Prefetching for sweep

ocaml · Sep 22, 2020 · b419956 · b419956
1 parent 2383971
commit b419956
Show file tree

Hide file tree

Showing 3 changed files with 12 additions and 0 deletions.
diff --git a/runtime/caml/misc.h b/runtime/caml/misc.h
@@ -113,6 +113,16 @@ CAMLdeprecated_typedef(addr, char *);
 #error "How do I align values on this platform?"
 #endif
 
+/* Prefetching */
+
+#ifdef CAML_INTERNALS
+#if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
+#define caml_prefetch(p) __builtin_prefetch((p), 1, 3)
+#else
+#define caml_prefetch(p)
+#endif
+#endif
+
 /* CAMLunused is preserved for compatibility reasons.
    Instead of the legacy GCC/Clang-only
      CAMLunused foo;

diff --git a/runtime/freelist.c b/runtime/freelist.c
@@ -1662,6 +1662,7 @@ static header_t *bf_merge_block (value bp, char *limit)
     }
     caml_fl_cur_wsz += Whsize_val (cur);
   next:
+    caml_prefetch(Hp_val(cur + 4096));
     cur = Next_in_mem (cur);
     if (Hp_val (cur) >= (header_t *) limit){
       CAMLassert (Hp_val (cur) == (header_t *) limit);

diff --git a/runtime/major_gc.c b/runtime/major_gc.c
@@ -695,6 +695,7 @@ static void sweep_slice (intnat work)
   limit = sweep_chunk + Chunk_size(sweep_chunk);
   while (work > 0){
     if (sweep_hp < limit){
+      caml_prefetch(sweep_hp + 4096);
       hp = sweep_hp;
       hd = Hd_hp (hp);
       work -= Whsize_hd (hd);