diff --git a/drama/src/rev-mc.c b/drama/src/rev-mc.c
index dc94777..4c765a9 100644
--- a/drama/src/rev-mc.c
+++ b/drama/src/rev-mc.c
@@ -172,7 +172,7 @@ reduce_masks(std::vector<uint64_t> masks)
 
 //----------------------------------------------------------
 // from https://graphics.stanford.edu/~seander/bithacks.html#NextBitPermutation
-uint64_t
+inline uint64_t
 next_bit_permutation(uint64_t v)
 {
     uint64_t t = v | (v - 1);
@@ -300,8 +300,9 @@ find_row_mask(std::vector<set_t> &sets,
     row_mask <<= CL_SHIFT;// skip the lowest 6 bits since they're used for CL addressing
 
     while (row_mask < last_mask) {
+        uint64_t next_row_mask = next_bit_permutation(row_mask);
         if (row_mask & LS_BITMASK(CL_SHIFT)) {
-            row_mask = next_bit_permutation(row_mask);
+            row_mask = next_row_mask;
             continue;
         }
 
@@ -316,7 +317,7 @@ find_row_mask(std::vector<set_t> &sets,
         break;
 
     next_mask:
-        row_mask = next_bit_permutation(row_mask);
+        row_mask = next_row_mask;
     }
 
     // super hackish way to recover the real row mask
@@ -330,6 +331,9 @@ find_row_mask(std::vector<set_t> &sets,
 }
 
 //----------------------------------------------------------
+#pragma push_options
+#pragma optimize("O3")
+
 void
 rev_mc(size_t sets_cnt, size_t threshold, size_t rounds, size_t m_size, char *o_file, uint64_t flags)
 {
@@ -413,6 +417,8 @@ rev_mc(size_t sets_cnt, size_t threshold, size_t rounds, size_t m_size, char *o_
     free_buffer(&mem);
 }
 
+#pragma pop_options
+
 // Fin.
 
 //----------------------------------------------------------