X-Git-Url: http://mj.ucw.cz/gitweb/?a=blobdiff_plain;ds=inline;f=lib%2Fsorter%2Fs-internal.h;h=ef2da2402958cfa62bfdac3a69d575390ce7e0a8;hb=7cf300f543b5023ad46a909ed6fde0feba5d1acd;hp=c05ca077e574868bb7e27cd94c85ddc6cf24ee10;hpb=c15baf7b8cf0dc6a07b3c57ad346aee68ac6f19f;p=libucw.git

diff --git a/lib/sorter/s-internal.h b/lib/sorter/s-internal.h
index c05ca077..ef2da240 100644
--- a/lib/sorter/s-internal.h
+++ b/lib/sorter/s-internal.h
@@ -7,23 +7,98 @@
  *	of the GNU Lesser General Public License.
  */
 
+#include "lib/stkstring.h"
+
+#ifdef SORT_INTERNAL_RADIX
+/* Keep copies of the items' hashes to save cache misses */
+#define SORT_COPY_HASH
+#endif
+
 typedef struct {
   P(key) *key;
-  // FIXME: Add the hash here to save cache misses
+#ifdef SORT_COPY_HASH
+  P(hash_t) hash;
+#endif
 } P(internal_item_t);
 
 #define ASORT_PREFIX(x) SORT_PREFIX(array_##x)
 #define ASORT_KEY_TYPE P(internal_item_t)
-#define ASORT_ELT(i) ary[i]
-#define ASORT_LT(x,y) (P(compare)((x).key, (y).key) < 0)
-#define ASORT_EXTRA_ARGS , P(internal_item_t) *ary
-#include "lib/arraysort.h"
+#ifdef SORT_COPY_HASH
+#  ifdef SORT_INT
+#    define ASORT_LT(x,y) ((x).hash < (y).hash)		// In this mode, the hash is the value
+#  else
+#    define ASORT_LT(x,y) ((x).hash < (y).hash || (x).hash == (y).hash && P(compare)((x).key, (y).key) < 0)
+#  endif
+#else
+#  define ASORT_LT(x,y) (P(compare)((x).key, (y).key) < 0)
+#endif
+#ifdef SORT_INTERNAL_RADIX
+#    ifdef SORT_COPY_HASH
+#      define ASORT_HASH(x) (x).hash
+#    else
+#      define ASORT_HASH(x) P(hash)((x).key)
+#    endif
+#    ifdef SORT_LONG_HASH
+#      define ASORT_LONG_HASH
+#    endif
+#endif
+#include "lib/sorter/array.h"
+
+/*
+ *  The big_buf has the following layout:
+ *
+ *	+-------------------------------------------------------------------------------+
+ *	| array of internal_item's							|
+ *	+-------------------------------------------------------------------------------+
+ *	| padding to make the following part page-aligned				|
+ *	+--------------------------------+----------------------------------------------+
+ *	| shadow copy of item array      | array of pointers to data for write_merged() |
+ *	| used if radix-sorting          +----------------------------------------------+
+ *	|                                | workspace for write_merged()			|
+ *	+--------------------------------+----------------------------------------------+
+ *	|	       +---------+							|
+ *	|              | key     |							|
+ *	|              +---------+							|
+ *	| sequence of  | padding |							|
+ *	| items        +---------+							|
+ *	|              | data    |							|
+ *	|              +---------+							|
+ *	|              | padding |							|
+ *	|              +---------+							|
+ *	+-------------------------------------------------------------------------------+
+ *
+ *  (the data which are in different columns are never accessed simultaneously,
+ *   so we use a single buffer for both)
+ */
+
+static inline void *P(internal_get_data)(P(key) *key)
+{
+  uns ksize = SORT_KEY_SIZE(*key);
+#ifdef SORT_UNIFY
+  ksize = ALIGN_TO(ksize, CPU_STRUCT_ALIGN);
+#endif
+  return (byte *) key + ksize;
+}
+
+static inline size_t P(internal_workspace)(P(key) *key UNUSED)
+{
+  size_t ws = 0;
+#ifdef SORT_UNIFY
+  ws += sizeof(void *);
+#endif
+#ifdef SORT_UNIFY_WORKSPACE
+  ws += SORT_UNIFY_WORKSPACE(*key);
+#endif
+#ifdef SORT_INTERNAL_RADIX
+  ws = MAX(ws, sizeof(P(internal_item_t)));
+#endif
+  return ws;
+}
 
 static int P(internal)(struct sort_context *ctx, struct sort_bucket *bin, struct sort_bucket *bout, struct sort_bucket *bout_only)
 {
   sorter_alloc_buf(ctx);
-  ASSERT(bin->fb);			// Expects the input bucket to be already open for reading
-  struct fastbuf *in = bin->fb;
+  struct fastbuf *in = sbuck_read(bin);
 
   P(key) key, *keybuf = ctx->key_buf;
   if (!keybuf)
@@ -36,23 +111,22 @@ static int P(internal)(struct sort_context *ctx, struct sort_bucket *bin, struct
   else if (!P(read_key)(in, &key))
     return 0;
 
+  size_t bufsize = ctx->big_buf_size;
 #ifdef SORT_VAR_DATA
-  if (sizeof(key) + 1024 + SORT_DATA_SIZE(key) > ctx->big_buf_half_size)
+  if (sizeof(key) + 2*CPU_PAGE_SIZE + SORT_DATA_SIZE(key) + P(internal_workspace)(&key) > bufsize)
     {
-      SORT_XTRACE("s-internal: Generating a giant run");
-      struct fastbuf *out = sorter_open_write(bout); /* FIXME: Using a non-direct buffer would be nice here */
+      SORT_XTRACE(3, "s-internal: Generating a giant run");
+      struct fastbuf *out = sbuck_write(bout);
       P(copy_data)(&key, in, out);
       bout->runs++;
       return 1;				// We don't know, but 1 is always safe
     }
 #endif
 
-  size_t bufsize = ctx->big_buf_half_size;	/* FIXME: In some cases, we can use the whole buffer */
-  bufsize = MIN((u64)bufsize, (u64)~0U * sizeof(P(internal_item_t)));	// The number of records must fit in uns
-
-  SORT_XTRACE("s-internal: Reading (bufsize=%zd)", bufsize);
+  SORT_XTRACE(4, "s-internal: Reading");
   P(internal_item_t) *item_array = ctx->big_buf, *item = item_array, *last_item;
   byte *end = (byte *) ctx->big_buf + bufsize;
+  size_t remains = bufsize - CPU_PAGE_SIZE;
   do
     {
       uns ksize = SORT_KEY_SIZE(key);
@@ -63,44 +137,113 @@ static int P(internal)(struct sort_context *ctx, struct sort_bucket *bin, struct
 #endif
       uns dsize = SORT_DATA_SIZE(key);
       uns recsize = ALIGN_TO(ksize_aligned + dsize, CPU_STRUCT_ALIGN);
-      if (unlikely(sizeof(P(internal_item_t)) + recsize > (size_t)(end - (byte *) item)))
+      size_t totalsize = recsize + sizeof(P(internal_item_t) *) + P(internal_workspace)(&key);
+      if (unlikely(totalsize > remains
+#ifdef CPU_64BIT_POINTERS
+		   || item >= item_array + ~0U		// The number of items must fit in an uns
+#endif
+	 ))
 	{
 	  ctx->more_keys = 1;
 	  *keybuf = key;
 	  break;
 	}
+      remains -= totalsize;
       end -= recsize;
       memcpy(end, &key, ksize);
 #ifdef SORT_VAR_DATA
       breadb(in, end + ksize_aligned, dsize);
 #endif
       item->key = (P(key)*) end;
+#ifdef SORT_COPY_HASH
+      item->hash = P(hash)(item->key);
+#endif
       item++;
     }
   while (P(read_key)(in, &key));
   last_item = item;
 
   uns count = last_item - item_array;
-  SORT_XTRACE("s-internal: Sorting %d items", count);
-  P(array_sort)(count, item_array);
+  void *workspace UNUSED = ALIGN_PTR(last_item, CPU_PAGE_SIZE);
+  SORT_XTRACE(3, "s-internal: Read %u items (%s items, %s workspace, %s data)",
+	count,
+	stk_fsize((byte*)last_item - (byte*)item_array),
+	stk_fsize(end - (byte*)last_item - remains),
+	stk_fsize((byte*)ctx->big_buf + bufsize - end));
+  timestamp_t timer;
+  init_timer(&timer);
+  item_array = P(array_sort)(item_array, count
+#ifdef SORT_INTERNAL_RADIX
+    , workspace, bin->hash_bits
+#endif
+    );
+  ctx->total_int_time += get_timer(&timer);
 
-  SORT_XTRACE("s-internal: Writing");
+  SORT_XTRACE(4, "s-internal: Writing");
   if (!ctx->more_keys)
     bout = bout_only;
   struct fastbuf *out = sbuck_write(bout);
   bout->runs++;
-  /* FIXME: No unification done yet */
+  uns merged UNUSED = 0;
   for (item = item_array; item < last_item; item++)
     {
-      P(write_key)(out, item->key);
-#ifdef SORT_VAR_DATA
-      uns ksize = SORT_KEY_SIZE(*item->key);
 #ifdef SORT_UNIFY
-      ksize = ALIGN_TO(ksize, CPU_STRUCT_ALIGN);
+      if (item < last_item - 1 && !P(compare)(item->key, item[1].key))
+	{
+	  // Rewrite the item structures with just pointers to keys and place
+	  // pointers to data in the workspace.
+	  P(key) **key_array = (void *) item;
+	  void **data_array = workspace;
+	  key_array[0] = item[0].key;
+	  data_array[0] = P(internal_get_data)(key_array[0]);
+	  uns cnt;
+	  for (cnt=1; item+cnt < last_item && !P(compare)(key_array[0], item[cnt].key); cnt++)
+	    {
+	      key_array[cnt] = item[cnt].key;
+	      data_array[cnt] = P(internal_get_data)(key_array[cnt]);
+	    }
+	  P(write_merged)(out, key_array, data_array, cnt, data_array+cnt);
+	  item += cnt - 1;
+	  merged += cnt - 1;
+	  continue;
+	}
+#endif
+#ifdef SORT_ASSERT_UNIQUE
+      ASSERT(item == last_item-1 || P(compare)(item->key, item[1].key) < 0);
 #endif
-      bwrite(out, (byte *) item->key + ksize, SORT_DATA_SIZE(*item->key));
+      P(write_key)(out, item->key);
+#ifdef SORT_VAR_DATA
+      bwrite(out, P(internal_get_data)(item->key), SORT_DATA_SIZE(*item->key));
 #endif
     }
+#ifdef SORT_UNIFY
+  SORT_XTRACE(3, "Merging reduced %u records", merged);
+#endif
 
   return ctx->more_keys;
 }
+
+static u64
+P(internal_estimate)(struct sort_context *ctx, struct sort_bucket *b UNUSED)
+{
+  // Most of this is just wild guesses
+#ifdef SORT_VAR_KEY
+  uns avg = ALIGN_TO(sizeof(P(key))/4, CPU_STRUCT_ALIGN);
+#else
+  uns avg = ALIGN_TO(sizeof(P(key)), CPU_STRUCT_ALIGN);
+#endif
+  uns ws = 0;
+#ifdef SORT_UNIFY
+  ws += sizeof(void *);
+#endif
+#ifdef SORT_UNIFY_WORKSPACE
+  ws += avg;
+#endif
+#ifdef SORT_INTERNAL_RADIX
+  ws = MAX(ws, sizeof(P(internal_item_t)));
+#endif
+  // We ignore the data part of records, it probably won't make the estimate much worse
+  return (ctx->big_buf_size / (avg + ws + sizeof(P(internal_item_t))) * avg);
+}
+
+#undef SORT_COPY_HASH