Implement Branchless Filter for small primitive lists

Adds a const-generic unrolled comparison chain that avoids CPU branching. Outperforms hash lookups for very small lists. Triggers for primitives when list size <= 32 (4-byte), 16 (8-byte), or 4 (16-byte).

Author: geoffreyclaude

datafusion/physical-expr/src/expressions/in_list/primitive_filter.rs

@@ -173,6 +173,91 @@ impl<C: BitmapFilterConfig> StaticFilter for BitmapFilter<C> {
     }
 }
 
+// =============================================================================
+// BRANCHLESS FILTER (Const Generic for Small Lists)
+// =============================================================================
+
+/// A branchless filter for very small IN lists (0-16 elements).
+///
+/// Uses const generics to unroll the membership check into a fixed-size
+/// comparison chain, outperforming hash lookups for small lists due to:
+/// - No branching (uses bitwise OR to combine comparisons)
+/// - Better CPU pipelining
+/// - No hash computation overhead
+pub(crate) struct BranchlessFilter<T: ArrowPrimitiveType, const N: usize> {
+    null_count: usize,
+    values: [T::Native; N],
+}
+
+impl<T: ArrowPrimitiveType, const N: usize> BranchlessFilter<T, N>
+where
+    T::Native: Copy + PartialEq,
+{
+    /// Try to create a branchless filter if the array has exactly N non-null values.
+    pub(crate) fn try_new(in_array: &ArrayRef) -> Option<Result<Self>> {
+        let in_array = in_array.as_primitive_opt::<T>()?;
+        let non_null_count = in_array.len() - in_array.null_count();
+        if non_null_count != N {
+            return None;
+        }
+        let values: Vec<_> = in_array.iter().flatten().collect();
+        // Use default_value() from ArrowPrimitiveType trait instead of Default::default()
+        let mut arr = [T::default_value(); N];
+        arr.copy_from_slice(&values);
+        Some(Ok(Self {
+            null_count: in_array.null_count(),
+            values: arr,
+        }))
+    }
+
+    /// Branchless membership check using OR-chain.
+    #[inline(always)]
+    fn check(&self, needle: T::Native) -> bool {
+        self.values
+            .iter()
+            .fold(false, |acc, &v| acc | (v == needle))
+    }
+
+    /// Check membership using a raw values slice (zero-copy path for type reinterpretation).
+    #[inline]
+    pub(crate) fn contains_slice(
+        &self,
+        values: &[T::Native],
+        nulls: Option<&NullBuffer>,
+        negated: bool,
+    ) -> BooleanArray {
+        build_in_list_result(values.len(), nulls, self.null_count > 0, negated, |i| {
+            self.check(unsafe { *values.get_unchecked(i) })
+        })
+    }
+}
+
+impl<T: ArrowPrimitiveType, const N: usize> StaticFilter for BranchlessFilter<T, N>
+where
+    T::Native: Copy + PartialEq + Send + Sync,
+{
+    fn null_count(&self) -> usize {
+        self.null_count
+    }
+
+    fn contains(&self, v: &dyn Array, negated: bool) -> Result<BooleanArray> {
+        handle_dictionary!(self, v, negated);
+        let v = v.as_primitive_opt::<T>().ok_or_else(|| {
+            exec_datafusion_err!("Failed to downcast array to primitive type")
+        })?;
+        let input_values = v.values();
+        Ok(build_in_list_result(
+            v.len(),
+            v.nulls(),
+            self.null_count > 0,
+            negated,
+            // SAFETY: i is in bounds since we iterate 0..v.len()
+            #[inline(always)]
+            |i| self.check(unsafe { *input_values.get_unchecked(i) }),
+        ))
+    }
+}
+
 // =============================================================================
 // LEGACY FILTERS (to be replaced by optimized ones in subsequent commits)
 // =============================================================================

datafusion/physical-expr/src/expressions/in_list/strategy.rs

@@ -18,30 +18,125 @@
 use std::sync::Arc;
 
 use arrow::array::ArrayRef;
-use arrow::datatypes::DataType;
-use datafusion_common::Result;
+use arrow::datatypes::*;
+use datafusion_common::{Result, exec_datafusion_err};
 
 use super::array_static_filter::ArrayStaticFilter;
 use super::primitive_filter::*;
 use super::static_filter::StaticFilter;
-use super::transform::make_bitmap_filter;
+use super::transform::{make_bitmap_filter, make_branchless_filter};
 
+// =============================================================================
+// LOOKUP STRATEGY THRESHOLDS (tuned via microbenchmarks)
+// =============================================================================
+
+/// Maximum list size for branchless lookup on 4-byte primitives (Int32, UInt32, Float32).
+const BRANCHLESS_MAX_4B: usize = 32;
+
+/// Maximum list size for branchless lookup on 8-byte primitives (Int64, UInt64, Float64).
+const BRANCHLESS_MAX_8B: usize = 16;
+
+/// Maximum list size for branchless lookup on 16-byte types (Decimal128).
+const BRANCHLESS_MAX_16B: usize = 4;
+
+// =============================================================================
+// FILTER STRATEGY SELECTION
+// =============================================================================
+
+/// The lookup strategy to use for a given data type and list size.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum FilterStrategy {
+    /// Bitmap filter for u8/u16 - O(1) bit test, always fastest for these types.
+    Bitmap1B,
+    Bitmap2B,
+    /// Branchless OR-chain for small lists.
+    Branchless,
+    /// Generic ArrayStaticFilter fallback.
+    Generic,
+}
+
+/// Determines the optimal lookup strategy based on data type and list size.
+fn select_strategy(dt: &DataType, len: usize) -> FilterStrategy {
+    match dt.primitive_width() {
+        Some(1) => FilterStrategy::Bitmap1B,
+        Some(2) => FilterStrategy::Bitmap2B,
+        Some(4) => {
+            if len <= BRANCHLESS_MAX_4B {
+                FilterStrategy::Branchless
+            } else {
+                FilterStrategy::Generic
+            }
+        }
+        Some(8) => {
+            if len <= BRANCHLESS_MAX_8B {
+                FilterStrategy::Branchless
+            } else {
+                FilterStrategy::Generic
+            }
+        }
+        Some(16) => {
+            if len <= BRANCHLESS_MAX_16B {
+                FilterStrategy::Branchless
+            } else {
+                FilterStrategy::Generic
+            }
+        }
+        _ => FilterStrategy::Generic,
+    }
+}
+
+// =============================================================================
+// FILTER INSTANTIATION
+// =============================================================================
+
+/// Creates the optimal static filter for the given array.
 pub(super) fn instantiate_static_filter(
     in_array: ArrayRef,
 ) -> Result<Arc<dyn StaticFilter + Send + Sync>> {
-    match in_array.data_type() {
-        DataType::Int8 | DataType::UInt8 => make_bitmap_filter::<U8Config>(&in_array),
-        DataType::Int16 | DataType::UInt16 => make_bitmap_filter::<U16Config>(&in_array),
-        DataType::Int32 => Ok(Arc::new(Int32StaticFilter::try_new(&in_array)?)),
-        DataType::Int64 => Ok(Arc::new(Int64StaticFilter::try_new(&in_array)?)),
-        DataType::UInt32 => Ok(Arc::new(UInt32StaticFilter::try_new(&in_array)?)),
-        DataType::UInt64 => Ok(Arc::new(UInt64StaticFilter::try_new(&in_array)?)),
-        // Float primitive types (use ordered wrappers for Hash/Eq)
-        DataType::Float32 => Ok(Arc::new(Float32StaticFilter::try_new(&in_array)?)),
-        DataType::Float64 => Ok(Arc::new(Float64StaticFilter::try_new(&in_array)?)),
-        _ => {
-            /* fall through to generic implementation for unsupported types (Struct, etc.) */
-            Ok(Arc::new(ArrayStaticFilter::try_new(in_array)?))
-        }
+    use FilterStrategy::*;
+
+    let len = in_array.len();
+    let dt = in_array.data_type();
+    let strategy = select_strategy(dt, len);
+
+    match (dt, strategy) {
+        // Bitmap filters for 1-byte and 2-byte types
+        (_, Bitmap1B) => make_bitmap_filter::<U8Config>(&in_array),
+        (_, Bitmap2B) => make_bitmap_filter::<U16Config>(&in_array),
+
+        // Branchless filters for small lists of primitives
+        (_, Branchless) => dispatch_branchless(&in_array).ok_or_else(|| {
+            exec_datafusion_err!(
+                "Branchless strategy selected but no filter for {:?}",
+                dt
+            )
+        })?,
+
+        // Fallback for larger primitive lists or complex types.
+        (_, Generic) => match dt {
+            DataType::Int32 => Ok(Arc::new(Int32StaticFilter::try_new(&in_array)?)),
+            DataType::Int64 => Ok(Arc::new(Int64StaticFilter::try_new(&in_array)?)),
+            DataType::UInt32 => Ok(Arc::new(UInt32StaticFilter::try_new(&in_array)?)),
+            DataType::UInt64 => Ok(Arc::new(UInt64StaticFilter::try_new(&in_array)?)),
+            DataType::Float32 => Ok(Arc::new(Float32StaticFilter::try_new(&in_array)?)),
+            DataType::Float64 => Ok(Arc::new(Float64StaticFilter::try_new(&in_array)?)),
+            _ => Ok(Arc::new(ArrayStaticFilter::try_new(in_array)?)),
+        },
+    }
+}
+
+// =============================================================================
+// TYPE DISPATCH
+// =============================================================================
+
+fn dispatch_branchless(
+    arr: &ArrayRef,
+) -> Option<Result<Arc<dyn StaticFilter + Send + Sync>>> {
+    // Dispatch to width-specific branchless filter.
+    match arr.data_type().primitive_width() {
+        Some(4) => Some(make_branchless_filter::<UInt32Type>(arr, 4)),
+        Some(8) => Some(make_branchless_filter::<UInt64Type>(arr, 8)),
+        Some(16) => Some(make_branchless_filter::<Decimal128Type>(arr, 16)),
+        _ => None,
     }
 }