pragma unroll 4

main.cpp

@@ -115,7 +115,7 @@ int main() {
 	test_basics(mapmap, &mi);
 
 	/* Performance tests */
-	int i = 10000;
+	int i = 1000;
 	keystore(i, true);
 	datastore(i, true);
 

makefile

@@ -2,8 +2,6 @@ debug:
 	g++ main.cpp -g -Wall -o main
 release:
 	g++ main.cpp -O2 -Wall -o main
-release-native:
-	g++ main.cpp -fopt-info-vec-missed -march=native -O3 -Wall -o main
 release-avx2:
 	g++ main.cpp -fopt-info-vec-missed -mavx2 -O3 -Wall -o main
 release-avx2-asm:

simap.h

@@ -95,6 +95,7 @@ typedef union simap_ptr64 simap_ptr64;
 
 // TODO: We can possibly hand-optimise this with intrinsics maybe - but I hope autovectorization (does not seem to happen???)
 static inline SM_ALWAYS_INLINE auint64 *make_tipp(auint64 *tip, auint64 prefix, auint64 *end) {
+	#pragma GCC unroll 4
 	while((tip < end) && (*tip != prefix)) ++tip;
 	return tip;
 }
@@ -111,8 +112,6 @@ static inline simap_ptr64 *simap_search_internal(simap_instance *map, const char
 	/* Construct keyremains - might point to the \0 terminator only if smallkey or 8 bytes exactly */
 	const char *keyremains = key + prefixlen;
 
-	/* TODO: Maybe this is buggy when we access behind our own data? */
-	/* TODO: Maybe I should create separate function for fast-lookup returning "next" pointer from a pointer to autovectorize? */
 	/* Lookup prefix (fast-key) - hopefully this gets vectorized (should be)!!! */
 	auint64 *base = map->base;
 	auint64 *end = (auint64 *)((uint8_t *)base + (map->usage_end));