--- deliantra/server/common/shstr.C	2006/09/03 00:18:40	1.3
+++ deliantra/server/common/shstr.C	2006/09/10 16:00:23	1.13
@@ -1,3 +1,4 @@
+
 /*
  * shstr.C
  */
@@ -5,105 +6,130 @@
 #include <cstring>
 #include <cstdlib>
 
+#include <glib.h>
+
 #include <tr1/unordered_set>
 
 #include "shstr.h"
+#include "util.h"
 
-struct hash
-{
-  std::size_t operator ()(const char *s) const
-  {
-    unsigned long hash = 0;
-    unsigned int i = 0;
-
-    /* use the one-at-a-time hash function, which supposedly is
-     * better than the djb2-like one used by perl5.005, but
-     * certainly is better then the bug used here before.
-     * see http://burtleburtle.net/bob/hash/doobs.html
-     */
-    while (*s)
-      {
-        hash += *s++;
-        hash += hash << 10;
-        hash ^= hash >>  6;
-      }
-
-    hash += hash <<  3;
-    hash ^= hash >> 11;
-    hash += hash << 15;
-
-    return hash;
-  }
-};
+typedef
+std::tr1::unordered_set < const char *,
+  str_hash,
+  str_equal >
+  HT;
 
-struct equal
+static HT
+  ht;
+
+static const char *
+makevec (const char *s)
 {
-  bool operator ()(const char *a, const char *b) const
-  {
-    return !strcmp (a, b);
-  }
-};
+  int
+    len = strlen (s);
+
+  const char *
+    v = (const char *) (2 + (int *) g_slice_alloc (sizeof (int) * 2 + len + 1));
+
+  shstr::length (v) = len;
+  shstr::refcnt (v) = 1;
 
-typedef std::tr1::unordered_set<const char *, hash, equal> HT;
+  memcpy ((char *) v, s, len + 1);
 
-static HT ht;
+  return v;
+}
+
+const char *
+  shstr::null = makevec ("<nil>");
+
+// what weird misoptimisation is this again?
+const shstr undead_name ("undead");
 
 const char *
 shstr::find (const char *s)
 {
+  if (!s)
+    return s;
+
   HT::iterator i = ht.find (s);
 
-  return i != ht.end ()
-    ? (char *)*i
-    : 0;
+  return i != ht.end ()? *i : 0;
 }
 
 const char *
 shstr::intern (const char *s)
 {
-  HT::iterator i = ht.find (s);
+  if (!s)
+    return null;
 
-  if (i != ht.end ())
-    return (char *)*i;
-
-  int len = strlen (s);
-
-  int *v = (int *)malloc (sizeof (int) * 2 + len + 1);
+  if (const char *found = find (s))
+    {
+      ++refcnt (found);
+      return found;
+    }
+
+  s = makevec (s);
+  ht.insert (s);
+  return s;
+}
 
-  v [0] = len;
-  v [1] = 0;
+// periodically test refcounts == 0 for a few strings
+// this is the ONLY thing that erases stuff from ht. keep it that way.
+void
+shstr::gc ()
+{
+  static const char *curpos;
 
-  v += 2;
+  HT::iterator i = curpos ? ht.find (curpos) : ht.begin ();
 
-  memcpy (v, s, len + 1);
+  if (i == ht.end ())
+    i = ht.begin ();
 
-  ht.insert ((char *)v);
+  // go through all strings roughly once every 4 minutes
+  int n = ht.size () / 256 + 16;
+
+  for (;;)
+    {
+      if (i == ht.end ())
+        {
+          curpos = 0;
+          return;
+        }
+      else if (!--n)
+        break;
+      else if (!refcnt (*i))
+        {
+          HT::iterator o = i++;
+          const char *s = *o;
+
+          ht.erase (o);
+
+          //printf ("GC %4d %3d %d >%s<%d\n", (int)ht.size (), n, shstr::refcnt (s), s, shstr::length (s));
+          g_slice_free1 (sizeof (int) * 2 + length (s) + 1, -2 + (int *) s);
+        }
+      else
+        ++i;
+    }
 
-  return (char *)v;
+  curpos = *i;
 }
 
-// TODO: periodically test refcounts == 0 for a few strings (e.g. one hash bucket,
-// exploiting the fatc that iterators stay valid for unordered_set).
-void
-shstr::gc ()
-{
-}
+//TODO: this should of course not be here
 
 /* buf_overflow() - we don't want to exceed the buffer size of 
  * buf1 by adding on buf2! Returns true if overflow will occur.
  */
 
-int 
+int
 buf_overflow (const char *buf1, const char *buf2, int bufsize)
 {
-    int len1 = 0, len2 = 0;
+  int len1 = 0, len2 = 0;
 
-    if (buf1)
-        len1 = strlen (buf1);
-    if (buf2)
-        len2 = strlen (buf2);
-    if ((len1 + len2) >= bufsize)
-        return 1;
-    return 0;
+  if (buf1)
+    len1 = strlen (buf1);
+  if (buf2)
+    len2 = strlen (buf2);
+  if ((len1 + len2) >= bufsize)
+    return 1;
+  return 0;
 }
-