enhance: [2.5] Optimize retrieve performance utilizing bitset SIMD methods (#39041)

Cherry pick from master
pr: #39004 #39037
Related to #39003

---------

Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>

internal/core/src/exec/expression/TermExpr.cpp

@@ -223,9 +223,10 @@ PhyTermFilterExpr::ExecPkTermImpl() {
     TargetBitmapView valid_res(res_vec->GetValidRawData(), real_batch_size);
     valid_res.set();
 
-    for (size_t i = 0; i < real_batch_size; ++i) {
-        res[i] = cached_bits_[current_data_chunk_pos_++];
-    }
+    auto current_chunk_view =
+        cached_bits_.view(current_data_chunk_pos_, real_batch_size);
+    res |= current_chunk_view;
+    current_data_chunk_pos_ += real_batch_size;
 
     return res_vec;
 }

internal/core/src/segcore/ChunkedSegmentSealedImpl.cpp

@@ -1255,20 +1255,24 @@ ChunkedSegmentSealedImpl::find_first(int64_t limit,
     std::vector<int64_t> seg_offsets;
     seg_offsets.reserve(limit);
 
+    // flip bitset since `find_next` is used to find true.
+    auto flipped = bitset.clone();
+    flipped.flip();
+
     int64_t offset = 0;
-    for (; hit_num < limit && offset < num_rows_.value(); offset++) {
+    std::optional<size_t> result = flipped.find_first();
+    while (result.has_value() && hit_num < limit) {
+        hit_num++;
+        seg_offsets.push_back(result.value());
+        offset = result.value();
         if (offset >= size) {
             // In fact, this case won't happen on sealed segments.
             continue;
         }
-
-        if (!bitset[offset]) {
-            seg_offsets.push_back(offset);
-            hit_num++;
-        }
+        result = flipped.find_next(offset);
     }
 
-    return {seg_offsets, more_hit_than_limit && offset != num_rows_.value()};
+    return {seg_offsets, more_hit_than_limit && result.has_value()};
 }
 
 ChunkedSegmentSealedImpl::ChunkedSegmentSealedImpl(

internal/core/src/segcore/SegmentSealedImpl.cpp

@@ -1719,20 +1719,25 @@ SegmentSealedImpl::find_first(int64_t limit, const BitsetType& bitset) const {
     std::vector<int64_t> seg_offsets;
     seg_offsets.reserve(limit);
 
+    // flip bitset since `find_first` & `find_next` is used to find true.
+    // could be optimized by support find false in bitset.
+    auto flipped = bitset.clone();
+    flipped.flip();
+
     int64_t offset = 0;
-    for (; hit_num < limit && offset < num_rows_.value(); offset++) {
+    std::optional<size_t> result = flipped.find_first();
+    while (result.has_value() && hit_num < limit) {
+        hit_num++;
+        seg_offsets.push_back(result.value());
+        offset = result.value();
         if (offset >= size) {
             // In fact, this case won't happen on sealed segments.
             continue;
         }
-
-        if (!bitset[offset]) {
-            seg_offsets.push_back(offset);
-            hit_num++;
-        }
+        result = flipped.find_next(offset);
     }
 
-    return {seg_offsets, more_hit_than_limit && offset != num_rows_.value()};
+    return {seg_offsets, more_hit_than_limit && result.has_value()};
 }
 
 SegcoreError