Bucket scan optimization

src/native/containers/dn-simdhash-specialization.h

@@ -146,12 +146,14 @@ DN_SIMDHASH_SCAN_BUCKET_INTERNAL (DN_SIMDHASH_T_PTR hash, bucket_t *bucket, DN_S
 #else
 	uint32_t index = find_first_matching_suffix(search_vector, bucket_suffixes, count);
 #endif
-	DN_SIMDHASH_KEY_T *key = &bucket->keys[index];
-
-	for (; index < count; index++, key++) {
+	for (; index < count; index++) {
 		// FIXME: Could be profitable to manually hoist the data load outside of the loop,
 		//  if not out of SCAN_BUCKET_INTERNAL entirely. Clang appears to do LICM on it.
-		if (DN_SIMDHASH_KEY_EQUALS(DN_SIMDHASH_GET_DATA(hash), needle, *key))
+		// It's better to index bucket->keys each iteration inside the loop than to precompute
+		//  a pointer outside and bump the pointer, because in many cases the bucket will be
+		//  empty, and in many other cases it will have one match. Putting the index inside the
+		//  loop means that for empty/no-match buckets we don't do the index calculation at all.
+		if (DN_SIMDHASH_KEY_EQUALS(DN_SIMDHASH_GET_DATA(hash), needle, bucket->keys[index]))
 			return index;
 	}
 
@@ -170,12 +172,11 @@ DN_SIMDHASH_SCAN_BUCKET_INTERNAL (DN_SIMDHASH_T_PTR hash, bucket_t *bucket, DN_S
 
 #define END_SCAN_BUCKETS(initial_index, bucket_index, bucket_address) \
 			bucket_index++; \
+			bucket_address++; \
 			/* Wrap around if we hit the last bucket. */ \
 			if (bucket_index >= buffers.buckets_length) { \
 				bucket_index = 0; \
 				bucket_address = address_of_bucket(buffers, 0); \
-			} else { \
-				bucket_address++; \
 			} \
 		/* if bucket_index == initial_index, we reached our starting point */ \
 		} while (bucket_index != initial_index); \

src/native/containers/dn-simdhash-test.c

@@ -17,7 +17,7 @@ typedef struct {
 	float f;
 } instance_data_t;
 
-static inline uint8_t
+static DN_FORCEINLINE(uint8_t)
 key_comparer (instance_data_t data, size_t lhs, size_t rhs) {
 	return ((data.f == 4.20f) || (lhs == rhs));
 }