premiere-libtorrent/ed25519/src/sc.cpp

// ignore warnings in this file
#include "libtorrent/aux_/disable_warnings_push.hpp"

#include "fixedint.h"
#include "sc.h"

static u64 load_3(const unsigned char *in) {
    u64 result;

    result = (u64) in[0];
    result |= ((u64) in[1]) << 8;
    result |= ((u64) in[2]) << 16;

    return result;
}

static u64 load_4(const unsigned char *in) {
    u64 result;

    result = (u64) in[0];
    result |= ((u64) in[1]) << 8;
    result |= ((u64) in[2]) << 16;
    result |= ((u64) in[3]) << 24;
    
    return result;
}

static inline i64 shift_left(i64 v, int s) {
	return i64(u64(v) << s);
}

/*
Input:
  s[0]+256*s[1]+...+256^63*s[63] = s

Output:
  s[0]+256*s[1]+...+256^31*s[31] = s mod l
  where l = 2^252 + 27742317777372353535851937790883648493.
  Overwrites s in place.
*/

void sc_reduce(unsigned char *s) {
    i64 s0 = 2097151 & load_3(s);
    i64 s1 = 2097151 & (load_4(s + 2) >> 5);
    i64 s2 = 2097151 & (load_3(s + 5) >> 2);
    i64 s3 = 2097151 & (load_4(s + 7) >> 7);
    i64 s4 = 2097151 & (load_4(s + 10) >> 4);
    i64 s5 = 2097151 & (load_3(s + 13) >> 1);
    i64 s6 = 2097151 & (load_4(s + 15) >> 6);
    i64 s7 = 2097151 & (load_3(s + 18) >> 3);
    i64 s8 = 2097151 & load_3(s + 21);
    i64 s9 = 2097151 & (load_4(s + 23) >> 5);
    i64 s10 = 2097151 & (load_3(s + 26) >> 2);
    i64 s11 = 2097151 & (load_4(s + 28) >> 7);
    i64 s12 = 2097151 & (load_4(s + 31) >> 4);
    i64 s13 = 2097151 & (load_3(s + 34) >> 1);
    i64 s14 = 2097151 & (load_4(s + 36) >> 6);
    i64 s15 = 2097151 & (load_3(s + 39) >> 3);
    i64 s16 = 2097151 & load_3(s + 42);
    i64 s17 = 2097151 & (load_4(s + 44) >> 5);
    i64 s18 = 2097151 & (load_3(s + 47) >> 2);
    i64 s19 = 2097151 & (load_4(s + 49) >> 7);
    i64 s20 = 2097151 & (load_4(s + 52) >> 4);
    i64 s21 = 2097151 & (load_3(s + 55) >> 1);
    i64 s22 = 2097151 & (load_4(s + 57) >> 6);
    i64 s23 = (load_4(s + 60) >> 3);
    i64 carry0;
    i64 carry1;
    i64 carry2;
    i64 carry3;
    i64 carry4;
    i64 carry5;
    i64 carry6;
    i64 carry7;
    i64 carry8;
    i64 carry9;
    i64 carry10;
    i64 carry11;
    i64 carry12;
    i64 carry13;
    i64 carry14;
    i64 carry15;
    i64 carry16;

    s11 += s23 * 666643;
    s12 += s23 * 470296;
    s13 += s23 * 654183;
    s14 -= s23 * 997805;
    s15 += s23 * 136657;
    s16 -= s23 * 683901;
    s23 = 0; // NOLINT
    s10 += s22 * 666643;
    s11 += s22 * 470296;
    s12 += s22 * 654183;
    s13 -= s22 * 997805;
    s14 += s22 * 136657;
    s15 -= s22 * 683901;
    s22 = 0; // NOLINT
    s9 += s21 * 666643;
    s10 += s21 * 470296;
    s11 += s21 * 654183;
    s12 -= s21 * 997805;
    s13 += s21 * 136657;
    s14 -= s21 * 683901;
    s21 = 0; // NOLINT
    s8 += s20 * 666643;
    s9 += s20 * 470296;
    s10 += s20 * 654183;
    s11 -= s20 * 997805;
    s12 += s20 * 136657;
    s13 -= s20 * 683901;
    s20 = 0; // NOLINT
    s7 += s19 * 666643;
    s8 += s19 * 470296;
    s9 += s19 * 654183;
    s10 -= s19 * 997805;
    s11 += s19 * 136657;
    s12 -= s19 * 683901;
    s19 = 0; // NOLINT
    s6 += s18 * 666643;
    s7 += s18 * 470296;
    s8 += s18 * 654183;
    s9 -= s18 * 997805;
    s10 += s18 * 136657;
    s11 -= s18 * 683901;
    s18 = 0; // NOLINT
    carry6 = (s6 + (1 << 20)) >> 21;
    s7 += carry6;
    s6 -= shift_left(carry6, 21);
    carry8 = (s8 + (1 << 20)) >> 21;
    s9 += carry8;
    s8 -= shift_left(carry8, 21);
    carry10 = (s10 + (1 << 20)) >> 21;
    s11 += carry10;
    s10 -= shift_left(carry10, 21);
    carry12 = (s12 + (1 << 20)) >> 21;
    s13 += carry12;
    s12 -= shift_left(carry12, 21);
    carry14 = (s14 + (1 << 20)) >> 21;
    s15 += carry14;
    s14 -= shift_left(carry14, 21);
    carry16 = (s16 + (1 << 20)) >> 21;
    s17 += carry16;
    s16 -= shift_left(carry16, 21);
    carry7 = (s7 + (1 << 20)) >> 21;
    s8 += carry7;
    s7 -= shift_left(carry7, 21);
    carry9 = (s9 + (1 << 20)) >> 21;
    s10 += carry9;
    s9 -= shift_left(carry9, 21);
    carry11 = (s11 + (1 << 20)) >> 21;
    s12 += carry11;
    s11 -= shift_left(carry11, 21);
    carry13 = (s13 + (1 << 20)) >> 21;
    s14 += carry13;
    s13 -= shift_left(carry13, 21);
    carry15 = (s15 + (1 << 20)) >> 21;
    s16 += carry15;
    s15 -= shift_left(carry15, 21);
    s5 += s17 * 666643;
    s6 += s17 * 470296;
    s7 += s17 * 654183;
    s8 -= s17 * 997805;
    s9 += s17 * 136657;
    s10 -= s17 * 683901;
    s17 = 0; // NOLINT
    s4 += s16 * 666643;
    s5 += s16 * 470296;
    s6 += s16 * 654183;
    s7 -= s16 * 997805;
    s8 += s16 * 136657;
    s9 -= s16 * 683901;
    s16 = 0; // NOLINT
    s3 += s15 * 666643;
    s4 += s15 * 470296;
    s5 += s15 * 654183;
    s6 -= s15 * 997805;
    s7 += s15 * 136657;
    s8 -= s15 * 683901;
    s15 = 0; // NOLINT
    s2 += s14 * 666643;
    s3 += s14 * 470296;
    s4 += s14 * 654183;
    s5 -= s14 * 997805;
    s6 += s14 * 136657;
    s7 -= s14 * 683901;
    s14 = 0; // NOLINT
    s1 += s13 * 666643;
    s2 += s13 * 470296;
    s3 += s13 * 654183;
    s4 -= s13 * 997805;
    s5 += s13 * 136657;
    s6 -= s13 * 683901;
    s13 = 0; // NOLINT
    s0 += s12 * 666643;
    s1 += s12 * 470296;
    s2 += s12 * 654183;
    s3 -= s12 * 997805;
    s4 += s12 * 136657;
    s5 -= s12 * 683901;
    s12 = 0;
    carry0 = (s0 + (1 << 20)) >> 21;
    s1 += carry0;
    s0 -= shift_left(carry0, 21);
    carry2 = (s2 + (1 << 20)) >> 21;
    s3 += carry2;
    s2 -= shift_left(carry2, 21);
    carry4 = (s4 + (1 << 20)) >> 21;
    s5 += carry4;
    s4 -= shift_left(carry4, 21);
    carry6 = (s6 + (1 << 20)) >> 21;
    s7 += carry6;
    s6 -= shift_left(carry6, 21);
    carry8 = (s8 + (1 << 20)) >> 21;
    s9 += carry8;
    s8 -= shift_left(carry8, 21);
    carry10 = (s10 + (1 << 20)) >> 21;
    s11 += carry10;
    s10 -= shift_left(carry10, 21);
    carry1 = (s1 + (1 << 20)) >> 21;
    s2 += carry1;
    s1 -= shift_left(carry1, 21);
    carry3 = (s3 + (1 << 20)) >> 21;
    s4 += carry3;
    s3 -= shift_left(carry3, 21);
    carry5 = (s5 + (1 << 20)) >> 21;
    s6 += carry5;
    s5 -= shift_left(carry5, 21);
    carry7 = (s7 + (1 << 20)) >> 21;
    s8 += carry7;
    s7 -= shift_left(carry7, 21);
    carry9 = (s9 + (1 << 20)) >> 21;
    s10 += carry9;
    s9 -= shift_left(carry9, 21);
    carry11 = (s11 + (1 << 20)) >> 21;
    s12 += carry11;
    s11 -= shift_left(carry11, 21);
    s0 += s12 * 666643;
    s1 += s12 * 470296;
    s2 += s12 * 654183;
    s3 -= s12 * 997805;
    s4 += s12 * 136657;
    s5 -= s12 * 683901;
    s12 = 0;
    carry0 = s0 >> 21;
    s1 += carry0;
    s0 -= shift_left(carry0, 21);
    carry1 = s1 >> 21;
    s2 += carry1;
    s1 -= shift_left(carry1, 21);
    carry2 = s2 >> 21;
    s3 += carry2;
    s2 -= shift_left(carry2, 21);
    carry3 = s3 >> 21;
    s4 += carry3;
    s3 -= shift_left(carry3, 21);
    carry4 = s4 >> 21;
    s5 += carry4;
    s4 -= shift_left(carry4, 21);
    carry5 = s5 >> 21;
    s6 += carry5;
    s5 -= shift_left(carry5, 21);
    carry6 = s6 >> 21;
    s7 += carry6;
    s6 -= shift_left(carry6, 21);
    carry7 = s7 >> 21;
    s8 += carry7;
    s7 -= shift_left(carry7, 21);
    carry8 = s8 >> 21;
    s9 += carry8;
    s8 -= shift_left(carry8, 21);
    carry9 = s9 >> 21;
    s10 += carry9;
    s9 -= shift_left(carry9, 21);
    carry10 = s10 >> 21;
    s11 += carry10;
    s10 -= shift_left(carry10, 21);
    carry11 = s11 >> 21;
    s12 += carry11;
    s11 -= shift_left(carry11, 21);
    s0 += s12 * 666643;
    s1 += s12 * 470296;
    s2 += s12 * 654183;
    s3 -= s12 * 997805;
    s4 += s12 * 136657;
    s5 -= s12 * 683901;
    s12 = 0; // NOLINT
    carry0 = s0 >> 21;
    s1 += carry0;
    s0 -= shift_left(carry0, 21);
    carry1 = s1 >> 21;
    s2 += carry1;
    s1 -= shift_left(carry1, 21);
    carry2 = s2 >> 21;
    s3 += carry2;
    s2 -= shift_left(carry2, 21);
    carry3 = s3 >> 21;
    s4 += carry3;
    s3 -= shift_left(carry3, 21);
    carry4 = s4 >> 21;
    s5 += carry4;
    s4 -= shift_left(carry4, 21);
    carry5 = s5 >> 21;
    s6 += carry5;
    s5 -= shift_left(carry5, 21);
    carry6 = s6 >> 21;
    s7 += carry6;
    s6 -= shift_left(carry6, 21);
    carry7 = s7 >> 21;
    s8 += carry7;
    s7 -= shift_left(carry7, 21);
    carry8 = s8 >> 21;
    s9 += carry8;
    s8 -= shift_left(carry8, 21);
    carry9 = s9 >> 21;
    s10 += carry9;
    s9 -= shift_left(carry9, 21);
    carry10 = s10 >> 21;
    s11 += carry10;
    s10 -= shift_left(carry10, 21);

    s[0] = (unsigned char) ((s0 >> 0) & 0xff);
    s[1] = (unsigned char) ((s0 >> 8) & 0xff);
    s[2] = (unsigned char) (((s0 >> 16) | (s1 << 5)) & 0xff);
    s[3] = (unsigned char) ((s1 >> 3) & 0xff);
    s[4] = (unsigned char) ((s1 >> 11) & 0xff);
    s[5] = (unsigned char) (((s1 >> 19) | (s2 << 2)) & 0xff);
    s[6] = (unsigned char) ((s2 >> 6) & 0xff);
    s[7] = (unsigned char) (((s2 >> 14) | (s3 << 7)) & 0xff);
    s[8] = (unsigned char) ((s3 >> 1) & 0xff);
    s[9] = (unsigned char) ((s3 >> 9) & 0xff);
    s[10] = (unsigned char) (((s3 >> 17) | (s4 << 4)) & 0xff);
    s[11] = (unsigned char) ((s4 >> 4) & 0xff);
    s[12] = (unsigned char) ((s4 >> 12) & 0xff);
    s[13] = (unsigned char) (((s4 >> 20) | (s5 << 1)) & 0xff);
    s[14] = (unsigned char) ((s5 >> 7) & 0xff);
    s[15] = (unsigned char) (((s5 >> 15) | (s6 << 6)) & 0xff);
    s[16] = (unsigned char) ((s6 >> 2) & 0xff);
    s[17] = (unsigned char) ((s6 >> 10) & 0xff);
    s[18] = (unsigned char) (((s6 >> 18) | (s7 << 3)) & 0xff);
    s[19] = (unsigned char) ((s7 >> 5) & 0xff);
    s[20] = (unsigned char) ((s7 >> 13) & 0xff);
    s[21] = (unsigned char) ((s8 >> 0) & 0xff);
    s[22] = (unsigned char) ((s8 >> 8) & 0xff);
    s[23] = (unsigned char) (((s8 >> 16) | (s9 << 5)) & 0xff);
    s[24] = (unsigned char) ((s9 >> 3) & 0xff);
    s[25] = (unsigned char) ((s9 >> 11) & 0xff);
    s[26] = (unsigned char) (((s9 >> 19) | (s10 << 2)) & 0xff);
    s[27] = (unsigned char) ((s10 >> 6) & 0xff);
    s[28] = (unsigned char) (((s10 >> 14) | (s11 << 7)) & 0xff);
    s[29] = (unsigned char) ((s11 >> 1) & 0xff);
    s[30] = (unsigned char) ((s11 >> 9) & 0xff);
    s[31] = (unsigned char) ((s11 >> 17) & 0xff);
}



/*
Input:
  a[0]+256*a[1]+...+256^31*a[31] = a
  b[0]+256*b[1]+...+256^31*b[31] = b
  c[0]+256*c[1]+...+256^31*c[31] = c

Output:
  s[0]+256*s[1]+...+256^31*s[31] = (ab+c) mod l
  where l = 2^252 + 27742317777372353535851937790883648493.
*/

void sc_muladd(unsigned char *s, const unsigned char *a, const unsigned char *b, const unsigned char *c) {
    i64 a0 = 2097151 & load_3(a);
    i64 a1 = 2097151 & (load_4(a + 2) >> 5);
    i64 a2 = 2097151 & (load_3(a + 5) >> 2);
    i64 a3 = 2097151 & (load_4(a + 7) >> 7);
    i64 a4 = 2097151 & (load_4(a + 10) >> 4);
    i64 a5 = 2097151 & (load_3(a + 13) >> 1);
    i64 a6 = 2097151 & (load_4(a + 15) >> 6);
    i64 a7 = 2097151 & (load_3(a + 18) >> 3);
    i64 a8 = 2097151 & load_3(a + 21);
    i64 a9 = 2097151 & (load_4(a + 23) >> 5);
    i64 a10 = 2097151 & (load_3(a + 26) >> 2);
    i64 a11 = (load_4(a + 28) >> 7);
    i64 b0 = 2097151 & load_3(b);
    i64 b1 = 2097151 & (load_4(b + 2) >> 5);
    i64 b2 = 2097151 & (load_3(b + 5) >> 2);
    i64 b3 = 2097151 & (load_4(b + 7) >> 7);
    i64 b4 = 2097151 & (load_4(b + 10) >> 4);
    i64 b5 = 2097151 & (load_3(b + 13) >> 1);
    i64 b6 = 2097151 & (load_4(b + 15) >> 6);
    i64 b7 = 2097151 & (load_3(b + 18) >> 3);
    i64 b8 = 2097151 & load_3(b + 21);
    i64 b9 = 2097151 & (load_4(b + 23) >> 5);
    i64 b10 = 2097151 & (load_3(b + 26) >> 2);
    i64 b11 = (load_4(b + 28) >> 7);
    i64 c0 = 2097151 & load_3(c);
    i64 c1 = 2097151 & (load_4(c + 2) >> 5);
    i64 c2 = 2097151 & (load_3(c + 5) >> 2);
    i64 c3 = 2097151 & (load_4(c + 7) >> 7);
    i64 c4 = 2097151 & (load_4(c + 10) >> 4);
    i64 c5 = 2097151 & (load_3(c + 13) >> 1);
    i64 c6 = 2097151 & (load_4(c + 15) >> 6);
    i64 c7 = 2097151 & (load_3(c + 18) >> 3);
    i64 c8 = 2097151 & load_3(c + 21);
    i64 c9 = 2097151 & (load_4(c + 23) >> 5);
    i64 c10 = 2097151 & (load_3(c + 26) >> 2);
    i64 c11 = (load_4(c + 28) >> 7);
    i64 s0;
    i64 s1;
    i64 s2;
    i64 s3;
    i64 s4;
    i64 s5;
    i64 s6;
    i64 s7;
    i64 s8;
    i64 s9;
    i64 s10;
    i64 s11;
    i64 s12;
    i64 s13;
    i64 s14;
    i64 s15;
    i64 s16;
    i64 s17;
    i64 s18;
    i64 s19;
    i64 s20;
    i64 s21;
    i64 s22;
    i64 s23;
    i64 carry0;
    i64 carry1;
    i64 carry2;
    i64 carry3;
    i64 carry4;
    i64 carry5;
    i64 carry6;
    i64 carry7;
    i64 carry8;
    i64 carry9;
    i64 carry10;
    i64 carry11;
    i64 carry12;
    i64 carry13;
    i64 carry14;
    i64 carry15;
    i64 carry16;
    i64 carry17;
    i64 carry18;
    i64 carry19;
    i64 carry20;
    i64 carry21;
    i64 carry22;

    s0 = c0 + a0 * b0;
    s1 = c1 + a0 * b1 + a1 * b0;
    s2 = c2 + a0 * b2 + a1 * b1 + a2 * b0;
    s3 = c3 + a0 * b3 + a1 * b2 + a2 * b1 + a3 * b0;
    s4 = c4 + a0 * b4 + a1 * b3 + a2 * b2 + a3 * b1 + a4 * b0;
    s5 = c5 + a0 * b5 + a1 * b4 + a2 * b3 + a3 * b2 + a4 * b1 + a5 * b0;
    s6 = c6 + a0 * b6 + a1 * b5 + a2 * b4 + a3 * b3 + a4 * b2 + a5 * b1 + a6 * b0;
    s7 = c7 + a0 * b7 + a1 * b6 + a2 * b5 + a3 * b4 + a4 * b3 + a5 * b2 + a6 * b1 + a7 * b0;
    s8 = c8 + a0 * b8 + a1 * b7 + a2 * b6 + a3 * b5 + a4 * b4 + a5 * b3 + a6 * b2 + a7 * b1 + a8 * b0;
    s9 = c9 + a0 * b9 + a1 * b8 + a2 * b7 + a3 * b6 + a4 * b5 + a5 * b4 + a6 * b3 + a7 * b2 + a8 * b1 + a9 * b0;
    s10 = c10 + a0 * b10 + a1 * b9 + a2 * b8 + a3 * b7 + a4 * b6 + a5 * b5 + a6 * b4 + a7 * b3 + a8 * b2 + a9 * b1 + a10 * b0;
    s11 = c11 + a0 * b11 + a1 * b10 + a2 * b9 + a3 * b8 + a4 * b7 + a5 * b6 + a6 * b5 + a7 * b4 + a8 * b3 + a9 * b2 + a10 * b1 + a11 * b0;
    s12 = a1 * b11 + a2 * b10 + a3 * b9 + a4 * b8 + a5 * b7 + a6 * b6 + a7 * b5 + a8 * b4 + a9 * b3 + a10 * b2 + a11 * b1;
    s13 = a2 * b11 + a3 * b10 + a4 * b9 + a5 * b8 + a6 * b7 + a7 * b6 + a8 * b5 + a9 * b4 + a10 * b3 + a11 * b2;
    s14 = a3 * b11 + a4 * b10 + a5 * b9 + a6 * b8 + a7 * b7 + a8 * b6 + a9 * b5 + a10 * b4 + a11 * b3;
    s15 = a4 * b11 + a5 * b10 + a6 * b9 + a7 * b8 + a8 * b7 + a9 * b6 + a10 * b5 + a11 * b4;
    s16 = a5 * b11 + a6 * b10 + a7 * b9 + a8 * b8 + a9 * b7 + a10 * b6 + a11 * b5;
    s17 = a6 * b11 + a7 * b10 + a8 * b9 + a9 * b8 + a10 * b7 + a11 * b6;
    s18 = a7 * b11 + a8 * b10 + a9 * b9 + a10 * b8 + a11 * b7;
    s19 = a8 * b11 + a9 * b10 + a10 * b9 + a11 * b8;
    s20 = a9 * b11 + a10 * b10 + a11 * b9;
    s21 = a10 * b11 + a11 * b10;
    s22 = a11 * b11;
    s23 = 0;
    carry0 = (s0 + (1 << 20)) >> 21;
    s1 += carry0;
    s0 -= shift_left(carry0, 21);
    carry2 = (s2 + (1 << 20)) >> 21;
    s3 += carry2;
    s2 -= shift_left(carry2, 21);
    carry4 = (s4 + (1 << 20)) >> 21;
    s5 += carry4;
    s4 -= shift_left(carry4, 21);
    carry6 = (s6 + (1 << 20)) >> 21;
    s7 += carry6;
    s6 -= shift_left(carry6, 21);
    carry8 = (s8 + (1 << 20)) >> 21;
    s9 += carry8;
    s8 -= shift_left(carry8, 21);
    carry10 = (s10 + (1 << 20)) >> 21;
    s11 += carry10;
    s10 -= shift_left(carry10, 21);
    carry12 = (s12 + (1 << 20)) >> 21;
    s13 += carry12;
    s12 -= shift_left(carry12, 21);
    carry14 = (s14 + (1 << 20)) >> 21;
    s15 += carry14;
    s14 -= shift_left(carry14, 21);
    carry16 = (s16 + (1 << 20)) >> 21;
    s17 += carry16;
    s16 -= shift_left(carry16, 21);
    carry18 = (s18 + (1 << 20)) >> 21;
    s19 += carry18;
    s18 -= shift_left(carry18, 21);
    carry20 = (s20 + (1 << 20)) >> 21;
    s21 += carry20;
    s20 -= shift_left(carry20, 21);
    carry22 = (s22 + (1 << 20)) >> 21;
    s23 += carry22;
    s22 -= shift_left(carry22, 21);
    carry1 = (s1 + (1 << 20)) >> 21;
    s2 += carry1;
    s1 -= shift_left(carry1, 21);
    carry3 = (s3 + (1 << 20)) >> 21;
    s4 += carry3;
    s3 -= shift_left(carry3, 21);
    carry5 = (s5 + (1 << 20)) >> 21;
    s6 += carry5;
    s5 -= shift_left(carry5, 21);
    carry7 = (s7 + (1 << 20)) >> 21;
    s8 += carry7;
    s7 -= shift_left(carry7, 21);
    carry9 = (s9 + (1 << 20)) >> 21;
    s10 += carry9;
    s9 -= shift_left(carry9, 21);
    carry11 = (s11 + (1 << 20)) >> 21;
    s12 += carry11;
    s11 -= shift_left(carry11, 21);
    carry13 = (s13 + (1 << 20)) >> 21;
    s14 += carry13;
    s13 -= shift_left(carry13, 21);
    carry15 = (s15 + (1 << 20)) >> 21;
    s16 += carry15;
    s15 -= shift_left(carry15, 21);
    carry17 = (s17 + (1 << 20)) >> 21;
    s18 += carry17;
    s17 -= shift_left(carry17, 21);
    carry19 = (s19 + (1 << 20)) >> 21;
    s20 += carry19;
    s19 -= shift_left(carry19, 21);
    carry21 = (s21 + (1 << 20)) >> 21;
    s22 += carry21;
    s21 -= shift_left(carry21, 21);
    s11 += s23 * 666643;
    s12 += s23 * 470296;
    s13 += s23 * 654183;
    s14 -= s23 * 997805;
    s15 += s23 * 136657;
    s16 -= s23 * 683901;
    s23 = 0; // NOLINT
    s10 += s22 * 666643;
    s11 += s22 * 470296;
    s12 += s22 * 654183;
    s13 -= s22 * 997805;
    s14 += s22 * 136657;
    s15 -= s22 * 683901;
    s22 = 0; // NOLINT
    s9 += s21 * 666643;
    s10 += s21 * 470296;
    s11 += s21 * 654183;
    s12 -= s21 * 997805;
    s13 += s21 * 136657;
    s14 -= s21 * 683901;
    s21 = 0; // NOLINT
    s8 += s20 * 666643;
    s9 += s20 * 470296;
    s10 += s20 * 654183;
    s11 -= s20 * 997805;
    s12 += s20 * 136657;
    s13 -= s20 * 683901;
    s20 = 0; // NOLINT
    s7 += s19 * 666643;
    s8 += s19 * 470296;
    s9 += s19 * 654183;
    s10 -= s19 * 997805;
    s11 += s19 * 136657;
    s12 -= s19 * 683901;
    s19 = 0; // NOLINT
    s6 += s18 * 666643;
    s7 += s18 * 470296;
    s8 += s18 * 654183;
    s9 -= s18 * 997805;
    s10 += s18 * 136657;
    s11 -= s18 * 683901;
    s18 = 0; // NOLINT
    carry6 = (s6 + (1 << 20)) >> 21;
    s7 += carry6;
    s6 -= shift_left(carry6, 21);
    carry8 = (s8 + (1 << 20)) >> 21;
    s9 += carry8;
    s8 -= shift_left(carry8, 21);
    carry10 = (s10 + (1 << 20)) >> 21;
    s11 += carry10;
    s10 -= shift_left(carry10, 21);
    carry12 = (s12 + (1 << 20)) >> 21;
    s13 += carry12;
    s12 -= shift_left(carry12, 21);
    carry14 = (s14 + (1 << 20)) >> 21;
    s15 += carry14;
    s14 -= shift_left(carry14, 21);
    carry16 = (s16 + (1 << 20)) >> 21;
    s17 += carry16;
    s16 -= shift_left(carry16, 21);
    carry7 = (s7 + (1 << 20)) >> 21;
    s8 += carry7;
    s7 -= shift_left(carry7, 21);
    carry9 = (s9 + (1 << 20)) >> 21;
    s10 += carry9;
    s9 -= shift_left(carry9, 21);
    carry11 = (s11 + (1 << 20)) >> 21;
    s12 += carry11;
    s11 -= shift_left(carry11, 21);
    carry13 = (s13 + (1 << 20)) >> 21;
    s14 += carry13;
    s13 -= shift_left(carry13, 21);
    carry15 = (s15 + (1 << 20)) >> 21;
    s16 += carry15;
    s15 -= shift_left(carry15, 21);
    s5 += s17 * 666643;
    s6 += s17 * 470296;
    s7 += s17 * 654183;
    s8 -= s17 * 997805;
    s9 += s17 * 136657;
    s10 -= s17 * 683901;
    s17 = 0; // NOLINT
    s4 += s16 * 666643;
    s5 += s16 * 470296;
    s6 += s16 * 654183;
    s7 -= s16 * 997805;
    s8 += s16 * 136657;
    s9 -= s16 * 683901;
    s16 = 0; // NOLINT
    s3 += s15 * 666643;
    s4 += s15 * 470296;
    s5 += s15 * 654183;
    s6 -= s15 * 997805;
    s7 += s15 * 136657;
    s8 -= s15 * 683901;
    s15 = 0; // NOLINT
    s2 += s14 * 666643;
    s3 += s14 * 470296;
    s4 += s14 * 654183;
    s5 -= s14 * 997805;
    s6 += s14 * 136657;
    s7 -= s14 * 683901;
    s14 = 0; // NOLINT
    s1 += s13 * 666643;
    s2 += s13 * 470296;
    s3 += s13 * 654183;
    s4 -= s13 * 997805;
    s5 += s13 * 136657;
    s6 -= s13 * 683901;
    s13 = 0; // NOLINT
    s0 += s12 * 666643;
    s1 += s12 * 470296;
    s2 += s12 * 654183;
    s3 -= s12 * 997805;
    s4 += s12 * 136657;
    s5 -= s12 * 683901;
    s12 = 0; // NOLINT
    carry0 = (s0 + (1 << 20)) >> 21;
    s1 += carry0;
    s0 -= shift_left(carry0, 21);
    carry2 = (s2 + (1 << 20)) >> 21;
    s3 += carry2;
    s2 -= shift_left(carry2, 21);
    carry4 = (s4 + (1 << 20)) >> 21;
    s5 += carry4;
    s4 -= shift_left(carry4, 21);
    carry6 = (s6 + (1 << 20)) >> 21;
    s7 += carry6;
    s6 -= shift_left(carry6, 21);
    carry8 = (s8 + (1 << 20)) >> 21;
    s9 += carry8;
    s8 -= shift_left(carry8, 21);
    carry10 = (s10 + (1 << 20)) >> 21;
    s11 += carry10;
    s10 -= shift_left(carry10, 21);
    carry1 = (s1 + (1 << 20)) >> 21;
    s2 += carry1;
    s1 -= shift_left(carry1, 21);
    carry3 = (s3 + (1 << 20)) >> 21;
    s4 += carry3;
    s3 -= shift_left(carry3, 21);
    carry5 = (s5 + (1 << 20)) >> 21;
    s6 += carry5;
    s5 -= shift_left(carry5, 21);
    carry7 = (s7 + (1 << 20)) >> 21;
    s8 += carry7;
    s7 -= shift_left(carry7, 21);
    carry9 = (s9 + (1 << 20)) >> 21;
    s10 += carry9;
    s9 -= shift_left(carry9, 21);
    carry11 = (s11 + (1 << 20)) >> 21;
    s12 += carry11;
    s11 -= shift_left(carry11, 21);
    s0 += s12 * 666643;
    s1 += s12 * 470296;
    s2 += s12 * 654183;
    s3 -= s12 * 997805;
    s4 += s12 * 136657;
    s5 -= s12 * 683901;
    s12 = 0;
    carry0 = s0 >> 21;
    s1 += carry0;
    s0 -= shift_left(carry0, 21);
    carry1 = s1 >> 21;
    s2 += carry1;
    s1 -= shift_left(carry1, 21);
    carry2 = s2 >> 21;
    s3 += carry2;
    s2 -= shift_left(carry2, 21);
    carry3 = s3 >> 21;
    s4 += carry3;
    s3 -= shift_left(carry3, 21);
    carry4 = s4 >> 21;
    s5 += carry4;
    s4 -= shift_left(carry4, 21);
    carry5 = s5 >> 21;
    s6 += carry5;
    s5 -= shift_left(carry5, 21);
    carry6 = s6 >> 21;
    s7 += carry6;
    s6 -= shift_left(carry6, 21);
    carry7 = s7 >> 21;
    s8 += carry7;
    s7 -= shift_left(carry7, 21);
    carry8 = s8 >> 21;
    s9 += carry8;
    s8 -= shift_left(carry8, 21);
    carry9 = s9 >> 21;
    s10 += carry9;
    s9 -= shift_left(carry9, 21);
    carry10 = s10 >> 21;
    s11 += carry10;
    s10 -= shift_left(carry10, 21);
    carry11 = s11 >> 21;
    s12 += carry11;
    s11 -= shift_left(carry11, 21);
    s0 += s12 * 666643;
    s1 += s12 * 470296;
    s2 += s12 * 654183;
    s3 -= s12 * 997805;
    s4 += s12 * 136657;
    s5 -= s12 * 683901;
    s12 = 0; // NOLINT
    carry0 = s0 >> 21;
    s1 += carry0;
    s0 -= shift_left(carry0, 21);
    carry1 = s1 >> 21;
    s2 += carry1;
    s1 -= shift_left(carry1, 21);
    carry2 = s2 >> 21;
    s3 += carry2;
    s2 -= shift_left(carry2, 21);
    carry3 = s3 >> 21;
    s4 += carry3;
    s3 -= shift_left(carry3, 21);
    carry4 = s4 >> 21;
    s5 += carry4;
    s4 -= shift_left(carry4, 21);
    carry5 = s5 >> 21;
    s6 += carry5;
    s5 -= shift_left(carry5, 21);
    carry6 = s6 >> 21;
    s7 += carry6;
    s6 -= shift_left(carry6, 21);
    carry7 = s7 >> 21;
    s8 += carry7;
    s7 -= shift_left(carry7, 21);
    carry8 = s8 >> 21;
    s9 += carry8;
    s8 -= shift_left(carry8, 21);
    carry9 = s9 >> 21;
    s10 += carry9;
    s9 -= shift_left(carry9, 21);
    carry10 = s10 >> 21;
    s11 += carry10;
    s10 -= shift_left(carry10, 21);
    
    s[0] = (unsigned char) ((s0 >> 0) & 0xff);
    s[1] = (unsigned char) ((s0 >> 8) & 0xff);
    s[2] = (unsigned char) (((s0 >> 16) | (s1 << 5)) & 0xff);
    s[3] = (unsigned char) ((s1 >> 3) & 0xff);
    s[4] = (unsigned char) ((s1 >> 11) & 0xff);
    s[5] = (unsigned char) (((s1 >> 19) | (s2 << 2)) & 0xff);
    s[6] = (unsigned char) ((s2 >> 6) & 0xff);
    s[7] = (unsigned char) (((s2 >> 14) | (s3 << 7)) & 0xff);
    s[8] = (unsigned char) ((s3 >> 1) & 0xff);
    s[9] = (unsigned char) ((s3 >> 9) & 0xff);
    s[10] = (unsigned char) (((s3 >> 17) | (s4 << 4)) & 0xff);
    s[11] = (unsigned char) ((s4 >> 4) & 0xff);
    s[12] = (unsigned char) ((s4 >> 12) & 0xff);
    s[13] = (unsigned char) (((s4 >> 20) | (s5 << 1)) & 0xff);
    s[14] = (unsigned char) ((s5 >> 7) & 0xff);
    s[15] = (unsigned char) (((s5 >> 15) | (s6 << 6)) & 0xff);
    s[16] = (unsigned char) ((s6 >> 2) & 0xff);
    s[17] = (unsigned char) ((s6 >> 10) & 0xff);
    s[18] = (unsigned char) (((s6 >> 18) | (s7 << 3)) & 0xff);
    s[19] = (unsigned char) ((s7 >> 5) & 0xff);
    s[20] = (unsigned char) ((s7 >> 13) & 0xff);
    s[21] = (unsigned char) ((s8 >> 0) & 0xff);
    s[22] = (unsigned char) ((s8 >> 8) & 0xff);
    s[23] = (unsigned char) (((s8 >> 16) | (s9 << 5)) & 0xff);
    s[24] = (unsigned char) ((s9 >> 3) & 0xff);
    s[25] = (unsigned char) ((s9 >> 11) & 0xff);
    s[26] = (unsigned char) (((s9 >> 19) | (s10 << 2)) & 0xff);
    s[27] = (unsigned char) ((s10 >> 6) & 0xff);
    s[28] = (unsigned char) (((s10 >> 14) | (s11 << 7)) & 0xff);
    s[29] = (unsigned char) ((s11 >> 1) & 0xff);
    s[30] = (unsigned char) ((s11 >> 9) & 0xff);
    s[31] = (unsigned char) ((s11 >> 17) & 0xff);
}