Simple movement of return calculation in lcs_upper_bound

This increased performance of the non-DO_SECOND_SUM code a _lot_ for something
so small.

1 files changed, 9 insertions, 8 deletions

diff --git a/lcs.c b/lcs.c
index 2346f12..d296ea7 100644
--- a/lcs.c
+++ b/lcs.c
@@ -158,7 +158,7 @@ int lcs_32_rough(unsigned char *str1, unsigned char *str2) {
 #endif
 
 #ifdef USE_PRE_LCS
-#define DO_SECOND_SUM
+//efine DO_SECOND_SUM
 
 #ifdef DO_SECOND_SUM
 #include "arpa/inet.h"
@@ -169,6 +169,7 @@ int lcs_upper_bound(const uint32_t * const str1_int, const uint32_t * const str2
     const unsigned char * const str1 = (const unsigned char * const)str1_int;
     const unsigned char * const str2 = (const unsigned char * const)str2_int;
     int first_sum;
+    int ret;
 #ifdef DO_SECOND_SUM
     uint32_t bo_str2_int[4];
     uint32_t shifted_str2_int[4];
@@ -207,6 +208,12 @@ int lcs_upper_bound(const uint32_t * const str1_int, const uint32_t * const str2
         + contains[str2_r[9]] + contains[str2_r[10]] + contains[str2_r[11]]
         + contains[str2_r[12]] + contains[str2_r[13]] + contains[str2_r[14]]
         + contains[str2_r[15]];
+
+    // can be off up to one nibble (e.g., "1110" and "2111")
+    ret = MAX(first_sum, second_sum) * 2 + 1;
+#else
+    // can be off up to two nibbles (e.g., "0110" and "1122")
+    ret = first_sum * 2 + 2;
 #endif
 
     contains[str1[0]] = contains[str1[1]] = contains[str1[2]]
@@ -216,13 +223,7 @@ int lcs_upper_bound(const uint32_t * const str1_int, const uint32_t * const str2
         = contains[str1[12]] = contains[str1[13]] = contains[str1[14]]
         = contains[str1[15]] = 0;
 
-#ifdef DO_SECOND_SUM
-    // can be off up to one nibble (e.g., "1110" and "2111")
-    return MAX(first_sum, second_sum) * 2 + 1;
-#else
-    // can be off up to two nibbles (e.g., "0110" and "1122")
-    return first_sum * 2 + 2;
-#endif
+    return ret;
 }
 #endif