RandomX: added huge-pages-jit config parameter

Set to false by default, gives 0.2% boost on Ryzen 7 3700X with 16 threads, but hashrate might be unstable on Ryzen between launches. Use with caution.

src/backend/cpu/CpuConfig.cpp

@@ -36,6 +36,7 @@ namespace xmrig {
 
 static const char *kEnabled             = "enabled";
 static const char *kHugePages           = "huge-pages";
+static const char *kHugePagesJit        = "huge-pages-jit";
 static const char *kHwAes               = "hw-aes";
 static const char *kMaxThreadsHint      = "max-threads-hint";
 static const char *kMemoryPool          = "memory-pool";
@@ -76,6 +77,7 @@ rapidjson::Value xmrig::CpuConfig::toJSON(rapidjson::Document &doc) const
 
     obj.AddMember(StringRef(kEnabled),      m_enabled, allocator);
     obj.AddMember(StringRef(kHugePages),    m_hugePages, allocator);
+    obj.AddMember(StringRef(kHugePagesJit), m_hugePagesJit, allocator);
     obj.AddMember(StringRef(kHwAes),        m_aes == AES_AUTO ? Value(kNullType) : Value(m_aes == AES_HW), allocator);
     obj.AddMember(StringRef(kPriority),     priority() != -1 ? Value(priority()) : Value(kNullType), allocator);
     obj.AddMember(StringRef(kMemoryPool),   m_memoryPool < 1 ? Value(m_memoryPool < 0) : Value(m_memoryPool), allocator);
@@ -134,6 +136,7 @@ void xmrig::CpuConfig::read(const rapidjson::Value &value)
     if (value.IsObject()) {
         m_enabled      = Json::getBool(value, kEnabled, m_enabled);
         m_hugePages    = Json::getBool(value, kHugePages, m_hugePages);
+        m_hugePagesJit = Json::getBool(value, kHugePagesJit, m_hugePagesJit);
         m_limit        = Json::getUint(value, kMaxThreadsHint, m_limit);
         m_yield        = Json::getBool(value, kYield, m_yield);
 

src/backend/cpu/CpuConfig.h

@@ -54,6 +54,7 @@ public:
 
     inline bool isEnabled() const                       { return m_enabled; }
     inline bool isHugePages() const                     { return m_hugePages; }
+    inline bool isHugePagesJit() const                  { return m_hugePagesJit; }
     inline bool isShouldSave() const                    { return m_shouldSave; }
     inline bool isYield() const                         { return m_yield; }
     inline const Assembly &assembly() const             { return m_assembly; }
@@ -76,6 +77,7 @@ private:
     bool m_astrobwtAVX2     = false;
     bool m_enabled          = true;
     bool m_hugePages        = true;
+    bool m_hugePagesJit     = false;
     bool m_shouldSave       = false;
     bool m_yield            = true;
     int m_astrobwtMaxSize   = 550;

src/config.json

@@ -27,6 +27,7 @@
     "cpu": {
         "enabled": true,
         "huge-pages": true,
+        "huge-pages-jit": false,
         "hw-aes": null,
         "priority": null,
         "memory-pool": false,

src/core/config/Config_default.h

@@ -61,6 +61,7 @@ R"===(
     "cpu": {
         "enabled": true,
         "huge-pages": true,
+        "huge-pages-jit": false,
         "hw-aes": null,
         "priority": null,
         "memory-pool": false,

src/crypto/cn/CnCtx.cpp

@@ -39,7 +39,7 @@ void xmrig::CnCtx::create(cryptonight_ctx **ctx, uint8_t *memory, size_t size, s
         cryptonight_ctx *c = static_cast<cryptonight_ctx *>(_mm_malloc(sizeof(cryptonight_ctx), 4096));
         c->memory          = memory + (i * size);
 
-        c->generated_code              = reinterpret_cast<cn_mainloop_fun_ms_abi>(VirtualMemory::allocateExecutableMemory(0x4000));
+        c->generated_code              = reinterpret_cast<cn_mainloop_fun_ms_abi>(VirtualMemory::allocateExecutableMemory(0x4000, false));
         c->generated_code_data.algo    = Algorithm::INVALID;
         c->generated_code_data.height  = std::numeric_limits<uint64_t>::max();
 

src/crypto/cn/CnHash.cpp

@@ -139,7 +139,7 @@ static void patchCode(T dst, U src, const uint32_t iterations, const uint32_t ma
 static void patchAsmVariants()
 {
     const int allocation_size = 81920;
-    auto base = static_cast<uint8_t *>(VirtualMemory::allocateExecutableMemory(allocation_size));
+    auto base = static_cast<uint8_t *>(VirtualMemory::allocateExecutableMemory(allocation_size, false));
 
     cn_half_mainloop_ivybridge_asm              = reinterpret_cast<cn_mainloop_fun>         (base + 0x0000);
     cn_half_mainloop_ryzen_asm                  = reinterpret_cast<cn_mainloop_fun>         (base + 0x1000);

src/crypto/common/MemoryPool.cpp

@@ -47,7 +47,11 @@ xmrig::MemoryPool::MemoryPool(size_t size, bool hugePages, uint32_t node)
         return;
     }
 
-    m_memory = new VirtualMemory(size * pageSize, hugePages, false, false, node);
+    constexpr size_t alignment = 1 << 24;
+
+    m_memory = new VirtualMemory(size * pageSize + alignment, hugePages, false, false, node);
+
+    m_alignOffset = (alignment - (((size_t)m_memory->scratchpad()) % alignment)) % alignment;
 }
 
 
@@ -71,7 +75,7 @@ uint8_t *xmrig::MemoryPool::get(size_t size, uint32_t)
         return nullptr;
     }
 
-    uint8_t *out = m_memory->scratchpad() + m_offset;
+    uint8_t *out = m_memory->scratchpad() + m_alignOffset + m_offset;
 
     m_offset += size;
     ++m_refs;

src/crypto/common/MemoryPool.h

@@ -54,6 +54,7 @@ protected:
 private:
     size_t m_refs           = 0;
     size_t m_offset         = 0;
+    size_t m_alignOffset    = 0;
     VirtualMemory *m_memory = nullptr;
 };
 

src/crypto/common/VirtualMemory.h

@@ -61,7 +61,7 @@ public:
     static bool isHugepagesAvailable();
     static bool isOneGbPagesAvailable();
     static uint32_t bindToNUMANode(int64_t affinity);
-    static void *allocateExecutableMemory(size_t size);
+    static void *allocateExecutableMemory(size_t size, bool hugePages);
     static void *allocateLargePagesMemory(size_t size);
     static void *allocateOneGbPagesMemory(size_t size);
     static void destroy();

src/crypto/common/VirtualMemory_unix.cpp

@@ -63,7 +63,7 @@ bool xmrig::VirtualMemory::isOneGbPagesAvailable()
 }
 
 
-void *xmrig::VirtualMemory::allocateExecutableMemory(size_t size)
+void *xmrig::VirtualMemory::allocateExecutableMemory(size_t size, bool hugePages)
 {
 #   if defined(__APPLE__)
     void *mem = mmap(0, size, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANON, -1, 0);
@@ -77,7 +77,12 @@ void *xmrig::VirtualMemory::allocateExecutableMemory(size_t size)
     constexpr int flag_2mb = 0;
 #   endif
 
-    void *mem = mmap(0, align(size), PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS | MAP_POPULATE | flag_2mb, -1, 0);
+    void *mem = nullptr;
+
+    if (hugePages) {
+        mem = mmap(0, align(size), PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS | MAP_POPULATE | flag_2mb, -1, 0);
+    }
+
     if (!mem) {
         void *mem = mmap(0, size, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
     }

src/crypto/common/VirtualMemory_win.cpp

@@ -162,9 +162,13 @@ bool xmrig::VirtualMemory::isOneGbPagesAvailable()
 }
 
 
-void *xmrig::VirtualMemory::allocateExecutableMemory(size_t size)
+void *xmrig::VirtualMemory::allocateExecutableMemory(size_t size, bool hugePages)
 {
-    void* result = VirtualAlloc(nullptr, align(size), MEM_COMMIT | MEM_RESERVE | MEM_LARGE_PAGES, PAGE_EXECUTE_READWRITE);
+    void* result = nullptr;
+
+    if (hugePages) {
+        result = VirtualAlloc(nullptr, align(size), MEM_COMMIT | MEM_RESERVE | MEM_LARGE_PAGES, PAGE_EXECUTE_READWRITE);
+    }
 
     if (!result) {
         result = VirtualAlloc(nullptr, size, MEM_COMMIT | MEM_RESERVE, PAGE_EXECUTE_READWRITE);

src/crypto/randomx/dataset.hpp

@@ -43,7 +43,7 @@ struct randomx_dataset {
 /* Global scope for C binding */
 struct randomx_cache {
 	uint8_t* memory = nullptr;
-	randomx::JitCompiler* jit;
+	randomx::JitCompiler* jit = nullptr;
 	randomx::CacheInitializeFunc* initialize;
 	randomx::DatasetInitFunc* datasetInit;
 	randomx::SuperscalarProgram programs[RANDOMX_CACHE_MAX_ACCESSES];

src/crypto/randomx/jit_compiler_a64.cpp

@@ -33,6 +33,8 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "crypto/randomx/reciprocal.h"
 #include "crypto/randomx/virtual_memory.hpp"
 
+void randomx_set_huge_pages_jit(bool) {}
+
 namespace ARMV8A {
 
 constexpr uint32_t B           = 0x14000000;
@@ -89,7 +91,7 @@ static size_t CalcDatasetItemSize()
 
 constexpr uint32_t IntRegMap[8] = { 4, 5, 6, 7, 12, 13, 14, 15 };
 
-JitCompilerA64::JitCompilerA64()
+JitCompilerA64::JitCompilerA64(bool)
 	: code((uint8_t*) allocExecutableMemory(CodeSize + CalcDatasetItemSize()))
 	, literalPos(ImulRcpLiteralsEnd)
 	, num32bitLiterals(0)

src/crypto/randomx/jit_compiler_a64.hpp

@@ -46,7 +46,7 @@ namespace randomx {
 
 	class JitCompilerA64 {
 	public:
-		JitCompilerA64();
+		explicit JitCompilerA64(bool);
 		~JitCompilerA64();
 
 		void prepare() {}

src/crypto/randomx/jit_compiler_fallback.hpp

@@ -41,7 +41,7 @@ namespace randomx {
 
 	class JitCompilerFallback {
 	public:
-		JitCompilerFallback() {
+		explicit JitCompilerFallback(bool) {
 			throw std::runtime_error("JIT compilation is not supported on this platform");
 		}
 		void prepare() {}

src/crypto/randomx/jit_compiler_x86.cpp

@@ -49,6 +49,13 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #   include <cpuid.h>
 #endif
 
+static bool hugePagesJIT = false;
+
+void randomx_set_huge_pages_jit(bool hugePages)
+{
+	hugePagesJIT = hugePages;
+}
+
 namespace randomx {
 	/*
 
@@ -175,8 +182,9 @@ namespace randomx {
 #	endif
 
 	static std::atomic<size_t> codeOffset;
+	constexpr size_t codeOffsetIncrement = 59 * 64;
 
-	JitCompilerX86::JitCompilerX86() {
+	JitCompilerX86::JitCompilerX86(bool hugePagesEnable) {
 		BranchesWithin32B = xmrig::Cpu::info()->jccErratum();
 
 		int32_t info[4];
@@ -186,9 +194,11 @@ namespace randomx {
 		cpuid(0x80000001, info);
 		hasXOP = ((info[2] & (1 << 11)) != 0);
 
-		allocatedCode = (uint8_t*)allocExecutableMemory(CodeSize * 2);
+		allocatedCode = (uint8_t*)allocExecutableMemory(CodeSize * 2, hugePagesJIT && hugePagesEnable);
+
 		// Shift code base address to improve caching - all threads will use different L2/L3 cache sets
-		code = allocatedCode + (codeOffset.fetch_add(59 * 64) % CodeSize);
+		code = allocatedCode + (codeOffset.fetch_add(codeOffsetIncrement) % CodeSize);
+
 		memcpy(code, codePrologue, prologueSize);
 		if (hasXOP) {
 			memcpy(code + prologueSize, codeLoopLoadXOP, loopLoadXOPSize);
@@ -207,6 +217,7 @@ namespace randomx {
 	}
 
 	JitCompilerX86::~JitCompilerX86() {
+		codeOffset.fetch_sub(codeOffsetIncrement);
 		freePagedMemory(allocatedCode, CodeSize);
 	}
 

src/crypto/randomx/jit_compiler_x86.hpp

@@ -47,7 +47,7 @@ namespace randomx {
 
 	class JitCompilerX86 {
 	public:
-		JitCompilerX86();
+		explicit JitCompilerX86(bool hugePagesEnable);
 		~JitCompilerX86();
 		void prepare();
 		void generateProgram(Program&, ProgramConfiguration&, uint32_t);

src/crypto/randomx/randomx.cpp

@@ -381,7 +381,7 @@ extern "C" {
 					break;
 
 				case RANDOMX_FLAG_JIT:
-					cache->jit          = new randomx::JitCompiler();
+					cache->jit          = new randomx::JitCompiler(false);
 					cache->initialize   = &randomx::initCacheCompile;
 					cache->datasetInit  = cache->jit->getDatasetInitFunc();
 					cache->memory       = memory;

src/crypto/randomx/randomx.h

@@ -169,6 +169,7 @@ void randomx_apply_config(const T& config)
 }
 
 void randomx_set_scratchpad_prefetch_mode(int mode);
+void randomx_set_huge_pages_jit(bool hugePages);
 
 #if defined(__cplusplus)
 extern "C" {

src/crypto/randomx/virtual_memory.cpp

@@ -33,8 +33,8 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "crypto/randomx/virtual_memory.hpp"
 
 
-void* allocExecutableMemory(std::size_t bytes) {
-    void *mem = xmrig::VirtualMemory::allocateExecutableMemory(bytes);
+void* allocExecutableMemory(std::size_t bytes, bool hugePages) {
+    void *mem = xmrig::VirtualMemory::allocateExecutableMemory(bytes, hugePages);
     if (mem == nullptr) {
         throw std::runtime_error("Failed to allocate executable memory");
     }

src/crypto/randomx/virtual_memory.hpp

@@ -30,6 +30,6 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include <cstddef>
 
-void* allocExecutableMemory(std::size_t);
+void* allocExecutableMemory(std::size_t, bool);
 void* allocLargePagesMemory(std::size_t);
 void freePagedMemory(void*, std::size_t);

src/crypto/randomx/vm_compiled.hpp

@@ -58,7 +58,7 @@ namespace randomx {
 	protected:
 		void execute();
 
-		JitCompiler compiler;
+		JitCompiler compiler{ true };
 	};
 
 	using CompiledVmDefault = CompiledVm<1>;

src/crypto/rx/Rx.cpp

@@ -100,6 +100,7 @@ bool xmrig::Rx::init(const T &seed, const RxConfig &config, const CpuConfig &cpu
     }
 
     randomx_set_scratchpad_prefetch_mode(config.scratchpadPrefetchMode());
+    randomx_set_huge_pages_jit(cpu.isHugePagesJit());
 
     if (isReady(seed)) {
         return true;