improvements: Made GetRowPointers even leaner

src/execution/join_hashtable.cpp

@@ -14,7 +14,7 @@ using ProbeSpill = JoinHashTable::ProbeSpill;
 using ProbeSpillLocalState = JoinHashTable::ProbeSpillLocalAppendState;
 
 JoinHashTable::ProbeState::ProbeState()
-    : ht_offsets_v(LogicalType::UBIGINT), ht_offsets_dense_v(LogicalType::UBIGINT),
+    : ht_offsets_v(LogicalType::UBIGINT), salt_v(LogicalType::UBIGINT), ht_offsets_dense_v(LogicalType::UBIGINT),
       row_ptr_insert_to_v(LogicalType::POINTER), non_empty_sel(STANDARD_VECTOR_SIZE),
       key_no_match_sel(STANDARD_VECTOR_SIZE), salt_match_sel(STANDARD_VECTOR_SIZE) {
 }
@@ -157,6 +157,7 @@ void JoinHashTable::GetRowPointers(DataChunk &keys, TupleDataChunkState &key_sta
 
 	auto hashes = UnifiedVectorFormat::GetData<hash_t>(hashes_v_unified);
 
+	auto salts = FlatVector::GetData<hash_t>(state.salt_v);
 	auto ht_offsets = FlatVector::GetData<idx_t>(state.ht_offsets_v);
 	auto ht_offsets_dense = FlatVector::GetData<idx_t>(state.ht_offsets_dense_v);
 
@@ -184,6 +185,11 @@ void JoinHashTable::GetRowPointers(DataChunk &keys, TupleDataChunkState &key_sta
 		idx_t dense_index = state.non_empty_sel.get_index(i);
 		const auto row_index = sel.get_index(dense_index);
 		state.non_empty_sel.set_index(i, row_index);
+
+		auto uvf_index = hashes_v_unified.sel->get_index(row_index);
+		auto hash = hashes[uvf_index];
+		hash_t row_salt = aggr_ht_entry_t::ExtractSalt(hash);
+		salts[row_index] = row_salt;
 	}
 
 	auto pointers_result = FlatVector::GetData<data_ptr_t>(pointers_result_v);
@@ -208,43 +214,38 @@ void JoinHashTable::GetRowPointers(DataChunk &keys, TupleDataChunkState &key_sta
 			auto &ht_offset = ht_offsets[row_index];
 
 			idx_t increment;
+			bool salt_match;
+			aggr_ht_entry_t entry;
 
 			// increment the ht_offset of the entry as long as next entry is occupied and salt does not match
 			do {
 
-				auto &entry = entries[ht_offset];
+				entry = entries[ht_offset];
 				bool occupied = entry.IsOccupied();
 
 				// no need to do anything, as the vector is zeroed
 				if (!occupied) {
 					break;
 				}
-				auto uvf_index = hashes_v_unified.sel->get_index(row_index);
-				auto hash = hashes[uvf_index];
-				hash_t row_salt = aggr_ht_entry_t::ExtractSalt(hash);
-				bool salt_match = entry.GetSalt() == row_salt;
 
-				// the entries we need to process in the next iteration are the ones that are occupied and the row_salt
-				// does not match, the ones that are empty need no further processing
-				state.salt_match_sel.set_index(salt_match_count, row_index);
-				salt_match_count += salt_match;
+				salt_match = entry.GetSalt() == salts[row_index];
 
 				// condition for incrementing the ht_offset: occupied and row_salt does not match -> move to next entry
 				increment = !salt_match;
 				IncrementAndWrap(ht_offset, increment, bitmask);
 
 			} while (increment);
+
+			// the entries we need to process in the next iteration are the ones that are occupied and the row_salt
+			// does not match, the ones that are empty need no further processing
+			state.salt_match_sel.set_index(salt_match_count, row_index);
+			salt_match_count += salt_match;
+			row_ptr_insert_to[row_index] = entry.GetPointer();
 		}
 
 		if (salt_match_count == 0) {
 			break;
 		} else {
-			// Get the pointers_result_v to the rows that need to be compared
-			for (idx_t need_compare_idx = 0; need_compare_idx < salt_match_count; need_compare_idx++) {
-				const auto row_index = state.salt_match_sel.get_index(need_compare_idx);
-				const auto &entry = entries[ht_offsets[row_index]];
-				row_ptr_insert_to[row_index] = entry.GetPointer();
-			}
 
 			// Perform row comparisons, after function call salt_match_sel will point to the keys that match
 			idx_t key_no_match_count = 0;

src/include/duckdb/execution/join_hashtable.hpp

@@ -126,6 +126,7 @@ class JoinHashTable {
 		Vector ht_offsets_v;
 		Vector ht_offsets_dense_v;
 		Vector row_ptr_insert_to_v;
+		Vector salt_v;
 
 		SelectionVector non_empty_sel;