Reduce marshal and unmarshal overhead (#1044)

* Reduce marshal and unmarshal overhead

Targeted optimizations to reduce performance overhead introduced by
recent feature additions and the unsafe removal.

Unmarshal:
- parseKeyval: access the node directly in the builder's slice to set
  Raw, bypassing NodeAt which triggers a GC write barrier for the
  nodes-pointer on every key-value expression.
- Iterator.Next: cache the *nodes slice dereference in a local variable
  to avoid repeated pointer-to-slice indirection in the hot loop.

Marshal:
- Guard shouldOmitZero calls with an inlineable options.omitzero check.
  shouldOmitZero has inlining cost 1145 (budget 80), so avoiding the
  function call when omitzero is not set removes per-field overhead.
- Inline the isNil check in encodeMap. isNil has inlining cost 93
  (budget 80), so expanding it at the single hot call site avoids
  per-map-entry function call overhead.

Update README benchmarks.

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

README.md

@@ -235,17 +235,17 @@ the AST level. See https://pkg.go.dev/github.com/pelletier/go-toml/v2/unstable.
 Execution time speedup compared to other Go TOML libraries:
 
 <table>
-	<thead>
-		<tr><th>Benchmark</th><th>go-toml v1</th><th>BurntSushi/toml</th></tr>
-	</thead>
-	<tbody>
-		<tr><td>Marshal/HugoFrontMatter-2</td><td>1.9x</td><td>2.2x</td></tr>
-		<tr><td>Marshal/ReferenceFile/map-2</td><td>1.7x</td><td>2.1x</td></tr>
-		<tr><td>Marshal/ReferenceFile/struct-2</td><td>2.2x</td><td>3.0x</td></tr>
-		<tr><td>Unmarshal/HugoFrontMatter-2</td><td>2.9x</td><td>2.7x</td></tr>
-		<tr><td>Unmarshal/ReferenceFile/map-2</td><td>2.6x</td><td>2.7x</td></tr>
-		<tr><td>Unmarshal/ReferenceFile/struct-2</td><td>4.6x</td><td>5.1x</td></tr>
-	 </tbody>
+    <thead>
+        <tr><th>Benchmark</th><th>go-toml v1</th><th>BurntSushi/toml</th></tr>
+    </thead>
+    <tbody>
+        <tr><td>Marshal/HugoFrontMatter-2</td><td>2.1x</td><td>2.0x</td></tr>
+        <tr><td>Marshal/ReferenceFile/map-2</td><td>2.0x</td><td>2.0x</td></tr>
+        <tr><td>Marshal/ReferenceFile/struct-2</td><td>2.3x</td><td>2.5x</td></tr>
+        <tr><td>Unmarshal/HugoFrontMatter-2</td><td>3.3x</td><td>2.8x</td></tr>
+        <tr><td>Unmarshal/ReferenceFile/map-2</td><td>2.9x</td><td>3.0x</td></tr>
+        <tr><td>Unmarshal/ReferenceFile/struct-2</td><td>4.8x</td><td>5.0x</td></tr>
+     </tbody>
 </table>
 <details><summary>See more</summary>
 <p>The table above has the results of the most common use-cases. The table below
@@ -253,22 +253,22 @@ contains the results of all benchmarks, including unrealistic ones. It is
 provided for completeness.</p>
 
 <table>
-	<thead>
-		<tr><th>Benchmark</th><th>go-toml v1</th><th>BurntSushi/toml</th></tr>
-	</thead>
-	<tbody>
-		<tr><td>Marshal/SimpleDocument/map-2</td><td>1.8x</td><td>2.7x</td></tr>
-		<tr><td>Marshal/SimpleDocument/struct-2</td><td>2.7x</td><td>3.8x</td></tr>
-		<tr><td>Unmarshal/SimpleDocument/map-2</td><td>3.8x</td><td>3.0x</td></tr>
-		<tr><td>Unmarshal/SimpleDocument/struct-2</td><td>5.6x</td><td>4.1x</td></tr>
-		<tr><td>UnmarshalDataset/example-2</td><td>3.0x</td><td>3.2x</td></tr>
-		<tr><td>UnmarshalDataset/code-2</td><td>2.3x</td><td>2.9x</td></tr>
-		<tr><td>UnmarshalDataset/twitter-2</td><td>2.6x</td><td>2.7x</td></tr>
-		<tr><td>UnmarshalDataset/citm_catalog-2</td><td>2.2x</td><td>2.3x</td></tr>
-		<tr><td>UnmarshalDataset/canada-2</td><td>1.8x</td><td>1.5x</td></tr>
-		<tr><td>UnmarshalDataset/config-2</td><td>4.1x</td><td>2.9x</td></tr>
-		<tr><td>geomean</td><td>2.7x</td><td>2.8x</td></tr>
-	 </tbody>
+    <thead>
+        <tr><th>Benchmark</th><th>go-toml v1</th><th>BurntSushi/toml</th></tr>
+    </thead>
+    <tbody>
+        <tr><td>Marshal/SimpleDocument/map-2</td><td>2.0x</td><td>2.9x</td></tr>
+        <tr><td>Marshal/SimpleDocument/struct-2</td><td>2.5x</td><td>3.6x</td></tr>
+        <tr><td>Unmarshal/SimpleDocument/map-2</td><td>4.2x</td><td>3.4x</td></tr>
+        <tr><td>Unmarshal/SimpleDocument/struct-2</td><td>5.9x</td><td>4.4x</td></tr>
+        <tr><td>UnmarshalDataset/example-2</td><td>3.2x</td><td>2.9x</td></tr>
+        <tr><td>UnmarshalDataset/code-2</td><td>2.4x</td><td>2.8x</td></tr>
+        <tr><td>UnmarshalDataset/twitter-2</td><td>2.7x</td><td>2.5x</td></tr>
+        <tr><td>UnmarshalDataset/citm_catalog-2</td><td>2.3x</td><td>2.3x</td></tr>
+        <tr><td>UnmarshalDataset/canada-2</td><td>1.9x</td><td>1.5x</td></tr>
+        <tr><td>UnmarshalDataset/config-2</td><td>5.4x</td><td>3.0x</td></tr>
+        <tr><td>geomean</td><td>2.9x</td><td>2.8x</td></tr>
+     </tbody>
 </table>
 <p>This table can be generated with <code>./ci.sh benchmark -a -html</code>.</p>
 </details>

ci.sh

@@ -147,7 +147,7 @@ bench() {
     pushd "$dir"
 
     if [ "${replace}" != "" ]; then
-        find ./benchmark/ -iname '*.go' -exec sed -i -E "s|github.com/pelletier/go-toml/v2|${replace}|g" {} \;
+        find ./benchmark/ -iname '*.go' -exec sed -i -E "s|github.com/pelletier/go-toml/v2\"|${replace}\"|g" {} \;
         go get "${replace}"
     fi
 
@@ -195,6 +195,11 @@ for line in reversed(lines[2:]):
         "%.1fx" % (float(line[3])/v2),  # v1
         "%.1fx" % (float(line[7])/v2),  # bs
     ])
+
+if not results:
+    print("No benchmark results to display.", file=sys.stderr)
+    sys.exit(1)
+
 # move geomean to the end
 results.append(results[0])
 del results[0]

marshaler.go

@@ -704,15 +704,18 @@ func (enc *Encoder) encodeMap(b []byte, ctx encoderCtx, v reflect.Value) ([]byte
 	for iter.Next() {
 		v := iter.Value()
 
-		if isNil(v) {
-			// For nil pointers, convert to zero value of the element type.
-			// This allows round-trip marshaling of maps with nil pointer values.
-			// For nil interfaces and nil maps, skip since we can't derive a type.
-			if v.Kind() == reflect.Ptr {
+		// Handle nil values: convert nil pointers to zero value,
+		// skip nil interfaces and nil maps.
+		switch v.Kind() {
+		case reflect.Ptr:
+			if v.IsNil() {
 				v = reflect.Zero(v.Type().Elem())
-			} else {
+			}
+		case reflect.Interface, reflect.Map:
+			if v.IsNil() {
 				continue
 			}
+		default:
 		}
 
 		k, err := enc.keyToString(iter.Key())
@@ -936,7 +939,7 @@ func (enc *Encoder) encodeTable(b []byte, ctx encoderCtx, t table) ([]byte, erro
 		if shouldOmitEmpty(kv.Options, kv.Value) {
 			continue
 		}
-		if shouldOmitZero(kv.Options, kv.Value) {
+		if kv.Options.omitzero && shouldOmitZero(kv.Options, kv.Value) {
 			continue
 		}
 		hasNonEmptyKV = true
@@ -958,7 +961,7 @@ func (enc *Encoder) encodeTable(b []byte, ctx encoderCtx, t table) ([]byte, erro
 		if shouldOmitEmpty(table.Options, table.Value) {
 			continue
 		}
-		if shouldOmitZero(table.Options, table.Value) {
+		if table.Options.omitzero && shouldOmitZero(table.Options, table.Value) {
 			continue
 		}
 		if first {
@@ -995,7 +998,7 @@ func (enc *Encoder) encodeTableInline(b []byte, ctx encoderCtx, t table) ([]byte
 		if shouldOmitEmpty(kv.Options, kv.Value) {
 			continue
 		}
-		if shouldOmitZero(kv.Options, kv.Value) {
+		if kv.Options.omitzero && shouldOmitZero(kv.Options, kv.Value) {
 			continue
 		}
 

unstable/ast.go

@@ -28,12 +28,16 @@ func (c *Iterator) Next() bool {
 	if c.nodes == nil {
 		return false
 	}
+	nodes := *c.nodes
 	if !c.started {
 		c.started = true
-	} else if c.idx >= 0 {
-		c.idx = (*c.nodes)[c.idx].next
+	} else {
+		idx := c.idx
+		if idx >= 0 && int(idx) < len(nodes) {
+			c.idx = nodes[idx].next
+		}
 	}
-	return c.idx >= 0 && int(c.idx) < len(*c.nodes)
+	return c.idx >= 0 && int(c.idx) < len(nodes)
 }
 
 // IsLast returns true if the current node of the iterator is the last

unstable/parser.go

@@ -363,9 +363,10 @@ func (p *Parser) parseKeyval(b []byte) (reference, []byte, error) {
 	p.builder.Chain(valRef, key)
 	p.builder.AttachChild(ref, valRef)
 
-	// Set Raw to span the entire key-value expression
-	node := p.builder.NodeAt(ref)
-	node.Raw = p.rangeOfToken(startB[:len(startB)-len(b)], b)
+	// Set Raw to span the entire key-value expression.
+	// Access the node directly in the slice to avoid the write barrier
+	// that NodeAt's nodes-pointer setup would trigger.
+	p.builder.tree.nodes[ref].Raw = p.rangeOfToken(startB[:len(startB)-len(b)], b)
 
 	return ref, b, err
 }