Use a different file access method for small inputs.

[libucw.git] / lib / sorter / govern.c

diff --git a/lib/sorter/govern.c b/lib/sorter/govern.c
index 13fc7a61404f801d5125f081423960cb9de0bebc..5ce884732a4209a44c962890c0df3de90d594804 100644 (file)
--- a/lib/sorter/govern.c
+++ b/lib/sorter/govern.c
@@ -48,6 +48,11 @@ 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);
@@ -278,6 +283,9 @@ sorter_multiway(struct sort_context *ctx, struct sort_bucket *b)
 static void
 sorter_radix(struct sort_context *ctx, struct sort_bucket *b, uns bits)
 {
+  // Add more bits if requested and allowed.
+  bits = MIN(bits + sorter_add_radix_bits, sorter_max_radix_bits);
+
   uns nbuck = 1 << bits;
   SORT_XTRACE(2, "Running radix split on %s with hash %d bits of %d (expecting %s buckets)",
              F_BSIZE(b), bits, b->hash_bits, stk_fsize(sbuck_size(b) / nbuck));
@@ -405,6 +413,7 @@ sorter_run(struct sort_context *ctx)
   bin->hash_bits = ctx->hash_bits;
   clist_add_tail(&ctx->bucket_list, &bin->n);
   SORT_XTRACE(2, "Input size: %s, %d hash bits", F_BSIZE(bin), bin->hash_bits);
+  ctx->fb_params = (bin->size < sorter_small_input) ? &sorter_small_fb_params : &sorter_fb_params;
 
   // Create bucket for the output
   struct sort_bucket *bout = sbuck_new(ctx);