btable: optimize entry address calculation

The btable uses division and modulo calculations, which are quiet
expensive in the most platforms, for the entry address calculations.

In case the allocation limit of a btable is power of 2 number,
the calculations can be optimized to use simple bit calculations.

Optimize the entry address calculations in the aforementioned case.

Partial-Bug: #1781402
Change-Id: Idfc56e14a6d499d23a306866cfa463f95eee1151
Signed-off-by: Matan Azrad <matan@mellanox.com>

dp-core/vr_btable.c

@@ -87,6 +87,20 @@ vr_btable_free(struct vr_btable *table)
     return;
 }
 
+static void
+vr_btable_fill_pow2_fields(struct vr_btable *table) {
+    unsigned int number = table->vb_alloc_limit;
+
+    if (!(table->vb_alloc_limit & (table->vb_alloc_limit - 1))) {
+        while (number >>= 1)
+            table->vb_alloc_limit_log++;
+        table->vb_alloc_limit_mask = table->vb_alloc_limit - 1;
+    } else {
+        table->vb_alloc_limit_log = 0;
+        table->vb_alloc_limit_mask = 0;
+    }
+}
+
 struct vr_btable *
 vr_btable_alloc(unsigned int num_entries, unsigned int entry_size)
 {
@@ -160,6 +174,8 @@ vr_btable_alloc(unsigned int num_entries, unsigned int entry_size)
     table->vb_entries = num_entries;
     table->vb_esize = entry_size;
 
+    vr_btable_fill_pow2_fields(table);
+
     return table;
 
 exit_alloc:
@@ -217,6 +233,8 @@ vr_btable_attach(struct iovec *iov, unsigned int iov_len,
     table->vb_entries = (total_size / esize);
     table->vb_flags |= VB_FLAG_MEMORY_ATTACHED;
 
+    vr_btable_fill_pow2_fields(table);
+
     return table;
 
 error:

include/vr_btable.h

@@ -21,6 +21,8 @@ struct vr_btable {
     unsigned short  vb_partitions;
     unsigned int    vb_flags;
     unsigned int    vb_alloc_limit;
+    unsigned short  vb_alloc_limit_log;
+    unsigned int    vb_alloc_limit_mask;
     void            **vb_mem;
     struct vr_btable_partition *vb_table_info;
 };
@@ -48,13 +50,21 @@ vr_btable_size(struct vr_btable *table)
 static inline void *
 vr_btable_get(struct vr_btable *table, unsigned int entry)
 {
-    unsigned int t_index, t_offset;
+    unsigned int t_index, t_offset, entry_cont;
 
     if (entry >= table->vb_entries)
         return NULL;
 
-    t_index = (entry * table->vb_esize) / table->vb_alloc_limit;
-    t_offset = (entry * table->vb_esize) % table->vb_alloc_limit;
+    entry_cont = entry * table->vb_esize;
+
+    if (table->vb_alloc_limit_mask) {
+        t_index = entry_cont >> table->vb_alloc_limit_log;
+        t_offset = entry_cont & table->vb_alloc_limit_mask;
+    } else {
+        t_index = entry_cont / table->vb_alloc_limit;
+        t_offset = entry_cont % table->vb_alloc_limit;
+    }
+
     if (t_index >= table->vb_partitions)
         return NULL;