Reduce size of heap block per-granule maps twice for 4KB heap blocks

If MAXOBJGRANULES-1 fits into a byte, we could use unsigned char
(instead of short) type for hb_map elements.

* docs/tree.md (A picture): Specify that element of GC_obj_map is of
either 8 or 16 bits (instead of 16 bits strictly); specify that hb_map
field of hblkhdr could be of uchar* type.
* include/private/gc_priv.h [!MARK_BIT_PER_OBJ] (hb_map_entry_t): New
type (defined depending on values of CPP_LOG_HBLKSIZE and
GC_GRANULE_BYTES); add comment.
* include/private/gc_priv.h [!MARK_BIT_PER_OBJ] (hblkhdr.hb_map,
_GC_arrays._obj_map[]): Change type from unsigned short* to
hb_map_entry_t*.
* obj_map.c [!MARK_BIT_PER_OBJ] (GC_add_map_entry): Change type of
new_map local variable from unsigned short* to hb_map_entry_t*; update
relevant casts and GC_scratch_alloc() argument accordingly; replace
0xffff to ~(hb_map_entry_t)0 in static assertion; expect GC_obj_map[lg]
is already non-NULL on entrance.

docs/tree.md

@@ -124,7 +124,7 @@ contributed originally by Dave Barrett.
            |                                         |   |<--OBJ_MAP_LEN-->|
            |                                         |   =HBLKSIZE/GC_GRANULE_BYTES
      HDR(p)| GC_find_header(p)                       |    (1024 elements on Alpha)
-           |                           \ from        |    (16 bits each)
+           |                           \ from        |    (8/16 bits each)
            |    (hdr) (struct hblkhdr) / alloc_hdr() |
            +--->+----------------------+             |
       GET_HDR(p)| struct hblk *hb_next |             |
@@ -141,7 +141,7 @@ contributed originally by Dave Barrett.
                 +----------------------+             |
                 | word   hb_descr      |             |
                 +----------------------+             |
-                | ushort *hb_map       |>------------+
+                | uchar/ushort *hb_map |>------------+
                 +----------------------+
                 | AO_t   hb_n_marks    |
        ---      +----------------------+

include/private/gc_priv.h

@@ -1088,6 +1088,17 @@ typedef word page_hash_table[PHT_SIZE];
                 /* initial group of mark bits, and it is safe to        */
                 /* allocate smaller header for large objects.           */
 
+#ifndef MARK_BIT_PER_OBJ
+  /* We maintain layout maps for heap blocks containing objects of      */
+  /* a given size.  Each entry in this map describes a byte offset      */
+  /* (displacement) and has the following type.                         */
+# if (1 << (CPP_LOG_HBLKSIZE - 1)) / GC_GRANULE_BYTES <= 0x100
+    typedef unsigned char hb_map_entry_t;
+# else
+    typedef unsigned short hb_map_entry_t;
+# endif
+#endif /* !MARK_BIT_PER_OBJ */
+
 struct hblkhdr {
     struct hblk * hb_next;      /* Link field for hblk free list         */
                                 /* and for lists of chunks waiting to be */
@@ -1143,7 +1154,7 @@ struct hblkhdr {
     word hb_descr;              /* Object descriptor for marking.       */
                                 /* See gc_mark.h.                       */
 #   ifndef MARK_BIT_PER_OBJ
-      unsigned short * hb_map;  /* Essentially a table of remainders    */
+      hb_map_entry_t *hb_map;   /* Essentially a table of remainders    */
                                 /* mod BYTES_TO_GRANULES(hb_sz), except */
                                 /* for large blocks.  See GC_obj_map.   */
 #   endif
@@ -1619,7 +1630,7 @@ struct _GC_arrays {
         /* the allocator lock held.                                     */
 # ifndef MARK_BIT_PER_OBJ
 #   define GC_obj_map GC_arrays._obj_map
-    unsigned short * _obj_map[MAXOBJGRANULES+1];
+    hb_map_entry_t *_obj_map[MAXOBJGRANULES+1];
                        /* If not NULL, then a pointer to a map of valid */
                        /* object addresses.  GC_obj_map[lg][i] is       */
                        /* i % lg.  This is now used purely to replace   */

obj_map.c

@@ -47,14 +47,15 @@ GC_INNER void GC_register_displacement_inner(size_t offset)
   GC_INNER GC_bool GC_add_map_entry(size_t lg)
   {
     size_t displ;
-    unsigned short * new_map;
+    hb_map_entry_t *new_map;
 
     GC_ASSERT(I_HOLD_LOCK());
-    GC_STATIC_ASSERT(MAXOBJGRANULES - 1U <= 0xffffU /* max ushort */);
+    GC_STATIC_ASSERT(MAXOBJGRANULES - 1 <= (size_t)(~(hb_map_entry_t)0));
     if (lg > MAXOBJGRANULES) lg = 0;
-    if (GC_obj_map[lg] != 0) return TRUE;
+    if (EXPECT(GC_obj_map[lg] != NULL, TRUE)) return TRUE;
 
-    new_map = (unsigned short *)GC_scratch_alloc(OBJ_MAP_LEN * sizeof(short));
+    new_map = (hb_map_entry_t *)GC_scratch_alloc(OBJ_MAP_LEN
+                                                 * sizeof(hb_map_entry_t));
     if (EXPECT(NULL == new_map, FALSE)) return FALSE;
 
     GC_COND_LOG_PRINTF(
@@ -66,7 +67,7 @@ GC_INNER void GC_register_displacement_inner(size_t offset)
       }
     } else {
       for (displ = 0; displ < OBJ_MAP_LEN; displ++) {
-        new_map[displ] = (unsigned short)(displ % lg);
+        new_map[displ] = (hb_map_entry_t)(displ % lg);
       }
     }
     GC_obj_map[lg] = new_map;