Added "bulk operations" to arrays

A few operations on arrays can be performed "in bulk", treating all
tags of a cell as a simple (or a few) word(s).

lgc.c

@@ -465,6 +465,46 @@ static void traverseweakvalue (global_State *g, Table *h) {
 }
 
 
+#define BK2(x)	cast(lua_Unsigned, ((x) << 8) | BIT_ISCOLLECTABLE)
+/*
+** Check whether some value in the cell starting at index 'i'
+** is collectable
+*/
+static int checkBulkCollectable (Table *h, unsigned i) {
+  const lua_Unsigned bitscoll = BK2(BK2(BK2(BK2(BK2(BK2(BK2(BK2(~0u))))))));
+  int j;
+  i /= NM;
+  for (j = 0; j < BKSZ; j++) {
+    if (h->array[i].u.bulk[j] & bitscoll)
+      return 1;
+  }
+  return 0;
+}
+
+
+/*
+** Traverse the array part of a table. The traversal is made by cells,
+** only traversing a cell if it has some collectable tag among its tags.
+*/
+static int traversearray (global_State *g, Table *h) {
+  unsigned asize = luaH_realasize(h);
+  int marked = 0;  /* true if some object is marked in this traversal */
+  unsigned i;
+  for (i = 0; i < asize; i += NM) {  /* traverse array in cells */
+    if (checkBulkCollectable(h, i)) {  /* something to mark in this cell? */
+      unsigned j;
+      for (j = 0; j < NM && i + j < asize; j++) {
+        GCObject *o = gcvalarr(h, i + j);
+        if (o != NULL && iswhite(o)) {
+          marked = 1;
+          reallymarkobject(g, o);
+        }
+      }
+    }
+  }
+  return marked;
+}
+
 /*
 ** Traverse an ephemeron table and link it to proper list. Returns true
 ** iff any object was marked during this traversal (which implies that
@@ -478,20 +518,11 @@ static void traverseweakvalue (global_State *g, Table *h) {
 ** by 'genlink'.
 */
 static int traverseephemeron (global_State *g, Table *h, int inv) {
-  int marked = 0;  /* true if an object is marked in this traversal */
   int hasclears = 0;  /* true if table has white keys */
   int hasww = 0;  /* true if table has entry "white-key -> white-value" */
   unsigned int i;
-  unsigned int asize = luaH_realasize(h);
   unsigned int nsize = sizenode(h);
-  /* traverse array part */
-  for (i = 0; i < asize; i++) {
-    GCObject *o = gcvalarr(h, i);
-    if (o != NULL && iswhite(o)) {
-      marked = 1;
-      reallymarkobject(g, o);
-    }
-  }
+  int marked = traversearray(g, h);  /* traverse array part */
   /* traverse hash part; if 'inv', traverse descending
      (see 'convergeephemerons') */
   for (i = 0; i < nsize; i++) {
@@ -523,13 +554,7 @@ static int traverseephemeron (global_State *g, Table *h, int inv) {
 
 static void traversestrongtable (global_State *g, Table *h) {
   Node *n, *limit = gnodelast(h);
-  unsigned int i;
-  unsigned int asize = luaH_realasize(h);
-  for (i = 0; i < asize; i++) {  /* traverse array part */
-    GCObject *o = gcvalarr(h, i);
-    if (o != NULL && iswhite(o))
-      reallymarkobject(g, o);
-  }
+  traversearray(g, h);
   for (n = gnode(h, 0); n < limit; n++) {  /* traverse hash part */
     if (isempty(gval(n)))  /* entry is empty? */
       clearkey(n);  /* clear its key */

ltable.c

@@ -653,6 +653,44 @@ static void exchangehashpart (Table *t1, Table *t2) {
 }
 
 
+/*
+** Re-insert into the new hash part of a table the elements from the
+** vanishing slice of the array part.
+*/
+static void reinsertOldSlice (lua_State *L, Table *t, unsigned oldasize,
+                                            unsigned newasize) {
+  unsigned i;
+  t->alimit = newasize;  /* pretend array has new size... */
+  for (i = newasize; i < oldasize; i++) {  /* traverse vanishing slice */
+    int tag = *getArrTag(t, i);
+    if (!tagisempty(tag)) {  /* a non-empty entry? */
+      TValue aux;
+      farr2val(t, i + 1, tag, &aux);
+      luaH_setint(L, t, i + 1, &aux);  /* re-insert it into the table */
+    }
+  }
+  t->alimit = oldasize;  /* restore current size... */
+}
+
+
+#define BK1(x)	cast(lua_Unsigned, ((x) << 8) | LUA_VEMPTY)
+
+/*
+** Clear new slice of the array, in bulk.
+*/
+static void clearNewSlice (Table *t, unsigned oldasize, unsigned newasize) {
+  int i, j;
+  int firstcell = (oldasize + NM - 1) / NM;
+  int lastcell = cast_int((newasize + NM - 1) / NM) - 1;
+  for (i = firstcell; i <= lastcell; i++) {
+    /* empty tag repeated for all tags in a word */
+    const lua_Unsigned empty = BK1(BK1(BK1(BK1(BK1(BK1(BK1(BK1(0))))))));
+    for (j = 0; j < BKSZ; j++)
+      t->array[i].u.bulk[j] = empty;
+  }
+}
+
+
 /*
 ** Resize table 't' for the new given sizes. Both allocations (for
 ** the hash part and for the array part) can fail, which creates some
@@ -668,7 +706,6 @@ static void exchangehashpart (Table *t1, Table *t2) {
 */
 void luaH_resize (lua_State *L, Table *t, unsigned int newasize,
                                           unsigned int nhsize) {
-  unsigned int i;
   Table newt;  /* to keep the new hash part */
   unsigned int oldasize = setlimittosize(t);
   ArrayCell *newarray;
@@ -678,19 +715,10 @@ void luaH_resize (lua_State *L, Table *t, unsigned int newasize,
   newt.flags = 0;
   setnodevector(L, &newt, nhsize);
   if (newasize < oldasize) {  /* will array shrink? */
-    t->alimit = newasize;  /* pretend array has new size... */
-    exchangehashpart(t, &newt);  /* and new hash */
     /* re-insert into the new hash the elements from vanishing slice */
-    for (i = newasize; i < oldasize; i++) {
-      int tag = *getArrTag(t, i);
-      if (!tagisempty(tag)) {  /* a non-empty entry? */
-        TValue aux;
-        farr2val(t, i + 1, tag, &aux);
-        luaH_setint(L, t, i + 1, &aux);
-      }
-    }
-    t->alimit = oldasize;  /* restore current size... */
-    exchangehashpart(t, &newt);  /* and hash (in case of errors) */
+    exchangehashpart(t, &newt);  /* pretend table has new hash */
+    reinsertOldSlice(L, t, oldasize, newasize);
+    exchangehashpart(t, &newt);  /* restore old hash (in case of errors) */
   }
   /* allocate new array */
   newarray = resizearray(L, t, oldasize, newasize);
@@ -702,8 +730,7 @@ void luaH_resize (lua_State *L, Table *t, unsigned int newasize,
   exchangehashpart(t, &newt);  /* 't' has the new hash ('newt' has the old) */
   t->array = newarray;  /* set new array part */
   t->alimit = newasize;
-  for (i = oldasize; i < newasize; i++)  /* clear new slice of the array */
-    *getArrTag(t, i) = LUA_VEMPTY;
+  clearNewSlice(t, oldasize, newasize);
   /* re-insert elements from old hash part into new parts */
   reinsert(L, &newt, t);  /* 'newt' now has the old hash */
   freehash(L, &newt);  /* free old hash part */

ltable.h

@@ -87,20 +87,32 @@
 
 
 /*
-** The array part of a table is represented by an array of cells.
+** The array part of a table is represented by an array of *cells*.
 ** Each cell is composed of NM tags followed by NM values, so that
 ** no space is wasted in padding.
 */
 #define NM      cast_uint(sizeof(Value))
 
+
+/*
+** A few operations on arrays can be performed "in bulk", treating all
+** tags of a cell as a simple (or a few) word(s). The next constant is
+** the number of words to cover the tags of a cell. (In conventional
+** architectures that will be 1 or 2.)
+*/
+#define BKSZ	cast_int((NM - 1) / sizeof(lua_Unsigned) + 1)
+
 struct ArrayCell {
-  lu_byte tag[NM];
+  union {
+    lua_Unsigned bulk[BKSZ];  /* for "bulk" operations */
+    lu_byte tag[NM];
+  } u;
   Value value[NM];
 };
 
 
 /* Computes the address of the tag for the abstract index 'k' */
-#define getArrTag(t,k)	(&(t)->array[(k)/NM].tag[(k)%NM])
+#define getArrTag(t,k)	(&(t)->array[(k)/NM].u.tag[(k)%NM])
 
 /* Computes the address of the value for the abstract index 'k' */
 #define getArrVal(t,k)	(&(t)->array[(k)/NM].value[(k)%NM])