@portll/cobolwork 0.2.117 → 0.2.150

package/lib/kernel/git.mjs

@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: AGPL-3.0-or-later
+// Reading a revision out of a repository with git plumbing only.
+import { spawnSync } from 'node:child_process';
+
+// A reviewed repository's .git/config can name commands; with fsmonitor off and plumbing only, git runs none.
+const GIT_ENV = { ...process.env, GIT_OPTIONAL_LOCKS: '0', GIT_TERMINAL_PROMPT: '0' };
+export const git = (repo, args, opts = {}) => spawnSync('git', ['-c', 'core.fsmonitor=false', '-C', repo, ...args], { env: GIT_ENV, maxBuffer: 256 * 1024 * 1024, ...opts });
+export const refused = (what, ref, why) => Object.assign(new Error(`${what} ${ref} failed: ${why}`), { code: 'EDIFFREF' });
+const BATCH_BYTES = 64 * 1024 * 1024;
+
+export function treeOf(repo, ref) {
+  if (!ref || String(ref).startsWith('-')) throw refused('git rev-parse', ref, 'a revision cannot be empty or start with "-"');
+  const r = git(repo, ['rev-parse', '--verify', '--quiet', '--end-of-options', `${ref}^{tree}`], { encoding: 'utf8' });
+  const oid = String(r.stdout || '').trim();
+  if (r.status !== 0 || !/^[0-9a-f]{40}([0-9a-f]{24})?$/.test(oid)) throw refused('git rev-parse', ref, String(r.stderr || '').trim() || 'not a revision');
+  return oid;
+}
+
+// The parts of a tree path, or null if a part is empty, climbs, names a drive or stream, or is .git.
+export function treePathParts(path) {
+  const parts = path.split('/');
+  return parts.some((p) => !p || p === '.' || p === '..' || /[\\:\0]/.test(p) || p.toLowerCase() === '.git') ? null : parts;
+}
+
+// Every file in the tree `oid` as committed, as [{ path, bytes }] in tree order: blobs as stored,
+// with no filter or line-ending conversion. Links and submodules are left out and counted.
+export function revisionBlobs(repo, oid, ref = oid) {
+  const listed = git(repo, ['ls-tree', '-r', '-z', '-l', '--full-tree', oid]);
+  if (listed.status !== 0) throw refused('git ls-tree', ref, String(listed.stderr || '').trim() || `exit ${listed.status}`);
+  const blobs = [];
+  let links = 0;
+  for (const entry of listed.stdout.toString('utf8').split('\0')) {
+    const m = /^(\d{6}) blob ([0-9a-f]+) +(\d+)\t(.+)$/s.exec(entry);
+    if (!m) continue;
+    if (m[1] === '120000') { links++; continue; }
+    blobs.push({ oid: m[2], size: Number(m[3]), path: m[4] });
+  }
+  // Batched by bytes as well as count, and each batch's buffer sized to hold it, so one large blob
+  // cannot overflow a buffer sized for the others.
+  const batches = [];
+  for (const b of blobs) {
+    const last = batches[batches.length - 1];
+    if (!last || last.length >= 500 || last.bytes + b.size > BATCH_BYTES) batches.push(Object.assign([b], { bytes: b.size }));
+    else { last.push(b); last.bytes += b.size; }
+  }
+  for (const batch of batches) {
+    const r = git(repo, ['cat-file', '--batch'], { input: batch.map((b) => b.oid).join('\n') + '\n', maxBuffer: batch.bytes + batch.length * 128 + 4096 });
+    if (r.status !== 0) throw refused('git cat-file', ref, String(r.stderr || '').trim() || (r.error ? r.error.code || r.error.message : `exit ${r.status}`));
+    let at = 0;
+    for (const b of batch) {
+      const nl = r.stdout.indexOf(0x0a, at);
+      const head = r.stdout.toString('utf8', at, nl).split(' ');
+      if (head[1] !== 'blob') throw refused('git cat-file', ref, `${b.oid} is ${head[1] || 'missing'}`);
+      const size = Number(head[2]);
+      b.bytes = r.stdout.subarray(nl + 1, nl + 1 + size);
+      at = nl + 1 + size + 1;
+    }
+  }
+  return { blobs, links };
+}

package/lib/kernel/identity.mjs

@@ -131,14 +131,15 @@ function codeText(line, kind, format) {
   return maskSecrets(s).replace(/\s+/g, ' ').trim();
 }
 
-// Reads each file once, however many findings it holds.
+// Reads each file once, however many findings it holds, from disk or through the reader the
+// caller's source tree supplies.
 function fileFacts(root, path, cache) {
   let facts = cache.get(path);
   if (facts) return facts;
   facts = { kind: 'other', lines: [], scopes: null, format: null };
   cache.set(path, facts);
   let src;
-  try { src = readSource(join(root, path)).text; } catch { return facts; }
+  try { src = cache.read ? cache.read(path) : readSource(join(root, path)).text; } catch { return facts; }
   facts.lines = src.split(/\r?\n/);
   if (isJcl(join(root, path))) {
     facts.kind = 'jcl';
@@ -171,8 +172,8 @@ const digest = (parts) => createHash('sha256').update(parts.join('\u0000')).dige
 // Stamps `fingerprint` on every finding and returns how many shared one with an earlier finding.
 // `repo` names the repository when one report holds several, where the same program id in two
 // repositories is two programs.
-export function stampFingerprints(findings, { root, repo = '' } = {}) {
-  const cache = new Map();
+export function stampFingerprints(findings, { root, repo = '', read = null } = {}) {
+  const cache = Object.assign(new Map(), { read });
   const seen = new Set();
   let shared = 0;
   for (const f of findings) {
@@ -188,8 +189,8 @@ export function stampFingerprints(findings, { root, repo = '' } = {}) {
 // Looser keys for one finding seen in two trees whose line was edited between them: `scope` is the
 // fingerprint without the line's text, and `route` names a data-flow finding by the statement its
 // trace starts at. Null where a finding has no route.
-export function pairingKeys(findings, { root, repo = '' } = {}) {
-  const cache = new Map();
+export function pairingKeys(findings, { root, repo = '', read = null } = {}) {
+  const cache = Object.assign(new Map(), { read });
   return findings.map((f) => {
     const { scope, subject } = partsOf(f, root, cache);
     const src = f.evidence === 'path' && f.related && f.related[0];

package/lib/kernel/source-tree.mjs

@@ -26,13 +26,16 @@
 //
 // An adapter that does not have a filesystem underneath it has to make that decision itself, and
 // `contains` is where it makes it. For a tree held in memory it is key membership. For a git
-// revision it is membership of the tree object, which cannot name a path outside itself. Anything
-// cleverer than those two deserves the six containment tests pointed at it before it ships:
-// test/sources.test.mjs:44, :63, :88, :101, :144 and test/review.test.mjs:114.
-import { readFileSync, realpathSync } from 'node:fs';
-import { dirname, resolve, sep } from 'node:path';
+// revision it is membership of the tree object, which cannot name a path outside itself. For a PDS
+// export it is membership of the members the walk found, each read with no link followed. Anything
+// cleverer than those deserves the six containment tests pointed at it before it ships:
+// test/sources.test.mjs:44, :63, :88, :101, :144 and test/review.test.mjs:114, as
+// test/git-tree.test.mjs and test/pds-export.test.mjs do.
+import { closeSync, constants, fstatSync, openSync, readFileSync, readSync, realpathSync } from 'node:fs';
+import { basename, dirname, relative, resolve, sep } from 'node:path';
 import { buildFileIndex, parseSource } from '../parser.mjs';
-import { readSource, relPath } from '../sources.mjs';
+import { classifyUnder, decodeSource, diskClassifier, heldClassifier, kindOfBytes, looksEbcdic, readSource, relPath } from '../sources.mjs';
+import { revisionBlobs, treeOf, treePathParts } from './git.mjs';
 
 // The shape every adapter answers to. Checked rather than documented, because an adapter missing a
 // method fails at the first rule set that happens to call it rather than at the boundary.
@@ -64,10 +67,14 @@ export function directoryTree(root, opts = {}) {
   const idx = buildFileIndex(root);
   const systemDirs = opts.systemDirs || [];
   const top = resolve(root);
+  const kindOf = diskClassifier();
+  classifyUnder(root, kindOf);
+  if (top !== root) classifyUnder(top, kindOf);
 
   return {
     kind: 'directory',
     root,
+    kindOf,
     // The index itself, for the two callers that need more than a list: the copybook set reads
     // copyDirs, and the inventory reports on unreadable directories and symlinks.
     index: idx,
@@ -92,40 +99,203 @@ export function directoryTree(root, opts = {}) {
   };
 }
 
-// A tree that was never on disk, for tests. `files` maps a path to its contents, as a string or a
-// Buffer. Paths are used exactly as given, so a test reads the way it writes.
-//
-// It cannot parse. parseSource resolves COPY statements against real directories and reads them
-// with readSource, so a program with copybooks needs a filesystem underneath it. A rule set that
-// only reads text - hidden, recon, vendor, jcl, build - works against this; one that parses does
-// not, and says so rather than parsing something unexpected.
-export function memoryTree(files, { root = '/memory' } = {}) {
-  const store = new Map(Object.entries(files).map(([p, v]) => [p, Buffer.isBuffer(v) ? v : Buffer.from(v, 'latin1')]));
+// A tree whose files are held rather than on disk. The parser finds copybooks in the index built
+// here and reads them through `readText`, so a COPY resolves to a held file or to a system copy
+// library outside the tree, never to a file on disk that happens to sit under this tree's root.
+function heldTree({ kind, root, store, rel, systemDirs = [], copyDirs = null, classify = heldClassifier,
+  symlinks = { followed: 0, outside: 0, broken: 0 }, unreadableDirs = [] }) {
+  const index = new Map();
+  const dirs = new Set(copyDirs || []);
+  for (const p of store.keys()) {
+    index.set(resolve(root, p).toLowerCase(), p);
+    if (!copyDirs && /\.(cpy|copy|inc|cbl|cob)$/i.test(p)) dirs.add(dirname(resolve(root, p)));
+  }
+  index.root = root;
+  const top = resolve(root);
+  const absent = (p) => Object.assign(new Error(`no such file in this tree: ${p}`), { code: 'ENOENT' });
+  const bytes = (p) => {
+    const b = store.get(p);
+    if (!b) throw absent(p);
+    return b;
+  };
+  const kindOf = classify(bytes);
+  classifyUnder(top, kindOf);
+  const text = (p) => decodeSource(bytes(p));
+  const readText = (p) => {
+    if (store.has(p)) return text(p).text;
+    const r = resolve(root, p);
+    if (r === top || r.startsWith(top + sep)) throw absent(p);
+    return readSource(p).text;
+  };
   return {
-    kind: 'memory',
+    kind,
     root,
-    index: { index: store, copyDirs: [], unreadableDirs: [], symlinks: { followed: 0, outside: 0, broken: 0 } },
+    kindOf,
+    index: { index, copyDirs: [...dirs].sort(), unreadableDirs, symlinks },
     list: () => [...store.keys()].sort(),
-    bytes: (p) => {
-      const b = store.get(p);
-      if (!b) throw Object.assign(new Error(`no such file in this tree: ${p}`), { code: 'ENOENT' });
-      return b;
-    },
-    text: (p) => {
-      const b = store.get(p);
-      if (!b) throw Object.assign(new Error(`no such file in this tree: ${p}`), { code: 'ENOENT' });
-      return { text: b.toString('latin1'), encoding: 'latin1' };
-    },
-    rel: (p) => String(p).replace(/\\/g, '/').replace(new RegExp('^' + root.replace(/[.*+?^${}()|[\]\\]/g, '\\$&') + '/?'), ''),
+    bytes,
+    text,
+    rel,
     // Containment by construction: a path this tree does not hold is a path outside it.
     contains: (p) => store.has(p),
-    parse: () => {
-      throw Object.assign(
-        new Error('a memory tree cannot parse: COPY resolution reads directories from disk'),
-        { code: 'ETREEPARSE' },
-      );
-    },
+    parse: (file, t) => parseSource(t ?? text(file).text, file, {
+      format: 'auto',
+      includeDirs: [...dirs].sort(),
+      fileIndex: index,
+      mainDir: dirname(resolve(root, file)),
+      copyFormat: 'auto',
+      systemDirs,
+      readText,
+    }),
+  };
+}
+
+// A tree that was never on disk, for tests. `files` maps a path to its contents, as a string or a
+// Buffer. Paths are used exactly as given, so a test reads the way it writes.
+export function memoryTree(files, { root = '/memory', systemDirs = [] } = {}) {
+  const store = new Map(Object.entries(files).map(([p, v]) => [p, Buffer.isBuffer(v) ? v : Buffer.from(v, 'latin1')]));
+  const rel = (p) => String(p).replace(/\\/g, '/').replace(new RegExp('^' + root.replace(/[.*+?^${}()|[\]\\]/g, '\\$&') + '/?'), '');
+  return heldTree({ kind: 'memory', root, store, rel, systemDirs });
+}
+
+// A git revision, read with plumbing into memory: nothing is written to disk. Its root is a path
+// beside the repository that does not exist, `<repo>@<tree>`, so every path it answers for is
+// absolute and none of them can be found on disk. The tree object is the containment: it cannot
+// name a path outside itself, and a path that would climb, name a drive or stream, or enter .git
+// is left out, as it is when a revision is written to disk. Links and submodules are not read.
+export function gitTree(repo, ref, opts = {}) {
+  const oid = treeOf(repo, ref);
+  const { blobs, links } = revisionBlobs(repo, oid, ref);
+  const root = `${resolve(repo)}@${oid.slice(0, 12)}`;
+  const store = new Map();
+  for (const b of blobs) {
+    const parts = treePathParts(b.path);
+    if (parts) store.set(resolve(root, ...parts), b.bytes);
+  }
+  const tree = heldTree({ kind: 'git', root, store, rel: (p) => relPath(root, p), systemDirs: opts.systemDirs || [],
+    symlinks: { followed: 0, outside: 0, broken: 0, notRead: links } });
+  return Object.assign(tree, { ref, oid });
+}
+
+// z/OSMF's record mode (zowe --record) writes each record after its length, four bytes big-endian,
+// and nothing between records. Read as that only when the lengths chain exactly to the end of the
+// file, and given back as lines; anything else is the file as it is.
+const MAX_LRECL = 32760;
+export function unframeRecords(buf) {
+  let records = 0;
+  for (let at = 0; at < buf.length; records++) {
+    if (buf.length - at < 4) return buf;
+    const n = buf.readUInt32BE(at);
+    if (n > MAX_LRECL || buf.length - at - 4 < n) return buf;
+    at += 4 + n;
+  }
+  if (!records || records * 4 === buf.length) return buf;
+  const lines = (end) => {
+    const out = Buffer.alloc(buf.length - records * 3);
+    let w = 0;
+    for (let at = 0; at < buf.length;) {
+      const n = buf.readUInt32BE(at);
+      w += buf.copy(out, w, at + 4, at + 4 + n);
+      out[w++] = end;
+      at += 4 + n;
+    }
+    return out;
   };
+  const ebcdic = lines(0x15);
+  return looksEbcdic(ebcdic) ? ebcdic : lines(0x0A);
+}
+
+// A member is read from the file the walk found. A link put in its place since, or a file that is
+// not a regular one, is refused rather than followed or waited on.
+function readMember(file) {
+  const fd = openSync(file, constants.O_RDONLY | (constants.O_NOFOLLOW || 0) | (constants.O_NONBLOCK || 0));
+  try {
+    const st = fstatSync(fd);
+    if (!st.isFile()) throw Object.assign(new Error(`not a regular file: ${file}`), { code: 'ENOTFILE' });
+    const buf = Buffer.alloc(st.size);
+    let n = 0;
+    while (n < st.size) {
+      const r = readSync(fd, buf, n, st.size - n, n);
+      if (!r) break;
+      n += r;
+    }
+    return unframeRecords(buf.subarray(0, n));
+  } finally { closeSync(fd); }
+}
+
+const QUALIFIER = /^[A-Z@#$][A-Z0-9@#$-]{0,7}$/;
+const MEMBER = /^([A-Z@#$][A-Z0-9@#$]{0,7})(?:\.[^.]*)?$/i;
+const MAX_LISTED = 8;
+
+// Partitioned data sets exported to a directory, a member to a file: what
+// `zowe zos-files download all-members` writes (ibmuser/new/cntl/member.txt, lower-cased unless
+// --preserve-original-letter-case, with whatever extension -e gave), or a directory named for the
+// data set with its members copied in. Each member is held as `DATA.SET.NAME/MEMBER` under a root
+// beside the export that does not exist, so a finding names the member as z/OS does however it was
+// downloaded, and nothing under that root is ever read from disk.
+//
+// The extension is the download's choice, not the language's, so members are classified by their
+// contents. A file whose directory is not a data set name or whose name is not a member name is not
+// read and is counted, as is a second file for a member already held. Every data set that holds
+// anything but JCL and maps is a copy library, searched in name order: the export does not say
+// what SYSLIB concatenated.
+export function pdsExportTree(dir, opts = {}) {
+  const idx = buildFileIndex(dir);
+  const top = realpathSync(dir);
+  const root = `${resolve(dir)}@pds`;
+  const origins = new Map();
+  const found = new Map();
+  const notMembers = [];
+  const duplicates = [];
+  const dataSets = new Set();
+  for (const file of [...idx.index.values()].sort()) {
+    const where = relative(dir, dirname(file)).split(sep).filter(Boolean);
+    const dsn = where.join('.').toUpperCase();
+    const member = MEMBER.exec(basename(file));
+    if (!member || dsn.length > 44 || (dsn && !dsn.split('.').every((q) => QUALIFIER.test(q)))) {
+      notMembers.push(relPath(dir, file));
+      continue;
+    }
+    const at = resolve(root, ...(dsn ? [dsn] : []), member[1].toUpperCase());
+    if (found.has(at)) { duplicates.push(`${relPath(dir, file)}: ${relPath(root, at)} is ${found.get(at)}`); continue; }
+    let real;
+    try { real = realpathSync(file); } catch { real = file; }
+    if (real !== top && !real.startsWith(top + sep)) { notMembers.push(relPath(dir, file)); continue; }
+    origins.set(at, real);
+    found.set(at, relPath(dir, file));
+    dataSets.add(dirname(at));
+  }
+  const kinds = new Map();
+  const byKind = {};
+  const copyDirs = new Set();
+  for (const [at, file] of origins) {
+    let kind = null;
+    try { kind = kindOfBytes(readMember(file)); } catch { kind = null; }
+    kinds.set(at, kind);
+    byKind[kind || 'other'] = (byKind[kind || 'other'] || 0) + 1;
+    if (kind !== 'jcl' && kind !== 'bms') copyDirs.add(dirname(at));
+  }
+  const store = {
+    keys: () => origins.keys(),
+    has: (p) => origins.has(p),
+    get: (p) => (origins.has(p) ? readMember(origins.get(p)) : undefined),
+  };
+  const tree = heldTree({
+    kind: 'pds-export', root, store, rel: (p) => relPath(root, p), systemDirs: opts.systemDirs || [],
+    copyDirs: [...copyDirs], classify: () => (p) => kinds.get(p) ?? null, symlinks: idx.symlinks,
+    unreadableDirs: idx.unreadableDirs.map((d) => resolve(root, relative(dir, d))),
+  });
+  return Object.assign(tree, {
+    dir,
+    origin: (p) => found.get(p) ?? null,
+    export: {
+      dataSets: dataSets.size,
+      members: origins.size,
+      byKind,
+      ...(notMembers.length ? { notMembers: notMembers.length, notMembersListed: notMembers.slice(0, MAX_LISTED) } : {}),
+      ...(duplicates.length ? { duplicates: duplicates.length, duplicatesListed: duplicates.slice(0, MAX_LISTED) } : {}),
+    },
+  });
 }
 
 // What a rule set calls. Given a tree it uses it; given none it builds the directory one, so every

package/lib/layout.mjs

@@ -0,0 +1,202 @@
+// SPDX-License-Identifier: AGPL-3.0-or-later
+// How a data description is laid out in storage: the bytes a PICTURE and USAGE take, where each item
+// starts within its record, and how wide a report group prints. Items are { level, occurs, redefines,
+// sync, children, ... } as lib/parser.mjs builds them.
+
+// The binary sizes a dialect gives COMP by digit count.
+export const BINARY_SIZE = { default: '1-2-4-8', ibm: '2-4-8', mf: '1--8' };
+
+function picInfo(pic, constants) {
+  const info = { digits: 0, display: 0, signed: false, alphanumeric: false, national: false };
+  const re = /(.)\(([A-Za-z0-9_-]+)\)|(.)/g;
+  let m;
+  while ((m = re.exec(pic))) {
+    const ch = (m[1] || m[3]).toUpperCase();
+    let n = 1;
+    if (m[1]) {
+      const raw = m[2];
+      n = /^\d+$/.test(raw) ? Number(raw) : Number(constants && constants.get(raw.toUpperCase()));
+      if (!Number.isFinite(n) || n < 0) n = 0;
+    }
+    if (ch === 'S') { info.signed = true; continue; }
+    if (ch === 'V' || ch === 'P') continue;
+    if (ch === '9') { info.digits += n; info.display += n; continue; }
+    if (ch === 'N' || ch === 'G') { info.national = true; info.display += 2 * n; continue; }
+    if (ch === 'X' || ch === 'A') info.alphanumeric = true;
+    info.display += n;
+  }
+  return info;
+}
+
+// Storage bytes of a literal: hexadecimal literals hold one byte per two digits, a Z literal adds a
+// terminating null, a national literal holds two bytes per character.
+export function literalBytes(tok) {
+  const prefix = tok.prefix || '';
+  if (prefix === 'X' || prefix === 'BX' || prefix === 'NX') return Math.floor(tok.v.length / (prefix === 'NX' ? 4 : 2)) * (prefix === 'NX' ? 2 : 1);
+  if (prefix === 'Z') return tok.v.length + 1;
+  if (prefix === 'N' || prefix === 'NC' || prefix === 'U') return tok.v.length * 2;
+  return tok.v.length;
+}
+
+function binaryBytes(digits, scheme) {
+  if (scheme === '2-4-8') return digits <= 4 ? 2 : digits <= 9 ? 4 : 8;
+  if (scheme === '1--8') return [1, 1, 1, 2, 2, 3, 3, 4, 4, 4, 5, 5, 6, 6, 6, 7, 7, 8, 8][Math.min(digits, 18)] || 8;
+  return digits <= 2 ? 1 : digits <= 4 ? 2 : digits <= 9 ? 4 : 8;
+}
+
+function elementarySize(item, scheme, constants) {
+  if (!item.picture && !item.usage) {
+    const constBytes = constants && constants.textBytes;
+    const fromValue = item.values.reduce((n, v) => n + (v.t === 'lit' ? literalBytes(v) : v.t === 'word' && constBytes && constBytes.has(v.u) ? constBytes.get(v.u) : 0), 0);
+    if (item.section === 'SCREEN' && !fromValue && item.screenRefItem) return item.screenRefItem.size || 0;
+    if (item.section === 'SCREEN') return Math.max(1, fromValue);
+    if (fromValue) return fromValue;
+  }
+  if (item.section === 'SCREEN' && !item.picture) return 1;
+  const usage = item.effectiveUsage || 'DISPLAY';
+  const p = item.picture ? picInfo(item.picture, constants) : null;
+  const digits = p ? p.digits : 0;
+  const u = usage.replace('COMPUTATIONAL', 'COMP');
+  if (u === 'COMP-1' || u === 'FLOAT-SHORT') return 4;
+  if (u === 'COMP-2' || u === 'FLOAT-LONG' || u === 'FLOAT-DECIMAL-16') return 8;
+  if (u === 'FLOAT-DECIMAL-34') return 16;
+  if (u === 'INDEX') return 4;
+  if (u === 'POINTER' || u === 'PROGRAM-POINTER' || u === 'FUNCTION-POINTER' || u === 'PROCEDURE-POINTER') return 8;
+  if (u === 'BINARY-CHAR') return 1;
+  if (u === 'BINARY-SHORT' || u === 'SIGNED-SHORT' || u === 'UNSIGNED-SHORT') return 2;
+  if (u === 'BINARY-LONG' || u === 'BINARY-INT' || u === 'SIGNED-INT' || u === 'UNSIGNED-INT') return 4;
+  if (u === 'BINARY-DOUBLE' || u === 'BINARY-LONG-LONG' || u === 'BINARY-C-LONG' || u === 'SIGNED-LONG' || u === 'UNSIGNED-LONG') return 8;
+  if (u === 'COMP-3' || u === 'PACKED-DECIMAL') return Math.floor(digits / 2) + 1;
+  if (u === 'COMP-6') return Math.ceil(digits / 2);
+  if (u === 'COMP-X' || u === 'COMP-N' || (u === 'COMP-5' && p && p.alphanumeric)) {
+    if (p && p.alphanumeric) return p.display;
+    return Math.max(1, Math.ceil((digits * Math.log(10)) / Math.log(256)));
+  }
+  if (u === 'COMP' || u === 'COMP-4' || u === 'COMP-5' || u === 'BINARY') return binaryBytes(digits, scheme);
+  let len = p ? p.display : 0;
+  if (item.signSeparate && p && p.signed) len++;
+  return len;
+}
+
+// SYNCHRONIZED aligns a binary, floating-point, pointer or index item to its own length; packed,
+// COMP-X and display items are not moved.
+const ALIGNED_USAGE = /^(COMP|COMP-[1245]|BINARY(-[A-Z-]+)?|FLOAT-[A-Z0-9-]+|(PROGRAM-|FUNCTION-|PROCEDURE-)?POINTER|INDEX|(UN)?SIGNED-[A-Z]+)$/;
+
+export function computeSizes(roots, scheme, constants) {
+  // `at` is where the item starts, counted from the start of its record, because the compiler aligns
+  // a SYNCHRONIZED item against the record and puts the slack inside the group that holds it. A
+  // table's entry is laid out from its own start and rounded up to its widest alignment, so every
+  // occurrence aligns alike.
+  const visit = (item, inheritedUsage, inheritedSignSeparate, at) => {
+    item.effectiveUsage = item.usage || inheritedUsage || null;
+    if (inheritedSignSeparate && !item.signExplicit) item.signSeparate = true;
+    const structural = item.children.filter(c => c.level !== 88 && c.level !== 66 && c.level !== 78);
+    if (!structural.length) {
+      const one = elementarySize(item, scheme, constants);
+      item.contributes = one * item.occurs;
+      // The listing prints the whole table only for POINTER and INDEX; every other usage prints one occurrence.
+      const wholeTable = /^(POINTER|INDEX)$/.test(item.effectiveUsage || '');
+      item.size = wholeTable ? item.contributes : one;
+      const usage = (item.effectiveUsage || '').replace('COMPUTATIONAL', 'COMP');
+      item.align = item.sync && ALIGNED_USAGE.test(usage) ? Math.min(one, 8) : 1;
+      item.maxAlign = item.align;
+      return;
+    }
+    const base = item.occurs > 1 ? 0 : at;
+    let offset = 0;
+    let end = 0;
+    let maxAlign = 1;
+    const startOf = new Map();
+    for (const c of structural) {
+      if (c.redefines) {
+        // REDEFINES names a SIBLING. Resolving it by name across the whole program reached the
+        // first item of that name anywhere, which crossed records.
+        const known = startOf.has(c.redefines);
+        visit(c, item.effectiveUsage, item.signSeparate, base + (known ? startOf.get(c.redefines) : offset));
+        const start = known ? startOf.get(c.redefines) : offset - c.contributes;
+        startOf.set(c.name, start);
+        c.localStart = start;
+        end = Math.max(end, start + c.contributes);
+        // A REDEFINES larger than the item it redefines pushes the next sibling past its end.
+        offset = Math.max(offset, start + c.contributes);
+        maxAlign = Math.max(maxAlign, c.maxAlign);
+        continue;
+      }
+      visit(c, item.effectiveUsage, item.signSeparate, base + offset);
+      const skew = (base + offset) % c.align;
+      if (c.align > 1 && skew) offset += c.align - skew;
+      startOf.set(c.name, offset);
+      c.localStart = offset;
+      offset += c.contributes;
+      end = Math.max(end, offset);
+      maxAlign = Math.max(maxAlign, c.maxAlign);
+    }
+    if (item.occurs > 1 && end % maxAlign) end += maxAlign - (end % maxAlign);
+    item.size = end * item.occurs;
+    item.contributes = item.size;
+    item.align = 1;
+    item.maxAlign = maxAlign;
+  };
+  for (const r of roots) visit(r, null, false, 0);
+  // Offsets are assigned after sizing: a child's absolute start depends on its parent's, which is
+  // only known once the parent's own siblings have been laid out.
+  const place = (item, at) => {
+    item.offset = at;
+    for (const c of item.children) if (c.localStart != null) place(c, at + c.localStart);
+  };
+  for (const r of roots) place(r, 0);
+}
+
+// A report group is laid out by column, not by adding up its items: a line is as wide as the column
+// its rightmost item ends in, COLUMN PLUS counting on from the end of the item before. A group holds
+// its lines one after another, and every 01 group of a report, like the file the report is written
+// to, is as large as the largest group.
+export function layoutReport(rd) {
+  const structural = (x) => x.children.filter(c => c.level !== 88 && c.level !== 66 && c.level !== 78);
+  const mark = (x) => { x.rd = rd; for (const c of x.children) mark(c); };
+  for (const g of rd.groups) mark(g);
+  let width = 0;
+  // Items printed at the same column under PRESENT WHEN each keep their own storage, so a line is
+  // never smaller than its items laid end to end.
+  const lineWidth = (line) => {
+    let last = 0;
+    let right = 0;
+    let total = 0;
+    const place = (c) => {
+      const len = c.contributes || c.size || 0;
+      const start = c.rwColumn ? (c.rwColumn.at != null ? c.rwColumn.at : last + c.rwColumn.plus) : last + 1;
+      last = start + len - 1;
+      right = Math.max(right, last);
+      total += len;
+    };
+    const walk = (x) => { for (const c of structural(x)) { if (structural(c).length) walk(c); else place(c); } };
+    if (structural(line).length) walk(line);
+    else place(line);
+    return Math.max(right, total);
+  };
+  // The entries that open a line; a group with no LINE clause anywhere in it is one line.
+  for (const g of rd.groups) {
+    const lines = [];
+    const collect = (x) => { if (x.rwLine) { lines.push(x); return; } for (const c of structural(x)) collect(c); };
+    collect(g);
+    if (!lines.length) lines.push(g);
+    let total = 0;
+    for (const line of lines) {
+      const w = lineWidth(line);
+      if (line !== g) { line.size = w; line.contributes = w * (line.occurs || 1); }
+      total += w;
+    }
+    width = Math.max(width, total);
+  }
+  for (const g of rd.groups) { g.size = width; g.contributes = width; }
+  rd.width = width;
+}
+
+// Whether an item's bytes are read as decimal digits: zoned (DISPLAY with a numeric picture) or
+// packed. A group, an edited picture, binary and floating point are not.
+export function holdsDecimal(it) {
+  if ((it.children || []).some((c) => c.level !== 88)) return false;
+  const usage = String(it.effectiveUsage || 'DISPLAY').replace('COMPUTATIONAL', 'COMP');
+  if (usage === 'COMP-3' || usage === 'PACKED-DECIMAL' || usage === 'COMP-6') return true;
+  return usage === 'DISPLAY' && !!it.picture && /9/.test(it.picture) && /^[S9VP()0-9]+$/i.test(it.picture);
+}