Merge branch 'dev'

CHANGELOG.md

@@ -1,3 +1,13 @@
+# v3.1.1
+- [#1133](https://github.com/xmrig/xmrig/issues/1133) Fixed syslog regression.
+- [#1138](https://github.com/xmrig/xmrig/issues/1138) Fixed multiple network bugs.
+- [#1141](https://github.com/xmrig/xmrig/issues/1141) Fixed log in background mode.
+- [#1142](https://github.com/xmrig/xmrig/pull/1142) RandomX hashrate improved by 0.5-1.5% depending on variant and CPU.
+- [#1146](https://github.com/xmrig/xmrig/pull/1146) Fixed race condition in RandomX thread init.
+- [#1148](https://github.com/xmrig/xmrig/pull/1148) Fixed, on Linux linker marking entire executable as having an executable stack.
+- Fixed, for Argon2 algorithms command line options like `--threads` was ignored.
+- Fixed command line options for single pool, free order allowed again.
+
 # v3.1.0
 - [#1107](https://github.com/xmrig/xmrig/issues/1107#issuecomment-522235892) Added Argon2 algorithm family: `argon2/chukwa` and `argon2/wrkz`.
 

src/backend/cpu/CpuWorker.cpp

@@ -90,6 +90,8 @@ void xmrig::CpuWorker<N>::allocateRandomX_VM()
         if (Nonce::sequence(Nonce::CPU) == 0) {
             return;
         }
+
+        dataset = Rx::dataset(m_job.currentJob(), m_node);
     }
 
     if (!m_vm) {

src/base/io/log/Log.cpp

@@ -91,6 +91,10 @@ public:
 
         std::lock_guard<std::mutex> lock(m_mutex);
 
+        if (Log::background && m_backends.empty()) {
+            return;
+        }
+
         timestamp(level, size, offset);
         color(level, size);
 
@@ -111,7 +115,7 @@ public:
         if (!m_backends.empty()) {
             for (ILogBackend *backend : m_backends) {
                 backend->print(level, m_buf, offset, size, true);
-                backend->print(level, txt.c_str(), offset, txt.size(), false);
+                backend->print(level, txt.c_str(), offset ? (offset - 11) : 0, txt.size(), false);
             }
         }
         else {
@@ -190,8 +194,9 @@ private:
 };
 
 
-bool Log::colors   = true;
-LogPrivate *Log::d = new LogPrivate();
+bool Log::background = false;
+bool Log::colors     = true;
+LogPrivate *Log::d   = new LogPrivate();
 
 
 } /* namespace xmrig */

src/base/io/log/Log.h

@@ -54,6 +54,7 @@ public:
     static void print(const char *fmt, ...);
     static void print(Level level, const char *fmt, ...);
 
+    static bool background;
     static bool colors;
 
 private:

src/base/io/log/backends/SysLog.h

@@ -37,7 +37,7 @@ class SysLog : public ILogBackend
 {
 public:
     SysLog();
-    ~SysLog();
+    ~SysLog() override;
 
 protected:
     void print(int level, const char *line, size_t offset, size_t size, bool colors) override;

src/base/kernel/Base.cpp

@@ -184,7 +184,10 @@ int xmrig::Base::init()
     Platform::setProcessPriority(config()->cpu().priority());
 #   endif
 
-    if (!config()->isBackground()) {
+    if (config()->isBackground()) {
+        Log::background = true;
+    }
+    else {
         Log::add(new ConsoleLog());
     }
 

src/base/kernel/config/BaseTransform.cpp

@@ -33,11 +33,12 @@
 #endif
 
 
-#include "base/kernel/config/BaseTransform.h"
-#include "base/kernel/Process.h"
-#include "base/io/log/Log.h"
-#include "base/kernel/interfaces/IConfig.h"
 #include "base/io/json/JsonChain.h"
+#include "base/io/log/Log.h"
+#include "base/kernel/config/BaseTransform.h"
+#include "base/kernel/interfaces/IConfig.h"
+#include "base/kernel/Process.h"
+#include "base/net/stratum/Pool.h"
 #include "core/config/Config_platform.h"
 
 
@@ -138,7 +139,19 @@ void xmrig::BaseTransform::transform(rapidjson::Document &doc, int key, const ch
         break;
 
     case IConfig::UrlKey: /* --url */
-        return add(doc, kPools, "url", arg, true);
+    {
+        if (!doc.HasMember(kPools)) {
+            doc.AddMember(rapidjson::StringRef(kPools), rapidjson::kArrayType, doc.GetAllocator());
+        }
+
+        rapidjson::Value &array = doc[kPools];
+        if (array.Size() == 0 || Pool(array[array.Size() - 1]).isValid()) {
+            array.PushBack(rapidjson::kObjectType, doc.GetAllocator());
+        }
+
+        set(doc, array[array.Size() - 1], "url", arg);
+        break;
+    }
 
     case IConfig::UserKey: /* --user */
         return add(doc, kPools, "user", arg);

src/base/net/stratum/BaseClient.cpp

@@ -42,7 +42,8 @@ xmrig::BaseClient::BaseClient(int id, IClientListener *listener) :
     m_retries(5),
     m_failures(0),
     m_state(UnconnectedState),
-    m_retryPause(5000)
+    m_retryPause(5000),
+    m_enabled(true)
 {
 }
 

src/base/net/stratum/BaseClient.h

@@ -64,7 +64,8 @@ protected:
         HostLookupState,
         ConnectingState,
         ConnectedState,
-        ClosingState
+        ClosingState,
+        ReconnectingState
     };
 
     inline bool isQuiet() const { return m_quiet || m_failures >= m_retries; }

src/base/net/stratum/Client.cpp

@@ -70,21 +70,15 @@ static const char *states[] = {
     "host-lookup",
     "connecting",
     "connected",
-    "closing"
+    "closing",
+    "reconnecting"
 };
 #endif
 
 
 xmrig::Client::Client(int id, const char *agent, IClientListener *listener) :
     BaseClient(id, listener),
-    m_agent(agent),
-    m_tls(nullptr),
-    m_expire(0),
-    m_jobs(0),
-    m_keepAlive(0),
-    m_key(0),
-    m_stream(nullptr),
-    m_socket(nullptr)
+    m_agent(agent)
 {
     m_key = m_storage.add(this);
     m_dns = new Dns(this);
@@ -234,10 +228,16 @@ void xmrig::Client::tick(uint64_t now)
         else if (m_keepAlive && now > m_keepAlive) {
             ping();
         }
+
+        return;
     }
 
-    if (m_expire && now > m_expire && m_state == ConnectingState) {
-        connect();
+    if (m_state == ReconnectingState && m_expire && now > m_expire) {
+        return connect();
+    }
+
+    if (m_state == ConnectingState && m_expire && now > m_expire) {
+        return reconnect();
     }
 }
 
@@ -447,7 +447,6 @@ int xmrig::Client::resolve(const String &host)
 {
     setState(HostLookupState);
 
-    m_expire = 0;
     m_recvBuf.reset();
 
     if (m_failures == -1) {
@@ -754,6 +753,8 @@ void xmrig::Client::parseResponse(int64_t id, const rapidjson::Value &result, co
 void xmrig::Client::ping()
 {
     send(snprintf(m_sendBuf, sizeof(m_sendBuf), "{\"id\":%" PRId64 ",\"jsonrpc\":\"2.0\",\"method\":\"keepalived\",\"params\":{\"id\":\"%s\"}}\n", m_sequence, m_rpcId.data()));
+
+    m_keepAlive = 0;
 }
 
 
@@ -810,12 +811,10 @@ void xmrig::Client::reconnect()
         return m_listener->onClose(this, -1);
     }
 
-    setState(ConnectingState);
+    setState(ReconnectingState);
 
     m_failures++;
     m_listener->onClose(this, static_cast<int>(m_failures));
-
-    m_expire = Chrono::steadyMSecs() + m_retryPause;
 }
 
 
@@ -827,6 +826,23 @@ void xmrig::Client::setState(SocketState state)
         return;
     }
 
+    switch (state) {
+    case HostLookupState:
+        m_expire = 0;
+        break;
+
+    case ConnectingState:
+        m_expire = Chrono::steadyMSecs() + kConnectTimeout;
+        break;
+
+    case ReconnectingState:
+        m_expire = Chrono::steadyMSecs() + m_retryPause;
+        break;
+
+    default:
+        break;
+    }
+
     m_state = state;
 }
 

src/base/net/stratum/Client.h

@@ -56,12 +56,13 @@ class JobResult;
 class Client : public BaseClient, public IDnsListener, public ILineListener
 {
 public:
-    constexpr static int kResponseTimeout = 20 * 1000;
+    constexpr static uint64_t kConnectTimeout  = 20 * 1000;
+    constexpr static uint64_t kResponseTimeout = 20 * 1000;
 
 #   ifdef XMRIG_FEATURE_TLS
-    constexpr static int kInputBufferSize = 1024 * 16;
+    constexpr static size_t kInputBufferSize = 1024 * 16;
 #   else
-    constexpr static int kInputBufferSize = 1024 * 2;
+    constexpr static size_t kInputBufferSize = 1024 * 2;
 #   endif
 
     Client(int id, const char *agent, IClientListener *listener);
@@ -122,19 +123,19 @@ private:
 
     static inline Client *getClient(void *data) { return m_storage.get(data); }
 
-    char m_sendBuf[2048];
+    char m_sendBuf[2048] = { 0 };
     const char *m_agent;
     Dns *m_dns;
     RecvBuf<kInputBufferSize> m_recvBuf;
     std::bitset<EXT_MAX> m_extensions;
     String m_rpcId;
-    Tls *m_tls;
-    uint64_t m_expire;
-    uint64_t m_jobs;
-    uint64_t m_keepAlive;
-    uintptr_t m_key;
-    uv_stream_t *m_stream;
-    uv_tcp_t *m_socket;
+    Tls *m_tls                  = nullptr;
+    uint64_t m_expire           = 0;
+    uint64_t m_jobs             = 0;
+    uint64_t m_keepAlive        = 0;
+    uintptr_t m_key             = 0;
+    uv_stream_t *m_stream       = nullptr;
+    uv_tcp_t *m_socket          = nullptr;
 
     static Storage<Client> m_storage;
 };

src/base/net/stratum/Pool.cpp

@@ -200,6 +200,9 @@ bool xmrig::Pool::isEqual(const Pool &other) const
 bool xmrig::Pool::parse(const char *url)
 {
     assert(url != nullptr);
+    if (url == nullptr) {
+        return false;
+    }
 
     const char *p    = strstr(url, "://");
     const char *base = url;

src/core/config/ConfigTransform.cpp

@@ -96,6 +96,8 @@ void xmrig::ConfigTransform::finalize(rapidjson::Document &doc)
     BaseTransform::finalize(doc);
 
     if (m_threads) {
+        doc.AddMember("version", 1, allocator);
+
         if (!doc.HasMember(kCpu)) {
             doc.AddMember(StringRef(kCpu), Value(kObjectType), allocator);
         }

src/crypto/cn/asm/CryptonightR_template.S

@@ -1593,3 +1593,7 @@ FN_PREFIX(CryptonightR_instruction_mov254):
 FN_PREFIX(CryptonightR_instruction_mov255):
 
 FN_PREFIX(CryptonightR_instruction_mov256):
+
+#if defined(__linux__) && defined(__ELF__)
+.section .note.GNU-stack,"",%progbits
+#endif

src/crypto/cn/asm/cn_main_loop.S

@@ -71,3 +71,7 @@ FN_PREFIX(cnv2_rwz_double_mainloop_asm):
 	add rsp, 48
 	ret 0
 	mov eax, 3735929054
+
+#if defined(__linux__) && defined(__ELF__)
+.section .note.GNU-stack,"",%progbits
+#endif

src/crypto/randomx/argon2_core.c

@@ -263,19 +263,6 @@ int rxa2_validate_inputs(const argon2_context *context) {
 		return ARGON2_INCORRECT_PARAMETER;
 	}
 
-	if (NULL == context->out) {
-		return ARGON2_OUTPUT_PTR_NULL;
-	}
-
-	/* Validate output length */
-	if (ARGON2_MIN_OUTLEN > context->outlen) {
-		return ARGON2_OUTPUT_TOO_SHORT;
-	}
-
-	if (ARGON2_MAX_OUTLEN < context->outlen) {
-		return ARGON2_OUTPUT_TOO_LONG;
-	}
-
 	/* Validate password (required param) */
 	if (NULL == context->pwd) {
 		if (0 != context->pwdlen) {

src/crypto/randomx/blake2_generator.cpp

@@ -46,7 +46,7 @@ namespace randomx {
 		return data[dataIndex++];
 	}
 
-	uint32_t Blake2Generator::getInt32() {
+	uint32_t Blake2Generator::getUInt32() {
 		checkData(4);
 		auto ret = load32(&data[dataIndex]);
 		dataIndex += 4;

src/crypto/randomx/blake2_generator.hpp

@@ -36,7 +36,7 @@ namespace randomx {
 	public:
 		Blake2Generator(const void* seed, size_t seedSize, int nonce = 0);
 		uint8_t getByte();
-		uint32_t getInt32();
+		uint32_t getUInt32();
 	private:
 		void checkData(const size_t);
 

src/crypto/randomx/bytecode_machine.cpp

@@ -244,7 +244,7 @@ namespace randomx {
 
 		if (opcode < RandomX_CurrentConfig.CEIL_IMUL_RCP) {
 			uint64_t divisor = instr.getImm32();
-			if (!isPowerOf2(divisor)) {
+			if (!isZeroOrPowerOf2(divisor)) {
 				auto dst = instr.dst % RegistersCount;
 				ibc.type = InstructionType::IMUL_R;
 				ibc.idst = &nreg->r[dst];

src/crypto/randomx/common.hpp

@@ -137,7 +137,7 @@ namespace randomx {
 	constexpr int RegisterNeedsDisplacement = 5; //x86 r13 register
 	constexpr int RegisterNeedsSib = 4; //x86 r12 register
 
-	inline bool isPowerOf2(uint64_t x) {
+	inline bool isZeroOrPowerOf2(uint64_t x) {
 		return (x & (x - 1)) == 0;
 	}
 

src/crypto/randomx/instruction.hpp

@@ -77,13 +77,13 @@ namespace randomx {
 		void setImm32(uint32_t val) {
 			return store32(&imm32, val);
 		}
-		int getModMem() const {
-			return mod % 4; //bits 0-1
+		uint32_t getModMem() const {
+			return mod & 3; //bits 0-1
 		}
-		int getModShift() const {
-			return (mod >> 2) % 4; //bits 2-3
+		uint32_t getModShift() const {
+			return (mod >> 2) & 3; //bits 2-3
 		}
-		int getModCond() const {
+		uint32_t getModCond() const {
 			return mod >> 4; //bits 4-7
 		}
 		void setMod(uint8_t val) {

src/crypto/randomx/jit_compiler_x86.cpp

@@ -181,7 +181,7 @@ namespace randomx {
 	static const uint8_t REX_TEST[] = { 0x49, 0xF7 };
 	static const uint8_t JZ[] = { 0x0f, 0x84 };
 	static const uint8_t RET = 0xc3;
-	static const uint8_t LEA_32[] = { 0x67, 0x41, 0x8d };
+	static const uint8_t LEA_32[] = { 0x41, 0x8d };
 	static const uint8_t MOVNTI[] = { 0x4c, 0x0f, 0xc3 };
 	static const uint8_t ADD_EBX_I[] = { 0x81, 0xc3 };
 
@@ -197,7 +197,7 @@ namespace randomx {
 //	static const uint8_t* NOPX[] = { NOP1, NOP2, NOP3, NOP4, NOP5, NOP6, NOP7, NOP8 };
 
 	size_t JitCompilerX86::getCodeSize() {
-		return codePos - prologueSize;
+		return codePos < prologueSize ? 0 : codePos - prologueSize;
 	}
 
 	JitCompilerX86::JitCompilerX86() {
@@ -219,12 +219,12 @@ namespace randomx {
 
 	void JitCompilerX86::generateProgramLight(Program& prog, ProgramConfiguration& pcfg, uint32_t datasetOffset) {
 		generateProgramPrologue(prog, pcfg);
-		emit(RandomX_CurrentConfig.codeReadDatasetLightSshInitTweaked, readDatasetLightInitSize);
-		emit(ADD_EBX_I);
-		emit32(datasetOffset / CacheLineSize);
-		emitByte(CALL);
-		emit32(superScalarHashOffset - (codePos + 4));
-		emit(codeReadDatasetLightSshFin, readDatasetLightFinSize);
+		emit(RandomX_CurrentConfig.codeReadDatasetLightSshInitTweaked, readDatasetLightInitSize, code, codePos);
+		emit(ADD_EBX_I, code, codePos);
+		emit32(datasetOffset / CacheLineSize, code, codePos);
+		emitByte(CALL, code, codePos);
+		emit32(superScalarHashOffset - (codePos + 4), code, codePos);
+		emit(codeReadDatasetLightSshFin, readDatasetLightFinSize, code, codePos);
 		generateProgramEpilogue(prog);
 	}
 
@@ -238,23 +238,23 @@ namespace randomx {
 				Instruction& instr = prog(i);
 				generateSuperscalarCode(instr, reciprocalCache);
 			}
-			emit(codeShhLoad, codeSshLoadSize);
+			emit(codeShhLoad, codeSshLoadSize, code, codePos);
 			if (j < RandomX_CurrentConfig.CacheAccesses - 1) {
-				emit(REX_MOV_RR64);
-				emitByte(0xd8 + prog.getAddressRegister());
-				emit(RandomX_CurrentConfig.codeShhPrefetchTweaked, codeSshPrefetchSize);
+				emit(REX_MOV_RR64, code, codePos);
+				emitByte(0xd8 + prog.getAddressRegister(), code, codePos);
+				emit(RandomX_CurrentConfig.codeShhPrefetchTweaked, codeSshPrefetchSize, code, codePos);
 #ifdef RANDOMX_ALIGN
 				int align = (codePos % 16);
 				while (align != 0) {
 					int nopSize = 16 - align;
 					if (nopSize > 8) nopSize = 8;
-					emit(NOPX[nopSize - 1], nopSize);
+					emit(NOPX[nopSize - 1], nopSize, code, codePos);
 					align = (codePos % 16);
 				}
 #endif
 			}
 		}
-		emitByte(RET);
+		emitByte(RET, code, codePos);
 	}
 
 	template
@@ -265,508 +265,664 @@ namespace randomx {
 	}
 
 	void JitCompilerX86::generateProgramPrologue(Program& prog, ProgramConfiguration& pcfg) {
-		instructionOffsets.clear();
-		for (unsigned i = 0; i < 8; ++i) {
-			registerUsage[i] = -1;
-		}
+		memset(registerUsage, -1, sizeof(registerUsage));
 		codePos = prologueSize;
 		memcpy(code + codePos - 48, &pcfg.eMask, sizeof(pcfg.eMask));
-		emit(REX_XOR_RAX_R64);
-		emitByte(0xc0 + pcfg.readReg0);
-		emit(REX_XOR_RAX_R64);
-		emitByte(0xc0 + pcfg.readReg1);
+		emit(REX_XOR_RAX_R64, code, codePos);
+		emitByte(0xc0 + pcfg.readReg0, code, codePos);
+		emit(REX_XOR_RAX_R64, code, codePos);
+		emitByte(0xc0 + pcfg.readReg1, code, codePos);
 		memcpy(code + codePos, RandomX_CurrentConfig.codeLoopLoadTweaked, loopLoadSize);
 		codePos += loopLoadSize;
 		for (unsigned i = 0; i < prog.getSize(); ++i) {
 			Instruction& instr = prog(i);
 			instr.src %= RegistersCount;
 			instr.dst %= RegistersCount;
-			generateCode(instr, i);
+			instructionOffsets[i] = codePos;
+			(this->*(engine[instr.opcode]))(instr, i);
 		}
-		emit(REX_MOV_RR);
-		emitByte(0xc0 + pcfg.readReg2);
-		emit(REX_XOR_EAX);
-		emitByte(0xc0 + pcfg.readReg3);
+		emit(REX_MOV_RR, code, codePos);
+		emitByte(0xc0 + pcfg.readReg2, code, codePos);
+		emit(REX_XOR_EAX, code, codePos);
+		emitByte(0xc0 + pcfg.readReg3, code, codePos);
 	}
 
 	void JitCompilerX86::generateProgramEpilogue(Program& prog) {
 		memcpy(code + codePos, codeLoopStore, loopStoreSize);
 		codePos += loopStoreSize;
-		emit(SUB_EBX);
-		emit(JNZ);
-		emit32(prologueSize - codePos - 4);
-		emitByte(JMP);
-		emit32(epilogueOffset - codePos - 4);
-	}
-
-	void JitCompilerX86::generateCode(Instruction& instr, int i) {
-		instructionOffsets.push_back(codePos);
-		auto generator = engine[instr.opcode];
-		(this->*generator)(instr, i);
+		emit(SUB_EBX, code, codePos);
+		emit(JNZ, code, codePos);
+		emit32(prologueSize - codePos - 4, code, codePos);
+		emitByte(JMP, code, codePos);
+		emit32(epilogueOffset - codePos - 4, code, codePos);
 	}
 
 	void JitCompilerX86::generateSuperscalarCode(Instruction& instr, std::vector<uint64_t> &reciprocalCache) {
 		switch ((SuperscalarInstructionType)instr.opcode)
 		{
 		case randomx::SuperscalarInstructionType::ISUB_R:
-			emit(REX_SUB_RR);
-			emitByte(0xc0 + 8 * instr.dst + instr.src);
+			emit(REX_SUB_RR, code, codePos);
+			emitByte(0xc0 + 8 * instr.dst + instr.src, code, codePos);
 			break;
 		case randomx::SuperscalarInstructionType::IXOR_R:
-			emit(REX_XOR_RR);
-			emitByte(0xc0 + 8 * instr.dst + instr.src);
+			emit(REX_XOR_RR, code, codePos);
+			emitByte(0xc0 + 8 * instr.dst + instr.src, code, codePos);
 			break;
 		case randomx::SuperscalarInstructionType::IADD_RS:
-			emit(REX_LEA);
-			emitByte(0x04 + 8 * instr.dst);
-			genSIB(instr.getModShift(), instr.src, instr.dst);
+			emit(REX_LEA, code, codePos);
+			emitByte(0x04 + 8 * instr.dst, code, codePos);
+			genSIB(instr.getModShift(), instr.src, instr.dst, code, codePos);
 			break;
 		case randomx::SuperscalarInstructionType::IMUL_R:
-			emit(REX_IMUL_RR);
-			emitByte(0xc0 + 8 * instr.dst + instr.src);
+			emit(REX_IMUL_RR, code, codePos);
+			emitByte(0xc0 + 8 * instr.dst + instr.src, code, codePos);
 			break;
 		case randomx::SuperscalarInstructionType::IROR_C:
-			emit(REX_ROT_I8);
-			emitByte(0xc8 + instr.dst);
-			emitByte(instr.getImm32() & 63);
+			emit(REX_ROT_I8, code, codePos);
+			emitByte(0xc8 + instr.dst, code, codePos);
+			emitByte(instr.getImm32() & 63, code, codePos);
 			break;
 		case randomx::SuperscalarInstructionType::IADD_C7:
-			emit(REX_81);
-			emitByte(0xc0 + instr.dst);
-			emit32(instr.getImm32());
+			emit(REX_81, code, codePos);
+			emitByte(0xc0 + instr.dst, code, codePos);
+			emit32(instr.getImm32(), code, codePos);
 			break;
 		case randomx::SuperscalarInstructionType::IXOR_C7:
-			emit(REX_XOR_RI);
-			emitByte(0xf0 + instr.dst);
-			emit32(instr.getImm32());
+			emit(REX_XOR_RI, code, codePos);
+			emitByte(0xf0 + instr.dst, code, codePos);
+			emit32(instr.getImm32(), code, codePos);
 			break;
 		case randomx::SuperscalarInstructionType::IADD_C8:
-			emit(REX_81);
-			emitByte(0xc0 + instr.dst);
-			emit32(instr.getImm32());
+			emit(REX_81, code, codePos);
+			emitByte(0xc0 + instr.dst, code, codePos);
+			emit32(instr.getImm32(), code, codePos);
 #ifdef RANDOMX_ALIGN
-			emit(NOP1);
+			emit(NOP1, code, codePos);
 #endif
 			break;
 		case randomx::SuperscalarInstructionType::IXOR_C8:
-			emit(REX_XOR_RI);
-			emitByte(0xf0 + instr.dst);
-			emit32(instr.getImm32());
+			emit(REX_XOR_RI, code, codePos);
+			emitByte(0xf0 + instr.dst, code, codePos);
+			emit32(instr.getImm32(), code, codePos);
 #ifdef RANDOMX_ALIGN
-			emit(NOP1);
+			emit(NOP1, code, codePos);
 #endif
 			break;
 		case randomx::SuperscalarInstructionType::IADD_C9:
-			emit(REX_81);
-			emitByte(0xc0 + instr.dst);
-			emit32(instr.getImm32());
+			emit(REX_81, code, codePos);
+			emitByte(0xc0 + instr.dst, code, codePos);
+			emit32(instr.getImm32(), code, codePos);
 #ifdef RANDOMX_ALIGN
-			emit(NOP2);
+			emit(NOP2, code, codePos);
 #endif
 			break;
 		case randomx::SuperscalarInstructionType::IXOR_C9:
-			emit(REX_XOR_RI);
-			emitByte(0xf0 + instr.dst);
-			emit32(instr.getImm32());
+			emit(REX_XOR_RI, code, codePos);
+			emitByte(0xf0 + instr.dst, code, codePos);
+			emit32(instr.getImm32(), code, codePos);
 #ifdef RANDOMX_ALIGN
-			emit(NOP2);
+			emit(NOP2, code, codePos);
 #endif
 			break;
 		case randomx::SuperscalarInstructionType::IMULH_R:
-			emit(REX_MOV_RR64);
-			emitByte(0xc0 + instr.dst);
-			emit(REX_MUL_R);
-			emitByte(0xe0 + instr.src);
-			emit(REX_MOV_R64R);
-			emitByte(0xc2 + 8 * instr.dst);
+			emit(REX_MOV_RR64, code, codePos);
+			emitByte(0xc0 + instr.dst, code, codePos);
+			emit(REX_MUL_R, code, codePos);
+			emitByte(0xe0 + instr.src, code, codePos);
+			emit(REX_MOV_R64R, code, codePos);
+			emitByte(0xc2 + 8 * instr.dst, code, codePos);
 			break;
 		case randomx::SuperscalarInstructionType::ISMULH_R:
-			emit(REX_MOV_RR64);
-			emitByte(0xc0 + instr.dst);
-			emit(REX_MUL_R);
-			emitByte(0xe8 + instr.src);
-			emit(REX_MOV_R64R);
-			emitByte(0xc2 + 8 * instr.dst);
+			emit(REX_MOV_RR64, code, codePos);
+			emitByte(0xc0 + instr.dst, code, codePos);
+			emit(REX_MUL_R, code, codePos);
+			emitByte(0xe8 + instr.src, code, codePos);
+			emit(REX_MOV_R64R, code, codePos);
+			emitByte(0xc2 + 8 * instr.dst, code, codePos);
 			break;
 		case randomx::SuperscalarInstructionType::IMUL_RCP:
-			emit(MOV_RAX_I);
-			emit64(reciprocalCache[instr.getImm32()]);
-			emit(REX_IMUL_RM);
-			emitByte(0xc0 + 8 * instr.dst);
+			emit(MOV_RAX_I, code, codePos);
+			emit64(reciprocalCache[instr.getImm32()], code, codePos);
+			emit(REX_IMUL_RM, code, codePos);
+			emitByte(0xc0 + 8 * instr.dst, code, codePos);
 			break;
 		default:
 			UNREACHABLE;
 		}
 	}
 
-	void JitCompilerX86::genAddressReg(Instruction& instr, bool rax = true) {
-		emit(LEA_32);
-		emitByte(0x80 + instr.src + (rax ? 0 : 8));
+	void JitCompilerX86::genAddressReg(Instruction& instr, uint8_t* code, int& codePos, bool rax) {
+		emit(LEA_32, code, codePos);
+		emitByte(0x80 + instr.src + (rax ? 0 : 8), code, codePos);
 		if (instr.src == RegisterNeedsSib) {
-			emitByte(0x24);
+			emitByte(0x24, code, codePos);
 		}
-		emit32(instr.getImm32());
+		emit32(instr.getImm32(), code, codePos);
 		if (rax)
-			emitByte(AND_EAX_I);
+			emitByte(AND_EAX_I, code, codePos);
 		else
-			emit(AND_ECX_I);
-		emit32(instr.getModMem() ? ScratchpadL1Mask : ScratchpadL2Mask);
+			emit(AND_ECX_I, code, codePos);
+		emit32(instr.getModMem() ? ScratchpadL1Mask : ScratchpadL2Mask, code, codePos);
 	}
 
-	void JitCompilerX86::genAddressRegDst(Instruction& instr) {
-		emit(LEA_32);
-		emitByte(0x80 + instr.dst);
+	void JitCompilerX86::genAddressRegDst(Instruction& instr, uint8_t* code, int& codePos) {
+		emit(LEA_32, code, codePos);
+		emitByte(0x80 + instr.dst, code, codePos);
 		if (instr.dst == RegisterNeedsSib) {
-			emitByte(0x24);
+			emitByte(0x24, code, codePos);
 		}
-		emit32(instr.getImm32());
-		emitByte(AND_EAX_I);
+		emit32(instr.getImm32(), code, codePos);
+		emitByte(AND_EAX_I, code, codePos);
 		if (instr.getModCond() < StoreL3Condition) {
-			emit32(instr.getModMem() ? ScratchpadL1Mask : ScratchpadL2Mask);
+			emit32(instr.getModMem() ? ScratchpadL1Mask : ScratchpadL2Mask, code, codePos);
 		}
 		else {
-			emit32(ScratchpadL3Mask);
+			emit32(ScratchpadL3Mask, code, codePos);
 		}
 	}
 
-	void JitCompilerX86::genAddressImm(Instruction& instr) {
-		emit32(instr.getImm32() & ScratchpadL3Mask);
+	void JitCompilerX86::genAddressImm(Instruction& instr, uint8_t* code, int& codePos) {
+		emit32(instr.getImm32() & ScratchpadL3Mask, code, codePos);
 	}
 
+	static const uint32_t template_IADD_RS[8] = {
+		0x048d4f,
+		0x0c8d4f,
+		0x148d4f,
+		0x1c8d4f,
+		0x248d4f,
+		0xac8d4f,
+		0x348d4f,
+		0x3c8d4f,
+	};
+
 	void JitCompilerX86::h_IADD_RS(Instruction& instr, int i) {
+		int pos = codePos;
+		uint8_t* const p = code + pos;
+
 		registerUsage[instr.dst] = i;
-		emit(REX_LEA);
-		if (instr.dst == RegisterNeedsDisplacement)
-			emitByte(0xac);
-		else
-			emitByte(0x04 + 8 * instr.dst);
-		genSIB(instr.getModShift(), instr.src, instr.dst);
-		if (instr.dst == RegisterNeedsDisplacement)
-			emit32(instr.getImm32());
+
+		const uint32_t sib = (instr.getModShift() << 6) | (instr.src << 3) | instr.dst;
+		*(uint32_t*)(p) = template_IADD_RS[instr.dst] | (sib << 24);
+		*(uint32_t*)(p + 4) = instr.getImm32();
+
+		codePos = pos + ((instr.dst == RegisterNeedsDisplacement) ? 8 : 4);
 	}
 
+	static const uint32_t template_IADD_M[8] = {
+		0x0604034c,
+		0x060c034c,
+		0x0614034c,
+		0x061c034c,
+		0x0624034c,
+		0x062c034c,
+		0x0634034c,
+		0x063c034c,
+	};
+
 	void JitCompilerX86::h_IADD_M(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
 		if (instr.src != instr.dst) {
-			genAddressReg(instr);
-			emit(REX_ADD_RM);
-			emitByte(0x04 + 8 * instr.dst);
-			emitByte(0x06);
+			genAddressReg(instr, p, pos);
+			emit32(template_IADD_M[instr.dst], p, pos);
 		}
 		else {
-			emit(REX_ADD_RM);
-			emitByte(0x86 + 8 * instr.dst);
-			genAddressImm(instr);
+			emit(REX_ADD_RM, p, pos);
+			emitByte(0x86 + 8 * instr.dst, p, pos);
+			genAddressImm(instr, p, pos);
 		}
+
+		codePos = pos;
 	}
 
-	void JitCompilerX86::genSIB(int scale, int index, int base) {
-		emitByte((scale << 6) | (index << 3) | base);
+	void JitCompilerX86::genSIB(int scale, int index, int base, uint8_t* code, int& codePos) {
+		emitByte((scale << 6) | (index << 3) | base, code, codePos);
 	}
 
 	void JitCompilerX86::h_ISUB_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
 		if (instr.src != instr.dst) {
-			emit(REX_SUB_RR);
-			emitByte(0xc0 + 8 * instr.dst + instr.src);
+			emit(REX_SUB_RR, p, pos);
+			emitByte(0xc0 + 8 * instr.dst + instr.src, p, pos);
 		}
 		else {
-			emit(REX_81);
-			emitByte(0xe8 + instr.dst);
-			emit32(instr.getImm32());
+			emit(REX_81, p, pos);
+			emitByte(0xe8 + instr.dst, p, pos);
+			emit32(instr.getImm32(), p, pos);
 		}
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_ISUB_M(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
 		if (instr.src != instr.dst) {
-			genAddressReg(instr);
-			emit(REX_SUB_RM);
-			emitByte(0x04 + 8 * instr.dst);
-			emitByte(0x06);
+			genAddressReg(instr, p, pos);
+			emit(REX_SUB_RM, p, pos);
+			emitByte(0x04 + 8 * instr.dst, p, pos);
+			emitByte(0x06, p, pos);
 		}
 		else {
-			emit(REX_SUB_RM);
-			emitByte(0x86 + 8 * instr.dst);
-			genAddressImm(instr);
+			emit(REX_SUB_RM, p, pos);
+			emitByte(0x86 + 8 * instr.dst, p, pos);
+			genAddressImm(instr, p, pos);
 		}
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_IMUL_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
 		if (instr.src != instr.dst) {
-			emit(REX_IMUL_RR);
-			emitByte(0xc0 + 8 * instr.dst + instr.src);
+			emit(REX_IMUL_RR, p, pos);
+			emitByte(0xc0 + 8 * instr.dst + instr.src, p, pos);
 		}
 		else {
-			emit(REX_IMUL_RRI);
-			emitByte(0xc0 + 9 * instr.dst);
-			emit32(instr.getImm32());
+			emit(REX_IMUL_RRI, p, pos);
+			emitByte(0xc0 + 9 * instr.dst, p, pos);
+			emit32(instr.getImm32(), p, pos);
 		}
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_IMUL_M(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
 		if (instr.src != instr.dst) {
-			genAddressReg(instr);
-			emit(REX_IMUL_RM);
-			emitByte(0x04 + 8 * instr.dst);
-			emitByte(0x06);
+			genAddressReg(instr, p, pos);
+			emit(REX_IMUL_RM, p, pos);
+			emitByte(0x04 + 8 * instr.dst, p, pos);
+			emitByte(0x06, p, pos);
 		}
 		else {
-			emit(REX_IMUL_RM);
-			emitByte(0x86 + 8 * instr.dst);
-			genAddressImm(instr);
+			emit(REX_IMUL_RM, p, pos);
+			emitByte(0x86 + 8 * instr.dst, p, pos);
+			genAddressImm(instr, p, pos);
 		}
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_IMULH_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+
 		registerUsage[instr.dst] = i;
-		emit(REX_MOV_RR64);
-		emitByte(0xc0 + instr.dst);
-		emit(REX_MUL_R);
-		emitByte(0xe0 + instr.src);
-		emit(REX_MOV_R64R);
-		emitByte(0xc2 + 8 * instr.dst);
+		emit(REX_MOV_RR64, p, pos);
+		emitByte(0xc0 + instr.dst, p, pos);
+		emit(REX_MUL_R, p, pos);
+		emitByte(0xe0 + instr.src, p, pos);
+		emit(REX_MOV_R64R, p, pos);
+		emitByte(0xc2 + 8 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_IMULH_M(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, false);
-			emit(REX_MOV_RR64);
-			emitByte(0xc0 + instr.dst);
-			emit(REX_MUL_MEM);
+			genAddressReg(instr, p, pos, false);
+			emit(REX_MOV_RR64, p, pos);
+			emitByte(0xc0 + instr.dst, p, pos);
+			emit(REX_MUL_MEM, p, pos);
 		}
 		else {
-			emit(REX_MOV_RR64);
-			emitByte(0xc0 + instr.dst);
-			emit(REX_MUL_M);
-			emitByte(0xa6);
-			genAddressImm(instr);
+			emit(REX_MOV_RR64, p, pos);
+			emitByte(0xc0 + instr.dst, p, pos);
+			emit(REX_MUL_M, p, pos);
+			emitByte(0xa6, p, pos);
+			genAddressImm(instr, p, pos);
 		}
-		emit(REX_MOV_R64R);
-		emitByte(0xc2 + 8 * instr.dst);
+		emit(REX_MOV_R64R, p, pos);
+		emitByte(0xc2 + 8 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_ISMULH_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
-		emit(REX_MOV_RR64);
-		emitByte(0xc0 + instr.dst);
-		emit(REX_MUL_R);
-		emitByte(0xe8 + instr.src);
-		emit(REX_MOV_R64R);
-		emitByte(0xc2 + 8 * instr.dst);
+		emit(REX_MOV_RR64, p, pos);
+		emitByte(0xc0 + instr.dst, p, pos);
+		emit(REX_MUL_R, p, pos);
+		emitByte(0xe8 + instr.src, p, pos);
+		emit(REX_MOV_R64R, p, pos);
+		emitByte(0xc2 + 8 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_ISMULH_M(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
 		if (instr.src != instr.dst) {
-			genAddressReg(instr, false);
-			emit(REX_MOV_RR64);
-			emitByte(0xc0 + instr.dst);
-			emit(REX_IMUL_MEM);
+			genAddressReg(instr, p, pos, false);
+			emit(REX_MOV_RR64, p, pos);
+			emitByte(0xc0 + instr.dst, p, pos);
+			emit(REX_IMUL_MEM, p, pos);
 		}
 		else {
-			emit(REX_MOV_RR64);
-			emitByte(0xc0 + instr.dst);
-			emit(REX_MUL_M);
-			emitByte(0xae);
-			genAddressImm(instr);
+			emit(REX_MOV_RR64, p, pos);
+			emitByte(0xc0 + instr.dst, p, pos);
+			emit(REX_MUL_M, p, pos);
+			emitByte(0xae, p, pos);
+			genAddressImm(instr, p, pos);
 		}
-		emit(REX_MOV_R64R);
-		emitByte(0xc2 + 8 * instr.dst);
+		emit(REX_MOV_R64R, p, pos);
+		emitByte(0xc2 + 8 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_IMUL_RCP(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		uint64_t divisor = instr.getImm32();
-		if (!isPowerOf2(divisor)) {
+		if (!isZeroOrPowerOf2(divisor)) {
 			registerUsage[instr.dst] = i;
-			emit(MOV_RAX_I);
-			emit64(randomx_reciprocal_fast(divisor));
-			emit(REX_IMUL_RM);
-			emitByte(0xc0 + 8 * instr.dst);
+			emit(MOV_RAX_I, p, pos);
+			emit64(randomx_reciprocal_fast(divisor), p, pos);
+			emit(REX_IMUL_RM, p, pos);
+			emitByte(0xc0 + 8 * instr.dst, p, pos);
 		}
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_INEG_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
-		emit(REX_NEG);
-		emitByte(0xd8 + instr.dst);
+		emit(REX_NEG, p, pos);
+		emitByte(0xd8 + instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_IXOR_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
 		if (instr.src != instr.dst) {
-			emit(REX_XOR_RR);
-			emitByte(0xc0 + 8 * instr.dst + instr.src);
+			emit(REX_XOR_RR, p, pos);
+			emitByte(0xc0 + 8 * instr.dst + instr.src, p, pos);
 		}
 		else {
-			emit(REX_XOR_RI);
-			emitByte(0xf0 + instr.dst);
-			emit32(instr.getImm32());
+			emit(REX_XOR_RI, p, pos);
+			emitByte(0xf0 + instr.dst, p, pos);
+			emit32(instr.getImm32(), p, pos);
 		}
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_IXOR_M(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
 		if (instr.src != instr.dst) {
-			genAddressReg(instr);
-			emit(REX_XOR_RM);
-			emitByte(0x04 + 8 * instr.dst);
-			emitByte(0x06);
+			genAddressReg(instr, p, pos);
+			emit(REX_XOR_RM, p, pos);
+			emitByte(0x04 + 8 * instr.dst, p, pos);
+			emitByte(0x06, p, pos);
 		}
 		else {
-			emit(REX_XOR_RM);
-			emitByte(0x86 + 8 * instr.dst);
-			genAddressImm(instr);
+			emit(REX_XOR_RM, p, pos);
+			emitByte(0x86 + 8 * instr.dst, p, pos);
+			genAddressImm(instr, p, pos);
 		}
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_IROR_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		registerUsage[instr.dst] = i;
 		if (instr.src != instr.dst) {
-			emit(REX_MOV_RR);
-			emitByte(0xc8 + instr.src);
-			emit(REX_ROT_CL);
-			emitByte(0xc8 + instr.dst);
+			emit(REX_MOV_RR, p, pos);
+			emitByte(0xc8 + instr.src, p, pos);
+			emit(REX_ROT_CL, p, pos);
+			emitByte(0xc8 + instr.dst, p, pos);
 		}
 		else {
-			emit(REX_ROT_I8);
-			emitByte(0xc8 + instr.dst);
-			emitByte(instr.getImm32() & 63);
+			emit(REX_ROT_I8, p, pos);
+			emitByte(0xc8 + instr.dst, p, pos);
+			emitByte(instr.getImm32() & 63, p, pos);
 		}
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_IROL_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+
 		registerUsage[instr.dst] = i;
 		if (instr.src != instr.dst) {
-			emit(REX_MOV_RR);
-			emitByte(0xc8 + instr.src);
-			emit(REX_ROT_CL);
-			emitByte(0xc0 + instr.dst);
+			emit(REX_MOV_RR, p, pos);
+			emitByte(0xc8 + instr.src, p, pos);
+			emit(REX_ROT_CL, p, pos);
+			emitByte(0xc0 + instr.dst, p, pos);
 		}
 		else {
-			emit(REX_ROT_I8);
-			emitByte(0xc0 + instr.dst);
-			emitByte(instr.getImm32() & 63);
+			emit(REX_ROT_I8, p, pos);
+			emitByte(0xc0 + instr.dst, p, pos);
+			emitByte(instr.getImm32() & 63, p, pos);
 		}
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_ISWAP_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		if (instr.src != instr.dst) {
 			registerUsage[instr.dst] = i;
 			registerUsage[instr.src] = i;
-			emit(REX_XCHG);
-			emitByte(0xc0 + instr.src + 8 * instr.dst);
+			emit(REX_XCHG, p, pos);
+			emitByte(0xc0 + instr.src + 8 * instr.dst, p, pos);
 		}
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_FSWAP_R(Instruction& instr, int i) {
-		emit(SHUFPD);
-		emitByte(0xc0 + 9 * instr.dst);
-		emitByte(1);
+		uint8_t* const p = code;
+		int pos = codePos;
+		
+		emit(SHUFPD, p, pos);
+		emitByte(0xc0 + 9 * instr.dst, p, pos);
+		emitByte(1, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_FADD_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+
 		instr.dst %= RegisterCountFlt;
 		instr.src %= RegisterCountFlt;
-		emit(REX_ADDPD);
-		emitByte(0xc0 + instr.src + 8 * instr.dst);
+		emit(REX_ADDPD, p, pos);
+		emitByte(0xc0 + instr.src + 8 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_FADD_M(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		instr.dst %= RegisterCountFlt;
-		genAddressReg(instr);
-		emit(REX_CVTDQ2PD_XMM12);
-		emit(REX_ADDPD);
-		emitByte(0xc4 + 8 * instr.dst);
+		genAddressReg(instr, p, pos);
+		emit(REX_CVTDQ2PD_XMM12, p, pos);
+		emit(REX_ADDPD, p, pos);
+		emitByte(0xc4 + 8 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_FSUB_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		instr.dst %= RegisterCountFlt;
 		instr.src %= RegisterCountFlt;
-		emit(REX_SUBPD);
-		emitByte(0xc0 + instr.src + 8 * instr.dst);
+		emit(REX_SUBPD, p, pos);
+		emitByte(0xc0 + instr.src + 8 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_FSUB_M(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		instr.dst %= RegisterCountFlt;
-		genAddressReg(instr);
-		emit(REX_CVTDQ2PD_XMM12);
-		emit(REX_SUBPD);
-		emitByte(0xc4 + 8 * instr.dst);
+		genAddressReg(instr, p, pos);
+		emit(REX_CVTDQ2PD_XMM12, p, pos);
+		emit(REX_SUBPD, p, pos);
+		emitByte(0xc4 + 8 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_FSCAL_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		instr.dst %= RegisterCountFlt;
-		emit(REX_XORPS);
-		emitByte(0xc7 + 8 * instr.dst);
+		emit(REX_XORPS, p, pos);
+		emitByte(0xc7 + 8 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_FMUL_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		instr.dst %= RegisterCountFlt;
 		instr.src %= RegisterCountFlt;
-		emit(REX_MULPD);
-		emitByte(0xe0 + instr.src + 8 * instr.dst);
+		emit(REX_MULPD, p, pos);
+		emitByte(0xe0 + instr.src + 8 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_FDIV_M(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		instr.dst %= RegisterCountFlt;
-		genAddressReg(instr);
-		emit(REX_CVTDQ2PD_XMM12);
-		emit(REX_ANDPS_XMM12);
-		emit(REX_DIVPD);
-		emitByte(0xe4 + 8 * instr.dst);
+		genAddressReg(instr, p, pos);
+		emit(REX_CVTDQ2PD_XMM12, p, pos);
+		emit(REX_ANDPS_XMM12, p, pos);
+		emit(REX_DIVPD, p, pos);
+		emitByte(0xe4 + 8 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_FSQRT_R(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		instr.dst %= RegisterCountFlt;
-		emit(SQRTPD);
-		emitByte(0xe4 + 9 * instr.dst);
+		emit(SQRTPD, p, pos);
+		emitByte(0xe4 + 9 * instr.dst, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_CFROUND(Instruction& instr, int i) {
-		emit(REX_MOV_RR64);
-		emitByte(0xc0 + instr.src);
+		uint8_t* const p = code;
+		int pos = codePos;
+
+		emit(REX_MOV_RR64, p, pos);
+		emitByte(0xc0 + instr.src, p, pos);
 		int rotate = (13 - (instr.getImm32() & 63)) & 63;
 		if (rotate != 0) {
-			emit(ROL_RAX);
-			emitByte(rotate);
+			emit(ROL_RAX, p, pos);
+			emitByte(rotate, p, pos);
 		}
-		emit(AND_OR_MOV_LDMXCSR);
+		emit(AND_OR_MOV_LDMXCSR, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_CBRANCH(Instruction& instr, int i) {
+		uint8_t* const p = code;
+		int pos = codePos;
+		
 		int reg = instr.dst;
 		int target = registerUsage[reg] + 1;
-		emit(REX_ADD_I);
-		emitByte(0xc0 + reg);
+		emit(REX_ADD_I, p, pos);
+		emitByte(0xc0 + reg, p, pos);
 		int shift = instr.getModCond() + RandomX_CurrentConfig.JumpOffset;
 		uint32_t imm = instr.getImm32() | (1UL << shift);
 		if (RandomX_CurrentConfig.JumpOffset > 0 || shift > 0)
 			imm &= ~(1UL << (shift - 1));
-		emit32(imm);
-		emit(REX_TEST);
-		emitByte(0xc0 + reg);
-		emit32(RandomX_CurrentConfig.ConditionMask_Calculated << shift);
-		emit(JZ);
-		emit32(instructionOffsets[target] - (codePos + 4));
+		emit32(imm, p, pos);
+		emit(REX_TEST, p, pos);
+		emitByte(0xc0 + reg, p, pos);
+		emit32(RandomX_CurrentConfig.ConditionMask_Calculated << shift, p, pos);
+		emit(JZ, p, pos);
+		emit32(instructionOffsets[target] - (pos + 4), p, pos);
 		//mark all registers as used
-		for (unsigned j = 0; j < RegistersCount; ++j) {
-			registerUsage[j] = i;
+		uint64_t* r = (uint64_t*) registerUsage;
+		uint64_t k = i;
+		k |= k << 32;
+		for (unsigned j = 0; j < RegistersCount / 2; ++j) {
+			r[j] = k;
 		}
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_ISTORE(Instruction& instr, int i) {
-		genAddressRegDst(instr);
-		emit(REX_MOV_MR);
-		emitByte(0x04 + 8 * instr.src);
-		emitByte(0x06);
+		uint8_t* const p = code;
+		int pos = codePos;
+
+		genAddressRegDst(instr, p, pos);
+		emit(REX_MOV_MR, p, pos);
+		emitByte(0x04 + 8 * instr.src, p, pos);
+		emitByte(0x06, p, pos);
+
+		codePos = pos;
 	}
 
 	void JitCompilerX86::h_NOP(Instruction& instr, int i) {
-		emit(NOP1);
+		emit(NOP1, code, codePos);
 	}
 
 	InstructionGeneratorX86 JitCompilerX86::engine[256] = {};

src/crypto/randomx/jit_compiler_x86.hpp

@@ -66,42 +66,41 @@ namespace randomx {
 		size_t getCodeSize();
 
 		static InstructionGeneratorX86 engine[256];
-		std::vector<int32_t> instructionOffsets;
+		int32_t instructionOffsets[512];
 		int registerUsage[RegistersCount];
 		uint8_t* code;
 		int32_t codePos;
 
 		void generateProgramPrologue(Program&, ProgramConfiguration&);
 		void generateProgramEpilogue(Program&);
-		void genAddressReg(Instruction&, bool);
-		void genAddressRegDst(Instruction&);
-		void genAddressImm(Instruction&);
-		void genSIB(int scale, int index, int base);
+		static void genAddressReg(Instruction&, uint8_t* code, int& codePos, bool rax = true);
+		static void genAddressRegDst(Instruction&, uint8_t* code, int& codePos);
+		static void genAddressImm(Instruction&, uint8_t* code, int& codePos);
+		static void genSIB(int scale, int index, int base, uint8_t* code, int& codePos);
 
-		void generateCode(Instruction&, int);
 		void generateSuperscalarCode(Instruction &, std::vector<uint64_t> &);
 
-		void emitByte(uint8_t val) {
+		static void emitByte(uint8_t val, uint8_t* code, int& codePos) {
 			code[codePos] = val;
-			codePos++;
+			++codePos;
 		}
 
-		void emit32(uint32_t val) {
+		static void emit32(uint32_t val, uint8_t* code, int& codePos) {
 			memcpy(code + codePos, &val, sizeof val);
 			codePos += sizeof val;
 		}
 
-		void emit64(uint64_t val) {
+		static void emit64(uint64_t val, uint8_t* code, int& codePos) {
 			memcpy(code + codePos, &val, sizeof val);
 			codePos += sizeof val;
 		}
 
 		template<size_t N>
-		void emit(const uint8_t (&src)[N]) {
-			emit(src, N);
+		static void emit(const uint8_t (&src)[N], uint8_t* code, int& codePos) {
+			emit(src, N, code, codePos);
 		}
 
-		void emit(const uint8_t* src, size_t count) {
+		static void emit(const uint8_t* src, size_t count, uint8_t* code, int& codePos) {
 			memcpy(code + codePos, src, count);
 			codePos += count;
 		}

src/crypto/randomx/jit_compiler_x86_static.S

@@ -206,3 +206,7 @@ DECL(randomx_reciprocal_fast):
 	mov rcx, rdi
 #endif
 	#include "asm/randomx_reciprocal.inc"
+
+#if defined(__linux__) && defined(__ELF__)
+.section .note.GNU-stack,"",%progbits
+#endif

src/crypto/randomx/randomx.cpp

@@ -44,12 +44,14 @@ RandomX_ConfigurationWownero::RandomX_ConfigurationWownero()
 	ScratchpadL2_Size = 131072;
 	ScratchpadL3_Size = 1048576;
 
+	RANDOMX_FREQ_IADD_RS = 25;
 	RANDOMX_FREQ_IROR_R = 10;
 	RANDOMX_FREQ_IROL_R = 0;
 	RANDOMX_FREQ_FSWAP_R = 8;
 	RANDOMX_FREQ_FADD_R = 20;
 	RANDOMX_FREQ_FSUB_R = 20;
 	RANDOMX_FREQ_FMUL_R = 20;
+	RANDOMX_FREQ_CBRANCH = 16;
 
 	fillAes4Rx4_Key[0] = rx_set_int_vec_i128(0xcf359e95, 0x141f82b7, 0x7ffbe4a6, 0xf890465d);
 	fillAes4Rx4_Key[1] = rx_set_int_vec_i128(0x6741ffdc, 0xbd5c5ac3, 0xfee8278a, 0x6a55c450);
@@ -68,6 +70,9 @@ RandomX_ConfigurationLoki::RandomX_ConfigurationLoki()
 	ArgonSalt = "RandomXL\x12";
 	ProgramSize = 320;
 	ProgramCount = 7;
+
+	RANDOMX_FREQ_IADD_RS = 25;
+	RANDOMX_FREQ_CBRANCH = 16;
 }
 
 RandomX_ConfigurationBase::RandomX_ConfigurationBase()
@@ -87,7 +92,7 @@ RandomX_ConfigurationBase::RandomX_ConfigurationBase()
 	, ProgramCount(8)
 	, JumpBits(8)
 	, JumpOffset(8)
-	, RANDOMX_FREQ_IADD_RS(25)
+	, RANDOMX_FREQ_IADD_RS(16)
 	, RANDOMX_FREQ_IADD_M(7)
 	, RANDOMX_FREQ_ISUB_R(16)
 	, RANDOMX_FREQ_ISUB_M(7)
@@ -113,7 +118,7 @@ RandomX_ConfigurationBase::RandomX_ConfigurationBase()
 	, RANDOMX_FREQ_FMUL_R(32)
 	, RANDOMX_FREQ_FDIV_M(4)
 	, RANDOMX_FREQ_FSQRT_R(6)
-	, RANDOMX_FREQ_CBRANCH(16)
+	, RANDOMX_FREQ_CBRANCH(25)
 	, RANDOMX_FREQ_CFROUND(1)
 	, RANDOMX_FREQ_ISTORE(16)
 	, RANDOMX_FREQ_NOP(0)

src/crypto/randomx/superscalar.cpp

@@ -329,7 +329,7 @@ namespace randomx {
 			return false;
 
 		if (availableRegisters.size() > 1) {
-			index = gen.getInt32() % availableRegisters.size();
+			index = gen.getUInt32() % availableRegisters.size();
 		}
 		else {
 			index = 0;
@@ -442,7 +442,7 @@ namespace randomx {
 			case SuperscalarInstructionType::IADD_C8:
 			case SuperscalarInstructionType::IADD_C9: {
 				mod_ = 0;
-				imm32_ = gen.getInt32();
+				imm32_ = gen.getUInt32();
 				opGroup_ = SuperscalarInstructionType::IADD_C7;
 				opGroupPar_ = -1;
 			} break;
@@ -451,7 +451,7 @@ namespace randomx {
 			case SuperscalarInstructionType::IXOR_C8:
 			case SuperscalarInstructionType::IXOR_C9: {
 				mod_ = 0;
-				imm32_ = gen.getInt32();
+				imm32_ = gen.getUInt32();
 				opGroup_ = SuperscalarInstructionType::IXOR_C7;
 				opGroupPar_ = -1;
 			} break;
@@ -461,7 +461,7 @@ namespace randomx {
 				mod_ = 0;
 				imm32_ = 0;
 				opGroup_ = SuperscalarInstructionType::IMULH_R;
-				opGroupPar_ = gen.getInt32();
+				opGroupPar_ = gen.getUInt32();
 			} break;
 
 			case SuperscalarInstructionType::ISMULH_R: {
@@ -469,14 +469,14 @@ namespace randomx {
 				mod_ = 0;
 				imm32_ = 0;
 				opGroup_ = SuperscalarInstructionType::ISMULH_R;
-				opGroupPar_ = gen.getInt32();
+				opGroupPar_ = gen.getUInt32();
 			} break;
 
 			case SuperscalarInstructionType::IMUL_RCP: {
 				mod_ = 0;
 				do {
-					imm32_ = gen.getInt32();
-				} while ((imm32_ & (imm32_ - 1)) == 0);
+					imm32_ = gen.getUInt32();
+				} while (isZeroOrPowerOf2(imm32_));
 				opGroup_ = SuperscalarInstructionType::IMUL_RCP;
 				opGroupPar_ = -1;
 			} break;

src/version.h

@@ -28,7 +28,7 @@
 #define APP_ID        "xmrig"
 #define APP_NAME      "XMRig"
 #define APP_DESC      "XMRig CPU miner"
-#define APP_VERSION   "3.1.0"
+#define APP_VERSION   "3.1.1-dev"
 #define APP_DOMAIN    "xmrig.com"
 #define APP_SITE      "www.xmrig.com"
 #define APP_COPYRIGHT "Copyright (C) 2016-2019 xmrig.com"
@@ -36,7 +36,7 @@
 
 #define APP_VER_MAJOR  3
 #define APP_VER_MINOR  1
-#define APP_VER_PATCH  0
+#define APP_VER_PATCH  1
 
 #ifdef _MSC_VER
 #   if (_MSC_VER >= 1920)