fix: Optimize ContainerRowSerde deserialization for string, array and…

… map (#12362)

Summary:

1. Optimize `deserializeString` in case of inlined strings
2. Optimize `deserializeArray` to avoid heap allocation for intermediate null-like data
3. Inline `FlatVector<StringView>::setNoCopy`

Differential Revision: D69750091

velox/exec/ContainerRowSerde.cpp

@@ -224,14 +224,17 @@ void deserializeString(
   VELOX_CHECK_EQ(result.encoding(), VectorEncoding::Simple::FLAT);
   auto values = result.asUnchecked<FlatVector<StringView>>();
   auto size = in.read<int32_t>();
-  auto buffer = values->getBufferWithSpace(size);
-  auto start = buffer->asMutable<char>() + buffer->size();
-  in.readBytes(start, size);
-  // If the string is not inlined in string view, we need to advance the buffer.
-  if (not StringView::isInline(size)) {
+  if (StringView::isInline(size)) {
+    char data[StringView::kInlineSize];
+    in.readBytes(data, size);
+    values->setNoCopy(index, StringView(data, size));
+  } else {
+    auto* buffer = values->getBufferWithSpace(size);
+    auto* start = buffer->asMutable<char>() + buffer->size();
+    in.readBytes(start, size);
     buffer->setSize(buffer->size() + size);
+    values->setNoCopy(index, StringView(start, size));
   }
-  values->setNoCopy(index, StringView(start, size));
 }
 
 template <>
@@ -250,12 +253,17 @@ void deserializeOne<TypeKind::VARBINARY>(
   deserializeString(in, index, result);
 }
 
-std::vector<uint64_t> readNulls(ByteInputStream& in, int32_t size) {
+void readNulls(ByteInputStream& in, int32_t size, uint64_t* out) {
   auto n = bits::nwords(size);
-  std::vector<uint64_t> nulls(n);
   for (auto i = 0; i < n; ++i) {
-    nulls[i] = in.read<uint64_t>();
+    out[i] = in.read<uint64_t>();
   }
+}
+
+std::vector<uint64_t> readNulls(ByteInputStream& in, int32_t size) {
+  auto n = bits::nwords(size);
+  std::vector<uint64_t> nulls(n);
+  readNulls(in, size, nulls.data());
   return nulls;
 }
 
@@ -292,11 +300,20 @@ vector_size_t deserializeArray(
     BaseVector& elements,
     vector_size_t& offset) {
   auto size = in.read<int32_t>();
-  auto nulls = readNulls(in, size);
   offset = elements.size();
   elements.resize(offset + size);
+  uint64_t smallNulls;
+  std::vector<uint64_t> largeNulls;
+  const uint64_t* nulls;
+  if (size <= 64) {
+    readNulls(in, size, &smallNulls);
+    nulls = &smallNulls;
+  } else {
+    largeNulls = readNulls(in, size);
+    nulls = largeNulls.data();
+  }
   for (auto i = 0; i < size; ++i) {
-    if (bits::isBitSet(nulls.data(), i)) {
+    if (bits::isBitSet(nulls, i)) {
       elements.setNull(i + offset, true);
     } else {
       deserializeSwitch(in, i + offset, elements);

velox/vector/FlatVector.cpp

@@ -119,23 +119,6 @@ void FlatVector<StringView>::set(vector_size_t idx, StringView value) {
   }
 }
 
-/// For types that requires buffer allocation this should be called only if
-/// value is inlined or if value is already allocated in a buffer within the
-/// vector. Used by StringWriter to allow UDFs to write directly into the
-/// buffers and avoid copying.
-template <>
-void FlatVector<StringView>::setNoCopy(
-    const vector_size_t idx,
-    const StringView& value) {
-  VELOX_DCHECK_LT(idx, BaseVector::length_);
-  ensureValues();
-  VELOX_DCHECK(!values_->isView());
-  if (BaseVector::nulls_) {
-    BaseVector::setNull(idx, false);
-  }
-  rawValues_[idx] = value;
-}
-
 template <>
 void FlatVector<StringView>::acquireSharedStringBuffers(
     const BaseVector* source) {

velox/vector/FlatVector.h

@@ -610,10 +610,22 @@ Range<bool> FlatVector<bool>::asRange() const;
 template <>
 void FlatVector<StringView>::set(vector_size_t idx, StringView value);
 
+/// For types that requires buffer allocation this should be called only if
+/// value is inlined or if value is already allocated in a buffer within the
+/// vector. Used by StringWriter to allow UDFs to write directly into the
+/// buffers and avoid copying.
 template <>
-void FlatVector<StringView>::setNoCopy(
+inline void FlatVector<StringView>::setNoCopy(
     const vector_size_t idx,
-    const StringView& value);
+    const StringView& value) {
+  VELOX_DCHECK_LT(idx, BaseVector::length_);
+  ensureValues();
+  VELOX_DCHECK(!values_->isView());
+  if (BaseVector::nulls_) {
+    BaseVector::setNull(idx, false);
+  }
+  rawValues_[idx] = value;
+}
 
 template <>
 void FlatVector<bool>::set(vector_size_t idx, bool value);