This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: GNU ObjC runtime class table rewritten


Your work is really good, thanks for your contribution! Few
observations, though.

1. In class_table_insert(), when you insert the node in the hash table
that already has a linked list for the a given index, you can insert
it at the beginning of the list. Thus you have constant time for
insertion. However this will not help much during searches, so
probably a better idea is to insert the node in the linked list so
that it is sorted based on the length of the class name. You can than
modify the class_table_get_safe() function to take into account this
modification.

Maybe a smarter arrangement of the linked list could also be
done. Moving the node corresponding to a class to the beginning of the
linked list, based on the access frequency may help in cases where the
program has a large number of classes. However the bookkeeping may be
too costly to be done on each class lookup. Probably a better approach
is to modify the size of the hash table dynamically, as the number of
classes in the program increases.

2. The class_table_replace() does an exhaustive search on the hash
table. I think you could instead use the hash code for the
old_class_pointer to directly identify the linked list that contains
the class. This would speed up things considerably.


Just to satisfy my curiosity, did you do any benchmarks using
class_table_print_histogram to see how good that hash function
performs in a normal OpenStep environment? I'm curious to see how
similar prefixes for the class names affects the hash function.

Other than the above points, the code looks very good!

Thanks,
Ovidiu

On Sun, 27 May 2001 00:24:02 +0100 (BST), Nicola Pero <n.pero@mi.flashnet.it> wrote:

> Hi
> 
> As you probably know, one of the traditional weaknesses of the GNU
> Objective-C runtime is that class method invocations are very slow.  A
> class method invocation takes approximately 12 times an instance
> method invocation on my machine.
> 
> The main reason is that when you write
> 
>  array = [NSArray new];
> 
> this gets basically compiled into the equivalent of 
> 
>  array = [(objc_get_class ("NSArray")) new];
> 
> objc_get_class returns the class pointer corresponding to the string
> `NSArray'; and the lookup is very slow.
> 
> Most high performance Objective-C code I had the opportunity to read
> (or write) work around this problem by caching the class pointer:
> 
>  Class arrayClass = [NSArray class];
> 
>  /* later on */
> 
>  array = [arrayClass new];
>  array = [arrayClass new];
>  array = [arrayClass new];
> 
> In this case, you always perform a class lookup (the first one), but
> then all the [arrayClass new] methods run exactly as fast as an
> instance method invocation.  It helps if you have many class method
> invocations to the same class.
> 
> The long-term solution to this problem would be to modify the compiler
> to output tables of class pointers corresponding to all the class
> method invocations, and to add code to the runtime to update these
> tables - that should in the end allow class method invocations to
> perform precisely as fast as instance method invocations, because no
> class lookup would be involved.  I think the Apple Objective-C runtime
> uses this technique.  Doing this involves synchronized modifications
> in the runtime and in the compiler.  I have not yet looked at doing
> it.
> 
> I took a more direct approach - I studied the way the runtime was
> performing class lookup, and I thought - hey, I can write code which
> does the same, but faster. :-)
> 
> So - I have redesigned and rewritten the class table in the GNU
> Objective-C runtime to be faster (it took me four days in total I
> think).  This doesn't give as much speed as the other approach, but a
> class method invocation now takes approximately 4.5 times an instance
> method invocation on my machine.  That's a considerable improvement
> (it means the core class lookup now takes 1/3 of the original time).
> If I get the time - which I doubt in the short-term - I could have a
> look at the other approach.
> 
> One of the main reason the new class lookup is so faster is because I
> implemented it in a way that can safely run multithreaded without
> using locks - a so-called `lock-free' data structure.  The atomic
> operation is pointer assignment.  The reason why in this problem
> lock-free data structures work so well is that you never remove
> classes from the table - and the difficult thing with lock-free data
> structures is freeing data when is removed from the structures.
> 
> By tricking with C and modifying it after studying the produced
> assembly, my original code would in the end go some 15% faster than
> the code I am submitting.  The fact is, I decided the added complexity
> is not worth this 15% in speed, this 15% in speed is going to become
> obsolete on future/other processors/compilers, and when we implement
> the long-term solution I explained above, all this extreme C mucking
> and tricking for speed would be out of place.  I preferred plain
> straightforward C code, which should hopefully be easy to read and
> maintain, should compile not exceptionally but reasonably well on all
> machines, and will still be good and appropriate even when we
> implement the solution above - class lookup is sometimes done
> independently from class method invocation anyway.
> 
> So - here is the new code.
> 
> This code is *NOT* for the 3.0 release.  I tested it, debugged it,
> profiled it, used it, but you never have enough of that - it's
> experimental code.  It's for the main trunk.
> 
> ChangeLog:
> ==========
> 
> Sat May 26 23:18:41 2001  Nicola Pero  <n.pero@mi.flashnet.it>
> 
> 	* class.c: Rewritten the class table to use optimized, lock-free
> 	lookup.  This more than doubles the speed of class method
> 	invocation.  (class_table_setup), (class_table_insert),
> 	(class_table_replace), (class_table_get_safe),
> 	(class_table_next), (class_table_print),
> 	(class_table_print_histogram): New functions.
> 	(__objc_init_class_tables): Use class_table_setup.
> 	(__objc_add_class_to_hash): Use class_table_get_safe and
> 	class_table_insert.  (objc_lookup_class), (objc_get_class): Do not
> 	assert the existence of the table; do not lock the runtime; use
> 	class_table_get_safe.  (objc_next_class): Use class_table_next.
> 	(__objc_resolve_class_links): Use class_table_next.
> 	(class_pose_as): Use class_table_replace.
> 
> class.c:
> ========
> 	
> /* GNU Objective C Runtime class related functions
>    Copyright (C) 1993, 1995, 1996, 1997, 2001 Free Software Foundation, Inc.
>    Contributed by Kresten Krab Thorup and Dennis Glatting.
> 
>    Lock-free class table code designed and written from scratch by
>    Nicola Pero, 2001.
> 
> This file is part of GNU CC.
> 
> GNU CC is free software; you can redistribute it and/or modify it under the
> terms of the GNU General Public License as published by the Free Software
> Foundation; either version 2, or (at your option) any later version.
> 
> GNU CC is distributed in the hope that it will be useful, but WITHOUT ANY
> WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
> FOR A PARTICULAR PURPOSE.  See the GNU General Public License for more
> details.
> 
> You should have received a copy of the GNU General Public License along with
> GNU CC; see the file COPYING.  If not, write to the Free Software
> Foundation, 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.  */
> 
> /* As a special exception, if you link this library with files compiled with
>    GCC to produce an executable, this does not cause the resulting executable
>    to be covered by the GNU General Public License. This exception does not
>    however invalidate any other reasons why the executable file might be
>    covered by the GNU General Public License.  */
> 
> #include "runtime.h"		/* the kitchen sink */
> #include "sarray.h"
> 
> #include <objc/objc.h>
> #include <objc/objc-api.h>
> #include <objc/thr.h>
> 
> /* We use a table which maps a class name to the corresponding class
>  * pointer.  The first part of this file defines this table, and
>  * functions to do basic operations on the table.  The second part of
>  * the file implements some higher level Objective-C functionality for
>  * classes by using the functions provided in the first part to manage
>  * the table. */
> 
> /**
>  ** Class Table Internals
>  **/
> 
> /* A node holding a class */
> typedef struct class_node
> {
>   struct class_node *next;      /* Pointer to next entry on the list.
> 				   NULL indicates end of list. */
>   
>   const char *name;             /* The class name string */
>   int length;                   /* The class name string length */
>   Class pointer;                /* The Class pointer */
>   
> } *class_node_ptr;
> 
> /* A table containing classes is a class_node_ptr (pointing to the
>    first entry in the table - if it is NULL, then the table is
>    empty). */
> 
> /* We have 1024 tables.  Each table contains all class names which
>    have the same hash (which is a number between 0 and 1023).  To look
>    up a class_name, we compute its hash, and get the corresponding
>    table.  Once we have the table, we simply compare strings directly
>    till we find the one which we want (using the length first).  The
>    number of tables is quite big on purpose (a normal big application
>    has less than 1000 classes), so that you shouldn't normally get any
>    collisions, and get away with a single comparison (which we can't
>    avoid since we need to know that you have got the right thing).  */
> #define CLASS_TABLE_SIZE 1024
> #define CLASS_TABLE_MASK 1023
> 
> static class_node_ptr class_table_array[CLASS_TABLE_SIZE];
> 
> /* The table writing mutex - we lock on writing to avoid conflicts
>    between different writers, but we read without locks.  That is
>    possible because we assume pointer assignment to be an atomic
>    operation. */
> static objc_mutex_t __class_table_lock = NULL;
> 
> /* CLASS_TABLE_HASH is how we compute the hash of a class name.  It is
>    a macro - *not* a function - arguments *are* modified directly.
> 
>    INDEX should be a variable holding an int;
>    HASH should be a variable holding an int;
>    CLASS_NAME should be a variable holding a (char *) to the class_name.
> 
>    After the macro is executed, INDEX contains the length of the
>    string, and HASH the computed hash of the string; CLASS_NAME is
>    untouched.  */
> 
> #define CLASS_TABLE_HASH(INDEX, HASH, CLASS_NAME)          \
>   HASH = 0;                                                  \
>   for (INDEX = 0; CLASS_NAME[INDEX] != '\0'; INDEX++)        \
>     {                                                        \
>       HASH = (HASH << 4) ^ (HASH >> 28) ^ CLASS_NAME[INDEX]; \
>     }                                                        \
>                                                              \
>   HASH = (HASH ^ (HASH >> 10) ^ (HASH >> 20)) & CLASS_TABLE_MASK;
> 
> /* Setup the table */
> static void
> class_table_setup ()
> {
>   /* Start - nothing in the table */
>   memset (class_table_array, 0, sizeof(class_node_ptr) * CLASS_TABLE_SIZE);
> 
>   /* The table writing mutex */
>   __class_table_lock = objc_mutex_allocate ();
> }
> 
> /* Insert a class in the table (used when a new class is registered) */
> static void 
> class_table_insert (const char *class_name, Class class_pointer)
> {
>   int hash, length;
>   class_node_ptr new_node;
>   class_node_ptr node;
>   class_node_ptr next;
> 
>   /* Find out the class name's hash and length */
>   CLASS_TABLE_HASH (length, hash, class_name);
>   
>   /* Prepare the new node holding the class */
>   new_node = objc_malloc (sizeof (struct class_node));
>   new_node->next = NULL;
>   new_node->name = class_name;
>   new_node->length = length;
>   new_node->pointer = class_pointer;
> 
>   /* Lock the table for modifications */
>   objc_mutex_lock (__class_table_lock);
>   
>   /* Insert the new node in the table */
>   node = class_table_array[hash];
>   if (node == NULL)
>     {
>       class_table_array[hash] = new_node;
>     }
>   else
>     {
>       while ((next = node->next) != NULL)
> 	{
> 	  node = next;
> 	}
> 
>       node->next = new_node;
>     }
>   
>   objc_mutex_unlock (__class_table_lock);
> }
> 
> /* Replace a class in the table (used only by poseAs:) */
> static void 
> class_table_replace (Class old_class_pointer, Class new_class_pointer)
> {
>   int hash;
>   class_node_ptr node;
> 
>   objc_mutex_lock (__class_table_lock);
>   
>   hash = 0;
>   node = class_table_array[hash];
>   
>   while (hash < CLASS_TABLE_SIZE)
>     {
>       if (node == NULL)
> 	{
> 	  hash++;
> 	  if (hash < CLASS_TABLE_SIZE)
> 	    {
> 	      node = class_table_array[hash];
> 	    }
> 	}
>       else
> 	{
> 	  Class class1 = node->pointer;
> 
> 	  if (class1 == old_class_pointer)
> 	    {
> 	      node->pointer = new_class_pointer;
> 	    }
> 	  node = node->next;
> 	}
>     }
> 
>   objc_mutex_unlock (__class_table_lock);
> }
> 
> 
> /* Get a class from the table.  This does not need mutex protection.
>    Currently, this function is called each time you call a static
>    method, this is why it must be very fast. */
> static inline Class 
> class_table_get_safe (const char *class_name)
> {
>   class_node_ptr node;  
>   int length, hash;
> 
>   /* Compute length and hash */
>   CLASS_TABLE_HASH (length, hash, class_name);
>   
>   node = class_table_array[hash];
>   
>   if (node != NULL)
>     {
>       do
> 	{
> 	  if (node->length == length)
> 	    {
> 	      /* Compare the class names */
> 	      int i;
> 
> 	      for (i = 0; i < length; i++)
> 		{
> 		  if ((node->name)[i] != class_name[i]) 
> 		    {
> 		      break;
> 		    }
> 		}
> 	      
> 	      if (i == length)
> 		{
> 		  /* They are equal! */
> 		  return node->pointer;
> 		}
> 	    }
> 	}
>       while ((node = node->next) != NULL);
>     }
> 
>   return Nil;
> }
> 
> /* Enumerate over the class table */
> struct class_table_enumerator
> {
>   int hash;
>   class_node_ptr node;
> };
> 
> 
> static Class
> class_table_next (struct class_table_enumerator **e)
> {
>   struct class_table_enumerator *enumerator = *e;
>   class_node_ptr next;
>   
>   if (enumerator == NULL)
>     {
>        *e = objc_malloc (sizeof (struct class_table_enumerator));
>       enumerator = *e;
>       enumerator->hash = 0;
>       enumerator->node = NULL;
> 
>       next = class_table_array[enumerator->hash];
>     }
>   else
>     {
>       next = enumerator->node->next;
>     }
>   
>   if (next != NULL)
>     {
>       enumerator->node = next;
>       return enumerator->node->pointer;
>     }
>   else 
>     {
>       enumerator->hash++;
>      
>       while (enumerator->hash < CLASS_TABLE_SIZE)
> 	{
> 	  next = class_table_array[enumerator->hash];
> 	  if (next != NULL)
> 	    {
> 	      enumerator->node = next;
> 	      return enumerator->node->pointer;
> 	    }
> 	  enumerator->hash++;
> 	}
>       
>       /* Ok - table finished - done */
>       objc_free (enumerator);
>       return Nil;
>     }
> }
> 
> #if 0 /* DEBUGGING FUNCTIONS */
> /* Debugging function - print the class table */
> void
> class_table_print ()
> {
>   int i;
>   
>   for (i = 0; i < CLASS_TABLE_SIZE; i++)
>     {
>       class_node_ptr node;
>       
>       printf ("%d:\n", i);
>       node = class_table_array[i];
>       
>       while (node != NULL)
> 	{
> 	  printf ("\t%s\n", node->name);
> 	  node = node->next;
> 	}
>     }
> }
> 
> /* Debugging function - print an histogram of number of classes in
>    function of hash key values.  Useful to evaluate the hash function
>    in real cases. */
> void
> class_table_print_histogram ()
> {
>   int i, j;
>   int counter = 0;
>   
>   for (i = 0; i < CLASS_TABLE_SIZE; i++)
>     {
>       class_node_ptr node;
>       
>       node = class_table_array[i];
>       
>       while (node != NULL)
> 	{
> 	  counter++;
> 	  node = node->next;
> 	}
>       if (((i + 1) % 50) == 0)
> 	{
> 	  printf ("%4d:", i + 1);
> 	  for (j = 0; j < counter; j++)
> 	    {
> 	      printf ("X");
> 	    }
> 	  printf ("\n");
> 	  counter = 0;
> 	}
>     }
>   printf ("%4d:", i + 1);
>   for (j = 0; j < counter; j++)
>     {
>       printf ("X");
>     }
>   printf ("\n");
> }
> #endif /* DEBUGGING FUNCTIONS */
> 
> /**
>  ** Objective-C runtime functions
>  **/
> 
> /* From now on, the only access to the class table data structure
>    should be via the class_table_* functions */
> 
> /* This is a hook which is called by objc_get_class and 
>    objc_lookup_class if the runtime is not able to find the class.
>    This may e.g. try to load in the class using dynamic loading */
> Class (*_objc_lookup_class)(const char* name) = 0;      /* !T:SAFE */
> 
> 
> /* True when class links has been resolved */     
> BOOL __objc_class_links_resolved = NO;                  /* !T:UNUSED */
> 
> 
> void __objc_init_class_tables()
> {
>   /* Allocate the class hash table */
>   
>   if(__class_table_lock)
>     return;
>   
>   objc_mutex_lock(__objc_runtime_mutex);
>   
>   class_table_setup ();
> 
>   objc_mutex_unlock(__objc_runtime_mutex);
> }  
> 
> /* This function adds a class to the class hash table, and assigns the 
>    class a number, unless it's already known */
> void
> __objc_add_class_to_hash(Class class)
> {
>   Class h_class;
> 
>   objc_mutex_lock(__objc_runtime_mutex);
> 
>   /* make sure the table is there */
>   assert(__class_table_lock);
> 
>   /* make sure it's not a meta class */
>   assert(CLS_ISCLASS(class));
> 
>   /* Check to see if the class is already in the hash table.  */
>   h_class = class_table_get_safe (class->name);
>   if (!h_class)
>     {
>       /* The class isn't in the hash table.  Add the class and assign a class
>          number.  */
>       static unsigned int class_number = 1;
> 
>       CLS_SETNUMBER(class, class_number);
>       CLS_SETNUMBER(class->class_pointer, class_number);
> 
>       ++class_number;
>       class_table_insert (class->name, class);
>     }
> 
>   objc_mutex_unlock(__objc_runtime_mutex);
> }
> 
> /* Get the class object for the class named NAME.  If NAME does not
>    identify a known class, the hook _objc_lookup_class is called.  If
>    this fails, nil is returned */
> Class objc_lookup_class (const char* name)
> {
>   Class class;
> 
>   class = class_table_get_safe (name);
> 
>   if (class)
>     return class;
> 
>   if (_objc_lookup_class)
>     return (*_objc_lookup_class)(name);
>   else
>     return 0;
> }
> 
> /* Get the class object for the class named NAME.  If NAME does not
>    identify a known class, the hook _objc_lookup_class is called.  If
>    this fails,  an error message is issued and the system aborts */
> Class
> objc_get_class (const char *name)
> {
>   Class class;
> 
>   class = class_table_get_safe (name);
> 
>   if (class)
>     return class;
> 
>   if (_objc_lookup_class)
>     class = (*_objc_lookup_class)(name);
> 
>   if(class)
>     return class;
>   
>   objc_error(nil, OBJC_ERR_BAD_CLASS, 
> 	     "objc runtime: cannot find class %s\n", name);
>   return 0;
> }
> 
> MetaClass
> objc_get_meta_class(const char *name)
> {
>   return objc_get_class(name)->class_pointer;
> }
> 
> /* This function provides a way to enumerate all the classes in the
>    executable.  Pass *ENUM_STATE == NULL to start the enumeration.  The
>    function will return 0 when there are no more classes.  
>    For example: 
>        id class; 
>        void *es = NULL;
>        while ((class = objc_next_class(&es)))
>          ... do something with class; 
> */
> Class
> objc_next_class(void **enum_state)
> {
>   Class class;
> 
>   objc_mutex_lock(__objc_runtime_mutex);
>   
>   /* make sure the table is there */
>   assert(__class_table_lock);
> 
>   class = class_table_next ((struct class_table_enumerator **)enum_state);
> 
>   objc_mutex_unlock(__objc_runtime_mutex);
>   
>   return class;
> }
> 
> /* Resolve super/subclass links for all classes.  The only thing we 
>    can be sure of is that the class_pointer for class objects point 
>    to the right meta class objects */
> void __objc_resolve_class_links()
> {
>   struct class_table_enumerator *es = NULL;
>   Class object_class = objc_get_class ("Object");
>   Class class1;
> 
>   assert(object_class);
> 
>   objc_mutex_lock(__objc_runtime_mutex);
> 
>   /* Assign subclass links */
>   while ((class1 = class_table_next (&es)))
>     {
>       /* Make sure we have what we think we have.  */
>       assert (CLS_ISCLASS(class1));
>       assert (CLS_ISMETA(class1->class_pointer));
> 
>       /* The class_pointer of all meta classes point to Object's meta class. */
>       class1->class_pointer->class_pointer = object_class->class_pointer;
> 
>       if (!(CLS_ISRESOLV(class1)))
>         {
>           CLS_SETRESOLV(class1);
>           CLS_SETRESOLV(class1->class_pointer);
>               
>           if(class1->super_class)
>             {   
>               Class a_super_class 
>                 = objc_get_class ((char *) class1->super_class);
>               
>               assert (a_super_class);
>               
>               DEBUG_PRINTF ("making class connections for: %s\n",
>                             class1->name);
>               
>               /* assign subclass links for superclass */
>               class1->sibling_class = a_super_class->subclass_list;
>               a_super_class->subclass_list = class1;
>               
>               /* Assign subclass links for meta class of superclass */
>               if (a_super_class->class_pointer)
>                 {
>                   class1->class_pointer->sibling_class
>                     = a_super_class->class_pointer->subclass_list;
>                   a_super_class->class_pointer->subclass_list 
>                     = class1->class_pointer;
>                 }
>             }
>           else                  /* a root class, make its meta object */
>                                 /* be a subclass of Object */
>             {
>               class1->class_pointer->sibling_class 
>                 = object_class->subclass_list;
>               object_class->subclass_list = class1->class_pointer;
>             }
>         }
>     }
> 
>   /* Assign superclass links */
>    es = NULL;
>    while ((class1 = class_table_next (&es)))
>     {
>       Class sub_class;
>       for (sub_class = class1->subclass_list; sub_class;
>            sub_class = sub_class->sibling_class)
>         {
>           sub_class->super_class = class1;
>           if(CLS_ISCLASS(sub_class))
>             sub_class->class_pointer->super_class = class1->class_pointer;
>         }
>     }
> 
>   objc_mutex_unlock(__objc_runtime_mutex);
> }
> 
> 
> 
> #define CLASSOF(c) ((c)->class_pointer)
> 
> Class
> class_pose_as (Class impostor, Class super_class)
> {
>   if (!CLS_ISRESOLV (impostor))
>     __objc_resolve_class_links ();
> 
>   /* preconditions */
>   assert (impostor);
>   assert (super_class);
>   assert (impostor->super_class == super_class);
>   assert (CLS_ISCLASS (impostor));
>   assert (CLS_ISCLASS (super_class));
>   assert (impostor->instance_size == super_class->instance_size);
> 
>   {
>     Class *subclass = &(super_class->subclass_list);
> 
>     /* move subclasses of super_class to impostor */
>     while (*subclass)
>       {
> 	Class nextSub = (*subclass)->sibling_class;
> 
> 	if (*subclass != impostor)
> 	  {
> 	    Class sub = *subclass;
> 
> 	    /* classes */
> 	    sub->sibling_class = impostor->subclass_list;
> 	    sub->super_class = impostor;
> 	    impostor->subclass_list = sub;
> 
> 	    /* It will happen that SUB is not a class object if it is 
> 	       the top of the meta class hierarchy chain.  (root
> 	       meta-class objects inherit their class object)  If that is
> 	       the case... don't mess with the meta-meta class. */ 
> 	    if (CLS_ISCLASS (sub))
> 	      {
> 		/* meta classes */
> 		CLASSOF (sub)->sibling_class = 
> 		  CLASSOF (impostor)->subclass_list;
> 		CLASSOF (sub)->super_class = CLASSOF (impostor);
> 		CLASSOF (impostor)->subclass_list = CLASSOF (sub);
> 	      }
> 	  }
> 
> 	*subclass = nextSub;
>       }
> 
>     /* set subclasses of superclass to be impostor only */
>     super_class->subclass_list = impostor;
>     CLASSOF (super_class)->subclass_list = CLASSOF (impostor);
>     
>     /* set impostor to have no sibling classes */
>     impostor->sibling_class = 0;
>     CLASSOF (impostor)->sibling_class = 0;
>   }
>   
>   /* check relationship of impostor and super_class is kept. */
>   assert (impostor->super_class == super_class);
>   assert (CLASSOF (impostor)->super_class == CLASSOF (super_class));
> 
>   /* This is how to update the lookup table. Regardless of
>      what the keys of the hashtable is, change all values that are
>      superclass into impostor. */
> 
>   objc_mutex_lock(__objc_runtime_mutex);
> 
>   class_table_replace (super_class, impostor);
> 
>   objc_mutex_unlock(__objc_runtime_mutex);
> 
>   /* next, we update the dispatch tables... */
>   __objc_update_dispatch_table_for_class (CLASSOF (impostor));
>   __objc_update_dispatch_table_for_class (impostor);
> 
>   return impostor;
> }
>   
> 
> 


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]