improve performance

include/pocketpy/common/memorypool.h

@@ -9,7 +9,7 @@ typedef struct FixedMemoryPool {
     int exceeded_bytes;
 
     char** _free_list;
-    char** _free_list_end;
+    int _free_list_length;
 } FixedMemoryPool;
 
 void FixedMemoryPool__ctor(FixedMemoryPool* self, int BlockSize, int BlockCount);

src/common/memorypool.c

@@ -11,7 +11,7 @@ void FixedMemoryPool__ctor(FixedMemoryPool* self, int BlockSize, int BlockCount)
     self->data = PK_MALLOC(BlockSize * BlockCount);
     self->data_end = self->data + BlockSize * BlockCount;
     self->_free_list = PK_MALLOC(sizeof(void*) * BlockCount);
-    self->_free_list_end = self->_free_list;
+    self->_free_list_length = BlockCount;
     for(int i = 0; i < BlockCount; i++) {
         self->_free_list[i] = self->data + i * BlockSize;
     }
@@ -23,9 +23,9 @@ void FixedMemoryPool__dtor(FixedMemoryPool* self) {
 }
 
 void* FixedMemoryPool__alloc(FixedMemoryPool* self) {
-    if(self->_free_list_end != self->_free_list) {
-        self->_free_list_end--;
-        return *self->_free_list_end;
+    if(self->_free_list_length > 0) {
+        self->_free_list_length--;
+        return self->_free_list[self->_free_list_length];
     } else {
         self->exceeded_bytes += self->BlockSize;
         return PK_MALLOC(self->BlockSize);
@@ -35,8 +35,8 @@ void* FixedMemoryPool__alloc(FixedMemoryPool* self) {
 void FixedMemoryPool__dealloc(FixedMemoryPool* self, void* p) {
     bool is_valid = (char*)p >= self->data && (char*)p < self->data_end;
     if(is_valid) {
-        *self->_free_list_end = p;
-        self->_free_list_end++;
+        self->_free_list[self->_free_list_length] = p;
+        self->_free_list_length++;
     } else {
         self->exceeded_bytes -= self->BlockSize;
         PK_FREE(p);

src/interpreter/ceval.c

@@ -199,13 +199,14 @@ FrameResult VM__run_top_frame(VM* self) {
                 /*****************************************/
             case OP_LOAD_FAST: {
                 assert(!frame->is_locals_special);
-                PUSH(&frame->locals[byte.arg]);
-                if(py_isnil(TOP())) {
-                    py_Name name = c11__getitem(uint16_t, &frame->co->varnames, byte.arg);
-                    UnboundLocalError(name);
-                    goto __ERROR;
+                py_Ref val = &frame->locals[byte.arg];
+                if(!py_isnil(val)) {
+                    PUSH(val);
+                    DISPATCH();
                 }
-                DISPATCH();
+                py_Name name = c11__getitem(uint16_t, &frame->co->varnames, byte.arg);
+                UnboundLocalError(name);
+                goto __ERROR;
             }
             case OP_LOAD_NAME: {
                 assert(frame->is_locals_special);

src/interpreter/vm.c

@@ -447,7 +447,9 @@ static bool
 }
 
 FrameResult VM__vectorcall(VM* self, uint16_t argc, uint16_t kwargc, bool opcall) {
+#ifndef NDEBUG
     pk_print_stack(self, self->top_frame, (Bytecode){0});
+#endif
 
     py_Ref p1 = self->stack.sp - kwargc * 2;
     py_Ref p0 = p1 - argc - 2;