适配aarch64架构机器及性能优化。

.gitignore

@@ -1,4 +1,5 @@
 README.md
+.vscode
 
 # Prerequisites
 *.d
@@ -32,3 +33,5 @@ README.md
 *.exe
 *.out
 *.app
+.claude/
+.opencode/

CMakeLists.txt

@@ -8,48 +8,21 @@ set(CMAKE_MACOSX_RPATH 0)
 set(LIBCO_VERSION   0.5)
 
 # Set cflags
-set(CMAKE_C_FLAGS ${CMAKE_C_FLAGS} -g -fno-strict-aliasing -O2 -Wall -export-dynamic -Wall -pipe  -D_GNU_SOURCE -D_REENTRANT -fPIC -Wno-deprecated -m64)
-
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -g -fno-strict-aliasing -O2 -Wall -export-dynamic -Wall -pipe -D_GNU_SOURCE -D_REENTRANT -fPIC -Wno-deprecated")
+
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64")
+    message(STATUS "x86_64 architecture detected")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -m64 -mtune=generic")
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64")
+    message(STATUS "ARM64 architecture detected")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=armv8-a -mtune=generic")
+else()
+    message(STATUS "Unknown architecture: ${CMAKE_SYSTEM_PROCESSOR}")
+endif()
 # Use c and asm
 enable_language(C ASM)
 
-# Add source files
-set(SOURCE_FILES
-        co_epoll.cpp
-        co_hook_sys_call.cpp
-        co_routine.cpp
-        coctx.cpp
-        coctx_swap.S)
-
-# Add static and shared library target
-add_library(colib_static STATIC ${SOURCE_FILES})
-add_library(colib_shared SHARED ${SOURCE_FILES})
-
-# Set library output name
-set_target_properties(colib_static PROPERTIES OUTPUT_NAME colib)
-set_target_properties(colib_shared PROPERTIES OUTPUT_NAME colib)
-
-set_target_properties(colib_static PROPERTIES CLEAN_DIRECT_OUTPUT 1)
-set_target_properties(colib_shared PROPERTIES CLEAN_DIRECT_OUTPUT 1)
-
-# Set shared library version, will generate libcolib.${LIBCO_VERSION}.so and a symbol link named libcolib.so
-# For mac osx, the extension name will be .dylib
-set_target_properties(colib_shared PROPERTIES VERSION ${LIBCO_VERSION} SOVERSION ${LIBCO_VERSION})
-
-
-
-# Macro for add example target
-macro(add_example_target EXAMPLE_TARGET)
-    add_executable("example_${EXAMPLE_TARGET}" "example_${EXAMPLE_TARGET}.cpp")
-    target_link_libraries("example_${EXAMPLE_TARGET}" colib_static pthread dl)
-endmacro(add_example_target)
+set(LIBCO_INCLUDE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/lib")
 
-add_example_target(closure)
-add_example_target(cond)
-add_example_target(copystack)
-add_example_target(echocli)
-add_example_target(echosvr)
-add_example_target(poll)
-add_example_target(setenv)
-add_example_target(specific)
-add_example_target(thread)
+add_subdirectory(lib)
+add_subdirectory(examples)

examples/CMakeLists.txt

@@ -0,0 +1,19 @@
+cmake_minimum_required(VERSION 2.8)
+
+macro(add_example_target EXAMPLE_TARGET)
+    get_filename_component(EXAMPLE_NAME ${EXAMPLE_TARGET} NAME)
+    if(${EXAMPLE_TARGET} MATCHES "\\.cpp$")
+        string(REPLACE ".cpp" "" EXAMPLE_NAME ${EXAMPLE_NAME})
+    endif()
+
+    add_executable(${EXAMPLE_NAME} ${EXAMPLE_TARGET})
+    target_include_directories(${EXAMPLE_NAME} PUBLIC ${LIBCO_INCLUDE_DIR})
+    target_link_libraries(${EXAMPLE_NAME} PUBLIC colib_static pthread dl)
+
+    add_test(NAME ${EXAMPLE_NAME} COMMAND ${EXAMPLE_NAME})
+endmacro()
+
+file(GLOB_RECURSE EXAMPLE_TARGETS "example_*")
+foreach(EXAMPLE_TARGET ${EXAMPLE_TARGETS})
+    add_example_target(${EXAMPLE_TARGET})
+endforeach()

lib/CMakeLists.txt

@@ -0,0 +1,32 @@
+cmake_minimum_required(VERSION 2.8)
+
+# Use c and asm
+enable_language(C ASM)
+
+# Add source files
+set(SOURCE_FILES
+        co_epoll.cpp
+        co_hook_sys_call.cpp
+        co_routine.cpp
+        coctx.cpp
+        co_comm.cpp
+        co_memcpy.c
+        coctx_swap.S)
+
+# Add static and shared library target
+add_library(colib_static STATIC ${SOURCE_FILES})
+add_library(colib_shared SHARED ${SOURCE_FILES})
+
+target_compile_definitions(colib_static PRIVATE _GNU_SOURCE _REENTRANT)
+target_compile_definitions(colib_shared PRIVATE _GNU_SOURCE _REENTRANT)
+
+# Set library output name
+set_target_properties(colib_static PROPERTIES OUTPUT_NAME colib)
+set_target_properties(colib_shared PROPERTIES OUTPUT_NAME colib)
+
+set_target_properties(colib_static PROPERTIES CLEAN_DIRECT_OUTPUT 1)
+set_target_properties(colib_shared PROPERTIES CLEAN_DIRECT_OUTPUT 1)
+
+# Set shared library version, will generate libcolib.${LIBCO_VERSION}.so and a symbol link named libcolib.so
+# For mac osx, the extension name will be .dylib
+set_target_properties(colib_shared PROPERTIES VERSION ${LIBCO_VERSION} SOVERSION ${LIBCO_VERSION})

co_closure.h → lib/co_closure.h

@@ -18,6 +18,11 @@
 
 #ifndef __CO_CLOSURE_H__
 #define __CO_CLOSURE_H__
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 struct stCoClosure_t 
 {
 public:
@@ -92,6 +97,8 @@ public:\
 
 #define co_func_end }
 
-
+#ifdef __cplusplus
+}
 #endif
 
+#endif

co_hook_sys_call.cpp → lib/co_hook_sys_call.cpp

@@ -98,7 +98,6 @@ typedef int (*unsetenv_pfn_t)(const char *name);
 typedef char *(*getenv_pfn_t)(const char *name);
 typedef hostent* (*gethostbyname_pfn_t)(const char *name);
 typedef res_state (*__res_state_pfn_t)();
-typedef int (*__poll_pfn_t)(struct pollfd fds[], nfds_t nfds, int timeout);
 typedef int (*gethostbyname_r_pfn_t)(const char* __restrict name, struct hostent* __restrict __result_buf, char* __restrict __buf, size_t __buflen, struct hostent** __restrict __result, int* __restrict __h_errnop);
 
 static socket_pfn_t g_sys_socket_func 	= (socket_pfn_t)dlsym(RTLD_NEXT,"socket");
@@ -128,9 +127,6 @@ static __res_state_pfn_t g_sys___res_state_func  = (__res_state_pfn_t)dlsym(RTLD
 static gethostbyname_pfn_t g_sys_gethostbyname_func = (gethostbyname_pfn_t)dlsym(RTLD_NEXT, "gethostbyname");
 static gethostbyname_r_pfn_t g_sys_gethostbyname_r_func = (gethostbyname_r_pfn_t)dlsym(RTLD_NEXT, "gethostbyname_r");
 
-static __poll_pfn_t g_sys___poll_func = (__poll_pfn_t)dlsym(RTLD_NEXT, "__poll");
-
-
 /*
 static pthread_getspecific_pfn_t g_sys_pthread_getspecific_func 
 			= (pthread_getspecific_pfn_t)dlsym(RTLD_NEXT,"pthread_getspecific");
@@ -970,10 +966,6 @@ extern "C"
 
 		return &(__co_state_wrap->state);
 	}
-	int __poll(struct pollfd fds[], nfds_t nfds, int timeout)
-	{
-		return poll(fds, nfds, timeout);
-	}
 }
 
 struct hostbuf_wrap 
@@ -1035,5 +1027,4 @@ void co_enable_hook_sys() //这函数必须在这里,否则本文件会被忽略
 	{
 		co->cEnableSysHook = 1;
 	}
-}
-
+}

lib/co_memcpy.c

@@ -0,0 +1,138 @@
+#include "co_memcpy.h"
+#include <string.h>
+#include "memcpy_def.h"
+#include "memcpy_neon_impl.h"
+
+#define SIZE_128B 128
+#define SIZE_1KB 1024
+#define SIZE_2KB 2048
+#define SIZE_4KB 4096
+#define SIZE_8KB 8192
+#define SIZE_12KB 12288
+
+static void *co_memcpy_save_aarch64(void *restrict dst, const void *restrict src, unsigned long len)
+{
+    void *dst_ret = dst;
+#ifdef __aarch64__
+    if (len < SIZE_128B) {
+        memcpy_rem(dst, src, len);
+        return dst_ret;
+    }
+
+    unsigned long align_offset = (unsigned long)dst & 0x3f;
+    if (align_offset) {
+        memcpy_rem(dst, src, align_offset);
+        dst += align_offset;
+        src += align_offset;
+        len -= align_offset;
+    }
+
+    int len_count = len / BLOCK_SIZE;
+    int rem = len % BLOCK_SIZE;
+    if (len <= SIZE_1KB) {
+        memcpy_blk(dst, src, len_count, memcpy_blk_save);
+        if (rem) {
+            memcpy_rem(dst, src, rem);
+        }
+    } else if (len <= SIZE_2KB) {
+        memcpy_256B_save(dst, src, len_count, rem);
+        if (rem) {
+            memcpy_rem(dst, src, rem);
+        }
+    } else if (len <= SIZE_4KB) {
+        memcpy_384B_save(dst, src, len_count, rem);
+        if (rem) {
+            memcpy_rem(dst, src, rem);
+        }
+    } else if (len <= SIZE_8KB) {
+        memcpy_512B_save(dst, src, len_count, rem);
+        if (rem) {
+            memcpy_rem(dst, src, rem);
+        }
+    } else if (len <= SIZE_12KB) {
+        memcpy_1024B_save(dst, src, len_count, rem);
+        if (rem) {
+            memcpy_rem(dst, src, rem);
+        }
+    } else {
+        memcpy_1536B_save(dst, src, len_count, rem);
+        if (rem) {
+            memcpy_rem(dst, src, rem);
+        }
+    }
+#endif
+    return dst_ret;
+}
+
+static void *co_memcpy_swap_aarch64(void *restrict dst, const void *restrict src, unsigned long len)
+{
+    void *dst_ret = dst;
+#ifdef __aarch64__
+    if (len < SIZE_128B) {
+        memcpy_rem(dst, src, len);
+        return dst_ret;
+    }
+
+    unsigned long align_offset = (unsigned long)dst & 0x3f;
+    if (align_offset) {
+        memcpy_rem(dst, src, align_offset);
+        dst += align_offset;
+        src += align_offset;
+        len -= align_offset;
+    }
+
+    int len_count = len / BLOCK_SIZE;
+    int rem = len % BLOCK_SIZE;
+    if (len <= SIZE_1KB) {
+        memcpy_blk(dst, src, len_count, memcpy_blk_swap);
+        if (rem) {
+            memcpy_rem(dst, src, rem);
+        }
+    } else if (len <= SIZE_2KB) {
+        memcpy_384B_swap(dst, src, len_count, rem);
+        if (rem) {
+            memcpy_rem(dst, src, rem);
+        }
+    } else if (len <= SIZE_8KB) {
+        memcpy_512B_swap(dst, src, len_count, rem);
+        if (rem) {
+            memcpy_rem(dst, src, rem);
+        }
+    } else if (len <= SIZE_12KB) {
+        memcpy_1024B_swap(dst, src, len_count, rem);
+        if (rem) {
+            memcpy_rem(dst, src, rem);
+        }
+    } else {
+        memcpy_1536B_swap(dst, src, len_count, rem);
+        if (rem) {
+            memcpy_rem(dst, src, rem);
+        }
+    }
+#endif
+    return dst_ret;
+}
+
+typedef void *(*memcpy_func_name)(void *restrict dst, const void *restrict src, unsigned long len);
+
+static memcpy_func_name memcpy_save_func = (memcpy_func_name)memcpy;
+static memcpy_func_name memcpy_swap_func = (memcpy_func_name)memcpy;
+
+static void __attribute__((constructor)) model_init(void)
+{
+#ifdef __aarch64__
+    memcpy_save_func = (memcpy_func_name)co_memcpy_save_aarch64;
+    memcpy_swap_func = (memcpy_func_name)co_memcpy_swap_aarch64;
+#endif
+    return;
+}
+
+void *co_memcpy_save(void *restrict dst, const void *restrict src, unsigned long len)
+{
+    return memcpy_save_func(dst, src, len);
+}
+
+void *co_memcpy_swap(void *restrict dst, const void *restrict src, unsigned long len)
+{
+    return memcpy_swap_func(dst, src, len);
+}

lib/co_memcpy.h

@@ -0,0 +1,15 @@
+#ifndef CO_MEMCPY_H
+#define CO_MEMCPY_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void *co_memcpy_save(void *__restrict dst, const void *__restrict src, unsigned long len);
+void *co_memcpy_swap(void *__restrict dst, const void *__restrict src, unsigned long len);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* CO_MEMCPY_H */