fix(unstable): correct Parser.Range for non-suffix highlights

Parser.Range used len(data)-len(highlight), which only matches suffix
slices. Single-byte highlights like b[0:1] are subslices of the
remaining buffer, so the wrong offset pointed at the end of the
document (issue #1047). Use pointer-based subslice offset like
wrapDecodeError.

Add regression tests for unstable.Parser and toml.Unmarshal error
positions and human-readable output.

Co-authored-by: Thomas Pelletier <thomas@pelletier.dev>

.github/workflows/cifuzz.yml

@@ -19,7 +19,7 @@ jobs:
         dry-run: false
         language: go
     - name: Upload Crash
-      uses: actions/upload-artifact@v6
+      uses: actions/upload-artifact@v7
       if: failure() && steps.build.outcome == 'success'
       with:
         name: artifacts

.github/workflows/coverage.yml

@@ -15,6 +15,6 @@ jobs:
       - name: Setup go
         uses: actions/setup-go@v6
         with:
-          go-version: "1.25"
+          go-version: "1.26"
       - name: Run tests with coverage
         run: ./ci.sh coverage -d "${GITHUB_BASE_REF-HEAD}"

.github/workflows/go-versions-test.yml

@@ -20,7 +20,7 @@ jobs:
           fetch-depth: 0
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
+        uses: docker/setup-buildx-action@v4
 
       - name: Run Go versions compatibility test
         run: |
@@ -28,7 +28,7 @@ jobs:
           ./test-go-versions.sh --output ./test-results $VERSIONS
 
       - name: Upload test results
-        uses: actions/upload-artifact@v6
+        uses: actions/upload-artifact@v7
         with:
           name: go-versions-test-results
           path: |

.github/workflows/lint.yml

@@ -15,7 +15,7 @@ jobs:
       - name: Setup go
         uses: actions/setup-go@v6
         with:
-          go-version: "1.24"
+          go-version: "1.26"
       - name: Run golangci-lint
         uses: golangci/golangci-lint-action@v9
         with:

.github/workflows/release.yml

@@ -22,15 +22,15 @@ jobs:
       - name: Set up Go
         uses: actions/setup-go@v6
         with:
-          go-version: "1.25"
+          go-version: "1.26"
       - name: Login to GitHub Container Registry
-        uses: docker/login-action@v3
+        uses: docker/login-action@v4
         with:
           registry: ghcr.io
           username: ${{ github.actor }}
           password: ${{ secrets.GITHUB_TOKEN }}
       - name: Run GoReleaser
-        uses: goreleaser/goreleaser-action@v6
+        uses: goreleaser/goreleaser-action@v7
         with:
           distribution: goreleaser
           version: '~> v2'

.github/workflows/workflow.yml

@@ -10,9 +10,10 @@ on:
 jobs:
   build:
     strategy:
+      fail-fast: false
       matrix:
         os: [ 'ubuntu-latest', 'windows-latest', 'macos-latest', 'macos-14' ]
-        go: [ '1.24', '1.25' ]
+        go: [ '1.25', '1.26' ]
     runs-on: ${{ matrix.os }}
     name: ${{ matrix.go }}/${{ matrix.os }}
     steps:

.goreleaser.yaml

@@ -22,7 +22,6 @@ builds:
       - linux_riscv64
       - windows_amd64
       - windows_arm64
-      - windows_arm
       - darwin_amd64
       - darwin_arm64
   - id: tomljson
@@ -42,7 +41,6 @@ builds:
       - linux_riscv64
       - windows_amd64
       - windows_arm64
-      - windows_arm
       - darwin_amd64
       - darwin_arm64
   - id: jsontoml
@@ -62,7 +60,6 @@ builds:
       - linux_arm
       - windows_amd64
       - windows_arm64
-      - windows_arm
       - darwin_amd64
       - darwin_arm64
 universal_binaries:

README.md

@@ -239,12 +239,12 @@ Execution time speedup compared to other Go TOML libraries:
         <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>
+        <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>
@@ -257,17 +257,17 @@ provided for completeness.</p>
         <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>
+        <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>

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]

errors_test.go

@@ -202,6 +202,38 @@ func TestDecodeError_Accessors(t *testing.T) {
 	assert.Equal(t, "bar", e.String())
 }
 
+func TestDecodeError_InvalidKeyStartAfterComment(t *testing.T) {
+	// Regression for https://github.com/pelletier/go-toml/issues/1047: the "="
+	// that starts an invalid keyval must be reported on line 2, column 1, with
+	// the human-readable context pointing at that byte (not the document end).
+	doc := "# comment\n= \"value\""
+
+	var v map[string]any
+	err := Unmarshal([]byte(doc), &v)
+	if err == nil {
+		t.Fatal("expected an error")
+	}
+
+	var derr *DecodeError
+	if !errors.As(err, &derr) {
+		t.Fatalf("expected *DecodeError, got %T", err)
+	}
+
+	row, col := derr.Position()
+	if row != 2 || col != 1 {
+		t.Errorf("Position(): got row %d col %d, want row 2 col 1", row, col)
+	}
+
+	human := derr.String()
+	if !strings.Contains(human, `2| = "value"`) {
+		t.Errorf("human output should show the error line; got:\n%s", human)
+	}
+	// Caret line uses line-number column width padding; only the "| ~" part is stable here.
+	if !strings.Contains(human, "| ~ invalid character at start of key") {
+		t.Errorf("human output should underline '=' and include the parser message; got:\n%s", human)
+	}
+}
+
 func TestDecodeError_DuplicateContent(t *testing.T) {
 	// This test verifies that when the same content appears multiple times
 	// in the document, the error correctly points to the actual location
@@ -259,6 +291,12 @@ func TestDecodeError_Position(t *testing.T) {
 			expectedRow: 3,
 			minCol:      5,
 		},
+		{
+			name:        "missing equals on last line without trailing newline",
+			doc:         "a = 1\nb = 2\nc",
+			expectedRow: 3,
+			minCol:      1,
+		},
 	}
 
 	for _, e := range examples {

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
 		}
 

test-go-versions.sh

@@ -9,7 +9,7 @@ YELLOW='\033[1;33m'
 BLUE='\033[0;34m'
 NC='\033[0m' # No Color
 
-# Go versions to test (1.11 through 1.25)
+# Go versions to test (1.11 through 1.26)
 GO_VERSIONS=(
     "1.11"
     "1.12"
@@ -26,6 +26,7 @@ GO_VERSIONS=(
     "1.23"
     "1.24"
     "1.25"
+    "1.26"
 )
 
 # Default values
@@ -64,7 +65,7 @@ EXAMPLES:
     $0                          # Test all Go versions in parallel
     $0 --sequential             # Test all Go versions sequentially
     $0 1.21 1.22 1.23          # Test specific versions
-    $0 --verbose --output ./results 1.24 1.25  # Verbose output to custom directory
+    $0 --verbose --output ./results 1.25 1.26  # Verbose output to custom directory
 
 EXIT CODES:
     0                   Recent Go versions pass (good compatibility)
@@ -136,8 +137,8 @@ fi
 
 # Validate Go versions
 for version in "${GO_VERSIONS[@]}"; do
-    if ! [[ "$version" =~ ^1\.(1[1-9]|2[0-5])$ ]]; then
-        log_error "Invalid Go version: $version. Supported versions: 1.11-1.25"
+    if ! [[ "$version" =~ ^1\.(1[1-9]|2[0-6])$ ]]; then
+        log_error "Invalid Go version: $version. Supported versions: 1.11-1.26"
         exit 1
     fi
 done

unmarshaler.go

@@ -56,13 +56,18 @@ func (d *Decoder) DisallowUnknownFields() *Decoder {
 
 // EnableUnmarshalerInterface allows to enable unmarshaler interface.
 //
-// With this feature enabled, types implementing the unstable/Unmarshaler
+// With this feature enabled, types implementing the unstable.Unmarshaler
 // interface can be decoded from any structure of the document. It allows types
 // that don't have a straightforward TOML representation to provide their own
 // decoding logic.
 //
-// Currently, types can only decode from a single value. Tables and array tables
-// are not supported.
+// The UnmarshalTOML method receives raw TOML bytes:
+//   - For single values: the raw value bytes (e.g., `"hello"` for a string)
+//   - For tables: all key-value lines belonging to that table
+//   - For inline tables/arrays: the raw bytes of the inline structure
+//
+// The unstable.RawMessage type can be used to capture raw TOML bytes for
+// later processing, similar to json.RawMessage.
 //
 // *Unstable:* This method does not follow the compatibility guarantees of
 // semver. It can be changed or removed without a new major version being
@@ -599,9 +604,8 @@ func (d *decoder) handleArrayTablePart(key unstable.Iterator, v reflect.Value) (
 // cannot handle it.
 func (d *decoder) handleTable(key unstable.Iterator, v reflect.Value) (reflect.Value, error) {
 	if v.Kind() == reflect.Slice {
-		if v.Len() == 0 {
-			return reflect.Value{}, unstable.NewParserError(key.Node().Data, "cannot store a table in a slice")
-		}
+		// For non-empty slices, work with the last element
+		if v.Len() > 0 {
 			elem := v.Index(v.Len() - 1)
 			x, err := d.handleTable(key, elem)
 			if err != nil {
@@ -612,6 +616,17 @@ func (d *decoder) handleTable(key unstable.Iterator, v reflect.Value) (reflect.V
 			}
 			return reflect.Value{}, nil
 		}
+		// Empty slice - check if it implements Unmarshaler (e.g., RawMessage)
+		// and we're at the end of the key path
+		if d.unmarshalerInterface && !key.Next() {
+			if v.CanAddr() && v.Addr().CanInterface() {
+				if outi, ok := v.Addr().Interface().(unstable.Unmarshaler); ok {
+					return d.handleKeyValuesUnmarshaler(outi)
+				}
+			}
+		}
+		return reflect.Value{}, unstable.NewParserError(key.Node().Data, "cannot store a table in a slice")
+	}
 	if key.Next() {
 		// Still scoping the key
 		return d.handleTablePart(key, v)
@@ -624,6 +639,24 @@ func (d *decoder) handleTable(key unstable.Iterator, v reflect.Value) (reflect.V
 // Handle root expressions until the end of the document or the next
 // non-key-value.
 func (d *decoder) handleKeyValues(v reflect.Value) (reflect.Value, error) {
+	// Check if target implements Unmarshaler before processing key-values.
+	// This allows types to handle entire tables themselves.
+	if d.unmarshalerInterface {
+		vv := v
+		for vv.Kind() == reflect.Ptr {
+			if vv.IsNil() {
+				vv.Set(reflect.New(vv.Type().Elem()))
+			}
+			vv = vv.Elem()
+		}
+		if vv.CanAddr() && vv.Addr().CanInterface() {
+			if outi, ok := vv.Addr().Interface().(unstable.Unmarshaler); ok {
+				// Collect all key-value expressions for this table
+				return d.handleKeyValuesUnmarshaler(outi)
+			}
+		}
+	}
+
 	var rv reflect.Value
 	for d.nextExpr() {
 		expr := d.expr()
@@ -653,6 +686,41 @@ func (d *decoder) handleKeyValues(v reflect.Value) (reflect.Value, error) {
 	return rv, nil
 }
 
+// handleKeyValuesUnmarshaler collects all key-value expressions for a table
+// and passes them to the Unmarshaler as raw TOML bytes.
+func (d *decoder) handleKeyValuesUnmarshaler(u unstable.Unmarshaler) (reflect.Value, error) {
+	// Collect raw bytes from all key-value expressions for this table.
+	// We use the Raw field on each KeyValue expression to preserve the
+	// original formatting (whitespace, quoting style, etc.) from the document.
+	var buf []byte
+
+	for d.nextExpr() {
+		expr := d.expr()
+		if expr.Kind != unstable.KeyValue {
+			d.stashExpr()
+			break
+		}
+
+		_, err := d.seen.CheckExpression(expr)
+		if err != nil {
+			return reflect.Value{}, err
+		}
+
+		// Use the raw bytes from the original document to preserve formatting
+		if expr.Raw.Length > 0 {
+			raw := d.p.Raw(expr.Raw)
+			buf = append(buf, raw...)
+		}
+		buf = append(buf, '\n')
+	}
+
+	if err := u.UnmarshalTOML(buf); err != nil {
+		return reflect.Value{}, err
+	}
+
+	return reflect.Value{}, nil
+}
+
 type (
 	handlerFn    func(key unstable.Iterator, v reflect.Value) (reflect.Value, error)
 	valueMakerFn func() reflect.Value
@@ -697,7 +765,8 @@ func (d *decoder) handleValue(value *unstable.Node, v reflect.Value) error {
 	if d.unmarshalerInterface {
 		if v.CanAddr() && v.Addr().CanInterface() {
 			if outi, ok := v.Addr().Interface().(unstable.Unmarshaler); ok {
-				return outi.UnmarshalTOML(value)
+				// Pass raw bytes from the original document
+				return outi.UnmarshalTOML(d.p.Raw(value.Raw))
 			}
 		}
 	}
@@ -1201,7 +1270,8 @@ func (d *decoder) handleKeyValuePart(key unstable.Iterator, value *unstable.Node
 			if d.unmarshalerInterface {
 				if v.CanAddr() && v.Addr().CanInterface() {
 					if outi, ok := v.Addr().Interface().(unstable.Unmarshaler); ok {
-						return reflect.Value{}, outi.UnmarshalTOML(value)
+						// Pass raw bytes from the original document
+						return reflect.Value{}, outi.UnmarshalTOML(d.p.Raw(value.Raw))
 					}
 				}
 			}

unmarshaler_test.go

@@ -96,6 +96,132 @@ func ExampleUnmarshal() {
 	// tags: [go toml]
 }
 
+// pluginConfig demonstrates how to implement dynamic unmarshaling
+// based on a "type" field. This pattern is useful for plugin systems
+// or polymorphic configuration.
+type pluginConfig struct {
+	Type   string
+	Config any
+}
+
+func (p *pluginConfig) UnmarshalTOML(data []byte) error {
+	// First, decode just the type field
+	var typeOnly struct {
+		Type string `toml:"type"`
+	}
+	if err := toml.Unmarshal(data, &typeOnly); err != nil {
+		return err
+	}
+	p.Type = typeOnly.Type
+
+	// Now decode the config based on the type
+	switch typeOnly.Type {
+	case "database":
+		var cfg struct {
+			Type string `toml:"type"`
+			Host string `toml:"host"`
+			Port int    `toml:"port"`
+		}
+		if err := toml.Unmarshal(data, &cfg); err != nil {
+			return err
+		}
+		p.Config = map[string]any{"host": cfg.Host, "port": cfg.Port}
+	case "cache":
+		var cfg struct {
+			Type string `toml:"type"`
+			TTL  int    `toml:"ttl"`
+		}
+		if err := toml.Unmarshal(data, &cfg); err != nil {
+			return err
+		}
+		p.Config = map[string]any{"ttl": cfg.TTL}
+	}
+	return nil
+}
+
+// This example demonstrates dynamic unmarshaling based on a discriminator
+// field. The pluginConfig type uses UnmarshalTOML to first read the "type"
+// field, then decode the rest of the configuration based on that type.
+// This pattern is useful for plugin systems or configuration that varies
+// by type.
+func ExampleDecoder_EnableUnmarshalerInterface_dynamicConfig() {
+	doc := `
+[[plugins]]
+type = "database"
+host = "localhost"
+port = 5432
+
+[[plugins]]
+type = "cache"
+ttl = 300
+`
+	type Config struct {
+		Plugins []pluginConfig `toml:"plugins"`
+	}
+
+	var cfg Config
+	err := toml.NewDecoder(strings.NewReader(doc)).
+		EnableUnmarshalerInterface().
+		Decode(&cfg)
+	if err != nil {
+		panic(err)
+	}
+
+	for _, p := range cfg.Plugins {
+		fmt.Printf("type=%s config=%v\n", p.Type, p.Config)
+	}
+	// Output:
+	// type=database config=map[host:localhost port:5432]
+	// type=cache config=map[ttl:300]
+}
+
+// This example demonstrates using RawMessage to capture raw TOML bytes
+// for later processing. RawMessage is similar to json.RawMessage - it
+// delays decoding so you can inspect the raw content or decode it
+// differently based on context.
+func ExampleDecoder_EnableUnmarshalerInterface_rawMessage() {
+	doc := `
+[plugin]
+name = "example"
+version = "1.0"
+enabled = true
+`
+
+	type Config struct {
+		Plugin unstable.RawMessage `toml:"plugin"`
+	}
+
+	var cfg Config
+	err := toml.NewDecoder(strings.NewReader(doc)).
+		EnableUnmarshalerInterface().
+		Decode(&cfg)
+	if err != nil {
+		panic(err)
+	}
+
+	// cfg.Plugin contains the raw TOML bytes
+	fmt.Printf("Raw TOML captured:\n%s", cfg.Plugin)
+
+	// You can later decode it into a specific type
+	var plugin struct {
+		Name    string `toml:"name"`
+		Version string `toml:"version"`
+		Enabled bool   `toml:"enabled"`
+	}
+	if err := toml.Unmarshal(cfg.Plugin, &plugin); err != nil {
+		panic(err)
+	}
+	fmt.Printf("Decoded: name=%s version=%s enabled=%v\n",
+		plugin.Name, plugin.Version, plugin.Enabled)
+
+	// Output:
+	// Raw TOML captured:
+	// name = "example"
+	// version = "1.0"
+	// enabled = true
+	// Decoded: name=example version=1.0 enabled=true
+}
+
 type badReader struct{}
 
 func (r *badReader) Read([]byte) (int, error) {
@@ -3900,8 +4026,8 @@ type CustomUnmarshalerKey struct {
 	A int64
 }
 
-func (k *CustomUnmarshalerKey) UnmarshalTOML(value *unstable.Node) error {
-	item, err := strconv.ParseInt(string(value.Data), 10, 64)
+func (k *CustomUnmarshalerKey) UnmarshalTOML(data []byte) error {
+	item, err := strconv.ParseInt(string(data), 10, 64)
 	if err != nil {
 		return fmt.Errorf("error converting to int64, %w", err)
 	}
@@ -3989,7 +4115,7 @@ foo = "bar"`,
 
 type doc994 struct{}
 
-func (d *doc994) UnmarshalTOML(*unstable.Node) error {
+func (d *doc994) UnmarshalTOML([]byte) error {
 	return errors.New("expected-error")
 }
 
@@ -4012,8 +4138,8 @@ type doc994ok struct {
 	S string
 }
 
-func (d *doc994ok) UnmarshalTOML(value *unstable.Node) error {
-	d.S = string(value.Data) + " from unmarshaler"
+func (d *doc994ok) UnmarshalTOML(data []byte) error {
+	d.S = string(data) + " from unmarshaler"
 	return nil
 }
 
@@ -4026,7 +4152,8 @@ func TestIssue994_OK(t *testing.T) {
 		Decode(&d)
 
 	assert.NoError(t, err)
-	assert.Equal(t, "bar from unmarshaler", d.S)
+	// With bytes-based interface, raw TOML bytes are passed including quotes
+	assert.Equal(t, "\"bar\" from unmarshaler", d.S)
 }
 
 func TestIssue995(t *testing.T) {
@@ -4385,3 +4512,265 @@ func TestIssue1028(t *testing.T) {
 		assert.Error(t, err)
 	})
 }
+
+// Tests for issue #873 - Bring back toml.Unmarshaler for tables and arrays
+
+type customTable873 struct {
+	Keys   []string
+	Values map[string]string
+}
+
+func (c *customTable873) UnmarshalTOML(data []byte) error {
+	c.Keys = []string{}
+	c.Values = make(map[string]string)
+
+	// Parse the raw TOML bytes into a map to extract keys in order
+	// For this test, we use a simple line-by-line parser to preserve order
+	lines := bytes.Split(data, []byte{'\n'})
+	for _, line := range lines {
+		line = bytes.TrimSpace(line)
+		if len(line) == 0 {
+			continue
+		}
+		// Skip table headers
+		if line[0] == '[' {
+			continue
+		}
+		// Parse key = value
+		eqIdx := bytes.Index(line, []byte{'='})
+		if eqIdx < 0 {
+			continue
+		}
+		key := string(bytes.TrimSpace(line[:eqIdx]))
+		// Remove quotes from quoted keys
+		if len(key) >= 2 && key[0] == '"' && key[len(key)-1] == '"' {
+			key = key[1 : len(key)-1]
+		}
+		valueBytes := bytes.TrimSpace(line[eqIdx+1:])
+		// Remove quotes from string values
+		if len(valueBytes) >= 2 && valueBytes[0] == '"' && valueBytes[len(valueBytes)-1] == '"' {
+			valueBytes = valueBytes[1 : len(valueBytes)-1]
+		}
+		c.Keys = append(c.Keys, key)
+		c.Values[key] = string(valueBytes)
+	}
+
+	return nil
+}
+
+// Test for split tables - when the same parent table is defined in multiple places
+// This is a key requirement for issue #873: if type A implements Unmarshaler,
+// and [a.b] and [a.d] are defined with another table [x] in between,
+// A should receive content for both b and d, but not x.
+func TestIssue873_SplitTables(t *testing.T) {
+	// For this test, we expect each sub-table to be handled separately
+	// The parent doesn't receive the sub-tables directly - each sub-table
+	// (b and d) gets its own call to handleKeyValues
+	type Config struct {
+		A struct {
+			B customTable873 `toml:"b"`
+			D customTable873 `toml:"d"`
+		} `toml:"a"`
+		X customTable873 `toml:"x"`
+	}
+
+	doc := `
+[a.b]
+C = "1"
+
+[x]
+Y = "100"
+
+[a.d]
+E = "2"
+`
+
+	var cfg Config
+	err := toml.NewDecoder(bytes.NewReader([]byte(doc))).
+		EnableUnmarshalerInterface().
+		Decode(&cfg)
+
+	assert.NoError(t, err)
+	// Each sub-table should have received its own key-values
+	assert.Equal(t, []string{"C"}, cfg.A.B.Keys)
+	assert.Equal(t, "1", cfg.A.B.Values["C"])
+	assert.Equal(t, []string{"E"}, cfg.A.D.Keys)
+	assert.Equal(t, "2", cfg.A.D.Values["E"])
+	assert.Equal(t, []string{"Y"}, cfg.X.Keys)
+	assert.Equal(t, "100", cfg.X.Values["Y"])
+}
+
+// Test using RawMessage to capture raw TOML bytes
+func TestIssue873_RawMessage(t *testing.T) {
+	type Config struct {
+		Plugin unstable.RawMessage `toml:"plugin"`
+	}
+
+	doc := `
+[plugin]
+name = "example"
+version = "1.0"
+`
+
+	var cfg Config
+	err := toml.NewDecoder(bytes.NewReader([]byte(doc))).
+		EnableUnmarshalerInterface().
+		Decode(&cfg)
+
+	assert.NoError(t, err)
+	// RawMessage should contain the raw key-value bytes
+	expected := "name = \"example\"\nversion = \"1.0\"\n"
+	assert.Equal(t, expected, string(cfg.Plugin))
+}
+
+// Test keys that need quoting (contain special characters)
+func TestIssue873_QuotedKeys(t *testing.T) {
+	type Config struct {
+		Section customTable873 `toml:"section"`
+	}
+
+	doc := `
+[section]
+"key with spaces" = "value1"
+"key.with.dots" = "value2"
+`
+
+	var cfg Config
+	err := toml.NewDecoder(bytes.NewReader([]byte(doc))).
+		EnableUnmarshalerInterface().
+		Decode(&cfg)
+
+	assert.NoError(t, err)
+	assert.Equal(t, 2, len(cfg.Section.Keys))
+	assert.Equal(t, "value1", cfg.Section.Values["key with spaces"])
+	assert.Equal(t, "value2", cfg.Section.Values["key.with.dots"])
+}
+
+// errorUnmarshaler873 is used to test error propagation from UnmarshalTOML
+type errorUnmarshaler873 struct{}
+
+func (e *errorUnmarshaler873) UnmarshalTOML([]byte) error {
+	return errors.New("intentional error")
+}
+
+// Test error propagation from UnmarshalTOML
+func TestIssue873_UnmarshalerError(t *testing.T) {
+	doc := `
+[section]
+key = "value"
+`
+
+	type Config struct {
+		Section errorUnmarshaler873 `toml:"section"`
+	}
+
+	var cfg Config
+	err := toml.NewDecoder(bytes.NewReader([]byte(doc))).
+		EnableUnmarshalerInterface().
+		Decode(&cfg)
+
+	assert.Error(t, err)
+	assert.True(t, strings.Contains(err.Error(), "intentional error"))
+}
+
+// Test dotted keys in a table (e.g., a.b = value)
+func TestIssue873_DottedKeys(t *testing.T) {
+	type Config struct {
+		Section customTable873 `toml:"section"`
+	}
+
+	doc := `
+[section]
+sub.key = "value1"
+another.nested.key = "value2"
+`
+
+	var cfg Config
+	err := toml.NewDecoder(bytes.NewReader([]byte(doc))).
+		EnableUnmarshalerInterface().
+		Decode(&cfg)
+
+	assert.NoError(t, err)
+	assert.Equal(t, 2, len(cfg.Section.Keys))
+	// The dotted keys should be preserved in the raw output
+	assert.Equal(t, "value1", cfg.Section.Values["sub.key"])
+	assert.Equal(t, "value2", cfg.Section.Values["another.nested.key"])
+}
+
+// Test pointer to pointer to Unmarshaler (covers pointer dereferencing loop)
+func TestIssue873_DoublePointerUnmarshaler(t *testing.T) {
+	type Config struct {
+		Section **customTable873 `toml:"section"`
+	}
+
+	doc := `
+[section]
+key = "value"
+`
+
+	var cfg Config
+	err := toml.NewDecoder(bytes.NewReader([]byte(doc))).
+		EnableUnmarshalerInterface().
+		Decode(&cfg)
+
+	assert.NoError(t, err)
+	assert.True(t, cfg.Section != nil)
+	assert.True(t, *cfg.Section != nil)
+	assert.Equal(t, []string{"key"}, (*cfg.Section).Keys)
+	assert.Equal(t, "value", (*cfg.Section).Values["key"])
+}
+
+// formattingCapture captures the raw TOML bytes to verify formatting preservation
+type formattingCapture struct {
+	RawBytes string
+}
+
+func (f *formattingCapture) UnmarshalTOML(data []byte) error {
+	f.RawBytes = string(data)
+	return nil
+}
+
+func TestIssue873_FormattingPreservation(t *testing.T) {
+	type Config struct {
+		Section *formattingCapture `toml:"section"`
+	}
+
+	// Test that various formatting styles are preserved:
+	// - Extra spaces around '='
+	// - Literal strings (single quotes)
+	// - Hex numbers
+	// - Inline tables
+	doc := `[section]
+key1   =   "value with spaces"
+key2 = 'literal string'
+hex_val = 0xDEADBEEF
+inline = { a = 1, b = 2 }
+`
+
+	var cfg Config
+	err := toml.NewDecoder(bytes.NewReader([]byte(doc))).
+		EnableUnmarshalerInterface().
+		Decode(&cfg)
+
+	assert.NoError(t, err)
+	assert.True(t, cfg.Section != nil)
+
+	// The raw bytes should preserve original formatting
+	raw := cfg.Section.RawBytes
+
+	// Check that extra spaces around '=' are preserved
+	assert.True(t, strings.Contains(raw, "key1   =   \"value with spaces\""),
+		"Expected spacing to be preserved, got: %s", raw)
+
+	// Check that literal string style is preserved
+	assert.True(t, strings.Contains(raw, "key2 = 'literal string'"),
+		"Expected literal string to be preserved, got: %s", raw)
+
+	// Check that hex format is preserved
+	assert.True(t, strings.Contains(raw, "hex_val = 0xDEADBEEF"),
+		"Expected hex format to be preserved, got: %s", raw)
+
+	// Check that inline table is preserved
+	assert.True(t, strings.Contains(raw, "inline = { a = 1, b = 2 }"),
+		"Expected inline table to be preserved, got: %s", raw)
+}

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

@@ -3,6 +3,7 @@ package unstable
 import (
 	"bytes"
 	"fmt"
+	"reflect"
 	"unicode"
 
 	"github.com/pelletier/go-toml/v2/internal/characters"
@@ -69,7 +70,7 @@ func (p *Parser) Data() []byte {
 // panics.
 func (p *Parser) Range(b []byte) Range {
 	return Range{
-		Offset: uint32(p.subsliceOffset(b)), //nolint:gosec // TOML documents are small
+		Offset: uint32(subsliceOffset(p.data, b)), //nolint:gosec // TOML documents are small
 		Length: uint32(len(b)),                    //nolint:gosec // TOML documents are small
 	}
 }
@@ -82,11 +83,21 @@ func (p *Parser) rangeOfToken(token, rest []byte) Range {
 	return Range{Offset: uint32(offset), Length: uint32(len(token))} //nolint:gosec // TOML documents are small
 }
 
-// subsliceOffset returns the byte offset of subslice b within p.data.
-// b must be a suffix (tail) of p.data.
-func (p *Parser) subsliceOffset(b []byte) int {
-	// b is a suffix of p.data, so its offset is len(p.data) - len(b)
-	return len(p.data) - len(b)
+// subsliceOffset returns the byte offset of subslice b within data.
+// b must share the same backing array as data (any subslice of data).
+func subsliceOffset(data, b []byte) int {
+	if len(b) == 0 {
+		return 0
+	}
+
+	dataPtr := reflect.ValueOf(data).Pointer()
+	bPtr := reflect.ValueOf(b).Pointer()
+
+	offset := int(bPtr - dataPtr)
+	if offset < 0 || offset > len(data) {
+		panic("subslice is not within data")
+	}
+	return offset
 }
 
 // Raw returns the slice corresponding to the bytes in the given range.
@@ -328,6 +339,9 @@ func (p *Parser) parseStdTable(b []byte) (reference, []byte, error) {
 
 func (p *Parser) parseKeyval(b []byte) (reference, []byte, error) {
 	// keyval = key keyval-sep val
+	// Track the start position for Raw range
+	startB := b
+
 	ref := p.builder.Push(Node{
 		Kind: KeyValue,
 	})
@@ -342,7 +356,7 @@ func (p *Parser) parseKeyval(b []byte) (reference, []byte, error) {
 	b = p.parseWhitespace(b)
 
 	if len(b) == 0 {
-		return invalidReference, nil, NewParserError(b, "expected = after a key, but the document ends there")
+		return invalidReference, nil, NewParserError(startB[:len(startB)-len(b)], "expected = after a key, but the document ends there")
 	}
 
 	b, err = expect('=', b)
@@ -360,6 +374,11 @@ 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.
+	// 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
 }
 

unstable/parser_range_test.go

@@ -0,0 +1,36 @@
+package unstable
+
+import (
+	"errors"
+	"testing"
+)
+
+// Regression test for https://github.com/pelletier/go-toml/issues/1047:
+// Parser.Range must use the real slice offset, not len(data)-len(slice).
+func TestParser_Range_HighlightAfterComment(t *testing.T) {
+	input := []byte("# comment\n= \"value\"")
+
+	var p Parser
+	p.Reset(input)
+	for p.NextExpression() {
+	}
+	err := p.Error()
+	if err == nil {
+		t.Fatal("expected an error")
+	}
+
+	var perr *ParserError
+	if !errors.As(err, &perr) {
+		t.Fatalf("expected *ParserError, got %T", err)
+	}
+
+	r := p.Range(perr.Highlight)
+	shape := p.Shape(r)
+
+	if r.Offset != 10 {
+		t.Errorf("Range offset: got %d, want 10", r.Offset)
+	}
+	if shape.Start.Line != 2 || shape.Start.Column != 1 {
+		t.Errorf("position: got %d:%d, want 2:1", shape.Start.Line, shape.Start.Column)
+	}
+}

unstable/parser_test.go

@@ -539,7 +539,7 @@ key5 = [ # Next to start of inline array.
 	// ---
 	// 6:1->6:22 (105->126)      | Comment [# Above simple value.]
 	// ---
-	// 1:1->1:1 (0->0)           | KeyValue []
+	// 7:1->7:14 (127->140)      | KeyValue []
 	// 7:7->7:14 (133->140)      |   String [value]
 	// 7:1->7:4 (127->130)       |   Key [key]
 	// 7:15->7:38 (141->164)     | Comment [# Next to simple value.]
@@ -552,12 +552,12 @@ key5 = [ # Next to start of inline array.
 	// ---
 	// 14:1->14:22 (252->273)    | Comment [# Above inline table.]
 	// ---
-	// 1:1->1:1 (0->0)           | KeyValue []
+	// 15:1->15:50 (274->323)    | KeyValue []
 	// 15:8->15:9 (281->282)     |   InlineTable []
-	// 1:1->1:1 (0->0)           |     KeyValue []
+	// 15:10->15:23 (283->296)   |     KeyValue []
 	// 15:18->15:23 (291->296)   |       String [Tom]
 	// 15:10->15:15 (283->288)   |       Key [first]
-	// 1:1->1:1 (0->0)           |     KeyValue []
+	// 15:25->15:48 (298->321)   |     KeyValue []
 	// 15:32->15:48 (305->321)   |       String [Preston-Werner]
 	// 15:25->15:29 (298->302)   |       Key [last]
 	// 15:1->15:5 (274->278)     |   Key [name]
@@ -567,7 +567,7 @@ key5 = [ # Next to start of inline array.
 	// ---
 	// 18:1->18:15 (371->385)    | Comment [# Above array.]
 	// ---
-	// 1:1->1:1 (0->0)           | KeyValue []
+	// 19:1->19:20 (386->405)    | KeyValue []
 	// 1:1->1:1 (0->0)           |   Array []
 	// 19:11->19:12 (396->397)   |     Integer [1]
 	// 19:14->19:15 (399->400)   |     Integer [2]
@@ -579,7 +579,7 @@ key5 = [ # Next to start of inline array.
 	// ---
 	// 22:1->22:26 (448->473)    | Comment [# Above multi-line array.]
 	// ---
-	// 1:1->1:1 (0->0)           | KeyValue []
+	// 23:1->31:2 (474->694)     | KeyValue []
 	// 1:1->1:1 (0->0)           |   Array []
 	// 23:10->23:42 (483->515)   |     Comment [# Next to start of inline array.]
 	// 24:3->24:38 (518->553)    |       Comment [# Second line before array content.]

unstable/unmarshaler.go

@@ -1,7 +1,32 @@
 package unstable
 
-// The Unmarshaler interface may be implemented by types to customize their
-// behavior when being unmarshaled from a TOML document.
+// Unmarshaler is implemented by types that can unmarshal a TOML
+// description of themselves. The input is a valid TOML document
+// containing the relevant portion of the parsed document.
+//
+// For tables (including split tables defined in multiple places),
+// the data contains the raw key-value bytes from the original document
+// with adjusted table headers to be relative to the unmarshaling target.
 type Unmarshaler interface {
-	UnmarshalTOML(value *Node) error
+	UnmarshalTOML(data []byte) error
+}
+
+// RawMessage is a raw encoded TOML value. It implements Unmarshaler
+// and can be used to delay TOML decoding or capture raw content.
+//
+// Example usage:
+//
+//	type Config struct {
+//	    Plugin RawMessage `toml:"plugin"`
+//	}
+//
+//	var cfg Config
+//	toml.NewDecoder(r).EnableUnmarshalerInterface().Decode(&cfg)
+//	// cfg.Plugin now contains the raw TOML bytes for [plugin]
+type RawMessage []byte
+
+// UnmarshalTOML implements Unmarshaler.
+func (m *RawMessage) UnmarshalTOML(data []byte) error {
+	*m = append((*m)[0:0], data...)
+	return nil
 }