Merge pull request #3253

e136bc6b tweaks to the monerov1 cryptonight algorithm (Lee Clagett)
d58c9ec9 slow-hash: optimized version (SChernykh)
608fd6f1 Monero Cryptonight variants, and add one for v7 (moneromooo-monero)

src/crypto/chacha.h

@@ -69,10 +69,10 @@ namespace crypto {
     chacha20(data, length, key.data(), reinterpret_cast<const uint8_t*>(&iv), cipher);
   }
 
-  inline void generate_chacha_key(const void *data, size_t size, chacha_key& key, bool prehashed=false) {
+  inline void generate_chacha_key(const void *data, size_t size, chacha_key& key, int cn_variant = 0, bool prehashed=false) {
     static_assert(sizeof(chacha_key) <= sizeof(hash), "Size of hash must be at least that of chacha_key");
     tools::scrubbed_arr<char, HASH_SIZE> pwd_hash;
-    crypto::cn_slow_hash_pre(data, size, pwd_hash.data(), prehashed);
+    crypto::cn_slow_hash_pre(data, size, pwd_hash.data(), cn_variant, prehashed);
     memcpy(&key, pwd_hash.data(), sizeof(key));
   }
 

src/crypto/hash-ops.h

@@ -79,8 +79,8 @@ enum {
 };
 
 void cn_fast_hash(const void *data, size_t length, char *hash);
-void cn_slow_hash(const void *data, size_t length, char *hash);
+void cn_slow_hash(const void *data, size_t length, char *hash, int variant);
-void cn_slow_hash_pre(const void *data, size_t length, char *hash, bool pre);
+void cn_slow_hash_pre(const void *data, size_t length, char *hash, int variant, bool pre);
 
 void hash_extra_blake(const void *data, size_t length, char *hash);
 void hash_extra_groestl(const void *data, size_t length, char *hash);

src/crypto/hash.h

@@ -71,8 +71,8 @@ namespace crypto {
     return h;
   }
 
-  inline void cn_slow_hash(const void *data, std::size_t length, hash &hash) {
+  inline void cn_slow_hash(const void *data, std::size_t length, hash &hash, int variant = 0) {
-    cn_slow_hash(data, length, reinterpret_cast<char *>(&hash));
+    cn_slow_hash(data, length, reinterpret_cast<char *>(&hash), variant);
   }
 
   inline void tree_hash(const hash *hashes, std::size_t count, hash &root_hash) {

src/crypto/slow-hash.c

@@ -32,6 +32,8 @@
 #include <stddef.h>
 #include <stdint.h>
 #include <string.h>
+#include <stdio.h>
+#include <unistd.h>
 
 #include "common/int-util.h"
 #include "hash-ops.h"
@@ -47,6 +49,46 @@
 extern int aesb_single_round(const uint8_t *in, uint8_t*out, const uint8_t *expandedKey);
 extern int aesb_pseudo_round(const uint8_t *in, uint8_t *out, const uint8_t *expandedKey);
 
+#define VARIANT1_1(p) \
+  do if (variant > 0) \
+  { \
+    const uint8_t tmp = ((const uint8_t*)(p))[11]; \
+    static const uint32_t table = 0x75310; \
+    const uint8_t index = (((tmp >> 3) & 6) | (tmp & 1)) << 1; \
+    ((uint8_t*)(p))[11] = tmp ^ ((table >> index) & 0x30); \
+  } while(0)
+
+#define VARIANT1_2(p) \
+  do if (variant > 0) \
+  { \
+    xor64(p, tweak1_2); \
+  } while(0)
+
+#define VARIANT1_CHECK() \
+  do if (length < 43) \
+  { \
+    fprintf(stderr, "Cryptonight variants need at least 43 bytes of data"); \
+    _exit(1); \
+  } while(0)
+
+#define NONCE_POINTER (((const uint8_t*)data)+35)
+
+#define VARIANT1_PORTABLE_INIT() \
+  uint8_t tweak1_2[8]; \
+  do if (variant > 0) \
+  { \
+    VARIANT1_CHECK(); \
+    memcpy(&tweak1_2, &state.hs.b[192], sizeof(tweak1_2)); \
+    xor64(tweak1_2, NONCE_POINTER); \
+  } while(0)
+
+#define VARIANT1_INIT64() \
+  if (variant > 0) \
+  { \
+    VARIANT1_CHECK(); \
+  } \
+  const uint64_t tweak1_2 = variant > 0 ? (state.hs.w[24] ^ (*((const uint64_t*)NONCE_POINTER))) : 0
+
 #if !defined NO_AES && (defined(__x86_64__) || (defined(_MSC_VER) && defined(_WIN64)))
 // Optimised code below, uses x86-specific intrinsics, SSE2, AES-NI
 // Fall back to more portable code is down at the bottom
@@ -125,6 +167,7 @@ extern int aesb_pseudo_round(const uint8_t *in, uint8_t *out, const uint8_t *exp
   _mm_store_si128(R128(c), _c); \
   _b = _mm_xor_si128(_b, _c); \
   _mm_store_si128(R128(&hp_state[j]), _b); \
+  VARIANT1_1(&hp_state[j]); \
   j = state_index(c); \
   p = U64(&hp_state[j]); \
   b[0] = p[0]; b[1] = p[1]; \
@@ -133,6 +176,7 @@ extern int aesb_pseudo_round(const uint8_t *in, uint8_t *out, const uint8_t *exp
   p = U64(&hp_state[j]); \
   p[0] = a[0];  p[1] = a[1]; \
   a[0] ^= b[0]; a[1] ^= b[1]; \
+  VARIANT1_2(p + 1); \
   _b = _c; \
 
 #if defined(_MSC_VER)
@@ -183,6 +227,11 @@ STATIC INLINE void xor_blocks(uint8_t *a, const uint8_t *b)
     U64(a)[1] ^= U64(b)[1];
 }
 
+STATIC INLINE void xor64(uint64_t *a, const uint64_t b)
+{
+    *a ^= b;
+}
+
 /**
  * @brief uses cpuid to determine if the CPU supports the AES instructions
  * @return true if the CPU supports AES, false otherwise
@@ -515,11 +564,11 @@ void slow_hash_free_state(void)
  * @param length the length in bytes of the data
  * @param hash a pointer to a buffer in which the final 256 bit hash will be stored
  */
-void cn_slow_hash(const void *data, size_t length, char *hash) {
+void cn_slow_hash(const void *data, size_t length, char *hash, int variant) {
-    cn_slow_hash_pre(data,length,hash,false);
+    cn_slow_hash_pre(data,length,hash,variant,false);
 }
 
-void cn_slow_hash_pre(const void *data, size_t length, char *hash, bool prehashed)
+void cn_slow_hash_pre(const void *data, size_t length, char *hash, int variant, bool prehashed)
 {
     RDATA_ALIGN16 uint8_t expandedKey[240];  /* These buffers are aligned to use later with SSE functions */
 
@@ -553,6 +602,8 @@ void cn_slow_hash_pre(const void *data, size_t length, char *hash, bool prehashe
     }
     memcpy(text, state.init, INIT_SIZE_BYTE);
 
+    VARIANT1_INIT64();
+
     /* CryptoNight Step 2:  Iteratively encrypt the results from Keccak to fill
      * the 2MB large random access buffer.
      */
@@ -676,6 +727,11 @@ void slow_hash_free_state(void)
 
 #define U64(x) ((uint64_t *) (x))
 
+STATIC INLINE void xor64(uint64 *a, const uint64 b)
+{
+    *a ^= b;
+}
+
 #pragma pack(push, 1)
 union cn_slow_hash_state
 {
@@ -712,6 +768,7 @@ union cn_slow_hash_state
   vst1q_u8((uint8_t *)c, _c); \
   _b = veorq_u8(_b, _c); \
   vst1q_u8(&hp_state[j], _b); \
+  VARIANT1_1(&hp_state[j]); \
   j = state_index(c); \
   p = U64(&hp_state[j]); \
   b[0] = p[0]; b[1] = p[1]; \
@@ -720,6 +777,7 @@ union cn_slow_hash_state
   p = U64(&hp_state[j]); \
   p[0] = a[0];  p[1] = a[1]; \
   a[0] ^= b[0]; a[1] ^= b[1]; \
+  VARIANT1_2(p + 1); \
   _b = _c; \
 
 
@@ -851,7 +909,7 @@ STATIC INLINE void aes_pseudo_round_xor(const uint8_t *in, uint8_t *out, const u
 	}
 }
 
-void cn_slow_hash(const void *data, size_t length, char *hash)
+void cn_slow_hash(const void *data, size_t length, char *hash, int variant)
 {
     RDATA_ALIGN16 uint8_t expandedKey[240];
     RDATA_ALIGN16 uint8_t hp_state[MEMORY];
@@ -877,6 +935,8 @@ void cn_slow_hash(const void *data, size_t length, char *hash)
     hash_process(&state.hs, data, length);
     memcpy(text, state.init, INIT_SIZE_BYTE);
 
+    VARIANT1_INIT64();
+
     /* CryptoNight Step 2:  Iteratively encrypt the results from Keccak to fill
      * the 2MB large random access buffer.
      */
@@ -1045,7 +1105,7 @@ STATIC INLINE void xor_blocks(uint8_t* a, const uint8_t* b)
   U64(a)[1] ^= U64(b)[1];
 }
 
-void cn_slow_hash(const void *data, size_t length, char *hash)
+void cn_slow_hash(const void *data, size_t length, char *hash, int variant)
 {
     uint8_t text[INIT_SIZE_BYTE];
     uint8_t a[AES_BLOCK_SIZE];
@@ -1074,6 +1134,8 @@ void cn_slow_hash(const void *data, size_t length, char *hash)
     hash_process(&state.hs, data, length);
     memcpy(text, state.init, INIT_SIZE_BYTE);
 
+    VARIANT1_INIT64();
+
     aes_ctx = (oaes_ctx *) oaes_alloc();
     oaes_key_import_data(aes_ctx, state.hs.b, AES_KEY_SIZE);
 
@@ -1103,6 +1165,7 @@ void cn_slow_hash(const void *data, size_t length, char *hash)
       xor_blocks(b, p);
       swap_blocks(b, p);
       swap_blocks(a, b);
+      VARIANT1_1(p);
 
       // Iteration 2
       p = &long_state[state_index(a)];
@@ -1112,6 +1175,7 @@ void cn_slow_hash(const void *data, size_t length, char *hash)
       swap_blocks(b, p);
       xor_blocks(b, p);
       swap_blocks(a, b);
+      VARIANT1_2(U64(p) + 1);
     }
 
     memcpy(text, state.init, INIT_SIZE_BYTE);
@@ -1206,6 +1270,15 @@ static void xor_blocks(uint8_t* a, const uint8_t* b) {
   }
 }
 
+static void xor64(uint8_t* left, const uint8_t* right)
+{
+  size_t i;
+  for (i = 0; i < 8; ++i)
+  {
+    left[i] ^= right[i];
+  }
+}
+
 #pragma pack(push, 1)
 union cn_slow_hash_state {
   union hash_state hs;
@@ -1216,7 +1289,7 @@ union cn_slow_hash_state {
 };
 #pragma pack(pop)
 
-void cn_slow_hash(const void *data, size_t length, char *hash) {
+void cn_slow_hash(const void *data, size_t length, char *hash, int variant) {
   uint8_t long_state[MEMORY];
   union cn_slow_hash_state state;
   uint8_t text[INIT_SIZE_BYTE];
@@ -1233,6 +1306,8 @@ void cn_slow_hash(const void *data, size_t length, char *hash) {
   memcpy(aes_key, state.hs.b, AES_KEY_SIZE);
   aes_ctx = (oaes_ctx *) oaes_alloc();
 
+  VARIANT1_PORTABLE_INIT();
+
   oaes_key_import_data(aes_ctx, aes_key, AES_KEY_SIZE);
   for (i = 0; i < MEMORY / INIT_SIZE_BYTE; i++) {
     for (j = 0; j < INIT_SIZE_BLK; j++) {
@@ -1260,6 +1335,7 @@ void cn_slow_hash(const void *data, size_t length, char *hash) {
     copy_block(&long_state[j * AES_BLOCK_SIZE], c);
     assert(j == e2i(a, MEMORY / AES_BLOCK_SIZE));
     swap_blocks(a, b);
+    VARIANT1_1(&long_state[j * AES_BLOCK_SIZE]);
     /* Iteration 2 */
     j = e2i(a, MEMORY / AES_BLOCK_SIZE);
     copy_block(c, &long_state[j * AES_BLOCK_SIZE]);
@@ -1267,6 +1343,7 @@ void cn_slow_hash(const void *data, size_t length, char *hash) {
     sum_half_blocks(b, d);
     swap_blocks(b, c);
     xor_blocks(b, c);
+    VARIANT1_2(c + 8);
     copy_block(&long_state[j * AES_BLOCK_SIZE], c);
     assert(j == e2i(a, MEMORY / AES_BLOCK_SIZE));
     swap_blocks(a, b);

src/cryptonote_basic/cryptonote_format_utils.cpp

@@ -1022,7 +1022,8 @@ namespace cryptonote
       return true;
     }
     blobdata bd = get_block_hashing_blob(b);
-    crypto::cn_slow_hash(bd.data(), bd.size(), res);
+    const int cn_variant = b.major_version >= 7 ? b.major_version - 6 : 0;
+    crypto::cn_slow_hash(bd.data(), bd.size(), res, cn_variant);
     return true;
   }
   //---------------------------------------------------------------

src/device/device_ledger.cpp

@@ -511,7 +511,7 @@ namespace hw {
 
         char prekey[200];
         memmove(prekey, &this->buffer_recv[0], 200);
-        crypto::generate_chacha_key(&prekey[0], sizeof(prekey), key, true);
+        crypto::generate_chacha_key(&prekey[0], sizeof(prekey), key, 0, true);
 
         #ifdef DEBUG_HWDEVICE
         hw::ledger::check32("generate_chacha_key", "key", (char*)key_x.data(), (char*)key.data());

tests/hash/CMakeLists.txt

@@ -43,7 +43,7 @@ set_property(TARGET hash-tests
   PROPERTY
     FOLDER "tests")
 
-foreach (hash IN ITEMS fast slow tree extra-blake extra-groestl extra-jh extra-skein)
+foreach (hash IN ITEMS fast slow slow-1 tree extra-blake extra-groestl extra-jh extra-skein)
   add_test(
     NAME    "hash-${hash}"
     COMMAND hash-tests "${hash}" "${CMAKE_CURRENT_SOURCE_DIR}/tests-${hash}.txt")

tests/hash/main.cpp

@@ -51,6 +51,12 @@ extern "C" {
     }
     tree_hash((const char (*)[32]) data, length >> 5, hash);
   }
+  static void cn_slow_hash_0(const void *data, size_t length, char *hash) {
+    return cn_slow_hash(data, length, hash, 0);
+  }
+  static void cn_slow_hash_1(const void *data, size_t length, char *hash) {
+    return cn_slow_hash(data, length, hash, 1);
+  }
 }
 POP_WARNINGS
 
@@ -58,9 +64,10 @@ extern "C" typedef void hash_f(const void *, size_t, char *);
 struct hash_func {
   const string name;
   hash_f &f;
-} hashes[] = {{"fast", cn_fast_hash}, {"slow", cn_slow_hash}, {"tree", hash_tree},
+} hashes[] = {{"fast", cn_fast_hash}, {"slow", cn_slow_hash_0}, {"tree", hash_tree},
   {"extra-blake", hash_extra_blake}, {"extra-groestl", hash_extra_groestl},
-  {"extra-jh", hash_extra_jh}, {"extra-skein", hash_extra_skein}};
+  {"extra-jh", hash_extra_jh}, {"extra-skein", hash_extra_skein},
+  {"slow-1", cn_slow_hash_1}};
 
 int main(int argc, char *argv[]) {
   hash_f *f;

tests/hash/tests-slow-1.txt

@@ -0,0 +1,5 @@
+b5a7f63abb94d07d1a6445c36c07c7e8327fe61b1647e391b4c7edae5de57a3d 00000000000000000000000000000000000000000000000000000000000000000000000000000000000000
+80563c40ed46575a9e44820d93ee095e2851aa22483fd67837118c6cd951ba61 00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
+5bb40c5880cef2f739bdb6aaaf16161eaae55530e7b10d7ea996b751a299e949 8519e039172b0d70e5ca7b3383d6b3167315a422747b73f019cf9528f0fde341fd0f2a63030ba6450525cf6de31837669af6f1df8131faf50aaab8d3a7405589
+613e638505ba1fd05f428d5c9f8e08f8165614342dac419adc6a47dce257eb3e 37a636d7dafdf259b7287eddca2f58099e98619d2f99bdb8969d7b14498102cc065201c8be90bd777323f449848b215d2977c92c4c1c2da36ab46b2e389689ed97c18fec08cd3b03235c5e4c62a37ad88c7b67932495a71090e85dd4020a9300
+ed082e49dbd5bbe34a3726a0d1dad981146062b39d36d62c71eb1ed8ab49459b 38274c97c45a172cfc97679870422e3a1ab0784960c60514d816271415c306ee3a3ed1a77e31f6a885c3cb