Optimize bucket scanning

Author: kg

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

@@ -83,21 +83,11 @@ static DN_FORCEINLINE(dn_simdhash_suffixes)
 build_search_vector (uint8_t needle)
 {
 	dn_simdhash_suffixes result;
-	// this produces a splat and then .const, .and in wasm, and the other architectures are fine too
+	// this produces a splat in wasm, and the other architectures are fine too
 	dn_u8x16 needles = {
 		needle, needle, needle, needle, needle, needle, needle, needle,
 		needle, needle, needle, needle, needle, needle, needle, needle
 	};
-	/*
-	// TODO: Evaluate whether the & mask is actually worth it. In the C# prototype, it wasn't.
-	// Not doing it means there is a ~1% chance of a false positive in the data bytes near the
-	//  end of the bucket, but we check the index against the bucket count anyway, so...
-	dn_u8x16 mask = {
-		0xFFu, 0xFFu, 0xFFu, 0xFFu, 0xFFu, 0xFFu, 0xFFu, 0xFFu,
-		0xFFu, 0xFFu, 0xFFu, 0xFFu, 0xFFu, 0xFFu, 0x00u, 0x00u
-	};
-	result.vec = needles & mask;
-	*/
 	result.vec = needles;
 	return result;
 }
@@ -169,9 +159,7 @@ build_search_vector (uint8_t needle)
 static DN_FORCEINLINE(uint32_t)
 find_first_matching_suffix (dn_simdhash_suffixes needle, dn_simdhash_suffixes haystack, uint32_t count)
 {
-	// FIXME: Completely untested.
-	__m128i match_vector = _mm_cmpeq_epi8(needle.m128, haystack.m128);
-	return ctz(_mm_movemask_epi8(match_vector));
+	return ctz(_mm_movemask_epi8(_mm_cmpeq_epi8(needle.m128, haystack.m128)));
 }
 
 #else // unknown compiler and/or unknown non-simd arch

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

@@ -165,15 +165,18 @@ DN_SIMDHASH_SCAN_BUCKET_INTERNAL (DN_SIMDHASH_T_PTR hash, bucket_t *bucket, DN_S
 #define BEGIN_SCAN_BUCKETS(initial_index, bucket_index, bucket_address) \
 	{ \
 		uint32_t bucket_index = initial_index; \
-		bucket_t *bucket_address; \
-		do { \
-			bucket_address = address_of_bucket(buffers, bucket_index);
+		bucket_t *bucket_address = address_of_bucket(buffers, bucket_index); \
+		do {
 
 #define END_SCAN_BUCKETS(initial_index, bucket_index, bucket_address) \
 			bucket_index++; \
 			/* Wrap around if we hit the last bucket. */ \
-			if (bucket_index >= buffers.buckets_length) \
+			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); \
 	}