perf(mustache): Improved performance.

AGENTS.md

@@ -32,6 +32,7 @@ thor/
 ├── frontmatter.odin    # JSON frontmatter parser (supports nested og + lastmod)
 ├── defaults.odin       # DEFAULTS_PATH constant (#directory)
 ├── main.odin           # Entry point
+├── bench/              # Template rendering benchmark
 └── defaults/layouts/   # Bundled default templates
 ```
 
@@ -64,6 +65,7 @@ thor/
 | | `sectionate.odin` | `wrap_sections` — splits HTML at `<h2>` into `<section>` wrappers |
 | | `highlight.odin` | Syntax highlighting via tree-sitter. Imports `../treesitter`. |
 | `mustache/` | See [Mustache engine](#mustache-engine) below | Template engine |
+| `bench/` | `bench.odin` + `templates/` | Standalone template rendering benchmark. Generates 500 posts + 100 comments, renders with indented partials + inheritance + pipes. `--dump <path>` for output validation, positional arg for iteration count (default 250). |
 
 Icon SVGs live as HTML partials in `layouts/partials/icons/` (home, github, rss, chevron_up, star).
 
@@ -333,6 +335,16 @@ odin test .                  # main package tests (site, frontmatter)
 odin test . -all-packages    # includes mustache specs, lambdas, pipes, diagnostics, markdown tests
 ```
 
+### Benchmark
+
+```bash
+cd thor
+odin build bench -o:speed
+./bench.bin                    # 250 iterations, prints timing
+./bench.bin --dump output.html # render once, write to file for diff validation
+./bench.bin 1000               # custom iteration count
+```
+
 ## Mustache engine
 
 Spec-compliant implementation at `mustache/`. See `mustache/SPEC.md` for the implementation specification, `mustache/EXTENSIONS.md` for non-standard extensions (pipes), and `mustache/diagnostic.odin` for the rust-style error formatter.
@@ -341,11 +353,11 @@ Spec-compliant implementation at `mustache/`. See `mustache/SPEC.md` for the imp
 
 | File | Responsibility |
 |---|---|
-| `mustache.odin` | Public API (`parse`, `render`, `Template`), parser (`parse_section`), renderer (`render_nodes`, takes `Template` by value), template inheritance (`merge_block_overrides`), `delete_template`/`delete_partials`. Pipe support in Variable/Unescaped/Section/Inverted tags. |
+| `mustache.odin` | Public API (`parse`, `render`, `Template`), parser (`parse_section`), renderer (`render_nodes` with `Indent_State` for partial indentation), template inheritance (`merge_block_overrides`), `delete_template`/`delete_partials`. Pipe support in Variable/Unescaped/Section/Inverted tags. |
 | `tokenizer.odin` | Tokenizer (template string → `[]Token`), standalone whitespace detection |
 | `data.odin` | Reflection-based data model: `base_value` (peels union/any/nested-any layers), `lookup_in` (structs + maps, handles `Type_Info_Any` value kind in maps), `resolve_name`, `is_truthy`, `any_to_string`, `list_info`, `extract_list_element`, `call_interp_lambda`/`call_section_lambda` |
 | `pipes.odin` | Pipes extension: `Pipe_Filter` AST, `parse_pipeline` (takes `pos`), `apply_pipeline`, `apply_filter` (switch dispatch: `group_by` + `format`), `apply_group_by`, `apply_format`. Stored on `Node.filters`; render-scoped results in temp allocator. |
-| `diagnostic.odin` | Rust-style error formatter: `format_error` (multi-line context, ANSI colors via `core:terminal/ansi`, `colorize` param), `format_render_error` (dispatch on `Render_Error`), `line_col`, `line_text`, `context_extent`, `count_lines`, `digit_count`, `should_colorize`. |
+| `diagnostic.odin` | Rust-style error formatter: `format_error` (multi-line context, ANSI colors via `core:terminal/ansi`, `colorize` param), `format_render_error` (formats `Error`), `line_col`, `line_text`, `context_extent`, `count_lines`, `digit_count`, `should_colorize`. |
 | `suggest.odin` | Strict-warning helpers: `validate_key_path` (walks dotted path, crosses maps silently), `suggest_correction` (Levenshtein via `core:strings/levenshtein_distance`), `collect_struct_keys` (via reflection, recurses into `using`), `struct_has_field` (distinguishes missing field from nil value — needed for `Maybe(bool)`), `collect_partial_names`, `collect_block_names`. |
 | `spec_test.odin` | JSON spec test runner — loads `spec/specs/*.json`, runs each test case. Uses `log.nil_logger()` to suppress expected warnings. |
 | `lambda_test.odin` | Spec lambda tests |
@@ -361,7 +373,7 @@ render(tmpl, data, partials) → render_nodes (walks flat node array against con
 ```
 
 - **Two-phase API**: `parse()` produces a reusable `Template`, `render()` walks it against data. Templates parsed once, rendered many times.
-- **Flat `[dynamic]Node` array** with `first_child`/`child_count` indices — pre-order layout. Each `Node` carries `pos: int` (byte offset into source) for diagnostics.
+- **Flat `[dynamic]Node` array** with `children: []Node` slices (pre-order layout; slices point into the backing array). Each `Node` carries `pos: int` (byte offset into source) for diagnostics.
 - **`Template`** carries `source` and `path` — used by diagnostics to show file location and source context.
 - **Context stack**: `^[dynamic]any` with `append`/`pop` for section push/pop.
 - **`render_nodes` takes `Template` by value** (not `^Template`) — Odin's calling convention promotes to pointer when efficient. Eliminates "local copy" patterns at call sites.
@@ -370,12 +382,13 @@ render(tmpl, data, partials) → render_nodes (walks flat node array against con
 - **`lookup_in`** resolves keys on structs (via `reflect.struct_field_value_by_name` with `allow_using = true`) and maps. Detects `Type_Info_Any` value kind in maps and reads the inner any directly to avoid double-wrap.
 - **Template inheritance**: `{{<parent}}` loads parent from partials, `{{$block}}` defines overridable sections. `merge_block_overrides` propagates overrides through multi-level chains.
 - **Dynamic partial names**: `{{>*key}}` resolves partial name from data context at render time.
+- **Render-time partial indentation**: `Indent_State` threads `at_line_start` through `render_nodes` so partial indent is applied at render time (via `write_indented` on Text nodes) instead of reparsing the partial's source. `render_template` writes initial indent, creates state, calls `render_nodes`. Data-injected newlines don't pick up indent (Variables don't update `at_line_start`).
 
 ### Diagnostics
 
 Rust-style error messages with multi-line source context, caret underlines, and Levenshtein suggestions. ANSI colors via `core:terminal/ansi`, gated on `should_colorize()` (TTY detection on stderr).
 
-**Error types**: `Syntax_Error{msg, pos}` and `Data_Error{msg, pos}` — both carry byte offset into template source. (`Partial_Error` was removed — dead code.)
+**Error types**: `Error_Body{msg, pos, kind}` where `kind` is `Error_Kind.Syntax` (parse-time) or `Error_Kind.Data` (render-time). `Error` is a single-variant union wrapping `Error_Body` (nilable for `!= nil` / `or_return`).
 
 **Strict-by-default warnings** — `render_nodes` emits `log.warnf` diagnostics for:
 - Unknown keys in `{{k}}`, `{{{k}}}`, `{{#k}}`, `{{^k}}` (via `validate_key_path` + `suggest_correction`)
@@ -424,7 +437,6 @@ See `mustache/EXTENSIONS.md`.
 
 You may never, *ever* remove `TODO:` or `FIXME:` comments. Those are for humans, not machines.
 See `HUGO.md` for analysis of why thor doesn't need Hugo's shortcode context isolation.
-See `mustache/PARTIAL_INDENT.md` for whitespace handling analysis.
 See `mustache/SPEC.md` for the original implementation specification.
 See `mustache/EXTENSIONS.md` for non-standard extensions (pipes).
 

mustache/mustache.odin

@@ -81,6 +81,14 @@ Block_Override :: struct {
 	source: Template,
 }
 
+// Indent_State threads partial-indent tracking through render_nodes so
+// the renderer can apply indentation at render time instead of reparsing
+// the partial's source with indentation baked in.
+Indent_State :: struct {
+	indent:        string,
+	at_line_start: bool,
+}
+
 delete_template :: proc(tmpl: ^Template) {
 	if tmpl != nil && len(tmpl.nodes) > 0 {
 		delete(tmpl.nodes)
@@ -393,20 +401,29 @@ find_common_indent :: proc(children: []Node) -> string {
 			text := children[i].text
 			if len(text) > 0 {
 				line_start := 0
-				for j in 0 ..= len(text) {
-					if j == len(text) || text[j] == '\n' {
-						line := text[line_start:j]
-						if len(strings.trim_space(line)) > 0 {
-							ws := leading_whitespace(line)
-							if !found {
-								common = ws
-								found = true
-							} else if len(ws) < len(common) {
-								common = ws
-							}
-						}
-						line_start = j + 1
+				for {
+					// Bulk-scan to next newline (AVX2-backed) instead of byte-by-byte.
+					rel := strings.index_byte(text[line_start:], '\n')
+					line_end: int
+					if rel < 0 {
+						line_end = len(text)
+					} else {
+						line_end = line_start + rel
 					}
+					line := text[line_start:line_end]
+					if len(strings.trim_space(line)) > 0 {
+						ws := leading_whitespace(line)
+						if !found {
+							common = ws
+							found = true
+						} else if len(ws) < len(common) {
+							common = ws
+						}
+					}
+					if rel < 0 {
+						break
+					}
+					line_start = line_end + 1
 				}
 			}
 		}
@@ -442,16 +459,19 @@ remove_line_indent :: proc(s: string, indent: string, allocator := context.alloc
 		if at_line_start {
 			if i + len(indent) <= len(s) && s[i:i + len(indent)] == indent {
 				i += len(indent)
-				at_line_start = false
-				continue
 			}
 			at_line_start = false
 		}
-		append(&buf, s[i])
-		if s[i] == '\n' {
-			at_line_start = true
+		// Bulk-append up to and including the next newline (AVX2-backed).
+		rel := strings.index_byte(s[i:], '\n')
+		if rel < 0 {
+			append(&buf, s[i:])
+			break
 		}
-		i += 1
+		next := i + rel
+		append(&buf, s[i:next + 1])
+		i = next + 1
+		at_line_start = true
 	}
 
 	return string(buf[:])
@@ -461,34 +481,6 @@ remove_line_indent :: proc(s: string, indent: string, allocator := context.alloc
 // Indentation helpers
 // ---------------------------------------------------------------------------
 
-indent_lines :: proc(source: string, indent: string) -> string {
-	if len(indent) == 0 || len(source) == 0 {
-		return source
-	}
-	b: strings.Builder
-	strings.builder_init(&b, context.temp_allocator)
-	write_indented(&b, indent, source)
-	return strings.to_string(b)
-}
-
-write_indented :: proc(b: ^strings.Builder, indent: string, content: string) {
-	if len(indent) == 0 || len(content) == 0 {
-		strings.write_string(b, content)
-		return
-	}
-	at_line_start := true
-	for i in 0 ..< len(content) {
-		if at_line_start {
-			strings.write_string(b, indent)
-			at_line_start = false
-		}
-		strings.write_byte(b, content[i])
-		if content[i] == '\n' {
-			at_line_start = true
-		}
-	}
-}
-
 render_template :: proc(
 	pt: Template,
 	ctx: ^[dynamic]any,
@@ -497,21 +489,36 @@ render_template :: proc(
 	blocks: map[string]Block_Override,
 	indent: string,
 ) -> Error {
-	if len(indent) > 0 && len(pt.source) > 0 {
-		// Per Mustache spec: the partial's source is indented before rendering,
-		// not its output. This is necessary so that data-injected newlines
-		// (e.g. from `{{{content}}}` where content contains `\n`) do NOT pick
-		// up the indent — only source-level line breaks do.
-		indented := indent_lines(pt.source, indent)
-		reparse := parse(
-			indented,
-			pt.path,
-			context.temp_allocator,
-			context.temp_allocator,
-		) or_return
-		return render_nodes(reparse, reparse.nodes[:], ctx, partials, b, blocks)
+	if len(indent) > 0 {
+		state := Indent_State{indent = indent, at_line_start = false}
+		strings.write_string(b, indent) // first line always gets indent
+		return render_nodes(pt, pt.nodes[:], ctx, partials, b, blocks, &state)
+	}
+	return render_nodes(pt, pt.nodes[:], ctx, partials, b, blocks, nil)
+}
+
+write_indented :: proc(b: ^strings.Builder, indent: string, content: string, at_line_start: ^bool) {
+	if len(indent) == 0 || len(content) == 0 {
+		strings.write_string(b, content)
+		return
+	}
+	i := 0
+	for i < len(content) {
+		if at_line_start^ {
+			strings.write_string(b, indent)
+			at_line_start^ = false
+		}
+		// Bulk-write up to and including the next newline (AVX2-backed).
+		rel := strings.index_byte(content[i:], '\n')
+		if rel < 0 {
+			strings.write_string(b, content[i:])
+			return
+		}
+		next := i + rel
+		strings.write_string(b, content[i:next + 1])
+		i = next + 1
+		at_line_start^ = true
 	}
-	return render_nodes(pt, pt.nodes[:], ctx, partials, b, blocks)
 }
 
 // ---------------------------------------------------------------------------
@@ -525,16 +532,25 @@ render_nodes :: proc(
 	partials: map[string]Template,
 	b: ^strings.Builder,
 	blocks: map[string]Block_Override = nil,
+	indent_state: ^Indent_State = nil,
 ) -> Error {
 	i := 0
 	for i < len(nodes) {
 		node := nodes[i]
 		switch node.kind {
 		case .Text:
-			strings.write_string(b, node.text)
+			if indent_state != nil {
+				write_indented(b, indent_state.indent, node.text, &indent_state.at_line_start)
+			} else {
+				strings.write_string(b, node.text)
+			}
 			i += 1
 
 		case .Variable:
+			if indent_state != nil && indent_state.at_line_start {
+				strings.write_string(b, indent_state.indent)
+				indent_state.at_line_start = false
+			}
 			val := resolve_name(node.key, ctx[:])
 			if val == nil {
 				warn_unknown_key(current, ctx[:], node)
@@ -563,6 +579,7 @@ render_nodes :: proc(
 					partials,
 					&temp,
 					blocks,
+					nil,
 				) or_return
 				write_value(b, strings.to_string(temp), escape = true)
 				}
@@ -572,6 +589,10 @@ render_nodes :: proc(
 			i += 1
 
 		case .Unescaped:
+			if indent_state != nil && indent_state.at_line_start {
+				strings.write_string(b, indent_state.indent)
+				indent_state.at_line_start = false
+			}
 			val := resolve_name(node.key, ctx[:])
 			if val == nil {
 				warn_unknown_key(current, ctx[:], node)
@@ -600,6 +621,7 @@ render_nodes :: proc(
 					partials,
 					&temp,
 					blocks,
+					nil,
 				) or_return
 				write_value(b, strings.to_string(temp), escape = false)
 				}
@@ -635,6 +657,7 @@ render_nodes :: proc(
 					partials,
 					b,
 					blocks,
+					nil,
 				) or_return
 				}
 		} else if is_truthy(val) {
@@ -652,12 +675,13 @@ render_nodes :: proc(
 						partials,
 						b,
 						blocks,
+						indent_state,
 					) or_return
 				}
 			} else {
 				append(ctx, val)
 				defer pop(ctx)
-				render_nodes(current, children, ctx, partials, b, blocks) or_return
+				render_nodes(current, children, ctx, partials, b, blocks, indent_state) or_return
 			}
 		}
 		i += 1 + len(node.children)
@@ -675,7 +699,7 @@ render_nodes :: proc(
 				val = transformed
 			}
 		if !is_truthy(val) {
-			render_nodes(current, node.children, ctx, partials, b, blocks) or_return
+			render_nodes(current, node.children, ctx, partials, b, blocks, indent_state) or_return
 		}
 		i += 1 + len(node.children)
 
@@ -690,6 +714,9 @@ render_nodes :: proc(
 				warn_missing_partial(current, partials, node, name)
 			} else {
 				render_template(pt, ctx, partials, b, nil, node.indent) or_return
+				if indent_state != nil {
+					indent_state.at_line_start = false
+				}
 			}
 			i += 1
 
@@ -720,8 +747,10 @@ render_nodes :: proc(
 				partials,
 				&temp,
 				content_blocks,
+				nil,
 			) or_return
-			write_indented(b, node.indent, strings.to_string(temp))
+			at_ls := true
+			write_indented(b, node.indent, strings.to_string(temp), &at_ls)
 		} else {
 			render_nodes(
 				render_current,
@@ -730,6 +759,7 @@ render_nodes :: proc(
 				partials,
 				b,
 				content_blocks,
+				indent_state,
 			) or_return
 		}
 		i += 1 + len(node.children)
@@ -743,6 +773,9 @@ render_nodes :: proc(
 		} else {
 			warn_unmatched_block_overrides(current, pt, parent_children)
 			render_template(pt, ctx, partials, b, merged, node.indent) or_return
+			if indent_state != nil {
+				indent_state.at_line_start = false
+			}
 		}
 		i += 1 + len(node.children)
 		}

mustache/tokenizer.odin

@@ -36,60 +36,73 @@ tokenize :: proc(
 	text_start := 0
 
 	for i < len(src) {
-		if src[i] == '{' && i + 1 < len(src) && src[i + 1] == '{' {
-			if i > text_start {
-				append(&tokens, Token{kind = .Text, value = src[text_start:i], pos = text_start})
+		// Find next '{' via AVX2/SSE-backed memchr instead of byte-by-byte scan.
+		rel := strings.index_byte(src[i:], '{')
+		if rel < 0 {
+			i = len(src)
+			break
+		}
+		next := i + rel
+		// Single '{' (not '{{') — advance past it and keep scanning.
+		if next + 1 >= len(src) || src[next + 1] != '{' {
+			i = next + 1
+			continue
+		}
+		i = next
+
+		if i > text_start {
+			append(&tokens, Token{kind = .Text, value = src[text_start:i], pos = text_start})
+		}
+
+		tag_pos := i
+
+		if i + 2 < len(src) && src[i + 2] == '{' {
+			content_start := i + 3
+			idx := strings.index(src[content_start:], "}}}")
+			if idx < 0 {
+				return tokens, Error_Body {
+					msg = "unclosed triple mustache '{{{'",
+					pos = tag_pos,
+					kind = .Syntax,
+				}
+			}
+			close := content_start + idx
+			key := strings.trim_space(src[content_start:close])
+			append(&tokens, Token{kind = .Unescaped, value = key, pos = tag_pos})
+			i = close + 3
+			text_start = i
+		} else {
+			content_start := i + 2
+			sigil: byte = 0
+			if content_start < len(src) {
+				sigil = src[content_start]
 			}
 
-			tag_pos := i
+			kind: Token_Kind
+			key_start := content_start
 
-			if i + 2 < len(src) && src[i + 2] == '{' {
-				content_start := i + 3
-				idx := strings.index(src[content_start:], "}}}")
-				if idx < 0 {
-					return tokens, Error_Body {
-						msg = "unclosed triple mustache '{{{'",
-						pos = tag_pos,
-						kind = .Syntax,
-					}
-				}
-				close := content_start + idx
-				key := strings.trim_space(src[content_start:close])
-				append(&tokens, Token{kind = .Unescaped, value = key, pos = tag_pos})
-				i = close + 3
-				text_start = i
-			} else {
-				content_start := i + 2
-				sigil: byte = 0
-				if content_start < len(src) {
-					sigil = src[content_start]
-				}
+			switch sigil {
+			case '&':
+				kind = .Unescaped; key_start = content_start + 1
+			case '#':
+				kind = .Section_Open; key_start = content_start + 1
+			case '^':
+				kind = .Inverted_Open; key_start = content_start + 1
+			case '/':
+				kind = .Section_Close; key_start = content_start + 1
+			case '!':
+				kind = .Comment; key_start = content_start + 1
+			case '>':
+				kind = .Partial; key_start = content_start + 1
+			case '<':
+				kind = .Parent; key_start = content_start + 1
+			case '$':
+				kind = .Block_Open; key_start = content_start + 1
+			case:
+				kind = .Variable
+			}
 
-				kind: Token_Kind
-				key_start := content_start
-
-				switch sigil {
-				case '&':
-					kind = .Unescaped; key_start = content_start + 1
-				case '#':
-					kind = .Section_Open; key_start = content_start + 1
-				case '^':
-					kind = .Inverted_Open; key_start = content_start + 1
-				case '/':
-					kind = .Section_Close; key_start = content_start + 1
-				case '!':
-					kind = .Comment; key_start = content_start + 1
-				case '>':
-					kind = .Partial; key_start = content_start + 1
-				case '<':
-					kind = .Parent; key_start = content_start + 1
-				case '$':
-					kind = .Block_Open; key_start = content_start + 1
-				case:
-					kind = .Variable
-				}
-
-				close_idx := strings.index(src[key_start:], "}}")
+			close_idx := strings.index(src[key_start:], "}}")
 			if close_idx < 0 {
 				return tokens, Error_Body {
 					msg = "unclosed tag '{{'",
@@ -97,40 +110,37 @@ tokenize :: proc(
 					kind = .Syntax,
 				}
 			}
-				close := key_start + close_idx
+			close := key_start + close_idx
 
-				content := src[key_start:close]
+			content := src[key_start:close]
 
-				if kind == .Comment {
-					append(&tokens, Token{kind = .Comment, value = content, pos = tag_pos})
-				} else if kind == .Partial {
-					trimmed := strings.trim_space(content)
-					is_dyn := false
-					if len(trimmed) > 0 && trimmed[0] == '*' {
-						is_dyn = true
-						trimmed = strings.trim_space(trimmed[1:])
-					}
-					append(
-						&tokens,
-						Token {
-							kind = .Partial,
-							value = trimmed,
-							is_dynamic = is_dyn,
-							pos = tag_pos,
-						},
-					)
-				} else {
-					append(
-						&tokens,
-						Token{kind = kind, value = strings.trim_space(content), pos = tag_pos},
-					)
+			if kind == .Comment {
+				append(&tokens, Token{kind = .Comment, value = content, pos = tag_pos})
+			} else if kind == .Partial {
+				trimmed := strings.trim_space(content)
+				is_dyn := false
+				if len(trimmed) > 0 && trimmed[0] == '*' {
+					is_dyn = true
+					trimmed = strings.trim_space(trimmed[1:])
 				}
-
-				i = close + 2
-				text_start = i
+				append(
+					&tokens,
+					Token {
+						kind = .Partial,
+						value = trimmed,
+						is_dynamic = is_dyn,
+						pos = tag_pos,
+					},
+				)
+			} else {
+				append(
+					&tokens,
+					Token{kind = kind, value = strings.trim_space(content), pos = tag_pos},
+				)
 			}
-		} else {
-			i += 1
+
+			i = close + 2
+			text_start = i
 		}
 	}