more on thread-local allocation

svn: r16879

src/mzscheme/gc2/vm_osx.c

@@ -279,25 +279,16 @@ void exception_thread(void *shared_thread_state)
   }
 }
 
-/* this initializes the subsystem (sets the exception port, starts the
+void GC_attach_current_thread_exceptions_to_handler()
-   exception handling thread, etc) */
-static void macosx_init_exception_handler() 
 {
   mach_port_t thread_self, exc_port_s;
   mach_msg_type_name_t type;
   kern_return_t retval;
 
-  /* get ids for ourself */
+  if (!task_self) return;
-  if(!task_self) task_self = mach_task_self();
-  thread_self = mach_thread_self();
 
-  /* allocate the port we're going to get exceptions on */
+  /* get ids for ourself */
-  retval = mach_port_allocate(task_self, MACH_PORT_RIGHT_RECEIVE, &exc_port);
+  thread_self = mach_thread_self();
-  if(retval != KERN_SUCCESS) {
-    GCPRINT(GCOUTF, "Couldn't allocate exception port: %s\n", 
-	   mach_error_string(retval));
-    abort();
-  }
 
   /* extract out the send rights for that port, which the OS needs */
   retval = mach_port_extract_right(task_self, exc_port, MACH_MSG_TYPE_MAKE_SEND,
@@ -315,6 +306,25 @@ static void macosx_init_exception_handler()
     GCPRINT(GCOUTF, "Couldn't set exception ports: %s\n", mach_error_string(retval));
     abort();
   }
+}
+
+/* this initializes the subsystem (sets the exception port, starts the
+   exception handling thread, etc) */
+static void macosx_init_exception_handler() 
+{
+  kern_return_t retval;
+
+  if(!task_self) task_self = mach_task_self();
+
+  /* allocate the port we're going to get exceptions on */
+  retval = mach_port_allocate(task_self, MACH_PORT_RIGHT_RECEIVE, &exc_port);
+  if(retval != KERN_SUCCESS) {
+    GCPRINT(GCOUTF, "Couldn't allocate exception port: %s\n", 
+	   mach_error_string(retval));
+    abort();
+  }
+
+  GC_attach_current_thread_exceptions_to_handler();
 
 #ifdef PPC_HAND_ROLLED_THREAD 
   /* Old hand-rolled thread creation. */

src/mzscheme/include/schthread.h

@@ -181,7 +181,8 @@ typedef struct Thread_Local_Variables {
   int num_running_threads_;
   int swap_no_setjmp_;
   int thread_swap_count_;
-  int did_gc_count_;
+  int scheme_did_gc_count_;
+  int worker_gc_counter_;
   struct Scheme_Object **scheme_current_runstack_start_;
   struct Scheme_Object **scheme_current_runstack_;
   MZ_MARK_STACK_TYPE scheme_current_cont_mark_stack_;
@@ -344,7 +345,8 @@ XFORM_GC_VARIABLE_STACK_THROUGH_THREAD_LOCAL;
 #define num_running_threads XOA (scheme_get_thread_local_variables()->num_running_threads_)
 #define swap_no_setjmp XOA (scheme_get_thread_local_variables()->swap_no_setjmp_)
 #define thread_swap_count XOA (scheme_get_thread_local_variables()->thread_swap_count_)
-#define did_gc_count XOA (scheme_get_thread_local_variables()->did_gc_count_)
+#define scheme_did_gc_count XOA (scheme_get_thread_local_variables()->scheme_did_gc_count_)
+#define worker_gc_counter XOA (scheme_get_thread_local_variables()->worker_gc_counter_)
 #define scheme_current_runstack_start XOA (scheme_get_thread_local_variables()->scheme_current_runstack_start_)
 #define scheme_current_runstack XOA (scheme_get_thread_local_variables()->scheme_current_runstack_)
 #define scheme_current_cont_mark_stack XOA (scheme_get_thread_local_variables()->scheme_current_cont_mark_stack_)

src/mzscheme/src/future.c

@@ -75,10 +75,12 @@ static int gc_not_ok;
 THREAD_LOCAL_DECL(extern unsigned long GC_gen0_alloc_page_ptr);
 #endif
 
-future_t **g_current_ft;
+static future_t **g_current_ft;
-Scheme_Object ***g_scheme_current_runstack;
+static Scheme_Object ***g_scheme_current_runstack;
-Scheme_Object ***g_scheme_current_runstack_start;
+static Scheme_Object ***g_scheme_current_runstack_start;
-void **g_jit_future_storage;
+static void **g_jit_future_storage;
+static int *gc_counter_ptr;
+THREAD_LOCAL_DECL(static int worker_gc_counter);
 
 static void register_traversers(void);
 extern void scheme_on_demand_generate_lambda(Scheme_Native_Closure *nc, int argc, Scheme_Object **argv);
@@ -307,6 +309,7 @@ void futures_init(void)
 	pthread_attr_setstacksize(&attr, INITIAL_C_STACK_SIZE); 
   for (i = 0; i < THREAD_POOL_SIZE; i++)
   {
+      gc_counter_ptr = &scheme_did_gc_count;
       pthread_create(&threadid, &attr, worker_thread_future_loop, &i);
       sema_wait(&ready_sema);
 	
@@ -326,6 +329,12 @@ static void start_gc_not_ok()
   pthread_mutex_lock(&gc_ok_m);
   gc_not_ok++;
   pthread_mutex_unlock(&gc_ok_m);
+#ifdef MZ_PRECISE_GC
+  if (worker_gc_counter != *gc_counter_ptr) {
+    GC_gen0_alloc_page_ptr = 0; /* forces future to ask for memory */
+    worker_gc_counter = *gc_counter_ptr;
+  }
+#endif
 }
 
 static void end_gc_not_ok()
@@ -581,7 +590,6 @@ Scheme_Object *touch(int argc, Scheme_Object *argv[])
     return retval;
 }
 
-
 //Entry point for a worker thread allocated for
 //executing futures.  This function will never terminate
 //(until the process dies).
@@ -616,10 +624,6 @@ void *worker_thread_future_loop(void *arg)
         g_fuel_pointers[id] = &scheme_fuel_counter;
         g_stack_boundary_pointers[id] = &scheme_jit_stack_boundary;
 
-				#ifdef MZ_PRECISE_GC
-        GC_gen0_alloc_page_ptr = 0; /* disables inline allocation */
-				#endif
-
 				g_current_ft = &current_ft;
 				g_scheme_current_runstack = &scheme_current_runstack;
 				g_scheme_current_runstack_start = &scheme_current_runstack_start;
@@ -744,6 +748,7 @@ int future_do_runtimecall(
     future->rt_prim = NULL;
 
     retval = future->rt_prim_retval;
+    future->rt_prim_retval = NULL;
     pthread_mutex_unlock(&g_future_queue_mutex);
     return 1;
 }
@@ -777,19 +782,21 @@ int rtcall_void_void(void (*f)())
 }
 
 
-int rtcall_void_pvoid(void (*f)(), void **retval)
+int rtcall_alloc_void_pvoid(void (*f)(), void **retval)
 {
 	START_XFORM_SKIP;
 	future_t *future;
 	prim_data_t data;
+
+        while (1) {
           memset(&data, 0, sizeof(prim_data_t));
           if (!IS_WORKER_THREAD)
             {
               return 0;
             }
 
-	data.void_pvoid = f;
+          data.alloc_void_pvoid = f;
-	data.sigtype = SIG_VOID_PVOID;
+          data.sigtype = SIG_ALLOC_VOID_PVOID;
 
           future = get_my_future();
           future->rt_prim = (void*)f;
@@ -797,7 +804,12 @@ int rtcall_void_pvoid(void (*f)(), void **retval)
 
           future_do_runtimecall((void*)f, NULL);
 
-        *retval = future->prim_data.retval;
+          *retval = future->alloc_retval;
+          future->alloc_retval = NULL;
+
+          if (*gc_counter_ptr == future->alloc_retval_counter)
+            break;
+        }
 
 	return 1;
 	END_XFORM_SKIP;
@@ -840,6 +852,7 @@ int rtcall_obj_int_pobj_obj(
 
 	future_do_runtimecall((void*)f, NULL);
 	*retval = future->prim_data.retval;
+        future->prim_data.retval = NULL;
 
 	return 1;
 	END_XFORM_SKIP;
@@ -880,6 +893,7 @@ int rtcall_int_pobj_obj(
 
 	future_do_runtimecall((void*)f, NULL);
 	*retval = future->prim_data.retval;
+        future->prim_data.retval = NULL;
 
 	return 1;
 	END_XFORM_SKIP;
@@ -962,6 +976,7 @@ int rtcall_int_pobj_obj_obj(
 
 	future_do_runtimecall((void*)f, NULL);
 	*retval = future->prim_data.retval;
+        future->prim_data.retval = NULL;
 
 	return 1;
 	END_XFORM_SKIP;
@@ -995,11 +1010,12 @@ void *invoke_rtcall(future_t *future)
       pret = &dummy_ret;
       break;
 		}
-    case SIG_VOID_PVOID:
+    case SIG_ALLOC_VOID_PVOID:
 		{
-			prim_void_pvoid_t func = pdata->void_pvoid;
+			prim_alloc_void_pvoid_t func = pdata->alloc_void_pvoid;
 			ret = func();
-                        pdata->retval = ret;
+                        future->alloc_retval = ret;
+                        future->alloc_retval_counter = scheme_did_gc_count;
                         break;
 		}
     case SIG_OBJ_INT_POBJ_OBJ:

src/mzscheme/src/future.h

@@ -33,7 +33,7 @@ extern int future_do_runtimecall(void *func, void *retval);
 extern void futures_init(void);
 
 typedef void (*prim_void_void_t)(void);
-typedef void *(*prim_void_pvoid_t)(void);
+typedef void *(*prim_alloc_void_pvoid_t)(void);
 typedef Scheme_Object* (*prim_obj_int_pobj_obj_t)(Scheme_Object*, int, Scheme_Object**);
 typedef Scheme_Object* (*prim_int_pobj_obj_t)(int, Scheme_Object**);
 typedef Scheme_Object* (*prim_int_pobj_obj_obj_t)(int, Scheme_Object**, Scheme_Object*);
@@ -43,7 +43,7 @@ typedef struct {
 	unsigned int sigtype;
 
 	prim_void_void_t void_void;
-	prim_void_pvoid_t void_pvoid;
+	prim_alloc_void_pvoid_t alloc_void_pvoid;
 	prim_obj_int_pobj_obj_t obj_int_pobj_obj;
 	prim_int_pobj_obj_t int_pobj_obj;
 	prim_int_pobj_obj_obj_t int_pobj_obj_obj;
@@ -90,6 +90,8 @@ typedef struct future {
 	void *rt_prim_retval;
 
 	prim_data_t prim_data;
+  void *alloc_retval;
+  int alloc_retval_counter;
 
 	Scheme_Object *retval;
 	struct future *prev;
@@ -152,7 +154,7 @@ extern void print_ms_and_us(void);
 //Signature flags for primitive invocations
 //Here the convention is SIG_[arg1type]_[arg2type]..._[return type]
 #define SIG_VOID_VOID 1 						//void -> void
-#define SIG_VOID_PVOID 2 						//void -> void*
+#define SIG_ALLOC_VOID_PVOID 2 						//void -> void*
 #define SIG_OBJ_INT_POBJ_OBJ 3 			//Scheme_Object* -> int -> Scheme_Object** -> Scheme_Object*
 #define SIG_INT_OBJARR_OBJ 4 				//int -> Scheme_Object*[] -> Scheme_Object
 #define SIG_INT_POBJ_OBJ_OBJ 17			//int -> Scheme_Object** -> Scheme_Object* -> Scheme_Object* 
@@ -169,7 +171,7 @@ extern void print_ms_and_us(void);
 															}
 
 extern int rtcall_void_void(void (*f)());
-extern int rtcall_void_pvoid(void (*f)(), void **retval);
+extern int rtcall_alloc_void_pvoid(void (*f)(), void **retval);
 extern int rtcall_obj_int_pobj_obj(
 	Scheme_Object* (*f)(Scheme_Object*, int, Scheme_Object**), 
 	Scheme_Object *a, 
@@ -195,7 +197,7 @@ extern int rtcall_int_pobj_obj(
 #define LOG_THISCALL LOG(__FUNCTION__)
 
 #define LOG_RTCALL_VOID_VOID(f) LOG("(function=%p)", f)
-#define LOG_RTCALL_VOID_PVOID(f) LOG("(function=%p)", f)
+#define LOG_RTCALL_ALLOC_VOID_PVOID(f) LOG("(function=%p)", f)
 #define LOG_RTCALL_OBJ_INT_POBJ_OBJ(f,a,b,c) LOG("(function = %p, a=%p, b=%d, c=%p)", f, a, b, c)
 #define LOG_RTCALL_OBJ_INT_POBJ_VOID(a,b,c) LOG("(%p, %d, %p)", a, b,c)
 #define LOG_RTCALL_INT_OBJARR_OBJ(a,b) LOG("(%d, %p)", a, b)
@@ -218,7 +220,7 @@ extern int rtcall_int_pobj_obj(
 #define LOG_THISCALL
 
 #define LOG_RTCALL_VOID_VOID(f)
-#define LOG_RTCALL_VOID_PVOID(f)
+#define LOG_RTCALL_ALLOC_VOID_PVOID(f)
 #define LOG_RTCALL_OBJ_INT_POBJ_OBJ(f,a,b,c)
 #define LOG_RTCALL_OBJ_INT_POBJ_VOID(a,b,c)
 #define LOG_RTCALL_INT_OBJARR_OBJ(a,b)

src/mzscheme/src/jit.c

@@ -2256,7 +2256,7 @@ static void *ts_prepare_retry_alloc(void *p, void *p2)
   LOG_PRIM_START(&prepare_retry_alloc);
   jit_future_storage[0] = p;
   jit_future_storage[1] = p2;
-  if (rtcall_void_pvoid(GC_make_jit_nursery_page,
+  if (rtcall_alloc_void_pvoid(GC_make_jit_nursery_page,
                               &ret)) {
     GC_gen0_alloc_page_ptr = ret;
     retry_alloc_r1 = jit_future_storage[1];
@@ -2284,6 +2284,7 @@ static void *ts_prepare_retry_alloc(void *p, void *p2)
 # define ts_scheme_apply_from_native _scheme_apply_from_native
 # define ts_scheme_tail_apply_from_native _scheme_tail_apply_from_native
 # define ts_on_demand on_demand
+# define ts_prepare_retry_alloc prepare_retry_alloc
 #endif
 
 static int generate_direct_prim_tail_call(mz_jit_state *jitter, int num_rands)

src/mzscheme/src/salloc.c

@@ -220,6 +220,9 @@ void scheme_set_report_out_of_memory(Scheme_Report_Out_Of_Memory_Proc p)
 
 #ifdef OS_X
 #include <mach/mach.h>
+# ifdef MZ_PRECISE_GC
+extern void GC_attach_current_thread_exceptions_to_handler();
+# endif
 #endif
 
 #ifdef MZ_XFORM
@@ -241,6 +244,11 @@ void scheme_init_os_thread()
   }
 # endif
 #endif
+#ifdef OS_X
+# ifdef MZ_PRECISE_GC
+  GC_attach_current_thread_exceptions_to_handler();
+# endif
+#endif
 }
 #ifdef MZ_XFORM
 END_XFORM_SKIP;

src/mzscheme/src/thread.c

@@ -160,7 +160,7 @@ volatile int *scheme_fuel_counter_ptr;
 THREAD_LOCAL_DECL(static int swap_no_setjmp = 0);
 
 THREAD_LOCAL_DECL(static int thread_swap_count);
-THREAD_LOCAL_DECL(static int did_gc_count);
+THREAD_LOCAL_DECL(int scheme_did_gc_count);
 
 static int init_load_on_demand = 1;
 
@@ -7390,7 +7390,7 @@ static void get_ready_for_GC()
     scheme_current_thread->gmp_tls_data = data;
   }
 
-  did_gc_count++;
+  scheme_did_gc_count++;
 }
 
 extern int GC_words_allocd;
@@ -7575,7 +7575,7 @@ static Scheme_Object *current_stats(int argc, Scheme_Object *argv[])
     case 5:
       SCHEME_VEC_ELS(v)[4] = scheme_make_integer(thread_swap_count);
     case 4:
-      SCHEME_VEC_ELS(v)[3] = scheme_make_integer(did_gc_count);
+      SCHEME_VEC_ELS(v)[3] = scheme_make_integer(scheme_did_gc_count);
     case 3:
       SCHEME_VEC_ELS(v)[2] = scheme_make_integer(gcend);
     case 2: