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 #include "lib/sorter/array.h"
24 * The big_buf has the following layout:
26 * +-------------------------------------------------------------------------------+
27 * | array of internal_item's |
28 * +-------------------------------------------------------------------------------+
29 * | padding to make the following part page-aligned |
30 * +--------------------------------+----------------------------------------------+
31 * | shadow copy of item array | array of pointers to data for write_merged() |
32 * | used if radix-sorting +----------------------------------------------+
33 * | | workspace for write_merged() |
34 * +--------------------------------+----------------------------------------------+
38 * | sequence of | padding | |
39 * | items +---------+ |
44 * +-------------------------------------------------------------------------------+
46 * (the data which are in different columns are never accessed simultaneously,
47 * so we use a single buffer for both)
50 static inline void *P(internal_get_data)(P(key) *key)
52 uns ksize = SORT_KEY_SIZE(*key);
54 ksize = ALIGN_TO(ksize, CPU_STRUCT_ALIGN);
56 return (byte *) key + ksize;
59 static inline size_t P(internal_workspace)(P(key) *key UNUSED)
65 #ifdef SORT_UNIFY_WORKSPACE
66 ws += SORT_UNIFY_WORKSPACE(*key);
69 ws = MAX(ws, sizeof(P(internal_item_t)));
74 static int P(internal)(struct sort_context *ctx, struct sort_bucket *bin, struct sort_bucket *bout, struct sort_bucket *bout_only)
76 sorter_alloc_buf(ctx);
77 struct fastbuf *in = sbuck_read(bin);
79 P(key) key, *keybuf = ctx->key_buf;
81 keybuf = ctx->key_buf = sorter_alloc(ctx, sizeof(key));
87 else if (!P(read_key)(in, &key))
90 size_t bufsize = ctx->big_buf_size;
92 if (sizeof(key) + 2*CPU_PAGE_SIZE + SORT_DATA_SIZE(key) + P(internal_workspace)(&key) > bufsize)
94 SORT_XTRACE(3, "s-internal: Generating a giant run");
95 struct fastbuf *out = sbuck_write(bout);
96 P(copy_data)(&key, in, out);
98 return 1; // We don't know, but 1 is always safe
102 SORT_XTRACE(4, "s-internal: Reading");
103 P(internal_item_t) *item_array = ctx->big_buf, *item = item_array, *last_item;
104 byte *end = (byte *) ctx->big_buf + bufsize;
105 size_t remains = bufsize - CPU_PAGE_SIZE;
108 uns ksize = SORT_KEY_SIZE(key);
110 uns ksize_aligned = ALIGN_TO(ksize, CPU_STRUCT_ALIGN);
112 uns ksize_aligned = ksize;
114 uns dsize = SORT_DATA_SIZE(key);
115 uns recsize = ALIGN_TO(ksize_aligned + dsize, CPU_STRUCT_ALIGN);
116 size_t totalsize = recsize + sizeof(P(internal_item_t) *) + P(internal_workspace)(&key);
117 if (unlikely(totalsize > remains
118 #ifdef CPU_64BIT_POINTERS
119 || item >= item_array + ~0U // The number of items must fit in an uns
127 remains -= totalsize;
129 memcpy(end, &key, ksize);
131 breadb(in, end + ksize_aligned, dsize);
133 item->key = (P(key)*) end;
136 while (P(read_key)(in, &key));
139 uns count = last_item - item_array;
140 void *workspace UNUSED = ALIGN_PTR(last_item, CPU_PAGE_SIZE);
141 SORT_XTRACE(3, "s-internal: Read %u items (%s items, %s workspace, %s data)",
143 stk_fsize((byte*)last_item - (byte*)item_array),
144 stk_fsize(end - (byte*)last_item - remains),
145 stk_fsize((byte*)ctx->big_buf + bufsize - end));
148 item_array = P(array_sort)(item_array, count,
149 #ifdef SORT_HASH_BITS
150 workspace, bin->hash_bits
155 ctx->total_int_time += get_timer(&timer);
157 SORT_XTRACE(4, "s-internal: Writing");
160 struct fastbuf *out = sbuck_write(bout);
162 uns merged UNUSED = 0;
163 for (item = item_array; item < last_item; item++)
166 if (item < last_item - 1 && !P(compare)(item->key, item[1].key))
168 // Rewrite the item structures with just pointers to keys and place
169 // pointers to data in the workspace.
170 P(key) **key_array = (void *) item;
171 void **data_array = workspace;
172 key_array[0] = item[0].key;
173 data_array[0] = P(internal_get_data)(key_array[0]);
175 for (cnt=1; item+cnt < last_item && !P(compare)(key_array[0], item[cnt].key); cnt++)
177 key_array[cnt] = item[cnt].key;
178 data_array[cnt] = P(internal_get_data)(key_array[cnt]);
180 P(write_merged)(out, key_array, data_array, cnt, data_array+cnt);
186 #ifdef SORT_ASSERT_UNIQUE
187 ASSERT(item == last_item-1 || P(compare)(item->key, item[1].key) < 0);
189 P(write_key)(out, item->key);
191 bwrite(out, P(internal_get_data)(item->key), SORT_DATA_SIZE(*item->key));
195 SORT_XTRACE(3, "Merging reduced %u records", merged);
198 return ctx->more_keys;
202 P(internal_estimate)(struct sort_context *ctx, struct sort_bucket *b UNUSED)
205 uns avg = ALIGN_TO(sizeof(P(key))/4, CPU_STRUCT_ALIGN); // Wild guess...
207 uns avg = ALIGN_TO(sizeof(P(key)), CPU_STRUCT_ALIGN);
209 // We ignore the data part of records, it probably won't make the estimate much worse
210 size_t bufsize = ctx->big_buf_size;
211 #ifdef SORT_UNIFY_WORKSPACE // FIXME: Or if radix-sorting
214 return (bufsize / (avg + sizeof(P(internal_item_t))) * avg);