2 * UCW Library -- Universal Sorter: Internal Sorting Module
4 * (c) 2007 Martin Mares <mj@ucw.cz>
6 * This software may be freely distributed and used according to the terms
7 * of the GNU Lesser General Public License.
10 #include "lib/stkstring.h"
14 // FIXME: Add the hash here to save cache misses
17 #define ASORT_PREFIX(x) SORT_PREFIX(array_##x)
18 #define ASORT_KEY_TYPE P(internal_item_t)
19 #define ASORT_LT(x,y) (P(compare)((x).key, (y).key) < 0)
20 #define ASORT_PAGE_ALIGNED
21 #ifdef SORT_INTERNAL_RADIX
22 # define ASORT_HASH(x) P(hash)((x).key)
23 # ifdef SORT_LONG_HASH
24 # define ASORT_LONG_HASH
27 #include "lib/sorter/array.h"
30 * The big_buf has the following layout:
32 * +-------------------------------------------------------------------------------+
33 * | array of internal_item's |
34 * +-------------------------------------------------------------------------------+
35 * | padding to make the following part page-aligned |
36 * +--------------------------------+----------------------------------------------+
37 * | shadow copy of item array | array of pointers to data for write_merged() |
38 * | used if radix-sorting +----------------------------------------------+
39 * | | workspace for write_merged() |
40 * +--------------------------------+----------------------------------------------+
44 * | sequence of | padding | |
45 * | items +---------+ |
50 * +-------------------------------------------------------------------------------+
52 * (the data which are in different columns are never accessed simultaneously,
53 * so we use a single buffer for both)
56 static inline void *P(internal_get_data)(P(key) *key)
58 uns ksize = SORT_KEY_SIZE(*key);
60 ksize = ALIGN_TO(ksize, CPU_STRUCT_ALIGN);
62 return (byte *) key + ksize;
65 static inline size_t P(internal_workspace)(P(key) *key UNUSED)
71 #ifdef SORT_UNIFY_WORKSPACE
72 ws += SORT_UNIFY_WORKSPACE(*key);
74 #ifdef SORT_INTERNAL_RADIX
75 ws = MAX(ws, sizeof(P(internal_item_t)));
80 static int P(internal)(struct sort_context *ctx, struct sort_bucket *bin, struct sort_bucket *bout, struct sort_bucket *bout_only)
82 sorter_alloc_buf(ctx);
83 struct fastbuf *in = sbuck_read(bin);
85 P(key) key, *keybuf = ctx->key_buf;
87 keybuf = ctx->key_buf = sorter_alloc(ctx, sizeof(key));
93 else if (!P(read_key)(in, &key))
96 size_t bufsize = ctx->big_buf_size;
98 if (sizeof(key) + 2*CPU_PAGE_SIZE + SORT_DATA_SIZE(key) + P(internal_workspace)(&key) > bufsize)
100 SORT_XTRACE(3, "s-internal: Generating a giant run");
101 struct fastbuf *out = sbuck_write(bout);
102 P(copy_data)(&key, in, out);
104 return 1; // We don't know, but 1 is always safe
108 SORT_XTRACE(4, "s-internal: Reading");
109 P(internal_item_t) *item_array = ctx->big_buf, *item = item_array, *last_item;
110 byte *end = (byte *) ctx->big_buf + bufsize;
111 size_t remains = bufsize - CPU_PAGE_SIZE;
114 uns ksize = SORT_KEY_SIZE(key);
116 uns ksize_aligned = ALIGN_TO(ksize, CPU_STRUCT_ALIGN);
118 uns ksize_aligned = ksize;
120 uns dsize = SORT_DATA_SIZE(key);
121 uns recsize = ALIGN_TO(ksize_aligned + dsize, CPU_STRUCT_ALIGN);
122 size_t totalsize = recsize + sizeof(P(internal_item_t) *) + P(internal_workspace)(&key);
123 if (unlikely(totalsize > remains
124 #ifdef CPU_64BIT_POINTERS
125 || item >= item_array + ~0U // The number of items must fit in an uns
133 remains -= totalsize;
135 memcpy(end, &key, ksize);
137 breadb(in, end + ksize_aligned, dsize);
139 item->key = (P(key)*) end;
142 while (P(read_key)(in, &key));
145 uns count = last_item - item_array;
146 void *workspace UNUSED = ALIGN_PTR(last_item, CPU_PAGE_SIZE);
147 SORT_XTRACE(3, "s-internal: Read %u items (%s items, %s workspace, %s data)",
149 stk_fsize((byte*)last_item - (byte*)item_array),
150 stk_fsize(end - (byte*)last_item - remains),
151 stk_fsize((byte*)ctx->big_buf + bufsize - end));
154 item_array = P(array_sort)(item_array, count,
155 #ifdef SORT_INTERNAL_RADIX
156 workspace, bin->hash_bits
161 ctx->total_int_time += get_timer(&timer);
163 SORT_XTRACE(4, "s-internal: Writing");
166 struct fastbuf *out = sbuck_write(bout);
168 uns merged UNUSED = 0;
169 for (item = item_array; item < last_item; item++)
172 if (item < last_item - 1 && !P(compare)(item->key, item[1].key))
174 // Rewrite the item structures with just pointers to keys and place
175 // pointers to data in the workspace.
176 P(key) **key_array = (void *) item;
177 void **data_array = workspace;
178 key_array[0] = item[0].key;
179 data_array[0] = P(internal_get_data)(key_array[0]);
181 for (cnt=1; item+cnt < last_item && !P(compare)(key_array[0], item[cnt].key); cnt++)
183 key_array[cnt] = item[cnt].key;
184 data_array[cnt] = P(internal_get_data)(key_array[cnt]);
186 P(write_merged)(out, key_array, data_array, cnt, data_array+cnt);
192 #ifdef SORT_ASSERT_UNIQUE
193 ASSERT(item == last_item-1 || P(compare)(item->key, item[1].key) < 0);
195 P(write_key)(out, item->key);
197 bwrite(out, P(internal_get_data)(item->key), SORT_DATA_SIZE(*item->key));
201 SORT_XTRACE(3, "Merging reduced %u records", merged);
204 return ctx->more_keys;
208 P(internal_estimate)(struct sort_context *ctx, struct sort_bucket *b UNUSED)
211 uns avg = ALIGN_TO(sizeof(P(key))/4, CPU_STRUCT_ALIGN); // Wild guess...
213 uns avg = ALIGN_TO(sizeof(P(key)), CPU_STRUCT_ALIGN);
215 // We ignore the data part of records, it probably won't make the estimate much worse
216 size_t bufsize = ctx->big_buf_size;
217 #ifdef SORT_UNIFY_WORKSPACE // FIXME: Or if radix-sorting
220 return (bufsize / (avg + sizeof(P(internal_item_t))) * avg);