perfmon: compile dispatch wrapper once for each number of counters

A bit ugly, but generates faster and less noisy code which
should be important for this particular use case.

Type: improvement
Change-Id: If2bba947dac33ffedb4236a5b3fb50fc783668e1
Signed-off-by: Damjan Marion <damarion@cisco.com>

3 files changed, 27 insertions, 35 deletions

diff --git a/src/plugins/perfmon/dispatch_wrapper.c b/src/plugins/perfmon/dispatch_wrapper.c
index 64df7f56928..3ae1219fe04 100644
--- a/src/plugins/perfmon/dispatch_wrapper.c
+++ b/src/plugins/perfmon/dispatch_wrapper.c
@@ -28,38 +28,13 @@
 static_always_inline void
 perfmon_read_pmcs (u64 *counters, u32 *indexes, u8 n_counters)
 {
-  switch (n_counters)
-    {
-    case 12:
-      counters[11] = _rdpmc (indexes[11] - 1);
-    case 11:
-      counters[10] = _rdpmc (indexes[10] - 1);
-    case 10:
-      counters[9] = _rdpmc (indexes[9] - 1);
-    case 9:
-      counters[8] = _rdpmc (indexes[8] - 1);
-    case 8:
-      counters[7] = _rdpmc (indexes[7] - 1);
-    case 7:
-      counters[6] = _rdpmc (indexes[6] - 1);
-    case 6:
-      counters[5] = _rdpmc (indexes[5] - 1);
-    case 5:
-      counters[4] = _rdpmc (indexes[4] - 1);
-    case 4:
-      counters[3] = _rdpmc (indexes[3] - 1);
-    case 3:
-      counters[2] = _rdpmc (indexes[2] - 1);
-    case 2:
-      counters[1] = _rdpmc (indexes[1] - 1);
-    case 1:
-      counters[0] = _rdpmc (indexes[0] - 1);
-    }
+  for (int i = 0; i < n_counters; i++)
+    counters[i] = _rdpmc (indexes[i] - 1);
 }
 
-uword
-perfmon_dispatch_wrapper (vlib_main_t *vm, vlib_node_runtime_t *node,
-			  vlib_frame_t *frame)
+static_always_inline uword
+perfmon_dispatch_wrapper_inline (vlib_main_t *vm, vlib_node_runtime_t *node,
+				 vlib_frame_t *frame, u8 n_events)
 {
   perfmon_main_t *pm = &perfmon_main;
   perfmon_thread_runtime_t *rt =
@@ -67,8 +42,6 @@ perfmon_dispatch_wrapper (vlib_main_t *vm, vlib_node_runtime_t *node,
   perfmon_node_stats_t *s =
     vec_elt_at_index (rt->node_stats, node->node_index);
 
-  u8 n_events = rt->n_events;
-
   struct
   {
     u64 t[2][PERF_MAX_EVENTS];
@@ -102,3 +75,22 @@ perfmon_dispatch_wrapper (vlib_main_t *vm, vlib_node_runtime_t *node,
 
   return rv;
 }
+
+#define foreach_n_events                                                      \
+  _ (1) _ (2) _ (3) _ (4) _ (5) _ (6) _ (7) _ (8) _ (9) _ (10) _ (11) _ (12)
+
+#define _(x)                                                                  \
+  static uword perfmon_dispatch_wrapper##x (                                  \
+    vlib_main_t *vm, vlib_node_runtime_t *node, vlib_frame_t *frame)          \
+  {                                                                           \
+    return perfmon_dispatch_wrapper_inline (vm, node, frame, x);              \
+  }
+
+foreach_n_events
+#undef _
+
+  vlib_node_function_t *perfmon_dispatch_wrappers[PERF_MAX_EVENTS + 1] = {
+#define _(x) [x] = &perfmon_dispatch_wrapper##x,
+    foreach_n_events
+#undef _
+  };
diff --git a/src/plugins/perfmon/perfmon.c b/src/plugins/perfmon/perfmon.c
index 09178929218..0f89708567c 100644
--- a/src/plugins/perfmon/perfmon.c
+++ b/src/plugins/perfmon/perfmon.c
@@ -285,8 +285,8 @@ perfmon_start (vlib_main_t *vm, perfmon_bundle_t *b)
 	}
 
       for (int i = 0; i < vlib_get_n_threads (); i++)
-	vlib_node_set_dispatch_wrapper (vlib_get_main_by_index (i),
-					perfmon_dispatch_wrapper);
+	vlib_node_set_dispatch_wrapper (
+	  vlib_get_main_by_index (i), perfmon_dispatch_wrappers[b->n_events]);
     }
   pm->sample_time = vlib_time_now (vm);
   pm->is_running = 1;
diff --git a/src/plugins/perfmon/perfmon.h b/src/plugins/perfmon/perfmon.h
index 5003e59015e..58d971f559a 100644
--- a/src/plugins/perfmon/perfmon.h
+++ b/src/plugins/perfmon/perfmon.h
@@ -78,7 +78,7 @@ typedef struct
 } perfmon_instance_type_t;
 
 struct perfmon_source;
-vlib_node_function_t perfmon_dispatch_wrapper;
+extern vlib_node_function_t *perfmon_dispatch_wrappers[PERF_MAX_EVENTS + 1];
 
 typedef clib_error_t *(perfmon_source_init_fn_t) (vlib_main_t *vm,
 						  struct perfmon_source *);