packages feed

crypton-2.1.2: cbits/s2n/x86_att/p256_scalarmulbase_alt.S

// 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 x86-64 ABI: RDI = res, RSI = scalar, RDX = blocksize, RCX = table
// Microsoft x64 ABI:   RCX = res, RDX = scalar, R8 = blocksize, R9 = table
// ----------------------------------------------------------------------------

#include "_internal_s2n_bignum_x86_att.h"


        S2N_BN_SYM_VISIBILITY_DIRECTIVE(p256_scalarmulbase_alt)
        S2N_BN_FUNCTION_TYPE_DIRECTIVE(p256_scalarmulbase_alt)
        S2N_BN_SYM_PRIVACY_DIRECTIVE(p256_scalarmulbase_alt)


        .text
        .balign 32

// Size of individual field elements

#define NUMSIZE 32

// 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

#define RSCALAR (0*NUMSIZE)
#define ACC (1*NUMSIZE)
#define NACC (4*NUMSIZE)
#define TABENT (7*NUMSIZE)
#define Z2 (4*NUMSIZE)
#define Z3 (5*NUMSIZE)

#define rscalar RSCALAR(%rsp)
#define acc ACC(%rsp)
#define nacc NACC(%rsp)
#define tabent TABENT(%rsp)

#define z2 Z2(%rsp)
#define z3 Z3(%rsp)

#define res (9*NUMSIZE)(%rsp)
#define blocksize (9*NUMSIZE+8)(%rsp)
#define table (9*NUMSIZE+16)(%rsp)
#define i (9*NUMSIZE+24)(%rsp)
#define bf (9*NUMSIZE+32)(%rsp)
#define cf (9*NUMSIZE+40)(%rsp)
#define j (9*NUMSIZE+48)(%rsp)

#define NSPACE 11*NUMSIZE

S2N_BN_SYMBOL(p256_scalarmulbase_alt):
        CFI_START
        _CET_ENDBR

// The Windows version literally calls the standard ABI version.
// This simplifies the proofs since subroutine offsets are fixed.

#if WINDOWS_ABI
        CFI_PUSH(%rdi)
        CFI_PUSH(%rsi)
        movq    %rcx, %rdi
        movq    %rdx, %rsi
        movq    %r8, %rdx
        movq    %r9, %rcx
        CFI_CALL(Lp256_scalarmulbase_alt_standard)
        CFI_POP(%rsi)
        CFI_POP(%rdi)
        CFI_RET

S2N_BN_SIZE_DIRECTIVE(p256_scalarmulbase_alt)

S2N_BN_FUNCTION_TYPE_DIRECTIVE(Lp256_scalarmulbase_alt_standard)

Lp256_scalarmulbase_alt_standard:
        CFI_START
#endif

// Real start of the standard ABI code.

        CFI_PUSH(%r15)
        CFI_PUSH(%r14)
        CFI_PUSH(%r13)
        CFI_PUSH(%r12)
        CFI_PUSH(%rbp)
        CFI_PUSH(%rbx)

        CFI_DEC_RSP(NSPACE)

// 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.

        movq    %rdi, res
        movq    %rdx, blocksize
        movq    %rcx, table

// Load the digits of group order n_256 = [%r15;%r14;%r13;%r12]

        movq    $0xf3b9cac2fc632551, %r12
        movq    $0xbce6faada7179e84, %r13
        movq    $0xffffffffffffffff, %r14
        movq    $0xffffffff00000000, %r15

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

        movq    (%rsi), %r8
        subq    %r12, %r8
        movq    8(%rsi), %r9
        sbbq    %r13, %r9
        movq    16(%rsi), %r10
        sbbq    %r14, %r10
        movq    24(%rsi), %r11
        sbbq    %r15, %r11

        cmovcq  (%rsi), %r8
        cmovcq  8(%rsi), %r9
        cmovcq  16(%rsi), %r10
        cmovcq  24(%rsi), %r11

        movq    %r8, RSCALAR(%rsp)
        movq    %r9, RSCALAR+8(%rsp)
        movq    %r10, RSCALAR+16(%rsp)
        movq    %r11, RSCALAR+24(%rsp)

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

        xorl    %eax, %eax

        movq    %rax, ACC(%rsp)
        movq    %rax, ACC+8(%rsp)
        movq    %rax, ACC+16(%rsp)
        movq    %rax, ACC+24(%rsp)
        movq    %rax, ACC+32(%rsp)
        movq    %rax, ACC+40(%rsp)
        movq    %rax, ACC+48(%rsp)
        movq    %rax, ACC+56(%rsp)
        movq    %rax, ACC+64(%rsp)
        movq    %rax, ACC+72(%rsp)
        movq    %rax, ACC+80(%rsp)
        movq    %rax, ACC+88(%rsp)

        movq    %rax, cf

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

        movq    %rax, i

Lp256_scalarmulbase_alt_loop:

// 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).

        movq    RSCALAR(%rsp), %r8
        movq    RSCALAR+8(%rsp), %r9
        movq    RSCALAR+16(%rsp), %r10
        movq    RSCALAR+24(%rsp), %r11

        movq    blocksize, %rcx
        movl    $1, %eax
        shlq    %cl, %rax
        decq    %rax
        andq    %r8, %rax

        shrdq   %cl, %r9, %r8
        shrdq   %cl, %r10, %r9
        shrdq   %cl, %r11, %r10
        shrq    %cl, %r11

        addq    cf, %rax
        movq    %rax, bf

        movq   %r8, RSCALAR(%rsp)
        movq   %r9, RSCALAR+8(%rsp)
        movq   %r10, RSCALAR+16(%rsp)
        movq   %r11, RSCALAR+24(%rsp)

// 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.

        movl    $1, %eax
        movq    blocksize, %rcx
        shlq    %cl, %rax
        movq    %rax, %rbx
        shrq    $1, %rax

        subq    bf, %rbx
        cmpq    bf, %rax

        cmovncq bf, %rbx
        sbbq    %rax, %rax
        movq    %rbx, j
        negq    %rax
        movq    %rax, cf

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

        movq    blocksize, %rcx
        decq    %rcx
        movl    $1, %esi
        shlq    %cl, %rsi
        movq    j, %r12
        movq    table, %rbp

Lp256_scalarmulbase_alt_tabloop:
        subq    $1, %r12
        cmovzq  (%rbp), %rax
        cmovzq  8(%rbp), %rbx
        cmovzq  16(%rbp), %rcx
        cmovzq  24(%rbp), %rdx
        cmovzq  32(%rbp), %r8
        cmovzq  40(%rbp), %r9
        cmovzq  48(%rbp), %r10
        cmovzq  56(%rbp), %r11

        addq    $64, %rbp
        decq    %rsi
        jnz     Lp256_scalarmulbase_alt_tabloop

        movq    %rbp, table

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

        xorl    %r14d, %r14d
        leaq    -1(%r14), %r12
        movq    $0x00000000ffffffff, %r15
        movq    %r15, %r13
        negq    %r15

        subq    %r8, %r12
        sbbq    %r9, %r13
        sbbq    %r10, %r14
        sbbq    %r11, %r15

        movq    %rax, TABENT(%rsp)
        movq    %rbx, TABENT+8(%rsp)
        movq    %rcx, TABENT+16(%rsp)
        movq    %rdx, TABENT+24(%rsp)

        movq    cf, %rax
        testq   %rax, %rax
        cmovnzq %r12, %r8
        cmovnzq %r13, %r9
        cmovnzq %r14, %r10
        cmovnzq %r15, %r11

        movq    %r8, TABENT+32(%rsp)
        movq    %r9, TABENT+40(%rsp)
        movq    %r10, TABENT+48(%rsp)
        movq    %r11, TABENT+56(%rsp)

// Add the adjusted table point to the accumulator

        leaq    NACC(%rsp), %rdi
        leaq    ACC(%rsp), %rsi
        leaq    TABENT(%rsp), %rdx
        CFI_CALL(Lp256_scalarmulbase_alt_local_p256_montjmixadd)

// 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.

        movq    j, %rax
        testq   %rax, %rax

        movq    ACC(%rsp), %rax
        cmovnzq NACC(%rsp), %rax
        movq    %rax, ACC(%rsp)

        movq    ACC+8(%rsp), %rax
        cmovnzq NACC+8(%rsp), %rax
        movq    %rax, ACC+8(%rsp)

        movq    ACC+16(%rsp), %rax
        cmovnzq NACC+16(%rsp), %rax
        movq    %rax, ACC+16(%rsp)

        movq    ACC+24(%rsp), %rax
        cmovnzq NACC+24(%rsp), %rax
        movq    %rax, ACC+24(%rsp)

        movq    ACC+32(%rsp), %rax
        cmovnzq NACC+32(%rsp), %rax
        movq    %rax, ACC+32(%rsp)

        movq    ACC+40(%rsp), %rax
        cmovnzq NACC+40(%rsp), %rax
        movq    %rax, ACC+40(%rsp)

        movq    ACC+48(%rsp), %rax
        cmovnzq NACC+48(%rsp), %rax
        movq    %rax, ACC+48(%rsp)

        movq    ACC+56(%rsp), %rax
        cmovnzq NACC+56(%rsp), %rax
        movq    %rax, ACC+56(%rsp)

        movq    ACC+64(%rsp), %rax
        cmovnzq NACC+64(%rsp), %rax
        movq    %rax, ACC+64(%rsp)

        movq    ACC+72(%rsp), %rax
        cmovnzq NACC+72(%rsp), %rax
        movq    %rax, ACC+72(%rsp)

        movq    ACC+80(%rsp), %rax
        cmovnzq NACC+80(%rsp), %rax
        movq    %rax, ACC+80(%rsp)

        movq    ACC+88(%rsp), %rax
        cmovnzq NACC+88(%rsp), %rax
        movq    %rax, ACC+88(%rsp)

// Loop while blocksize * i <= 256

        movq    i, %rax
        incq    %rax
        movq    %rax, i

        imulq   blocksize, %rax
        cmpq    $257, %rax
        jc      Lp256_scalarmulbase_alt_loop

// 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

        leaq    Z2(%rsp), %rdi
        leaq    ACC+64(%rsp), %rsi
        CFI_CALL(Lp256_scalarmulbase_alt_local_montsqr_p256)

        leaq    Z3(%rsp), %rdi
        leaq    ACC+64(%rsp), %rsi
        leaq    Z2(%rsp), %rdx
        CFI_CALL(Lp256_scalarmulbase_alt_local_montmul_p256)

        leaq    Z2(%rsp), %rdi
        leaq    Z3(%rsp), %rsi
        CFI_CALL(Lp256_scalarmulbase_alt_local_demont_p256)

        leaq    Z3(%rsp), %rdi
        leaq    Z2(%rsp), %rsi
        CFI_CALL(Lp256_scalarmulbase_alt_local_inv_p256)

        leaq    Z2(%rsp), %rdi
        leaq    ACC+64(%rsp), %rsi
        leaq    Z3(%rsp), %rdx
        CFI_CALL(Lp256_scalarmulbase_alt_local_montmul_p256)

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

        movq    res, %rdi
        leaq    ACC(%rsp), %rsi
        leaq    Z2(%rsp), %rdx
        movq    %rdi, %rbx
        CFI_CALL(Lp256_scalarmulbase_alt_local_montmul_p256)

        leaq    32(%rbx), %rdi
        leaq    ACC+32(%rsp), %rsi
        leaq    Z3(%rsp), %rdx
        CFI_CALL(Lp256_scalarmulbase_alt_local_montmul_p256)

// Restore stack and registers and return

        CFI_INC_RSP(NSPACE)
        CFI_POP(%rbx)
        CFI_POP(%rbp)
        CFI_POP(%r12)
        CFI_POP(%r13)
        CFI_POP(%r14)
        CFI_POP(%r15)
        CFI_RET

#if WINDOWS_ABI
S2N_BN_SIZE_DIRECTIVE(Lp256_scalarmulbase_alt_standard)
#else
S2N_BN_SIZE_DIRECTIVE(p256_scalarmulbase_alt)
#endif

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

S2N_BN_FUNCTION_TYPE_DIRECTIVE(Lp256_scalarmulbase_alt_local_demont_p256)

Lp256_scalarmulbase_alt_local_demont_p256:
        CFI_START
        movq    (%rsi), %r8
        movq    0x8(%rsi), %r9
        movq    0x10(%rsi), %r10
        movq    0x18(%rsi), %r11
        movabsq $0x100000000, %rcx
        movq    %r8, %rax
        mulq    %rcx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %rsi, %rsi
        movq    %r9, %rax
        mulq    %rcx
        subq    %rsi, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %rsi, %rsi
        negq    %rcx
        negq    %rsi
        incq    %rcx
        movq    %r8, %rax
        mulq    %rcx
        addq    %rax, %r11
        adcq    %rdx, %rsi
        sbbq    %r8, %r8
        negq    %r8
        movq    %r9, %rax
        mulq    %rcx
        addq    %rax, %rsi
        adcq    %rdx, %r8
        negq    %rcx
        incq    %rcx
        movq    %r10, %rax
        mulq    %rcx
        addq    %rax, %r11
        adcq    %rdx, %rsi
        sbbq    %r9, %r9
        movq    %r11, %rax
        mulq    %rcx
        subq    %r9, %rdx
        addq    %rax, %rsi
        adcq    %rdx, %r8
        sbbq    %r9, %r9
        negq    %rcx
        negq    %r9
        incq    %rcx
        movq    %r10, %rax
        mulq    %rcx
        addq    %rax, %r8
        adcq    %rdx, %r9
        sbbq    %r10, %r10
        negq    %r10
        movq    %r11, %rax
        mulq    %rcx
        addq    %rax, %r9
        adcq    %rdx, %r10
        movq    %rsi, (%rdi)
        movq    %r8, 0x8(%rdi)
        movq    %r9, 0x10(%rdi)
        movq    %r10, 0x18(%rdi)
        CFI_RET

S2N_BN_SIZE_DIRECTIVE(Lp256_scalarmulbase_alt_local_demont_p256)

S2N_BN_FUNCTION_TYPE_DIRECTIVE(Lp256_scalarmulbase_alt_local_inv_p256)

Lp256_scalarmulbase_alt_local_inv_p256:
        CFI_START
        CFI_PUSH(%rbx)
        CFI_PUSH(%rbp)
        CFI_PUSH(%r12)
        CFI_PUSH(%r13)
        CFI_PUSH(%r14)
        CFI_PUSH(%r15)
        CFI_DEC_RSP(240)
        movq    %rdi, 0xe0(%rsp)
        xorl    %ecx, %ecx
        movl    $0xffffffff, %edx
        movq    %rdx, %rbx
        leaq    -0x1(%rcx), %rax
        negq    %rdx
        movq    %rax, (%rsp)
        movq    %rbx, 0x8(%rsp)
        movq    %rcx, 0x10(%rsp)
        movq    %rdx, 0x18(%rsp)
        movq    %rcx, 0x20(%rsp)
        movq    (%rsi), %r8
        movq    0x8(%rsi), %r9
        movq    0x10(%rsi), %r10
        movq    0x18(%rsi), %r11
        leaq    0x1(%rcx), %rax
        addq    %r8, %rax
        leaq    -0x1(%rdx), %rbx
        adcq    %r9, %rbx
        notq    %rcx
        adcq    %r10, %rcx
        notq    %rdx
        adcq    %r11, %rdx
        cmovaeq %r8, %rax
        cmovaeq %r9, %rbx
        cmovaeq %r10, %rcx
        cmovaeq %r11, %rdx
        movq    %rax, 0x28(%rsp)
        movq    %rbx, 0x30(%rsp)
        movq    %rcx, 0x38(%rsp)
        movq    %rdx, 0x40(%rsp)
        xorl    %eax, %eax
        movq    %rax, 0x48(%rsp)
        xorl    %eax, %eax
        movq    %rax, 0x50(%rsp)
        movq    %rax, 0x58(%rsp)
        movq    %rax, 0x60(%rsp)
        movq    %rax, 0x68(%rsp)
        movabsq $0x4000000000000, %rcx
        movq    %rcx, 0x78(%rsp)
        movq    %rax, 0x80(%rsp)
        movq    %rax, 0x88(%rsp)
        movq    %rax, 0x90(%rsp)
        movq    $0xa,  0xb0(%rsp)
        movq    $0x1,  0xb8(%rsp)
        jmp     Lp256_scalarmulbase_alt_inv_midloop
Lp256_scalarmulbase_alt_inv_loop:
        movq    %r8, %r9
        sarq    $0x3f, %r9
        xorq    %r9, %r8
        subq    %r9, %r8
        movq    %r10, %r11
        sarq    $0x3f, %r11
        xorq    %r11, %r10
        subq    %r11, %r10
        movq    %r12, %r13
        sarq    $0x3f, %r13
        xorq    %r13, %r12
        subq    %r13, %r12
        movq    %r14, %r15
        sarq    $0x3f, %r15
        xorq    %r15, %r14
        subq    %r15, %r14
        movq    %r8, %rax
        andq    %r9, %rax
        movq    %r10, %rdi
        andq    %r11, %rdi
        addq    %rax, %rdi
        movq    %rdi, 0xa0(%rsp)
        movq    %r12, %rax
        andq    %r13, %rax
        movq    %r14, %rsi
        andq    %r15, %rsi
        addq    %rax, %rsi
        movq    %rsi, 0xa8(%rsp)
        xorl    %ebx, %ebx
        movq    (%rsp), %rax
        xorq    %r9, %rax
        mulq    %r8
        addq    %rax, %rdi
        adcq    %rdx, %rbx
        movq    0x28(%rsp), %rax
        xorq    %r11, %rax
        mulq    %r10
        addq    %rax, %rdi
        adcq    %rdx, %rbx
        xorl    %ebp, %ebp
        movq    (%rsp), %rax
        xorq    %r13, %rax
        mulq    %r12
        addq    %rax, %rsi
        adcq    %rdx, %rbp
        movq    0x28(%rsp), %rax
        xorq    %r15, %rax
        mulq    %r14
        addq    %rax, %rsi
        adcq    %rdx, %rbp
        xorl    %ecx, %ecx
        movq    0x8(%rsp), %rax
        xorq    %r9, %rax
        mulq    %r8
        addq    %rax, %rbx
        adcq    %rdx, %rcx
        movq    0x30(%rsp), %rax
        xorq    %r11, %rax
        mulq    %r10
        addq    %rax, %rbx
        adcq    %rdx, %rcx
        shrdq   $0x3b, %rbx, %rdi
        movq    %rdi, (%rsp)
        xorl    %edi, %edi
        movq    0x8(%rsp), %rax
        xorq    %r13, %rax
        mulq    %r12
        addq    %rax, %rbp
        adcq    %rdx, %rdi
        movq    0x30(%rsp), %rax
        xorq    %r15, %rax
        mulq    %r14
        addq    %rax, %rbp
        adcq    %rdx, %rdi
        shrdq   $0x3b, %rbp, %rsi
        movq    %rsi, 0x28(%rsp)
        xorl    %esi, %esi
        movq    0x10(%rsp), %rax
        xorq    %r9, %rax
        mulq    %r8
        addq    %rax, %rcx
        adcq    %rdx, %rsi
        movq    0x38(%rsp), %rax
        xorq    %r11, %rax
        mulq    %r10
        addq    %rax, %rcx
        adcq    %rdx, %rsi
        shrdq   $0x3b, %rcx, %rbx
        movq    %rbx, 0x8(%rsp)
        xorl    %ebx, %ebx
        movq    0x10(%rsp), %rax
        xorq    %r13, %rax
        mulq    %r12
        addq    %rax, %rdi
        adcq    %rdx, %rbx
        movq    0x38(%rsp), %rax
        xorq    %r15, %rax
        mulq    %r14
        addq    %rax, %rdi
        adcq    %rdx, %rbx
        shrdq   $0x3b, %rdi, %rbp
        movq    %rbp, 0x30(%rsp)
        movq    0x18(%rsp), %rax
        xorq    %r9, %rax
        movq    0x20(%rsp), %rbp
        xorq    %r9, %rbp
        andq    %r8, %rbp
        negq    %rbp
        mulq    %r8
        addq    %rax, %rsi
        adcq    %rdx, %rbp
        movq    0x40(%rsp), %rax
        xorq    %r11, %rax
        movq    0x48(%rsp), %rdx
        xorq    %r11, %rdx
        andq    %r10, %rdx
        subq    %rdx, %rbp
        mulq    %r10
        addq    %rax, %rsi
        adcq    %rdx, %rbp
        shrdq   $0x3b, %rsi, %rcx
        movq    %rcx, 0x10(%rsp)
        shrdq   $0x3b, %rbp, %rsi
        sarq    $0x3b, %rbp
        movq    0x18(%rsp), %rax
        movq    %rsi, 0x18(%rsp)
        movq    0x20(%rsp), %rsi
        movq    %rbp, 0x20(%rsp)
        xorq    %r13, %rax
        xorq    %r13, %rsi
        andq    %r12, %rsi
        negq    %rsi
        mulq    %r12
        addq    %rax, %rbx
        adcq    %rdx, %rsi
        movq    0x40(%rsp), %rax
        xorq    %r15, %rax
        movq    0x48(%rsp), %rdx
        xorq    %r15, %rdx
        andq    %r14, %rdx
        subq    %rdx, %rsi
        mulq    %r14
        addq    %rax, %rbx
        adcq    %rdx, %rsi
        shrdq   $0x3b, %rbx, %rdi
        movq    %rdi, 0x38(%rsp)
        shrdq   $0x3b, %rsi, %rbx
        movq    %rbx, 0x40(%rsp)
        sarq    $0x3b, %rsi
        movq    %rsi, 0x48(%rsp)
        movq    0xa0(%rsp), %rbx
        movq    0xa8(%rsp), %rbp
        xorl    %ecx, %ecx
        movq    0x50(%rsp), %rax
        xorq    %r9, %rax
        mulq    %r8
        addq    %rax, %rbx
        adcq    %rdx, %rcx
        movq    0x78(%rsp), %rax
        xorq    %r11, %rax
        mulq    %r10
        addq    %rax, %rbx
        adcq    %rdx, %rcx
        xorl    %esi, %esi
        movq    0x50(%rsp), %rax
        xorq    %r13, %rax
        mulq    %r12
        movq    %rbx, 0x50(%rsp)
        addq    %rax, %rbp
        adcq    %rdx, %rsi
        movq    0x78(%rsp), %rax
        xorq    %r15, %rax
        mulq    %r14
        addq    %rax, %rbp
        adcq    %rdx, %rsi
        movq    %rbp, 0x78(%rsp)
        xorl    %ebx, %ebx
        movq    0x58(%rsp), %rax
        xorq    %r9, %rax
        mulq    %r8
        addq    %rax, %rcx
        adcq    %rdx, %rbx
        movq    0x80(%rsp), %rax
        xorq    %r11, %rax
        mulq    %r10
        addq    %rax, %rcx
        adcq    %rdx, %rbx
        xorl    %ebp, %ebp
        movq    0x58(%rsp), %rax
        xorq    %r13, %rax
        mulq    %r12
        movq    %rcx, 0x58(%rsp)
        addq    %rax, %rsi
        adcq    %rdx, %rbp
        movq    0x80(%rsp), %rax
        xorq    %r15, %rax
        mulq    %r14
        addq    %rax, %rsi
        adcq    %rdx, %rbp
        movq    %rsi, 0x80(%rsp)
        xorl    %ecx, %ecx
        movq    0x60(%rsp), %rax
        xorq    %r9, %rax
        mulq    %r8
        addq    %rax, %rbx
        adcq    %rdx, %rcx
        movq    0x88(%rsp), %rax
        xorq    %r11, %rax
        mulq    %r10
        addq    %rax, %rbx
        adcq    %rdx, %rcx
        xorl    %esi, %esi
        movq    0x60(%rsp), %rax
        xorq    %r13, %rax
        mulq    %r12
        movq    %rbx, 0x60(%rsp)
        addq    %rax, %rbp
        adcq    %rdx, %rsi
        movq    0x88(%rsp), %rax
        xorq    %r15, %rax
        mulq    %r14
        addq    %rax, %rbp
        adcq    %rdx, %rsi
        movq    %rbp, 0x88(%rsp)
        movq    0x68(%rsp), %rax
        xorq    %r9, %rax
        movq    %r9, %rbx
        andq    %r8, %rbx
        negq    %rbx
        mulq    %r8
        addq    %rax, %rcx
        adcq    %rdx, %rbx
        movq    0x90(%rsp), %rax
        xorq    %r11, %rax
        movq    %r11, %rdx
        andq    %r10, %rdx
        subq    %rdx, %rbx
        mulq    %r10
        addq    %rax, %rcx
        adcq    %rbx, %rdx
        movq    0x68(%rsp), %rax
        movq    %rcx, 0x68(%rsp)
        movq    %rdx, 0x70(%rsp)
        xorq    %r13, %rax
        movq    %r13, %rcx
        andq    %r12, %rcx
        negq    %rcx
        mulq    %r12
        addq    %rax, %rsi
        adcq    %rdx, %rcx
        movq    0x90(%rsp), %rax
        xorq    %r15, %rax
        movq    %r15, %rdx
        andq    %r14, %rdx
        subq    %rdx, %rcx
        mulq    %r14
        addq    %rax, %rsi
        adcq    %rcx, %rdx
        movq    %rsi, 0x90(%rsp)
        movq    %rdx, 0x98(%rsp)
        movabsq $0xe000000000000000, %r8
        addq    0x50(%rsp), %r8
        movq    $0xffffffffffffffff, %r9
        adcq    0x58(%rsp), %r9
        movq    $0x1fffffff, %r10
        adcq    0x60(%rsp), %r10
        movabsq $0x2000000000000000, %r11
        adcq    0x68(%rsp), %r11
        movabsq $0x1fffffffe0000000, %r12
        adcq    0x70(%rsp), %r12
        movq    %r8, %rbx
        shlq    $0x20, %rbx
        movabsq $0xffffffff00000001, %rax
        mulq    %r8
        shrq    $0x20, %r8
        addq    %rbx, %r9
        adcq    %r8, %r10
        adcq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rax, %rax
        movl    $0xffffffff, %ebx
        andq    %rax, %rbx
        movabsq $0xffffffff00000001, %rdx
        andq    %rax, %rdx
        subq    %rax, %r9
        movq    %r9, 0x50(%rsp)
        sbbq    %rbx, %r10
        movq    %r10, 0x58(%rsp)
        sbbq    $0x0, %r11
        movq    %r11, 0x60(%rsp)
        sbbq    %rdx, %r12
        movq    %r12, 0x68(%rsp)
        movabsq $0xe000000000000000, %r8
        addq    0x78(%rsp), %r8
        movq    $0xffffffffffffffff, %r9
        adcq    0x80(%rsp), %r9
        movq    $0x1fffffff, %r10
        adcq    0x88(%rsp), %r10
        movabsq $0x2000000000000000, %r11
        adcq    0x90(%rsp), %r11
        movabsq $0x1fffffffe0000000, %r12
        adcq    0x98(%rsp), %r12
        movq    %r8, %rbx
        shlq    $0x20, %rbx
        movabsq $0xffffffff00000001, %rax
        mulq    %r8
        shrq    $0x20, %r8
        addq    %rbx, %r9
        adcq    %r8, %r10
        adcq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rax, %rax
        movl    $0xffffffff, %ebx
        andq    %rax, %rbx
        movabsq $0xffffffff00000001, %rdx
        andq    %rax, %rdx
        subq    %rax, %r9
        movq    %r9, 0x78(%rsp)
        sbbq    %rbx, %r10
        movq    %r10, 0x80(%rsp)
        sbbq    $0x0, %r11
        movq    %r11, 0x88(%rsp)
        sbbq    %rdx, %r12
        movq    %r12, 0x90(%rsp)
Lp256_scalarmulbase_alt_inv_midloop:
        movq    0xb8(%rsp), %rsi
        movq    (%rsp), %rdx
        movq    0x28(%rsp), %rcx
        movq    %rdx, %rbx
        andq    $0xfffff, %rbx
        movabsq $0xfffffe0000000000, %rax
        orq     %rax, %rbx
        andq    $0xfffff, %rcx
        movabsq $0xc000000000000000, %rax
        orq     %rax, %rcx
        movq    $0xfffffffffffffffe, %rax
        xorl    %ebp, %ebp
        movl    $0x2, %edx
        movq    %rbx, %rdi
        movq    %rax, %r8
        testq   %rsi, %rsi
        cmovs   %rbp, %r8
        testq   $0x1, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        sarq    $1, %rcx
        movl    $0x100000, %eax
        leaq    (%rbx,%rax), %rdx
        leaq    (%rcx,%rax), %rdi
        shlq    $0x16, %rdx
        shlq    $0x16, %rdi
        sarq    $0x2b, %rdx
        sarq    $0x2b, %rdi
        movabsq $0x20000100000, %rax
        leaq    (%rbx,%rax), %rbx
        leaq    (%rcx,%rax), %rcx
        sarq    $0x2a, %rbx
        sarq    $0x2a, %rcx
        movq    %rdx, 0xc0(%rsp)
        movq    %rbx, 0xc8(%rsp)
        movq    %rdi, 0xd0(%rsp)
        movq    %rcx, 0xd8(%rsp)
        movq    (%rsp), %r12
        imulq   %r12, %rdi
        imulq   %rdx, %r12
        movq    0x28(%rsp), %r13
        imulq   %r13, %rbx
        imulq   %rcx, %r13
        addq    %rbx, %r12
        addq    %rdi, %r13
        sarq    $0x14, %r12
        sarq    $0x14, %r13
        movq    %r12, %rbx
        andq    $0xfffff, %rbx
        movabsq $0xfffffe0000000000, %rax
        orq     %rax, %rbx
        movq    %r13, %rcx
        andq    $0xfffff, %rcx
        movabsq $0xc000000000000000, %rax
        orq     %rax, %rcx
        movq    $0xfffffffffffffffe, %rax
        movl    $0x2, %edx
        movq    %rbx, %rdi
        movq    %rax, %r8
        testq   %rsi, %rsi
        cmovs   %rbp, %r8
        testq   $0x1, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        sarq    $1, %rcx
        movl    $0x100000, %eax
        leaq    (%rbx,%rax), %r8
        leaq    (%rcx,%rax), %r10
        shlq    $0x16, %r8
        shlq    $0x16, %r10
        sarq    $0x2b, %r8
        sarq    $0x2b, %r10
        movabsq $0x20000100000, %rax
        leaq    (%rbx,%rax), %r15
        leaq    (%rcx,%rax), %r11
        sarq    $0x2a, %r15
        sarq    $0x2a, %r11
        movq    %r13, %rbx
        movq    %r12, %rcx
        imulq   %r8, %r12
        imulq   %r15, %rbx
        addq    %rbx, %r12
        imulq   %r11, %r13
        imulq   %r10, %rcx
        addq    %rcx, %r13
        sarq    $0x14, %r12
        sarq    $0x14, %r13
        movq    %r12, %rbx
        andq    $0xfffff, %rbx
        movabsq $0xfffffe0000000000, %rax
        orq     %rax, %rbx
        movq    %r13, %rcx
        andq    $0xfffff, %rcx
        movabsq $0xc000000000000000, %rax
        orq     %rax, %rcx
        movq    0xc0(%rsp), %rax
        imulq   %r8, %rax
        movq    0xd0(%rsp), %rdx
        imulq   %r15, %rdx
        imulq   0xc8(%rsp), %r8
        imulq   0xd8(%rsp), %r15
        addq    %r8, %r15
        leaq    (%rax,%rdx), %r9
        movq    0xc0(%rsp), %rax
        imulq   %r10, %rax
        movq    0xd0(%rsp), %rdx
        imulq   %r11, %rdx
        imulq   0xc8(%rsp), %r10
        imulq   0xd8(%rsp), %r11
        addq    %r10, %r11
        leaq    (%rax,%rdx), %r13
        movq    $0xfffffffffffffffe, %rax
        movl    $0x2, %edx
        movq    %rbx, %rdi
        movq    %rax, %r8
        testq   %rsi, %rsi
        cmovs   %rbp, %r8
        testq   $0x1, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        cmovs   %rbp, %r8
        movq    %rbx, %rdi
        testq   %rdx, %rcx
        cmoveq  %rbp, %r8
        cmoveq  %rbp, %rdi
        sarq    $1, %rcx
        xorq    %r8, %rdi
        xorq    %r8, %rsi
        btq     $0x3f, %r8
        cmovbq  %rcx, %rbx
        movq    %rax, %r8
        subq    %rax, %rsi
        leaq    (%rcx,%rdi), %rcx
        sarq    $1, %rcx
        movl    $0x100000, %eax
        leaq    (%rbx,%rax), %r8
        leaq    (%rcx,%rax), %r12
        shlq    $0x15, %r8
        shlq    $0x15, %r12
        sarq    $0x2b, %r8
        sarq    $0x2b, %r12
        movabsq $0x20000100000, %rax
        leaq    (%rbx,%rax), %r10
        leaq    (%rcx,%rax), %r14
        sarq    $0x2b, %r10
        sarq    $0x2b, %r14
        movq    %r9, %rax
        imulq   %r8, %rax
        movq    %r13, %rdx
        imulq   %r10, %rdx
        imulq   %r15, %r8
        imulq   %r11, %r10
        addq    %r8, %r10
        leaq    (%rax,%rdx), %r8
        movq    %r9, %rax
        imulq   %r12, %rax
        movq    %r13, %rdx
        imulq   %r14, %rdx
        imulq   %r15, %r12
        imulq   %r11, %r14
        addq    %r12, %r14
        leaq    (%rax,%rdx), %r12
        movq    %rsi, 0xb8(%rsp)
        decq     0xb0(%rsp)
        jne     Lp256_scalarmulbase_alt_inv_loop
        movq    (%rsp), %rax
        movq    0x28(%rsp), %rcx
        imulq   %r8, %rax
        imulq   %r10, %rcx
        addq    %rcx, %rax
        sarq    $0x3f, %rax
        movq    %r8, %r9
        sarq    $0x3f, %r9
        xorq    %r9, %r8
        subq    %r9, %r8
        xorq    %rax, %r9
        movq    %r10, %r11
        sarq    $0x3f, %r11
        xorq    %r11, %r10
        subq    %r11, %r10
        xorq    %rax, %r11
        movq    %r12, %r13
        sarq    $0x3f, %r13
        xorq    %r13, %r12
        subq    %r13, %r12
        xorq    %rax, %r13
        movq    %r14, %r15
        sarq    $0x3f, %r15
        xorq    %r15, %r14
        subq    %r15, %r14
        xorq    %rax, %r15
        movq    %r8, %rax
        andq    %r9, %rax
        movq    %r10, %r12
        andq    %r11, %r12
        addq    %rax, %r12
        xorl    %r13d, %r13d
        movq    0x50(%rsp), %rax
        xorq    %r9, %rax
        mulq    %r8
        addq    %rax, %r12
        adcq    %rdx, %r13
        movq    0x78(%rsp), %rax
        xorq    %r11, %rax
        mulq    %r10
        addq    %rax, %r12
        adcq    %rdx, %r13
        xorl    %r14d, %r14d
        movq    0x58(%rsp), %rax
        xorq    %r9, %rax
        mulq    %r8
        addq    %rax, %r13
        adcq    %rdx, %r14
        movq    0x80(%rsp), %rax
        xorq    %r11, %rax
        mulq    %r10
        addq    %rax, %r13
        adcq    %rdx, %r14
        xorl    %r15d, %r15d
        movq    0x60(%rsp), %rax
        xorq    %r9, %rax
        mulq    %r8
        addq    %rax, %r14
        adcq    %rdx, %r15
        movq    0x88(%rsp), %rax
        xorq    %r11, %rax
        mulq    %r10
        addq    %rax, %r14
        adcq    %rdx, %r15
        movq    0x68(%rsp), %rax
        xorq    %r9, %rax
        andq    %r8, %r9
        negq    %r9
        mulq    %r8
        addq    %rax, %r15
        adcq    %rdx, %r9
        movq    0x90(%rsp), %rax
        xorq    %r11, %rax
        movq    %r11, %rdx
        andq    %r10, %rdx
        subq    %rdx, %r9
        mulq    %r10
        addq    %rax, %r15
        adcq    %rdx, %r9
        movq    %r12, 0x50(%rsp)
        movq    %r13, 0x58(%rsp)
        movq    %r14, 0x60(%rsp)
        movq    %r15, 0x68(%rsp)
        movq    %r9, 0x70(%rsp)
        movabsq $0xe000000000000000, %r8
        addq    0x50(%rsp), %r8
        movq    $0xffffffffffffffff, %r9
        adcq    0x58(%rsp), %r9
        movq    $0x1fffffff, %r10
        adcq    0x60(%rsp), %r10
        movabsq $0x2000000000000000, %r11
        adcq    0x68(%rsp), %r11
        movabsq $0x1fffffffe0000000, %r12
        adcq    0x70(%rsp), %r12
        movq    %r8, %rbx
        shlq    $0x20, %rbx
        movabsq $0xffffffff00000001, %rax
        mulq    %r8
        shrq    $0x20, %r8
        addq    %rbx, %r9
        adcq    %r8, %r10
        adcq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rax, %rax
        movl    $0xffffffff, %ebx
        andq    %rax, %rbx
        movabsq $0xffffffff00000001, %rdx
        andq    %rax, %rdx
        subq    %rax, %r9
        movq    %r9, 0x50(%rsp)
        sbbq    %rbx, %r10
        movq    %r10, 0x58(%rsp)
        sbbq    $0x0, %r11
        movq    %r11, 0x60(%rsp)
        sbbq    %rdx, %r12
        movq    %r12, 0x68(%rsp)
        movq    0x50(%rsp), %r8
        movq    0x58(%rsp), %r9
        movq    0x60(%rsp), %r10
        movq    0x68(%rsp), %r11
        movl    $0x1, %eax
        movl    $0xffffffff, %ebx
        leaq    -0x2(%rax), %rcx
        leaq    -0x1(%rbx), %rdx
        notq    %rbx
        addq    %r8, %rax
        adcq    %r9, %rbx
        adcq    %r10, %rcx
        adcq    %r11, %rdx
        cmovaeq %r8, %rax
        cmovaeq %r9, %rbx
        cmovaeq %r10, %rcx
        cmovaeq %r11, %rdx
        movq    0xe0(%rsp), %rdi
        movq    %rax, (%rdi)
        movq    %rbx, 0x8(%rdi)
        movq    %rcx, 0x10(%rdi)
        movq    %rdx, 0x18(%rdi)
        CFI_INC_RSP(240)
        CFI_POP(%r15)
        CFI_POP(%r14)
        CFI_POP(%r13)
        CFI_POP(%r12)
        CFI_POP(%rbp)
        CFI_POP(%rbx)
        CFI_RET

S2N_BN_SIZE_DIRECTIVE(Lp256_scalarmulbase_alt_local_inv_p256)

S2N_BN_FUNCTION_TYPE_DIRECTIVE(Lp256_scalarmulbase_alt_local_montmul_p256)

Lp256_scalarmulbase_alt_local_montmul_p256:
        CFI_START
        CFI_PUSH(%rbx)
        CFI_PUSH(%r12)
        CFI_PUSH(%r13)
        CFI_PUSH(%r14)
        CFI_PUSH(%r15)
        movq    %rdx, %rcx
        movq    (%rcx), %rbx
        movq    (%rsi), %rax
        mulq    %rbx
        movq    %rax, %r8
        movq    %rdx, %r9
        movq    0x8(%rsi), %rax
        mulq    %rbx
        xorl    %r10d, %r10d
        addq    %rax, %r9
        adcq    %rdx, %r10
        movq    0x10(%rsi), %rax
        mulq    %rbx
        xorl    %r11d, %r11d
        addq    %rax, %r10
        adcq    %rdx, %r11
        movq    0x18(%rsi), %rax
        mulq    %rbx
        xorl    %r12d, %r12d
        addq    %rax, %r11
        adcq    %rdx, %r12
        movq    0x8(%rcx), %rbx
        xorl    %r13d, %r13d
        movq    (%rsi), %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r14, %r14
        movq    0x8(%rsi), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r14, %r14
        movq    0x10(%rsi), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r14, %r14
        movq    0x18(%rsi), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        xorl    %r14d, %r14d
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r14, %r14
        movq    0x10(%rcx), %rbx
        xorl    %r15d, %r15d
        movq    (%rsi), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r8, %r8
        movq    0x8(%rsi), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r8, %r8
        movq    0x10(%rsi), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r8, %r8
        movq    0x18(%rsi), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        adcq    %r15, %r15
        movq    0x18(%rcx), %rbx
        xorl    %r8d, %r8d
        movq    (%rsi), %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r9, %r9
        movq    0x8(%rsi), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r9, %r9
        movq    0x10(%rsi), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %r9, %r9
        movq    0x18(%rsi), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r8, %r8
        xorl    %r9d, %r9d
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        decq    %rbx
        adcq    %r13, %rbx
        decq    %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, (%rdi)
        movq    %r13, 0x8(%rdi)
        movq    %r14, 0x10(%rdi)
        movq    %r15, 0x18(%rdi)
        CFI_POP(%r15)
        CFI_POP(%r14)
        CFI_POP(%r13)
        CFI_POP(%r12)
        CFI_POP(%rbx)
        CFI_RET

S2N_BN_SIZE_DIRECTIVE(Lp256_scalarmulbase_alt_local_montmul_p256)

S2N_BN_FUNCTION_TYPE_DIRECTIVE(Lp256_scalarmulbase_alt_local_montsqr_p256)

Lp256_scalarmulbase_alt_local_montsqr_p256:
        CFI_START
        CFI_PUSH(%rbx)
        CFI_PUSH(%r12)
        CFI_PUSH(%r13)
        CFI_PUSH(%r14)
        CFI_PUSH(%r15)
        movq    (%rsi), %rax
        movq    %rax, %rbx
        mulq    %rax
        movq    %rax, %r8
        movq    %rdx, %r15
        movq    0x8(%rsi), %rax
        mulq    %rbx
        movq    %rax, %r9
        movq    %rdx, %r10
        movq    0x18(%rsi), %rax
        movq    %rax, %r13
        mulq    %rbx
        movq    %rax, %r11
        movq    %rdx, %r12
        movq    0x10(%rsi), %rax
        movq    %rax, %rbx
        mulq    %r13
        movq    %rax, %r13
        movq    %rdx, %r14
        movq    (%rsi), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %rcx, %rcx
        movq    0x8(%rsi), %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    0x18(%rsi), %rbx
        movq    0x8(%rsi), %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    $0x0, %r14
        xorl    %ecx, %ecx
        addq    %r9, %r9
        adcq    %r10, %r10
        adcq    %r11, %r11
        adcq    %r12, %r12
        adcq    %r13, %r13
        adcq    %r14, %r14
        adcq    %rcx, %rcx
        movq    0x8(%rsi), %rax
        mulq    %rax
        addq    %r15, %r9
        adcq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        movq    0x10(%rsi), %rax
        mulq    %rax
        negq    %r15
        adcq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r15, %r15
        movq    0x18(%rsi), %rax
        mulq    %rax
        negq    %r15
        adcq    %rax, %r14
        adcq    %rcx, %rdx
        movq    %rdx, %r15
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %rcx, %rcx
        movq    %r9, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        xorl    %r8d, %r8d
        movq    %r9, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r8, %r14
        adcq    %r8, %r15
        adcq    %r8, %r8
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        xorl    %r9d, %r9d
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        leaq    -0x1(%rbx), %rbx
        adcq    %r13, %rbx
        leaq    -0x1(%r9), %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, (%rdi)
        movq    %r13, 0x8(%rdi)
        movq    %r14, 0x10(%rdi)
        movq    %r15, 0x18(%rdi)
        CFI_POP(%r15)
        CFI_POP(%r14)
        CFI_POP(%r13)
        CFI_POP(%r12)
        CFI_POP(%rbx)
        CFI_RET

S2N_BN_SIZE_DIRECTIVE(Lp256_scalarmulbase_alt_local_montsqr_p256)

S2N_BN_FUNCTION_TYPE_DIRECTIVE(Lp256_scalarmulbase_alt_local_p256_montjmixadd)

Lp256_scalarmulbase_alt_local_p256_montjmixadd:
        CFI_START
        CFI_PUSH(%rbx)
        CFI_PUSH(%rbp)
        CFI_PUSH(%r12)
        CFI_PUSH(%r13)
        CFI_PUSH(%r14)
        CFI_PUSH(%r15)
        CFI_DEC_RSP(192)
        movq    %rdx, %rbp
        movq    0x40(%rsi), %rax
        movq    %rax, %rbx
        mulq    %rax
        movq    %rax, %r8
        movq    %rdx, %r15
        movq    0x48(%rsi), %rax
        mulq    %rbx
        movq    %rax, %r9
        movq    %rdx, %r10
        movq    0x58(%rsi), %rax
        movq    %rax, %r13
        mulq    %rbx
        movq    %rax, %r11
        movq    %rdx, %r12
        movq    0x50(%rsi), %rax
        movq    %rax, %rbx
        mulq    %r13
        movq    %rax, %r13
        movq    %rdx, %r14
        movq    0x40(%rsi), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %rcx, %rcx
        movq    0x48(%rsi), %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    0x58(%rsi), %rbx
        movq    0x48(%rsi), %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    $0x0, %r14
        xorl    %ecx, %ecx
        addq    %r9, %r9
        adcq    %r10, %r10
        adcq    %r11, %r11
        adcq    %r12, %r12
        adcq    %r13, %r13
        adcq    %r14, %r14
        adcq    %rcx, %rcx
        movq    0x48(%rsi), %rax
        mulq    %rax
        addq    %r15, %r9
        adcq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        movq    0x50(%rsi), %rax
        mulq    %rax
        negq    %r15
        adcq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r15, %r15
        movq    0x58(%rsi), %rax
        mulq    %rax
        negq    %r15
        adcq    %rax, %r14
        adcq    %rcx, %rdx
        movq    %rdx, %r15
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %rcx, %rcx
        movq    %r9, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        xorl    %r8d, %r8d
        movq    %r9, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r8, %r14
        adcq    %r8, %r15
        adcq    %r8, %r8
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        xorl    %r9d, %r9d
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        leaq    -0x1(%rbx), %rbx
        adcq    %r13, %rbx
        leaq    -0x1(%r9), %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, (%rsp)
        movq    %r13, 0x8(%rsp)
        movq    %r14, 0x10(%rsp)
        movq    %r15, 0x18(%rsp)
        movq    0x20(%rbp), %rbx
        movq    0x40(%rsi), %rax
        mulq    %rbx
        movq    %rax, %r8
        movq    %rdx, %r9
        movq    0x48(%rsi), %rax
        mulq    %rbx
        xorl    %r10d, %r10d
        addq    %rax, %r9
        adcq    %rdx, %r10
        movq    0x50(%rsi), %rax
        mulq    %rbx
        xorl    %r11d, %r11d
        addq    %rax, %r10
        adcq    %rdx, %r11
        movq    0x58(%rsi), %rax
        mulq    %rbx
        xorl    %r12d, %r12d
        addq    %rax, %r11
        adcq    %rdx, %r12
        movq    0x28(%rbp), %rbx
        xorl    %r13d, %r13d
        movq    0x40(%rsi), %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r14, %r14
        movq    0x48(%rsi), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r14, %r14
        movq    0x50(%rsi), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r14, %r14
        movq    0x58(%rsi), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        xorl    %r14d, %r14d
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r14, %r14
        movq    0x30(%rbp), %rbx
        xorl    %r15d, %r15d
        movq    0x40(%rsi), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r8, %r8
        movq    0x48(%rsi), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r8, %r8
        movq    0x50(%rsi), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r8, %r8
        movq    0x58(%rsi), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        adcq    %r15, %r15
        movq    0x38(%rbp), %rbx
        xorl    %r8d, %r8d
        movq    0x40(%rsi), %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r9, %r9
        movq    0x48(%rsi), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r9, %r9
        movq    0x50(%rsi), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %r9, %r9
        movq    0x58(%rsi), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r8, %r8
        xorl    %r9d, %r9d
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        decq    %rbx
        adcq    %r13, %rbx
        decq    %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, 0x20(%rsp)
        movq    %r13, 0x28(%rsp)
        movq    %r14, 0x30(%rsp)
        movq    %r15, 0x38(%rsp)
        movq    0x0(%rbp), %rbx
        movq    (%rsp), %rax
        mulq    %rbx
        movq    %rax, %r8
        movq    %rdx, %r9
        movq    0x8(%rsp), %rax
        mulq    %rbx
        xorl    %r10d, %r10d
        addq    %rax, %r9
        adcq    %rdx, %r10
        movq    0x10(%rsp), %rax
        mulq    %rbx
        xorl    %r11d, %r11d
        addq    %rax, %r10
        adcq    %rdx, %r11
        movq    0x18(%rsp), %rax
        mulq    %rbx
        xorl    %r12d, %r12d
        addq    %rax, %r11
        adcq    %rdx, %r12
        movq    0x8(%rbp), %rbx
        xorl    %r13d, %r13d
        movq    (%rsp), %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r14, %r14
        movq    0x8(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r14, %r14
        movq    0x10(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r14, %r14
        movq    0x18(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        xorl    %r14d, %r14d
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r14, %r14
        movq    0x10(%rbp), %rbx
        xorl    %r15d, %r15d
        movq    (%rsp), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r8, %r8
        movq    0x8(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r8, %r8
        movq    0x10(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r8, %r8
        movq    0x18(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        adcq    %r15, %r15
        movq    0x18(%rbp), %rbx
        xorl    %r8d, %r8d
        movq    (%rsp), %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r9, %r9
        movq    0x8(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r9, %r9
        movq    0x10(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %r9, %r9
        movq    0x18(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r8, %r8
        xorl    %r9d, %r9d
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        decq    %rbx
        adcq    %r13, %rbx
        decq    %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, 0x40(%rsp)
        movq    %r13, 0x48(%rsp)
        movq    %r14, 0x50(%rsp)
        movq    %r15, 0x58(%rsp)
        movq    0x20(%rsp), %rbx
        movq    (%rsp), %rax
        mulq    %rbx
        movq    %rax, %r8
        movq    %rdx, %r9
        movq    0x8(%rsp), %rax
        mulq    %rbx
        xorl    %r10d, %r10d
        addq    %rax, %r9
        adcq    %rdx, %r10
        movq    0x10(%rsp), %rax
        mulq    %rbx
        xorl    %r11d, %r11d
        addq    %rax, %r10
        adcq    %rdx, %r11
        movq    0x18(%rsp), %rax
        mulq    %rbx
        xorl    %r12d, %r12d
        addq    %rax, %r11
        adcq    %rdx, %r12
        movq    0x28(%rsp), %rbx
        xorl    %r13d, %r13d
        movq    (%rsp), %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r14, %r14
        movq    0x8(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r14, %r14
        movq    0x10(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r14, %r14
        movq    0x18(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        xorl    %r14d, %r14d
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r14, %r14
        movq    0x30(%rsp), %rbx
        xorl    %r15d, %r15d
        movq    (%rsp), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r8, %r8
        movq    0x8(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r8, %r8
        movq    0x10(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r8, %r8
        movq    0x18(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        adcq    %r15, %r15
        movq    0x38(%rsp), %rbx
        xorl    %r8d, %r8d
        movq    (%rsp), %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r9, %r9
        movq    0x8(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r9, %r9
        movq    0x10(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %r9, %r9
        movq    0x18(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r8, %r8
        xorl    %r9d, %r9d
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        decq    %rbx
        adcq    %r13, %rbx
        decq    %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, 0x20(%rsp)
        movq    %r13, 0x28(%rsp)
        movq    %r14, 0x30(%rsp)
        movq    %r15, 0x38(%rsp)
        movq    0x40(%rsp), %rax
        subq    (%rsi), %rax
        movq    0x48(%rsp), %rcx
        sbbq    0x8(%rsi), %rcx
        movq    0x50(%rsp), %r8
        sbbq    0x10(%rsi), %r8
        movq    0x58(%rsp), %r9
        sbbq    0x18(%rsi), %r9
        movl    $0xffffffff, %r10d
        sbbq    %r11, %r11
        xorq    %rdx, %rdx
        andq    %r11, %r10
        subq    %r10, %rdx
        addq    %r11, %rax
        movq    %rax, 0xa0(%rsp)
        adcq    %r10, %rcx
        movq    %rcx, 0xa8(%rsp)
        adcq    $0x0, %r8
        movq    %r8, 0xb0(%rsp)
        adcq    %rdx, %r9
        movq    %r9, 0xb8(%rsp)
        movq    0x20(%rsp), %rax
        subq    0x20(%rsi), %rax
        movq    0x28(%rsp), %rcx
        sbbq    0x28(%rsi), %rcx
        movq    0x30(%rsp), %r8
        sbbq    0x30(%rsi), %r8
        movq    0x38(%rsp), %r9
        sbbq    0x38(%rsi), %r9
        movl    $0xffffffff, %r10d
        sbbq    %r11, %r11
        xorq    %rdx, %rdx
        andq    %r11, %r10
        subq    %r10, %rdx
        addq    %r11, %rax
        movq    %rax, 0x20(%rsp)
        adcq    %r10, %rcx
        movq    %rcx, 0x28(%rsp)
        adcq    $0x0, %r8
        movq    %r8, 0x30(%rsp)
        adcq    %rdx, %r9
        movq    %r9, 0x38(%rsp)
        movq    0xa0(%rsp), %rax
        movq    %rax, %rbx
        mulq    %rax
        movq    %rax, %r8
        movq    %rdx, %r15
        movq    0xa8(%rsp), %rax
        mulq    %rbx
        movq    %rax, %r9
        movq    %rdx, %r10
        movq    0xb8(%rsp), %rax
        movq    %rax, %r13
        mulq    %rbx
        movq    %rax, %r11
        movq    %rdx, %r12
        movq    0xb0(%rsp), %rax
        movq    %rax, %rbx
        mulq    %r13
        movq    %rax, %r13
        movq    %rdx, %r14
        movq    0xa0(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %rcx, %rcx
        movq    0xa8(%rsp), %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    0xb8(%rsp), %rbx
        movq    0xa8(%rsp), %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    $0x0, %r14
        xorl    %ecx, %ecx
        addq    %r9, %r9
        adcq    %r10, %r10
        adcq    %r11, %r11
        adcq    %r12, %r12
        adcq    %r13, %r13
        adcq    %r14, %r14
        adcq    %rcx, %rcx
        movq    0xa8(%rsp), %rax
        mulq    %rax
        addq    %r15, %r9
        adcq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        movq    0xb0(%rsp), %rax
        mulq    %rax
        negq    %r15
        adcq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r15, %r15
        movq    0xb8(%rsp), %rax
        mulq    %rax
        negq    %r15
        adcq    %rax, %r14
        adcq    %rcx, %rdx
        movq    %rdx, %r15
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %rcx, %rcx
        movq    %r9, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        xorl    %r8d, %r8d
        movq    %r9, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r8, %r14
        adcq    %r8, %r15
        adcq    %r8, %r8
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        xorl    %r9d, %r9d
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        leaq    -0x1(%rbx), %rbx
        adcq    %r13, %rbx
        leaq    -0x1(%r9), %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, 0x60(%rsp)
        movq    %r13, 0x68(%rsp)
        movq    %r14, 0x70(%rsp)
        movq    %r15, 0x78(%rsp)
        movq    0x20(%rsp), %rax
        movq    %rax, %rbx
        mulq    %rax
        movq    %rax, %r8
        movq    %rdx, %r15
        movq    0x28(%rsp), %rax
        mulq    %rbx
        movq    %rax, %r9
        movq    %rdx, %r10
        movq    0x38(%rsp), %rax
        movq    %rax, %r13
        mulq    %rbx
        movq    %rax, %r11
        movq    %rdx, %r12
        movq    0x30(%rsp), %rax
        movq    %rax, %rbx
        mulq    %r13
        movq    %rax, %r13
        movq    %rdx, %r14
        movq    0x20(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %rcx, %rcx
        movq    0x28(%rsp), %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    0x38(%rsp), %rbx
        movq    0x28(%rsp), %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    $0x0, %r14
        xorl    %ecx, %ecx
        addq    %r9, %r9
        adcq    %r10, %r10
        adcq    %r11, %r11
        adcq    %r12, %r12
        adcq    %r13, %r13
        adcq    %r14, %r14
        adcq    %rcx, %rcx
        movq    0x28(%rsp), %rax
        mulq    %rax
        addq    %r15, %r9
        adcq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        movq    0x30(%rsp), %rax
        mulq    %rax
        negq    %r15
        adcq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r15, %r15
        movq    0x38(%rsp), %rax
        mulq    %rax
        negq    %r15
        adcq    %rax, %r14
        adcq    %rcx, %rdx
        movq    %rdx, %r15
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %rcx, %rcx
        movq    %r9, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        xorl    %r8d, %r8d
        movq    %r9, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r8, %r14
        adcq    %r8, %r15
        adcq    %r8, %r8
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        xorl    %r9d, %r9d
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        leaq    -0x1(%rbx), %rbx
        adcq    %r13, %rbx
        leaq    -0x1(%r9), %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, (%rsp)
        movq    %r13, 0x8(%rsp)
        movq    %r14, 0x10(%rsp)
        movq    %r15, 0x18(%rsp)
        movq    (%rsi), %rbx
        movq    0x60(%rsp), %rax
        mulq    %rbx
        movq    %rax, %r8
        movq    %rdx, %r9
        movq    0x68(%rsp), %rax
        mulq    %rbx
        xorl    %r10d, %r10d
        addq    %rax, %r9
        adcq    %rdx, %r10
        movq    0x70(%rsp), %rax
        mulq    %rbx
        xorl    %r11d, %r11d
        addq    %rax, %r10
        adcq    %rdx, %r11
        movq    0x78(%rsp), %rax
        mulq    %rbx
        xorl    %r12d, %r12d
        addq    %rax, %r11
        adcq    %rdx, %r12
        movq    0x8(%rsi), %rbx
        xorl    %r13d, %r13d
        movq    0x60(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r14, %r14
        movq    0x68(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r14, %r14
        movq    0x70(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r14, %r14
        movq    0x78(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        xorl    %r14d, %r14d
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r14, %r14
        movq    0x10(%rsi), %rbx
        xorl    %r15d, %r15d
        movq    0x60(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r8, %r8
        movq    0x68(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r8, %r8
        movq    0x70(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r8, %r8
        movq    0x78(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        adcq    %r15, %r15
        movq    0x18(%rsi), %rbx
        xorl    %r8d, %r8d
        movq    0x60(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r9, %r9
        movq    0x68(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r9, %r9
        movq    0x70(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %r9, %r9
        movq    0x78(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r8, %r8
        xorl    %r9d, %r9d
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        decq    %rbx
        adcq    %r13, %rbx
        decq    %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, 0x80(%rsp)
        movq    %r13, 0x88(%rsp)
        movq    %r14, 0x90(%rsp)
        movq    %r15, 0x98(%rsp)
        movq    0x40(%rsp), %rbx
        movq    0x60(%rsp), %rax
        mulq    %rbx
        movq    %rax, %r8
        movq    %rdx, %r9
        movq    0x68(%rsp), %rax
        mulq    %rbx
        xorl    %r10d, %r10d
        addq    %rax, %r9
        adcq    %rdx, %r10
        movq    0x70(%rsp), %rax
        mulq    %rbx
        xorl    %r11d, %r11d
        addq    %rax, %r10
        adcq    %rdx, %r11
        movq    0x78(%rsp), %rax
        mulq    %rbx
        xorl    %r12d, %r12d
        addq    %rax, %r11
        adcq    %rdx, %r12
        movq    0x48(%rsp), %rbx
        xorl    %r13d, %r13d
        movq    0x60(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r14, %r14
        movq    0x68(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r14, %r14
        movq    0x70(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r14, %r14
        movq    0x78(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        xorl    %r14d, %r14d
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r14, %r14
        movq    0x50(%rsp), %rbx
        xorl    %r15d, %r15d
        movq    0x60(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r8, %r8
        movq    0x68(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r8, %r8
        movq    0x70(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r8, %r8
        movq    0x78(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        adcq    %r15, %r15
        movq    0x58(%rsp), %rbx
        xorl    %r8d, %r8d
        movq    0x60(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r9, %r9
        movq    0x68(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r9, %r9
        movq    0x70(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %r9, %r9
        movq    0x78(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r8, %r8
        xorl    %r9d, %r9d
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        decq    %rbx
        adcq    %r13, %rbx
        decq    %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, 0x40(%rsp)
        movq    %r13, 0x48(%rsp)
        movq    %r14, 0x50(%rsp)
        movq    %r15, 0x58(%rsp)
        movq    (%rsp), %rax
        subq    0x80(%rsp), %rax
        movq    0x8(%rsp), %rcx
        sbbq    0x88(%rsp), %rcx
        movq    0x10(%rsp), %r8
        sbbq    0x90(%rsp), %r8
        movq    0x18(%rsp), %r9
        sbbq    0x98(%rsp), %r9
        movl    $0xffffffff, %r10d
        sbbq    %r11, %r11
        xorq    %rdx, %rdx
        andq    %r11, %r10
        subq    %r10, %rdx
        addq    %r11, %rax
        movq    %rax, (%rsp)
        adcq    %r10, %rcx
        movq    %rcx, 0x8(%rsp)
        adcq    $0x0, %r8
        movq    %r8, 0x10(%rsp)
        adcq    %rdx, %r9
        movq    %r9, 0x18(%rsp)
        movq    0x40(%rsp), %rax
        subq    0x80(%rsp), %rax
        movq    0x48(%rsp), %rcx
        sbbq    0x88(%rsp), %rcx
        movq    0x50(%rsp), %r8
        sbbq    0x90(%rsp), %r8
        movq    0x58(%rsp), %r9
        sbbq    0x98(%rsp), %r9
        movl    $0xffffffff, %r10d
        sbbq    %r11, %r11
        xorq    %rdx, %rdx
        andq    %r11, %r10
        subq    %r10, %rdx
        addq    %r11, %rax
        movq    %rax, 0x60(%rsp)
        adcq    %r10, %rcx
        movq    %rcx, 0x68(%rsp)
        adcq    $0x0, %r8
        movq    %r8, 0x70(%rsp)
        adcq    %rdx, %r9
        movq    %r9, 0x78(%rsp)
        movq    0x40(%rsi), %rbx
        movq    0xa0(%rsp), %rax
        mulq    %rbx
        movq    %rax, %r8
        movq    %rdx, %r9
        movq    0xa8(%rsp), %rax
        mulq    %rbx
        xorl    %r10d, %r10d
        addq    %rax, %r9
        adcq    %rdx, %r10
        movq    0xb0(%rsp), %rax
        mulq    %rbx
        xorl    %r11d, %r11d
        addq    %rax, %r10
        adcq    %rdx, %r11
        movq    0xb8(%rsp), %rax
        mulq    %rbx
        xorl    %r12d, %r12d
        addq    %rax, %r11
        adcq    %rdx, %r12
        movq    0x48(%rsi), %rbx
        xorl    %r13d, %r13d
        movq    0xa0(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r14, %r14
        movq    0xa8(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r14, %r14
        movq    0xb0(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r14, %r14
        movq    0xb8(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        xorl    %r14d, %r14d
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r14, %r14
        movq    0x50(%rsi), %rbx
        xorl    %r15d, %r15d
        movq    0xa0(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r8, %r8
        movq    0xa8(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r8, %r8
        movq    0xb0(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r8, %r8
        movq    0xb8(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        adcq    %r15, %r15
        movq    0x58(%rsi), %rbx
        xorl    %r8d, %r8d
        movq    0xa0(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r9, %r9
        movq    0xa8(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r9, %r9
        movq    0xb0(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %r9, %r9
        movq    0xb8(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r8, %r8
        xorl    %r9d, %r9d
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        decq    %rbx
        adcq    %r13, %rbx
        decq    %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, 0xa0(%rsp)
        movq    %r13, 0xa8(%rsp)
        movq    %r14, 0xb0(%rsp)
        movq    %r15, 0xb8(%rsp)
        movq    (%rsp), %rax
        subq    0x40(%rsp), %rax
        movq    0x8(%rsp), %rcx
        sbbq    0x48(%rsp), %rcx
        movq    0x10(%rsp), %r8
        sbbq    0x50(%rsp), %r8
        movq    0x18(%rsp), %r9
        sbbq    0x58(%rsp), %r9
        movl    $0xffffffff, %r10d
        sbbq    %r11, %r11
        xorq    %rdx, %rdx
        andq    %r11, %r10
        subq    %r10, %rdx
        addq    %r11, %rax
        movq    %rax, (%rsp)
        adcq    %r10, %rcx
        movq    %rcx, 0x8(%rsp)
        adcq    $0x0, %r8
        movq    %r8, 0x10(%rsp)
        adcq    %rdx, %r9
        movq    %r9, 0x18(%rsp)
        movq    0x80(%rsp), %rax
        subq    (%rsp), %rax
        movq    0x88(%rsp), %rcx
        sbbq    0x8(%rsp), %rcx
        movq    0x90(%rsp), %r8
        sbbq    0x10(%rsp), %r8
        movq    0x98(%rsp), %r9
        sbbq    0x18(%rsp), %r9
        movl    $0xffffffff, %r10d
        sbbq    %r11, %r11
        xorq    %rdx, %rdx
        andq    %r11, %r10
        subq    %r10, %rdx
        addq    %r11, %rax
        movq    %rax, 0x80(%rsp)
        adcq    %r10, %rcx
        movq    %rcx, 0x88(%rsp)
        adcq    $0x0, %r8
        movq    %r8, 0x90(%rsp)
        adcq    %rdx, %r9
        movq    %r9, 0x98(%rsp)
        movq    0x20(%rsi), %rbx
        movq    0x60(%rsp), %rax
        mulq    %rbx
        movq    %rax, %r8
        movq    %rdx, %r9
        movq    0x68(%rsp), %rax
        mulq    %rbx
        xorl    %r10d, %r10d
        addq    %rax, %r9
        adcq    %rdx, %r10
        movq    0x70(%rsp), %rax
        mulq    %rbx
        xorl    %r11d, %r11d
        addq    %rax, %r10
        adcq    %rdx, %r11
        movq    0x78(%rsp), %rax
        mulq    %rbx
        xorl    %r12d, %r12d
        addq    %rax, %r11
        adcq    %rdx, %r12
        movq    0x28(%rsi), %rbx
        xorl    %r13d, %r13d
        movq    0x60(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r14, %r14
        movq    0x68(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r14, %r14
        movq    0x70(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r14, %r14
        movq    0x78(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        xorl    %r14d, %r14d
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r14, %r14
        movq    0x30(%rsi), %rbx
        xorl    %r15d, %r15d
        movq    0x60(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r8, %r8
        movq    0x68(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r8, %r8
        movq    0x70(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r8, %r8
        movq    0x78(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        adcq    %r15, %r15
        movq    0x38(%rsi), %rbx
        xorl    %r8d, %r8d
        movq    0x60(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r9, %r9
        movq    0x68(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r9, %r9
        movq    0x70(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %r9, %r9
        movq    0x78(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r8, %r8
        xorl    %r9d, %r9d
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        decq    %rbx
        adcq    %r13, %rbx
        decq    %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, 0x60(%rsp)
        movq    %r13, 0x68(%rsp)
        movq    %r14, 0x70(%rsp)
        movq    %r15, 0x78(%rsp)
        movq    0x80(%rsp), %rbx
        movq    0x20(%rsp), %rax
        mulq    %rbx
        movq    %rax, %r8
        movq    %rdx, %r9
        movq    0x28(%rsp), %rax
        mulq    %rbx
        xorl    %r10d, %r10d
        addq    %rax, %r9
        adcq    %rdx, %r10
        movq    0x30(%rsp), %rax
        mulq    %rbx
        xorl    %r11d, %r11d
        addq    %rax, %r10
        adcq    %rdx, %r11
        movq    0x38(%rsp), %rax
        mulq    %rbx
        xorl    %r12d, %r12d
        addq    %rax, %r11
        adcq    %rdx, %r12
        movq    0x88(%rsp), %rbx
        xorl    %r13d, %r13d
        movq    0x20(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r14, %r14
        movq    0x28(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r14, %r14
        movq    0x30(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r14, %r14
        movq    0x38(%rsp), %rax
        mulq    %rbx
        subq    %r14, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        xorl    %r14d, %r14d
        movabsq $0x100000000, %rbx
        movq    %r8, %rax
        mulq    %rbx
        addq    %rax, %r9
        adcq    %rdx, %r10
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r15, %r15
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r8, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r15, %r15
        movq    %r9, %rax
        mulq    %rbx
        subq    %r15, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        adcq    %r14, %r14
        movq    0x90(%rsp), %rbx
        xorl    %r15d, %r15d
        movq    0x20(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r10
        adcq    %rdx, %r11
        sbbq    %r8, %r8
        movq    0x28(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r8, %r8
        movq    0x30(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r8, %r8
        movq    0x38(%rsp), %rax
        mulq    %rbx
        subq    %r8, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        adcq    %r15, %r15
        movq    0x98(%rsp), %rbx
        xorl    %r8d, %r8d
        movq    0x20(%rsp), %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %r9, %r9
        movq    0x28(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %r9, %r9
        movq    0x30(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %r9, %r9
        movq    0x38(%rsp), %rax
        mulq    %rbx
        subq    %r9, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r8, %r8
        xorl    %r9d, %r9d
        movabsq $0x100000000, %rbx
        movq    %r10, %rax
        mulq    %rbx
        addq    %rax, %r11
        adcq    %rdx, %r12
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r12
        adcq    %rdx, %r13
        sbbq    %rcx, %rcx
        notq    %rbx
        leaq    0x2(%rbx), %rbx
        movq    %r10, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r13
        adcq    %rdx, %r14
        sbbq    %rcx, %rcx
        movq    %r11, %rax
        mulq    %rbx
        subq    %rcx, %rdx
        addq    %rax, %r14
        adcq    %rdx, %r15
        adcq    %r9, %r8
        movl    $0x1, %ecx
        addq    %r12, %rcx
        decq    %rbx
        adcq    %r13, %rbx
        decq    %r9
        movq    %r9, %rax
        adcq    %r14, %r9
        movl    $0xfffffffe, %r11d
        adcq    %r15, %r11
        adcq    %r8, %rax
        cmovbq  %rcx, %r12
        cmovbq  %rbx, %r13
        cmovbq  %r9, %r14
        cmovbq  %r11, %r15
        movq    %r12, 0x80(%rsp)
        movq    %r13, 0x88(%rsp)
        movq    %r14, 0x90(%rsp)
        movq    %r15, 0x98(%rsp)
        movq    0x80(%rsp), %rax
        subq    0x60(%rsp), %rax
        movq    0x88(%rsp), %rcx
        sbbq    0x68(%rsp), %rcx
        movq    0x90(%rsp), %r8
        sbbq    0x70(%rsp), %r8
        movq    0x98(%rsp), %r9
        sbbq    0x78(%rsp), %r9
        movl    $0xffffffff, %r10d
        sbbq    %r11, %r11
        xorq    %rdx, %rdx
        andq    %r11, %r10
        subq    %r10, %rdx
        addq    %r11, %rax
        movq    %rax, 0x80(%rsp)
        adcq    %r10, %rcx
        movq    %rcx, 0x88(%rsp)
        adcq    $0x0, %r8
        movq    %r8, 0x90(%rsp)
        adcq    %rdx, %r9
        movq    %r9, 0x98(%rsp)
        movq    0x40(%rsi), %rax
        movq    0x48(%rsi), %rdx
        orq     0x50(%rsi), %rax
        orq     0x58(%rsi), %rdx
        orq     %rdx, %rax
        movq    (%rsp), %r8
        movq    0x0(%rbp), %rax
        cmoveq  %rax, %r8
        movq    0x8(%rsp), %r9
        movq    0x8(%rbp), %rax
        cmoveq  %rax, %r9
        movq    0x10(%rsp), %r10
        movq    0x10(%rbp), %rax
        cmoveq  %rax, %r10
        movq    0x18(%rsp), %r11
        movq    0x18(%rbp), %rax
        cmoveq  %rax, %r11
        movq    0x80(%rsp), %r12
        movq    0x20(%rbp), %rax
        cmoveq  %rax, %r12
        movq    0x88(%rsp), %r13
        movq    0x28(%rbp), %rax
        cmoveq  %rax, %r13
        movq    0x90(%rsp), %r14
        movq    0x30(%rbp), %rax
        cmoveq  %rax, %r14
        movq    0x98(%rsp), %r15
        movq    0x38(%rbp), %rax
        cmoveq  %rax, %r15
        movq    %r8, (%rdi)
        movq    %r9, 0x8(%rdi)
        movq    %r10, 0x10(%rdi)
        movq    %r11, 0x18(%rdi)
        movq    %r12, 0x20(%rdi)
        movq    %r13, 0x28(%rdi)
        movq    %r14, 0x30(%rdi)
        movq    %r15, 0x38(%rdi)
        movq    0xa0(%rsp), %r8
        movq    0xa8(%rsp), %r9
        movq    0xb0(%rsp), %r10
        movq    0xb8(%rsp), %r11
        movl    $0x1, %eax
        cmoveq  %rax, %r8
        movabsq $0xffffffff00000000, %rax
        cmoveq  %rax, %r9
        movq    $0xffffffffffffffff, %rax
        cmoveq  %rax, %r10
        movl    $0xfffffffe, %eax
        cmoveq  %rax, %r11
        movq    %r8, 0x40(%rdi)
        movq    %r9, 0x48(%rdi)
        movq    %r10, 0x50(%rdi)
        movq    %r11, 0x58(%rdi)
        CFI_INC_RSP(192)
        CFI_POP(%r15)
        CFI_POP(%r14)
        CFI_POP(%r13)
        CFI_POP(%r12)
        CFI_POP(%rbp)
        CFI_POP(%rbx)
        CFI_RET

S2N_BN_SIZE_DIRECTIVE(Lp256_scalarmulbase_alt_local_p256_montjmixadd)

#if defined(__linux__) && defined(__ELF__)
.section .note.GNU-stack, "", %progbits
#endif