The current implementation of swapping could use a lot of work.  Swapping
an object in/out isn't that expensive CPU wise, and can save a decent amount
of memory.  The problem is that cloned programs can't swap, which disables
swapping for much of the mud ...

-Beek

-----

7. It would be more real, if a new call_out happen after all the previous
   call_outs of the same delay. (because it were called later)
   example: call_out("foo",1);
            call_out("bar",1);
   now bar will be called first and not foo! It's unreal.
   could be possible a call_out with 0 delay ? we would use it for
   safe destruct object.

Comment:

Ideally, a better fix would be for call_out to have a smaller
granularity, so the second call_out follows the first by whatever the
time difference between the two was.  This problem will always happen
if the call_outs are within the granularity, and real small
granularities (microsecond?) may/may not be desirable.

-----

The names of includes which define no code don't have to be saved for
line number purposes (although they are still needed for binaries; see
the other note about merging the two lists)

-----

Master applies should be cached like sefuns are.  Currently, APPLY_*
are string constants so they always miss the call_other cache, making
master applies slow.  APPLY_CREATE also always misses the cache, but
it's always going to miss anyway so no problem :-).  Others like APPLY_ID
etc should be changed to use shared strings so they can hit.

-----

There should be a NO_HEART_BEATS option

-----

we need to made a single-user option for MudOS that doesn't use sockets
at all so that people wanting to run MudOS at schools where programs
aren't allowed to listen on socket (anti-mud policies) can use MudOS
for developing LPC code for another mud.

--john

-----

From: John Garnett <garnett@gestalt.austin.tx.us>


It'd be cool if there were efuns like this:

mapping x;

attach_mapping(mapping x, string filename);

detach_mapping(x);

which would attach mapping x to the database named by filename.

any accesses to the mapping x would in fact manipulate the database.

This would work much the way attached mappings work in Perl.

We could use the dbm database code in UNIX(tm).  Only problem is
not all UNIX's have dbm's that support multiple databases per
UNIX process.

-----

In message <9303080120.AA26940@rock.concert.net> Dank writes:
>powerful.  This seems "grossly unfair" to our more serious gamers.  That's
>why i need to know the best way to implement (either in the driver or in LPC)
>an hb_call_out() efun, that operates exactly as call_out() does, except that
>it calls its function after a set number of _heartbeats_ have elapsed, not
>after a set number of seconds.
>  I ran this by Mobydick, and he thought the solution should be sought from
>you driver folks rather than in the mudlib (course, that may be because
>he's a mudlib coder :)

I agree; it does seem that we need a variant of call_out that operates in terms
of number of heartbeats rather than absolute # of seconds.  A few ways
to do this.

// optionally change the way call_out() works so that muds can change it if
// they like.
#define CALL_OUT_UNITS_EQUAL_HEARTBEAT /* or some shorter similar thing */

Or we could make a call_out2() efun that operates in the desired heartbeat units
(name it something more appropriate than call_out2()).

-----

hmm, save_object() should probably somehow save the subtype of numbers.
(ie null, undefined, error, etc)
 
ucs_brf@pip.shsu.edu

Comment: with some of the optimizations lately, the subtype of numbers is
	not always reliable longterm ...

-----

An option for persistent swap files (that the driver can restore/recover
from), saving clones and the contents of variables.

### This seems to be a fad and a bad idea, but I'll get flamed for saying
	 that ...

-----

mudlib_stats uses a linked list, it is fast for insertion; however
  lookup is O(n)...this should be hacked to use an AVL tree (ala uid.c)

-- robo

-----

#pragma no_inherit
#pragma no_clone

-- Cygnus & Descartes

Comment:
    Unlike regular C, should these two pragmas should generate warnings
    (or error out) if the driver wasn't compiled with this option enabled?

    Also, adapt Tensor's idea for uniquep() efun...which queries these
    flags, telling whether an object is cloneable and/or inheritable
    (or not, depending on your point of view).

    These aren't object flags as first believed...they have to be stored
    in the program structure, so that it can be restored from binaries.
    pragma_no_inherit & pragma_no_clone flags should probably be in
    simulate.c, disjoint from those in lex.c.
        clone_object()
          ob = find_object_no_load()
          if (ob->prog->flags & P_NO_CLONE)
            error("Illegal object to clone\n");

          find_object()
            find_object2() <- avoid calling this again (for performance)
            load_object()

        inheritance: (compiler/grammar & lpc_compiler/grammar)
          ob = find_object2()

          if (ob->prog->flags & P_NO_INHERIT)
             yyerror("Illegal object to inherit\n");

        lex.c
          init pragma_no_inherit
               pragma_no_clone

        compiler/grammar & lpc_compiler/grammar
          set prog flags

-----

Implement compile stacks so we don't have to dump an object (and later
restart the compile) if an inherited object isn't defined yet.

Note: This would entail requiring the use of bison or pre-yacc'ed grammars
since most versions of yacc parsers aren't reentrant. -Beek

Actually, I take this back.  One could make the inherit rule look like:

inheritance: type_modifier_list L_INHERIT string_con1 ';'
    {
       ...
       ob = find_object2(file);
       if (ob == 0) {
           push_compiler_globals();
           lex_inherit_hook(file);
       } else {
           lex_loaded_inherit_hook();
       }
    }
  program terminator
    {
       ... import symbols ...
    }

terminator:
    L_ALREADY_LOADED
  | L_JUST_LOADED
    {
       pop_compiler_globals();
    }

etc ... lex_inherit_hook() cues lex to push it's globals and start a new
file; lex_loaded_inherit_hook() cues lex to just return L_ALREADY_LOADED
as the next token (note: in deep inherits this could require a rather
deep yacc stack)

---

Set up grammar dir for pre-byacc'd and/or pre-bison'd parsers for people
  with broken 'yacc's.

Support L_ARRAY_CLOSE, L_FUNCTION_CLOSE, and L_MAPPING_CLOSE for consistency;
L_MAPPING_CLOSE requires a bit more work...

-- robo

-----

Threads :)

-----

Exceptions :)

-----

A make_mapping() efun that obeys the tautology:

make_mapping( keys(m), values(m) ) == m

-----

include_list() efun; also, BINARIES saves a list of includes, but error
line numbers do as well; the former should be modified to use the latter's
list (which was added after binaries were)

-----

maybe:

foreach (i, x in ({ 1, 2 }))

foreach (x in ([ 1 : 2, 3 : 4 ]))

foreach (i, j in ({ 1, 2 }), ({ 3, 4, 5 }))
   for parallel walking

---

The handling of || and && is still suboptimal; it tests twice, and does
a uneccessary pop for expressions like: (x && y); it completely blows it
on expressions like 1 || 1 || 1;

---

Profiling shows that a significant amount of CPU is wasted copying stack
values around in the parser due to the large size of %union.  In most
cases, only 4 (assuming 32 bit) bytes is needed:

4: /* most stuff */
5: decl
8: func_block

so this is a huge waste.  The larger ones could be managed from temporary
pools (they might even have FIFO ordering, but be careful about parse errors)
decl seems unavoidable, but 6 should be possible, cutting the size of
the compile stack and mem that needs to be copied down a bit.
func_block is only used by function pointers anyway, so isn't that common.
There is already a stack for the other function temporary and a depth limit
...

-Beek