| 1 | package export_test |
| 2 | |
| 3 | // A small PDF reader, for tests only. |
| 4 | // |
| 5 | // The point of it is that a PDF export cannot be checked by comparing |
| 6 | // bytes — the interesting properties are "a reader can find the |
| 7 | // objects", "the pages are there", and "the text can be selected and |
| 8 | // copied back out", and none of those is visible in a byte comparison. |
| 9 | // This walks the file the way a reader does: through the trailer to |
| 10 | // the cross-reference table, through the offsets in it to the objects, |
| 11 | // and through each font's ToUnicode CMap to turn the glyph ids in a |
| 12 | // content stream back into words. |
| 13 | // |
| 14 | // It follows the file's own xref offsets rather than scanning for |
| 15 | // "obj", which means a wrong offset fails the test rather than being |
| 16 | // routed around — the offsets are exactly the kind of thing that is |
| 17 | // silently wrong in a hand-written PDF writer. |
| 18 | |
| 19 | import ( |
| 20 | "bytes" |
| 21 | "compress/zlib" |
| 22 | "fmt" |
| 23 | "io" |
| 24 | "regexp" |
| 25 | "strconv" |
| 26 | "strings" |
| 27 | "testing" |
| 28 | ) |
| 29 | |
| 30 | type pdfObject struct { |
| 31 | dict string |
| 32 | stream []byte // already decompressed |
| 33 | } |
| 34 | |
| 35 | type pdfFile struct { |
| 36 | objects map[int]pdfObject |
| 37 | } |
| 38 | |
| 39 | func readPDF(t *testing.T, data []byte) pdfFile { |
| 40 | t.Helper() |
| 41 | |
| 42 | if !bytes.HasPrefix(data, []byte("%PDF-1.")) { |
| 43 | t.Fatalf("not a PDF: starts with %q", firstLine(data)) |
| 44 | } |
| 45 | if !bytes.HasSuffix(bytes.TrimRight(data, "\n"), []byte("%%EOF")) { |
| 46 | t.Fatal("file does not end with the end-of-file marker") |
| 47 | } |
| 48 | |
| 49 | // startxref, then the offset on the following line. |
| 50 | idx := bytes.LastIndex(data, []byte("startxref")) |
| 51 | if idx < 0 { |
| 52 | t.Fatal("no startxref") |
| 53 | } |
| 54 | var xrefAt int |
| 55 | if _, err := fmt.Sscanf(string(data[idx:]), "startxref\n%d", &xrefAt); err != nil { |
| 56 | t.Fatalf("unreadable startxref: %v", err) |
| 57 | } |
| 58 | if xrefAt <= 0 || xrefAt >= len(data) { |
| 59 | t.Fatalf("startxref points outside the file: %d of %d", xrefAt, len(data)) |
| 60 | } |
| 61 | |
| 62 | var first, count int |
| 63 | if _, err := fmt.Sscanf(string(data[xrefAt:]), "xref\n%d %d", &first, &count); err != nil { |
| 64 | t.Fatalf("unreadable xref header at %d: %v", xrefAt, err) |
| 65 | } |
| 66 | if first != 0 { |
| 67 | t.Fatalf("xref starts at object %d, want 0", first) |
| 68 | } |
| 69 | |
| 70 | // Entries are exactly 20 bytes each, starting after the two header |
| 71 | // lines. Finding the entry table by scanning past the header keeps |
| 72 | // this honest about the header's own length. |
| 73 | entriesAt := xrefAt + strings.Index(string(data[xrefAt:]), "\n") + 1 |
| 74 | entriesAt += strings.Index(string(data[entriesAt:]), "\n") + 1 |
| 75 | |
| 76 | out := pdfFile{objects: map[int]pdfObject{}} |
| 77 | for i := 1; i < count; i++ { |
| 78 | at := entriesAt + i*20 |
| 79 | if at+20 > len(data) { |
| 80 | t.Fatalf("xref entry %d runs past the end of the file", i) |
| 81 | } |
| 82 | entry := string(data[at : at+20]) |
| 83 | off, err := strconv.Atoi(strings.TrimSpace(entry[:10])) |
| 84 | if err != nil { |
| 85 | t.Fatalf("xref entry %d is not a number: %q", i, entry) |
| 86 | } |
| 87 | out.objects[i] = readObject(t, data, i, off) |
| 88 | } |
| 89 | return out |
| 90 | } |
| 91 | |
| 92 | var lengthRe = regexp.MustCompile(`/Length (\d+)`) |
| 93 | |
| 94 | func readObject(t *testing.T, data []byte, num, off int) pdfObject { |
| 95 | t.Helper() |
| 96 | if off <= 0 || off >= len(data) { |
| 97 | t.Fatalf("object %d: offset %d is outside the file", num, off) |
| 98 | } |
| 99 | want := fmt.Sprintf("%d 0 obj\n", num) |
| 100 | if !bytes.HasPrefix(data[off:], []byte(want)) { |
| 101 | t.Fatalf("object %d: xref points at %q, not %q", num, firstLine(data[off:]), strings.TrimSpace(want)) |
| 102 | } |
| 103 | body := data[off+len(want):] |
| 104 | |
| 105 | streamAt := bytes.Index(body, []byte("stream\n")) |
| 106 | endObj := bytes.Index(body, []byte("\nendobj")) |
| 107 | if streamAt < 0 || (endObj >= 0 && endObj < streamAt) { |
| 108 | if endObj < 0 { |
| 109 | t.Fatalf("object %d: no endobj", num) |
| 110 | } |
| 111 | return pdfObject{dict: string(body[:endObj])} |
| 112 | } |
| 113 | |
| 114 | dict := string(body[:streamAt]) |
| 115 | m := lengthRe.FindStringSubmatch(dict) |
| 116 | if m == nil { |
| 117 | t.Fatalf("object %d: a stream with no /Length", num) |
| 118 | } |
| 119 | n, _ := strconv.Atoi(m[1]) |
| 120 | start := streamAt + len("stream\n") |
| 121 | if start+n > len(body) { |
| 122 | t.Fatalf("object %d: /Length %d runs past the end of the file", num, n) |
| 123 | } |
| 124 | raw := body[start : start+n] |
| 125 | |
| 126 | if !strings.Contains(dict, "/FlateDecode") { |
| 127 | return pdfObject{dict: dict, stream: raw} |
| 128 | } |
| 129 | zr, err := zlib.NewReader(bytes.NewReader(raw)) |
| 130 | if err != nil { |
| 131 | t.Fatalf("object %d: stream is not valid zlib: %v", num, err) |
| 132 | } |
| 133 | plain, err := io.ReadAll(zr) |
| 134 | if err != nil { |
| 135 | t.Fatalf("object %d: decompressing: %v", num, err) |
| 136 | } |
| 137 | return pdfObject{dict: dict, stream: plain} |
| 138 | } |
| 139 | |
| 140 | func firstLine(b []byte) string { |
| 141 | if i := bytes.IndexByte(b, '\n'); i >= 0 { |
| 142 | b = b[:i] |
| 143 | } |
| 144 | if len(b) > 60 { |
| 145 | b = b[:60] |
| 146 | } |
| 147 | return string(b) |
| 148 | } |
| 149 | |
| 150 | // pages returns every page object's dictionary, in /Kids order. |
| 151 | func (f pdfFile) pages(t *testing.T) []pdfObject { |
| 152 | t.Helper() |
| 153 | var pagesDict string |
| 154 | for _, o := range f.objects { |
| 155 | if strings.Contains(o.dict, "/Type /Pages") { |
| 156 | pagesDict = o.dict |
| 157 | } |
| 158 | } |
| 159 | if pagesDict == "" { |
| 160 | t.Fatal("no /Type /Pages node") |
| 161 | } |
| 162 | kids := regexp.MustCompile(`(\d+) 0 R`).FindAllStringSubmatch( |
| 163 | betweenBrackets(t, pagesDict, "/Kids ["), -1) |
| 164 | var out []pdfObject |
| 165 | for _, k := range kids { |
| 166 | n, _ := strconv.Atoi(k[1]) |
| 167 | o, ok := f.objects[n] |
| 168 | if !ok { |
| 169 | t.Fatalf("/Kids names object %d, which the xref does not list", n) |
| 170 | } |
| 171 | if !strings.Contains(o.dict, "/Type /Page") { |
| 172 | t.Fatalf("object %d is in /Kids but is not a /Page", n) |
| 173 | } |
| 174 | out = append(out, o) |
| 175 | } |
| 176 | return out |
| 177 | } |
| 178 | |
| 179 | func betweenBrackets(t *testing.T, s, open string) string { |
| 180 | t.Helper() |
| 181 | i := strings.Index(s, open) |
| 182 | if i < 0 { |
| 183 | t.Fatalf("%q not found in %q", open, s) |
| 184 | } |
| 185 | rest := s[i+len(open):] |
| 186 | j := strings.Index(rest, "]") |
| 187 | if j < 0 { |
| 188 | t.Fatalf("unterminated %q", open) |
| 189 | } |
| 190 | return rest[:j] |
| 191 | } |
| 192 | |
| 193 | var ( |
| 194 | fontResRe = regexp.MustCompile(`/(F\d+) (\d+) 0 R`) |
| 195 | contentsRe = regexp.MustCompile(`/Contents (\d+) 0 R`) |
| 196 | toUnicodeRe = regexp.MustCompile(`/ToUnicode (\d+) 0 R`) |
| 197 | bfcharRe = regexp.MustCompile(`<([0-9A-Fa-f]+)> <([0-9A-Fa-f]+)>`) |
| 198 | // One drawn run, exactly as internal/pdf's Page.Text writes it: |
| 199 | // the text matrix carries the baseline position, so the x and y a |
| 200 | // run actually landed at are readable here and nowhere else. |
| 201 | showRe = regexp.MustCompile( |
| 202 | `/(F\d+) [\d.-]+ Tf 1 0 0 1 ([\d.-]+) ([\d.-]+) Tm <([0-9A-Fa-f]*)> Tj`) |
| 203 | // One filled rectangle (Page.Rule) — the blockquote's edge is the |
| 204 | // only non-text drawing this exporter does. |
| 205 | ruleRe = regexp.MustCompile(`([\d.-]+) ([\d.-]+) ([\d.-]+) ([\d.-]+) re f`) |
| 206 | ) |
| 207 | |
| 208 | // run is one drawn piece of text with where it was drawn: x and y are |
| 209 | // the baseline's own coordinates, in points from the bottom-left of the |
| 210 | // page, which is PDF's origin and not the layout's. |
| 211 | type run struct { |
| 212 | x, y float64 |
| 213 | text string |
| 214 | } |
| 215 | |
| 216 | // cmaps maps each font resource named by a page to its own ToUnicode |
| 217 | // table — the glyph-id-to-rune mapping a reader uses to let a person |
| 218 | // copy text out. |
| 219 | func (f pdfFile) cmaps(t *testing.T, page pdfObject) map[string]map[uint16]rune { |
| 220 | t.Helper() |
| 221 | out := map[string]map[uint16]rune{} |
| 222 | for _, m := range fontResRe.FindAllStringSubmatch(page.dict, -1) { |
| 223 | n, _ := strconv.Atoi(m[2]) |
| 224 | fontObj, ok := f.objects[n] |
| 225 | if !ok { |
| 226 | t.Fatalf("page resources name font object %d, which does not exist", n) |
| 227 | } |
| 228 | tu := toUnicodeRe.FindStringSubmatch(fontObj.dict) |
| 229 | if tu == nil { |
| 230 | t.Fatalf("font %s has no /ToUnicode — its text cannot be copied out", m[1]) |
| 231 | } |
| 232 | tn, _ := strconv.Atoi(tu[1]) |
| 233 | cmap := map[uint16]rune{} |
| 234 | for _, e := range bfcharRe.FindAllStringSubmatch(string(f.objects[tn].stream), -1) { |
| 235 | gid, _ := strconv.ParseUint(e[1], 16, 16) |
| 236 | cmap[uint16(gid)] = decodeUTF16BE(e[2]) |
| 237 | } |
| 238 | out[m[1]] = cmap |
| 239 | } |
| 240 | return out |
| 241 | } |
| 242 | |
| 243 | func (f pdfFile) content(t *testing.T, page pdfObject) string { |
| 244 | t.Helper() |
| 245 | cm := contentsRe.FindStringSubmatch(page.dict) |
| 246 | if cm == nil { |
| 247 | t.Fatal("page has no /Contents") |
| 248 | } |
| 249 | cn, _ := strconv.Atoi(cm[1]) |
| 250 | return string(f.objects[cn].stream) |
| 251 | } |
| 252 | |
| 253 | // runs returns every drawn piece of text on the page WITH its baseline |
| 254 | // position, in the order the content stream draws them. |
| 255 | // |
| 256 | // text below throws the positions away, which is why a whole-page |
| 257 | // horizontal offset was invisible to every test in this package until |
| 258 | // this existed: the words all copied back out correctly, they were just |
| 259 | // drawn in the margin. |
| 260 | func (f pdfFile) runs(t *testing.T, page pdfObject) []run { |
| 261 | t.Helper() |
| 262 | cmaps := f.cmaps(t, page) |
| 263 | |
| 264 | var out []run |
| 265 | for _, m := range showRe.FindAllStringSubmatch(f.content(t, page), -1) { |
| 266 | cmap, ok := cmaps[m[1]] |
| 267 | if !ok { |
| 268 | t.Fatalf("content stream uses font %s, which the page resources do not declare", m[1]) |
| 269 | } |
| 270 | x, err := strconv.ParseFloat(m[2], 64) |
| 271 | if err != nil { |
| 272 | t.Fatalf("unreadable x in the text matrix: %q", m[2]) |
| 273 | } |
| 274 | y, err := strconv.ParseFloat(m[3], 64) |
| 275 | if err != nil { |
| 276 | t.Fatalf("unreadable y in the text matrix: %q", m[3]) |
| 277 | } |
| 278 | var b strings.Builder |
| 279 | hex := m[4] |
| 280 | for i := 0; i+4 <= len(hex); i += 4 { |
| 281 | gid, _ := strconv.ParseUint(hex[i:i+4], 16, 16) |
| 282 | r, ok := cmap[uint16(gid)] |
| 283 | if !ok { |
| 284 | t.Errorf("glyph %d is drawn but has no ToUnicode entry", gid) |
| 285 | continue |
| 286 | } |
| 287 | b.WriteRune(r) |
| 288 | } |
| 289 | out = append(out, run{x: x, y: y, text: b.String()}) |
| 290 | } |
| 291 | return out |
| 292 | } |
| 293 | |
| 294 | // rect is one filled rectangle, in PDF coordinates. |
| 295 | type rect struct{ x, y, w, h float64 } |
| 296 | |
| 297 | // rules returns every filled rectangle on the page — the blockquote |
| 298 | // edge, which is drawn at an x of its own and would sit in the margin |
| 299 | // alongside the text if that x were wrong. |
| 300 | func (f pdfFile) rules(t *testing.T, page pdfObject) []rect { |
| 301 | t.Helper() |
| 302 | var out []rect |
| 303 | for _, m := range ruleRe.FindAllStringSubmatch(f.content(t, page), -1) { |
| 304 | var r rect |
| 305 | for i, dst := range []*float64{&r.x, &r.y, &r.w, &r.h} { |
| 306 | v, err := strconv.ParseFloat(m[i+1], 64) |
| 307 | if err != nil { |
| 308 | t.Fatalf("unreadable rectangle coordinate %q", m[i+1]) |
| 309 | } |
| 310 | *dst = v |
| 311 | } |
| 312 | out = append(out, r) |
| 313 | } |
| 314 | return out |
| 315 | } |
| 316 | |
| 317 | // text reconstructs what a person would get by selecting everything on |
| 318 | // the page and copying it: the glyph ids in the content stream, mapped |
| 319 | // back through each font's own ToUnicode CMap. If that mapping is |
| 320 | // wrong or missing, this returns mojibake, which is exactly what a |
| 321 | // reader would show. |
| 322 | func (f pdfFile) text(t *testing.T, page pdfObject) string { |
| 323 | t.Helper() |
| 324 | var b strings.Builder |
| 325 | for _, r := range f.runs(t, page) { |
| 326 | b.WriteString(r.text) |
| 327 | } |
| 328 | return b.String() |
| 329 | } |
| 330 | |
| 331 | func decodeUTF16BE(hex string) rune { |
| 332 | var units []rune |
| 333 | for i := 0; i+4 <= len(hex); i += 4 { |
| 334 | v, _ := strconv.ParseUint(hex[i:i+4], 16, 16) |
| 335 | units = append(units, rune(v)) |
| 336 | } |
| 337 | if len(units) == 2 && units[0] >= 0xD800 && units[0] <= 0xDBFF { |
| 338 | return ((units[0] - 0xD800) << 10) + (units[1] - 0xDC00) + 0x10000 |
| 339 | } |
| 340 | if len(units) == 0 { |
| 341 | return 0 |
| 342 | } |
| 343 | return units[0] |
| 344 | } |
| 345 | |
amadan