Re: gc.c -- possible logic error?

On Tue, 2 Oct 2007, Sylvain Joyeux wrote:

> I think I found the explanation of (parts of) the problem
> 
> Most objects which are allocated are allocated by define_method ... The 
> profiler says 320_000 objects for define_method, which is roughly the count 
> of objects remaining after garbage collection. 
> 
> My guess would be that the problem related to recursion I found is somewhere 
> in there too

Interesting.  In eval.c there is rb_mod_define_method, which has:

 9639         Data_Get_Struct(body, struct BLOCK, block);

and block gets made from 

 9617     if (argc == 1) {
 9618         id = rb_to_id(argv[0]);
 9619         body = proc_lambda();
 9620     }

where proc_lamba does the allocation,

 8538 static VALUE
 8539 proc_lambda()
 8540 {
 8541     return proc_alloc(rb_cProc, Qtrue);
 8542 }

and that uses Data_Make_Struct

 8454     block = Data_Make_Struct(klass, struct BLOCK, blk_mark, blk_free, data      );
 
All well and good, but what's bothering me is the asymmetry in the mark
and free functions:


 8209
 8210 static void
 8211 blk_mark(data)
 8212     struct BLOCK *data;
 8213 {
 8214     while (data) {
 8215         rb_gc_mark_frame(&data->frame);
 8216         rb_gc_mark((VALUE)data->scope);
 8217         rb_gc_mark((VALUE)data->var);
 8218         rb_gc_mark((VALUE)data->body);
 8219         rb_gc_mark((VALUE)data->self);
 8220         rb_gc_mark((VALUE)data->dyna_vars);
 8221         rb_gc_mark((VALUE)data->cref);
 8222         rb_gc_mark(data->wrapper);
 8223         rb_gc_mark(data->block_obj);
 8224         data = data->prev;
 8225     }
 8226 }
 8227
 8228 static void
 8229 frame_free(frame)
 8230     struct FRAME *frame;
 8231 {
 8232     struct FRAME *tmp;
 8233
 8234     frame = frame->prev;
 8235     while (frame) {
 8236         tmp = frame;
 8237         frame = frame->prev;
 8238         free(tmp);
 8239     }
 8240 }
 8241
 8242 static void
 8243 blk_free(data)
 8244     struct BLOCK *data;
 8245 {
 8246     void *tmp;
 8247
 8248     while (data) {
 8249         frame_free(&data->frame);
 8250         tmp = data;
 8251         data = data->prev;
 8252         free(tmp);
 8253     }
 8254 }
 8255

So a whole load of components of data get marked by the GC, but don't
get freed in the free function, because only the frame component
gets that treatment.  Components var, body and cref are all pointers,
so where do they get unmarked and freed?  I can't match /free.*cref/
in this file (eval.c).

I've never gone to this depth before, so I'm definitely getting
out of my depth now.  

> -- 
> Sylvain

        Hugh