path: root/web/render.go

// render.go — source to HTML: tree-sitter highlighting, coverage classes and
// note anchors baked into one <ol> per chunk.
package web

import (
	"bytes"
	"html/template"
	"path/filepath"
	"sort"
	"strconv"
	"strings"
	"unsafe"

	tszig "github.com/tree-sitter-grammars/tree-sitter-zig/bindings/go"
	sitter "github.com/tree-sitter/go-tree-sitter"
	tsbash "github.com/tree-sitter/tree-sitter-bash/bindings/go"
	tsc "github.com/tree-sitter/tree-sitter-c/bindings/go"
	tscpp "github.com/tree-sitter/tree-sitter-cpp/bindings/go"
	tsgo "github.com/tree-sitter/tree-sitter-go/bindings/go"
	tsjs "github.com/tree-sitter/tree-sitter-javascript/bindings/go"
	tsjson "github.com/tree-sitter/tree-sitter-json/bindings/go"
	tspython "github.com/tree-sitter/tree-sitter-python/bindings/go"
	tsrust "github.com/tree-sitter/tree-sitter-rust/bindings/go"
	tsts "github.com/tree-sitter/tree-sitter-typescript/bindings/go"
)

// Lines per content-visibility chunk: the browser skips layout and paint of
// offscreen chunks wholesale, instead of managing one containment context
// per line.
const chunkLines = 800

// approximate rendered line height, for contain-intrinsic-size placeholders
const lineHeightPx = 19

// render builds the code HTML: <ol> chunks of chunkLines lines, each line an
// <li> with syntax spans, coverage classes and note anchors. One pre-sized
// builder, no fmt, minimal escaping — large files are browser-bound, so the
// output is kept as small and flat as possible.
func (f *File) render(src []byte, hl *highlighter) {
	classes := hl.classify(f.Path, src)
	f.Total = bytes.Count(src, []byte{'\n'})
	if len(src) > 0 && src[len(src)-1] != '\n' {
		f.Total++
	}
	f.Covered = 0
	for i := 1; i < len(f.Cov) && i <= f.Total; i++ {
		if f.Cov[i] != 0 {
			f.Covered++
		}
	}
	if f.Total > 0 {
		f.Pct = f.Covered * 100 / f.Total
	}
	sort.Slice(f.Notes, func(i, j int) bool { return f.Notes[i].Start < f.Notes[j].Start })
	for ni, n := range f.Notes {
		for l := n.Start; l <= n.End && l > 0; l++ {
			if _, taken := f.NoteAt[l]; !taken {
				f.NoteAt[l] = ni
			}
		}
	}
	f.Body = f.emit(src, classes, hl.classNames, 1, f.Total)
}

// emit writes lines [from..to] as <ol> chunks. from is where the chunking (and
// so the browser's content-visibility windows) starts counting, which is what
// lets an excerpt come out as one chunk rather than a slice of the file's.
func (f *File) emit(src []byte, classes []int16, names []string, from, to int) template.HTML {
	if from < 1 {
		from = 1
	}
	if to > f.Total {
		to = f.Total
	}
	if from > to {
		return ""
	}
	start := 0
	for ln := 1; ln < from; ln++ { // skip to the first line asked for
		start = pastEOL(src, start)
	}
	stop := start
	for ln := from; ln <= to; ln++ {
		stop = pastEOL(src, stop)
	}
	var b strings.Builder
	span := stop - start
	b.Grow(span + span/2 + (to-from+1)*48)
	var num []byte
	writeInt := func(n int) { num = strconv.AppendInt(num[:0], int64(n), 10); b.Write(num) }
	for ln := from; ln <= to; ln++ {
		end := start
		for end < len(src) && src[end] != '\n' {
			end++
		}
		opens := (ln-from)%chunkLines == 0
		if opens {
			if ln > from {
				b.WriteString("</ol>")
			}
			rem := to - (ln - 1)
			if rem > chunkLines {
				rem = chunkLines
			}
			b.WriteString(`<ol class="code" style="--h: `)
			writeInt(rem * lineHeightPx)
			b.WriteString(`px">`)
		}
		b.WriteString(`<li id="L`)
		writeInt(ln)
		b.WriteByte('"')
		// Where the chunk's numbering starts. It has to be counter-set on the
		// first line rather than counter-reset on the <ol>, because
		// content-visibility: auto brings style containment with it and a
		// contained element's own counter-reset comes out as 0 no matter what
		// value it names — so every chunk but the first would count from 1.
		if opens {
			b.WriteString(` style="counter-set: ln `)
			writeInt(ln)
			b.WriteByte('"')
		}
		mask := uint8(0)
		if ln < len(f.Cov) {
			mask = f.Cov[ln]
		}
		ni, noted := f.NoteAt[ln]
		if mask != 0 || noted {
			b.WriteString(` class="`)
			if mask != 0 {
				b.WriteString("cov am")
				writeInt(int(mask))
				if noted {
					b.WriteString(" noted")
				}
			} else {
				b.WriteString("noted")
			}
			b.WriteByte('"')
		}
		if noted {
			b.WriteString(` data-note="note-`)
			writeInt(f.Notes[ni].ID)
			b.WriteByte('"')
		}
		b.WriteByte('>')
		emitLine(&b, src[start:end], classes[start:end], names)
		b.WriteString("</li>")
		start = end + 1
	}
	b.WriteString("</ol>")
	return template.HTML(b.String())
}

// pastEOL is the offset just past the line starting at i.
func pastEOL(src []byte, i int) int {
	for i < len(src) && src[i] != '\n' {
		i++
	}
	if i < len(src) {
		i++
	}
	return i
}

// excerptContext is how many lines either side of a note's range its own page
// shows, so the noted lines are read in context rather than in isolation.
const excerptContext = 4

// excerptFor renders just the lines a note is about, the way the code view
// would: the file's real line numbers, its coverage tint, and the noted lines
// marked. This is what lets a shared permalink stand on its own — whoever
// opens it sees the note and the code it is about together, without having to
// find their way into the file first.
//
// cov comes from the model and is read-only here, so the caller can pass the
// live *File's slice rather than copying it.
func excerptFor(path string, src []byte, cov []uint8, hl *highlighter, n *Note) (body template.HTML, from, to int) {
	total := bytes.Count(src, []byte{'\n'})
	if len(src) > 0 && src[len(src)-1] != '\n' {
		total++
	}
	end := n.End
	if end < n.Start {
		end = n.Start
	}
	if n.Start < 1 || total == 0 {
		return "", 0, 0
	}
	f := &File{Path: path, Cov: cov, Total: total, Notes: []*Note{n}, NoteAt: map[int]int{}}
	for l := n.Start; l <= end; l++ {
		f.NoteAt[l] = 0
	}
	from, to = n.Start-excerptContext, end+excerptContext
	if from < 1 {
		from = 1
	}
	if to > total {
		to = total
	}
	if from > total {
		return "", 0, 0
	}
	return f.emit(src, hl.classify(path, src), hl.classNames, from, to), from, to
}

// emitLine writes one line's spans, merging runs of the same class across
// whitespace-only gaps ("pub fn" is one span, not two) to keep the DOM flat.
func emitLine(b *strings.Builder, line []byte, cls []int16, names []string) {
	open := int16(0)
	i := 0
	for i < len(line) {
		j := i
		c := cls[i]
		for j < len(line) && cls[j] == c {
			j++
		}
		if c == 0 {
			allSpace := true
			for k := i; k < j; k++ {
				if line[k] != ' ' && line[k] != '\t' {
					allSpace = false
					break
				}
			}
			if !(allSpace && open != 0 && j < len(line) && cls[j] == open) {
				if open != 0 {
					b.WriteString("</span>")
					open = 0
				}
			}
			escapeTo(b, line[i:j])
		} else {
			if open != c {
				if open != 0 {
					b.WriteString("</span>")
				}
				b.WriteString(`<span class="s-`)
				b.WriteString(names[c])
				b.WriteString(`">`)
				open = c
			}
			escapeTo(b, line[i:j])
		}
		i = j
	}
	if open != 0 {
		b.WriteString("</span>")
	}
}

// escapeTo escapes the three characters that matter in text content.
func escapeTo(b *strings.Builder, s []byte) {
	last := 0
	for i, c := range s {
		var rep string
		switch c {
		case '&':
			rep = "&amp;"
		case '<':
			rep = "&lt;"
		case '>':
			rep = "&gt;"
		default:
			continue
		}
		b.Write(s[last:i])
		b.WriteString(rep)
		last = i + 1
	}
	b.Write(s[last:])
}

// ── highlighting ──────────────────────────────────────────────────────────

// grammars lists every language the site can highlight: the extensions it
// claims, the vendored queries that paint it, and the cgo grammar. Query files
// concatenate the way upstream composes them — C's patterns underpin C++,
// JavaScript's underpin TypeScript — and the concatenation keeps the
// later-patterns-win order, so the supplement wins where both match.
var grammars = []struct {
	exts []string
	scms []string
	lang func() unsafe.Pointer
}{
	{[]string{".rs"}, []string{"rust"}, tsrust.Language},
	{[]string{".go"}, []string{"go"}, tsgo.Language},
	{[]string{".py", ".pyi"}, []string{"python"}, tspython.Language},
	{[]string{".js", ".mjs", ".cjs", ".jsx"}, []string{"javascript", "javascript-params", "jsx"}, tsjs.Language},
	{[]string{".ts", ".mts", ".cts"}, []string{"javascript", "typescript"}, tsts.LanguageTypescript},
	{[]string{".tsx"}, []string{"javascript", "typescript", "jsx"}, tsts.LanguageTSX},
	{[]string{".c"}, []string{"c"}, tsc.Language},
	// .h is C++'s upstream claim, and the C++ grammar is a superset: C headers
	// parse the same under it, C++ headers only parse under it
	{[]string{".cc", ".cpp", ".cxx", ".h", ".hh", ".hpp", ".hxx"}, []string{"c", "cpp"}, tscpp.Language},
	{[]string{".zig"}, []string{"zig"}, tszig.Language},
	{[]string{".json"}, []string{"json"}, tsjson.Language},
	{[]string{".sh", ".bash"}, []string{"bash"}, tsbash.Language},
}

// language is one grammar's share of the highlighter. Everything here is built
// at startup and never written again, which is what lets classify run from any
// number of goroutines at once.
type language struct {
	lang       *sitter.Language
	query      *sitter.Query
	captureCls []int16 // capture index -> class id
}

type highlighter struct {
	byExt      map[string]*language
	classNames []string
}

// resolveKey finds the theme syntax key for a capture name by stripping dot
// segments; "" when the theme has no entry at all.
func resolveKey(t *Theme, name string) string {
	for key := name; key != ""; {
		if s, ok := t.Syntax[key]; ok && s.Color != "" {
			return key
		}
		if i := strings.LastIndex(key, "."); i >= 0 {
			key = key[:i]
		} else {
			break
		}
	}
	return ""
}

// nearFg reports whether a capture would render indistinguishably from plain
// foreground text in the theme. Spans for such captures are pure DOM weight.
func nearFg(t *Theme, name string) bool {
	key := resolveKey(t, name)
	if key == "" {
		return true // no color -> falls through to fg anyway
	}
	return colorDist(hex6(t.Syntax[key].Color), hex6(t.Fg)) < 2000
}

func newHighlighter(dark, light *Theme) *highlighter {
	h := &highlighter{byExt: map[string]*language{}, classNames: []string{""}}
	for _, g := range grammars {
		var scm strings.Builder
		for _, name := range g.scms {
			b, err := tfs.ReadFile("assets/" + name + ".scm")
			if err != nil {
				fatal("%v", err)
			}
			scm.Write(b)
		}
		l := &language{lang: sitter.NewLanguage(g.lang())}
		q, qerr := sitter.NewQuery(l.lang, scm.String())
		if qerr != nil {
			fatal("%s query: %v", strings.Join(g.scms, "+"), qerr)
		}
		l.query = q
		for _, name := range q.CaptureNames() {
			l.captureCls = append(l.captureCls, h.classFor(dark, light, name))
		}
		for _, ext := range g.exts {
			h.byExt[ext] = l
		}
	}
	return h
}

// classFor interns the class id of a capture name, which doubles as a zed
// theme syntax key. Captures both themes paint (nearly) as foreground get id 0
// and no span at all — punctuation and friends are the bulk of all spans.
func (h *highlighter) classFor(dark, light *Theme, name string) int16 {
	if nearFg(dark, name) && nearFg(light, name) {
		return 0
	}
	key := resolveKey(dark, name)
	if key == "" {
		key = resolveKey(light, name)
	}
	cls := strings.ReplaceAll(key, ".", "-")
	for i, n := range h.classNames {
		if n == cls {
			return int16(i)
		}
	}
	h.classNames = append(h.classNames, cls)
	return int16(len(h.classNames) - 1)
}

type paint struct {
	start, end uint
	pattern    uint
	cls        int16
}

// classify paints a file. Extensions no grammar claims come back unclassified
// rather than guessed at.
func (h *highlighter) classify(path string, src []byte) []int16 {
	return h.classifyWith(h.byExt[strings.ToLower(filepath.Ext(path))], src)
}

// fenceAliases maps the fence languages that are not already one of the
// extensions in grammars. Most are ("go", "rs", "py", "ts", "json", "bash"),
// so the grammar table does the bulk of the lookup on its own.
var fenceAliases = map[string]string{
	"golang": ".go", "rust": ".rs", "python": ".py", "python3": ".py",
	"javascript": ".js", "node": ".js", "typescript": ".ts",
	"c++": ".cc", "shell": ".sh", "zsh": ".sh", "console": ".sh",
	"jsonc": ".json", "json5": ".json",
}

// classifyLang paints a fenced code block, whose language arrives as a name
// rather than a path. Unknown or absent languages come back unclassified, the
// same as an extension no grammar claims.
func (h *highlighter) classifyLang(lang string, src []byte) []int16 {
	if h == nil {
		return make([]int16, len(src))
	}
	name := strings.ToLower(strings.TrimSpace(lang))
	l := h.byExt["."+name]
	if l == nil {
		l = h.byExt[fenceAliases[name]]
	}
	return h.classifyWith(l, src)
}

// classifyWith is safe for concurrent use: the shared per-language query is
// immutable, and parser and cursor are per-call.
//
// Matches arrive already filtered by their #eq?/#match?/#any-of? predicates —
// go-tree-sitter evaluates those in QueryMatches.Next — so every capture here
// is one the query really meant.
func (h *highlighter) classifyWith(l *language, src []byte) []int16 {
	classes := make([]int16, len(src))
	if l == nil {
		return classes
	}
	parser := sitter.NewParser()
	defer parser.Close()
	parser.SetLanguage(l.lang)
	tree := parser.Parse(src, nil)
	defer tree.Close()

	qc := sitter.NewQueryCursor()
	defer qc.Close()
	matches := qc.Matches(l.query, tree.RootNode(), src)
	var paints []paint
	for m := matches.Next(); m != nil; m = matches.Next() {
		for _, c := range m.Captures {
			if cls := l.captureCls[c.Index]; cls != 0 {
				paints = append(paints, paint{c.Node.StartByte(), c.Node.EndByte(), m.PatternIndex, cls})
			}
		}
	}
	// later patterns in the query paint over earlier ones; two captures of one
	// pattern must not overlap, since this sort leaves their order arbitrary
	sort.Slice(paints, func(i, j int) bool { return paints[i].pattern < paints[j].pattern })
	for _, p := range paints {
		for i := p.start; i < p.end && int(i) < len(classes); i++ {
			classes[i] = p.cls
		}
	}
	return classes
}