GC笔记--fast/slow allocator比较

p_obj = gc_alloc_fast(size, vt);
if(p_obj == null){
prepare_for_native_call();
gc_alloc(size, vt);
clean_after_native_call();
}

 

//this routine does not deal with any slow path operations,
//but returns null if unsuccessful.
Object* gc_alloc_fast(unsigned size, Vtable* vt)
{
//return if object to be allocated has finalizer
if(type_has_finalizer(vt)) return NULL;
//return if it is large object
if ( size > GC_OBJ_SIZE_THRESHOLD ) return NULL;
Object* p_obj = null;
Allocator* allocator = thread_get_allocator();
int free = (int)allocator->free;
int ceiling = (int)allocator->ceiling;
int new_free = free + size;
if (new_free <= ceiling){
p_obj = (Object*)free;
allocator->free= (void*)new_free;
}else{
return null;
}
//install vtable pointer to the object header
obj_set_vt(p_obj, vt);
return p_obj;
}

 

这是在thread local上的实现bump-the-pointer。因为需要找线程私有的allocator，这一部分一般通过thread_local_top寻找。thread_local_top可以固定放在某一个寄存器。

 

fast和普通的slow模式相差非常大。如果fast失败，中间还要插入safepoint点。这一段就会有上百的instruction，开销远远大于fast的allocator。