refactor: optimize chunk processing and introduce fast path for no-quote scenarios

Author: nnnkkk7

field_parser.go

@@ -294,12 +294,20 @@ func processChunkMasks(
 	state *parserState, result *parseResult,
 	rowFirstField, lineNum *int,
 ) {
-	for {
-		combined := sepMask | nlMask | quoteMask
-		if combined == 0 {
-			return
-		}
+	combined := sepMask | nlMask | quoteMask
+	if combined == 0 {
+		return
+	}
+
+	// Fast path: no quotes in this chunk and not inside a quoted field.
+	// Avoids quote-related event classification overhead.
+	if quoteMask == 0 && !state.quoted {
+		processChunkMasksNoQuotes(buf, offset, sepMask, nlMask, state, result, rowFirstField, lineNum)
+		return
+	}
 
+	// Standard path: process all structural characters in position order
+	for combined != 0 {
 		pos := bits.TrailingZeros64(combined)
 		bit := uint64(1) << pos
 		absPos := offset + uint64(pos)
@@ -318,6 +326,36 @@ func processChunkMasks(
 			handleNewlineEvent(buf, absPos, state, result, rowFirstField, lineNum)
 			nlMask &^= bit
 		}
+
+		combined = sepMask | nlMask | quoteMask
+	}
+}
+
+// processChunkMasksNoQuotes is a fast path for chunks without quotes.
+// Avoids quote-related checks and event classification overhead.
+func processChunkMasksNoQuotes(
+	buf []byte, offset uint64,
+	sepMask, nlMask uint64,
+	state *parserState, result *parseResult,
+	rowFirstField, lineNum *int,
+) {
+	combined := sepMask | nlMask
+	for combined != 0 {
+		pos := bits.TrailingZeros64(combined)
+		bit := uint64(1) << pos
+		absPos := offset + uint64(pos)
+
+		if sepMask&bit != 0 {
+			// Separator - always record field (not quoted)
+			recordField(buf, absPos, state, result, false)
+			sepMask &^= bit
+		} else {
+			// Newline - record field and row
+			handleNewlineEvent(buf, absPos, state, result, rowFirstField, lineNum)
+			nlMask &^= bit
+		}
+
+		combined = sepMask | nlMask
 	}
 }
 

parse.go

@@ -97,19 +97,23 @@ func buildRecords(buf []byte, pr *parseResult, hasCR bool) [][]string {
 }
 
 // buildRecordZeroCopy creates a record with zero-copy strings from buf.
-// Only safe when no transformation (unescape/CRLF) is needed.
+// Safety: Only call when no transformation (unescape/CRLF) is needed.
+// The returned strings reference buf directly, so buf must outlive the record.
 func buildRecordZeroCopy(buf []byte, pr *parseResult, row rowInfo) []string {
 	if row.fieldCount == 0 {
 		return nil
 	}
+
 	record := make([]string, row.fieldCount)
 	bufLen := uint32(len(buf))
-	for i := 0; i < row.fieldCount; i++ {
-		fieldIdx := row.firstField + i
-		if fieldIdx >= len(pr.fields) {
-			break
-		}
-		field := pr.fields[fieldIdx]
+
+	endIdx := row.firstField + row.fieldCount
+	if endIdx > len(pr.fields) {
+		endIdx = len(pr.fields)
+	}
+	fields := pr.fields[row.firstField:endIdx]
+
+	for i, field := range fields {
 		if field.length == 0 {
 			continue
 		}

record_builder.go

@@ -65,7 +65,8 @@ func (r *Reader) buildRecordWithValidation(row rowInfo, rowIdx int) ([]string, e
 	return r.buildFinalRecord(fieldCount), nil
 }
 
-// buildRecordWithValidationZeroCopy builds a record with zero-copy strings while still validating.
+// buildRecordWithValidationZeroCopy builds a record with zero-copy strings while validating.
+// Zero-copy is safe here because rawBuffer outlives the returned record strings.
 func (r *Reader) buildRecordWithValidationZeroCopy(row rowInfo, fields []fieldInfo) ([]string, error) {
 	fieldCount := row.fieldCount
 	record := r.allocateRecord(fieldCount)
@@ -75,108 +76,126 @@ func (r *Reader) buildRecordWithValidationZeroCopy(row rowInfo, fields []fieldIn
 	bufLen := uint32(len(buf))
 
 	for i, field := range fields {
-		// Validate even in zero-copy path
 		if err := r.validateFieldIfNeeded(field, row.lineNum); err != nil {
 			return record[:i], err
 		}
 
-		start := field.start
-		end := start + field.length
-		if start >= bufLen {
-			record[i] = ""
-		} else {
-			if end > bufLen {
-				end = bufLen
-			}
-			// Zero-copy string from rawBuffer
-			record[i] = unsafe.String(&buf[start], int(end-start))
-		}
+		record[i] = r.extractFieldString(buf, bufLen, field)
 		r.state.fieldPositions[i] = position{line: row.lineNum, column: int(field.rawStart()) + 1}
 	}
 	return record, nil
 }
 
+// extractFieldString returns a zero-copy string for the field content.
+// Returns empty string if field is out of bounds.
+func (r *Reader) extractFieldString(buf []byte, bufLen uint32, field fieldInfo) string {
+	start := field.start
+	end := start + field.length
+	if start >= bufLen {
+		return ""
+	}
+	if end > bufLen {
+		end = bufLen
+	}
+	return unsafe.String(&buf[start], int(end-start))
+}
+
 // buildRecordNoQuotes builds a record when the input contains no quotes.
 // Uses a single row string to avoid per-field allocations.
+// Zero-copy when TrimLeadingSpace is disabled; copies when trimming is needed.
 func (r *Reader) buildRecordNoQuotes(row rowInfo) []string {
 	fieldCount := row.fieldCount
 	record := r.allocateRecord(fieldCount)
 	r.state.fieldPositions = r.ensureFieldPositionsCapacity(fieldCount)
 
 	fields := r.getFieldsForRow(row, fieldCount)
-	buf := r.state.rawBuffer
-	bufLen := uint32(len(buf))
-
 	if len(fields) == 0 {
 		return record
 	}
 
+	buf := r.state.rawBuffer
+	bufLen := uint32(len(buf))
+
+	// Calculate row span in buffer
 	rowStart := fields[0].rawStart()
 	rowEnd := fields[len(fields)-1].rawEnd()
+
+	// Handle out-of-bounds row
 	if rowStart >= bufLen {
 		for i, field := range fields {
 			record[i] = ""
 			r.state.fieldPositions[i] = position{line: row.lineNum, column: int(field.rawStart()) + 1}
 		}
 		return record
 	}
-	if rowEnd > bufLen {
-		rowEnd = bufLen
+
+	// Clamp row bounds
+	rowEnd = clampUint32(rowEnd, rowStart, bufLen)
+
+	// Create row string (copy if trimming, zero-copy otherwise)
+	rowStr := r.createRowString(buf, rowStart, rowEnd)
+	rowStrLen := len(rowStr)
+
+	// Extract fields from row string
+	for i, field := range fields {
+		record[i] = r.extractFieldFromRow(buf, bufLen, rowStr, rowStrLen, rowStart, field)
+		r.state.fieldPositions[i] = position{line: row.lineNum, column: int(field.rawStart()) + 1}
+	}
+	return record
+}
+
+// clampUint32 clamps value to [minVal, maxVal].
+func clampUint32(value, minVal, maxVal uint32) uint32 {
+	if value < minVal {
+		return minVal
 	}
-	if rowEnd < rowStart {
-		rowEnd = rowStart
+	if value > maxVal {
+		return maxVal
 	}
+	return value
+}
 
-	var rowStr string
+// createRowString creates a string for the row span.
+// Copies when TrimLeadingSpace is enabled; zero-copy otherwise.
+func (r *Reader) createRowString(buf []byte, rowStart, rowEnd uint32) string {
 	if r.TrimLeadingSpace {
-		rowStr = string(buf[rowStart:rowEnd])
-	} else {
-		// Zero-copy string from rawBuffer - safe because rawBuffer outlives record.
-		rowStr = unsafe.String(&buf[rowStart], int(rowEnd-rowStart))
+		return string(buf[rowStart:rowEnd])
 	}
-	rowStrLen := len(rowStr)
-
-	for i, field := range fields {
-		start := field.start
-		end := start + field.length
-		rawStart := field.rawStart()
+	return unsafe.String(&buf[rowStart], int(rowEnd-rowStart))
+}
 
-		if start < bufLen {
-			if end > bufLen {
-				end = bufLen
-			}
-			if r.TrimLeadingSpace && start < end {
-				for start < end && (buf[start] == ' ' || buf[start] == '\t') {
-					start++
-				}
-			}
-		}
+// extractFieldFromRow extracts a field string from the row string.
+func (r *Reader) extractFieldFromRow(buf []byte, bufLen uint32, rowStr string, rowStrLen int, rowStart uint32, field fieldInfo) string {
+	start := field.start
+	end := start + field.length
 
-		if start < rowStart {
-			start = rowStart
-		}
-		if end < start {
-			end = start
-		}
-		relStart := int(start - rowStart)
-		relEnd := int(end - rowStart)
-		if relStart < 0 {
-			relStart = 0
+	// Apply trimming if needed
+	if r.TrimLeadingSpace && start < bufLen && start < end {
+		trimEnd := end
+		if trimEnd > bufLen {
+			trimEnd = bufLen
 		}
-		if relStart > rowStrLen {
-			relStart = rowStrLen
-		}
-		if relEnd < relStart {
-			relEnd = relStart
-		}
-		if relEnd > rowStrLen {
-			relEnd = rowStrLen
+		for start < trimEnd && (buf[start] == ' ' || buf[start] == '\t') {
+			start++
 		}
+	}
 
-		record[i] = rowStr[relStart:relEnd]
-		r.state.fieldPositions[i] = position{line: row.lineNum, column: int(rawStart) + 1}
+	// Calculate relative positions in row string
+	relStart := clampInt(int(start)-int(rowStart), 0, rowStrLen)
+	relEnd := clampInt(int(end)-int(rowStart), relStart, rowStrLen)
+
+	return rowStr[relStart:relEnd]
+}
+
+// clampInt clamps value to [minVal, maxVal].
+func clampInt(value, minVal, maxVal int) int {
+	if value < minVal {
+		return minVal
 	}
-	return record
+	if value > maxVal {
+		return maxVal
+	}
+	return value
 }
 
 // getFieldsForRow extracts the slice of fieldInfo for the given row.

simd_scanner.go

@@ -101,6 +101,8 @@ var scanResultPool = sync.Pool{
 }
 
 // scanResultLargeCache retains a single large scanResult across GC cycles.
+// This prevents repeated large allocations when processing files > 1MB,
+// as sync.Pool may evict large objects during GC.
 var scanResultLargeCache struct {
 	mu sync.Mutex
 	sr *scanResult
@@ -123,20 +125,28 @@ func (sr *scanResult) reset() {
 }
 
 // releaseScanResult returns a scanResult to the pool for reuse.
+// Large results (>= scanResultLargeThreshold) are cached separately to survive GC.
 func releaseScanResult(sr *scanResult) {
-	if sr != nil {
-		sr.reset()
-		if cap(sr.quoteMasks) >= scanResultLargeThreshold {
-			scanResultLargeCache.mu.Lock()
-			if scanResultLargeCache.sr == nil || cap(scanResultLargeCache.sr.quoteMasks) < cap(sr.quoteMasks) {
-				scanResultLargeCache.sr = sr
-				scanResultLargeCache.mu.Unlock()
-				return
-			}
+	if sr == nil {
+		return
+	}
+
+	sr.reset()
+
+	// Cache large results separately to prevent GC eviction
+	if cap(sr.quoteMasks) >= scanResultLargeThreshold {
+		scanResultLargeCache.mu.Lock()
+		shouldCache := scanResultLargeCache.sr == nil ||
+			cap(scanResultLargeCache.sr.quoteMasks) < cap(sr.quoteMasks)
+		if shouldCache {
+			scanResultLargeCache.sr = sr
 			scanResultLargeCache.mu.Unlock()
+			return
 		}
-		scanResultPool.Put(sr)
+		scanResultLargeCache.mu.Unlock()
 	}
+
+	scanResultPool.Put(sr)
 }
 
 // =============================================================================

validation.go

@@ -2,6 +2,8 @@
 
 package simdcsv
 
+import "bytes"
+
 // =============================================================================
 // Validation Policy - Configurable behavior decisions
 // =============================================================================
@@ -164,24 +166,14 @@ func (r *Reader) validateQuotedField(raw []byte, rawStart uint64, lineNum int) e
 // validateUnquotedField validates a field that does not start with a quote.
 // Reports ErrBareQuote if quotes appear in unquoted fields.
 func (r *Reader) validateUnquotedField(raw []byte, rawStart uint64, lineNum int) error {
-	quotePos := findBareQuote(raw)
+	quotePos := bytes.IndexByte(raw, '"')
 	if quotePos == -1 {
 		return nil
 	}
 	col := int(rawStart) + quotePos + 1 //nolint:gosec // G115
 	return &ParseError{StartLine: lineNum, Line: lineNum, Column: col, Err: ErrBareQuote}
 }
 
-// findBareQuote returns the index of the first quote in data, or -1 if none found.
-func findBareQuote(data []byte) int {
-	for i, b := range data {
-		if b == '"' {
-			return i
-		}
-	}
-	return -1
-}
-
 // =============================================================================
 // Quote Structure Validators - Single responsibility functions
 // =============================================================================