Cryptonight variant 4 aka CryptonightR

It introduces random integer math into the main loop.

src/crypto/chacha.h

@@ -73,18 +73,18 @@ namespace crypto {
   inline void generate_chacha_key(const void *data, size_t size, chacha_key& key, uint64_t kdf_rounds) {
     static_assert(sizeof(chacha_key) <= sizeof(hash), "Size of hash must be at least that of chacha_key");
     epee::mlocked<tools::scrubbed_arr<char, HASH_SIZE>> pwd_hash;
-    crypto::cn_slow_hash(data, size, pwd_hash.data(), 0/*variant*/, 0/*prehashed*/);
+    crypto::cn_slow_hash(data, size, pwd_hash.data(), 0/*variant*/, 0/*prehashed*/, 0/*height*/);
     for (uint64_t n = 1; n < kdf_rounds; ++n)
-      crypto::cn_slow_hash(pwd_hash.data(), pwd_hash.size(), pwd_hash.data(), 0/*variant*/, 0/*prehashed*/);
+      crypto::cn_slow_hash(pwd_hash.data(), pwd_hash.size(), pwd_hash.data(), 0/*variant*/, 0/*prehashed*/, 0/*height*/);
     memcpy(&unwrap(unwrap(key)), pwd_hash.data(), sizeof(key));
   }
 
   inline void generate_chacha_key_prehashed(const void *data, size_t size, chacha_key& key, uint64_t kdf_rounds) {
     static_assert(sizeof(chacha_key) <= sizeof(hash), "Size of hash must be at least that of chacha_key");
     epee::mlocked<tools::scrubbed_arr<char, HASH_SIZE>> pwd_hash;
-    crypto::cn_slow_hash(data, size, pwd_hash.data(), 0/*variant*/, 1/*prehashed*/);
+    crypto::cn_slow_hash(data, size, pwd_hash.data(), 0/*variant*/, 1/*prehashed*/, 0/*height*/);
     for (uint64_t n = 1; n < kdf_rounds; ++n)
-      crypto::cn_slow_hash(pwd_hash.data(), pwd_hash.size(), pwd_hash.data(), 0/*variant*/, 0/*prehashed*/);
+      crypto::cn_slow_hash(pwd_hash.data(), pwd_hash.size(), pwd_hash.data(), 0/*variant*/, 0/*prehashed*/, 0/*height*/);
     memcpy(&unwrap(unwrap(key)), pwd_hash.data(), sizeof(key));
   }
 

src/crypto/hash-ops.h

@@ -79,7 +79,7 @@ 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, int variant, int prehashed);
+void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int prehashed, uint64_t height);
 
 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

@@ -71,12 +71,12 @@ namespace crypto {
     return h;
   }
 
-  inline void cn_slow_hash(const void *data, std::size_t length, hash &hash, int variant = 0) {
+  inline void cn_slow_hash(const void *data, std::size_t length, hash &hash, int variant = 0, uint64_t height = 0) {
-    cn_slow_hash(data, length, reinterpret_cast<char *>(&hash), variant, 0/*prehashed*/);
+    cn_slow_hash(data, length, reinterpret_cast<char *>(&hash), variant, 0/*prehashed*/, height);
   }
 
-  inline void cn_slow_hash_prehashed(const void *data, std::size_t length, hash &hash, int variant = 0) {
+  inline void cn_slow_hash_prehashed(const void *data, std::size_t length, hash &hash, int variant = 0, uint64_t height = 0) {
-    cn_slow_hash(data, length, reinterpret_cast<char *>(&hash), variant, 1/*prehashed*/);
+    cn_slow_hash(data, length, reinterpret_cast<char *>(&hash), variant, 1/*prehashed*/, height);
   }
 
   inline void tree_hash(const hash *hashes, std::size_t count, hash &root_hash) {

src/crypto/slow-hash.c

@@ -39,6 +39,7 @@
 #include "hash-ops.h"
 #include "oaes_lib.h"
 #include "variant2_int_sqrt.h"
+#include "variant4_random_math.h"
 
 #define MEMORY         (1 << 21) // 2MB scratchpad
 #define ITER           (1 << 20)
@@ -172,7 +173,7 @@ extern void aesb_pseudo_round(const uint8_t *in, uint8_t *out, const uint8_t *ex
   const uint64_t sqrt_input = SWAP64LE(((uint64_t*)(ptr))[0]) + division_result
 
 #define VARIANT2_INTEGER_MATH_SSE2(b, ptr) \
-  do if (variant >= 2) \
+  do if ((variant == 2) || (variant == 3)) \
   { \
     VARIANT2_INTEGER_MATH_DIVISION_STEP(b, ptr); \
     VARIANT2_INTEGER_MATH_SQRT_STEP_SSE2(); \
@@ -182,7 +183,7 @@ extern void aesb_pseudo_round(const uint8_t *in, uint8_t *out, const uint8_t *ex
 #if defined DBL_MANT_DIG && (DBL_MANT_DIG >= 50)
   // double precision floating point type has enough bits of precision on current platform
   #define VARIANT2_PORTABLE_INTEGER_MATH(b, ptr) \
-    do if (variant >= 2) \
+    do if ((variant == 2) || (variant == 3)) \
     { \
       VARIANT2_INTEGER_MATH_DIVISION_STEP(b, ptr); \
       VARIANT2_INTEGER_MATH_SQRT_STEP_FP64(); \
@@ -192,7 +193,7 @@ extern void aesb_pseudo_round(const uint8_t *in, uint8_t *out, const uint8_t *ex
   // double precision floating point type is not good enough on current platform
   // fall back to the reference code (integer only)
   #define VARIANT2_PORTABLE_INTEGER_MATH(b, ptr) \
-    do if (variant >= 2) \
+    do if ((variant == 2) || (variant == 3)) \
     { \
       VARIANT2_INTEGER_MATH_DIVISION_STEP(b, ptr); \
       VARIANT2_INTEGER_MATH_SQRT_STEP_REF(); \
@@ -214,6 +215,47 @@ extern void aesb_pseudo_round(const uint8_t *in, uint8_t *out, const uint8_t *ex
     lo ^= SWAP64LE(*(U64(hp_state + (j ^ 0x20)) + 1)); \
   } while (0)
 
+#define V4_REG_LOAD(dst, src) \
+  do { \
+    memcpy((dst), (src), sizeof(v4_reg)); \
+    if (sizeof(v4_reg) == sizeof(uint32_t)) \
+      *(dst) = SWAP32LE(*(dst)); \
+    else \
+      *(dst) = SWAP64LE(*(dst)); \
+  } while (0)
+
+#define VARIANT4_RANDOM_MATH_INIT() \
+  v4_reg r[9]; \
+  struct V4_Instruction code[NUM_INSTRUCTIONS_MAX + 1]; \
+  do if (variant >= 4) \
+  { \
+    for (int i = 0; i < 4; ++i) \
+      V4_REG_LOAD(r + i, (uint8_t*)(state.hs.w + 12) + sizeof(v4_reg) * i); \
+    v4_random_math_init(code, height); \
+  } while (0)
+
+#define VARIANT4_RANDOM_MATH(a, b, r, _b, _b1) \
+  do if (variant >= 4) \
+  { \
+    uint64_t t; \
+    memcpy(&t, b, sizeof(uint64_t)); \
+    \
+    if (sizeof(v4_reg) == sizeof(uint32_t)) \
+      t ^= SWAP64LE((r[0] + r[1]) | ((uint64_t)(r[2] + r[3]) << 32)); \
+    else \
+      t ^= SWAP64LE((r[0] + r[1]) ^ (r[2] + r[3])); \
+    \
+    memcpy(b, &t, sizeof(uint64_t)); \
+    \
+    V4_REG_LOAD(r + 4, a); \
+    V4_REG_LOAD(r + 5, (uint64_t*)(a) + 1); \
+    V4_REG_LOAD(r + 6, _b); \
+    V4_REG_LOAD(r + 7, _b1); \
+    V4_REG_LOAD(r + 8, (uint64_t*)(_b1) + 1); \
+    \
+    v4_random_math(code, r); \
+  } while (0)
+
 
 #if !defined NO_AES && (defined(__x86_64__) || (defined(_MSC_VER) && defined(_WIN64)))
 // Optimised code below, uses x86-specific intrinsics, SSE2, AES-NI
@@ -298,6 +340,7 @@ extern void aesb_pseudo_round(const uint8_t *in, uint8_t *out, const uint8_t *ex
   p = U64(&hp_state[j]); \
   b[0] = p[0]; b[1] = p[1]; \
   VARIANT2_INTEGER_MATH_SSE2(b, c); \
+  VARIANT4_RANDOM_MATH(a, b, r, &_b, &_b1); \
   __mul(); \
   VARIANT2_2(); \
   VARIANT2_SHUFFLE_ADD_SSE2(hp_state, j); \
@@ -694,7 +737,7 @@ 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, int variant, int prehashed)
+void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int prehashed, uint64_t height)
 {
     RDATA_ALIGN16 uint8_t expandedKey[240];  /* These buffers are aligned to use later with SSE functions */
 
@@ -730,6 +773,7 @@ void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int
 
     VARIANT1_INIT64();
     VARIANT2_INIT64();
+    VARIANT4_RANDOM_MATH_INIT();
 
     /* CryptoNight Step 2:  Iteratively encrypt the results from Keccak to fill
      * the 2MB large random access buffer.
@@ -901,6 +945,7 @@ union cn_slow_hash_state
   p = U64(&hp_state[j]); \
   b[0] = p[0]; b[1] = p[1]; \
   VARIANT2_PORTABLE_INTEGER_MATH(b, c); \
+  VARIANT4_RANDOM_MATH(a, b, r, &_b, &_b1); \
   __mul(); \
   VARIANT2_2(); \
   VARIANT2_SHUFFLE_ADD_NEON(hp_state, j); \
@@ -1063,7 +1108,7 @@ STATIC INLINE void aligned_free(void *ptr)
 }
 #endif /* FORCE_USE_HEAP */
 
-void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int prehashed)
+void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int prehashed, uint64_t height)
 {
     RDATA_ALIGN16 uint8_t expandedKey[240];
 
@@ -1100,6 +1145,7 @@ void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int
 
     VARIANT1_INIT64();
     VARIANT2_INIT64();
+    VARIANT4_RANDOM_MATH_INIT();
 
     /* CryptoNight Step 2:  Iteratively encrypt the results from Keccak to fill
      * the 2MB large random access buffer.
@@ -1278,7 +1324,7 @@ 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, int variant, int prehashed)
+void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int prehashed, uint64_t height)
 {
     uint8_t text[INIT_SIZE_BYTE];
     uint8_t a[AES_BLOCK_SIZE];
@@ -1317,6 +1363,7 @@ void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int
 
     VARIANT1_INIT64();
     VARIANT2_INIT64();
+    VARIANT4_RANDOM_MATH_INIT();
 
     // use aligned data
     memcpy(expandedKey, aes_ctx->key->exp_data, aes_ctx->key->exp_data_len);
@@ -1353,6 +1400,7 @@ void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int
       copy_block(c, p);
 
       VARIANT2_PORTABLE_INTEGER_MATH(c, c1);
+      VARIANT4_RANDOM_MATH(a, c, r, b, b + AES_BLOCK_SIZE);
       mul(c1, c, d);
       VARIANT2_2_PORTABLE();
       VARIANT2_PORTABLE_SHUFFLE_ADD(long_state, j);
@@ -1476,7 +1524,7 @@ union cn_slow_hash_state {
 };
 #pragma pack(pop)
 
-void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int prehashed) {
+void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int prehashed, uint64_t height) {
 #ifndef FORCE_USE_HEAP
   uint8_t long_state[MEMORY];
 #else
@@ -1505,6 +1553,7 @@ void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int
 
   VARIANT1_PORTABLE_INIT();
   VARIANT2_PORTABLE_INIT();
+  VARIANT4_RANDOM_MATH_INIT();
 
   oaes_key_import_data(aes_ctx, aes_key, AES_KEY_SIZE);
   for (i = 0; i < MEMORY / INIT_SIZE_BYTE; i++) {
@@ -1537,6 +1586,7 @@ void cn_slow_hash(const void *data, size_t length, char *hash, int variant, int
     j = e2i(c1, MEMORY / AES_BLOCK_SIZE) * AES_BLOCK_SIZE;
     copy_block(c2, &long_state[j]);
     VARIANT2_PORTABLE_INTEGER_MATH(c2, c1);
+    VARIANT4_RANDOM_MATH(a, c2, r, b, b + AES_BLOCK_SIZE);
     mul(c1, c2, d);
     VARIANT2_2_PORTABLE();
     VARIANT2_PORTABLE_SHUFFLE_ADD(long_state, j);

src/crypto/variant4_random_math.h

@@ -0,0 +1,441 @@
+#ifndef VARIANT4_RANDOM_MATH_H
+#define VARIANT4_RANDOM_MATH_H
+
+// Register size can be configured to either 32 bit (uint32_t) or 64 bit (uint64_t)
+typedef uint32_t v4_reg;
+
+enum V4_Settings
+{
+	// Generate code with minimal theoretical latency = 45 cycles, which is equivalent to 15 multiplications
+	TOTAL_LATENCY = 15 * 3,
+	
+	// Always generate at least 60 instructions
+	NUM_INSTRUCTIONS_MIN = 60,
+
+	// Never generate more than 70 instructions (final RET instruction doesn't count here)
+	NUM_INSTRUCTIONS_MAX = 70,
+
+	// Available ALUs for MUL
+	// Modern CPUs typically have only 1 ALU which can do multiplications
+	ALU_COUNT_MUL = 1,
+
+	// Total available ALUs
+	// Modern CPUs have 4 ALUs, but we use only 3 because random math executes together with other main loop code
+	ALU_COUNT = 3,
+};
+
+enum V4_InstructionList
+{
+	MUL,	// a*b
+	ADD,	// a+b + C, C is an unsigned 32-bit constant
+	SUB,	// a-b
+	ROR,	// rotate right "a" by "b & 31" bits
+	ROL,	// rotate left "a" by "b & 31" bits
+	XOR,	// a^b
+	RET,	// finish execution
+	V4_INSTRUCTION_COUNT = RET,
+};
+
+// V4_InstructionDefinition is used to generate code from random data
+// Every random sequence of bytes is a valid code
+//
+// There are 8 registers in total:
+// - 4 variable registers
+// - 4 constant registers initialized from loop variables
+//
+// This is why dst_index is 2 bits
+enum V4_InstructionDefinition
+{
+	V4_OPCODE_BITS = 3,
+	V4_DST_INDEX_BITS = 2,
+	V4_SRC_INDEX_BITS = 3,
+};
+
+struct V4_Instruction
+{
+	uint8_t opcode;
+	uint8_t dst_index;
+	uint8_t src_index;
+	uint32_t C;
+};
+
+#ifndef FORCEINLINE
+#if defined(__GNUC__)
+#define FORCEINLINE __attribute__((always_inline)) inline
+#elif defined(_MSC_VER)
+#define FORCEINLINE __forceinline
+#else
+#define FORCEINLINE inline
+#endif
+#endif
+
+#ifndef UNREACHABLE_CODE
+#if defined(__GNUC__)
+#define UNREACHABLE_CODE __builtin_unreachable()
+#elif defined(_MSC_VER)
+#define UNREACHABLE_CODE __assume(false)
+#else
+#define UNREACHABLE_CODE
+#endif
+#endif
+
+// Random math interpreter's loop is fully unrolled and inlined to achieve 100% branch prediction on CPU:
+// every switch-case will point to the same destination on every iteration of Cryptonight main loop
+//
+// This is about as fast as it can get without using low-level machine code generation
+static FORCEINLINE void v4_random_math(const struct V4_Instruction* code, v4_reg* r)
+{
+	enum
+	{
+		REG_BITS = sizeof(v4_reg) * 8,
+	};
+
+#define V4_EXEC(i) \
+	{ \
+		const struct V4_Instruction* op = code + i; \
+		const v4_reg src = r[op->src_index]; \
+		v4_reg* dst = r + op->dst_index; \
+		switch (op->opcode) \
+		{ \
+		case MUL: \
+			*dst *= src; \
+			break; \
+		case ADD: \
+			*dst += src + op->C; \
+			break; \
+		case SUB: \
+			*dst -= src; \
+			break; \
+		case ROR: \
+			{ \
+				const uint32_t shift = src % REG_BITS; \
+				*dst = (*dst >> shift) | (*dst << ((REG_BITS - shift) % REG_BITS)); \
+			} \
+			break; \
+		case ROL: \
+			{ \
+				const uint32_t shift = src % REG_BITS; \
+				*dst = (*dst << shift) | (*dst >> ((REG_BITS - shift) % REG_BITS)); \
+			} \
+			break; \
+		case XOR: \
+			*dst ^= src; \
+			break; \
+		case RET: \
+			return; \
+		default: \
+			UNREACHABLE_CODE; \
+			break; \
+		} \
+	}
+
+#define V4_EXEC_10(j) \
+	V4_EXEC(j + 0) \
+	V4_EXEC(j + 1) \
+	V4_EXEC(j + 2) \
+	V4_EXEC(j + 3) \
+	V4_EXEC(j + 4) \
+	V4_EXEC(j + 5) \
+	V4_EXEC(j + 6) \
+	V4_EXEC(j + 7) \
+	V4_EXEC(j + 8) \
+	V4_EXEC(j + 9)
+
+	// Generated program can have 60 + a few more (usually 2-3) instructions to achieve required latency
+	// I've checked all block heights < 10,000,000 and here is the distribution of program sizes:
+	//
+	// 60      27960
+	// 61      105054
+	// 62      2452759
+	// 63      5115997
+	// 64      1022269
+	// 65      1109635
+	// 66      153145
+	// 67      8550
+	// 68      4529
+	// 69      102
+
+	// Unroll 70 instructions here
+	V4_EXEC_10(0);		// instructions 0-9
+	V4_EXEC_10(10);		// instructions 10-19
+	V4_EXEC_10(20);		// instructions 20-29
+	V4_EXEC_10(30);		// instructions 30-39
+	V4_EXEC_10(40);		// instructions 40-49
+	V4_EXEC_10(50);		// instructions 50-59
+	V4_EXEC_10(60);		// instructions 60-69
+
+#undef V4_EXEC_10
+#undef V4_EXEC
+}
+
+// If we don't have enough data available, generate more
+static FORCEINLINE void check_data(size_t* data_index, const size_t bytes_needed, int8_t* data, const size_t data_size)
+{
+	if (*data_index + bytes_needed > data_size)
+	{
+		hash_extra_blake(data, data_size, (char*) data);
+		*data_index = 0;
+	}
+}
+
+// Generates as many random math operations as possible with given latency and ALU restrictions
+// "code" array must have space for NUM_INSTRUCTIONS_MAX+1 instructions
+static inline int v4_random_math_init(struct V4_Instruction* code, const uint64_t height)
+{
+	// MUL is 3 cycles, 3-way addition and rotations are 2 cycles, SUB/XOR are 1 cycle
+	// These latencies match real-life instruction latencies for Intel CPUs starting from Sandy Bridge and up to Skylake/Coffee lake
+	//
+	// AMD Ryzen has the same latencies except 1-cycle ROR/ROL, so it'll be a bit faster than Intel Sandy Bridge and newer processors
+	// Surprisingly, Intel Nehalem also has 1-cycle ROR/ROL, so it'll also be faster than Intel Sandy Bridge and newer processors
+	// AMD Bulldozer has 4 cycles latency for MUL (slower than Intel) and 1 cycle for ROR/ROL (faster than Intel), so average performance will be the same
+	// Source: https://www.agner.org/optimize/instruction_tables.pdf
+	const int op_latency[V4_INSTRUCTION_COUNT] = { 3, 2, 1, 2, 2, 1 };
+
+	// Instruction latencies for theoretical ASIC implementation
+	const int asic_op_latency[V4_INSTRUCTION_COUNT] = { 3, 1, 1, 1, 1, 1 };
+
+	// Available ALUs for each instruction
+	const int op_ALUs[V4_INSTRUCTION_COUNT] = { ALU_COUNT_MUL, ALU_COUNT, ALU_COUNT, ALU_COUNT, ALU_COUNT, ALU_COUNT };
+
+	int8_t data[32];
+	memset(data, 0, sizeof(data));
+	uint64_t tmp = SWAP64LE(height);
+	memcpy(data, &tmp, sizeof(uint64_t));
+
+	// Set data_index past the last byte in data
+	// to trigger full data update with blake hash
+	// before we start using it
+	size_t data_index = sizeof(data);
+
+	int code_size;
+
+	// There is a small chance (1.8%) that register R8 won't be used in the generated program
+	// So we keep track of it and try again if it's not used
+	bool r8_used;
+	do {
+		int latency[9];
+		int asic_latency[9];
+
+		// Tracks previous instruction and value of the source operand for registers R0-R3 throughout code execution
+		// byte 0: current value of the destination register
+		// byte 1: instruction opcode
+		// byte 2: current value of the source register
+		//
+		// Registers R4-R8 are constant and are treated as having the same value because when we do
+		// the same operation twice with two constant source registers, it can be optimized into a single operation
+		uint32_t inst_data[9] = { 0, 1, 2, 3, 0xFFFFFF, 0xFFFFFF, 0xFFFFFF, 0xFFFFFF, 0xFFFFFF };
+
+		bool alu_busy[TOTAL_LATENCY + 1][ALU_COUNT];
+		bool is_rotation[V4_INSTRUCTION_COUNT];
+		bool rotated[4];
+		int rotate_count = 0;
+
+		memset(latency, 0, sizeof(latency));
+		memset(asic_latency, 0, sizeof(asic_latency));
+		memset(alu_busy, 0, sizeof(alu_busy));
+		memset(is_rotation, 0, sizeof(is_rotation));
+		memset(rotated, 0, sizeof(rotated));
+		is_rotation[ROR] = true;
+		is_rotation[ROL] = true;
+
+		int num_retries = 0;
+		code_size = 0;
+
+		int total_iterations = 0;
+		r8_used = false;
+
+		// Generate random code to achieve minimal required latency for our abstract CPU
+		// Try to get this latency for all 4 registers
+		while (((latency[0] < TOTAL_LATENCY) || (latency[1] < TOTAL_LATENCY) || (latency[2] < TOTAL_LATENCY) || (latency[3] < TOTAL_LATENCY)) && (num_retries < 64))
+		{
+			// Fail-safe to guarantee loop termination
+			++total_iterations;
+			if (total_iterations > 256)
+				break;
+
+			check_data(&data_index, 1, data, sizeof(data));
+
+			const uint8_t c = ((uint8_t*)data)[data_index++];
+
+			// MUL = opcodes 0-2
+			// ADD = opcode 3
+			// SUB = opcode 4
+			// ROR/ROL = opcode 5, shift direction is selected randomly
+			// XOR = opcodes 6-7
+			uint8_t opcode = c & ((1 << V4_OPCODE_BITS) - 1);
+			if (opcode == 5)
+			{
+				check_data(&data_index, 1, data, sizeof(data));
+				opcode = (data[data_index++] >= 0) ? ROR : ROL;
+			}
+			else if (opcode >= 6)
+			{
+				opcode = XOR;
+			}
+			else
+			{
+				opcode = (opcode <= 2) ? MUL : (opcode - 2);
+			}
+
+			uint8_t dst_index = (c >> V4_OPCODE_BITS) & ((1 << V4_DST_INDEX_BITS) - 1);
+			uint8_t src_index = (c >> (V4_OPCODE_BITS + V4_DST_INDEX_BITS)) & ((1 << V4_SRC_INDEX_BITS) - 1);
+
+			const int a = dst_index;
+			int b = src_index;
+
+			// Don't do ADD/SUB/XOR with the same register
+			if (((opcode == ADD) || (opcode == SUB) || (opcode == XOR)) && (a == b))
+			{
+				// Use register R8 as source instead
+				b = 8;
+				src_index = 8;
+			}
+
+			// Don't do rotation with the same destination twice because it's equal to a single rotation
+			if (is_rotation[opcode] && rotated[a])
+			{
+				continue;
+			}
+
+			// Don't do the same instruction (except MUL) with the same source value twice because all other cases can be optimized:
+			// 2xADD(a, b, C) = ADD(a, b*2, C1+C2), same for SUB and rotations
+			// 2xXOR(a, b) = NOP
+			if ((opcode != MUL) && ((inst_data[a] & 0xFFFF00) == (opcode << 8) + ((inst_data[b] & 255) << 16)))
+			{
+				continue;
+			}
+
+			// Find which ALU is available (and when) for this instruction
+			int next_latency = (latency[a] > latency[b]) ? latency[a] : latency[b];
+			int alu_index = -1;
+			while (next_latency < TOTAL_LATENCY)
+			{
+				for (int i = op_ALUs[opcode] - 1; i >= 0; --i)
+				{
+					if (!alu_busy[next_latency][i])
+					{
+						// ADD is implemented as two 1-cycle instructions on a real CPU, so do an additional availability check
+						if ((opcode == ADD) && alu_busy[next_latency + 1][i])
+						{
+							continue;
+						}
+
+						// Rotation can only start when previous rotation is finished, so do an additional availability check
+						if (is_rotation[opcode] && (next_latency < rotate_count * op_latency[opcode]))
+						{
+							continue;
+						}
+
+						alu_index = i;
+						break;
+					}
+				}
+				if (alu_index >= 0)
+				{
+					break;
+				}
+				++next_latency;
+			}
+
+			// Don't generate instructions that leave some register unchanged for more than 7 cycles
+			if (next_latency > latency[a] + 7)
+			{
+				continue;
+			}
+
+			next_latency += op_latency[opcode];
+
+			if (next_latency <= TOTAL_LATENCY)
+			{
+				if (is_rotation[opcode])
+				{
+					++rotate_count;
+				}
+
+				// Mark ALU as busy only for the first cycle when it starts executing the instruction because ALUs are fully pipelined
+				alu_busy[next_latency - op_latency[opcode]][alu_index] = true;
+				latency[a] = next_latency;
+
+				// ASIC is supposed to have enough ALUs to run as many independent instructions per cycle as possible, so latency calculation for ASIC is simple
+				asic_latency[a] = ((asic_latency[a] > asic_latency[b]) ? asic_latency[a] : asic_latency[b]) + asic_op_latency[opcode];
+
+				rotated[a] = is_rotation[opcode];
+
+				inst_data[a] = code_size + (opcode << 8) + ((inst_data[b] & 255) << 16);
+
+				code[code_size].opcode = opcode;
+				code[code_size].dst_index = dst_index;
+				code[code_size].src_index = src_index;
+				code[code_size].C = 0;
+
+				if (src_index == 8)
+				{
+					r8_used = true;
+				}
+
+				if (opcode == ADD)
+				{
+					// ADD instruction is implemented as two 1-cycle instructions on a real CPU, so mark ALU as busy for the next cycle too
+					alu_busy[next_latency - op_latency[opcode] + 1][alu_index] = true;
+
+					// ADD instruction requires 4 more random bytes for 32-bit constant "C" in "a = a + b + C"
+					check_data(&data_index, sizeof(uint32_t), data, sizeof(data));
+					uint32_t t;
+					memcpy(&t, data + data_index, sizeof(uint32_t));
+					code[code_size].C = SWAP32LE(t);
+					data_index += sizeof(uint32_t);
+				}
+
+				++code_size;
+				if (code_size >= NUM_INSTRUCTIONS_MIN)
+				{
+					break;
+				}
+			}
+			else
+			{
+				++num_retries;
+			}
+		}
+
+		// ASIC has more execution resources and can extract as much parallelism from the code as possible
+		// We need to add a few more MUL and ROR instructions to achieve minimal required latency for ASIC
+		// Get this latency for at least 1 of the 4 registers
+		const int prev_code_size = code_size;
+		while ((code_size < NUM_INSTRUCTIONS_MAX) && (asic_latency[0] < TOTAL_LATENCY) && (asic_latency[1] < TOTAL_LATENCY) && (asic_latency[2] < TOTAL_LATENCY) && (asic_latency[3] < TOTAL_LATENCY))
+		{
+			int min_idx = 0;
+			int max_idx = 0;
+			for (int i = 1; i < 4; ++i)
+			{
+				if (asic_latency[i] < asic_latency[min_idx]) min_idx = i;
+				if (asic_latency[i] > asic_latency[max_idx]) max_idx = i;
+			}
+
+			const uint8_t pattern[3] = { ROR, MUL, MUL };
+			const uint8_t opcode = pattern[(code_size - prev_code_size) % 3];
+			latency[min_idx] = latency[max_idx] + op_latency[opcode];
+			asic_latency[min_idx] = asic_latency[max_idx] + asic_op_latency[opcode];
+
+			code[code_size].opcode = opcode;
+			code[code_size].dst_index = min_idx;
+			code[code_size].src_index = max_idx;
+			code[code_size].C = 0;
+			++code_size;
+		}
+
+	// There is ~98.15% chance that loop condition is false, so this loop will execute only 1 iteration most of the time
+	// It never does more than 4 iterations for all block heights < 10,000,000
+	}  while (!r8_used || (code_size < NUM_INSTRUCTIONS_MIN) || (code_size > NUM_INSTRUCTIONS_MAX));
+
+	// It's guaranteed that NUM_INSTRUCTIONS_MIN <= code_size <= NUM_INSTRUCTIONS_MAX here
+	// Add final instruction to stop the interpreter
+	code[code_size].opcode = RET;
+	code[code_size].dst_index = 0;
+	code[code_size].src_index = 0;
+	code[code_size].C = 0;
+
+	return code_size;
+}
+
+#endif

src/cryptonote_basic/cryptonote_format_utils.cpp

@@ -1174,7 +1174,7 @@ namespace cryptonote
     }
     blobdata bd = get_block_hashing_blob(b);
     const int cn_variant = b.major_version >= 7 ? b.major_version - 6 : 0;
-    crypto::cn_slow_hash(bd.data(), bd.size(), res, cn_variant);
+    crypto::cn_slow_hash(bd.data(), bd.size(), res, cn_variant, height);
     return true;
   }
   //---------------------------------------------------------------

tests/hash/CMakeLists.txt

@@ -43,7 +43,7 @@ set_property(TARGET hash-tests
   PROPERTY
     FOLDER "tests")
 
-foreach (hash IN ITEMS fast slow slow-1 slow-2 tree extra-blake extra-groestl extra-jh extra-skein)
+foreach (hash IN ITEMS fast slow slow-1 slow-2 slow-4 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

@@ -44,6 +44,13 @@ using namespace std;
 using namespace crypto;
 typedef crypto::hash chash;
 
+struct V4_Data
+{
+  const void* data;
+  size_t length;
+  uint64_t height;
+};
+
 PUSH_WARNINGS
 DISABLE_VS_WARNINGS(4297)
 extern "C" {
@@ -54,13 +61,17 @@ extern "C" {
     tree_hash((const char (*)[crypto::HASH_SIZE]) 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/*variant*/, 0/*prehashed*/);
+    return cn_slow_hash(data, length, hash, 0/*variant*/, 0/*prehashed*/, 0/*height*/);
   }
   static void cn_slow_hash_1(const void *data, size_t length, char *hash) {
-    return cn_slow_hash(data, length, hash, 1/*variant*/, 0/*prehashed*/);
+    return cn_slow_hash(data, length, hash, 1/*variant*/, 0/*prehashed*/, 0/*height*/);
   }
   static void cn_slow_hash_2(const void *data, size_t length, char *hash) {
-    return cn_slow_hash(data, length, hash, 2/*variant*/, 0/*prehashed*/);
+    return cn_slow_hash(data, length, hash, 2/*variant*/, 0/*prehashed*/, 0/*height*/);
+  }
+  static void cn_slow_hash_4(const void *data, size_t, char *hash) {
+    const V4_Data* p = reinterpret_cast<const V4_Data*>(data);
+    return cn_slow_hash(p->data, p->length, hash, 4/*variant*/, 0/*prehashed*/, p->height);
   }
 }
 POP_WARNINGS
@@ -72,7 +83,7 @@ struct hash_func {
 } 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},
-  {"slow-1", cn_slow_hash_1}, {"slow-2", cn_slow_hash_2}};
+  {"slow-1", cn_slow_hash_1}, {"slow-2", cn_slow_hash_2}, {"slow-4", cn_slow_hash_4}};
 
 int test_variant2_int_sqrt();
 int test_variant2_int_sqrt_ref();
@@ -140,7 +151,15 @@ int main(int argc, char *argv[]) {
     input.exceptions(ios_base::badbit | ios_base::failbit | ios_base::eofbit);
     input.clear(input.rdstate());
     get(input, data);
-    f(data.data(), data.size(), (char *) &actual);
+    if (f == cn_slow_hash_4) {
+      V4_Data d;
+      d.data = data.data();
+      d.length = data.size();
+      get(input, d.height);
+      f(&d, 0, (char *) &actual);
+    } else {
+      f(data.data(), data.size(), (char *) &actual);
+    }
     if (expected != actual) {
       size_t i;
       cerr << "Hash mismatch on test " << test << endl << "Input: ";

tests/hash/tests-slow-4.txt

@@ -0,0 +1,10 @@
+47c996e2d6aa453f50b15a6e829a8c6e5070500c08ba2426019510753e31af42 5468697320697320612074657374205468697320697320612074657374205468697320697320612074657374 1806260
+eb17f755e8f394ff911603826b0e2a37c3f40a5990693a1be7e39cd5c178f0b4 4c6f72656d20697073756d20646f6c6f722073697420616d65742c20636f6e73656374657475722061646970697363696e67 1806261
+f4de6adc61efa498fd4929ed00e88ed8e12caa2907f99cb42442567d3da9daec 656c69742c2073656420646f20656975736d6f642074656d706f7220696e6369646964756e74207574206c61626f7265 1806262
+03004aaa1cdbda343fcbc835aaca191b8577c21267dadd0e4e86a57e68614a71 657420646f6c6f7265206d61676e6120616c697175612e20557420656e696d206164206d696e696d2076656e69616d2c 1806263
+2081cb5646549b44356f5c81787c529367751bc1cdd5f1ea8c0a333b5e49e220 71756973206e6f737472756420657865726369746174696f6e20756c6c616d636f206c61626f726973206e697369 1806264
+653c57f666f6fa1121b82f217485f6fda64ce58bf311d664e92da9119c7d5b95 757420616c697175697020657820656120636f6d6d6f646f20636f6e7365717561742e20447569732061757465 1806265
+20c4b9f8ded7fd1e348341ce2b6297e5ac330e588e5f34985446fd20346b69e3 697275726520646f6c6f7220696e20726570726568656e646572697420696e20766f6c7570746174652076656c6974 1806266
+82e35ae2f7258fb5cb6b53b332f898ac19c385b49fc35e32ae0c5ab56025f763 657373652063696c6c756d20646f6c6f726520657520667567696174206e756c6c612070617269617475722e 1806267
+ddcf95b7d0066668a1d36d4115de1a2bc52dd1f95d94366ec34c8c3c7196e5dd 4578636570746575722073696e74206f6363616563617420637570696461746174206e6f6e2070726f6964656e742c 1806268
+882c2bddf05736ab1072c678c3b75661813709a2ac1fd6e861f2dcc65d466b90 73756e7420696e2063756c706120717569206f666669636961206465736572756e74206d6f6c6c697420616e696d20696420657374206c61626f72756d2e 1806269