#include "target.h"
#include <emmintrin.h>
#include <iostream>
#include <iomanip>
#include <string.h>
#include <stdlib.h>
#include <time.h>
using namespace std;
#include "cpucycles.h"

long long innerloopcycles;

class uint32x4 {
  __m128i x;
public:
  inline uint32x4() { }
  inline uint32x4(unsigned int u) { x = _mm_set_epi32(u,u,u,u); }
  inline uint32x4(__m128i u) { x = u; }
  inline uint32x4(const uint32x4 &a) { x = a.x; }
  unsigned int uint0() const { return _mm_cvtsi128_si32(x); }
  unsigned int uint1() const { return _mm_cvtsi128_si32(_mm_shuffle_epi32(x,0x39)); }
  unsigned int uint2() const { return _mm_cvtsi128_si32(_mm_shuffle_epi32(x,0x4e)); }
  unsigned int uint3() const { return _mm_cvtsi128_si32(_mm_shuffle_epi32(x,0x93)); }
  friend inline uint32x4 operator+(uint32x4 a,uint32x4 b) { return _mm_add_epi32(a.x,b.x); }
  friend inline uint32x4 operator|(uint32x4 a,uint32x4 b) { return _mm_or_si128(a.x,b.x); }
  friend inline uint32x4 operator&(uint32x4 a,uint32x4 b) { return _mm_and_si128(a.x,b.x); }
  friend inline uint32x4 operator^(uint32x4 a,uint32x4 b) { return _mm_xor_si128(a.x,b.x); }
  friend inline uint32x4 operator>>(uint32x4 a,int b) { return _mm_srli_epi32(a.x,b); }
  friend inline uint32x4 andnot(uint32x4 a,uint32x4 b) { return _mm_andnot_si128(b.x,a.x); }
  friend inline uint32x4 rotate1(uint32x4 a) {
    return _mm_xor_si128(_mm_slli_epi32(a.x,1),_mm_srli_epi32(a.x,31)); }
  friend inline uint32x4 rotate5(uint32x4 a) {
    return _mm_xor_si128(_mm_slli_epi32(a.x,5),_mm_srli_epi32(a.x,27)); }
  friend inline uint32x4 rotate30(uint32x4 a) {
    return _mm_xor_si128(_mm_slli_epi32(a.x,30),_mm_srli_epi32(a.x,2)); }
  friend ostream& operator<<(ostream& o,const uint32x4& u) {
    unsigned int r = u.uint0();
    o << hex << setw(2) << setfill('0') << ((r >> 24) & 255);
    o << hex << setw(2) << setfill('0') << ((r >> 16) & 255);
    o << hex << setw(2) << setfill('0') << ((r >> 8) & 255);
    o << hex << setw(2) << setfill('0') << (r & 255);
    return o;
  }
} ;

class hash {
  uint32x4 state[5];
public:
  hash() { }
  hash(const hash &x) {
    state[0] = x.state[0];
    state[1] = x.state[1];
    state[2] = x.state[2];
    state[3] = x.state[3];
    state[4] = x.state[4];
  }
  void init()
  {
    state[0] = 0x67452301;
    state[1] = 0xefcdab89;
    state[2] = 0x98badcfe;
    state[3] = 0x10325476;
    state[4] = 0xc3d2e1f0;
  }
  uint32x4 hammingdistance(hash b) {
    uint32x4 x0 = state[0] ^ b.state[0];
    uint32x4 x1 = state[1] ^ b.state[1];
    uint32x4 x2 = state[2] ^ b.state[2];
    uint32x4 x3 = state[3] ^ b.state[3];
    uint32x4 x4 = state[4] ^ b.state[4];
    uint32x4 mask;
    // 32 1-bit chunks
    mask = 0x55555555;
    x0 = (x0 & mask) + ((x0 >> 1) & mask);
    x1 = (x1 & mask) + ((x1 >> 1) & mask);
    x2 = (x2 & mask) + ((x2 >> 1) & mask);
    x3 = (x3 & mask) + ((x3 >> 1) & mask);
    x4 = (x4 & mask) + ((x4 >> 1) & mask);
    // 16 2-bit chunks: 012,012,012,012,012,012,012,012,012,012,012,012,012,012,012,012
    mask = 0x33333333;
    x0 = (x0 & mask) + ((x0 >> 2) & mask);
    x1 = (x1 & mask) + ((x1 >> 2) & mask);
    x2 = (x2 & mask) + ((x2 >> 2) & mask);
    x3 = (x3 & mask) + ((x3 >> 2) & mask);
    x4 = (x4 & mask) + ((x4 >> 2) & mask);
    // 8 4-bit chunks: 01234,01234,01234,01234,01234,01234,01234,01234
    mask = 0x0f0f0f0f;
    x0 = (x0 & mask) + ((x0 >> 4) & mask);
    x1 = (x1 & mask) + ((x1 >> 4) & mask);
    x2 = (x2 & mask) + ((x2 >> 4) & mask);
    x3 = (x3 & mask) + ((x3 >> 4) & mask);
    x4 = (x4 & mask) + ((x4 >> 4) & mask);
    // 4 8-bit chunks: 012345678,012345678,012345678,012345678
    mask = 0x00ff00ff;
    x0 = (x0 & mask) + ((x0 >> 8) & mask);
    x1 = (x1 & mask) + ((x1 >> 8) & mask);
    x2 = (x2 & mask) + ((x2 >> 8) & mask);
    x3 = (x3 & mask) + ((x3 >> 8) & mask);
    x4 = (x4 & mask) + ((x4 >> 8) & mask);
    mask = 0x0000ffff;
    x0 = (x0 & mask) + ((x0 >> 16) & mask);
    x1 = (x1 & mask) + ((x1 >> 16) & mask);
    x2 = (x2 & mask) + ((x2 >> 16) & mask);
    x3 = (x3 & mask) + ((x3 >> 16) & mask);
    x4 = (x4 & mask) + ((x4 >> 16) & mask);
    return x0 + x1 + x2 + x3 + x4;
  }
  friend ostream& operator<<(ostream& o,const hash& h) {
    o << h.state[0];
    o << h.state[1];
    o << h.state[2];
    o << h.state[3];
    o << h.state[4];
    return o;
  }

#define X(x0,x2,x8,x13) \
  x0 = x0 ^ x8; \
  x0 = x0 ^ x13; \
  x0 = x0 ^ x2; \
  x0 = rotate1(x0);

#define R1(x0,a,b,c,d,e) \
  e = e + delta; \
  e = e + (c & b); \
  e = e + andnot(d,b); \
  e = e + x0; \
  b = rotate30(b); \
  e = e + rotate5(a); \

#define R2(x0,a,b,c,d,e) \
  e = e + delta; \
  e = e + (b ^ c ^ d); \
  e = e + x0; \
  b = rotate30(b); \
  e = e + rotate5(a); \

#define R3(x0,a,b,c,d,e) \
  e = e + delta; \
  e = e + (((b | d) & c) | (b & d)); \
  e = e + x0; \
  b = rotate30(b); \
  e = e + rotate5(a); \

#define R10(a,b,c,d,e) \
  e = e + delta; \
  e = e + (c & b); \
  e = e + andnot(d,b); \
  b = rotate30(b); \
  e = e + rotate5(a); \

#define R20(a,b,c,d,e) \
  e = e + delta; \
  e = e + (b ^ c ^ d); \
  b = rotate30(b); \
  e = e + rotate5(a); \

#define R30(a,b,c,d,e) \
  e = e + delta; \
  e = e + (((b | d) & c) | (b & d)); \
  b = rotate30(b); \
  e = e + rotate5(a); \

  void update(const uint32x4 *in,unsigned long long inblocks) {
    uint32x4 a = state[0];
    uint32x4 b = state[1];
    uint32x4 c = state[2];
    uint32x4 d = state[3];
    uint32x4 e = state[4];
    uint32x4 delta;
    uint32x4 x0,x1,x2,x3,x4,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15;
  
    while (inblocks > 0) {
      delta = 0x5a827999;
      x0 = in[0]; R1(x0,a,b,c,d,e)
      x1 = in[1]; R1(x1,e,a,b,c,d)
      x2 = in[2]; R1(x2,d,e,a,b,c)
      x3 = in[3]; R1(x3,c,d,e,a,b)
      x4 = in[4]; R1(x4,b,c,d,e,a)
      x5 = in[5]; R1(x5,a,b,c,d,e)
      x6 = in[6]; R1(x6,e,a,b,c,d)
      x7 = in[7]; R1(x7,d,e,a,b,c)
      x8 = in[8]; R1(x8,c,d,e,a,b)
      x9 = in[9]; R1(x9,b,c,d,e,a)
      x10 = in[10]; R1(x10,a,b,c,d,e)
      x11 = in[11]; R1(x11,e,a,b,c,d)
      x12 = in[12]; R1(x12,d,e,a,b,c)
      x13 = in[13]; R1(x13,c,d,e,a,b)
      x14 = in[14]; R1(x14,b,c,d,e,a)
      x15 = in[15]; R1(x15,a,b,c,d,e)
      X(x0,x2,x8,x13)  R1(x0,e,a,b,c,d)
      X(x1,x3,x9,x14)  R1(x1,d,e,a,b,c)
      X(x2,x4,x10,x15) R1(x2,c,d,e,a,b)
      X(x3,x5,x11,x0)  R1(x3,b,c,d,e,a)
      delta = 0x6ed9eba1;
      X(x4,x6,x12,x1) R2(x4,a,b,c,d,e)
      X(x5,x7,x13,x2) R2(x5,e,a,b,c,d)
      X(x6,x8,x14,x3) R2(x6,d,e,a,b,c)
      X(x7,x9,x15,x4) R2(x7,c,d,e,a,b)
      X(x8,x10,x0,x5) R2(x8,b,c,d,e,a)
      X(x9,x11,x1,x6) R2(x9,a,b,c,d,e)
      X(x10,x12,x2,x7) R2(x10,e,a,b,c,d)
      X(x11,x13,x3,x8) R2(x11,d,e,a,b,c)
      X(x12,x14,x4,x9) R2(x12,c,d,e,a,b)
      X(x13,x15,x5,x10) R2(x13,b,c,d,e,a)
      X(x14,x0,x6,x11) R2(x14,a,b,c,d,e)
      X(x15,x1,x7,x12) R2(x15,e,a,b,c,d)
      X(x0,x2,x8,x13) R2(x0,d,e,a,b,c)
      X(x1,x3,x9,x14) R2(x1,c,d,e,a,b)
      X(x2,x4,x10,x15) R2(x2,b,c,d,e,a)
      X(x3,x5,x11,x0) R2(x3,a,b,c,d,e)
      X(x4,x6,x12,x1) R2(x4,e,a,b,c,d)
      X(x5,x7,x13,x2) R2(x5,d,e,a,b,c)
      X(x6,x8,x14,x3) R2(x6,c,d,e,a,b)
      X(x7,x9,x15,x4) R2(x7,b,c,d,e,a)
      delta = 0x8f1bbcdc;
      X(x8,x10,x0,x5) R3(x8,a,b,c,d,e)
      X(x9,x11,x1,x6) R3(x9,e,a,b,c,d)
      X(x10,x12,x2,x7) R3(x10,d,e,a,b,c)
      X(x11,x13,x3,x8) R3(x11,c,d,e,a,b)
      X(x12,x14,x4,x9) R3(x12,b,c,d,e,a)
      X(x13,x15,x5,x10) R3(x13,a,b,c,d,e)
      X(x14,x0,x6,x11) R3(x14,e,a,b,c,d)
      X(x15,x1,x7,x12) R3(x15,d,e,a,b,c)
      X(x0,x2,x8,x13) R3(x0,c,d,e,a,b)
      X(x1,x3,x9,x14) R3(x1,b,c,d,e,a)
      X(x2,x4,x10,x15) R3(x2,a,b,c,d,e)
      X(x3,x5,x11,x0) R3(x3,e,a,b,c,d)
      X(x4,x6,x12,x1) R3(x4,d,e,a,b,c)
      X(x5,x7,x13,x2) R3(x5,c,d,e,a,b)
      X(x6,x8,x14,x3) R3(x6,b,c,d,e,a)
      X(x7,x9,x15,x4) R3(x7,a,b,c,d,e)
      X(x8,x10,x0,x5) R3(x8,e,a,b,c,d)
      X(x9,x11,x1,x6) R3(x9,d,e,a,b,c)
      X(x10,x12,x2,x7) R3(x10,c,d,e,a,b)
      X(x11,x13,x3,x8) R3(x11,b,c,d,e,a)
      delta = 0xca62c1d6;
      X(x12,x14,x4,x9) R2(x12,a,b,c,d,e)
      X(x13,x15,x5,x10) R2(x13,e,a,b,c,d)
      X(x14,x0,x6,x11) R2(x14,d,e,a,b,c)
      X(x15,x1,x7,x12) R2(x15,c,d,e,a,b)
      X(x0,x2,x8,x13) R2(x0,b,c,d,e,a)
      X(x1,x3,x9,x14) R2(x1,a,b,c,d,e)
      X(x2,x4,x10,x15) R2(x2,e,a,b,c,d)
      X(x3,x5,x11,x0) R2(x3,d,e,a,b,c)
      X(x4,x6,x12,x1) R2(x4,c,d,e,a,b)
      X(x5,x7,x13,x2) R2(x5,b,c,d,e,a)
      X(x6,x8,x14,x3) R2(x6,a,b,c,d,e)
      X(x7,x9,x15,x4) R2(x7,e,a,b,c,d)
      X(x8,x10,x0,x5) R2(x8,d,e,a,b,c)
      X(x9,x11,x1,x6) R2(x9,c,d,e,a,b)
      X(x10,x12,x2,x7) R2(x10,b,c,d,e,a)
      X(x11,x13,x3,x8) R2(x11,a,b,c,d,e)
      X(x12,x14,x4,x9) R2(x12,e,a,b,c,d)
      X(x13,x15,x5,x10) R2(x13,d,e,a,b,c)
      X(x14,x0,x6,x11) R2(x14,c,d,e,a,b)
      X(x15,x1,x7,x12) R2(x15,b,c,d,e,a)
  
      --inblocks;
      in += 16; 
  
      a = a + state[0];
      b = b + state[1];
      c = c + state[2];
      d = d + state[3];
      e = e + state[4];
      state[0] = a;
      state[1] = b;
      state[2] = c;
      state[3] = d;
      state[4] = e;
    }
  }
  void update1(const uint32x4 *in) {
    uint32x4 a = state[0];
    uint32x4 b = state[1];
    uint32x4 c = state[2];
    uint32x4 d = state[3];
    uint32x4 e = state[4];
    uint32x4 delta;
    uint32x4 x0,x1,x2,x3,x4,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15;
  
    delta = 0x5a827999;
    x0 = in[0]; R1(x0,a,b,c,d,e)
    x1 = in[1]; R1(x1,e,a,b,c,d)
    R10(d,e,a,b,c)
    R10(c,d,e,a,b)
    R10(b,c,d,e,a)
    R10(a,b,c,d,e)
    R10(e,a,b,c,d)
    R10(d,e,a,b,c)
    R10(c,d,e,a,b)
    R10(b,c,d,e,a)
    R10(a,b,c,d,e)
    R10(e,a,b,c,d)
    R10(d,e,a,b,c)
    R10(c,d,e,a,b)
    R10(b,c,d,e,a)
    x15 = in[15]; R1(x15,a,b,c,d,e)
    x0 = rotate1(x0);  R1(x0,e,a,b,c,d)
    x1 = rotate1(x1);  R1(x1,d,e,a,b,c)
    x2 = rotate1(x15); R1(x2,c,d,e,a,b)
    x3 = rotate1(x0);  R1(x3,b,c,d,e,a)
    delta = 0x6ed9eba1;
    x4 = rotate1(x1); R2(x4,a,b,c,d,e)
    x5 = rotate1(x2); R2(x5,e,a,b,c,d)
    x6 = rotate1(x3); R2(x6,d,e,a,b,c)
    x7 = x15 ^ x4;
    x7 = rotate1(x7); R2(x7,c,d,e,a,b)
    x8 = x0 ^ x5;
    x8 = rotate1(x8); R2(x8,b,c,d,e,a)
    x9 = x1 ^ x6;
    x9 = rotate1(x9); R2(x9,a,b,c,d,e)
    x10 = x2 ^ x7;
    x10 = rotate1(x10); R2(x10,e,a,b,c,d)
    x11 = x3 ^ x8;
    x11 = rotate1(x11); R2(x11,d,e,a,b,c)
    x12 = x4 ^ x9;
    x12 = rotate1(x12); R2(x12,c,d,e,a,b)
    x13 = x5 ^ x15;
    x13 = x13 ^ x10;
    x13 = rotate1(x13); R2(x13,b,c,d,e,a)
    x14 = x6 ^ x0;
    x14 = x14 ^ x11;
    x14 = rotate1(x14); R2(x14,a,b,c,d,e)
    x15 = x15 ^ x7;
    x15 = x15 ^ x1;
    x15 = x15 ^ x12;
    x15 = rotate1(x15); R2(x15,e,a,b,c,d)
    x0 = x0 ^ x8;
    x0 = x0 ^ x2;
    x0 = x0 ^ x13;
    x0 = rotate1(x0); R2(x0,d,e,a,b,c)
    x1 = x1 ^ x9;
    x1 = x1 ^ x3;
    x1 = x1 ^ x14;
    x1 = rotate1(x1); R2(x1,c,d,e,a,b)
    x2 = x2 ^ x10;
    x2 = x2 ^ x4;
    x2 = x2 ^ x15;
    x2 = rotate1(x2); R2(x2,b,c,d,e,a)
    x3 = x3 ^ x11;
    x3 = x3 ^ x5;
    x3 = x3 ^ x0;
    x3 = rotate1(x3); R2(x3,a,b,c,d,e)
    x4 = x4 ^ x12;
    x4 = x4 ^ x6;
    x4 = x4 ^ x1;
    x4 = rotate1(x4); R2(x4,e,a,b,c,d)
    x5 = x5 ^ x13;
    x5 = x5 ^ x7;
    x5 = x5 ^ x2;
    x5 = rotate1(x5); R2(x5,d,e,a,b,c)
    x6 = x6 ^ x14;
    x6 = x6 ^ x8;
    x6 = x6 ^ x3;
    x6 = rotate1(x6); R2(x6,c,d,e,a,b)
    x7 = x7 ^ x15;
    x7 = x7 ^ x9;
    x7 = x7 ^ x4;
    x7 = rotate1(x7); R2(x7,b,c,d,e,a)
    delta = 0x8f1bbcdc;
    x8 = x8 ^ x0;
    x8 = x8 ^ x10;
    x8 = x8 ^ x5;
    x8 = rotate1(x8); R3(x8,a,b,c,d,e)
    x9 = x9 ^ x1;
    x9 = x9 ^ x11;
    x9 = x9 ^ x6;
    x9 = rotate1(x9); R3(x9,e,a,b,c,d)
    x10 = x10 ^ x2;
    x10 = x10 ^ x12;
    x10 = x10 ^ x7;
    x10 = rotate1(x10); R3(x10,d,e,a,b,c)
    x11 = x11 ^ x3;
    x11 = x11 ^ x13;
    x11 = x11 ^ x8;
    x11 = rotate1(x11); R3(x11,c,d,e,a,b)
    x12 = x12 ^ x4;
    x12 = x12 ^ x14;
    x12 = x12 ^ x9;
    x12 = rotate1(x12); R3(x12,b,c,d,e,a)
    x13 = x13 ^ x5;
    x13 = x13 ^ x15;
    x13 = x13 ^ x10;
    x13 = rotate1(x13); R3(x13,a,b,c,d,e)
    x14 = x14 ^ x6;
    x14 = x14 ^ x0;
    x14 = x14 ^ x11;
    x14 = rotate1(x14); R3(x14,e,a,b,c,d)
    x15 = x15 ^ x7;
    x15 = x15 ^ x1;
    x15 = x15 ^ x12;
    x15 = rotate1(x15); R3(x15,d,e,a,b,c)
    x0 = x0 ^ x8;
    x0 = x0 ^ x2;
    x0 = x0 ^ x13;
    x0 = rotate1(x0); R3(x0,c,d,e,a,b)
    x1 = x1 ^ x9;
    x1 = x1 ^ x3;
    x1 = x1 ^ x14;
    x1 = rotate1(x1); R3(x1,b,c,d,e,a)
    x2 = x2 ^ x10;
    x2 = x2 ^ x4;
    x2 = x2 ^ x15;
    x2 = rotate1(x2); R3(x2,a,b,c,d,e)
    x3 = x3 ^ x11;
    x3 = x3 ^ x5;
    x3 = x3 ^ x0;
    x3 = rotate1(x3); R3(x3,e,a,b,c,d)
    x4 = x4 ^ x12;
    x4 = x4 ^ x6;
    x4 = x4 ^ x1;
    x4 = rotate1(x4); R3(x4,d,e,a,b,c)
    x5 = x5 ^ x13;
    x5 = x5 ^ x7;
    x5 = x5 ^ x2;
    x5 = rotate1(x5); R3(x5,c,d,e,a,b)
    x6 = x6 ^ x14;
    x6 = x6 ^ x8;
    x6 = x6 ^ x3;
    x6 = rotate1(x6); R3(x6,b,c,d,e,a)
    x7 = x7 ^ x15;
    x7 = x7 ^ x9;
    x7 = x7 ^ x4;
    x7 = rotate1(x7); R3(x7,a,b,c,d,e)
    x8 = x8 ^ x0;
    x8 = x8 ^ x10;
    x8 = x8 ^ x5;
    x8 = rotate1(x8); R3(x8,e,a,b,c,d)
    x9 = x9 ^ x1;
    x9 = x9 ^ x11;
    x9 = x9 ^ x6;
    x9 = rotate1(x9); R3(x9,d,e,a,b,c)
    x10 = x10 ^ x2;
    x10 = x10 ^ x12;
    x10 = x10 ^ x7;
    x10 = rotate1(x10); R3(x10,c,d,e,a,b)
    x11 = x11 ^ x3;
    x11 = x11 ^ x13;
    x11 = x11 ^ x8;
    x11 = rotate1(x11); R3(x11,b,c,d,e,a)
    delta = 0xca62c1d6;
    x12 = x12 ^ x4;
    x12 = x12 ^ x14;
    x12 = x12 ^ x9;
    x12 = rotate1(x12); R2(x12,a,b,c,d,e)
    x13 = x13 ^ x5;
    x13 = x13 ^ x15;
    x13 = x13 ^ x10;
    x13 = rotate1(x13); R2(x13,e,a,b,c,d)
    x14 = x14 ^ x6;
    x14 = x14 ^ x0;
    x14 = x14 ^ x11;
    x14 = rotate1(x14); R2(x14,d,e,a,b,c)
    x15 = x15 ^ x7;
    x15 = x15 ^ x1;
    x15 = x15 ^ x12;
    x15 = rotate1(x15); R2(x15,c,d,e,a,b)
    x0 = x0 ^ x8;
    x0 = x0 ^ x2;
    x0 = x0 ^ x13;
    x0 = rotate1(x0); R2(x0,b,c,d,e,a)
    x1 = x1 ^ x9;
    x1 = x1 ^ x3;
    x1 = x1 ^ x14;
    x1 = rotate1(x1); R2(x1,a,b,c,d,e)
    x2 = x2 ^ x10;
    x2 = x2 ^ x4;
    x2 = x2 ^ x15;
    x2 = rotate1(x2); R2(x2,e,a,b,c,d)
    x3 = x3 ^ x11;
    x3 = x3 ^ x5;
    x3 = x3 ^ x0;
    x3 = rotate1(x3); R2(x3,d,e,a,b,c)
    x4 = x4 ^ x12;
    x4 = x4 ^ x6;
    x4 = x4 ^ x1;
    x4 = rotate1(x4); R2(x4,c,d,e,a,b)
    x5 = x5 ^ x13;
    x5 = x5 ^ x7;
    x5 = x5 ^ x2;
    x5 = rotate1(x5); R2(x5,b,c,d,e,a)
    x6 = x6 ^ x14;
    x6 = x6 ^ x8;
    x6 = x6 ^ x3;
    x6 = rotate1(x6); R2(x6,a,b,c,d,e)
    x7 = x7 ^ x15;
    x7 = x7 ^ x9;
    x7 = x7 ^ x4;
    x7 = rotate1(x7); R2(x7,e,a,b,c,d)
    x8 = x8 ^ x0;
    x8 = x8 ^ x10;
    x8 = x8 ^ x5;
    x8 = rotate1(x8); R2(x8,d,e,a,b,c)
    x9 = x9 ^ x1;
    x9 = x9 ^ x11;
    x9 = x9 ^ x6;
    x9 = rotate1(x9); R2(x9,c,d,e,a,b)
    x10 = x10 ^ x2;
    x10 = x10 ^ x12;
    x10 = x10 ^ x7;
    x10 = rotate1(x10); R2(x10,b,c,d,e,a)
    x11 = x11 ^ x3;
    x11 = x11 ^ x13;
    x11 = x11 ^ x8;
    x11 = rotate1(x11); R2(x11,a,b,c,d,e)
    x12 = x12 ^ x4;
    x12 = x12 ^ x14;
    x12 = x12 ^ x9;
    x12 = rotate1(x12); R2(x12,e,a,b,c,d)
    x13 = x13 ^ x5;
    x13 = x13 ^ x15;
    x13 = x13 ^ x10;
    x13 = rotate1(x13); R2(x13,d,e,a,b,c)
    x14 = x14 ^ x6;
    x14 = x14 ^ x0;
    x14 = x14 ^ x11;
    x14 = rotate1(x14); R2(x14,c,d,e,a,b)
    x15 = x15 ^ x7;
    x15 = x15 ^ x1;
    x15 = x15 ^ x12;
    x15 = rotate1(x15); R2(x15,b,c,d,e,a)
  
    state[0] = state[0] + a;
    state[1] = state[1] + b;
    state[2] = state[2] + c;
    state[3] = state[3] + d;
    state[4] = state[4] + e;
  }
} ;
const char ALPHABET[64] =
"abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ/_";

#define ALPHABETUSED 32
// has to be a multiple of 4 for vectorization

#define WORDS (64 + 16 * ((sizeof s + sizeof target) / 64))
unsigned int words[WORDS];
uint32x4 wordsx4[WORDS];

int main()
{
  int i;
  int c0;
  int c1;
  int c2;
  int c3;
  int c4;
  int pos;
  long long startcycles = cpucycles();
  long long hashes = 1;
  
  long long targetlen = 0;
  while (target[targetlen]) ++targetlen;
  long long targetblocks = (targetlen + 72) / 64;

  for (i = 0;i < WORDS;++i) words[i] = 0;
  for (i = 0;i < targetlen;++i) ((unsigned char *) words)[i ^ 3] = target[i];
  ((unsigned char *) words)[targetlen ^ 3] = 0x80;
  words[targetblocks * 16 - 1] = targetlen * 8;
  for (i = 0;i < WORDS;++i) wordsx4[i] = words[i];

  hash targethash;
  targethash.init();
  targethash.update(wordsx4,targetblocks);

  cout << 0 << " " << targethash << " " << target << "\n";

  unsigned char flip[sizeof s];

  long long slen = 0;
  while (s[slen]) ++slen;
  if (slen < 5) return 100;
  long long sblocks = (slen + 72) / 64;
  long long sblockspre = (slen - 5) / 64;

  if (slen != 69) return 100;

  for (i = 0;i < slen;++i) flip[i] = 0;
#ifndef NONRANDOM
  srandom(cpucycles()); // XXX: randomize better
#endif
  for (i = 0;i < slen;++i) if (random() & 1) if (s[i] != ' ') s[i] ^= 32;
#ifndef NONRANDOM
  srandom(cpucycles() >> 10);
  for (i = 0;i < slen;++i) if (random() & 1) if (s[i] != ' ') s[i] ^= 32;
#endif

  for (i = 0;i < sizeof words;++i) ((unsigned char *) words)[i] = 0;
  ((unsigned char *) words)[slen ^ 3] = 0x80;
  words[sblocks * 16 - 1] = slen * 8;
  for (i = 0;i < slen;++i) ((unsigned char *) words)[i ^ 3] = s[i];

  for (;;) {
    for (i = 0;i < WORDS;++i) wordsx4[i] = words[i];

    hash shashpre;
    shashpre.init();
    shashpre.update(wordsx4,sblockspre);

    for (i = 0;i < slen;++i) {
      pos = i ^ 3;
      s[i] = ((unsigned char *) (wordsx4 + pos / 4))[3 & pos];
    }
    cout << "cycles/hash " << dec
      << (cpucycles() - startcycles) / hashes << " "
      << (innerloopcycles) / hashes << " "
      << s << "\n" << flush;
    for (c0 = 0;c0 < ALPHABETUSED;++c0) {
      int pos = (slen - 5) ^ 3;
      ((unsigned char *) (wordsx4 + pos / 4))[3 & pos] = ALPHABET[c0];
      ((unsigned char *) (wordsx4 + pos / 4))[4 + (3 & pos)] = ALPHABET[c0];
      ((unsigned char *) (wordsx4 + pos / 4))[8 + (3 & pos)] = ALPHABET[c0];
      ((unsigned char *) (wordsx4 + pos / 4))[12 + (3 & pos)] = ALPHABET[c0];
      for (c1 = 0;c1 < ALPHABETUSED;++c1) {
        pos = (slen - 4) ^ 3;
        ((unsigned char *) (wordsx4 + pos / 4))[3 & pos] = ALPHABET[c1];
        ((unsigned char *) (wordsx4 + pos / 4))[4 + (3 & pos)] = ALPHABET[c1];
        ((unsigned char *) (wordsx4 + pos / 4))[8 + (3 & pos)] = ALPHABET[c1];
        ((unsigned char *) (wordsx4 + pos / 4))[12 + (3 & pos)] = ALPHABET[c1];
        for (c2 = 0;c2 < ALPHABETUSED;++c2) {
          pos = (slen - 3) ^ 3;
          ((unsigned char *) (wordsx4 + pos / 4))[3 & pos] = ALPHABET[c2];
          ((unsigned char *) (wordsx4 + pos / 4))[4 + (3 & pos)] = ALPHABET[c2];
          ((unsigned char *) (wordsx4 + pos / 4))[8 + (3 & pos)] = ALPHABET[c2];
          ((unsigned char *) (wordsx4 + pos / 4))[12 + (3 & pos)] = ALPHABET[c2];
          innerloopcycles -= cpucycles();
          for (c3 = 0;c3 < ALPHABETUSED;++c3) {
            pos = (slen - 2) ^ 3;
            ((unsigned char *) (wordsx4 + pos / 4))[3 & pos] = ALPHABET[c3];
            ((unsigned char *) (wordsx4 + pos / 4))[4 + (3 & pos)] = ALPHABET[c3];
            ((unsigned char *) (wordsx4 + pos / 4))[8 + (3 & pos)] = ALPHABET[c3];
            ((unsigned char *) (wordsx4 + pos / 4))[12 + (3 & pos)] = ALPHABET[c3];
            for (c4 = 0;c4 + 4 <= ALPHABETUSED;c4 += 4) {
              pos = (slen - 1) ^ 3;
              ((unsigned char *) (wordsx4 + pos / 4))[3 & pos] = ALPHABET[c4];
              ((unsigned char *) (wordsx4 + pos / 4))[4 + (3 & pos)] = ALPHABET[c4 + 1];
              ((unsigned char *) (wordsx4 + pos / 4))[8 + (3 & pos)] = ALPHABET[c4 + 2];
              ((unsigned char *) (wordsx4 + pos / 4))[12 + (3 & pos)] = ALPHABET[c4 + 3];

              hash shash(shashpre);
              shash.update1(wordsx4 + sblockspre * 16);
              uint32x4 d = shash.hammingdistance(targethash);

              if (d.uint0() < 40) {
		pos = (slen - 1) ^ 3;
                ((unsigned char *) (wordsx4 + pos / 4))[3 & pos] = ALPHABET[c4];
                hash shash(shashpre);
                shash.update1(wordsx4 + sblockspre * 16);
                for (i = 0;i < slen;++i) {
		  pos = i ^ 3;
		  s[i] = ((unsigned char *) (wordsx4 + pos / 4))[3 & pos];
		}
                cout << dec << d.uint0() << " " << shash << " " << s << "\n" << flush;
              }
	      if (d.uint1() < 40) {
		pos = (slen - 1) ^ 3;
                ((unsigned char *) (wordsx4 + pos / 4))[3 & pos] = ALPHABET[c4 + 1];
                hash shash(shashpre);
                shash.update1(wordsx4 + sblockspre * 16);
                for (i = 0;i < slen;++i) {
		  pos = i ^ 3;
		  s[i] = ((unsigned char *) (wordsx4 + pos / 4))[3 & pos];
		}
                cout << dec << d.uint1() << " " << shash << " " << s << "\n" << flush;
              }
	      if (d.uint2() < 40) {
		pos = (slen - 1) ^ 3;
                ((unsigned char *) (wordsx4 + pos / 4))[3 & pos] = ALPHABET[c4 + 2];
                hash shash(shashpre);
                shash.update1(wordsx4 + sblockspre * 16);
                for (i = 0;i < slen;++i) {
		  pos = i ^ 3;
		  s[i] = ((unsigned char *) (wordsx4 + pos / 4))[3 & pos];
		}
                cout << dec << d.uint2() << " " << shash << " " << s << "\n" << flush;
              }
	      if (d.uint3() < 40) {
		pos = (slen - 1) ^ 3;
                ((unsigned char *) (wordsx4 + pos / 4))[3 & pos] = ALPHABET[c4 + 3];
                hash shash(shashpre);
                shash.update1(wordsx4 + sblockspre * 16);
                for (i = 0;i < slen;++i) {
		  pos = i ^ 3;
		  s[i] = ((unsigned char *) (wordsx4 + pos / 4))[3 & pos];
		}
                cout << dec << d.uint3() << " " << shash << " " << s << "\n" << flush;
              }
            }
            hashes += c4;
          }
          innerloopcycles += cpucycles();
        }
      }
    }

    for (i = 0;i < slen - 5;++i) if (s[i] != ' ') {
      ((unsigned char *) words)[i ^ 3] ^= 32;
      flip[i] ^= 32;
      if (flip[i]) break;
    }

    if (i == slen - 5) return 0;
  }
}