x10"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    umulh           " "x11, x4, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    umulh           " "x11, x4, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x4, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    adc             " "x3, x3, xzr"),
        Q!("    mul             " "x11, x5, x7"),
        Q!("    adds            " "x14, x14, x11"),
        Q!("    mul             " "x11, x5, x8"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x5, x9"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x5, x10"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x4, x5, x10"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    umulh           " "x11, x5, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    umulh           " "x11, x5, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x5, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    adc             " "x4, x4, xzr"),
        Q!("    mul             " "x11, x6, x7"),
        Q!("    adds            " "x0, x0, x11"),
        Q!("    mul             " "x11, x6, x8"),
        Q!("    adcs            " "x1, x1, x11"),
        Q!("    mul             " "x11, x6, x9"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    mul             " "x11, x6, x10"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    umulh           " "x5, x6, x10"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    mov             " "x10, #0xffffffff00000001"),
        Q!("    adds            " "x13, x13, x12, lsl #32"),
        Q!("    lsr             " "x11, x12, #32"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    mul             " "x11, x12, x10"),
        Q!("    umulh           " "x12, x12, x10"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    adc             " "x12, x12, xzr"),
        Q!("    umulh           " "x11, x6, x7"),
        Q!("    adds            " "x1, x1, x11"),
        Q!("    umulh           " "x11, x6, x8"),
        Q!("    adcs            " "x3, x3, x11"),
        Q!("    umulh           " "x11, x6, x9"),
        Q!("    adcs            " "x4, x4, x11"),
        Q!("    adc             " "x5, x5, xzr"),
        Q!("    adds            " "x14, x14, x13, lsl #32"),
        Q!("    lsr             " "x11, x13, #32"),
        Q!("    adcs            " "x0, x0, x11"),
        Q!("    mul             " "x11, x13, x10"),
        Q!("    umulh           " "x13, x13, x10"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    adc             " "x13, x13, xzr"),
        Q!("    adds            " "x0, x0, x14, lsl #32"),
        Q!("    lsr             " "x11, x14, #32"),
        Q!("    adcs            " "x12, x12, x11"),
        Q!("    mul             " "x11, x14, x10"),
        Q!("    umulh           " "x14, x14, x10"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    adc             " "x14, x14, xzr"),
        Q!("    adds            " "x12, x12, x0, lsl #32"),
        Q!("    lsr             " "x11, x0, #32"),
        Q!("    adcs            " "x13, x13, x11"),
        Q!("    mul             " "x11, x0, x10"),
        Q!("    umulh           " "x0, x0, x10"),
        Q!("    adcs            " "x14, x14, x11"),
        Q!("    adc             " "x0, x0, xzr"),
        Q!("    adds            " "x12, x12, x1"),
        Q!("    adcs            " "x13, x13, x3"),
        Q!("    adcs            " "x14, x14, x4"),
        Q!("    adcs            " "x0, x0, x5"),
        Q!("    cset            " "x8, cs"),
        Q!("    mov             " "x11, #0xffffffff"),
        Q!("    adds            " "x1, x12, #0x1"),
        Q!("    sbcs            " "x3, x13, x11"),
        Q!("    sbcs            " "x4, x14, xzr"),
        Q!("    sbcs            " "x5, x0, x10"),
        Q!("    sbcs            " "xzr, x8, xzr"),
        Q!("    csel            " "x12, x12, x1, cc"),
        Q!("    csel            " "x13, x13, x3, cc"),
        Q!("    csel            " "x14, x14, x4, cc"),
        Q!("    csel            " "x0, x0, x5, cc"),
        Q!("    stp             " "x12, x13, [sp, #128]"),
        Q!("    stp             " "x14, x0, [sp, #144]"),
        Q!("    ldp             " "x5, x6, [sp, #128]"),
        Q!("    ldp             " "x4, x3, [sp, #96]"),
        Q!("    subs            " "x5, x5, x4"),
        Q!("    sbcs            " "x6, x6, x3"),
        Q!("    ldp             " "x7, x8, [sp, #144]"),
        Q!("    ldp             " "x4, x3, [sp, #112]"),
        Q!("    sbcs            " "x7, x7, x4"),
        Q!("    sbcs            " "x8, x8, x3"),
        Q!("    csetm           " "x3, cc"),
        Q!("    adds            " "x5, x5, x3"),
        Q!("    mov             " "x4, #0xffffffff"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adcs            " "x6, x6, x4"),
        Q!("    adcs            " "x7, x7, xzr"),
        Q!("    mov             " "x4, #0xffffffff00000001"),
        Q!("    and             " "x4, x4, x3"),
        Q!("    adc             " "x8, x8, x4"),
        Q!("    stp             " "x5, x6, [sp, #128]"),
        Q!("    stp             " "x7, x8, [sp, #144]"),
        Q!("    ldp             " "x0, x1, [x16, #64]"),
        Q!("    ldp             " "x2, x3, [x16, #80]"),
        Q!("    orr             " "x4, x0, x1"),
        Q!("    orr             " "x5, x2, x3"),
        Q!("    orr             " "x4, x4, x5"),
        Q!("    cmp             " "x4, xzr"),
        Q!("    ldp             " "x0, x1, [sp]"),
        Q!("    ldp             " "x12, x13, [x17]"),
        Q!("    csel            " "x0, x0, x12, ne"),
        Q!("    csel            " "x1, x1, x13, ne"),
        Q!("    ldp             " "x2, x3, [sp, #16]"),
        Q!("    ldp             " "x12, x13, [x17, #16]"),
        Q!("    csel            " "x2, x2, x12, ne"),
        Q!("    csel            " "x3, x3, x13, ne"),
        Q!("    ldp             " "x4, x5, [sp, #128]"),
        Q!("    ldp             " "x12, x13, [x17, #32]"),
        Q!("    csel            " "x4, x4, x12, ne"),
        Q!("    csel            " "x5, x5, x13, ne"),
        Q!("    ldp             " "x6, x7, [sp, #144]"),
        Q!("    ldp             " "x12, x13, [x17, #48]"),
        Q!("    csel            " "x6, x6, x12, ne"),
        Q!("    csel            " "x7, x7, x13, ne"),
        Q!("    ldp             " "x8, x9, [sp, #160]"),
        Q!("    mov             " "x12, #0x1"),
        Q!("    mov             " "x13, #0xffffffff00000000"),
        Q!("    csel            " "x8, x8, x12, ne"),
        Q!("    csel            " "x9, x9, x13, ne"),
        Q!("    ldp             " "x10, x11, [sp, #176]"),
        Q!("    mov             " "x12, #0xffffffffffffffff"),
        Q!("    mov             " "x13, #0xfffffffe"),
        Q!("    csel            " "x10, x10, x12, ne"),
        Q!("    csel            " "x11, x11, x13, ne"),
        Q!("    stp             " "x0, x1, [x15]"),
        Q!("    stp             " "x2, x3, [x15, #16]"),
        Q!("    stp             " "x4, x5, [x15, #32]"),
        Q!("    stp             " "x6, x7, [x15, #48]"),
        Q!("    stp             " "x8, x9, [x15, #64]"),
        Q!("    stp             " "x10, x11, [x15, #80]"),
        Q!("    add             " "sp, sp, # (192 + 0)"),
        Q!("    ret             " ),
        Q!(Label!("hoist_finish", 9) ":"),
        inout("x0") res.as_mut_ptr() => _,
        inout("x1") scalar.as_ptr() => _,
        inout("x2") 6u64 => _,
        inout("x3") table.as_ptr() => _,
        // clobbers
        out("x10") _,
        out("x11") _,
        out("x12") _,
        out("x13") _,
        out("x14") _,
        out("x15") _,
        out("x16") _,
        out("x17") _,
        out("x20") _,
        out("x21") _,
        out("x22") _,
        out("x23") _,
        out("x24") _,
        out("x25") _,
        out("x30") _,
        out("x4") _,
        out("x5") _,
        out("x6") _,
        out("x7") _,
        out("x8") _,
        out("x9") _,
            )
    };
}