RandomX fixes

Intel JCC erratum fix and various other improvements, see more here: https://www.phoronix.com/scan.php?page=article&item=intel-jcc-microcode&num=1

src/App.cpp

@@ -33,6 +33,7 @@
 #include "base/io/Console.h"
 #include "base/io/log/Log.h"
 #include "base/kernel/Signals.h"
+#include "base/kernel/Platform.h"
 #include "core/config/Config.h"
 #include "core/Controller.h"
 #include "core/Miner.h"
@@ -89,6 +90,8 @@ int xmrig::App::exec()
 
     m_controller->start();
 
+    Platform::setThreadPriority(5);
+
     rc = uv_run(uv_default_loop(), UV_RUN_DEFAULT);
     uv_loop_close(uv_default_loop());
 

src/backend/common/Workers.cpp

@@ -109,6 +109,11 @@ void xmrig::Workers<T>::start(const std::vector<T> &data)
 
     for (Thread<T> *worker : m_workers) {
         worker->start(Workers<T>::onReady);
+
+        // This sleep is important for optimal caching!
+        // Threads must allocate scratchpads in order so that adjacent cores will use adjacent scratchpads
+        // Sub-optimal caching can result in up to 0.5% hashrate penalty
+        std::this_thread::sleep_for(std::chrono::milliseconds(20));
     }
 }
 

src/backend/cpu/CpuWorker.cpp

@@ -185,8 +185,20 @@ void xmrig::CpuWorker<N>::start()
             consumeJob();
         }
 
+        uint64_t storeStatsMask = 7;
+
+#       ifdef XMRIG_ALGO_RANDOMX
+        bool first = true;
+        uint64_t tempHash[8] = {};
+
+        // RandomX is faster, we don't need to store stats so often
+        if (m_job.currentJob().algorithm().family() == Algorithm::RANDOM_X) {
+            storeStatsMask = 63;
+        }
+#       endif
+
         while (!Nonce::isOutdated(Nonce::CPU, m_job.sequence())) {
-            if ((m_count & 0x7) == 0) {
+            if ((m_count & storeStatsMask) == 0) {
                 storeStats();
             }
 
@@ -196,26 +208,34 @@ void xmrig::CpuWorker<N>::start()
                 break;
             }
 
+            uint32_t current_job_nonces[N];
+            for (size_t i = 0; i < N; ++i) {
+                current_job_nonces[i] = *m_job.nonce(i);
+            }
+
 #           ifdef XMRIG_ALGO_RANDOMX
             if (job.algorithm().family() == Algorithm::RANDOM_X) {
-                randomx_calculate_hash(m_vm->get(), m_job.blob(), job.size(), m_hash);
+                if (first) {
+                    first = false;
+                    randomx_calculate_hash_first(m_vm->get(), tempHash, m_job.blob(), job.size());
+                }
+                m_job.nextRound(kReserveCount, 1);
+                randomx_calculate_hash_next(m_vm->get(), tempHash, m_job.blob(), job.size(), m_hash);
             }
             else
 #           endif
             {
                 fn(job.algorithm())(m_job.blob(), job.size(), m_hash, m_ctx, job.height());
+                m_job.nextRound(kReserveCount, 1);
             }
 
             for (size_t i = 0; i < N; ++i) {
                 if (*reinterpret_cast<uint64_t*>(m_hash + (i * 32) + 24) < job.target()) {
-                    JobResults::submit(job, *m_job.nonce(i), m_hash + (i * 32));
+                    JobResults::submit(job, current_job_nonces[i], m_hash + (i * 32));
                 }
             }
 
-            m_job.nextRound(kReserveCount, 1);
             m_count += N;
-
-            std::this_thread::yield();
         }
 
         consumeJob();

src/crypto/randomx/aes_hash.cpp

@@ -212,3 +212,84 @@ void fillAes4Rx4(void *state, size_t outputSize, void *buffer) {
 
 template void fillAes4Rx4<true>(void *state, size_t outputSize, void *buffer);
 template void fillAes4Rx4<false>(void *state, size_t outputSize, void *buffer);
+
+template<bool softAes>
+void hashAndFillAes1Rx4(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state) {
+	uint8_t* scratchpadPtr = (uint8_t*)scratchpad;
+	const uint8_t* scratchpadEnd = scratchpadPtr + scratchpadSize;
+
+	// initial state
+	rx_vec_i128 hash_state0 = rx_set_int_vec_i128(AES_HASH_1R_STATE0);
+	rx_vec_i128 hash_state1 = rx_set_int_vec_i128(AES_HASH_1R_STATE1);
+	rx_vec_i128 hash_state2 = rx_set_int_vec_i128(AES_HASH_1R_STATE2);
+	rx_vec_i128 hash_state3 = rx_set_int_vec_i128(AES_HASH_1R_STATE3);
+
+	const rx_vec_i128 key0 = rx_set_int_vec_i128(AES_GEN_1R_KEY0);
+	const rx_vec_i128 key1 = rx_set_int_vec_i128(AES_GEN_1R_KEY1);
+	const rx_vec_i128 key2 = rx_set_int_vec_i128(AES_GEN_1R_KEY2);
+	const rx_vec_i128 key3 = rx_set_int_vec_i128(AES_GEN_1R_KEY3);
+
+	rx_vec_i128 fill_state0 = rx_load_vec_i128((rx_vec_i128*)fill_state + 0);
+	rx_vec_i128 fill_state1 = rx_load_vec_i128((rx_vec_i128*)fill_state + 1);
+	rx_vec_i128 fill_state2 = rx_load_vec_i128((rx_vec_i128*)fill_state + 2);
+	rx_vec_i128 fill_state3 = rx_load_vec_i128((rx_vec_i128*)fill_state + 3);
+
+	constexpr int PREFETCH_DISTANCE = 4096;
+	const char* prefetchPtr = ((const char*)scratchpad) + PREFETCH_DISTANCE;
+	scratchpadEnd -= PREFETCH_DISTANCE;
+
+	for (int i = 0; i < 2; ++i) {
+		//process 64 bytes at a time in 4 lanes
+		while (scratchpadPtr < scratchpadEnd) {
+			hash_state0 = aesenc<softAes>(hash_state0, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 0));
+			hash_state1 = aesdec<softAes>(hash_state1, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 1));
+			hash_state2 = aesenc<softAes>(hash_state2, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 2));
+			hash_state3 = aesdec<softAes>(hash_state3, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 3));
+
+			fill_state0 = aesdec<softAes>(fill_state0, key0);
+			fill_state1 = aesenc<softAes>(fill_state1, key1);
+			fill_state2 = aesdec<softAes>(fill_state2, key2);
+			fill_state3 = aesenc<softAes>(fill_state3, key3);
+
+			rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 0, fill_state0);
+			rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 1, fill_state1);
+			rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 2, fill_state2);
+			rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 3, fill_state3);
+
+			rx_prefetch_t0(prefetchPtr);
+
+			scratchpadPtr += 64;
+			prefetchPtr += 64;
+		}
+		prefetchPtr = (const char*) scratchpad;
+		scratchpadEnd += PREFETCH_DISTANCE;
+	}
+
+	rx_store_vec_i128((rx_vec_i128*)fill_state + 0, fill_state0);
+	rx_store_vec_i128((rx_vec_i128*)fill_state + 1, fill_state1);
+	rx_store_vec_i128((rx_vec_i128*)fill_state + 2, fill_state2);
+	rx_store_vec_i128((rx_vec_i128*)fill_state + 3, fill_state3);
+
+	//two extra rounds to achieve full diffusion
+	rx_vec_i128 xkey0 = rx_set_int_vec_i128(AES_HASH_1R_XKEY0);
+	rx_vec_i128 xkey1 = rx_set_int_vec_i128(AES_HASH_1R_XKEY1);
+
+	hash_state0 = aesenc<softAes>(hash_state0, xkey0);
+	hash_state1 = aesdec<softAes>(hash_state1, xkey0);
+	hash_state2 = aesenc<softAes>(hash_state2, xkey0);
+	hash_state3 = aesdec<softAes>(hash_state3, xkey0);
+
+	hash_state0 = aesenc<softAes>(hash_state0, xkey1);
+	hash_state1 = aesdec<softAes>(hash_state1, xkey1);
+	hash_state2 = aesenc<softAes>(hash_state2, xkey1);
+	hash_state3 = aesdec<softAes>(hash_state3, xkey1);
+
+	//output hash
+	rx_store_vec_i128((rx_vec_i128*)hash + 0, hash_state0);
+	rx_store_vec_i128((rx_vec_i128*)hash + 1, hash_state1);
+	rx_store_vec_i128((rx_vec_i128*)hash + 2, hash_state2);
+	rx_store_vec_i128((rx_vec_i128*)hash + 3, hash_state3);
+}
+
+template void hashAndFillAes1Rx4<false>(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state);
+template void hashAndFillAes1Rx4<true>(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state);

src/crypto/randomx/aes_hash.hpp

@@ -38,3 +38,6 @@ void fillAes1Rx4(void *state, size_t outputSize, void *buffer);
 
 template<bool softAes>
 void fillAes4Rx4(void *state, size_t outputSize, void *buffer);
+
+template<bool softAes>
+void hashAndFillAes1Rx4(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state);

src/crypto/randomx/intrin_portable.h

@@ -102,6 +102,7 @@ typedef __m128d rx_vec_f128;
 #define rx_aligned_alloc(a, b) _mm_malloc(a,b)
 #define rx_aligned_free(a) _mm_free(a)
 #define rx_prefetch_nta(x) _mm_prefetch((const char *)(x), _MM_HINT_NTA)
+#define rx_prefetch_t0(x) _mm_prefetch((const char *)(x), _MM_HINT_T0)
 
 #define rx_load_vec_f128 _mm_load_pd
 #define rx_store_vec_f128 _mm_store_pd
@@ -201,6 +202,7 @@ typedef union{
 #define rx_aligned_alloc(a, b) malloc(a)
 #define rx_aligned_free(a) free(a)
 #define rx_prefetch_nta(x)
+#define rx_prefetch_t0(x)
 
 /* Splat 64-bit long long to 2 64-bit long longs */
 FORCE_INLINE __m128i vec_splat2sd (int64_t scalar)
@@ -399,6 +401,10 @@ inline void rx_prefetch_nta(void* ptr) {
 	asm volatile ("prfm pldl1strm, [%0]\n" : : "r" (ptr));
 }
 
+inline void rx_prefetch_t0(const void* ptr) {
+	asm volatile ("prfm pldl1strm, [%0]\n" : : "r" (ptr));
+}
+
 FORCE_INLINE rx_vec_f128 rx_load_vec_f128(const double* pd) {
 	return vld1q_f64((const float64_t*)pd);
 }
@@ -532,6 +538,7 @@ typedef union {
 #define rx_aligned_alloc(a, b) malloc(a)
 #define rx_aligned_free(a) free(a)
 #define rx_prefetch_nta(x)
+#define rx_prefetch_t0(x)
 
 FORCE_INLINE rx_vec_f128 rx_load_vec_f128(const double* pd) {
 	rx_vec_f128 x;

src/crypto/randomx/jit_compiler_x86.cpp

@@ -29,6 +29,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include <stdexcept>
 #include <cstring>
 #include <climits>
+#include <atomic>
 #include "crypto/randomx/jit_compiler_x86.hpp"
 #include "crypto/randomx/jit_compiler_x86_static.hpp"
 #include "crypto/randomx/superscalar.hpp"
@@ -36,6 +37,12 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "crypto/randomx/reciprocal.h"
 #include "crypto/randomx/virtual_memory.hpp"
 
+#ifdef _MSC_VER
+#   include <intrin.h>
+#else
+#   include <cpuid.h>
+#endif
+
 namespace randomx {
 	/*
 
@@ -108,7 +115,7 @@ namespace randomx {
 	const int32_t codeSshPrefetchSize = codeShhEnd - codeShhPrefetch;
 	const int32_t codeSshInitSize = codeProgramEnd - codeShhInit;
 
-	const int32_t epilogueOffset = CodeSize - epilogueSize;
+	const int32_t epilogueOffset = (CodeSize - epilogueSize) & ~63;
 	constexpr int32_t superScalarHashOffset = 32768;
 
 	static const uint8_t REX_ADD_RR[] = { 0x4d, 0x03 };
@@ -183,6 +190,7 @@ namespace randomx {
 	static const uint8_t REX_ADD_I[] = { 0x49, 0x81 };
 	static const uint8_t REX_TEST[] = { 0x49, 0xF7 };
 	static const uint8_t JZ[] = { 0x0f, 0x84 };
+	static const uint8_t JZ_SHORT = 0x74;
 	static const uint8_t RET = 0xc3;
 	static const uint8_t LEA_32[] = { 0x41, 0x8d };
 	static const uint8_t MOVNTI[] = { 0x4c, 0x0f, 0xc3 };
@@ -197,20 +205,100 @@ namespace randomx {
 	static const uint8_t NOP7[] = { 0x0F, 0x1F, 0x80, 0x00, 0x00, 0x00, 0x00 };
 	static const uint8_t NOP8[] = { 0x0F, 0x1F, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00 };
 
-//	static const uint8_t* NOPX[] = { NOP1, NOP2, NOP3, NOP4, NOP5, NOP6, NOP7, NOP8 };
+	static const uint8_t* NOPX[] = { NOP1, NOP2, NOP3, NOP4, NOP5, NOP6, NOP7, NOP8 };
+
+	static const uint8_t JMP_ALIGN_PREFIX[14][16] = {
+		{},
+		{0x2E},
+		{0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x90, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x66, 0x90, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x66, 0x66, 0x90, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x0F, 0x1F, 0x40, 0x00, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+		{0x0F, 0x1F, 0x44, 0x00, 0x00, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E, 0x2E},
+	};
+
+	bool JitCompilerX86::BranchesWithin32B = false;
 
 	size_t JitCompilerX86::getCodeSize() {
 		return codePos < prologueSize ? 0 : codePos - prologueSize;
 	}
 
+    static inline void cpuid(uint32_t level, int32_t output[4])
+    {
+        memset(output, 0, sizeof(int32_t) * 4);
+
+#   ifdef _MSC_VER
+        __cpuid(output, static_cast<int>(level));
+#   else
+        __cpuid_count(level, 0, output[0], output[1], output[2], output[3]);
+#   endif
+    }
+
+    // CPU-specific tweaks
+	void JitCompilerX86::applyTweaks() {
+		int32_t info[4];
+		cpuid(0, info);
+
+		int32_t manufacturer[4];
+		manufacturer[0] = info[1];
+		manufacturer[1] = info[3];
+		manufacturer[2] = info[2];
+		manufacturer[3] = 0;
+
+		if (strcmp((const char*)manufacturer, "GenuineIntel") == 0) {
+			struct
+			{
+				unsigned int stepping : 4;
+				unsigned int model : 4;
+				unsigned int family : 4;
+				unsigned int processor_type : 2;
+				unsigned int reserved1 : 2;
+				unsigned int ext_model : 4;
+				unsigned int ext_family : 8;
+				unsigned int reserved2 : 4;
+			} processor_info;
+
+			cpuid(1, info);
+			memcpy(&processor_info, info, sizeof(processor_info));
+
+			// Intel JCC erratum mitigation
+			if (processor_info.family == 6) {
+				const uint32_t model = processor_info.model | (processor_info.ext_model << 4);
+				const uint32_t stepping = processor_info.stepping;
+
+				// Affected CPU models and stepping numbers are taken from https://www.intel.com/content/dam/support/us/en/documents/processors/mitigations-jump-conditional-code-erratum.pdf
+				BranchesWithin32B =
+					((model == 0x4E) && (stepping == 0x3)) ||
+					((model == 0x55) && (stepping == 0x4)) ||
+					((model == 0x5E) && (stepping == 0x3)) ||
+					((model == 0x8E) && (stepping >= 0x9) && (stepping <= 0xC)) ||
+					((model == 0x9E) && (stepping >= 0x9) && (stepping <= 0xD)) ||
+					((model == 0xA6) && (stepping == 0x0)) ||
+					((model == 0xAE) && (stepping == 0xA));
+			}
+		}
+	}
+
+	static std::atomic<size_t> codeOffset;
+
 	JitCompilerX86::JitCompilerX86() {
-		code = (uint8_t*)allocExecutableMemory(CodeSize);
+		applyTweaks();
+		allocatedCode = (uint8_t*)allocExecutableMemory(CodeSize * 2);
+		// Shift code base address to improve caching - all threads will use different L2/L3 cache sets
+		code = allocatedCode + (codeOffset.fetch_add(59 * 64) % CodeSize);
 		memcpy(code, codePrologue, prologueSize);
 		memcpy(code + epilogueOffset, codeEpilogue, epilogueSize);
 	}
 
 	JitCompilerX86::~JitCompilerX86() {
-		freePagedMemory(code, CodeSize);
+		freePagedMemory(allocatedCode, CodeSize);
 	}
 
 	void JitCompilerX86::generateProgram(Program& prog, ProgramConfiguration& pcfg) {
@@ -307,6 +395,22 @@ namespace randomx {
 		emit(RandomX_CurrentConfig.codePrefetchScratchpadTweaked, prefetchScratchpadSize, code, codePos);
 		memcpy(code + codePos, codeLoopStore, loopStoreSize);
 		codePos += loopStoreSize;
+
+		if (BranchesWithin32B) {
+			const uint32_t branch_begin = static_cast<uint32_t>(codePos);
+			const uint32_t branch_end = static_cast<uint32_t>(branch_begin + 9);
+
+			// If the jump crosses or touches 32-byte boundary, align it
+			if ((branch_begin ^ branch_end) >= 32) {
+				uint32_t alignment_size = 32 - (branch_begin & 31);
+				if (alignment_size > 8) {
+					emit(NOPX[alignment_size - 9], alignment_size - 8, code, codePos);
+					alignment_size = 8;
+				}
+				emit(NOPX[alignment_size - 1], alignment_size, code, codePos);
+			}
+		}
+
 		emit(SUB_EBX, code, codePos);
 		emit(JNZ, code, codePos);
 		emit32(prologueSize - codePos - 4, code, codePos);
@@ -408,12 +512,13 @@ namespace randomx {
 		}
 	}
 
-	void JitCompilerX86::genAddressReg(const Instruction& instr, uint8_t* code, int& codePos, bool rax) {
+	template<bool rax>
-		emit(LEA_32, code, codePos);
+	FORCE_INLINE void JitCompilerX86::genAddressReg(const Instruction& instr, uint8_t* code, int& codePos) {
-		emitByte(0x80 + instr.src + (rax ? 0 : 8), code, codePos);
+		const uint32_t src = *((uint32_t*)&instr) & 0xFF0000;
-		if (instr.src == RegisterNeedsSib) {
+
-			emitByte(0x24, code, codePos);
+		*(uint32_t*)(code + codePos) = (rax ? 0x24808d41 : 0x24888d41) + src;
-		}
+		codePos += (src == (RegisterNeedsSib << 16)) ? 4 : 3;
+
 		emit32(instr.getImm32(), code, codePos);
 		if (rax)
 			emitByte(AND_EAX_I, code, codePos);
@@ -422,12 +527,14 @@ namespace randomx {
 		emit32(instr.getModMem() ? ScratchpadL1Mask : ScratchpadL2Mask, code, codePos);
 	}
 
-	void JitCompilerX86::genAddressRegDst(const Instruction& instr, uint8_t* code, int& codePos) {
+	template void JitCompilerX86::genAddressReg<false>(const Instruction& instr, uint8_t* code, int& codePos);
-		emit(LEA_32, code, codePos);
+	template void JitCompilerX86::genAddressReg<true>(const Instruction& instr, uint8_t* code, int& codePos);
-		emitByte(0x80 + instr.dst, code, codePos);
+
-		if (instr.dst == RegisterNeedsSib) {
+	FORCE_INLINE void JitCompilerX86::genAddressRegDst(const Instruction& instr, uint8_t* code, int& codePos) {
-			emitByte(0x24, code, codePos);
+		const uint32_t dst = static_cast<uint32_t>(instr.dst) << 16;
-		}
+		*(uint32_t*)(code + codePos) = 0x24808d41 + dst;
+		codePos += (dst == (RegisterNeedsSib << 16)) ? 4 : 3;
+
 		emit32(instr.getImm32(), code, codePos);
 		emitByte(AND_EAX_I, code, codePos);
 		if (instr.getModCond() < StoreL3Condition) {
@@ -438,7 +545,7 @@ namespace randomx {
 		}
 	}
 
-	void JitCompilerX86::genAddressImm(const Instruction& instr, uint8_t* code, int& codePos) {
+	FORCE_INLINE void JitCompilerX86::genAddressImm(const Instruction& instr, uint8_t* code, int& codePos) {
 		emit32(instr.getImm32() & ScratchpadL3Mask, code, codePos);
 	}
 
@@ -483,7 +590,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos);
+			genAddressReg<true>(instr, p, pos);
 			emit32(template_IADD_M[instr.dst], p, pos);
 		}
 		else {
@@ -523,7 +630,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos);
+			genAddressReg<true>(instr, p, pos);
 			emit(REX_SUB_RM, p, pos);
 			emitByte(0x04 + 8 * instr.dst, p, pos);
 			emitByte(0x06, p, pos);
@@ -561,7 +668,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos);
+			genAddressReg<true>(instr, p, pos);
 			emit(REX_IMUL_RM, p, pos);
 			emitByte(0x04 + 8 * instr.dst, p, pos);
 			emitByte(0x06, p, pos);
@@ -596,7 +703,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos, false);
+			genAddressReg<false>(instr, p, pos);
 			emit(REX_MOV_RR64, p, pos);
 			emitByte(0xc0 + instr.dst, p, pos);
 			emit(REX_MUL_MEM, p, pos);
@@ -635,7 +742,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos, false);
+			genAddressReg<false>(instr, p, pos);
 			emit(REX_MOV_RR64, p, pos);
 			emitByte(0xc0 + instr.dst, p, pos);
 			emit(REX_IMUL_MEM, p, pos);
@@ -704,7 +811,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, p, pos);
+			genAddressReg<true>(instr, p, pos);
 			emit(REX_XOR_RM, p, pos);
 			emitByte(0x04 + 8 * instr.dst, p, pos);
 			emitByte(0x06, p, pos);
@@ -801,7 +908,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		const uint32_t dst = instr.dst % RegisterCountFlt;
-		genAddressReg(instr, p, pos);
+		genAddressReg<true>(instr, p, pos);
 		emit(REX_CVTDQ2PD_XMM12, p, pos);
 		emit(REX_ADDPD, p, pos);
 		emitByte(0xc4 + 8 * dst, p, pos);
@@ -826,7 +933,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		const uint32_t dst = instr.dst % RegisterCountFlt;
-		genAddressReg(instr, p, pos);
+		genAddressReg<true>(instr, p, pos);
 		emit(REX_CVTDQ2PD_XMM12, p, pos);
 		emit(REX_SUBPD, p, pos);
 		emitByte(0xc4 + 8 * dst, p, pos);
@@ -862,7 +969,7 @@ namespace randomx {
 		int pos = codePos;
 		
 		const uint32_t dst = instr.dst % RegisterCountFlt;
-		genAddressReg(instr, p, pos);
+		genAddressReg<true>(instr, p, pos);
 		emit(REX_CVTDQ2PD_XMM12, p, pos);
 		emit(REX_ANDPS_XMM12, p, pos);
 		emit(REX_DIVPD, p, pos);
@@ -902,19 +1009,39 @@ namespace randomx {
 		uint8_t* const p = code;
 		int pos = codePos;
 		
-		int reg = instr.dst;
+		const int reg = instr.dst;
+		int32_t jmp_offset = registerUsage[reg] - (pos + 16);
+
+		if (BranchesWithin32B) {
+			const uint32_t branch_begin = static_cast<uint32_t>(pos + 7);
+			const uint32_t branch_end = static_cast<uint32_t>(branch_begin + ((jmp_offset >= -128) ? 9 : 13));
+
+			// If the jump crosses or touches 32-byte boundary, align it
+			if ((branch_begin ^ branch_end) >= 32) {
+				const uint32_t alignment_size = 32 - (branch_begin & 31);
+				jmp_offset -= alignment_size;
+				emit(JMP_ALIGN_PREFIX[alignment_size], alignment_size, p, pos);
+			}
+		}
+
 		emit(REX_ADD_I, p, pos);
 		emitByte(0xc0 + reg, p, pos);
-		int shift = instr.getModCond() + RandomX_CurrentConfig.JumpOffset;
+		const int shift = instr.getModCond() + RandomX_CurrentConfig.JumpOffset;
-		uint32_t imm = instr.getImm32() | (1UL << shift);
+		const uint32_t imm = (instr.getImm32() | (1UL << shift)) & ~(1UL << (shift - 1));
-		if (RandomX_CurrentConfig.JumpOffset > 0 || shift > 0)
-			imm &= ~(1UL << (shift - 1));
 		emit32(imm, p, pos);
 		emit(REX_TEST, p, pos);
 		emitByte(0xc0 + reg, p, pos);
 		emit32(RandomX_CurrentConfig.ConditionMask_Calculated << shift, p, pos);
+
+		if (jmp_offset >= -128) {
+			emitByte(JZ_SHORT, p, pos);
+			emitByte(jmp_offset, p, pos);
+		}
+		else {
 			emit(JZ, p, pos);
-		emit32(registerUsage[reg] - (pos + 4), p, pos);
+			emit32(jmp_offset - 4, p, pos);
+		}
+
 		//mark all registers as used
 		uint64_t* r = (uint64_t*) registerUsage;
 		uint64_t k = pos;

src/crypto/randomx/jit_compiler_x86.hpp

@@ -67,12 +67,17 @@ namespace randomx {
 
 		static InstructionGeneratorX86 engine[256];
 		int registerUsage[RegistersCount];
+		uint8_t* allocatedCode;
 		uint8_t* code;
 		int32_t codePos;
 
+		static bool BranchesWithin32B;
+
+		static void applyTweaks();
 		void generateProgramPrologue(Program&, ProgramConfiguration&);
 		void generateProgramEpilogue(Program&, ProgramConfiguration&);
-		static void genAddressReg(const Instruction&, uint8_t* code, int& codePos, bool rax = true);
+		template<bool rax>
+		static void genAddressReg(const Instruction&, uint8_t* code, int& codePos);
 		static void genAddressRegDst(const Instruction&, uint8_t* code, int& codePos);
 		static void genAddressImm(const Instruction&, uint8_t* code, int& codePos);
 		static void genSIB(int scale, int index, int base, uint8_t* code, int& codePos);

src/crypto/randomx/randomx.cpp

@@ -473,4 +473,22 @@ extern "C" {
 		machine->getFinalResult(output, RANDOMX_HASH_SIZE);
 	}
 
+	void randomx_calculate_hash_first(randomx_vm* machine, uint64_t (&tempHash)[8], const void* input, size_t inputSize) {
+		rx_blake2b(tempHash, sizeof(tempHash), input, inputSize, nullptr, 0);
+		machine->initScratchpad(tempHash);
+	}
+
+	void randomx_calculate_hash_next(randomx_vm* machine, uint64_t (&tempHash)[8], const void* nextInput, size_t nextInputSize, void* output) {
+		machine->resetRoundingMode();
+		for (uint32_t chain = 0; chain < RandomX_CurrentConfig.ProgramCount - 1; ++chain) {
+			machine->run(&tempHash);
+			rx_blake2b(tempHash, sizeof(tempHash), machine->getRegisterFile(), sizeof(randomx::RegisterFile), nullptr, 0);
+		}
+		machine->run(&tempHash);
+
+		// Finish current hash and fill the scratchpad for the next hash at the same time
+		rx_blake2b(tempHash, sizeof(tempHash), nextInput, nextInputSize, nullptr, 0);
+		machine->hashAndFill(output, RANDOMX_HASH_SIZE, tempHash);
+	}
+
 }

src/crypto/randomx/randomx.h

@@ -338,6 +338,9 @@ RANDOMX_EXPORT void randomx_destroy_vm(randomx_vm *machine);
 */
 RANDOMX_EXPORT void randomx_calculate_hash(randomx_vm *machine, const void *input, size_t inputSize, void *output);
 
+RANDOMX_EXPORT void randomx_calculate_hash_first(randomx_vm* machine, uint64_t (&tempHash)[8], const void* input, size_t inputSize);
+RANDOMX_EXPORT void randomx_calculate_hash_next(randomx_vm* machine, uint64_t (&tempHash)[8], const void* nextInput, size_t nextInputSize, void* output);
+
 #if defined(__cplusplus)
 }
 #endif

src/crypto/randomx/virtual_machine.cpp

@@ -114,6 +114,12 @@ namespace randomx {
         rx_blake2b(out, outSize, &reg, sizeof(RegisterFile), nullptr, 0);
 	}
 
+	template<bool softAes>
+	void VmBase<softAes>::hashAndFill(void* out, size_t outSize, uint64_t (&fill_state)[8]) {
+		hashAndFillAes1Rx4<softAes>(scratchpad, ScratchpadSize, &reg.a, fill_state);
+        rx_blake2b(out, outSize, &reg, sizeof(RegisterFile), nullptr, 0);
+	}
+
 	template<bool softAes>
 	void VmBase<softAes>::initScratchpad(void* seed) {
 		fillAes1Rx4<softAes>(seed, ScratchpadSize, scratchpad);

src/crypto/randomx/virtual_machine.hpp

@@ -39,6 +39,7 @@ public:
 	virtual ~randomx_vm() = 0;
 	virtual void setScratchpad(uint8_t *scratchpad) = 0;
 	virtual void getFinalResult(void* out, size_t outSize) = 0;
+	virtual void hashAndFill(void* out, size_t outSize, uint64_t (&fill_state)[8]) = 0;
 	virtual void setDataset(randomx_dataset* dataset) { }
 	virtual void setCache(randomx_cache* cache) { }
 	virtual void initScratchpad(void* seed) = 0;
@@ -82,6 +83,7 @@ namespace randomx {
 		void setScratchpad(uint8_t *scratchpad) override;
 		void initScratchpad(void* seed) override;
 		void getFinalResult(void* out, size_t outSize) override;
+		void hashAndFill(void* out, size_t outSize, uint64_t (&fill_state)[8]) override;
 
 	protected:
 		void generateProgram(void* seed);