Optimize null bitmap handling for better append performance

contrib/pax_storage/src/cpp/comm/bitmap.h

@@ -134,12 +134,31 @@ struct BitmapRaw final {
     static_assert(BM_WORD_BITS == (1 << BM_WORD_SHIFTS));
     return (index >> BM_WORD_SHIFTS) < size;
   }
-  inline bool Empty() const {
+
+
+  inline bool Empty(uint32 end_index) const {
     if (!bitmap) return true;
-    for (size_t i = 0; i < size; i++)
-      if (bitmap[i]) return false;
+
+    uint32 end_word = BM_INDEX_WORD_OFF(end_index);
+    uint32 end_bit_offset = BM_INDEX_BIT_OFF(end_index);
+
+    for (uint32 i = 0; i < end_word && i < size; i++) {
+      if (bitmap[i] != 0) return false;
+    }
+
+    // Check partial word at end
+    if (end_word < size && end_bit_offset > 0) {
+      T mask = (T(1) << end_bit_offset) - 1;
+      if (bitmap[end_word] & mask) return false;
+    }
+
     return true;
   }
+
+  inline bool Empty() const {
+    return Empty(size);
+  }
+
   BitmapRaw() = default;
   BitmapRaw(T *buffer, size_t size) : bitmap(buffer), size(size) {}
   BitmapRaw(const BitmapRaw &) = delete;
@@ -160,13 +179,14 @@ struct BitmapRaw final {
 template <typename T>
 class BitmapTpl final {
  public:
-  using BitmapMemoryPolicy = void (*)(BitmapRaw<T> &, uint32);
-  explicit BitmapTpl(uint32 initial_size = 16) {
+  using BitmapMemoryPolicy = void (*)(BitmapRaw<T> &, uint32, uint8);
+  explicit BitmapTpl(uint32 initial_size = 16, uint8 init_value = 0) {
     static_assert(sizeof(T) == 1 || sizeof(T) == 2 || sizeof(T) == 4 ||
                   sizeof(T) == 8);
     static_assert(BM_WORD_BITS == (1 << BM_WORD_SHIFTS));
     policy_ = DefaultBitmapMemoryPolicy;
-    policy_(raw_, Max(initial_size, 16));
+    policy_(raw_, Max(initial_size, 16), init_value);
+    init_value_ = init_value;
   }
   explicit BitmapTpl(const BitmapRaw<T> &raw) {
     static_assert(sizeof(T) == 1 || sizeof(T) == 2 || sizeof(T) == 4 ||
@@ -205,11 +225,11 @@ class BitmapTpl final {
 
   inline size_t WordBits() const { return BM_WORD_BITS; }
   inline void Set(uint32 index) {
-    if (unlikely(!raw_.HasEnoughSpace(index))) policy_(raw_, index);
+    if (unlikely(!raw_.HasEnoughSpace(index))) policy_(raw_, index, init_value_);
     raw_.Set(index);
   }
   inline void SetN(uint32 index) {
-    if (unlikely(!raw_.HasEnoughSpace(index))) policy_(raw_, index);
+    if (unlikely(!raw_.HasEnoughSpace(index))) policy_(raw_, index, init_value_);
     raw_.SetN(index);
   }
   inline void Clear(uint32 index) {
@@ -228,7 +248,7 @@ class BitmapTpl final {
   }
   // invert the bit and return the old value.
   inline bool Toggle(uint32 index) {
-    if (unlikely(!raw_.HasEnoughSpace(index))) policy_(raw_, index);
+    if (unlikely(!raw_.HasEnoughSpace(index))) policy_(raw_, index, init_value_);
     return raw_.Toggle(index);
   }
   // count bits in range [0, index]
@@ -247,24 +267,27 @@ class BitmapTpl final {
   }
 
   inline bool Empty() const { return raw_.Empty(); }
+  inline bool Empty(uint32 end_index) const {
+    return raw_.Empty(end_index);
+  }
 
   BitmapMemoryPolicy Policy() const { return policy_; }
 
   const BitmapRaw<T> &Raw() const { return raw_; }
   BitmapRaw<T> &Raw() { return raw_; }
 
-  static void DefaultBitmapMemoryPolicy(BitmapRaw<T> &raw, uint32 index) {
+  static void DefaultBitmapMemoryPolicy(BitmapRaw<T> &raw, uint32 index, uint8 init_value = 0) {
     auto old_bitmap = raw.bitmap;
     auto old_size = raw.size;
     auto size = Max(BM_INDEX_WORD_OFF(index) + 1, old_size * 2);
     auto p = PAX_NEW_ARRAY<T>(size);
     if (old_size > 0) memcpy(p, old_bitmap, sizeof(T) * old_size);
-    memset(&p[old_size], 0, sizeof(T) * (size - old_size));
+    memset(&p[old_size], init_value, sizeof(T) * (size - old_size));
     raw.bitmap = p;
     raw.size = size;
     PAX_DELETE_ARRAY(old_bitmap);
   }
-  static void ReadOnlyRefBitmap(BitmapRaw<T> & /*raw*/, uint32 /*index*/) {
+  static void ReadOnlyRefBitmap(BitmapRaw<T> & /*raw*/, uint32 /*index*/, uint8 /*init_value*/) {
     // raise
     CBDB_RAISE(cbdb::CException::kExTypeInvalidMemoryOperation);
   }
@@ -315,6 +338,7 @@ class BitmapTpl final {
 
   BitmapRaw<T> raw_;
   BitmapMemoryPolicy policy_;
+  uint8 init_value_ = 0;
 };
 
 using Bitmap8 = BitmapTpl<uint8>;

contrib/pax_storage/src/cpp/storage/columns/pax_column.cc

@@ -83,13 +83,21 @@ size_t PaxColumn::GetRangeNonNullRows(size_t start_pos, size_t len) {
 
 void PaxColumn::CreateNulls(size_t cap) {
   Assert(!null_bitmap_);
-  null_bitmap_ = std::make_unique<Bitmap8>(cap);
-  null_bitmap_->SetN(total_rows_);
+  // By default, initialize every bit in the null bitmap to 1.
+  // This is based on the assumption that null values are much less frequent
+  // than non-null values in most datasets. As a result, when appending non-null
+  // values, we can simply skip setting the bit to 1, since it is already set.
+  // Only when appending a null value do we need to explicitly clear the
+  // corresponding bit.
+  null_bitmap_ = std::make_unique<Bitmap8>(cap, 0xff);
 }
 
 void PaxColumn::AppendNull() {
   if (!null_bitmap_) {
-    CreateNulls(DEFAULT_CAPACITY);
+    // Ensure that the capacity of null_bitmap_ is pax_max_tuples_per_group.
+    // This design allows the use of raw_bitmap in normal cases without
+    // incurring the overhead of checking the bitmap's capacity.
+    CreateNulls(pax::pax_max_tuples_per_group);
   }
   null_bitmap_->Clear(total_rows_);
   ++total_rows_;
@@ -111,7 +119,6 @@ void PaxColumn::AppendToast(char *buffer, size_t size) {
 }
 
 void PaxColumn::Append(char * /*buffer*/, size_t /*size*/) {
-  if (null_bitmap_) null_bitmap_->Set(total_rows_);
   ++total_rows_;
   ++non_null_rows_;
 }

contrib/pax_storage/src/cpp/storage/columns/pax_column.h

@@ -46,7 +46,8 @@
 
 namespace pax {
 
-#define DEFAULT_CAPACITY MIN(2048, MAX(16, MAXALIGN(pax::pax_max_tuples_per_group)))
+#define DEFAULT_CAPACITY \
+  MIN(2048, MAX(16, MAXALIGN(pax::pax_max_tuples_per_group)))
 
 // Used to mapping pg_type
 enum PaxColumnTypeInMem {
@@ -230,7 +231,14 @@ class PaxColumn {
   inline bool HasNull() { return null_bitmap_ != nullptr; }
 
   // Are all values null?
-  inline bool AllNull() const { return null_bitmap_ && null_bitmap_->Empty(); }
+  // Check whether all bits in the specified range are zero.
+  // In pax_column, to avoid checking the capacity of the null bitmap, we
+  // allocate memory based on pax_max_tuples_per_group. As a result, the last
+  // group may contain fewer tuples than pax_max_tuples_per_group, so we need to
+  // check whether all bits in the range [0, total_rows_) are zero.
+  inline bool AllNull() const {
+    return null_bitmap_ && null_bitmap_->Empty(total_rows_);
+  }
 
   // Set the null bitmap
   inline void SetBitmap(std::unique_ptr<Bitmap8> null_bitmap) {