X-Git-Url: http://mj.ucw.cz/gitweb/?a=blobdiff_plain;f=lib%2Fsorter%2Fgovern.c;h=826417d1f6e3207c91eda5cc2a53efbb76390190;hb=89e1c88c0d904e474f134cde37b78c8dedd4b9d1;hp=7b67110ba21318360976ac95263ee7b964c3d711;hpb=d119c8b3262d795777679f55015588771453d51f;p=libucw.git

diff --git a/lib/sorter/govern.c b/lib/sorter/govern.c
index 7b67110b..826417d1 100644
--- a/lib/sorter/govern.c
+++ b/lib/sorter/govern.c
@@ -48,9 +48,14 @@ sorter_presort(struct sort_context *ctx, struct sort_bucket *in, struct sort_buc
   sorter_alloc_buf(ctx);
   if (in->flags & SBF_CUSTOM_PRESORT)
     {
+      /*
+       *  The trick with automatic joining, which we use for the normal presorter,
+       *  is not necessary with the custom presorter, because the custom presorter
+       *  is never called in the middle of the sorted data.
+       */
       struct fastbuf *f = sbuck_write(out);
       out->runs++;
-      return ctx->custom_presort(f, ctx->big_buf, ctx->big_buf_size);	// FIXME: out_only optimization?
+      return ctx->custom_presort(f, ctx->big_buf, ctx->big_buf_size);
     }
   return ctx->internal_sort(ctx, in, out, out_only);
 }
@@ -158,7 +163,7 @@ sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
   do {
     ++pass;
     sorter_start_timer(ctx);
-    if (ins[0]->runs == 1 && ins[1]->runs == 1 && join)
+    if (ins[0]->runs <= 1 && ins[1]->runs <= 1 && join)
       {
 	// This is guaranteed to produce a single run, so join if possible
 	outs[0] = join;
@@ -226,7 +231,7 @@ sorter_multiway(struct sort_context *ctx, struct sort_bucket *b)
 
   if (part_cnt <= 1)
     {
-      sh_off_t size = sbuck_ins_or_join(clist_head(&parts), list_pos, join, join_size);
+      sh_off_t size = sbuck_ins_or_join(clist_head(&parts), list_pos, (part_cnt ? NULL : join), join_size);
       SORT_XTRACE(trace_level, "Sorted in memory (%s, %dMB/s)", stk_fsize(size), sorter_speed(ctx, size));
       return;
     }
@@ -325,7 +330,7 @@ sorter_decide(struct sort_context *ctx, struct sort_bucket *b)
   // How many bits of bucket size we have to reduce before it fits in the RAM?
   // (this is insanely large if the input size is unknown, but it serves our purpose)
   u64 insize = sbuck_size(b);
-  u64 mem = ctx->internal_estimate(ctx, b) * 0.8;	// FIXME: Magical factor for various non-uniformities
+  u64 mem = ctx->internal_estimate(ctx, b) * 0.8;	// Magical factor accounting for various non-uniformities
   uns bits = 0;
   while ((insize >> bits) > mem)
     bits++;
@@ -391,6 +396,7 @@ sorter_run(struct sort_context *ctx)
   ctx->pool = mp_new(4096);
   clist_init(&ctx->bucket_list);
   sorter_prepare_buf(ctx);
+  asort_start_threads(0);
 
   // Create bucket containing the source
   struct sort_bucket *bin = sbuck_new(ctx);
@@ -419,6 +425,7 @@ sorter_run(struct sort_context *ctx)
   while (bout = clist_head(&ctx->bucket_list), b = clist_next(&ctx->bucket_list, &bout->n))
     sorter_decide(ctx, b);
 
+  asort_stop_threads();
   sorter_free_buf(ctx);
   sbuck_write(bout);		// Force empty bucket to a file
   SORT_XTRACE(2, "Final size: %s", F_BSIZE(bout));