⚡ add concurrent mark for experiment

CMakeLists.txt

@@ -40,6 +40,7 @@ add_executable(nasal ${CMAKE_SOURCE_DIR}/src/main.cpp)
 target_link_libraries(nasal nasal-object)
 if(NOT CMAKE_HOST_SYSTEM_NAME MATCHES "Windows")
     target_link_libraries(nasal dl)
+    target_link_libraries(nasal pthread)
 endif()
 target_include_directories(nasal PRIVATE ${CMAKE_SOURCE_DIR}/src)
 if(NOT CMAKE_HOST_SYSTEM_NAME MATCHES "Windows")

makefile

@@ -39,7 +39,7 @@ NASAL_OBJECT=\
 
 # for test
 nasal: $(NASAL_OBJECT) | build
-	$(CXX) $(NASAL_OBJECT) -O3 -o nasal -ldl
+	$(CXX) $(NASAL_OBJECT) -O3 -o nasal -ldl -lpthread
 
 nasal.exe: $(NASAL_OBJECT) | build
 	$(CXX) $(NASAL_OBJECT) -O3 -o nasal.exe

src/main.cpp

@@ -145,7 +145,7 @@ void execute(
     // get running time
     if (cmd&VM_TIME) {
         f64 tm = (clk::now()-start).count()*1.0/den;
-        std::clog << " process exited after " << tm << "s.\n\n";
+        std::clog << "process exited after " << tm << "s.\n\n";
     }
 }
 

src/nasal_builtin.cpp

@@ -1,4 +1,5 @@
 #include "nasal_builtin.h"
+#include <chrono>
 
 var builtin_print(var* local, gc& ngc) {
     for(auto& i : local[1].vec().elems) {
@@ -1234,6 +1235,24 @@ var builtin_gcextend(var* local, gc& ngc) {
     return nil;
 }
 
+var builtin_gcinfo(var* local, gc& ngc) {
+    auto den = std::chrono::high_resolution_clock::duration::period::den;
+    var res = ngc.alloc(vm_hash);
+
+    double total = 0;
+    for(u32 i = 0; i<gc_type_size; ++i) {
+        total += ngc.gcnt[i];
+    }
+    // using ms
+    auto& map = res.hash().elems;
+    map["total"] = var::num(ngc.worktime*1.0/den*1000);
+    map["average"] = var::num(ngc.worktime*1.0/den*1000/total);
+    map["max_gc"] = var::num(ngc.max_time*1.0/den*1000);
+    map["max_mark"] = var::num(ngc.max_mark_time*1.0/den*1000);
+    map["max_sweep"] = var::num(ngc.max_sweep_time*1.0/den*1000);
+    return res;
+}
+
 var builtin_logtime(var* local, gc& ngc) {
     time_t t = time(nullptr);
     tm* tm_t = localtime(&t);
@@ -1346,11 +1365,12 @@ nasal_builtin_table builtin[] = {
     {"__md5", builtin_md5},
     {"__cocreate", builtin_cocreate},
     {"__coresume", builtin_coresume},
-    {"__coyield", builtin_coyield },
+    {"__coyield", builtin_coyield},
     {"__costatus", builtin_costatus},
     {"__corun", builtin_corun},
     {"__millisec", builtin_millisec},
     {"__gcextd", builtin_gcextend},
+    {"__gcinfo", builtin_gcinfo},
     {"__logtime", builtin_logtime},
     {"__ghosttype", builtin_ghosttype},
     {nullptr, nullptr}

src/nasal_builtin.h

@@ -126,6 +126,7 @@ var builtin_costatus(var*, gc&);
 var builtin_corun(var*, gc&);
 var builtin_millisec(var*, gc&);
 var builtin_gcextend(var*, gc&);
+var builtin_gcinfo(var*, gc&);
 var builtin_logtime(var*, gc&);
 var builtin_ghosttype(var*, gc&);
 

src/nasal_gc.cpp

@@ -357,6 +357,18 @@ void gc::do_mark_sweep() {
 void gc::mark() {
     std::vector<var> bfs;
     mark_context(bfs);
+    if (memory.size()>1048576 && bfs.size()>4) {
+        usize size = bfs.size();
+        std::thread t0(&gc::concurrent_mark, this, std::ref(bfs), 0, size/4);
+        std::thread t1(&gc::concurrent_mark, this, std::ref(bfs), size/4, size/2);
+        std::thread t2(&gc::concurrent_mark, this, std::ref(bfs), size/2, size/4*3);
+        std::thread t3(&gc::concurrent_mark, this, std::ref(bfs), size/4*3, size);
+        t0.join();
+        t1.join();
+        t2.join();
+        t3.join();
+        return;
+    }
     
     while(!bfs.empty()) {
         var value = bfs.back();
@@ -369,6 +381,27 @@ void gc::mark() {
     }
 }
 
+void gc::concurrent_mark(std::vector<var>& vec, usize begin, usize end) {
+    std::vector<var> bfs;
+    for(auto i = begin; i<end; ++i) {
+        var value = vec[i];
+        if (value.type<=vm_num ||
+            value.val.gcobj->mark!=gc_status::uncollected) {
+            continue;
+        }
+        mark_var(bfs, value);
+    }
+    while(!bfs.empty()) {
+        var value = bfs.back();
+        bfs.pop_back();
+        if (value.type<=vm_num ||
+            value.val.gcobj->mark!=gc_status::uncollected) {
+            continue;
+        }
+        mark_var(bfs, value);
+    }
+}
+
 void gc::mark_context(std::vector<var>& bfs_queue) {
 
     // scan now running context, this context maybe related to coroutine or main
@@ -579,6 +612,8 @@ void gc::info() const {
     for(usize i = 0; i<indent; ++i) {
         indent_string += "-";
     }
+    auto last_line = indent_string + "+" +
+        indent_string + "-" + indent_string + "-" + indent_string;
     indent_string = indent_string + "+" +
         indent_string + "+" + indent_string + "+" + indent_string;
 
@@ -620,7 +655,7 @@ void gc::info() const {
     std::clog << " | " << max_mark_time*1.0/den*1000 << " ms\n";
     std::clog << " " << left << setw(indent) << setfill(' ') << "max sweep";
     std::clog << " | " << max_sweep_time*1.0/den*1000 << " ms\n";
-    std::clog << indent_string << "\n";
+    std::clog << last_line << "\n";
 }
 
 var gc::alloc(u8 type) {

src/nasal_gc.h

@@ -385,6 +385,7 @@ private:
     /* gc functions */
     void do_mark_sweep();
     void mark();
+    void concurrent_mark(std::vector<var>&, usize, usize);
     void mark_context(std::vector<var>&);
     void mark_var(std::vector<var>&, var&);
     void mark_vec(std::vector<var>&, nas_vec&);

src/nasal_vm.h

@@ -923,7 +923,7 @@ inline void vm::o_ret() {
     ctx.localr = ctx.top[-2].addr();
     ctx.upvalr = ctx.top[-3];
 
-    ctx.top=local-1;
+    ctx.top = local-1;
     ctx.funcr = ctx.top[0];
     ctx.top[0] = ret; // rewrite func with returned value
 

std/lib.nas

@@ -480,7 +480,8 @@ var runtime = {
     # command line arguments
     argv: func() {return globals.arg;},
     gc: {
-        extend: func(type) {return __gcextd;}
+        extend: func(type) {return __gcextd;},
+        info: func() {return __gcinfo;}
     }
 };
 

test/gc_test.nas

@@ -1,12 +1,79 @@
-var res = [];
+
-for(var i=0;i<1e6;i+=1) {
+var test_func = func(test_processes...) {
-    append(res, []);
+    var test_process_total = maketimestamp();
+    test_process_total.stamp();
+
+    var time_stamp = maketimestamp();
+    var info = runtime.gc.info();
+    var gc_total = info.total;
+    var duration = 0;
+    foreach(var f; test_processes) {
+        println("[", os.time(), "] testing ", id(f));
+        time_stamp.stamp();
+        f();
+        duration = time_stamp.elapsedMSec();
+        info = runtime.gc.info();
+        println("[", os.time(), "] ", duration, " ms, gc ",
+            (info.total-gc_total)*100/duration, "%, ",
+            int(1000/duration), " cps");
+        gc_total = info.total;
+    }
+
+    println("[", os.time(), "] test time: ",
+        test_process_total.elapsedMSec(), " ms");
+    
+    info = runtime.gc.info();
+    println("##-gc----------------");
+    println("total    : ", info.total, " ms");
+    println("average  : ", info.average, " ms");
+    println("max gc   : ", info.max_gc, " ms");
+    println("max mark : ", info.max_mark, " ms");
+    println("max sweep: ", info.max_sweep, " ms");
+    println("---------------------");
 }
-res = [];
+
-for(var i=0;i<1e6;i+=1) {
+var append_vec = func {
-    append(res, {});
+    var res = [];
+    for(var i=0; i<1e6; i+=1) {
+        append(res, [1]);
+    }
 }
-res = [];
+
-for(var i=0;i<1e6;i+=1) {
+var append_hash = func {
-    append(res, func {});
+    var res = [];
+    for(var i=0; i<1e6; i+=1) {
+        append(res, {a:1, b:2});
+    }
+}
+
+var append_func = func {
+    var res = [];
+    for(var i=0; i<1e6; i+=1) {
+        append(res, func {
+            println(arg);
+        });
+    }
+}
+
+var append_vec_in_vec = func {
+    var res = [];
+    for(var i=0; i<1e6; i+=1) {
+        append(res, [[]]);
+    }
+}
+
+var append_vec_in_hash = func {
+    for(var i=0; i<1e6; i+=1) {
+        append([], {a:[], b:[]});
+    }
+}
+
+for(var i=0; i<10; i+=1) {
+    test_func(
+        append_vec,
+        append_hash,
+        append_func,
+        append_vec_in_vec,
+        append_vec_in_hash
+    );
 }