You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@arrow.apache.org by ze...@apache.org on 2026/04/07 16:55:33 UTC

(arrow-go) branch main updated: fix: eliminate per-element heap allocs in is_in kernel for binary types (#745)

This is an automated email from the ASF dual-hosted git repository.

zeroshade pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/arrow-go.git


The following commit(s) were added to refs/heads/main by this push:
     new 3af6b398 fix: eliminate per-element heap allocs in is_in kernel for binary types (#745)
3af6b398 is described below

commit 3af6b3984fde2bfb18de5c2fd6eeacd11c613bff
Author: Matt Topol <zo...@gmail.com>
AuthorDate: Tue Apr 7 12:55:28 2026 -0400

    fix: eliminate per-element heap allocs in is_in kernel for binary types (#745)
    
    ## Summary
    
    - Add `BinaryMemoTable.ExistsDirect` that inlines the hash table probe
    loop, avoiding the closure in `HashTable.Lookup` that causes `val
    []byte` to escape to the heap
    - Add `isInBinaryDirect` specialized kernel path that bypasses the
    `visitBinary` → `VisitBitBlocksShort` closure chain by directly
    iterating with `OptionalBitBlockCounter`
    - Route `BinaryDataType` dispatch in `DispatchIsIn` to the new direct
    path (handles both int32 and int64 offsets)
    
    ## Motivation
    
    The `is_in` kernel for binary types allocated once per input element
    because the `[]byte` value escaped to the heap through a closure chain:
    
    1. `visitBinary` slices `rawBytes[offsets[pos]:offsets[pos+1]]` and
    passes to a callback
    2. The callback calls `BinaryMemoTable.Exists(v)`
    3. `Exists` calls `lookup` which creates a closure capturing `val`
    4. The closure is passed to `HashTable.Lookup`, causing escape analysis
    to move `val` to the heap
    
    Closes #736
    
    ## Benchmark (100k rows, 10-element value set)
    
    | Metric | Before | After | Improvement |
    |--------|--------|-------|-------------|
    | ns/op | 4,133,679 | 923,565 | **4.5x faster** |
    | B/op | 2,435,327 | 33,092 | **73x less memory** |
    | allocs/op | 100,075 | 70 | **1,430x fewer allocs** |
    
    All existing `TestIsInBinary` subtests pass (binary, large\_binary,
    utf8, large\_utf8 × all null matching behaviors).
---
 .../compute/internal/kernels/scalar_set_lookup.go  | 99 +++++++++++++++++++++-
 arrow/compute/scalar_set_lookup_test.go            | 41 +++++++++
 internal/hashing/xxh3_memo_table.go                | 26 ++++++
 3 files changed, 165 insertions(+), 1 deletion(-)

diff --git a/arrow/compute/internal/kernels/scalar_set_lookup.go b/arrow/compute/internal/kernels/scalar_set_lookup.go
index c3562675..fe9841ac 100644
--- a/arrow/compute/internal/kernels/scalar_set_lookup.go
+++ b/arrow/compute/internal/kernels/scalar_set_lookup.go
@@ -235,7 +235,12 @@ func DispatchIsIn(state lookupState, in *exec.ArraySpan, out *exec.ExecResult) e
 
 	switch ty := inType.(type) {
 	case arrow.BinaryDataType:
-		return isInKernelExec(state.(*SetLookupState[[]byte]), in, out)
+		switch ty.Layout().Buffers[1].ByteWidth {
+		case 8:
+			return isInBinaryDirect[int64](state.(*SetLookupState[[]byte]), in, out)
+		default:
+			return isInBinaryDirect[int32](state.(*SetLookupState[[]byte]), in, out)
+		}
 	case arrow.FixedWidthDataType:
 		switch ty.Bytes() {
 		case 1:
@@ -254,6 +259,98 @@ func DispatchIsIn(state lookupState, in *exec.ArraySpan, out *exec.ExecResult) e
 	}
 }
 
+// isInBinaryDirect is a specialized is_in path for binary/string types
+// that avoids the nested closure chain (isInKernelExec -> visitBinary ->
+// VisitBitBlocksShort) which causes per-element heap allocations due to
+// closure escape analysis. It inlines the bit block iteration and calls
+// BinaryMemoTable.ExistsDirect which also avoids closure-based lookup.
+func isInBinaryDirect[OffsetT int32 | int64](state *SetLookupState[[]byte], in *exec.ArraySpan, out *exec.ExecResult) error {
+	if in.Len == 0 {
+		return nil
+	}
+
+	writerBool := bitutil.NewBitmapWriter(out.Buffers[1].Buf, int(out.Offset), int(out.Len))
+	defer writerBool.Finish()
+	writerNulls := bitutil.NewBitmapWriter(out.Buffers[0].Buf, int(out.Offset), int(out.Len))
+	defer writerNulls.Finish()
+
+	valueSetHasNull := state.NullIndex != -1
+	rawBytes := in.Buffers[2].Buf
+	offsets := exec.GetSpanOffsets[OffsetT](in, 1)
+	lookup := state.Lookup.(*hashing.BinaryMemoTable)
+
+	bitmap := in.Buffers[0].Buf
+	counter := bitutils.NewOptionalBitBlockCounter(bitmap, in.Offset, in.Len)
+	pos := int64(0)
+	for pos < in.Len {
+		block := counter.NextBlock()
+		if block.AllSet() {
+			for i := 0; i < int(block.Len); i, pos = i+1, pos+1 {
+				val := rawBytes[offsets[pos]:offsets[pos+1]]
+				if lookup.ExistsDirect(val) {
+					writerBool.Set()
+					writerNulls.Set()
+				} else if state.NullBehavior == NullMatchingInconclusive && valueSetHasNull {
+					writerBool.Clear()
+					writerNulls.Clear()
+				} else {
+					writerBool.Clear()
+					writerNulls.Set()
+				}
+				writerBool.Next()
+				writerNulls.Next()
+			}
+		} else if block.NoneSet() {
+			for i := 0; i < int(block.Len); i, pos = i+1, pos+1 {
+				switch {
+				case state.NullBehavior == NullMatchingMatch && valueSetHasNull:
+					writerBool.Set()
+					writerNulls.Set()
+				case state.NullBehavior == NullMatchingSkip || (!valueSetHasNull && state.NullBehavior == NullMatchingMatch):
+					writerBool.Clear()
+					writerNulls.Set()
+				default:
+					writerBool.Clear()
+					writerNulls.Clear()
+				}
+				writerBool.Next()
+				writerNulls.Next()
+			}
+		} else {
+			for i := 0; i < int(block.Len); i, pos = i+1, pos+1 {
+				if bitutil.BitIsSet(bitmap, int(in.Offset+pos)) {
+					val := rawBytes[offsets[pos]:offsets[pos+1]]
+					if lookup.ExistsDirect(val) {
+						writerBool.Set()
+						writerNulls.Set()
+					} else if state.NullBehavior == NullMatchingInconclusive && valueSetHasNull {
+						writerBool.Clear()
+						writerNulls.Clear()
+					} else {
+						writerBool.Clear()
+						writerNulls.Set()
+					}
+				} else {
+					switch {
+					case state.NullBehavior == NullMatchingMatch && valueSetHasNull:
+						writerBool.Set()
+						writerNulls.Set()
+					case state.NullBehavior == NullMatchingSkip || (!valueSetHasNull && state.NullBehavior == NullMatchingMatch):
+						writerBool.Clear()
+						writerNulls.Set()
+					default:
+						writerBool.Clear()
+						writerNulls.Clear()
+					}
+				}
+				writerBool.Next()
+				writerNulls.Next()
+			}
+		}
+	}
+	return nil
+}
+
 func isInKernelExec[T hashing.MemoTypes](state *SetLookupState[T], in *exec.ArraySpan, out *exec.ExecResult) error {
 	writerBool := bitutil.NewBitmapWriter(out.Buffers[1].Buf, int(out.Offset), int(out.Len))
 	defer writerBool.Finish()
diff --git a/arrow/compute/scalar_set_lookup_test.go b/arrow/compute/scalar_set_lookup_test.go
index 58a666bf..b28c3d75 100644
--- a/arrow/compute/scalar_set_lookup_test.go
+++ b/arrow/compute/scalar_set_lookup_test.go
@@ -18,6 +18,7 @@ package compute_test
 
 import (
 	"context"
+	"fmt"
 	"strings"
 	"sync"
 	"testing"
@@ -639,3 +640,43 @@ func (ss *ScalarSetLookupSuite) TearDownTest() {
 func TestScalarSetLookup(t *testing.T) {
 	suite.Run(t, new(ScalarSetLookupSuite))
 }
+
+func BenchmarkIsInBinary(b *testing.B) {
+	const numRows = 100_000
+	const valueSetSize = 10
+
+	mem := memory.DefaultAllocator
+	ctx := compute.WithAllocator(context.TODO(), mem)
+
+	bldr := array.NewBinaryBuilder(mem, arrow.BinaryTypes.Binary)
+	defer bldr.Release()
+	for i := range numRows {
+		v := []byte(fmt.Sprintf("value-%08d", i%1000))
+		bldr.Append(v)
+	}
+	input := bldr.NewArray()
+	defer input.Release()
+
+	vsBldr := array.NewBinaryBuilder(mem, arrow.BinaryTypes.Binary)
+	defer vsBldr.Release()
+	for i := range valueSetSize {
+		v := []byte(fmt.Sprintf("value-%08d", i))
+		vsBldr.Append(v)
+	}
+	valueSet := vsBldr.NewArray()
+	defer valueSet.Release()
+
+	opts := compute.SetOptions{
+		ValueSet: compute.NewDatumWithoutOwning(valueSet),
+	}
+
+	b.ResetTimer()
+	b.ReportAllocs()
+	for range b.N {
+		result, err := compute.IsIn(ctx, opts, compute.NewDatumWithoutOwning(input))
+		if err != nil {
+			b.Fatal(err)
+		}
+		result.Release()
+	}
+}
diff --git a/internal/hashing/xxh3_memo_table.go b/internal/hashing/xxh3_memo_table.go
index f0ca131a..26b4ce64 100644
--- a/internal/hashing/xxh3_memo_table.go
+++ b/internal/hashing/xxh3_memo_table.go
@@ -225,6 +225,32 @@ func (b *BinaryMemoTable) Exists(val []byte) bool {
 	return ok
 }
 
+// ExistsDirect checks if val exists in the table by inlining the hash
+// table probe loop directly. This avoids the closure allocation that
+// occurs in the normal Exists -> lookup -> HashTable.Lookup path,
+// where the comparison closure captures val and causes it to escape
+// to the heap.
+func (b *BinaryMemoTable) ExistsDirect(val []byte) bool {
+	const perturbShift uint8 = 5
+
+	h := Hash(val, 0)
+	v := b.tbl.fixHash(h)
+	idx := v & b.tbl.capMask
+	perturb := (v >> uint64(perturbShift)) + 1
+
+	for {
+		e := &b.tbl.entries[idx]
+		if e.h == v && bytes.Equal(val, b.builder.Value(int(e.payload.val))) {
+			return true
+		}
+		if e.h == sentinel {
+			return false
+		}
+		idx = (idx + perturb) & b.tbl.capMask
+		perturb = (perturb >> uint64(perturbShift)) + 1
+	}
+}
+
 // Get returns the index of the specified value in the table or KeyNotFound,
 // and a boolean indicating whether it was found in the table.
 func (b *BinaryMemoTable) Get(val interface{}) (int, bool) {