debian/copyright: Fix and extend based on CFD research.

[runlisp] / lib.c

/* -*-c-*-
 *
 * Common definitions for `runlisp'
 *
 * (c) 2020 Mark Wooding
 */

/*----- Licensing notice --------------------------------------------------*
 *
 * This file is part of Runlisp, a tool for invoking Common Lisp scripts.
 *
 * Runlisp is free software: you can redistribute it and/or modify it
 * under the terms of the GNU General Public License as published by the
 * Free Software Foundation; either version 3 of the License, or (at your
 * option) any later version.
 *
 * Runlisp is distributed in the hope that it will be useful, but WITHOUT
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 * for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with Runlisp.  If not, see <https://www.gnu.org/licenses/>.
 */

/*----- Header files ------------------------------------------------------*/

#include "config.h"

#include <assert.h>

#include <ctype.h>
#include <errno.h>
#include <stdarg.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#include <unistd.h>

#include "lib.h"

/*----- Diagnostic utilities ----------------------------------------------*/

const char *progname = "???";
        /* Our program name, for use in error messages. */

/* Set `progname' from the pathname in PROG (typically from `argv[0]'). */
void set_progname(const char *prog)
{
  const char *p;

  p = strrchr(prog, '/');
  progname = p ? p + 1 : prog;
}

/* Report an error or warning in Unix style, given a captured argument
 * cursor.
 */
void vmoan(const char *msg, va_list ap)
{
  fprintf(stderr, "%s: ", progname);
  vfprintf(stderr, msg, ap);
  fputc('\n', stderr);
}

/* Issue a warning message. */
void moan(const char *msg, ...)
  { va_list ap; va_start(ap, msg); vmoan(msg, ap); va_end(ap); }

/* Issue a fatal error message and exit unsuccessfully. */
void lose(const char *msg, ...)
  { va_list ap; va_start(ap, msg); vmoan(msg, ap); va_end(ap); exit(127); }

/*----- Memory allocation -------------------------------------------------*/

/* Allocate and return a pointer to N bytes, or report a fatal error.
 *
 * Release the pointer using `free' as usual.  If N is zero, returns null
 * (but you are not expected to check for this).
 */
void *xmalloc(size_t n)
{
  void *p;

  if (!n) return (0);
  p = malloc(n); if (!p) lose("failed to allocate memory");
  return (p);
}

/* Resize the block at P (from `malloc' or `xmalloc') to be N bytes long.
 *
 * The block might (and probably will) move, so it returns the new address.
 * If N is zero, then the block is freed (if necessary) and a null pointer
 * returned; otherwise, if P is null then a fresh block is allocated.  If
 * allocation fails, then a fatal error is reported.
 */
void *xrealloc(void *p, size_t n)
{
  if (!n) { free(p); return (0); }
  else if (!p) return (xmalloc(n));
  p = realloc(p, n); if (!p) lose("failed to allocate memory");
  return (p);
}

/* Allocate and return a copy of the N-byte string starting at P.
 *
 * The new string is null-terminated, though P need not be.  If allocation
 * fails, then a fatal error is reported.
 */
char *xstrndup(const char *p, size_t n)
{
  char *q = xmalloc(n + 1);

  memcpy(q, p, n); q[n] = 0;
  return (q);
}

/* Allocate and return a copy of the null-terminated string starting at P.
 *
 * If allocation fails, then a fatal error is reported.
 */
char *xstrdup(const char *p) { return (xstrndup(p, strlen(p))); }

/*----- Dynamic strings ---------------------------------------------------*/

/* Initialize the string D.
 *
 * Usually you'd use the static initializer `DSTR_INIT'.
 */
void dstr_init(struct dstr *d) { d->p = 0; d->len = d->sz = 0; }

/* Reset string D so it's empty again. */
void dstr_reset(struct dstr *d) { d->len = 0; }

/* Ensure that D has at least N unused bytes available. */
void dstr_ensure(struct dstr *d, size_t n)
{
  size_t need = d->len + n, newsz;

  if (need <= d->sz) return;
  newsz = d->sz ? 2*d->sz : 16;
  while (newsz < need) newsz *= 2;
  d->p = xrealloc(d->p, newsz); d->sz = newsz;
}

/* Release the memory held by D.
 *
 * It must be reinitialized (e.g., by `dstr_init') before it can be used
 * again.
 */
void dstr_release(struct dstr *d) { free(d->p); }

/* Append the N-byte string at P to D.
 *
 * P need not be null-terminated.  D will not be null-terminated
 * afterwards.
 */
void dstr_putm(struct dstr *d, const void *p, size_t n)
  { dstr_ensure(d, n); memcpy(d->p + d->len, p, n); d->len += n; }

/* Append the null-terminated string P to D.
 *
 * D /is/ guaranteed to be null-terminated after this.
 */
void dstr_puts(struct dstr *d, const char *p)
{
  size_t n = strlen(p);

  dstr_ensure(d, n + 1);
  memcpy(d->p + d->len, p, n + 1);
  d->len += n;
}

/* Append the single character CH to D.
 *
 * D will not be null-terminated afterwards.
 */
void dstr_putc(struct dstr *d, int ch)
  { dstr_ensure(d, 1); d->p[d->len++] = ch; }

/* Append N copies of the character CH to D.
 *
 * D will not be null-terminated afterwards.
 */
void dstr_putcn(struct dstr *d, int ch, size_t n)
  { dstr_ensure(d, n); memset(d->p + d->len, ch, n); d->len += n; }

/* Null-terminate the string D.
 *
 * This doesn't change the length of D.  If further stuff is appended then
 * the null terminator will be overwritten.
 */
void dstr_putz(struct dstr *d)
  { dstr_ensure(d, 1); d->p[d->len] = 0; }

/* Append stuff to D, determined by printf(3) format string P and argument
 * tail AP.
 *
 * D will not be null-terminated afterwards.
 */
void dstr_vputf(struct dstr *d, const char *p, va_list ap)
{
  va_list ap2;
  size_t r;
  int n;

  r = d->sz - d->len;
  va_copy(ap2, ap);
  n = vsnprintf(d->p + d->len, r, p, ap2); assert(n >= 0);
  va_end(ap2);
  if (n >= r) {
    dstr_ensure(d, n + 1); r = d->sz - d->len;
    n = vsnprintf(d->p + d->len, r, p, ap); assert(n >= 0); assert(n < r);
  }
  d->len += n;
}

/* Append stuff to D, determined by printf(3) format string P and arguments.
 *
 * D will not be null-terminated afterwards.
 */
PRINTF_LIKE(2, 3) void dstr_putf(struct dstr *d, const char *p, ...)
  { va_list ap; va_start(ap, p); dstr_vputf(d, p, ap); va_end(ap); }

/* Append the next input line from FP to D.
 *
 * Return 0 on success, or -1 if reading immediately fails or encounters
 * end-of-file (call ferror(3) to distinguish).  Any trailing newline is
 * discarded: it is not possible to determine whether the last line was ended
 * with a newline.  D is guaranteed to be null-terminated afterwards.
 */
int dstr_readline(struct dstr *d, FILE *fp)
{
  char *p = d->p + d->len, *lim = d->p + d->sz;
  size_t n;
  int ch;

  ch = getc(fp); if (ch == EOF) return (-1);
  for (;;) {
    if (p == lim) {
      n = d->len = p - d->p; dstr_ensure(d, 1);
      p = d->p + n; lim = d->p + d->sz;
    }
    if (ch == EOF || ch == '\n') break;
    *p++ = ch; ch = getc(fp);
  }
  d->len = p - d->p; *p++ = 0;
  return (0);
}

/*----- Dynamic vectors of strings ----------------------------------------*/

/* Initialize the vector AV.
 *
 * Usually you'd use the static initializer `ARGV_INIT'.
 */
void argv_init(struct argv *av)
  { av->v = 0; av->o = av->n = av->sz = 0; }

/* Reset the vector AV so that it's empty again. */
void argv_reset(struct argv *av) { av->n = 0; }

/* Ensure that AV has at least N unused slots at the end. */
void argv_ensure(struct argv *av, size_t n)
{
  size_t need = av->n + av->o + n, newsz;

  if (need <= av->sz) return;
  newsz = av->sz ? 2*av->sz : 8;
  while (newsz < need) newsz *= 2;
  av->v = xrealloc(av->v - av->o, newsz*sizeof(char *)); av->v += av->o;
  av->sz = newsz;
}

/* Ensure that AV has at least N unused slots at the /start/. */
void argv_ensure_offset(struct argv *av, size_t n)
{
  size_t newoff;

  /* Stupid version.  We won't, in practice, be prepending lots of stuff, so
   * avoid the extra bookkeeping involved in trying to make a double-ended
   * extendable array asymptotically efficient.
   */
  if (av->o >= n) return;
  newoff = 16;
  while (newoff < n) newoff *= 2;
  argv_ensure(av, newoff - av->o);
  memmove(av->v + newoff - av->o, av->v, av->n*sizeof(char *));
  av->v += newoff - av->o; av->o = newoff;
}

/* Release the memory held by AV.
 *
 * It must be reinitialized (e.g., by `argv_init') before it can be used
 * again.
 */
void argv_release(struct argv *av) { free(av->v - av->o); }

/* Append the pointer P to AV. */
void argv_append(struct argv *av, char *p)
  { argv_ensure(av, 1); av->v[av->n++] = p; }

/* Append a null pointer to AV, without extending the vactor length.
 *
 * The null pointer will be overwritten when the next string is appended.
 */
void argv_appendz(struct argv *av)
  { argv_ensure(av, 1); av->v[av->n] = 0; }

/* Append a N-element vector V of pointers to AV. */
void argv_appendn(struct argv *av, char *const *v, size_t n)
{
  argv_ensure(av, n);
  memcpy(av->v + av->n, v, n*sizeof(const char *));
  av->n += n;
}

/* Append the variable-length vector BV to AV. */
void argv_appendav(struct argv *av, const struct argv *bv)
  { argv_appendn(av, bv->v, bv->n); }

/* Append the pointers from a variable-length argument list AP to AV.
 *
 * The list is terminated by a null pointer.
 */
void argv_appendv(struct argv *av, va_list ap)
{
  char *p;
  for (;;) { p = va_arg(ap, char *); if (!p) break; argv_append(av, p); }
}

/* Append the argument pointers, terminated by a null pointer, to AV. */
void argv_appendl(struct argv *av, ...)
  { va_list ap; va_start(ap, av); argv_appendv(av, ap); va_end(ap); }

/* Prepend the pointer P to AV. */
void argv_prepend(struct argv *av, char *p)
  { argv_ensure_offset(av, 1); *--av->v = p; av->o--; av->n++; }

/* Prepend a N-element vector V of pointers to AV. */
void argv_prependn(struct argv *av, char *const *v, size_t n)
{
  argv_ensure_offset(av, n);
  av->o -= n; av->v -= n; av->n += n;
  memcpy(av->v, v, n*sizeof(const char *));
}

/* Prepend the variable-length vector BV to AV. */
void argv_prependav(struct argv *av, const struct argv *bv)
  { argv_prependn(av, bv->v, bv->n); }

/* Prepend the pointers from a variable-length argument list AP to AV.
 *
 * The list is terminated by a null pointer.
 */
void argv_prependv(struct argv *av, va_list ap)
{
  char *p, **v;
  size_t n = 0;

  for (;;) {
    p = va_arg(ap, char *); if (!p) break;
    argv_prepend(av, p); n++;
  }
  v = av->v;
  while (n >= 2) {
    p = v[0]; v[0] = v[n - 1]; v[n - 1] = p;
    v++; n -= 2;
  }
}

/* Prepend the argument pointers, terminated by a null pointer, to AV. */
void argv_prependl(struct argv *av, ...)
  { va_list ap; va_start(ap, av); argv_prependv(av, ap); va_end(ap); }

/*----- Treaps ------------------------------------------------------------*/

/* Return nonzero if the AN-byte string A is strictly precedes the BN-byte
 * string B in a lexicographic ordering.
 *
 * All comparison of keys is handled by this function.
 */
static int str_lt(const char *a, size_t an, const char *b, size_t bn)
{
  /* This is a little subtle.  We need only compare the first N bytes of the
   * strings, where N is the length of the shorter string.  If this
   * distinguishes the two strings, then we're clearly done.  Otherwise, if
   * the prefixes are equal then the shorter string is the smaller one.  If
   * the two strings are the same length, then they're equal.
   *
   * Hence, if A is the strictly shorter string, then A precedes B if A
   * precedes or matches the prefix of B; otherwise A only precedes B if A
   * strictly precedes the prefix of B.
   */
  if (an < bn) return (MEMCMP(a, <=, b, an));
  else return (MEMCMP(a, <, b, bn));
}

/* Initialize the treap T.
 *
 * Usually you'd use the static initializer `TREAP_INIT'.
 */
void treap_init(struct treap *t) { t->root = 0; }

/* Look up the KN-byte key K in the treap T.
 *
 * Return a pointer to the matching node if one was found, or null otherwise.
 */
void *treap_lookup(const struct treap *t, const char *k, size_t kn)
{
  struct treap_node *n = t->root, *candidate = 0;

  /* This is a simple prototype for some of the search loops we'll encounter
   * later.  Notice that we use a strict one-sided comparison, rather than
   * the more conventional two-sided comparison.
   *
   * The main loop will find the largest key not greater than K.
   */
  while (n)
    /* Compare the node's key against our key.  If the node is too large,
     * then we ignore it and move left.  Otherwise remember this node for
     * later, and move right to see if we can find a better, larger node.
     */

    if (str_lt(k, kn, n->k, n->kn)) n = n->left;
    else { candidate = n; n = n->right; }

  /* If the candidate node is less than our key then we failed.  Otherwise,
   * by trichotomy, we have found the correct node.
   */
  if (!candidate || str_lt(candidate->k, candidate->kn, k, kn)) return (0);
  return (candidate);
}

/* Look up the KN-byte K in the treap T, recording a path in P.
 *
 * This is similar to `treap_lookup', in that it returns the requested node
 * if it already exists, or null otherwise, but it also records in P
 * information to be used by `treap_insert' to insert a new node with the
 * given key if it's not there already.
 */
void *treap_probe(struct treap *t, const char *k, size_t kn,
                  struct treap_path *p)
{
  struct treap_node **nn = &t->root, *candidate = 0;
  unsigned i = 0;

  /* This walk is similar to `treap_lookup' above, except that we also record
   * the address of each node pointer we visit along the way.
   */
  for (;;) {
    assert(i < TREAP_PATHMAX); p->path[i++] = nn;
    if (!*nn) break;
    if (str_lt(k, kn, (*nn)->k, (*nn)->kn)) nn = &(*nn)->left;
    else { candidate = *nn; nn = &(*nn)->right; }
  }
  p->nsteps = i;

  /* Check to see whether we found the right node. */
  if (!candidate || str_lt(candidate->k, candidate->kn, k, kn)) return (0);
  return (candidate);
}

/* Insert a new node N into T, associating it with the KN-byte key K.
 *
 * Use the path data P, from `treap_probe', to help with insertion.
 */
void treap_insert(struct treap *t, const struct treap_path *p,
                  struct treap_node *n, const char *k, size_t kn)
{
  size_t i = p->nsteps;
  struct treap_node **nn, **uu, *u;
  unsigned wt;

  /* Fill in the node structure. */
  n->k = xstrndup(k, kn); n->kn = kn;
  n->wt = wt = rand(); n->left = n->right = 0;

  /* Prepare for the insertion.
   *
   * The path actually points to each of the links traversed when searching
   * for the node, starting with the `root' pointer, then the `left' or
   * `right' pointer of the root node, and so on; `nsteps' will always be
   * nonzero, since the path will always pass through the root, and the final
   * step, `path->path[path->nsteps - 1]' will always be the address of a
   * null pointer onto which the freshly inserted node could be hooked in
   * order to satisfy the binary-search-tree ordering.  (Of course, this will
   * likely /not/ satisfy the heap condition, so more work needs to be done.)
   *
   * Throughout, NN is our current candidate for where to attach the node N.
   * As the loop progresses, NN will ascend to links further up the tree, and
   * N will be adjusted to accumulate pieces of the existing tree structure.
   * We'll stop when we find that the parent node's weight is larger than our
   * new node's weight, at which point we can just set *NN = N; or if we run
   * out of steps in the path, in which case *NN is the root pointer.
   */
  assert(i); nn = p->path[--i];
  while (i--) {

    /* Collect the next step in the path, and get the pointer to the node. */
    uu = p->path[i]; u = *uu;

    /* If this node's weight is higher, then we've found the right level and
     * we can stop.
     */
    if (wt <= u->wt) break;

    /* The node U is lighter than our new node N, so we must rotate in order
     * to fix things.  If we were currently planning to hook N as the left
     * subtree of U, then we rotate like this:
     *
     *                  |                   |
     *                  U                  (N)
     *                /   \               /   \
     *             (N)      Z   --->    X       U
     *            /   \                       /   \
     *          X       Y                   Y       Z
     *
     * On the other hand, if we were planning to hook N as the right subtree
     * of U, then we do the opposite rotation:
     *
     *              |                           |
     *              U                          (N)
     *            /   \                       /   \
     *          X      (N)      --->        U       Z
     *                /   \               /   \
     *              Y       Z           X       Y
     *
     * These transformations clearly preserve the ordering of nodes in the
     * binary search tree, and satisfy the heap condition in the subtree
     * headed by N.
     */
    if (nn == &u->left) { u->left = n->right; n->right = u; }
    else { u->right = n->left; n->left = u; }

    /* And this arrangement must be attached to UU, or some higher attachment
     * point.  The subtree satisfies the heap condition, and can be attached
     * safely at the selected place.
     */
    nn = uu;
  }

  /* We've found the right spot.  Hook the accumulated subtree into place. */
  *nn = n;
}

/* Remove the node with the KN-byte K from T.
 *
 * Return the address of the node we removed, or null if it couldn't be
 * found.
 */
void *treap_remove(struct treap *t, const char *k, size_t kn)
{
  struct treap_node **nn = &t->root, **candidate = 0, *n, *l, *r;

  /* Search for the matching node, but keep track of the address of the link
   * which points to our target node.
   */
  while (*nn)
    if (str_lt(k, kn, (*nn)->k, (*nn)->kn)) nn = &(*nn)->left;
    else { candidate = nn; nn = &(*nn)->right; }

  /* If this isn't the right node then give up. */
  if (!candidate || str_lt((*candidate)->k, (*candidate)->kn, k, kn))
    return (0);

  /* Now we need to disentangle the node from the tree.  This is essentially
   * the reverse of insertion: we pretend that this node is suddenly very
   * light, and mutate the tree so as to restore the heap condition until
   * eventually our node is a leaf and can be cut off without trouble.
   *
   * Throughout, the link *NN notionally points to N, but we don't actually
   * update it until we're certain what value it should finally take.
   */
  nn = candidate; n = *nn; l = n->left; r = n->right;
  for (;;)

    /* If its left subtree is empty then we can replace our node by its right
     * subtree and be done.  Similarly, if the right subtree is empty then we
     * replace the node by its left subtree.
     *
     *              |           |               |               |
     *             (N)  --->    R ;            (N)      --->    L
     *            /   \                       /   \
     *          *       R                   L       *
     */
    if (!l) { *nn = r; break; }
    else if (!r) { *nn = l; break; }

    /* Otherwise we need to rotate the pointers so that the heavier of the
     * two children takes the place of our node; thus we have either
     *
     *                  |                   |
     *                 (N)                  L
     *                /   \               /   \
     *              L       R   --->    X      (N)
     *            /   \                       /   \
     *          X       Y                   Y       R
     *
     * or
     *
     *              |                           |
     *             (N)                          R
     *            /   \                       /   \
     *          L       R       --->       (N)      Y
     *                /   \               /   \
     *              X       Y           L       X
     *
     * Again, these transformations clearly preserve the ordering of nodes in
     * the binary search tree, and the heap condition.
     */
    else if (l->wt > r->wt)
      { *nn = l; nn = &l->right; l = n->left = l->right; }
    else
      { *nn = r; nn = &r->left; r = n->right = r->left; }

  /* Release the key buffer, and return the node that we've now detached. */
  free(n->k); return (n);
}

/* Initialize an iterator I over T's nodes. */
void treap_start_iter(struct treap *t, struct treap_iter *i)
{
  struct treap_node *n = t->root;
  unsigned sp = 0;

  /* The `stack' in the iterator structure is an empty ascending stack of
   * nodes which have been encountered, and their left subtrees investigated,
   * but not yet visited by the iteration.
   *
   * Iteration begins by stacking the root node, its left child, and so on,
   * At the end of this, the topmost entry on the stack is the least node of
   * the tree, followed by its parent, grandparent, and so on up to the root.
   */
  while (n) {
    assert(sp < TREAP_PATHMAX);
    i->stack[sp++] = n; n = n->left;
  }
  i->sp = sp;
}

/* Return the next node from I, in ascending order by key.
 *
 * If there are no more nodes, then return null.
 */
void *treap_next(struct treap_iter *i)
{
  struct treap_node *n, *o;
  unsigned sp = i->sp;

  /* We say that a node is /visited/ once it's been returned by this
   * iterator.  To traverse a tree in order, then, we traverse its left
   * subtree, visit the tree root, and traverse its right subtree -- which is
   * a fine recursive definition, but we need a nonrecursive implementation.
   *
   * As is usual in this kind of essential structural recursion, we maintain
   * a stack.  The invariant that we'll maintain is as follows.
   *
   *   1. If the stack is empty, then all nodes have been visited.
   *
   *   2, If the stack is nonempty then the topmost entry on the stack is the
   *      least node which has not yet been visited -- and therefore is the
   *      next node to visit.
   *
   *   3. The earlier entries in the stack are, in (top to bottom) order,
   *      those of the topmost node's parent, grandparent, etc., up to the
   *      root, which have not yet been visited.  More specifically, a node
   *      appears in the stack if and only if some node in its left subtree
   *      is nearer the top of the stack.
   *
   * When we initialized the iterator state (in `treap_start_iter' above), we
   * traced a path to the leftmost leaf, stacking the root, its left-hand
   * child, and so on.  The leftmost leaf is clearly the first node to be
   * visited, and its entire ancestry is on the stack since none of these
   * nodes has yet been visited.  (If the tree is empty, then we have done
   * nothing, the stack is empty, and there are no nodes to visit.)  This
   * establishes the base case for the induction.
   */

  /* So, if the stack is empty now, then (1) all of the nodes have been
   * visited and there's nothing left to do.  Return null.
   */
  if (!sp) return (0);

  /* It's clear that, if we pop the topmost element of the stack, visit it,
   * and arrange to reestablish the invariant, then we'll visit the nodes in
   * the correct order, pretty much by definition.
   *
   * So, pop a node off the stack.  This is the node we shall return.  But
   * before we can do that, we must reestablish the above invariant.
   * Firstly, the current node is removed from the stack, because we're about
   * to visit it, and visited nodes don't belong on the stack.  Then there
   * are two cases to consider.
   *
   *   * If the current node's right subtree is not empty, then the next node
   *     to be visited is the leftmost node in that subtree.  All of the
   *     nodes on the stack are ancestors of the current node, and the right
   *     subtree consists of its descendants, so none of them are already on
   *     the stack; and they're all greater than the current node, and
   *     therefore haven't been visited.  Therefore, we must push the current
   *     node's right child, its /left/ child, and so on, proceeding
   *     leftwards until we fall off the bottom of the tree.
   *
   *   * Otherwise, we've finished traversing some subtree.  Either we are
   *     now done, or (3) we have just finished traversing the left subtree
   *     of the next topmost item on the stack.  This must therefore be the
   *     next node to visit.  The rest of the stack is already correct.
   */
  n = i->stack[--sp];
  o = n->right;
  while (o) {
    assert(sp < TREAP_PATHMAX);
    i->stack[sp++] = o; o = o->left;
  }
  i->sp = sp;
  return (n);
}

/* Recursively check the subtree headed by N.
 *
 * No node should have weight greater than MAXWT, to satisfy the heap
 * condition; if LO is not null, then all node keys should be strictly
 * greater than LO, and, similarly, if HI is not null, then all keys should
 * be strictly smaller than HI.
 */
static void check_subtree(struct treap_node *n, unsigned maxwt,
                          const char *klo, const char *khi)
{
  /* Check the heap condition. */
  assert(n->wt <= maxwt);

  /* Check that the key is in bounds.  (Use `strcmp' here to ensure that our
   * own `str_lt' is working correctly.)
   */
  if (klo) assert(STRCMP(n->k, >, klo));
  if (khi) assert(STRCMP(n->k, <, khi));

  /* Check the left subtree.  Node weights must be bounded above by our own
   * weight.  And every key in the left subtree must be smaller than our
   * current key.  We propagate the lower bound.
   */
  if (n->left) check_subtree(n->left, n->wt, klo, n->k);

  /* Finally, check the right subtree.  This time, every key must be larger
   * than our key, and we propagate the upper bound.
   */
  if (n->right) check_subtree(n->right, n->wt, n->k, khi);
}

/* Check the treap structure rules for T. */
void treap_check(struct treap *t)
  { if (t->root) check_subtree(t->root, t->root->wt, 0, 0); }

/* Recursively dump the subtree headed by N, indenting the output lines by
 * IND spaces.
 */
static void dump_node(struct treap_node *n, int ind)
{
  if (n->left) dump_node(n->left, ind + 1);
  printf(";;%*s [%10u] `%s'\n", 2*ind, "", n->wt, n->k);
  if (n->right) dump_node(n->right, ind + 1);
}

/* Dump the treap T to standard output, for debugging purposes. */
void treap_dump(struct treap *t) { if (t->root) dump_node(t->root, 0); }

/*----- Configuration file parsing ----------------------------------------*/

#ifndef DECL_ENVIRON
  extern char **environ;
#endif

/* Advance P past a syntactically valid name, but no further than L.
 *
 * Return the new pointer.  If no name is found, report an error, blaming
 * FILE and LINE; WHAT is an adjective for the kind of name that was
 * expected.
 */
static const char *scan_name(const char *what,
                             const char *p, const char *l,
                             const char *file, unsigned line)
{
  const char *q = p;

  while (q < l &&
         (ISALNUM(*q) || *q == '-' || *q == '_' || *q == '.' || *q == '/' ||
                         *q == '*' || *q == '+' || *q == '%' || *q == '@'))
    q++;
  if (q == p) lose("%s:%u: expected %s name", file, line, what);
  return (q);
}

/* Initialize the configuration state CONF.
 *
 * Usually you'd use the static initializer `CONFIG_INIT'.
 */
void config_init(struct config *conf)
  { treap_init(&conf->sections); }

/* Find and return the section with null-terminated NAME in CONF.
 *
 * If no section is found, the behaviour depends on whether `CF_CREAT' is set
 * in F: if so, an empty section is created and returned; otherwise, a null
 * pointer is returned.
 */
struct config_section *config_find_section(struct config *conf, unsigned f,
                                           const char *name)
  { return (config_find_section_n(conf, f, name, strlen(name))); }

/* Find and return the section with the SZ-byte NAME in CONF.
 *
 * This works like `config_find_section', but with an explicit length for the
 * NAME rather than null-termination.
 */
struct config_section *config_find_section_n(struct config *conf, unsigned f,
                                             const char *name, size_t sz)
{
  struct config_section *sect;
  struct treap_path path;

  if (!(f&CF_CREAT))
    sect = treap_lookup(&conf->sections, name, sz);
  else {
    sect = treap_probe(&conf->sections, name, sz, &path);
    if (!sect) {
      sect = xmalloc(sizeof(*sect));
      if (!conf->head) conf->tail = &conf->head;
      sect->next = 0; *conf->tail = sect; conf->tail = &sect->next;
      sect->parents = 0; sect->nparents = SIZE_MAX;
      treap_init(&sect->vars); treap_init(&sect->cache);
      treap_insert(&conf->sections, &path, &sect->_node, name, sz);
      config_set_var_n(conf, sect, CF_LITERAL, "@name", 5, name, sz);
    }
  }
  return (sect);
}

/* Set the fallback section for CONF to be SECT.
 *
 * That is, if a section has no explicit parents, then by default it will
 * have a single parent which is SECT.  If SECT is null then there is no
 * fallback section, and sections which don't have explicitly specified
 * parents have no parents at all.  (This is the default situation.)
 */
void config_set_fallback(struct config *conf, struct config_section *sect)
  { conf->fallback = sect; }

/* Arrange that SECT has PARENT as its single parent section.
 *
 * If PARENT is null, then arrange that SECT has no parents at all.  In
 * either case, any `@parents' setting will be ignored.
 */
void config_set_parent(struct config_section *sect,
                       struct config_section *parent)
{
  if (!parent)
    sect->nparents = 0;
  else {
    sect->parents = xmalloc(sizeof(*sect->parents));
    sect->parents[0] = parent; sect->nparents = 1;
  }
}

/* Initialize I to iterate over the sections defined in CONF. */
void config_start_section_iter(struct config *conf,
                               struct config_section_iter *i)
  { i->sect = conf->head; }

/* Return the next section from I, in order of creation.
 *
 * If there are no more sections, then return null.
 */
struct config_section *config_next_section(struct config_section_iter *i)
{
  struct config_section *sect;

  sect = i->sect;
  if (sect) i->sect = sect->next;
  return (sect);
}

/* Initialize the `parents' links of SECT, if they aren't set up already.
 *
 * If SECT contains a `@parents' setting then parse it to determine the
 * parents; otherwise use CONF's fallbeck section, as established by
 * `config_set_fallback'.
 */
static void set_config_section_parents(struct config *conf,
                                       struct config_section *sect)
{
  struct config_section *parent;
  struct config_var *var;
  const char *file; unsigned line;
  size_t i, n;
  char *p, *q, *l;
  struct argv av = ARGV_INIT;

  /* If the section already has parents established then there's nothing to
   * do.
   */
  if (sect->nparents != SIZE_MAX) return;

  /* Look up `@parents', without recursion! */
  var = treap_lookup(&sect->vars, "@parents", 8);
  if (!var) {
    /* No explicit setting: use the fallback setting. */

    if (!conf->fallback || conf->fallback == sect)
      sect->nparents = 0;
    else {
      sect->parents = xmalloc(sizeof(*sect->parents)); sect->nparents = 1;
      sect->parents[0] = conf->fallback;
    }
  } else {
    /* Found a `@parents' list: parse it and set the parents list. */

    file = var->file; line = var->line; if (!file) file = "<internal>";

    /* We do this in two phases.  First, we parse out the section names, and
     * record start/limit pointer pairs in `av'.
     */
    p = var->val; l = p + var->n; while (p < l && ISSPACE(*p)) p++;
    while (*p) {
      q = p;
      p = (/*unconst*/ char *)scan_name("parent section", p, l, file, line);
      argv_append(&av, q); argv_append(&av, p);
      while (p < l && ISSPACE(*p)) p++;
      if (p >= l) break;
      if (*p == ',') do p++; while (ISSPACE(*p));
    }

    /* Now that we've finished parsing, we know how many parents we're going
     * to have, so we can allocate the `parents' vector and fill it in.
     */
    sect->nparents = av.n/2;
    sect->parents = xmalloc(sect->nparents*sizeof(*sect->parents));
    for (i = 0; i < av.n; i += 2) {
      n = av.v[i + 1] - av.v[i];
      parent = config_find_section_n(conf, 0, av.v[i], n);
      if (!parent)
        lose("%s:%u: unknown parent section `%.*s'",
             file, line, (int)n, av.v[i]);
      sect->parents[i/2] = parent;
    }
  }

  /* All done. */
  argv_release(&av);
}

/* Find a setting of the SZ-byte variable NAME in CONF, starting from SECT.
 *
 * If successful, return a pointer to the variable; otherwise return null.
 * Inheritance cycles and ambiguous inheritance are diagnosed as fatal
 * errors.
 */
static struct config_var *search_recursive(struct config *conf,
                                           struct config_section *sect,
                                           const char *name, size_t sz)
{
  struct config_cache_entry *cache;
  struct treap_path path;
  struct config_var *var, *v;
  size_t i, j = j;

  /* If the variable is defined locally then we can just return it. */
  var = treap_lookup(&sect->vars, name, sz); if (var) return (var);

  /* If we have no parents then there's no way we can find it. */
  set_config_section_parents(conf, sect);
  if (!sect->parents) return (0);

  /* Otherwise we must visit the section's parents.  We can avoid paying for
   * this on every lookup by using a cache.  If there's already an entry for
   * this variable then we can return the result immediately (note that we
   * cache both positive and negative outcomes).  Otherwise we create a new
   * cache entry, do the full recursive search, and fill in the result when
   * we're done.
   *
   * The cache also helps us detect cycles: we set the `CF_OPEN' flag on a
   * new cache entry when it's first created, and clear it when we fill in
   * the result: if we encounter an open cache entry again, we know that
   * we've found a cycle.
   */
  cache = treap_probe(&sect->cache, name, sz, &path);
  if (!cache) {
    cache = xmalloc(sizeof(*cache)); cache->f = CF_OPEN;
    treap_insert(&sect->cache, &path, &cache->_node, name, sz);
  } else if (cache->f&CF_OPEN)
    lose("inheritance cycle through section `%s'",
         CONFIG_SECTION_NAME(sect));
  else
    return (cache->var);

  /* Recursively search in each parent.  We insist that all parents that find
   * a variable find the same binding; otherwise we declare ambiguous
   * inheritance.
   */
  for (i = 0; i < sect->nparents; i++) {
    v = search_recursive(conf, sect->parents[i], name, sz);
    if (!v);
    else if (!var) { var = v; j = i; }
    else if (var != v)
      lose("section `%s' inherits variable `%s' ambiguously "
           "via `%s' and `%s'",
           CONFIG_SECTION_NAME(sect), CONFIG_VAR_NAME(var),
           CONFIG_SECTION_NAME(sect->parents[j]),
           CONFIG_SECTION_NAME(sect->parents[i]));
  }

  /* All done: fill the cache entry in, clear the open flag, and return the
   * result.
   */
  cache->var = var; cache->f &= ~CF_OPEN;
  return (var);
}

/* Find and return the variable with null-terminated NAME in SECT.
 *
 * If `CF_INHERIT' is set in F, then the function searches the section's
 * parents recursively; otherwise, it only checks to see whether the variable
 * is set directly in SECT.
 *
 * If no variable is found, the behaviour depends on whether `CF_CREAT' is
 * set in F: if so, an empty variable is created and returned; otherwise, a
 * null pointer is returned.
 *
 * Setting both `CF_INHERIT' and `CF_CREAT' is not useful.
 */
struct config_var *config_find_var(struct config *conf,
                                   struct config_section *sect,
                                   unsigned f, const char *name)
  { return (config_find_var_n(conf, sect, f, name, strlen(name))); }

/* Find and return the variable with the given SZ-byte NAME in SECT.
 *
 * This works like `config_find_var', but with an explicit length for the
 * NAME rather than null-termination.
 */
struct config_var *config_find_var_n(struct config *conf,
                                     struct config_section *sect,
                                     unsigned f, const char *name, size_t sz)
{
  struct config_var *var;
  struct treap_path path;

  if (f&CF_INHERIT)
    var = search_recursive(conf, sect, name, sz);
  else if (!(f&CF_CREAT))
    var = treap_lookup(&sect->vars, name, sz);
  else {
    var = treap_probe(&sect->vars, name, sz, &path);
    if (!var) {
      var = xmalloc(sizeof(*var));
      var->val = 0; var->file = 0; var->f = 0; var->line = 1;
      treap_insert(&sect->vars, &path, &var->_node, name, sz);
    }
  }
  return (var);
}

/* Set variable NAME to VALUE in SECT, with associated flags F.
 *
 * The names are null-terminated.  The flags are variable flags: see `struct
 * config_var' for details.  Returns the variable.
 *
 * If the variable is already set and has the `CF_OVERRIDE' flag, then this
 * function does nothing unless `CF_OVERRIDE' is /also/ set in F.
 */
struct config_var *config_set_var(struct config *conf,
                                  struct config_section *sect,
                                  unsigned f,
                                  const char *name, const char *value)
{
  return (config_set_var_n(conf, sect, f,
                           name, strlen(name),
                           value, strlen(value)));
}

/* As `config_set_var', except that the variable NAME and VALUE have explicit
 * lengths (NAMELEN and VALUELEN, respectively) rather than being null-
 * terminated.
 */
struct config_var *config_set_var_n(struct config *conf,
                                    struct config_section *sect,
                                    unsigned f,
                                    const char *name, size_t namelen,
                                    const char *value, size_t valuelen)
{
  struct config_var *var =
    config_find_var_n(conf, sect, CF_CREAT, name, namelen);

  if (var->f&~f&CF_OVERRIDE) return (var);
  free(var->val); var->val = xstrndup(value, valuelen); var->n = valuelen;
  var->f = f;
  return (var);
}

/* Initialize I to iterate over the variables directly defined in SECT. */
void config_start_var_iter(struct config *conf, struct config_section *sect,
                           struct config_var_iter *i)
  { treap_start_iter(&sect->vars, &i->i); }

/* Return next variable from I, in ascending lexicographical order.
 *
 * If there are no more variables, then return null.
 */
struct config_var *config_next_var(struct config_var_iter *i)
  { return (treap_next(&i->i)); }

/* Read and parse configuration FILE, applying its settings to CONF.
 *
 * If all goes well, the function returns 0.  If the file is not found, then
 * the behaviour depends on whether `CF_NOENTOK' is set in F: if so, then the
 * function simply returns -1.  Otherwise, a fatal error is reported.  Note
 * that this /only/ applies if the file does not exist (specifically, opening
 * it fails with `ENOENT') -- any other problems are reported as fatal
 * errors regardless of the flag setting.
 */
int config_read_file(struct config *conf, const char *file, unsigned f)
{
  struct config_section *sect;
  struct config_var *var;
  struct dstr d = DSTR_INIT, dd = DSTR_INIT;
  unsigned line = 0;
  const char *p, *q, *r;
  FILE *fp;

  /* Try to open the file. */
  fp = fopen(file, "r");
  if (!fp) {
    if ((f&CF_NOENTOK) && errno == ENOENT) return (-1);
    lose("failed to open configuration file `%s': %s",
         file, strerror(errno));
  }

  /* Find the initial section. */
  sect = config_find_section(conf, CF_CREAT, "@CONFIG"); var = 0;

  /* Work through the file, line by line. */
  for (;;) {
    dstr_reset(&d); if (dstr_readline(&d, fp)) break;
    line++;

    /* Trim trailing spaces from the line.  The syntax is sensitive to
     * leading spaces, so we can't trim those yet.
     */
    while (d.len && ISSPACE(d.p[d.len - 1])) d.len--;
    d.p[d.len] = 0;

    if (!*d.p || *d.p == ';')
      /* Ignore comments entirely.  (In particular, a comment doesn't
       * interrupt a multiline variable value.)
       */
      ;

    else if (ISSPACE(d.p[0])) {
      /* The line starts with whitespace, so it's a continuation line. */

      /* Skip the initial whitespace. */
      p = d.p; while (ISSPACE(*p)) p++;

      /* If we aren't collecting a variable value then this is an error.
       * Otherwise, accumulate it into the current value.
       */
      if (!var)
        lose("%s:%u: continuation line, but no variable", file, line);
      if (dd.len) dstr_putc(&dd, ' ');
      dstr_putm(&dd, p, d.len - (p - d.p));

    } else {
      /* The line starts in the first column. */

      /* If there's a value value being collected then we must commit it to
       * its variable (unless there's already a setting there that says we
       * shouldn't).
       */
      if (var) {
        if (!(var->f&CF_OVERRIDE))
          { var->val = xstrndup(dd.p, dd.len); var->n = dd.len; }
        var = 0;
      }

      /* Now decide what kind of line this is. */
      if (d.p[0] == '[') {
        /* It's a section header. */

        /* Parse the header. */
        p = d.p + 1; while (ISSPACE(*p)) p++;
        q = scan_name("section", p, d.p + d.len, file, line);
        r = q; while (ISSPACE(*r)) r++;
        if (*r != ']')
          lose("%s:%u: expected `]' in section header", file, line);
        if (r[1])
          lose("%s:%u: trailing junk after `]' in section header",
               file, line);

        /* Create the new section. */
        sect = config_find_section_n(conf, CF_CREAT, p, q - p);

      } else {
        /* It's a variable assignment.  Parse the name out. */
        p = scan_name("variable", d.p, d.p + d.len, file, line);
        var = config_find_var_n(conf, sect, CF_CREAT, d.p, p - d.p);
        while (ISSPACE(*p)) p++;
        if (*p != '=') lose("%s:%u: missing `=' in assignment", file, line);
        p++; while (ISSPACE(*p)) p++;

        /* Clear out the variable's initial value, unless we shouldn't
         * override it.
         */
        if (!(var->f&CF_OVERRIDE)) {
          free(var->val); var->val = 0; var->f = 0;
          free(var->file); var->file = xstrdup(file); var->line = line;
        }
        dstr_reset(&dd); dstr_puts(&dd, p);
      }
    }
  }

  /* If there's a value under construction then commit the result. */
  if (var && !(var->f&CF_OVERRIDE))
    { var->val = xstrndup(dd.p, dd.len); var->n = dd.len; }

  /* Close the file. */
  if (fclose(fp))
    lose("error reading configuration file `%s': %s", file, strerror(errno));

  /* All done. */
  dstr_release(&d); dstr_release(&dd);
  return (0);
}

/* Populate SECT with environment variables.
 *
 * Environment variables are always set with `CF_LITERAL'.
 */
void config_read_env(struct config *conf, struct config_section *sect)
{
  const char *p, *v;
  size_t i;

  for (i = 0; (p = environ[i]) != 0; i++) {
    v = strchr(p, '='); if (!v) continue;
    config_set_var_n(conf, sect, CF_LITERAL, p, v - p, v + 1, strlen(v + 1));
  }
}

/*----- Substitution and quoting ------------------------------------------*/

/* The substitution and word-splitting state.
 *
 * This only keeps track of the immutable parameters for the substitution
 * task: stuff which changes (flags, filtering state, cursor position) is
 * maintained separately.
 */
struct subst {
  struct config *config;                /* configuration state */
  struct config_section *home;          /* home section for lookups */
  struct dstr *d;                       /* current word being constructed */
  struct argv *av;                      /* output word list */
};

/* Flags for `subst' and related functions. */
#define SF_SPLIT 0x0001u                /* split at (unquoted) whitespace */
#define SF_QUOT 0x0002u                 /* currently within double quotes */
#define SF_SUBST 0x0004u                /* apply `$-substitutions */
#define SF_SUBEXPR 0x0008u              /* stop at delimiter `|' or `}' */
#define SF_SPANMASK 0x00ffu             /* mask for the above */

#define SF_WORD 0x0100u                 /* output word under construction */
#define SF_SKIP 0x0200u                 /* not producing output */
#define SF_LITERAL 0x0400u              /* do not expand or substitute */
#define SF_UPCASE 0x0800u               /* convert to uppercase */
#define SF_DOWNCASE 0x1000u             /* convert to lowercase */
#define SF_CASEMASK 0x1800u             /* mask for case conversions */

/* Apply filters encoded in QFILT and F to the text from P to L, and output.
 *
 * SB is the substitution state which, in particular, explains where the
 * output should go.
 *
 * The filters are encoded as flags `SF_UPCASE' and `SF_DOWNCASE' for case
 * conversions, and a nesting depth QFILT for toothpick escaping.  (QFILT is
 * encoded as the number of toothpicks to print: see `subst' for how this
 * determined.)
 */
static void filter_string(const char *p, const char *l,
                          const struct subst *sb, unsigned qfilt, unsigned f)
{
  size_t r, n;
  char *q; const char *pp, *ll;

  if (!qfilt && !(f&SF_CASEMASK))
    /* Fast path: there's nothing to do: just write to the output. */
    dstr_putm(sb->d, p, l - p);

  else for (;;) {
    /* We must be a bit more circumspect. */

    /* Determine the length of the next span of characters which don't need
     * escaping.  (If QFILT is zero then this is everything.)
     */
    r = l - p; n = qfilt ? strcspn(p, "\"\\") : r;
    if (n > r) n = r;

    if (!(f&SF_CASEMASK))
      /* No case conversion: we can just emit this chunk. */

      dstr_putm(sb->d, p, n);

    else {
      /* Case conversion to do.  Arrange enough space for the output, and
       * convert it character by character.
       */

      dstr_ensure(sb->d, n); q = sb->d->p + sb->d->len; pp = p; ll = p + n;
      if (f&SF_DOWNCASE) while (pp < ll) *q++ = TOLOWER(*pp++);
      else if (f&SF_UPCASE) while (pp < ll) *q++ = TOUPPER(*pp++);
      sb->d->len += n;
    }

    /* If we've reached the end then stop. */
    if (n >= r) break;

    /* Otherwise we must have found a character which requires escaping.
     * Emit enough toothpicks.
     */
    dstr_putcn(sb->d, '\\', qfilt);

    /* This character is now done, so we can skip over and see if there's
     * another chunk of stuff we can do at high speed.
     */
    dstr_putc(sb->d, p[n]); p += n + 1;
  }
}

/* Scan and resolve a `[SECT:]VAR' specifier at P.
 *
 * Return the address of the next character following the specifier; and set
 * *VAR_OUT to point to the variable we found, or null if it's not there.  L
 * is a limit on the region of the buffer that we should process; SB is the
 * substitution state which provides the home section if none is given
 * explicitly; FILE and LINE are the source location to blame for problems.
 */
static const char *retrieve_varspec(const char *p, const char *l,
                                    const struct subst *sb,
                                    struct config_var **var_out,
                                    const char *file, unsigned line)
{
  struct config_section *sect = sb->home;
  const char *t;

  t = scan_name("section or variable", p, l, file, line);
  if (t < l && *t == ':') {
    sect = config_find_section_n(sb->config, 0, p, t - p);
    p = t + 1; t = scan_name("variable", p, l, file, line);
  }

  if (!sect) *var_out = 0;
  else *var_out = config_find_var_n(sb->config, sect, CF_INHERIT, p, t - p);
  return (t);
}

/* Substitute and/or word-split text.
 *
 * The input text starts at P, and continues to (just before) L.  Context for
 * the task is provided by SB; the source location to blame is FILE and LINE
 * (FILE may be null so that this can be passed directly from a `config_var'
 * without further checking); QFILT is the nesting depth in toothpick-
 * escaping; and F holds a mask of `SF_...' flags.
 */
static const char *subst(const char *p, const char *l,
                         const struct subst *sb,
                         const char *file, unsigned line,
                         unsigned qfilt, unsigned f)
{
  struct config_var *var;
  const char *q0, *q1, *t;
  unsigned subqfilt, ff;
  size_t n;

  /* It would be best if we could process literal text at high speed.  To
   * this end, we have a table, indexed by the low-order bits of F, to tell
   * us which special characters we need to stop at.  This way, we can use
   * `strcspn' to skip over literal text and stop at the next character which
   * needs special handling.  Entries in this table with a null pointer
   * correspond to impossible flag settings: notably, `SF_QUOT' can only be
   * set when `SF_SUBST' is also set.
   */
  static const char *const delimtab[] = {

#define ESCAPE "\\"                     /* always watch for `\'-escapes */
#define SUBST "$"                       /* check for `$' if `SF_SUBST' set */
#define WORDSEP " \f\r\n\t\v'\""        /* space, quotes if `SF_SPLIT' but
                                         * not `SF_QUOT' */
#define QUOT "\""                       /* only quotes if `SF_SPLIT' and
                                         * `SF_QUOT' */
#define DELIM "|}"                      /* end delimiters of `SF_SUBEXPR' */

    ESCAPE,                             /* --- */
    ESCAPE             WORDSEP,         /* SPLIT */
    0,                                  /*         QUOT */
    ESCAPE             QUOT,            /* SPLIT | QUOT */
    ESCAPE       SUBST,                 /*                SUBST */
    ESCAPE       SUBST WORDSEP,         /* SPLIT |        SUBST */
    0,                                  /*         QUOT | SUBST */
    ESCAPE       SUBST QUOT,            /* SPLIT | QUOT | SUBST */
    ESCAPE DELIM,                       /*                        SUBEXPR */
    ESCAPE DELIM       WORDSEP,         /* SPLIT |                SUBEXPR */
    0,                                  /*         QUOT |         SUBEXPR */
    ESCAPE DELIM       QUOT,            /* SPLIT | QUOT |         SUBEXPR */
    ESCAPE DELIM SUBST,                 /*                SUBST | SUBEXPR */
    ESCAPE DELIM SUBST WORDSEP,         /* SPLIT |        SUBST | SUBEXPR */
    0,                                  /*         QUOT | SUBST | SUBEXPR */
    ESCAPE DELIM SUBST QUOT             /* SPLIT | QUOT | SUBST | SUBEXPR */

#undef ESCAPE
#undef SUBST
#undef WORDSEP
#undef QUOT
#undef DELIM
  };

  /* Set FILE to be useful if it was null on entry. */
  if (!file) file = "<internal>";

  /* If the text is literal then hand off to `filter_string'.  This obviously
   * starts a word.
   */
  if (f&SF_LITERAL) {
    filter_string(p, l, sb, qfilt, f);
    f |= SF_WORD;
    goto done;
  }

  /* Chew through the input until it's all gone. */
  while (p < l) {

    if ((f&(SF_SPLIT | SF_QUOT)) == SF_SPLIT && ISSPACE(*p)) {
      /* This is whitespace, we're supposed to split, and we're not within
       * quotes, so we should split here.
       */

      /* If there's a word in progress then we should commit it. */
      if (f&SF_WORD) {
        if (!(f&SF_SKIP)) {
          argv_append(sb->av, xstrndup(sb->d->p, sb->d->len));
          dstr_reset(sb->d);
        }
        f &= ~SF_WORD;
      }

      /* Skip over further whitespace at high speed. */
      do p++; while (p < l && ISSPACE(*p));

    } else if (*p == '\\') {
      /* This is a toothpick, so start a new word and add the next character
       * to it.
       */

      /* If there's no next character then we should be upset. */
      p++; if (p >= l) lose("%s:%u: unfinished `\\' escape", file, line);

      if (!(f&SF_SKIP)) {

        /* If this is a double quote or backslash then check QFILT to see if
         * it needs escaping.
         */
        if (qfilt && (*p == '"' || *p == '\\'))
          dstr_putcn(sb->d, '\\', qfilt);

        /* Output the character. */
        if (f&SF_DOWNCASE) dstr_putc(sb->d, TOLOWER(*p));
        else if (f&SF_UPCASE) dstr_putc(sb->d, TOUPPER(*p));
        else dstr_putc(sb->d, *p);
      }

      /* Move past the escaped character.  Remember we started a word. */
      p++; f |= SF_WORD;

    } else if ((f&SF_SPLIT) && *p == '"') {
      /* This is a double quote, and we're word splitting.  We're definitely
       * in a word now.  Toggle whether we're within quotes.
       */

      f ^= SF_QUOT; f |= SF_WORD; p++;

    } else if ((f&(SF_SPLIT | SF_QUOT)) == SF_SPLIT && *p == '\'') {
      /* This is a single quote, and we're word splitting but not within
       * double quotes.  Find the matching end quote, and just output
       * everything between literally.
       */

      p++; t = strchr(p, '\'');
      if (!t || t >= l) lose("%s:%u: missing `''", file, line);
      if (!(f&SF_SKIP)) filter_string(p, t, sb, qfilt, f);
      p = t + 1; f |= SF_WORD;

    } else if ((f&SF_SUBEXPR) && (*p == '|' || *p == '}')) {
      /* This is an end delimiter, and we're supposed to stop here. */
      break;

    } else if ((f&SF_SUBST) && *p == '$') {
      /* This is a `$' and we're supposed to do substitution. */

      /* The kind of substitution is determined by the next character. */
      p++; if (p >= l) lose("%s:%u: incomplete substitution", file, line);

      /* Prepare flags for a recursive substitution.
       *
       * Hide our quote state from the recursive call.  If we're within a
       * word, then disable word-splitting.
       */
      ff = f&~(SF_QUOT | (f&SF_WORD ? SF_SPLIT : 0));

      /* Now dispatch based on the following character. */
      switch (*p) {

        case '?':
          /* A conditional expression: $?VAR{CONSEQ[|ALT]} */

          /* Skip initial space. */
          p++; while (p < l && ISSPACE(*p)) p++;

          /* Find the variable. */
          p = retrieve_varspec(p, l, sb, &var, file, line);

          /* Skip whitespace again. */
          while (p < l && ISSPACE(*p)) p++;

          /* Expect the opening `{'. */
          if (p > l || *p != '{') lose("%s:%u: expected `{'", file, line);
          p++;

          /* We'll process the parts recursively, but we need to come back
           * when we hit the appropriate delimiters, so arrange for that.
           */
          ff |= SF_SUBEXPR;

          /* Process the consequent (skip if the variable wasn't found). */
          p = subst(p, l, sb, file, line, qfilt,
                    ff | (var ? 0 : SF_SKIP));

          /* If there's a `|' then process the alternative too (skip if the
           * variable /was/ found).
           */
          if (p < l && *p == '|')
            p = subst(p + 1, l, sb, file, line, qfilt,
                      ff | (var ? SF_SKIP : 0));

          /* We should now be past the closing `}'. */
          if (p >= l || *p != '}') lose("%s:%u: missing `}'", file, line);
          p++;
          break;

        case '{':
          /* A variable substitution: ${VAR[|FILT]...[?ALT]} */

          /* Skip initial whitespace. */
          p++; while (p < l && ISSPACE(*p)) p++;

          /* Find the variable. */
          q0 = p; p = retrieve_varspec(p, l, sb, &var, file, line); q1 = p;

          /* Determine the filters to apply when substituting the variable
           * value.
           */
          subqfilt = qfilt;
          for (;;) {

            /* Skip spaces again. */
            while (p < l && ISSPACE(*p)) p++;

            /* If there's no `|' then there are no more filters, so stop. */
            if (p >= l || *p != '|') break;

            /* Skip the `|' and more spaces. */
            p++; while (p < l && ISSPACE(*p)) p++;

            /* Collect the filter name. */
            t = scan_name("filter", p, l, file, line);

            /* Dispatch on the filter name. */
            if (t - p == 1 && *p == 'q')
              /* `q' -- quote for Lisp string.
               *
               * We're currently adding Q `\' characters before each naughty
               * character.  But a backslash itself is naughty too, so that
               * makes Q + 1 naughty characters, each of which needs a
               * toothpick, so now we need Q + (Q + 1) = 2 Q + 1 toothpicks.
               *
               * Calculate this here rather than at each point toothpicks
               * need to be deployed.
               */

              subqfilt = 2*subqfilt + 1;

            else if (t - p == 1 && *p == 'l')
              /* `l' -- convert to lowercase.
               *
               * If a case conversion is already set, then that will override
               * whatever we do here, so don't bother.
               */

              { if (!(ff&SF_CASEMASK)) ff |= SF_DOWNCASE; }

            else if (t - p == 1 && *p == 'u')
              /* `u' -- convert to uppercase.
               *
               * If a case conversion is already set, then that will override
               * whatever we do here, so don't bother.
               */
              { if (!(ff&SF_CASEMASK)) ff |= SF_UPCASE; }

            else
              /* Something else we didn't understand. */
              lose("%s:%u: unknown filter `%.*s'",
                   file, line, (int)(t - p), p);

            /* Continue from after the filter name. */
            p = t;
          }

          /* If we're not skipping, and we found a variable, then substitute
           * its value.  This is the point where we need to be careful about
           * recursive expansion.
           */
          if (!(f&SF_SKIP) && var) {
            if (var->f&CF_EXPAND)
              lose("%s:%u: recursive expansion of variable `%.*s'",
                   file, line, (int)(q1 - q0), q0);
            var->f |= CF_EXPAND;
            subst(var->val, var->val + var->n, sb,
                  var->file, var->line, subqfilt,
                  ff | (var->f&CF_LITERAL ? SF_LITERAL : 0));
            var->f &= ~CF_EXPAND;
          }

          /* If there's an alternative, then we need to process (or maybe
           * skip) it.  Otherwise, we should complain if there was no
           * variable, and we're not skipping.
           */
          if (p < l && *p == '?')
            p = subst(p + 1, l, sb, file, line, subqfilt,
                      ff | SF_SUBEXPR | (var ? SF_SKIP : 0));
          else if (!var && !(f&SF_SKIP))
            lose("%s:%u: unknown variable `%.*s'",
                 file, line, (int)(q1 - q0), q0);

          /* Expect a `}' here.  (No need to skip spaces: we already did that
           * after scanning for filters, and either there was no alternative,
           * or we advanced to a delimiter character anyway.)
           */
          if (p >= l || *p != '}') lose("%s:%u: missing `}'", file, line);
          p++;
          break;

        default:
          /* Something else.  That's a shame. */
          lose("%s:%u: unexpected `$'-substitution `%c'", file, line, *p);
      }

      /* Complain if we started out in word-splitting state, and therefore
       * have added a whole number of words to the output, but there's a
       * word-fragment stuck onto the end of this substitution.
       */
      if (p < l && !(~f&~(SF_WORD | SF_SPLIT)) && !ISSPACE(*p) &&
          !((f&SF_SUBEXPR) && (*p == '|' || *p == '}')))
        lose("%s:%u: surprising word boundary "
             "after splicing substitution",
             file, line);
    }

    else {
      /* Something else.  Try to skip over this at high speed.
       *
       * This makes use of the table we set up earlier.
       */

      n = strcspn(p, delimtab[f&SF_SPANMASK]);
      if (n > l - p) n = l - p;
      if (!(f&SF_SKIP)) filter_string(p, p + n, sb, qfilt, f);
      p += n; f |= SF_WORD;
    }
  }

done:
  /* Sort out the wreckage. */

  /* If we're still within quotes then something has gone wrong. */
  if (f&SF_QUOT) lose("%s:%u: missing `\"'", file, line);

  /* If we're within a word, and should be splitting, then commit the word to
   * the output list.
   */
  if ((f&(SF_WORD | SF_SPLIT | SF_SKIP)) == (SF_SPLIT | SF_WORD)) {
    argv_append(sb->av, xstrndup(sb->d->p, sb->d->len));
    dstr_reset(sb->d);
  }

  /* And, with that, we're done. */
  return (p);
}

/* Expand substitutions in a string.
 *
 * Expand the null-terminated string P relative to the HOME section, using
 * configuration CONFIG, and appending the result to dynamic string D.  Blame
 * WHAT in any error messages.
 */
void config_subst_string(struct config *config, struct config_section *home,
                         const char *what, const char *p, struct dstr *d)
{
  struct subst sb;

  sb.config = config; sb.home = home; sb.d = d;
  subst(p, p + strlen(p), &sb, what, 0, 0, SF_SUBST);
  dstr_putz(d);
}

/* Expand substitutions in a string.
 *
 * Expand the null-terminated string P relative to the HOME section, using
 * configuration CONFIG, returning the result as a freshly malloc(3)ed
 * string.  Blame WHAT in any error messages.
 */
char *config_subst_string_alloc(struct config *config,
                                struct config_section *home,
                                const char *what, const char *p)
{
  struct dstr d = DSTR_INIT;
  char *q;

  config_subst_string(config, home, what, p, &d);
  q = xstrndup(d.p, d.len); dstr_release(&d); return (q);
}

/* Expand substitutions in a variable.
 *
 * Expand the value of the variable VAR relative to the HOME section, using
 * configuration CONFIG, appending the result to dynamic string D.
 */
void config_subst_var(struct config *config, struct config_section *home,
                      struct config_var *var, struct dstr *d)
{
  struct subst sb;

  sb.config = config; sb.home = home; sb.d = d;
  var->f |= CF_EXPAND;
  subst(var->val, var->val + var->n, &sb, var->file, var->line, 0,
        SF_SUBST | (var->f&CF_LITERAL ? SF_LITERAL : 0));
  var->f &= ~CF_EXPAND;
  dstr_putz(d);
}

/* Expand substitutions in a variable.
 *
 * Expand the value of the variable VAR relative to the HOME section, using
 * configuration CONFIG, returning the result as a freshly malloc(3)ed
 * string.
 */
char *config_subst_var_alloc(struct config *config,
                             struct config_section *home,
                             struct config_var *var)
{
  struct dstr d = DSTR_INIT;
  char *q;

  config_subst_var(config, home, var, &d);
  q = xstrndup(d.p, d.len); dstr_release(&d); return (q);
}

/* Expand substitutions in a variable and split into words.
 *
 * Expand and word-split the value of the variable VAR relative to the HOME
 * section, using configuration CONFIG, appending the resulting words into
 * the vector AV.
 */
void config_subst_split_var(struct config *config,
                            struct config_section *home,
                            struct config_var *var, struct argv *av)
{
  struct dstr d = DSTR_INIT;
  struct subst sb;

  sb.config = config; sb.home = home; sb.av = av; sb.d = &d;
  var->f |= CF_EXPAND;
  subst(var->val, var->val + var->n, &sb, var->file, var->line, 0,
        SF_SUBST | SF_SPLIT | (var->f&CF_LITERAL ? SF_LITERAL : 0));
  var->f &= ~CF_EXPAND;
  dstr_release(&d);
}

/*----- That's all, folks -------------------------------------------------*/