/[pcre]/code/trunk/pcre_compile.c

Diff of /code/trunk/pcre_compile.c

Parent Directory | Revision Log | View Patch Patch

-revision 91 by nigel,
Sat Feb 24 21:41:34 2007 UTC
+revision 93 by nigel,
Sat Feb 24 21:41:42 2007 UTC
 Line 42 
 POSSIBILITY OF SUCH DAMAGE.
  supporting internal functions that are not used by other modules. */
- #define NLBLOCK cd            /* The block containing newline information */
+ #define NLBLOCK cd             /* Block containing newline information */
+ #define PSSTART start_pattern  /* Field containing processed string start */
+ #define PSEND   end_pattern    /* Field containing processed string end */
  #include "pcre_internal.h"
-Line 54 
 used by pcretest. DEBUG is not defined w
+Line 58 
 used by pcretest. DEBUG is not defined w
  #endif
  /*************************************************
  *      Code parameters and static tables         *
  *************************************************/
- /* Maximum number of items on the nested bracket stacks at compile time. This
+ /* This value specifies the size of stack workspace that is used during the
- applies to the nesting of all kinds of parentheses. It does not limit
+ first pre-compile phase that determines how much memory is required. The regex
- un-nested, non-capturing parentheses. This number can be made bigger if
+ is partly compiled into this space, but the compiled parts are discarded as
- necessary - it is used to dimension one int and one unsigned char vector at
+ soon as they can be, so that hopefully there will never be an overrun. The code
- compile time. */
+ does, however, check for an overrun. The largest amount I've seen used is 218,
+ so this number is very generous.
+ The same workspace is used during the second, actual compile phase for
+ remembering forward references to groups so that they can be filled in at the
+ end. Each entry in this list occupies LINK_SIZE bytes, so even when LINK_SIZE
+ is 4 there is plenty of room. */
- #define BRASTACK_SIZE 200
+ #define COMPILE_WORK_SIZE (4096)
  /* Table for handling escaped characters in the range '0'-'z'. Positive returns
-Line 79 
 static const short int escapes[] = {
+Line 88 
 static const short int escapes[] = {
 ,      0,    ':',    ';',    '<',    '=',    '>',    '?',   /* 8 - ? */
     '@', -ESC_A, -ESC_B, -ESC_C, -ESC_D, -ESC_E,      0, -ESC_G,   /* @ - G */
 ,      0,      0,      0,      0,      0,      0,      0,   /* H - O */
- -ESC_P, -ESC_Q,      0, -ESC_S,      0,      0,      0, -ESC_W,   /* P - W */
+ -ESC_P, -ESC_Q, -ESC_R, -ESC_S,      0,      0,      0, -ESC_W,   /* P - W */
  -ESC_X,      0, -ESC_Z,    '[',   '\\',    ']',    '^',    '_',   /* X - _ */
     '`',      7, -ESC_b,      0, -ESC_d,  ESC_e,  ESC_f,      0,   /* ` - g */
-,      0,      0,      0,      0,      0,  ESC_n,      0,   /* h - o */
+,      0,      0, -ESC_k,      0,      0,  ESC_n,      0,   /* h - o */
  -ESC_p,      0,  ESC_r, -ESC_s,  ESC_tee,    0,      0, -ESC_w,   /* p - w */
 ,      0, -ESC_z                                            /* x - z */
  };
-Line 98 
 static const short int escapes[] = {
+Line 107 
 static const short int escapes[] = {
  /*  78 */     0,   '`',    ':',     '#',    '@',  '\'',    '=',    '"',
  /*  80 */     0,     7, -ESC_b,       0, -ESC_d, ESC_e,  ESC_f,      0,
  /*  88 */     0,     0,      0,     '{',      0,     0,      0,      0,
- /*  90 */     0,     0,      0,     'l',      0, ESC_n,      0, -ESC_p,
+ /*  90 */     0,     0, -ESC_k,     'l',      0, ESC_n,      0, -ESC_p,
  /*  98 */     0, ESC_r,      0,     '}',      0,     0,      0,      0,
  /*  A0 */     0,   '~', -ESC_s, ESC_tee,      0,     0, -ESC_w,      0,
  /*  A8 */     0,-ESC_z,      0,       0,      0,   '[',      0,      0,
-Line 107 
 static const short int escapes[] = {
+Line 116 
 static const short int escapes[] = {
  /*  C0 */   '{',-ESC_A, -ESC_B,  -ESC_C, -ESC_D,-ESC_E,      0, -ESC_G,
  /*  C8 */     0,     0,      0,       0,      0,     0,      0,      0,
  /*  D0 */   '}',     0,      0,       0,      0,     0,      0, -ESC_P,
- /*  D8 */-ESC_Q,     0,      0,       0,      0,     0,      0,      0,
+ /*  D8 */-ESC_Q,-ESC_R,      0,       0,      0,     0,      0,      0,
  /*  E0 */  '\\',     0, -ESC_S,       0,      0,     0, -ESC_W, -ESC_X,
  /*  E8 */     0,-ESC_Z,      0,       0,      0,     0,      0,      0,
  /*  F0 */     0,     0,      0,       0,      0,     0,      0,      0,
-Line 156 
 static const int posix_class_maps[] = {
+Line 165 
 static const int posix_class_maps[] = {
  };
+ #define STRING(a)  # a
+ #define XSTRING(s) STRING(s)
  /* The texts of compile-time error messages. These are "char *" because they
- are passed to the outside world. */
+ are passed to the outside world. Do not ever re-use any error number, because
+ they are documented. Always add a new error instead. Messages marked DEAD below
+ are no longer used. */
  static const char *error_texts[] = {
    "no error",
-Line 172 
 static const char *error_texts[] = {
+Line 186 
 static const char *error_texts[] = {
    "range out of order in character class",
    "nothing to repeat",
    /* 10 */
-   "operand of unlimited repeat could match the empty string",
+   "operand of unlimited repeat could match the empty string",  /** DEAD **/
    "internal error: unexpected repeat",
    "unrecognized character after (?",
    "POSIX named classes are supported only within a class",
-Line 182 
 static const char *error_texts[] = {
+Line 196 
 static const char *error_texts[] = {
    "erroffset passed as NULL",
    "unknown option bit(s) set",
    "missing ) after comment",
-   "parentheses nested too deeply",
+   "parentheses nested too deeply",  /** DEAD **/
    /* 20 */
    "regular expression too large",
    "failed to get memory",
-Line 199 
 static const char *error_texts[] = {
+Line 213 
 static const char *error_texts[] = {
    "unknown POSIX class name",
    "POSIX collating elements are not supported",
    "this version of PCRE is not compiled with PCRE_UTF8 support",
-   "spare error",
+   "spare error",  /** DEAD **/
    "character value in \\x{...} sequence is too large",
    /* 35 */
    "invalid condition (?(0)",
-Line 210 
 static const char *error_texts[] = {
+Line 224 
 static const char *error_texts[] = {
    /* 40 */
    "recursive call could loop indefinitely",
    "unrecognized character after (?P",
-   "syntax error after (?P",
+   "syntax error in subpattern name (missing terminator)",
    "two named subpatterns have the same name",
    "invalid UTF-8 string",
    /* 45 */
    "support for \\P, \\p, and \\X has not been compiled",
    "malformed \\P or \\p sequence",
    "unknown property name after \\P or \\p",
-   "subpattern name is too long (maximum 32 characters)",
+   "subpattern name is too long (maximum " XSTRING(MAX_NAME_SIZE) " characters)",
-   "too many named subpatterns (maximum 10,000)",
+   "too many named subpatterns (maximum " XSTRING(MAX_NAME_COUNT) ")",
    /* 50 */
    "repeated subpattern is too long",
-   "octal value is greater than \\377 (not in UTF-8 mode)"
+   "octal value is greater than \\377 (not in UTF-8 mode)",
+   "internal error: overran compiling workspace",
+   "internal error: previously-checked referenced subpattern not found",
+   "DEFINE group contains more than one branch",
+   /* 55 */
+   "repeating a DEFINE group is not allowed",
+   "inconsistent NEWLINE options",
+   "\\g is not followed by an (optionally braced) non-zero number"
  };
-Line 352 
 static const unsigned char ebcdic_charta
+Line 373 
 static const unsigned char ebcdic_charta
  /* Definition to allow mutual recursion */
  static BOOL
-   compile_regex(int, int, int *, uschar **, const uschar **, int *, BOOL, int,
+   compile_regex(int, int, uschar **, const uschar **, int *, BOOL, int, int *,
-     int *, int *, branch_chain *, compile_data *);
+     int *, branch_chain *, compile_data *, int *);
-Line 363 
 static BOOL
+Line 384 
 static BOOL
  /* This function is called when a \ has been encountered. It either returns a
  positive value for a simple escape such as \n, or a negative value which
- encodes one of the more complicated things such as \d. When UTF-8 is enabled,
+ encodes one of the more complicated things such as \d. A backreference to group
- a positive value greater than 255 may be returned. On entry, ptr is pointing at
+ n is returned as -(ESC_REF + n); ESC_REF is the highest ESC_xxx macro. When
- the \. On exit, it is on the final character of the escape sequence.
+ UTF-8 is enabled, a positive value greater than 255 may be returned. On entry,
+ ptr is pointing at the \. On exit, it is on the final character of the escape
+ sequence.
  Arguments:
    ptrptr         points to the pattern position pointer
-Line 412 
 else if ((i = escapes[c - 0x48]) != 0)
+Line 435 
 else if ((i = escapes[c - 0x48]) != 0)
  else
    {
    const uschar *oldptr;
+   BOOL braced, negated;
    switch (c)
      {
      /* A number of Perl escapes are not handled by PCRE. We give an explicit
-Line 425 
 else
+Line 450 
 else
      *errorcodeptr = ERR37;
      break;
+     /* \g must be followed by a number, either plain or braced. If positive, it
+     is an absolute backreference. If negative, it is a relative backreference.
+     This is a Perl 5.10 feature. */
+     case 'g':
+     if (ptr[1] == '{')
+       {
+       braced = TRUE;
+       ptr++;
+       }
+     else braced = FALSE;
+     if (ptr[1] == '-')
+       {
+       negated = TRUE;
+       ptr++;
+       }
+     else negated = FALSE;
+     c = 0;
+     while ((digitab[ptr[1]] & ctype_digit) != 0)
+       c = c * 10 + *(++ptr) - '0';
+     if (c == 0 || (braced && *(++ptr) != '}'))
+       {
+       *errorcodeptr = ERR57;
+       return 0;
+       }
+     if (negated)
+       {
+       if (c > bracount)
+         {
+         *errorcodeptr = ERR15;
+         return 0;
+         }
+       c = bracount - (c - 1);
+       }
+     c = -(ESC_REF + c);
+     break;
      /* The handling of escape sequences consisting of a string of digits
      starting with one that is not zero is not straightforward. By experiment,
      the way Perl works seems to be as follows:
-Line 532 
 else
+Line 599 
 else
        }
      break;
-     /* Other special escapes not starting with a digit are straightforward */
+     /* For \c, a following letter is upper-cased; then the 0x40 bit is flipped.
+     This coding is ASCII-specific, but then the whole concept of \cx is
+     ASCII-specific. (However, an EBCDIC equivalent has now been added.) */
      case 'c':
      c = *(++ptr);
-Line 542 
 else
+Line 611 
 else
        return 0;
        }
-     /* A letter is upper-cased; then the 0x40 bit is flipped. This coding
-     is ASCII-specific, but then the whole concept of \cx is ASCII-specific.
-     (However, an EBCDIC equivalent has now been added.) */
  #if !EBCDIC    /* ASCII coding */
      if (c >= 'a' && c <= 'z') c -= 32;
      c ^= 0x40;
-Line 772 
 return p;
+Line 837 
 return p;
  /*************************************************
- *     Find forward referenced named subpattern   *
+ *       Find forward referenced subpattern       *
  *************************************************/
- /* This function scans along a pattern looking for capturing subpatterns, and
+ /* This function scans along a pattern's text looking for capturing
- counting them. If it finds a named pattern that matches the name it is given,
+ subpatterns, and counting them. If it finds a named pattern that matches the
- it returns its number. This is used for forward references to named
+ name it is given, it returns its number. Alternatively, if the name is NULL, it
- subpatterns. We know that if (?P< is encountered, the name will be terminated
+ returns when it reaches a given numbered subpattern. This is used for forward
- by '>' because that is checked in the first pass.
+ references to subpatterns. We know that if (?P< is encountered, the name will
+ be terminated by '>' because that is checked in the first pass.
  Arguments:
-   pointer      current position in the pattern
+   ptr          current position in the pattern
-   count        current count of capturing parens
+   count        current count of capturing parens so far encountered
-   name         name to seek
+   name         name to seek, or NULL if seeking a numbered subpattern
-   namelen      name length
+   lorn         name length, or subpattern number if name is NULL
+   xmode        TRUE if we are in /x mode
  Returns:       the number of the named subpattern, or -1 if not found
  */
  static int
- find_named_parens(const uschar *ptr, int count, const uschar *name, int namelen)
+ find_parens(const uschar *ptr, int count, const uschar *name, int lorn,
+   BOOL xmode)
  {
  const uschar *thisname;
  for (; *ptr != 0; ptr++)
    {
-   if (*ptr == '\\' && ptr[1] != 0) { ptr++; continue; }
+   int term;
+   /* Skip over backslashed characters and also entire \Q...\E */
+   if (*ptr == '\\')
+     {
+     if (*(++ptr) == 0) return -1;
+     if (*ptr == 'Q') for (;;)
+       {
+       while (*(++ptr) != 0 && *ptr != '\\');
+       if (*ptr == 0) return -1;
+       if (*(++ptr) == 'E') break;
+       }
+     continue;
+     }
+   /* Skip over character classes */
+   if (*ptr == '[')
+     {
+     while (*(++ptr) != ']')
+       {
+       if (*ptr == '\\')
+         {
+         if (*(++ptr) == 0) return -1;
+         if (*ptr == 'Q') for (;;)
+           {
+           while (*(++ptr) != 0 && *ptr != '\\');
+           if (*ptr == 0) return -1;
+           if (*(++ptr) == 'E') break;
+           }
+         continue;
+         }
+       }
+     continue;
+     }
+   /* Skip comments in /x mode */
+   if (xmode && *ptr == '#')
+     {
+     while (*(++ptr) != 0 && *ptr != '\n');
+     if (*ptr == 0) return -1;
+     continue;
+     }
+   /* An opening parens must now be a real metacharacter */
    if (*ptr != '(') continue;
-   if (ptr[1] != '?') { count++; continue; }
+   if (ptr[1] != '?')
-   if (ptr[2] == '(') { ptr += 2; continue; }
+     {
-   if (ptr[2] != 'P' || ptr[3] != '<') continue;
+     count++;
+     if (name == NULL && count == lorn) return count;
+     continue;
+     }
+   ptr += 2;
+   if (*ptr == 'P') ptr++;                      /* Allow optional P */
+   /* We have to disambiguate (?<! and (?<= from (?<name> */
+   if ((*ptr != '<' || ptr[1] == '!' || ptr[1] == '=') &&
+        *ptr != '\'')
+     continue;
    count++;
-   ptr += 4;
+   if (name == NULL && count == lorn) return count;
+   term = *ptr++;
+   if (term == '<') term = '>';
    thisname = ptr;
-   while (*ptr != '>') ptr++;
+   while (*ptr != term) ptr++;
-   if (namelen == ptr - thisname && strncmp(name, thisname, namelen) == 0)
+   if (name != NULL && lorn == ptr - thisname &&
+       strncmp((const char *)name, (const char *)thisname, lorn) == 0)
      return count;
    }
  return -1;
  }
-Line 862 
 for (;;)
+Line 996 
 for (;;)
      case OP_CALLOUT:
      case OP_CREF:
-     case OP_BRANUMBER:
+     case OP_RREF:
+     case OP_DEF:
      code += _pcre_OP_lengths[*code];
      break;
-Line 907 
 for (;;)
+Line 1042 
 for (;;)
    {
    int d;
    register int op = *cc;
-   if (op >= OP_BRA) op = OP_BRA;
    switch (op)
      {
+     case OP_CBRA:
      case OP_BRA:
      case OP_ONCE:
      case OP_COND:
-     d = find_fixedlength(cc, options);
+     d = find_fixedlength(cc + ((op == OP_CBRA)? 2:0), options);
      if (d < 0) return d;
      branchlength += d;
      do cc += GET(cc, 1); while (*cc == OP_ALT);
-Line 949 
 for (;;)
+Line 1084 
 for (;;)
      /* Skip over things that don't match chars */
      case OP_REVERSE:
-     case OP_BRANUMBER:
      case OP_CREF:
+     case OP_RREF:
+     case OP_DEF:
      case OP_OPT:
      case OP_CALLOUT:
      case OP_SOD:
-Line 1094 
 for (;;)
+Line 1230 
 for (;;)
    if (c == OP_XCLASS) code += GET(code, 1);
-   /* Handle bracketed group */
+   /* Handle capturing bracket */
-   else if (c > OP_BRA)
+   else if (c == OP_CBRA)
      {
-     int n = c - OP_BRA;
+     int n = GET2(code, 1+LINK_SIZE);
-     if (n > EXTRACT_BASIC_MAX) n = GET2(code, 2+LINK_SIZE);
      if (n == number) return (uschar *)code;
-     code += _pcre_OP_lengths[OP_BRA];
+     code += _pcre_OP_lengths[c];
      }
-   /* Otherwise, we get the item's length from the table. In UTF-8 mode, opcodes
+   /* In UTF-8 mode, opcodes that are followed by a character may be followed by
-   that are followed by a character may be followed by a multi-byte character.
+   a multi-byte character. The length in the table is a minimum, so we have to
-   The length in the table is a minimum, so we have to scan along to skip the
+   arrange to skip the extra bytes. */
-   extra bytes. All opcodes are less than 128, so we can use relatively
-   efficient code. */
    else
      {
-Line 1120 
 for (;;)
+Line 1253 
 for (;;)
        case OP_EXACT:
        case OP_UPTO:
        case OP_MINUPTO:
+       case OP_POSUPTO:
        case OP_STAR:
        case OP_MINSTAR:
+       case OP_POSSTAR:
        case OP_PLUS:
        case OP_MINPLUS:
+       case OP_POSPLUS:
        case OP_QUERY:
        case OP_MINQUERY:
-       while ((*code & 0xc0) == 0x80) code++;
+       case OP_POSQUERY:
+       if (code[-1] >= 0xc0) code += _pcre_utf8_table4[code[-1] & 0x3f];
        break;
        }
      }
-Line 1164 
 for (;;)
+Line 1301 
 for (;;)
    if (c == OP_XCLASS) code += GET(code, 1);
-   /* All bracketed groups have the same length. */
-   else if (c > OP_BRA)
-     {
-     code += _pcre_OP_lengths[OP_BRA];
-     }
    /* Otherwise, we get the item's length from the table. In UTF-8 mode, opcodes
    that are followed by a character may be followed by a multi-byte character.
-   The length in the table is a minimum, so we have to scan along to skip the
+   The length in the table is a minimum, so we have to arrange to skip the extra
-   extra bytes. All opcodes are less than 128, so we can use relatively
+   bytes. */
-   efficient code. */
    else
      {
-Line 1187 
 for (;;)
+Line 1316 
 for (;;)
        case OP_EXACT:
        case OP_UPTO:
        case OP_MINUPTO:
+       case OP_POSUPTO:
        case OP_STAR:
        case OP_MINSTAR:
+       case OP_POSSTAR:
        case OP_PLUS:
        case OP_MINPLUS:
+       case OP_POSPLUS:
        case OP_QUERY:
        case OP_MINQUERY:
-       while ((*code & 0xc0) == 0x80) code++;
+       case OP_POSQUERY:
+       if (code[-1] >= 0xc0) code += _pcre_utf8_table4[code[-1] & 0x3f];
        break;
        }
      }
-Line 1207 
 for (;;)
+Line 1340 
 for (;;)
  *************************************************/
  /* This function scans through a branch of a compiled pattern to see whether it
- can match the empty string or not. It is called only from could_be_empty()
+ can match the empty string or not. It is called from could_be_empty()
- below. Note that first_significant_code() skips over assertions. If we hit an
+ below and from compile_branch() when checking for an unlimited repeat of a
- unclosed bracket, we return "empty" - this means we've struck an inner bracket
+ group that can match nothing. Note that first_significant_code() skips over
- whose current branch will already have been scanned.
+ assertions. If we hit an unclosed bracket, we return "empty" - this means we've
+ struck an inner bracket whose current branch will already have been scanned.
  Arguments:
    code        points to start of search
-Line 1224 
 static BOOL
+Line 1358 
 static BOOL
  could_be_empty_branch(const uschar *code, const uschar *endcode, BOOL utf8)
  {
  register int c;
- for (code = first_significant_code(code + 1 + LINK_SIZE, NULL, 0, TRUE);
+ for (code = first_significant_code(code + _pcre_OP_lengths[*code], NULL, 0, TRUE);
       code < endcode;
       code = first_significant_code(code + _pcre_OP_lengths[c], NULL, 0, TRUE))
    {
-Line 1232 
 for (code = first_significant_code(code
+Line 1366 
 for (code = first_significant_code(code
    c = *code;
-   if (c >= OP_BRA)
+   if (c == OP_BRA || c == OP_CBRA || c == OP_ONCE)
      {
      BOOL empty_branch;
      if (GET(code, 1) == 0) return TRUE;    /* Hit unclosed bracket */
-Line 1248 
 for (code = first_significant_code(code
+Line 1382 
 for (code = first_significant_code(code
        }
      while (*code == OP_ALT);
      if (!empty_branch) return FALSE;   /* All branches are non-empty */
-     code += 1 + LINK_SIZE;
-     c = *code;
+     /* Move past the KET and fudge things so that the increment in the "for"
+     above has no effect. */
+     c = OP_END;
+     code += 1 + LINK_SIZE - _pcre_OP_lengths[c];
+     continue;
      }
-   else switch (c)
+   /* Handle the other opcodes */
+   switch (c)
      {
      /* Check for quantifiers after a class */
-Line 1308 
 for (code = first_significant_code(code
+Line 1449 
 for (code = first_significant_code(code
      case OP_NOT:
      case OP_PLUS:
      case OP_MINPLUS:
+     case OP_POSPLUS:
      case OP_EXACT:
      case OP_NOTPLUS:
      case OP_NOTMINPLUS:
+     case OP_NOTPOSPLUS:
      case OP_NOTEXACT:
      case OP_TYPEPLUS:
      case OP_TYPEMINPLUS:
+     case OP_TYPEPOSPLUS:
      case OP_TYPEEXACT:
      return FALSE;
-Line 1325 
 for (code = first_significant_code(code
+Line 1469 
 for (code = first_significant_code(code
      case OP_ALT:
      return TRUE;
-     /* In UTF-8 mode, STAR, MINSTAR, QUERY, MINQUERY, UPTO, and MINUPTO  may be
+     /* In UTF-8 mode, STAR, MINSTAR, POSSTAR, QUERY, MINQUERY, POSQUERY, UPTO,
-     followed by a multibyte character */
+     MINUPTO, and POSUPTO may be followed by a multibyte character */
  #ifdef SUPPORT_UTF8
      case OP_STAR:
      case OP_MINSTAR:
+     case OP_POSSTAR:
      case OP_QUERY:
      case OP_MINQUERY:
+     case OP_POSQUERY:
      case OP_UPTO:
      case OP_MINUPTO:
+     case OP_POSUPTO:
      if (utf8) while ((code[2] & 0xc0) == 0x80) code++;
      break;
  #endif
-Line 1452 
 earlier groups that are outside the curr
+Line 1599 
 earlier groups that are outside the curr
  optional (i.e. the minimum quantifier is zero), OP_BRAZERO is inserted before
  it, after it has been compiled. This means that any OP_RECURSE items within it
  that refer to the group itself or any contained groups have to have their
- offsets adjusted. That is the job of this function. Before it is called, the
+ offsets adjusted. That one of the jobs of this function. Before it is called,
- partially compiled regex must be temporarily terminated with OP_END.
+ the partially compiled regex must be temporarily terminated with OP_END.
+ This function has been extended with the possibility of forward references for
+ recursions and subroutine calls. It must also check the list of such references
+ for the group we are dealing with. If it finds that one of the recursions in
+ the current group is on this list, it adjusts the offset in the list, not the
+ value in the reference (which is a group number).
  Arguments:
    group      points to the start of the group
    adjust     the amount by which the group is to be moved
    utf8       TRUE in UTF-8 mode
    cd         contains pointers to tables etc.
+   save_hwm   the hwm forward reference pointer at the start of the group
  Returns:     nothing
  */
  static void
- adjust_recurse(uschar *group, int adjust, BOOL utf8, compile_data *cd)
+ adjust_recurse(uschar *group, int adjust, BOOL utf8, compile_data *cd,
+   uschar *save_hwm)
  {
  uschar *ptr = group;
  while ((ptr = (uschar *)find_recurse(ptr, utf8)) != NULL)
    {
-   int offset = GET(ptr, 1);
+   int offset;
-   if (cd->start_code + offset >= group) PUT(ptr, 1, offset + adjust);
+   uschar *hc;
+   /* See if this recursion is on the forward reference list. If so, adjust the
+   reference. */
+   for (hc = save_hwm; hc < cd->hwm; hc += LINK_SIZE)
+     {
+     offset = GET(hc, 0);
+     if (cd->start_code + offset == ptr + 1)
+       {
+       PUT(hc, 0, offset + adjust);
+       break;
+       }
+     }
+   /* Otherwise, adjust the recursion offset if it's after the start of this
+   group. */
+   if (hc >= cd->hwm)
+     {
+     offset = GET(ptr, 1);
+     if (cd->start_code + offset >= group) PUT(ptr, 1, offset + adjust);
+     }
    ptr += 1 + LINK_SIZE;
    }
  }
-Line 1550 
 Yield:        TRUE when range returned;
+Line 1728 
 Yield:        TRUE when range returned;
  */
  static BOOL
- get_othercase_range(int *cptr, int d, int *ocptr, int *odptr)
+ get_othercase_range(unsigned int *cptr, unsigned int d, unsigned int *ocptr,
+   unsigned int *odptr)
  {
- int c, othercase, next;
+ unsigned int c, othercase, next;
  for (c = *cptr; c <= d; c++)
-   { if ((othercase = _pcre_ucp_othercase(c)) >= 0) break; }
+   { if ((othercase = _pcre_ucp_othercase(c)) != NOTACHAR) break; }
  if (c > d) return FALSE;
-Line 1576 
 return TRUE;
+Line 1755 
 return TRUE;
  #endif  /* SUPPORT_UCP */
+ /*************************************************
+ *     Check if auto-possessifying is possible    *
+ *************************************************/
+ /* This function is called for unlimited repeats of certain items, to see
+ whether the next thing could possibly match the repeated item. If not, it makes
+ sense to automatically possessify the repeated item.
+ Arguments:
+   op_code       the repeated op code
+   this          data for this item, depends on the opcode
+   utf8          TRUE in UTF-8 mode
+   utf8_char     used for utf8 character bytes, NULL if not relevant
+   ptr           next character in pattern
+   options       options bits
+   cd            contains pointers to tables etc.
+ Returns:        TRUE if possessifying is wanted
+ */
+ static BOOL
+ check_auto_possessive(int op_code, int item, BOOL utf8, uschar *utf8_char,
+   const uschar *ptr, int options, compile_data *cd)
+ {
+ int next;
+ /* Skip whitespace and comments in extended mode */
+ if ((options & PCRE_EXTENDED) != 0)
+   {
+   for (;;)
+     {
+     while ((cd->ctypes[*ptr] & ctype_space) != 0) ptr++;
+     if (*ptr == '#')
+       {
+       while (*(++ptr) != 0)
+         if (IS_NEWLINE(ptr)) { ptr += cd->nllen; break; }
+       }
+     else break;
+     }
+   }
+ /* If the next item is one that we can handle, get its value. A non-negative
+ value is a character, a negative value is an escape value. */
+ if (*ptr == '\\')
+   {
+   int temperrorcode = 0;
+   next = check_escape(&ptr, &temperrorcode, cd->bracount, options, FALSE);
+   if (temperrorcode != 0) return FALSE;
+   ptr++;    /* Point after the escape sequence */
+   }
+ else if ((cd->ctypes[*ptr] & ctype_meta) == 0)
+   {
+ #ifdef SUPPORT_UTF8
+   if (utf8) { GETCHARINC(next, ptr); } else
+ #endif
+   next = *ptr++;
+   }
+ else return FALSE;
+ /* Skip whitespace and comments in extended mode */
+ if ((options & PCRE_EXTENDED) != 0)
+   {
+   for (;;)
+     {
+     while ((cd->ctypes[*ptr] & ctype_space) != 0) ptr++;
+     if (*ptr == '#')
+       {
+       while (*(++ptr) != 0)
+         if (IS_NEWLINE(ptr)) { ptr += cd->nllen; break; }
+       }
+     else break;
+     }
+   }
+ /* If the next thing is itself optional, we have to give up. */
+ if (*ptr == '*' || *ptr == '?' || strncmp((char *)ptr, "{0,", 3) == 0)
+   return FALSE;
+ /* Now compare the next item with the previous opcode. If the previous is a
+ positive single character match, "item" either contains the character or, if
+ "item" is greater than 127 in utf8 mode, the character's bytes are in
+ utf8_char. */
+ /* Handle cases when the next item is a character. */
+ if (next >= 0) switch(op_code)
+   {
+   case OP_CHAR:
+ #ifdef SUPPORT_UTF8
+   if (utf8 && item > 127) { GETCHAR(item, utf8_char); }
+ #endif
+   return item != next;
+   /* For CHARNC (caseless character) we must check the other case. If we have
+   Unicode property support, we can use it to test the other case of
+   high-valued characters. */
+   case OP_CHARNC:
+ #ifdef SUPPORT_UTF8
+   if (utf8 && item > 127) { GETCHAR(item, utf8_char); }
+ #endif
+   if (item == next) return FALSE;
+ #ifdef SUPPORT_UTF8
+   if (utf8)
+     {
+     unsigned int othercase;
+     if (next < 128) othercase = cd->fcc[next]; else
+ #ifdef SUPPORT_UCP
+     othercase = _pcre_ucp_othercase((unsigned int)next);
+ #else
+     othercase = NOTACHAR;
+ #endif
+     return (unsigned int)item != othercase;
+     }
+   else
+ #endif  /* SUPPORT_UTF8 */
+   return (item != cd->fcc[next]);  /* Non-UTF-8 mode */
+   /* For OP_NOT, "item" must be a single-byte character. */
+   case OP_NOT:
+   if (next < 0) return FALSE;  /* Not a character */
+   if (item == next) return TRUE;
+   if ((options & PCRE_CASELESS) == 0) return FALSE;
+ #ifdef SUPPORT_UTF8
+   if (utf8)
+     {
+     unsigned int othercase;
+     if (next < 128) othercase = cd->fcc[next]; else
+ #ifdef SUPPORT_UCP
+     othercase = _pcre_ucp_othercase(next);
+ #else
+     othercase = NOTACHAR;
+ #endif
+     return (unsigned int)item == othercase;
+     }
+   else
+ #endif  /* SUPPORT_UTF8 */
+   return (item == cd->fcc[next]);  /* Non-UTF-8 mode */
+   case OP_DIGIT:
+   return next > 127 || (cd->ctypes[next] & ctype_digit) == 0;
+   case OP_NOT_DIGIT:
+   return next <= 127 && (cd->ctypes[next] & ctype_digit) != 0;
+   case OP_WHITESPACE:
+   return next > 127 || (cd->ctypes[next] & ctype_space) == 0;
+   case OP_NOT_WHITESPACE:
+   return next <= 127 && (cd->ctypes[next] & ctype_space) != 0;
+   case OP_WORDCHAR:
+   return next > 127 || (cd->ctypes[next] & ctype_word) == 0;
+   case OP_NOT_WORDCHAR:
+   return next <= 127 && (cd->ctypes[next] & ctype_word) != 0;
+   default:
+   return FALSE;
+   }
+ /* Handle the case when the next item is \d, \s, etc. */
+ switch(op_code)
+   {
+   case OP_CHAR:
+   case OP_CHARNC:
+ #ifdef SUPPORT_UTF8
+   if (utf8 && item > 127) { GETCHAR(item, utf8_char); }
+ #endif
+   switch(-next)
+     {
+     case ESC_d:
+     return item > 127 || (cd->ctypes[item] & ctype_digit) == 0;
+     case ESC_D:
+     return item <= 127 && (cd->ctypes[item] & ctype_digit) != 0;
+     case ESC_s:
+     return item > 127 || (cd->ctypes[item] & ctype_space) == 0;
+     case ESC_S:
+     return item <= 127 && (cd->ctypes[item] & ctype_space) != 0;
+     case ESC_w:
+     return item > 127 || (cd->ctypes[item] & ctype_word) == 0;
+     case ESC_W:
+     return item <= 127 && (cd->ctypes[item] & ctype_word) != 0;
+     default:
+     return FALSE;
+     }
+   case OP_DIGIT:
+   return next == -ESC_D || next == -ESC_s || next == -ESC_W;
+   case OP_NOT_DIGIT:
+   return next == -ESC_d;
+   case OP_WHITESPACE:
+   return next == -ESC_S || next == -ESC_d || next == -ESC_w;
+   case OP_NOT_WHITESPACE:
+   return next == -ESC_s;
+   case OP_WORDCHAR:
+   return next == -ESC_W || next == -ESC_s;
+   case OP_NOT_WORDCHAR:
+   return next == -ESC_w || next == -ESC_d;
+   default:
+   return FALSE;
+   }
+ /* Control does not reach here */
+ }
  /*************************************************
  *           Compile one branch                   *
  *************************************************/
- /* Scan the pattern, compiling it into the code vector. If the options are
+ /* Scan the pattern, compiling it into the a vector. If the options are
  changed during the branch, the pointer is used to change the external options
- bits.
+ bits. This function is used during the pre-compile phase when we are trying
+ to find out the amount of memory needed, as well as during the real compile
+ phase. The value of lengthptr distinguishes the two phases.
  Arguments:
    optionsptr     pointer to the option bits
-   brackets       points to number of extracting brackets used
    codeptr        points to the pointer to the current code point
    ptrptr         points to the current pattern pointer
    errorcodeptr   points to error code variable
-Line 1594 
 Arguments:
+Line 2005 
 Arguments:
    reqbyteptr     set to the last literal character required, else < 0
    bcptr          points to current branch chain
    cd             contains pointers to tables etc.
+   lengthptr      NULL during the real compile phase
+                  points to length accumulator during pre-compile phase
  Returns:         TRUE on success
                   FALSE, with *errorcodeptr set non-zero on error
  */
  static BOOL
- compile_branch(int *optionsptr, int *brackets, uschar **codeptr,
+ compile_branch(int *optionsptr, uschar **codeptr, const uschar **ptrptr,
-   const uschar **ptrptr, int *errorcodeptr, int *firstbyteptr,
+   int *errorcodeptr, int *firstbyteptr, int *reqbyteptr, branch_chain *bcptr,
-   int *reqbyteptr, branch_chain *bcptr, compile_data *cd)
+   compile_data *cd, int *lengthptr)
  {
  int repeat_type, op_type;
  int repeat_min = 0, repeat_max = 0;      /* To please picky compilers */
-Line 1613 
 int zeroreqbyte, zerofirstbyte;
+Line 2026 
 int zeroreqbyte, zerofirstbyte;
  int req_caseopt, reqvary, tempreqvary;
  int options = *optionsptr;
  int after_manual_callout = 0;
+ int length_prevgroup = 0;
  register int c;
  register uschar *code = *codeptr;
+ uschar *last_code = code;
+ uschar *orig_code = code;
  uschar *tempcode;
  BOOL inescq = FALSE;
  BOOL groupsetfirstbyte = FALSE;
-Line 1622 
 const uschar *ptr = *ptrptr;
+Line 2038 
 const uschar *ptr = *ptrptr;
  const uschar *tempptr;
  uschar *previous = NULL;
  uschar *previous_callout = NULL;
+ uschar *save_hwm = NULL;
  uschar classbits[32];
  #ifdef SUPPORT_UTF8
-Line 1631 
 uschar *class_utf8data;
+Line 2048 
 uschar *class_utf8data;
  uschar utf8_char[6];
  #else
  BOOL utf8 = FALSE;
+ uschar *utf8_char = NULL;
+ #endif
+ #ifdef DEBUG
+ if (lengthptr != NULL) DPRINTF((">> start branch\n"));
  #endif
  /* Set up the default and non-default settings for greediness */
-Line 1664 
 for (;; ptr++)
+Line 2086 
 for (;; ptr++)
    BOOL negate_class;
    BOOL possessive_quantifier;
    BOOL is_quantifier;
+   BOOL is_recurse;
    int class_charcount;
    int class_lastchar;
    int newoptions;
-Line 1671 
 for (;; ptr++)
+Line 2094 
 for (;; ptr++)
    int skipbytes;
    int subreqbyte;
    int subfirstbyte;
+   int terminator;
    int mclength;
    uschar mcbuffer[8];
-   /* Next byte in the pattern */
+   /* Get next byte in the pattern */
    c = *ptr;
+   /* If we are in the pre-compile phase, accumulate the length used for the
+   previous cycle of this loop. */
+   if (lengthptr != NULL)
+     {
+ #ifdef DEBUG
+     if (code > cd->hwm) cd->hwm = code;                 /* High water info */
+ #endif
+     if (code > cd->start_workspace + COMPILE_WORK_SIZE) /* Check for overrun */
+       {
+       *errorcodeptr = ERR52;
+       goto FAILED;
+       }
+     /* There is at least one situation where code goes backwards: this is the
+     case of a zero quantifier after a class (e.g. [ab]{0}). At compile time,
+     the class is simply eliminated. However, it is created first, so we have to
+     allow memory for it. Therefore, don't ever reduce the length at this point.
+     */
+     if (code < last_code) code = last_code;
+     *lengthptr += code - last_code;
+     DPRINTF(("length=%d added %d c=%c\n", *lengthptr, code - last_code, c));
+     /* If "previous" is set and it is not at the start of the work space, move
+     it back to there, in order to avoid filling up the work space. Otherwise,
+     if "previous" is NULL, reset the current code pointer to the start. */
+     if (previous != NULL)
+       {
+       if (previous > orig_code)
+         {
+         memmove(orig_code, previous, code - previous);
+         code -= previous - orig_code;
+         previous = orig_code;
+         }
+       }
+     else code = orig_code;
+     /* Remember where this code item starts so we can pick up the length
+     next time round. */
+     last_code = code;
+     }
+   /* In the real compile phase, just check the workspace used by the forward
+   reference list. */
+   else if (cd->hwm > cd->start_workspace + COMPILE_WORK_SIZE)
+     {
+     *errorcodeptr = ERR52;
+     goto FAILED;
+     }
    /* If in \Q...\E, check for the end; if not, we have a literal */
    if (inescq && c != 0)
-Line 1692 
 for (;; ptr++)
+Line 2170 
 for (;; ptr++)
        {
        if (previous_callout != NULL)
          {
-         complete_callout(previous_callout, ptr, cd);
+         if (lengthptr == NULL)  /* Don't attempt in pre-compile phase */
+           complete_callout(previous_callout, ptr, cd);
          previous_callout = NULL;
          }
        if ((options & PCRE_AUTO_CALLOUT) != 0)
-Line 1713 
 for (;; ptr++)
+Line 2192 
 for (;; ptr++)
    if (!is_quantifier && previous_callout != NULL &&
         after_manual_callout-- <= 0)
      {
-     complete_callout(previous_callout, ptr, cd);
+     if (lengthptr == NULL)      /* Don't attempt in pre-compile phase */
+       complete_callout(previous_callout, ptr, cd);
      previous_callout = NULL;
      }
-Line 1724 
 for (;; ptr++)
+Line 2204 
 for (;; ptr++)
      if ((cd->ctypes[c] & ctype_space) != 0) continue;
      if (c == '#')
        {
-       while (*(++ptr) != 0) if (IS_NEWLINE(ptr)) break;
+       while (*(++ptr) != 0)
-       if (*ptr != 0)
          {
-         ptr += cd->nllen - 1;
+         if (IS_NEWLINE(ptr)) { ptr += cd->nllen - 1; break; }
-         continue;
          }
+       if (*ptr != 0) continue;
        /* Else fall through to handle end of string */
        c = 0;
        }
-Line 1745 
 for (;; ptr++)
+Line 2225 
 for (;; ptr++)
    switch(c)
      {
-     /* The branch terminates at end of string, |, or ). */
+     /* ===================================================================*/
+     case 0:                        /* The branch terminates at string end */
-     case 0:
+     case '|':                      /* or | or ) */
-     case '|':
      case ')':
      *firstbyteptr = firstbyte;
      *reqbyteptr = reqbyte;
      *codeptr = code;
      *ptrptr = ptr;
+     if (lengthptr != NULL)
+       {
+       *lengthptr += code - last_code;   /* To include callout length */
+       DPRINTF((">> end branch\n"));
+       }
      return TRUE;
+     /* ===================================================================*/
      /* Handle single-character metacharacters. In multiline mode, ^ disables
      the setting of any following char as a first character. */
-Line 1784 
 for (;; ptr++)
+Line 2270 
 for (;; ptr++)
      *code++ = OP_ANY;
      break;
+     /* ===================================================================*/
      /* Character classes. If the included characters are all < 256, we build a
 -byte bitmap of the permitted characters, except in the special case
      where there is only one such character. For negated classes, we build the
-Line 1822 
 for (;; ptr++)
+Line 2310 
 for (;; ptr++)
        }
      /* Keep a count of chars with values < 256 so that we can optimize the case
-     of just a single character (as long as it's < 256). For higher valued UTF-8
+     of just a single character (as long as it's < 256). However, For higher
-     characters, we don't yet do any optimization. */
+     valued UTF-8 characters, we don't yet do any optimization. */
      class_charcount = 0;
      class_lastchar = -1;
+     /* Initialize the 32-char bit map to all zeros. We build the map in a
+     temporary bit of memory, in case the class contains only 1 character (less
+     than 256), because in that case the compiled code doesn't use the bit map.
+     */
+     memset(classbits, 0, 32 * sizeof(uschar));
  #ifdef SUPPORT_UTF8
      class_utf8 = FALSE;                       /* No chars >= 256 */
-     class_utf8data = code + LINK_SIZE + 34;   /* For UTF-8 items */
+     class_utf8data = code + LINK_SIZE + 2;    /* For UTF-8 items */
  #endif
-     /* Initialize the 32-char bit map to all zeros. We have to build the
-     map in a temporary bit of store, in case the class contains only 1
-     character (< 256), because in that case the compiled code doesn't use the
-     bit map. */
-     memset(classbits, 0, 32 * sizeof(uschar));
      /* Process characters until ] is reached. By writing this as a "do" it
-     means that an initial ] is taken as a data character. The first pass
+     means that an initial ] is taken as a data character. At the start of the
-     through the regex checked the overall syntax, so we don't need to be very
+     loop, c contains the first byte of the character. */
-     strict here. At the start of the loop, c contains the first byte of the
-     character. */
-     do
+     if (c != 0) do
        {
+       const uschar *oldptr;
  #ifdef SUPPORT_UTF8
        if (utf8 && c > 127)
          {                           /* Braces are required because the */
-Line 1859 
 for (;; ptr++)
+Line 2347 
 for (;; ptr++)
        if (inescq)
          {
-         if (c == '\\' && ptr[1] == 'E')
+         if (c == '\\' && ptr[1] == 'E')     /* If we are at \E */
            {
-           inescq = FALSE;
+           inescq = FALSE;                   /* Reset literal state */
-           ptr++;
+           ptr++;                            /* Skip the 'E' */
-           continue;
+           continue;                         /* Carry on with next */
            }
-         else goto LONE_SINGLE_CHARACTER;
+         goto CHECK_RANGE;                   /* Could be range if \E follows */
          }
        /* Handle POSIX class names. Perl allows a negation extension of the
-Line 1956 
 for (;; ptr++)
+Line 2444 
 for (;; ptr++)
          }
        /* Backslash may introduce a single character, or it may introduce one
-       of the specials, which just set a flag. Escaped items are checked for
+       of the specials, which just set a flag. The sequence \b is a special
-       validity in the pre-compiling pass. The sequence \b is a special case.
+       case. Inside a class (and only there) it is treated as backspace.
-       Inside a class (and only there) it is treated as backspace. Elsewhere
+       Elsewhere it marks a word boundary. Other escapes have preset maps ready
-       it marks a word boundary. Other escapes have preset maps ready to
+       to or into the one we are building. We assume they have more than one
-       or into the one we are building. We assume they have more than one
        character in them, so set class_charcount bigger than one. */
        if (c == '\\')
          {
-         c = check_escape(&ptr, errorcodeptr, *brackets, options, TRUE);
+         c = check_escape(&ptr, errorcodeptr, cd->bracount, options, TRUE);
+         if (*errorcodeptr != 0) goto FAILED;
          if (-c == ESC_b) c = '\b';       /* \b is backslash in a class */
          else if (-c == ESC_X) c = 'X';   /* \X is literal X in a class */
+         else if (-c == ESC_R) c = 'R';   /* \R is literal R in a class */
          else if (-c == ESC_Q)            /* Handle start of quoted string */
            {
            if (ptr[1] == '\\' && ptr[2] == 'E')
-Line 1983 
 for (;; ptr++)
+Line 2472 
 for (;; ptr++)
            {
            register const uschar *cbits = cd->cbits;
            class_charcount += 2;     /* Greater than 1 is what matters */
-           switch (-c)
+           /* Save time by not doing this in the pre-compile phase. */
+           if (lengthptr == NULL) switch (-c)
              {
              case ESC_d:
              for (c = 0; c < 32; c++) classbits[c] |= cbits[c+cbit_digit];
-Line 2011 
 for (;; ptr++)
+Line 2503 
 for (;; ptr++)
              classbits[1] |= 0x08;    /* Perl 5.004 onwards omits VT from \s */
              continue;
- #ifdef SUPPORT_UCP
+             case ESC_E: /* Perl ignores an orphan \E */
-             case ESC_p:
-             case ESC_P:
-               {
-               BOOL negated;
-               int pdata;
-               int ptype = get_ucp(&ptr, &negated, &pdata, errorcodeptr);
-               if (ptype < 0) goto FAILED;
-               class_utf8 = TRUE;
-               *class_utf8data++ = ((-c == ESC_p) != negated)?
-                 XCL_PROP : XCL_NOTPROP;
-               *class_utf8data++ = ptype;
-               *class_utf8data++ = pdata;
-               class_charcount -= 2;   /* Not a < 256 character */
-               }
              continue;
- #endif
-             /* Unrecognized escapes are faulted if PCRE is running in its
-             strict mode. By default, for compatibility with Perl, they are
-             treated as literals. */
-             default:
+             default:    /* Not recognized; fall through */
-             if ((options & PCRE_EXTRA) != 0)
+             break;      /* Need "default" setting to stop compiler warning. */
-               {
-               *errorcodeptr = ERR7;
-               goto FAILED;
-               }
-             c = *ptr;              /* The final character */
-             class_charcount -= 2;  /* Undo the default count from above */
              }
-           }
-         /* Fall through if we have a single character (c >= 0). This may be
+           /* In the pre-compile phase, just do the recognition. */
-         > 256 in UTF-8 mode. */
-         }   /* End of backslash handling */
+           else if (c == -ESC_d || c == -ESC_D || c == -ESC_w ||
+                    c == -ESC_W || c == -ESC_s || c == -ESC_S) continue;
+           /* We need to deal with \P and \p in both phases. */
+ #ifdef SUPPORT_UCP
+           if (-c == ESC_p || -c == ESC_P)
+             {
+             BOOL negated;
+             int pdata;
+             int ptype = get_ucp(&ptr, &negated, &pdata, errorcodeptr);
+             if (ptype < 0) goto FAILED;
+             class_utf8 = TRUE;
+             *class_utf8data++ = ((-c == ESC_p) != negated)?
+               XCL_PROP : XCL_NOTPROP;
+             *class_utf8data++ = ptype;
+             *class_utf8data++ = pdata;
+             class_charcount -= 2;   /* Not a < 256 character */
+             continue;
+             }
+ #endif
+           /* Unrecognized escapes are faulted if PCRE is running in its
+           strict mode. By default, for compatibility with Perl, they are
+           treated as literals. */
+           if ((options & PCRE_EXTRA) != 0)
+             {
+             *errorcodeptr = ERR7;
+             goto FAILED;
+             }
+           class_charcount -= 2;  /* Undo the default count from above */
+           c = *ptr;              /* Get the final character and fall through */
+           }
+         /* Fall through if we have a single character (c >= 0). This may be
+         greater than 256 in UTF-8 mode. */
+         }   /* End of backslash handling */
        /* A single character may be followed by '-' to form a range. However,
        Perl does not permit ']' to be the end of the range. A '-' character
-       here is treated as a literal. */
+       at the end is treated as a literal. Perl ignores orphaned \E sequences
+       entirely. The code for handling \Q and \E is messy. */
+       CHECK_RANGE:
+       while (ptr[1] == '\\' && ptr[2] == 'E')
+         {
+         inescq = FALSE;
+         ptr += 2;
+         }
+       oldptr = ptr;
-       if (ptr[1] == '-' && ptr[2] != ']')
+       if (!inescq && ptr[1] == '-')
          {
          int d;
          ptr += 2;
+         while (*ptr == '\\' && ptr[1] == 'E') ptr += 2;
+         /* If we hit \Q (not followed by \E) at this point, go into escaped
+         mode. */
+         while (*ptr == '\\' && ptr[1] == 'Q')
+           {
+           ptr += 2;
+           if (*ptr == '\\' && ptr[1] == 'E') { ptr += 2; continue; }
+           inescq = TRUE;
+           break;
+           }
+         if (*ptr == 0 || (!inescq && *ptr == ']'))
+           {
+           ptr = oldptr;
+           goto LONE_SINGLE_CHARACTER;
+           }
  #ifdef SUPPORT_UTF8
          if (utf8)
-Line 2071 
 for (;; ptr++)
+Line 2602 
 for (;; ptr++)
          not any of the other escapes. Perl 5.6 treats a hyphen as a literal
          in such circumstances. */
-         if (d == '\\')
+         if (!inescq && d == '\\')
            {
-           const uschar *oldptr = ptr;
+           d = check_escape(&ptr, errorcodeptr, cd->bracount, options, TRUE);
-           d = check_escape(&ptr, errorcodeptr, *brackets, options, TRUE);
+           if (*errorcodeptr != 0) goto FAILED;
-           /* \b is backslash; \X is literal X; any other special means the '-'
+           /* \b is backslash; \X is literal X; \R is literal R; any other
-           was literal */
+           special means the '-' was literal */
            if (d < 0)
              {
              if (d == -ESC_b) d = '\b';
-             else if (d == -ESC_X) d = 'X'; else
+             else if (d == -ESC_X) d = 'X';
+             else if (d == -ESC_R) d = 'R'; else
                {
-               ptr = oldptr - 2;
+               ptr = oldptr;
                goto LONE_SINGLE_CHARACTER;  /* A few lines below */
                }
              }
            }
-         /* The check that the two values are in the correct order happens in
+         /* Check that the two values are in the correct order. Optimize
-         the pre-pass. Optimize one-character ranges */
+         one-character ranges */
+         if (d < c)
+           {
+           *errorcodeptr = ERR8;
+           goto FAILED;
+           }
          if (d == c) goto LONE_SINGLE_CHARACTER;  /* A few lines below */
-Line 2112 
 for (;; ptr++)
+Line 2650 
 for (;; ptr++)
  #ifdef SUPPORT_UCP
            if ((options & PCRE_CASELESS) != 0)
              {
-             int occ, ocd;
+             unsigned int occ, ocd;
-             int cc = c;
+             unsigned int cc = c;
-             int origd = d;
+             unsigned int origd = d;
              while (get_othercase_range(&cc, origd, &occ, &ocd))
                {
                if (occ >= c && ocd <= d) continue;  /* Skip embedded ranges */
-Line 2172 
 for (;; ptr++)
+Line 2710 
 for (;; ptr++)
          ranges that lie entirely within 0-127 when there is UCP support; else
          for partial ranges without UCP support. */
-         for (; c <= d; c++)
+         class_charcount += d - c + 1;
+         class_lastchar = d;
+         /* We can save a bit of time by skipping this in the pre-compile. */
+         if (lengthptr == NULL) for (; c <= d; c++)
            {
            classbits[c/8] |= (1 << (c&7));
            if ((options & PCRE_CASELESS) != 0)
-Line 2180 
 for (;; ptr++)
+Line 2723 
 for (;; ptr++)
              int uc = cd->fcc[c];           /* flip case */
              classbits[uc/8] |= (1 << (uc&7));
              }
-           class_charcount++;                /* in case a one-char range */
-           class_lastchar = c;
            }
          continue;   /* Go get the next char in the class */
-Line 2205 
 for (;; ptr++)
+Line 2746 
 for (;; ptr++)
  #ifdef SUPPORT_UCP
          if ((options & PCRE_CASELESS) != 0)
            {
-           int othercase;
+           unsigned int othercase;
-           if ((othercase = _pcre_ucp_othercase(c)) >= 0)
+           if ((othercase = _pcre_ucp_othercase(c)) != NOTACHAR)
              {
              *class_utf8data++ = XCL_SINGLE;
              class_utf8data += _pcre_ord2utf8(othercase, class_utf8data);
-Line 2231 
 for (;; ptr++)
+Line 2772 
 for (;; ptr++)
          }
        }
-     /* Loop until ']' reached; the check for end of string happens inside the
+     /* Loop until ']' reached. This "while" is the end of the "do" above. */
-     loop. This "while" is the end of the "do" above. */
-     while ((c = *(++ptr)) != ']' || inescq);
+     while ((c = *(++ptr)) != 0 && (c != ']' || inescq));
+     if (c == 0)                          /* Missing terminating ']' */
+       {
+       *errorcodeptr = ERR6;
+       goto FAILED;
+       }
      /* If class_charcount is 1, we saw precisely one character whose value is
      less than 256. In non-UTF-8 mode we can always optimize. In UTF-8 mode, we
-Line 2298 
 for (;; ptr++)
+Line 2844 
 for (;; ptr++)
      /* If there are characters with values > 255, we have to compile an
      extended class, with its own opcode. If there are no characters < 256,
-     we can omit the bitmap. */
+     we can omit the bitmap in the actual compiled code. */
  #ifdef SUPPORT_UTF8
      if (class_utf8)
-Line 2308 
 for (;; ptr++)
+Line 2854 
 for (;; ptr++)
        code += LINK_SIZE;
        *code = negate_class? XCL_NOT : 0;
-       /* If the map is required, install it, and move on to the end of
+       /* If the map is required, move up the extra data to make room for it;
-       the extra data */
+       otherwise just move the code pointer to the end of the extra data. */
        if (class_charcount > 0)
          {
          *code++ |= XCL_MAP;
+         memmove(code + 32, code, class_utf8data - code);
          memcpy(code, classbits, 32);
-         code = class_utf8data;
+         code = class_utf8data + 32;
-         }
-       /* If the map is not required, slide down the extra data. */
-       else
-         {
-         int len = class_utf8data - (code + 33);
-         memmove(code + 1, code + 33, len);
-         code += len + 1;
          }
+       else code = class_utf8data;
        /* Now fill in the complete length of the item */
-Line 2342 
 for (;; ptr++)
+Line 2881 
 for (;; ptr++)
      if (negate_class)
        {
        *code++ = OP_NCLASS;
-       for (c = 0; c < 32; c++) code[c] = ~classbits[c];
+       if (lengthptr == NULL)    /* Save time in the pre-compile phase */
+         for (c = 0; c < 32; c++) code[c] = ~classbits[c];
        }
      else
        {
-Line 2352 
 for (;; ptr++)
+Line 2892 
 for (;; ptr++)
      code += 32;
      break;
+     /* ===================================================================*/
      /* Various kinds of repeat; '{' is not necessarily a quantifier, but this
      has been tested above. */
-Line 2419 
 for (;; ptr++)
+Line 2961 
 for (;; ptr++)
        }
      else repeat_type = greedy_default;
-     /* If previous was a recursion, we need to wrap it inside brackets so that
-     it can be replicated if necessary. */
-     if (*previous == OP_RECURSE)
-       {
-       memmove(previous + 1 + LINK_SIZE, previous, 1 + LINK_SIZE);
-       code += 1 + LINK_SIZE;
-       *previous = OP_BRA;
-       PUT(previous, 1, code - previous);
-       *code = OP_KET;
-       PUT(code, 1, code - previous);
-       code += 1 + LINK_SIZE;
-       }
      /* If previous was a character match, abolish the item and generate a
      repeat item instead. If a char item has a minumum of more than one, ensure
      that it is set in reqbyte - it might not be if a sequence such as x{3} is
-Line 2466 
 for (;; ptr++)
+Line 2994 
 for (;; ptr++)
          if (repeat_min > 1) reqbyte = c | req_caseopt | cd->req_varyopt;
          }
+       /* If the repetition is unlimited, it pays to see if the next thing on
+       the line is something that cannot possibly match this character. If so,
+       automatically possessifying this item gains some performance in the case
+       where the match fails. */
+       if (!possessive_quantifier &&
+           repeat_max < 0 &&
+           check_auto_possessive(*previous, c, utf8, utf8_char, ptr + 1,
+             options, cd))
+         {
+         repeat_type = 0;    /* Force greedy */
+         possessive_quantifier = TRUE;
+         }
        goto OUTPUT_SINGLE_REPEAT;   /* Code shared with single character types */
        }
      /* If previous was a single negated character ([^a] or similar), we use
      one of the special opcodes, replacing it. The code is shared with single-
      character repeats by setting opt_type to add a suitable offset into
-     repeat_type. OP_NOT is currently used only for single-byte chars. */
+     repeat_type. We can also test for auto-possessification. OP_NOT is
+     currently used only for single-byte chars. */
      else if (*previous == OP_NOT)
        {
        op_type = OP_NOTSTAR - OP_STAR;  /* Use "not" opcodes */
        c = previous[1];
+       if (!possessive_quantifier &&
+           repeat_max < 0 &&
+           check_auto_possessive(OP_NOT, c, utf8, NULL, ptr + 1, options, cd))
+         {
+         repeat_type = 0;    /* Force greedy */
+         possessive_quantifier = TRUE;
+         }
        goto OUTPUT_SINGLE_REPEAT;
        }
-Line 2495 
 for (;; ptr++)
+Line 3045 
 for (;; ptr++)
        op_type = OP_TYPESTAR - OP_STAR;  /* Use type opcodes */
        c = *previous;
+       if (!possessive_quantifier &&
+           repeat_max < 0 &&
+           check_auto_possessive(c, 0, utf8, NULL, ptr + 1, options, cd))
+         {
+         repeat_type = 0;    /* Force greedy */
+         possessive_quantifier = TRUE;
+         }
        OUTPUT_SINGLE_REPEAT:
        if (*previous == OP_PROP || *previous == OP_NOTPROP)
          {
-Line 2535 
 for (;; ptr++)
+Line 3093 
 for (;; ptr++)
          }
        /* A repeat minimum of 1 is optimized into some special cases. If the
-       maximum is unlimited, we use OP_PLUS. Otherwise, the original item it
+       maximum is unlimited, we use OP_PLUS. Otherwise, the original item is
        left in place and, if the maximum is greater than 1, we use OP_UPTO with
        one less than the maximum. */
-Line 2588 
 for (;; ptr++)
+Line 3146 
 for (;; ptr++)
            }
          /* Else insert an UPTO if the max is greater than the min, again
-         preceded by the character, for the previously inserted code. */
+         preceded by the character, for the previously inserted code. If the
+         UPTO is just for 1 instance, we can use QUERY instead. */
          else if (repeat_max != repeat_min)
            {
-Line 2607 
 for (;; ptr++)
+Line 3166 
 for (;; ptr++)
              *code++ = prop_value;
              }
            repeat_max -= repeat_min;
-           *code++ = OP_UPTO + repeat_type;
-           PUT2INC(code, 0, repeat_max);
+           if (repeat_max == 1)
+             {
+             *code++ = OP_QUERY + repeat_type;
+             }
+           else
+             {
+             *code++ = OP_UPTO + repeat_type;
+             PUT2INC(code, 0, repeat_max);
+             }
            }
          }
-Line 2675 
 for (;; ptr++)
+Line 3242 
 for (;; ptr++)
      /* If previous was a bracket group, we may have to replicate it in certain
      cases. */
-     else if (*previous >= OP_BRA || *previous == OP_ONCE ||
+     else if (*previous == OP_BRA  || *previous == OP_CBRA ||
-              *previous == OP_COND)
+              *previous == OP_ONCE || *previous == OP_COND)
        {
        register int i;
        int ketoffset = 0;
        int len = code - previous;
        uschar *bralink = NULL;
+       /* Repeating a DEFINE group is pointless */
+       if (*previous == OP_COND && previous[LINK_SIZE+1] == OP_DEF)
+         {
+         *errorcodeptr = ERR55;
+         goto FAILED;
+         }
+       /* This is a paranoid check to stop integer overflow later on */
+       if (len > MAX_DUPLENGTH)
+         {
+         *errorcodeptr = ERR50;
+         goto FAILED;
+         }
        /* If the maximum repeat count is unlimited, find the end of the bracket
        by scanning through from the start, and compute the offset back to it
        from the current code pointer. There may be an OP_OPT setting following
-Line 2717 
 for (;; ptr++)
+Line 3300 
 for (;; ptr++)
          /* If the maximum is 1 or unlimited, we just have to stick in the
          BRAZERO and do no more at this point. However, we do need to adjust
          any OP_RECURSE calls inside the group that refer to the group itself or
-         any internal group, because the offset is from the start of the whole
+         any internal or forward referenced group, because the offset is from
-         regex. Temporarily terminate the pattern while doing this. */
+         the start of the whole regex. Temporarily terminate the pattern while
+         doing this. */
          if (repeat_max <= 1)
            {
            *code = OP_END;
-           adjust_recurse(previous, 1, utf8, cd);
+           adjust_recurse(previous, 1, utf8, cd, save_hwm);
            memmove(previous+1, previous, len);
            code++;
            *previous++ = OP_BRAZERO + repeat_type;
-Line 2741 
 for (;; ptr++)
+Line 3325 
 for (;; ptr++)
            {
            int offset;
            *code = OP_END;
-           adjust_recurse(previous, 2 + LINK_SIZE, utf8, cd);
+           adjust_recurse(previous, 2 + LINK_SIZE, utf8, cd, save_hwm);
            memmove(previous + 2 + LINK_SIZE, previous, len);
            code += 2 + LINK_SIZE;
            *previous++ = OP_BRAZERO + repeat_type;
-Line 2761 
 for (;; ptr++)
+Line 3345 
 for (;; ptr++)
        /* If the minimum is greater than zero, replicate the group as many
        times as necessary, and adjust the maximum to the number of subsequent
        copies that we need. If we set a first char from the group, and didn't
-       set a required char, copy the latter from the former. */
+       set a required char, copy the latter from the former. If there are any
+       forward reference subroutine calls in the group, there will be entries on
+       the workspace list; replicate these with an appropriate increment. */
        else
          {
          if (repeat_min > 1)
            {
-           if (groupsetfirstbyte && reqbyte < 0) reqbyte = firstbyte;
+           /* In the pre-compile phase, we don't actually do the replication. We
-           for (i = 1; i < repeat_min; i++)
+           just adjust the length as if we had. */
+           if (lengthptr != NULL)
+             *lengthptr += (repeat_min - 1)*length_prevgroup;
+           /* This is compiling for real */
+           else
              {
-             memcpy(code, previous, len);
+             if (groupsetfirstbyte && reqbyte < 0) reqbyte = firstbyte;
-             code += len;
+             for (i = 1; i < repeat_min; i++)
+               {
+               uschar *hc;
+               uschar *this_hwm = cd->hwm;
+               memcpy(code, previous, len);
+               for (hc = save_hwm; hc < this_hwm; hc += LINK_SIZE)
+                 {
+                 PUT(cd->hwm, 0, GET(hc, 0) + len);
+                 cd->hwm += LINK_SIZE;
+                 }
+               save_hwm = this_hwm;
+               code += len;
+               }
              }
            }
          if (repeat_max > 0) repeat_max -= repeat_min;
          }
-Line 2781 
 for (;; ptr++)
+Line 3387 
 for (;; ptr++)
        the maximum is limited, it replicates the group in a nested fashion,
        remembering the bracket starts on a stack. In the case of a zero minimum,
        the first one was set up above. In all cases the repeat_max now specifies
-       the number of additional copies needed. */
+       the number of additional copies needed. Again, we must remember to
+       replicate entries on the forward reference list. */
        if (repeat_max >= 0)
          {
-         for (i = repeat_max - 1; i >= 0; i--)
+         /* In the pre-compile phase, we don't actually do the replication. We
+         just adjust the length as if we had. For each repetition we must add 1
+         to the length for BRAZERO and for all but the last repetition we must
+         add 2 + 2*LINKSIZE to allow for the nesting that occurs. */
+         if (lengthptr != NULL && repeat_max > 0)
+           *lengthptr += repeat_max * (length_prevgroup + 1 + 2 + 2*LINK_SIZE) -
+- 2*LINK_SIZE;  /* Last one doesn't nest */
+         /* This is compiling for real */
+         else for (i = repeat_max - 1; i >= 0; i--)
            {
+           uschar *hc;
+           uschar *this_hwm = cd->hwm;
            *code++ = OP_BRAZERO + repeat_type;
            /* All but the final copy start a new nesting, maintaining the
-Line 2802 
 for (;; ptr++)
+Line 3423 
 for (;; ptr++)
              }
            memcpy(code, previous, len);
+           for (hc = save_hwm; hc < this_hwm; hc += LINK_SIZE)
+             {
+             PUT(cd->hwm, 0, GET(hc, 0) + len + ((i != 0)? 2+LINK_SIZE : 1));
+             cd->hwm += LINK_SIZE;
+             }
+           save_hwm = this_hwm;
            code += len;
            }
-Line 2824 
 for (;; ptr++)
+Line 3451 
 for (;; ptr++)
        /* If the maximum is unlimited, set a repeater in the final copy. We
        can't just offset backwards from the current code point, because we
        don't know if there's been an options resetting after the ket. The
-       correct offset was computed above. */
+       correct offset was computed above.
+       Then, when we are doing the actual compile phase, check to see whether
+       this group is a non-atomic one that could match an empty string. If so,
+       convert the initial operator to the S form (e.g. OP_BRA -> OP_SBRA) so
+       that runtime checking can be done. [This check is also applied to
+       atomic groups at runtime, but in a different way.] */
-       else code[-ketoffset] = OP_KETRMAX + repeat_type;
+       else
+         {
+         uschar *ketcode = code - ketoffset;
+         uschar *bracode = ketcode - GET(ketcode, 1);
+         *ketcode = OP_KETRMAX + repeat_type;
+         if (lengthptr == NULL && *bracode != OP_ONCE)
+           {
+           uschar *scode = bracode;
+           do
+             {
+             if (could_be_empty_branch(scode, ketcode, utf8))
+               {
+               *bracode += OP_SBRA - OP_BRA;
+               break;
+               }
+             scode += GET(scode, 1);
+             }
+           while (*scode == OP_ALT);
+           }
+         }
        }
      /* Else there's some kind of shambles */
-Line 2837 
 for (;; ptr++)
+Line 3489 
 for (;; ptr++)
        goto FAILED;
        }
-     /* If the character following a repeat is '+', we wrap the entire repeated
+     /* If the character following a repeat is '+', or if certain optimization
-     item inside OP_ONCE brackets. This is just syntactic sugar, taken from
+     tests above succeeded, possessive_quantifier is TRUE. For some of the
-     Sun's Java package. The repeated item starts at tempcode, not at previous,
+     simpler opcodes, there is an special alternative opcode for this. For
-     which might be the first part of a string whose (former) last char we
+     anything else, we wrap the entire repeated item inside OP_ONCE brackets.
-     repeated. However, we don't support '+' after a greediness '?'. */
+     The '+' notation is just syntactic sugar, taken from Sun's Java package,
+     but the special opcodes can optimize it a bit. The repeated item starts at
+     tempcode, not at previous, which might be the first part of a string whose
+     (former) last char we repeated.
+     Possessifying an 'exact' quantifier has no effect, so we can ignore it. But
+     an 'upto' may follow. We skip over an 'exact' item, and then test the
+     length of what remains before proceeding. */
      if (possessive_quantifier)
        {
-       int len = code - tempcode;
+       int len;
-       memmove(tempcode + 1+LINK_SIZE, tempcode, len);
+       if (*tempcode == OP_EXACT || *tempcode == OP_TYPEEXACT ||
-       code += 1 + LINK_SIZE;
+           *tempcode == OP_NOTEXACT)
-       len += 1 + LINK_SIZE;
+         tempcode += _pcre_OP_lengths[*tempcode];
-       tempcode[0] = OP_ONCE;
+       len = code - tempcode;
-       *code++ = OP_KET;
+       if (len > 0) switch (*tempcode)
-       PUTINC(code, 0, len);
+         {
-       PUT(tempcode, 1, len);
+         case OP_STAR:  *tempcode = OP_POSSTAR; break;
+         case OP_PLUS:  *tempcode = OP_POSPLUS; break;
+         case OP_QUERY: *tempcode = OP_POSQUERY; break;
+         case OP_UPTO:  *tempcode = OP_POSUPTO; break;
+         case OP_TYPESTAR:  *tempcode = OP_TYPEPOSSTAR; break;
+         case OP_TYPEPLUS:  *tempcode = OP_TYPEPOSPLUS; break;
+         case OP_TYPEQUERY: *tempcode = OP_TYPEPOSQUERY; break;
+         case OP_TYPEUPTO:  *tempcode = OP_TYPEPOSUPTO; break;
+         case OP_NOTSTAR:  *tempcode = OP_NOTPOSSTAR; break;
+         case OP_NOTPLUS:  *tempcode = OP_NOTPOSPLUS; break;
+         case OP_NOTQUERY: *tempcode = OP_NOTPOSQUERY; break;
+         case OP_NOTUPTO:  *tempcode = OP_NOTPOSUPTO; break;
+         default:
+         memmove(tempcode + 1+LINK_SIZE, tempcode, len);
+         code += 1 + LINK_SIZE;
+         len += 1 + LINK_SIZE;
+         tempcode[0] = OP_ONCE;
+         *code++ = OP_KET;
+         PUTINC(code, 0, len);
+         PUT(tempcode, 1, len);
+         break;
+         }
        }
      /* In all case we no longer have a previous item. We also set the
-Line 2865 
 for (;; ptr++)
+Line 3548 
 for (;; ptr++)
      break;
-     /* Start of nested bracket sub-expression, or comment or lookahead or
+     /* ===================================================================*/
-     lookbehind or option setting or condition. First deal with special things
+     /* Start of nested parenthesized sub-expression, or comment or lookahead or
-     that can come after a bracket; all are introduced by ?, and the appearance
+     lookbehind or option setting or condition or all the other extended
-     of any of them means that this is not a referencing group. They were
+     parenthesis forms. First deal with the specials; all are introduced by ?,
-     checked for validity in the first pass over the string, so we don't have to
+     and the appearance of any of them means that this is not a capturing
-     check for syntax errors here.  */
+     group. */
      case '(':
      newoptions = options;
      skipbytes = 0;
+     bravalue = OP_CBRA;
+     save_hwm = cd->hwm;
      if (*(++ptr) == '?')
        {
-       int set, unset;
+       int i, set, unset, namelen;
        int *optset;
+       const uschar *name;
+       uschar *slot;
        switch (*(++ptr))
          {
          case '#':                 /* Comment; skip to ket */
          ptr++;
-         while (*ptr != ')') ptr++;
+         while (*ptr != 0 && *ptr != ')') ptr++;
+         if (*ptr == 0)
+           {
+           *errorcodeptr = ERR18;
+           goto FAILED;
+           }
          continue;
-         case ':':                 /* Non-extracting bracket */
+         /* ------------------------------------------------------------ */
+         case ':':                 /* Non-capturing bracket */
          bravalue = OP_BRA;
          ptr++;
          break;
+         /* ------------------------------------------------------------ */
          case '(':
          bravalue = OP_COND;       /* Conditional group */
-         /* A condition can be a number, referring to a numbered group, a name,
+         /* A condition can be an assertion, a number (referring to a numbered
-         referring to a named group, 'R', referring to recursion, or an
+         group), a name (referring to a named group), or 'R', referring to
-         assertion. There are two unfortunate ambiguities, caused by history.
+         recursion. R<digits> and R&name are also permitted for recursion tests.
-         (a) 'R' can be the recursive thing or the name 'R', and (b) a number
-         could be a name that consists of digits. In both cases, we look for a
+         There are several syntaxes for testing a named group: (?(name)) is used
-         name first; if not found, we try the other cases. If the first
+         by Python; Perl 5.10 onwards uses (?(<name>) or (?('name')).
-         character after (?( is a word character, we know the rest up to ) will
-         also be word characters because the syntax was checked in the first
+         There are two unfortunate ambiguities, caused by history. (a) 'R' can
-         pass. */
+         be the recursive thing or the name 'R' (and similarly for 'R' followed
+         by digits), and (b) a number could be a name that consists of digits.
-         if ((cd->ctypes[ptr[1]] & ctype_word) != 0)
+         In both cases, we look for a name first; if not found, we try the other
-           {
+         cases. */
-           int i, namelen;
-           int condref = 0;
+         /* For conditions that are assertions, check the syntax, and then exit
-           const uschar *name;
+         the switch. This will take control down to where bracketed groups,
-           uschar *slot = cd->name_table;
+         including assertions, are processed. */
-           /* This is needed for all successful cases. */
+         if (ptr[1] == '?' && (ptr[2] == '=' || ptr[2] == '!' || ptr[2] == '<'))
+           break;
-           skipbytes = 3;
+         /* Most other conditions use OP_CREF (a couple change to OP_RREF
+         below), and all need to skip 3 bytes at the start of the group. */
-           /* Read the name, but also get it as a number if it's all digits */
+         code[1+LINK_SIZE] = OP_CREF;
+         skipbytes = 3;
-           name = ++ptr;
+         /* Check for a test for recursion in a named group. */
-           while (*ptr != ')')
-             {
+         if (ptr[1] == 'R' && ptr[2] == '&')
-             if (condref >= 0)
+           {
-               condref = ((digitab[*ptr] & ctype_digit) != 0)?
+           terminator = -1;
-                 condref * 10 + *ptr - '0' : -1;
+           ptr += 2;
-             ptr++;
+           code[1+LINK_SIZE] = OP_RREF;    /* Change the type of test */
-             }
+           }
-           namelen = ptr - name;
+         /* Check for a test for a named group's having been set, using the Perl
+         syntax (?(<name>) or (?('name') */
+         else if (ptr[1] == '<')
+           {
+           terminator = '>';
            ptr++;
+           }
+         else if (ptr[1] == '\'')
+           {
+           terminator = '\'';
+           ptr++;
+           }
+         else terminator = 0;
-           for (i = 0; i < cd->names_found; i++)
+         /* We now expect to read a name; any thing else is an error */
-             {
-             if (strncmp((char *)name, (char *)slot+2, namelen) == 0) break;
-             slot += cd->name_entry_size;
-             }
-           /* Found a previous named subpattern */
+         if ((cd->ctypes[ptr[1]] & ctype_word) == 0)
+           {
+           ptr += 1;  /* To get the right offset */
+           *errorcodeptr = ERR28;
+           goto FAILED;
+           }
-           if (i < cd->names_found)
+         /* Read the name, but also get it as a number if it's all digits */
-             {
-             condref = GET2(slot, 0);
-             code[1+LINK_SIZE] = OP_CREF;
-             PUT2(code, 2+LINK_SIZE, condref);
-             }
-           /* Search the pattern for a forward reference */
+         recno = 0;
+         name = ++ptr;
+         while ((cd->ctypes[*ptr] & ctype_word) != 0)
+           {
+           if (recno >= 0)
+             recno = ((digitab[*ptr] & ctype_digit) != 0)?
+               recno * 10 + *ptr - '0' : -1;
+           ptr++;
+           }
+         namelen = ptr - name;
-           else if ((i = find_named_parens(ptr, *brackets, name, namelen)) > 0)
+         if ((terminator > 0 && *ptr++ != terminator) || *ptr++ != ')')
-             {
+           {
-             code[1+LINK_SIZE] = OP_CREF;
+           ptr--;      /* Error offset */
-             PUT2(code, 2+LINK_SIZE, i);
+           *errorcodeptr = ERR26;
-             }
+           goto FAILED;
+           }
-           /* Check for 'R' for recursion */
+         /* Do no further checking in the pre-compile phase. */
-           else if (namelen == 1 && *name == 'R')
+         if (lengthptr != NULL) break;
-             {
-             code[1+LINK_SIZE] = OP_CREF;
-             PUT2(code, 2+LINK_SIZE, CREF_RECURSE);
-             }
-           /* Check for a subpattern number */
+         /* In the real compile we do the work of looking for the actual
+         reference. */
-           else if (condref > 0)
+         slot = cd->name_table;
-             {
+         for (i = 0; i < cd->names_found; i++)
-             code[1+LINK_SIZE] = OP_CREF;
+           {
-             PUT2(code, 2+LINK_SIZE, condref);
+           if (strncmp((char *)name, (char *)slot+2, namelen) == 0) break;
-             }
+           slot += cd->name_entry_size;
+           }
-           /* Either an unidentified subpattern, or a reference to (?(0) */
+         /* Found a previous named subpattern */
-           else
+         if (i < cd->names_found)
+           {
+           recno = GET2(slot, 0);
+           PUT2(code, 2+LINK_SIZE, recno);
+           }
+         /* Search the pattern for a forward reference */
+         else if ((i = find_parens(ptr, cd->bracount, name, namelen,
+                         (options & PCRE_EXTENDED) != 0)) > 0)
+           {
+           PUT2(code, 2+LINK_SIZE, i);
+           }
+         /* If terminator == 0 it means that the name followed directly after
+         the opening parenthesis [e.g. (?(abc)...] and in this case there are
+         some further alternatives to try. For the cases where terminator != 0
+         [things like (?(<name>... or (?('name')... or (?(R&name)... ] we have
+         now checked all the possibilities, so give an error. */
+         else if (terminator != 0)
+           {
+           *errorcodeptr = ERR15;
+           goto FAILED;
+           }
+         /* Check for (?(R) for recursion. Allow digits after R to specify a
+         specific group number. */
+         else if (*name == 'R')
+           {
+           recno = 0;
+           for (i = 1; i < namelen; i++)
              {
-             *errorcodeptr = (condref == 0)? ERR35: ERR15;
+             if ((digitab[name[i]] & ctype_digit) == 0)
-             goto FAILED;
+               {
+               *errorcodeptr = ERR15;
+               goto FAILED;
+               }
+             recno = recno * 10 + name[i] - '0';
              }
+           if (recno == 0) recno = RREF_ANY;
+           code[1+LINK_SIZE] = OP_RREF;      /* Change test type */
+           PUT2(code, 2+LINK_SIZE, recno);
+           }
+         /* Similarly, check for the (?(DEFINE) "condition", which is always
+         false. */
+         else if (namelen == 6 && strncmp((char *)name, "DEFINE", 6) == 0)
+           {
+           code[1+LINK_SIZE] = OP_DEF;
+           skipbytes = 1;
+           }
+         /* Check for the "name" actually being a subpattern number. */
+         else if (recno > 0)
+           {
+           PUT2(code, 2+LINK_SIZE, recno);
            }
-         /* For conditions that are assertions, we just fall through, having
+         /* Either an unidentified subpattern, or a reference to (?(0) */
-         set bravalue above. */
+         else
+           {
+           *errorcodeptr = (recno == 0)? ERR35: ERR15;
+           goto FAILED;
+           }
          break;
+         /* ------------------------------------------------------------ */
          case '=':                 /* Positive lookahead */
          bravalue = OP_ASSERT;
          ptr++;
          break;
+         /* ------------------------------------------------------------ */
          case '!':                 /* Negative lookahead */
          bravalue = OP_ASSERT_NOT;
          ptr++;
          break;
-         case '<':                 /* Lookbehinds */
-         switch (*(++ptr))
+         /* ------------------------------------------------------------ */
+         case '<':                 /* Lookbehind or named define */
+         switch (ptr[1])
            {
            case '=':               /* Positive lookbehind */
            bravalue = OP_ASSERTBACK;
-           ptr++;
+           ptr += 2;
            break;
            case '!':               /* Negative lookbehind */
            bravalue = OP_ASSERTBACK_NOT;
-           ptr++;
+           ptr += 2;
            break;
+           default:                /* Could be name define, else bad */
+           if ((cd->ctypes[ptr[1]] & ctype_word) != 0) goto DEFINE_NAME;
+           ptr++;                  /* Correct offset for error */
+           *errorcodeptr = ERR24;
+           goto FAILED;
            }
          break;
+         /* ------------------------------------------------------------ */
          case '>':                 /* One-time brackets */
          bravalue = OP_ONCE;
          ptr++;
          break;
+         /* ------------------------------------------------------------ */
          case 'C':                 /* Callout - may be followed by digits; */
          previous_callout = code;  /* Save for later completion */
          after_manual_callout = 1; /* Skip one item before completing */
-         *code++ = OP_CALLOUT;     /* Already checked that the terminating */
+         *code++ = OP_CALLOUT;
-           {                       /* closing parenthesis is present. */
+           {
            int n = 0;
            while ((digitab[*(++ptr)] & ctype_digit) != 0)
              n = n * 10 + *ptr - '0';
+           if (*ptr != ')')
+             {
+             *errorcodeptr = ERR39;
+             goto FAILED;
+             }
            if (n > 255)
              {
              *errorcodeptr = ERR38;
-Line 3034 
 for (;; ptr++)
+Line 3830 
 for (;; ptr++)
          previous = NULL;
          continue;
-         case 'P':                 /* Named subpattern handling */
-         if (*(++ptr) == '<')      /* Definition */
+         /* ------------------------------------------------------------ */
+         case 'P':                 /* Python-style named subpattern handling */
+         if (*(++ptr) == '=' || *ptr == '>')  /* Reference or recursion */
+           {
+           is_recurse = *ptr == '>';
+           terminator = ')';
+           goto NAMED_REF_OR_RECURSE;
+           }
+         else if (*ptr != '<')    /* Test for Python-style definition */
+           {
+           *errorcodeptr = ERR41;
+           goto FAILED;
+           }
+         /* Fall through to handle (?P< as (?< is handled */
+         /* ------------------------------------------------------------ */
+         DEFINE_NAME:    /* Come here from (?< handling */
+         case '\'':
            {
-           int i, namelen;
+           terminator = (*ptr == '<')? '>' : '\'';
-           uschar *slot = cd->name_table;
+           name = ++ptr;
-           const uschar *name;     /* Don't amalgamate; some compilers */
-           name = ++ptr;           /* grumble at autoincrement in declaration */
+           while ((cd->ctypes[*ptr] & ctype_word) != 0) ptr++;
+           namelen = ptr - name;
-           while (*ptr++ != '>');
+           /* In the pre-compile phase, just do a syntax check. */
-           namelen = ptr - name - 1;
-           for (i = 0; i < cd->names_found; i++)
+           if (lengthptr != NULL)
+             {
+             if (*ptr != terminator)
+               {
+               *errorcodeptr = ERR42;
+               goto FAILED;
+               }
+             if (cd->names_found >= MAX_NAME_COUNT)
+               {
+               *errorcodeptr = ERR49;
+               goto FAILED;
+               }
+             if (namelen + 3 > cd->name_entry_size)
+               {
+               cd->name_entry_size = namelen + 3;
+               if (namelen > MAX_NAME_SIZE)
+                 {
+                 *errorcodeptr = ERR48;
+                 goto FAILED;
+                 }
+               }
+             }
+           /* In the real compile, create the entry in the table */
+           else
              {
-             int crc = memcmp(name, slot+2, namelen);
+             slot = cd->name_table;
-             if (crc == 0)
+             for (i = 0; i < cd->names_found; i++)
                {
-               if (slot[2+namelen] == 0)
+               int crc = memcmp(name, slot+2, namelen);
+               if (crc == 0)
                  {
-                 if ((options & PCRE_DUPNAMES) == 0)
+                 if (slot[2+namelen] == 0)
                    {
-                   *errorcodeptr = ERR43;
+                   if ((options & PCRE_DUPNAMES) == 0)
-                   goto FAILED;
+                     {
+                     *errorcodeptr = ERR43;
+                     goto FAILED;
+                     }
                    }
+                 else crc = -1;      /* Current name is substring */
                  }
-               else crc = -1;      /* Current name is substring */
+               if (crc < 0)
-               }
+                 {
-             if (crc < 0)
+                 memmove(slot + cd->name_entry_size, slot,
-               {
+                   (cd->names_found - i) * cd->name_entry_size);
-               memmove(slot + cd->name_entry_size, slot,
+                 break;
-                 (cd->names_found - i) * cd->name_entry_size);
+                 }
-               break;
+               slot += cd->name_entry_size;
                }
-             slot += cd->name_entry_size;
-             }
-           PUT2(slot, 0, *brackets + 1);
+             PUT2(slot, 0, cd->bracount + 1);
-           memcpy(slot + 2, name, namelen);
+             memcpy(slot + 2, name, namelen);
-           slot[2+namelen] = 0;
+             slot[2+namelen] = 0;
-           cd->names_found++;
+             }
-           goto NUMBERED_GROUP;
            }
-         if (*ptr == '=' || *ptr == '>')  /* Reference or recursion */
+         /* In both cases, count the number of names we've encountered. */
-           {
-           int i, namelen;
-           int type = *ptr++;
-           const uschar *name = ptr;
-           uschar *slot = cd->name_table;
-           while (*ptr != ')') ptr++;
+         ptr++;                    /* Move past > or ' */
-           namelen = ptr - name;
+         cd->names_found++;
+         goto NUMBERED_GROUP;
-           for (i = 0; i < cd->names_found; i++)
+         /* ------------------------------------------------------------ */
+         case '&':                 /* Perl recursion/subroutine syntax */
+         terminator = ')';
+         is_recurse = TRUE;
+         /* Fall through */
+         /* We come here from the Python syntax above that handles both
+         references (?P=name) and recursion (?P>name), as well as falling
+         through from the Perl recursion syntax (?&name). */
+         NAMED_REF_OR_RECURSE:
+         name = ++ptr;
+         while ((cd->ctypes[*ptr] & ctype_word) != 0) ptr++;
+         namelen = ptr - name;
+         /* In the pre-compile phase, do a syntax check and set a dummy
+         reference number. */
+         if (lengthptr != NULL)
+           {
+           if (*ptr != terminator)
              {
-             if (strncmp((char *)name, (char *)slot+2, namelen) == 0) break;
+             *errorcodeptr = ERR42;
-             slot += cd->name_entry_size;
+             goto FAILED;
              }
+           if (namelen > MAX_NAME_SIZE)
-           if (i < cd->names_found)         /* Back reference */
+             {
+             *errorcodeptr = ERR48;
+             goto FAILED;
+             }
+           recno = 0;
+           }
+         /* In the real compile, seek the name in the table */
+         else
+           {
+           slot = cd->name_table;
+           for (i = 0; i < cd->names_found; i++)
+             {
+             if (strncmp((char *)name, (char *)slot+2, namelen) == 0) break;
+             slot += cd->name_entry_size;
+             }
+           if (i < cd->names_found)         /* Back reference */
              {
              recno = GET2(slot, 0);
              }
            else if ((recno =                /* Forward back reference */
-                     find_named_parens(ptr, *brackets, name, namelen)) <= 0)
+                     find_parens(ptr, cd->bracount, name, namelen,
+                       (options & PCRE_EXTENDED) != 0)) <= 0)
              {
              *errorcodeptr = ERR15;
              goto FAILED;
              }
+           }
-           if (type == '>') goto HANDLE_RECURSION;  /* A few lines below */
+         /* In both phases, we can now go to the code than handles numerical
+         recursion or backreferences. */
-           /* Back reference */
-           previous = code;
+         if (is_recurse) goto HANDLE_RECURSION;
-           *code++ = OP_REF;
+           else goto HANDLE_REFERENCE;
-           PUT2INC(code, 0, recno);
-           cd->backref_map |= (recno < 32)? (1 << recno) : 1;
-           if (recno > cd->top_backref) cd->top_backref = recno;
-           continue;
-           }
-         /* Should never happen */
-         break;
-         case 'R':                 /* Pattern recursion */
+         /* ------------------------------------------------------------ */
+         case 'R':                 /* Recursion */
          ptr++;                    /* Same as (?0)      */
          /* Fall through */
-         /* Recursion or "subroutine" call */
-         case '0': case '1': case '2': case '3': case '4':
+         /* ------------------------------------------------------------ */
-         case '5': case '6': case '7': case '8': case '9':
+         case '0': case '1': case '2': case '3': case '4':   /* Recursion or */
+         case '5': case '6': case '7': case '8': case '9':   /* subroutine */
            {
            const uschar *called;
            recno = 0;
            while((digitab[*ptr] & ctype_digit) != 0)
              recno = recno * 10 + *ptr++ - '0';
+           if (*ptr != ')')
+             {
+             *errorcodeptr = ERR29;
+             goto FAILED;
+             }
            /* Come here from code above that handles a named recursion */
            HANDLE_RECURSION:
            previous = code;
+           called = cd->start_code;
-           /* Find the bracket that is being referenced. Temporarily end the
+           /* When we are actually compiling, find the bracket that is being
-           regex in case it doesn't exist. */
+           referenced. Temporarily end the regex in case it doesn't exist before
+           this point. If we end up with a forward reference, first check that
+           the bracket does occur later so we can give the error (and position)
+           now. Then remember this forward reference in the workspace so it can
+           be filled in at the end. */
-           *code = OP_END;
+           if (lengthptr == NULL)
-           called = (recno == 0)? cd->start_code :
-             find_bracket(cd->start_code, utf8, recno);
-           if (called == NULL)
              {
-             *errorcodeptr = ERR15;
+             *code = OP_END;
-             goto FAILED;
+             if (recno != 0) called = find_bracket(cd->start_code, utf8, recno);
-             }
-           /* If the subpattern is still open, this is a recursive call. We
+             /* Forward reference */
-           check to see if this is a left recursion that could loop for ever,
-           and diagnose that case. */
-           if (GET(called, 1) == 0 && could_be_empty(called, code, bcptr, utf8))
+             if (called == NULL)
-             {
+               {
-             *errorcodeptr = ERR40;
+               if (find_parens(ptr, cd->bracount, NULL, recno,
-             goto FAILED;
+                    (options & PCRE_EXTENDED) != 0) < 0)
+                 {
+                 *errorcodeptr = ERR15;
+                 goto FAILED;
+                 }
+               called = cd->start_code + recno;
+               PUTINC(cd->hwm, 0, code + 2 + LINK_SIZE - cd->start_code);
+               }
+             /* If not a forward reference, and the subpattern is still open,
+             this is a recursive call. We check to see if this is a left
+             recursion that could loop for ever, and diagnose that case. */
+             else if (GET(called, 1) == 0 &&
+                      could_be_empty(called, code, bcptr, utf8))
+               {
+               *errorcodeptr = ERR40;
+               goto FAILED;
+               }
              }
            /* Insert the recursion/subroutine item, automatically wrapped inside
-           "once" brackets. */
+           "once" brackets. Set up a "previous group" length so that a
+           subsequent quantifier will work. */
            *code = OP_ONCE;
            PUT(code, 1, 2 + 2*LINK_SIZE);
-Line 3174 
 for (;; ptr++)
+Line 4068 
 for (;; ptr++)
            *code = OP_KET;
            PUT(code, 1, 2 + 2*LINK_SIZE);
            code += 1 + LINK_SIZE;
+           length_prevgroup = 3 + 3*LINK_SIZE;
            }
+         /* Can't determine a first byte now */
+         if (firstbyte == REQ_UNSET) firstbyte = REQ_NONE;
          continue;
-         /* Character after (? not specially recognized */
-         default:                  /* Option setting */
+         /* ------------------------------------------------------------ */
+         default:              /* Other characters: check option setting */
          set = unset = 0;
          optset = &set;
-Line 3189 
 for (;; ptr++)
+Line 4089 
 for (;; ptr++)
              {
              case '-': optset = &unset; break;
+             case 'J':    /* Record that it changed in the external options */
+             *optset |= PCRE_DUPNAMES;
+             cd->external_options |= PCRE_JCHANGED;
+             break;
              case 'i': *optset |= PCRE_CASELESS; break;
-             case 'J': *optset |= PCRE_DUPNAMES; break;
              case 'm': *optset |= PCRE_MULTILINE; break;
              case 's': *optset |= PCRE_DOTALL; break;
              case 'x': *optset |= PCRE_EXTENDED; break;
              case 'U': *optset |= PCRE_UNGREEDY; break;
              case 'X': *optset |= PCRE_EXTRA; break;
+             default:  *errorcodeptr = ERR12;
+                       ptr--;    /* Correct the offset */
+                       goto FAILED;
              }
            }
-Line 3204 
 for (;; ptr++)
+Line 4112 
 for (;; ptr++)
          newoptions = (options | set) & (~unset);
          /* If the options ended with ')' this is not the start of a nested
-         group with option changes, so the options change at this level. Compile
+         group with option changes, so the options change at this level. If this
-         code to change the ims options if this setting actually changes any of
+         item is right at the start of the pattern, the options can be
-         them. We also pass the new setting back so that it can be put at the
+         abstracted and made external in the pre-compile phase, and ignored in
-         start of any following branches, and when this group ends (if we are in
+         the compile phase. This can be helpful when matching -- for instance in
-         a group), a resetting item can be compiled.
+         caseless checking of required bytes.
-         Note that if this item is right at the start of the pattern, the
+         If the code pointer is not (cd->start_code + 1 + LINK_SIZE), we are
-         options will have been abstracted and made global, so there will be no
+         definitely *not* at the start of the pattern because something has been
-         change to compile. */
+         compiled. In the pre-compile phase, however, the code pointer can have
+         that value after the start, because it gets reset as code is discarded
+         during the pre-compile. However, this can happen only at top level - if
+         we are within parentheses, the starting BRA will still be present. At
+         any parenthesis level, the length value can be used to test if anything
+         has been compiled at that level. Thus, a test for both these conditions
+         is necessary to ensure we correctly detect the start of the pattern in
+         both phases.
+         If we are not at the pattern start, compile code to change the ims
+         options if this setting actually changes any of them. We also pass the
+         new setting back so that it can be put at the start of any following
+         branches, and when this group ends (if we are in a group), a resetting
+         item can be compiled. */
          if (*ptr == ')')
            {
-           if ((options & PCRE_IMS) != (newoptions & PCRE_IMS))
+           if (code == cd->start_code + 1 + LINK_SIZE &&
+                (lengthptr == NULL || *lengthptr == 2 + 2*LINK_SIZE))
              {
-             *code++ = OP_OPT;
+             cd->external_options = newoptions;
-             *code++ = newoptions & PCRE_IMS;
+             options = newoptions;
              }
+          else
+             {
+             if ((options & PCRE_IMS) != (newoptions & PCRE_IMS))
+               {
+               *code++ = OP_OPT;
+               *code++ = newoptions & PCRE_IMS;
+               }
-           /* Change options at this level, and pass them back for use
+             /* Change options at this level, and pass them back for use
-           in subsequent branches. Reset the greedy defaults and the case
+             in subsequent branches. Reset the greedy defaults and the case
-           value for firstbyte and reqbyte. */
+             value for firstbyte and reqbyte. */
-           *optionsptr = options = newoptions;
+             *optionsptr = options = newoptions;
-           greedy_default = ((newoptions & PCRE_UNGREEDY) != 0);
+             greedy_default = ((newoptions & PCRE_UNGREEDY) != 0);
-           greedy_non_default = greedy_default ^ 1;
+             greedy_non_default = greedy_default ^ 1;
-           req_caseopt = ((options & PCRE_CASELESS) != 0)? REQ_CASELESS : 0;
+             req_caseopt = ((options & PCRE_CASELESS) != 0)? REQ_CASELESS : 0;
+             }
            previous = NULL;       /* This item can't be repeated */
            continue;              /* It is complete */
-Line 3242 
 for (;; ptr++)
+Line 4172 
 for (;; ptr++)
          bravalue = OP_BRA;
          ptr++;
-         }
+         }     /* End of switch for character following (? */
-       }
+       }       /* End of (? handling */
-     /* If PCRE_NO_AUTO_CAPTURE is set, all unadorned brackets become
+     /* Opening parenthesis not followed by '?'. If PCRE_NO_AUTO_CAPTURE is set,
-     non-capturing and behave like (?:...) brackets */
+     all unadorned brackets become non-capturing and behave like (?:...)
+     brackets. */
      else if ((options & PCRE_NO_AUTO_CAPTURE) != 0)
        {
        bravalue = OP_BRA;
        }
-     /* Else we have a referencing group; adjust the opcode. If the bracket
+     /* Else we have a capturing group. */
-     number is greater than EXTRACT_BASIC_MAX, we set the opcode one higher, and
-     arrange for the true number to follow later, in an OP_BRANUMBER item. */
      else
        {
        NUMBERED_GROUP:
-       if (++(*brackets) > EXTRACT_BASIC_MAX)
+       cd->bracount += 1;
-         {
+       PUT2(code, 1+LINK_SIZE, cd->bracount);
-         bravalue = OP_BRA + EXTRACT_BASIC_MAX + 1;
+       skipbytes = 2;
-         code[1+LINK_SIZE] = OP_BRANUMBER;
-         PUT2(code, 2+LINK_SIZE, *brackets);
-         skipbytes = 3;
-         }
-       else bravalue = OP_BRA + *brackets;
        }
-     /* Process nested bracketed re. Assertions may not be repeated, but other
+     /* Process nested bracketed regex. Assertions may not be repeated, but
-     kinds can be. We copy code into a non-register variable in order to be able
+     other kinds can be. All their opcodes are >= OP_ONCE. We copy code into a
-     to pass its address because some compilers complain otherwise. Pass in a
+     non-register variable in order to be able to pass its address because some
-     new setting for the ims options if they have changed. */
+     compilers complain otherwise. Pass in a new setting for the ims options if
+     they have changed. */
      previous = (bravalue >= OP_ONCE)? code : NULL;
      *code = bravalue;
      tempcode = code;
      tempreqvary = cd->req_varyopt;     /* Save value before bracket */
+     length_prevgroup = 0;              /* Initialize for pre-compile phase */
      if (!compile_regex(
           newoptions,                   /* The complete new option state */
           options & PCRE_IMS,           /* The previous ims option state */
-          brackets,                     /* Extracting bracket count */
           &tempcode,                    /* Where to put code (updated) */
           &ptr,                         /* Input pointer (updated) */
           errorcodeptr,                 /* Where to put an error message */
           (bravalue == OP_ASSERTBACK ||
            bravalue == OP_ASSERTBACK_NOT), /* TRUE if back assert */
-          skipbytes,                    /* Skip over OP_COND/OP_BRANUMBER */
+          skipbytes,                    /* Skip over bracket number */
           &subfirstbyte,                /* For possible first char */
           &subreqbyte,                  /* For possible last char */
           bcptr,                        /* Current branch chain */
-          cd))                          /* Tables block */
+          cd,                           /* Tables block */
+          (lengthptr == NULL)? NULL :   /* Actual compile phase */
+            &length_prevgroup           /* Pre-compile phase */
+          ))
        goto FAILED;
      /* At the end of compiling, code is still pointing to the start of the
-Line 3302 
 for (;; ptr++)
+Line 4230 
 for (;; ptr++)
      is on the bracket. */
      /* If this is a conditional bracket, check that there are no more than
-     two branches in the group. */
+     two branches in the group, or just one if it's a DEFINE group. */
-     else if (bravalue == OP_COND)
+     if (bravalue == OP_COND)
        {
        uschar *tc = code;
        int condcount = 0;
-Line 3315 
 for (;; ptr++)
+Line 4243 
 for (;; ptr++)
           }
        while (*tc != OP_KET);
-       if (condcount > 2)
+       /* A DEFINE group is never obeyed inline (the "condition" is always
+       false). It must have only one branch. */
+       if (code[LINK_SIZE+1] == OP_DEF)
          {
-         *errorcodeptr = ERR27;
+         if (condcount > 1)
-         goto FAILED;
+           {
+           *errorcodeptr = ERR54;
+           goto FAILED;
+           }
+         bravalue = OP_DEF;   /* Just a flag to suppress char handling below */
+         }
+       /* A "normal" conditional group. If there is just one branch, we must not
+       make use of its firstbyte or reqbyte, because this is equivalent to an
+       empty second branch. */
+       else
+         {
+         if (condcount > 2)
+           {
+           *errorcodeptr = ERR27;
+           goto FAILED;
+           }
+         if (condcount == 1) subfirstbyte = subreqbyte = REQ_NONE;
          }
+       }
+     /* Error if hit end of pattern */
-       /* If there is just one branch, we must not make use of its firstbyte or
+     if (*ptr != ')')
-       reqbyte, because this is equivalent to an empty second branch. */
+       {
+       *errorcodeptr = ERR14;
+       goto FAILED;
+       }
-       if (condcount == 1) subfirstbyte = subreqbyte = REQ_NONE;
+     /* In the pre-compile phase, update the length by the length of the nested
+     group, less the brackets at either end. Then reduce the compiled code to
+     just the brackets so that it doesn't use much memory if it is duplicated by
+     a quantifier. */
+     if (lengthptr != NULL)
+       {
+       *lengthptr += length_prevgroup - 2 - 2*LINK_SIZE;
+       code++;
+       PUTINC(code, 0, 1 + LINK_SIZE);
+       *code++ = OP_KET;
+       PUTINC(code, 0, 1 + LINK_SIZE);
        }
-     /* Handle updating of the required and first characters. Update for normal
+     /* Otherwise update the main code pointer to the end of the group. */
-     brackets of all kinds, and conditions with two branches (see code above).
-     If the bracket is followed by a quantifier with zero repeat, we have to
+     else code = tempcode;
-     back off. Hence the definition of zeroreqbyte and zerofirstbyte outside the
-     main loop so that they can be accessed for the back off. */
+     /* For a DEFINE group, required and first character settings are not
+     relevant. */
+     if (bravalue == OP_DEF) break;
+     /* Handle updating of the required and first characters for other types of
+     group. Update for normal brackets of all kinds, and conditions with two
+     branches (see code above). If the bracket is followed by a quantifier with
+     zero repeat, we have to back off. Hence the definition of zeroreqbyte and
+     zerofirstbyte outside the main loop so that they can be accessed for the
+     back off. */
      zeroreqbyte = reqbyte;
      zerofirstbyte = firstbyte;
      groupsetfirstbyte = FALSE;
-     if (bravalue >= OP_BRA || bravalue == OP_ONCE || bravalue == OP_COND)
+     if (bravalue >= OP_ONCE)
        {
        /* If we have not yet set a firstbyte in this branch, take it from the
        subpattern, remembering that it was set here so that a repeat of more
-Line 3378 
 for (;; ptr++)
+Line 4354 
 for (;; ptr++)
      firstbyte, looking for an asserted first char. */
      else if (bravalue == OP_ASSERT && subreqbyte >= 0) reqbyte = subreqbyte;
+     break;     /* End of processing '(' */
-     /* Now update the main code pointer to the end of the group. */
-     code = tempcode;
-     /* Error if hit end of pattern */
-     if (*ptr != ')')
-       {
-       *errorcodeptr = ERR14;
-       goto FAILED;
-       }
-     break;
-     /* Check \ for being a real metacharacter; if not, fall through and handle
-     it as a data character at the start of a string. Escape items are checked
-     for validity in the pre-compiling pass. */
-     case '\\':
-     tempptr = ptr;
-     c = check_escape(&ptr, errorcodeptr, *brackets, options, FALSE);
-     /* Handle metacharacters introduced by \. For ones like \d, the ESC_ values
+     /* ===================================================================*/
+     /* Handle metasequences introduced by \. For ones like \d, the ESC_ values
      are arranged to be the negation of the corresponding OP_values. For the
      back references, the values are ESC_REF plus the reference number. Only
      back references and those types that consume a character may be repeated.
      We can test for values between ESC_b and ESC_Z for the latter; this may
      have to change if any new ones are ever created. */
+     case '\\':
+     tempptr = ptr;
+     c = check_escape(&ptr, errorcodeptr, cd->bracount, options, FALSE);
+     if (*errorcodeptr != 0) goto FAILED;
      if (c < 0)
        {
        if (-c == ESC_Q)            /* Handle start of quoted string */
-Line 3416 
 for (;; ptr++)
+Line 4379 
 for (;; ptr++)
          continue;
          }
+       if (-c == ESC_E) continue;  /* Perl ignores an orphan \E */
        /* For metasequences that actually match a character, we disable the
        setting of a first character if it hasn't already been set. */
-Line 3427 
 for (;; ptr++)
+Line 4392 
 for (;; ptr++)
        zerofirstbyte = firstbyte;
        zeroreqbyte = reqbyte;
-       /* Back references are handled specially */
+       /* \k<name> or \k'name' is a back reference by name (Perl syntax) */
+       if (-c == ESC_k && (ptr[1] == '<' || ptr[1] == '\''))
+         {
+         is_recurse = FALSE;
+         terminator = (*(++ptr) == '<')? '>' : '\'';
+         goto NAMED_REF_OR_RECURSE;
+         }
+       /* Back references are handled specially; must disable firstbyte if
+       not set to cope with cases like (?=(\w+))\1: which would otherwise set
+       ':' later. */
        if (-c >= ESC_REF)
          {
-         int number = -c - ESC_REF;
+         recno = -c - ESC_REF;
+         HANDLE_REFERENCE:    /* Come here from named backref handling */
+         if (firstbyte == REQ_UNSET) firstbyte = REQ_NONE;
          previous = code;
          *code++ = OP_REF;
-         PUT2INC(code, 0, number);
+         PUT2INC(code, 0, recno);
+         cd->backref_map |= (recno < 32)? (1 << recno) : 1;
+         if (recno > cd->top_backref) cd->top_backref = recno;
          }
-       /* So are Unicode property matches, if supported. We know that get_ucp
+       /* So are Unicode property matches, if supported. */
-       won't fail because it was tested in the pre-pass. */
  #ifdef SUPPORT_UCP
        else if (-c == ESC_P || -c == ESC_p)
-Line 3446 
 for (;; ptr++)
+Line 4426 
 for (;; ptr++)
          BOOL negated;
          int pdata;
          int ptype = get_ucp(&ptr, &negated, &pdata, errorcodeptr);
+         if (ptype < 0) goto FAILED;
          previous = code;
          *code++ = ((-c == ESC_p) != negated)? OP_PROP : OP_NOTPROP;
          *code++ = ptype;
          *code++ = pdata;
          }
+ #else
+       /* If Unicode properties are not supported, \X, \P, and \p are not
+       allowed. */
+       else if (-c == ESC_X || -c == ESC_P || -c == ESC_p)
+         {
+         *errorcodeptr = ERR45;
+         goto FAILED;
+         }
  #endif
-       /* For the rest, we can obtain the OP value by negating the escape
+       /* For the rest (including \X when Unicode properties are supported), we
-       value */
+       can obtain the OP value by negating the escape value. */
        else
          {
-Line 3478 
 for (;; ptr++)
+Line 4469 
 for (;; ptr++)
       mcbuffer[0] = c;
       mclength = 1;
       }
      goto ONE_CHAR;
+     /* ===================================================================*/
      /* Handle a literal character. It is guaranteed not to be whitespace or #
      when the extended flag is set. If we are in UTF-8 mode, it may be a
      multi-byte literal character. */
-Line 3491 
 for (;; ptr++)
+Line 4483 
 for (;; ptr++)
      mcbuffer[0] = c;
  #ifdef SUPPORT_UTF8
-     if (utf8 && (c & 0xc0) == 0xc0)
+     if (utf8 && c >= 0xc0)
        {
        while ((ptr[1] & 0xc0) == 0x80)
          mcbuffer[mclength++] = *(++ptr);
-Line 3542 
 for (;; ptr++)
+Line 4534 
 for (;; ptr++)
      }
    }                   /* end of big loop */
  /* Control never reaches here by falling through, only by a goto for all the
  error states. Pass back the position in the pattern so that it can be displayed
  to the user for diagnosing the error. */
-Line 3558 
 return FALSE;
+Line 4551 
 return FALSE;
  *     Compile sequence of alternatives           *
  *************************************************/
- /* On entry, ptr is pointing past the bracket character, but on return
+ /* On entry, ptr is pointing past the bracket character, but on return it
- it points to the closing bracket, or vertical bar, or end of string.
+ points to the closing bracket, or vertical bar, or end of string. The code
- The code variable is pointing at the byte into which the BRA operator has been
+ variable is pointing at the byte into which the BRA operator has been stored.
- stored. If the ims options are changed at the start (for a (?ims: group) or
+ If the ims options are changed at the start (for a (?ims: group) or during any
- during any branch, we need to insert an OP_OPT item at the start of every
+ branch, we need to insert an OP_OPT item at the start of every following branch
- following branch to ensure they get set correctly at run time, and also pass
+ to ensure they get set correctly at run time, and also pass the new options
- the new options into every subsequent branch compile.
+ into every subsequent branch compile.
+ This function is used during the pre-compile phase when we are trying to find
+ out the amount of memory needed, as well as during the real compile phase. The
+ value of lengthptr distinguishes the two phases.
  Argument:
    options        option bits, including any changes for this subpattern
    oldims         previous settings of ims option bits
-   brackets       -> int containing the number of extracting brackets used
    codeptr        -> the address of the current code pointer
    ptrptr         -> the address of the current pattern pointer
    errorcodeptr   -> pointer to error code variable
    lookbehind     TRUE if this is a lookbehind assertion
-   skipbytes      skip this many bytes at start (for OP_COND, OP_BRANUMBER)
+   skipbytes      skip this many bytes at start (for brackets and OP_COND)
    firstbyteptr   place to put the first required character, or a negative number
    reqbyteptr     place to put the last required character, or a negative number
    bcptr          pointer to the chain of currently open branches
    cd             points to the data block with tables pointers etc.
+   lengthptr      NULL during the real compile phase
+                  points to length accumulator during pre-compile phase
- Returns:      TRUE on success
+ Returns:         TRUE on success
  */
  static BOOL
- compile_regex(int options, int oldims, int *brackets, uschar **codeptr,
+ compile_regex(int options, int oldims, uschar **codeptr, const uschar **ptrptr,
-   const uschar **ptrptr, int *errorcodeptr, BOOL lookbehind, int skipbytes,
+   int *errorcodeptr, BOOL lookbehind, int skipbytes, int *firstbyteptr,
-   int *firstbyteptr, int *reqbyteptr, branch_chain *bcptr, compile_data *cd)
+   int *reqbyteptr, branch_chain *bcptr, compile_data *cd, int *lengthptr)
  {
  const uschar *ptr = *ptrptr;
  uschar *code = *codeptr;
-Line 3595 
 uschar *start_bracket = code;
+Line 4593 
 uschar *start_bracket = code;
  uschar *reverse_count = NULL;
  int firstbyte, reqbyte;
  int branchfirstbyte, branchreqbyte;
+ int length;
  branch_chain bc;
  bc.outer = bcptr;
-Line 3602 
 bc.current = code;
+Line 4601 
 bc.current = code;
  firstbyte = reqbyte = REQ_UNSET;
+ /* Accumulate the length for use in the pre-compile phase. Start with the
+ length of the BRA and KET and any extra bytes that are required at the
+ beginning. We accumulate in a local variable to save frequent testing of
+ lenthptr for NULL. We cannot do this by looking at the value of code at the
+ start and end of each alternative, because compiled items are discarded during
+ the pre-compile phase so that the work space is not exceeded. */
+ length = 2 + 2*LINK_SIZE + skipbytes;
+ /* WARNING: If the above line is changed for any reason, you must also change
+ the code that abstracts option settings at the start of the pattern and makes
+ them global. It tests the value of length for (2 + 2*LINK_SIZE) in the
+ pre-compile phase to find out whether anything has yet been compiled or not. */
  /* Offset is set zero to mark that this bracket is still open */
  PUT(code, 1, 0);
-Line 3617 
 for (;;)
+Line 4630 
 for (;;)
      {
      *code++ = OP_OPT;
      *code++ = options & PCRE_IMS;
+     length += 2;
      }
    /* Set up dummy OP_REVERSE if lookbehind assertion */
-Line 3626 
 for (;;)
+Line 4640 
 for (;;)
      *code++ = OP_REVERSE;
      reverse_count = code;
      PUTINC(code, 0, 0);
+     length += 1 + LINK_SIZE;
      }
-   /* Now compile the branch */
+   /* Now compile the branch; in the pre-compile phase its length gets added
+   into the length. */
-   if (!compile_branch(&options, brackets, &code, &ptr, errorcodeptr,
+   if (!compile_branch(&options, &code, &ptr, errorcodeptr, &branchfirstbyte,
-         &branchfirstbyte, &branchreqbyte, &bc, cd))
+         &branchreqbyte, &bc, cd, (lengthptr == NULL)? NULL : &length))
      {
      *ptrptr = ptr;
      return FALSE;
      }
-   /* If this is the first branch, the firstbyte and reqbyte values for the
+   /* In the real compile phase, there is some post-processing to be done. */
-   branch become the values for the regex. */
-   if (*last_branch != OP_ALT)
+   if (lengthptr == NULL)
      {
-     firstbyte = branchfirstbyte;
+     /* If this is the first branch, the firstbyte and reqbyte values for the
-     reqbyte = branchreqbyte;
+     branch become the values for the regex. */
-     }
-   /* If this is not the first branch, the first char and reqbyte have to
+     if (*last_branch != OP_ALT)
-   match the values from all the previous branches, except that if the previous
+       {
-   value for reqbyte didn't have REQ_VARY set, it can still match, and we set
+       firstbyte = branchfirstbyte;
-   REQ_VARY for the regex. */
+       reqbyte = branchreqbyte;
+       }
-   else
+     /* If this is not the first branch, the first char and reqbyte have to
-     {
+     match the values from all the previous branches, except that if the
-     /* If we previously had a firstbyte, but it doesn't match the new branch,
+     previous value for reqbyte didn't have REQ_VARY set, it can still match,
-     we have to abandon the firstbyte for the regex, but if there was previously
+     and we set REQ_VARY for the regex. */
-     no reqbyte, it takes on the value of the old firstbyte. */
-     if (firstbyte >= 0 && firstbyte != branchfirstbyte)
+     else
        {
-       if (reqbyte < 0) reqbyte = firstbyte;
+       /* If we previously had a firstbyte, but it doesn't match the new branch,
-       firstbyte = REQ_NONE;
+       we have to abandon the firstbyte for the regex, but if there was
-       }
+       previously no reqbyte, it takes on the value of the old firstbyte. */
+       if (firstbyte >= 0 && firstbyte != branchfirstbyte)
+         {
+         if (reqbyte < 0) reqbyte = firstbyte;
+         firstbyte = REQ_NONE;
+         }
-     /* If we (now or from before) have no firstbyte, a firstbyte from the
+       /* If we (now or from before) have no firstbyte, a firstbyte from the
-     branch becomes a reqbyte if there isn't a branch reqbyte. */
+       branch becomes a reqbyte if there isn't a branch reqbyte. */
-     if (firstbyte < 0 && branchfirstbyte >= 0 && branchreqbyte < 0)
+       if (firstbyte < 0 && branchfirstbyte >= 0 && branchreqbyte < 0)
-         branchreqbyte = branchfirstbyte;
+           branchreqbyte = branchfirstbyte;
-     /* Now ensure that the reqbytes match */
+       /* Now ensure that the reqbytes match */
-     if ((reqbyte & ~REQ_VARY) != (branchreqbyte & ~REQ_VARY))
+       if ((reqbyte & ~REQ_VARY) != (branchreqbyte & ~REQ_VARY))
-       reqbyte = REQ_NONE;
+         reqbyte = REQ_NONE;
-     else reqbyte |= branchreqbyte;   /* To "or" REQ_VARY */
+       else reqbyte |= branchreqbyte;   /* To "or" REQ_VARY */
-     }
+       }
-   /* If lookbehind, check that this branch matches a fixed-length string,
+     /* If lookbehind, check that this branch matches a fixed-length string, and
-   and put the length into the OP_REVERSE item. Temporarily mark the end of
+     put the length into the OP_REVERSE item. Temporarily mark the end of the
-   the branch with OP_END. */
+     branch with OP_END. */
-   if (lookbehind)
+     if (lookbehind)
-     {
-     int length;
-     *code = OP_END;
-     length = find_fixedlength(last_branch, options);
-     DPRINTF(("fixed length = %d\n", length));
-     if (length < 0)
        {
-       *errorcodeptr = (length == -2)? ERR36 : ERR25;
+       int fixed_length;
-       *ptrptr = ptr;
+       *code = OP_END;
-       return FALSE;
+       fixed_length = find_fixedlength(last_branch, options);
+       DPRINTF(("fixed length = %d\n", fixed_length));
+       if (fixed_length < 0)
+         {
+         *errorcodeptr = (fixed_length == -2)? ERR36 : ERR25;
+         *ptrptr = ptr;
+         return FALSE;
+         }
+       PUT(reverse_count, 0, fixed_length);
        }
-     PUT(reverse_count, 0, length);
      }
    /* Reached end of expression, either ')' or end of pattern. Go back through
-Line 3706 
 for (;;)
+Line 4727 
 for (;;)
    if (*ptr != '|')
      {
-     int length = code - last_branch;
+     int branch_length = code - last_branch;
      do
        {
        int prev_length = GET(last_branch, 1);
-       PUT(last_branch, 1, length);
+       PUT(last_branch, 1, branch_length);
-       length = prev_length;
+       branch_length = prev_length;
-       last_branch -= length;
+       last_branch -= branch_length;
        }
-     while (length > 0);
+     while (branch_length > 0);
      /* Fill in the ket */
-Line 3728 
 for (;;)
+Line 4749 
 for (;;)
        {
        *code++ = OP_OPT;
        *code++ = oldims;
+       length += 2;
        }
      /* Set values to pass back */
-Line 3736 
 for (;;)
+Line 4758 
 for (;;)
      *ptrptr = ptr;
      *firstbyteptr = firstbyte;
      *reqbyteptr = reqbyte;
+     if (lengthptr != NULL) *lengthptr += length;
      return TRUE;
      }
-Line 3749 
 for (;;)
+Line 4772 
 for (;;)
    bc.current = last_branch = code;
    code += 1 + LINK_SIZE;
    ptr++;
+   length += 1 + LINK_SIZE;
    }
  /* Control never reaches here */
  }
-Line 3799 
 is_anchored(register const uschar *code,
+Line 4823 
 is_anchored(register const uschar *code,
    unsigned int backref_map)
  {
  do {
-    const uschar *scode =
+    const uschar *scode = first_significant_code(code + _pcre_OP_lengths[*code],
-      first_significant_code(code + 1+LINK_SIZE, options, PCRE_MULTILINE, FALSE);
+      options, PCRE_MULTILINE, FALSE);
     register int op = *scode;
+    /* Non-capturing brackets */
+    if (op == OP_BRA)
+      {
+      if (!is_anchored(scode, options, bracket_map, backref_map)) return FALSE;
+      }
     /* Capturing brackets */
-    if (op > OP_BRA)
+    else if (op == OP_CBRA)
       {
-      int new_map;
+      int n = GET2(scode, 1+LINK_SIZE);
-      op -= OP_BRA;
+      int new_map = bracket_map | ((n < 32)? (1 << n) : 1);
-      if (op > EXTRACT_BASIC_MAX) op = GET2(scode, 2+LINK_SIZE);
-      new_map = bracket_map | ((op < 32)? (1 << op) : 1);
       if (!is_anchored(scode, options, new_map, backref_map)) return FALSE;
       }
     /* Other brackets */
-    else if (op == OP_BRA || op == OP_ASSERT || op == OP_ONCE || op == OP_COND)
+    else if (op == OP_ASSERT || op == OP_ONCE || op == OP_COND)
       {
       if (!is_anchored(scode, options, bracket_map, backref_map)) return FALSE;
       }
-Line 3824 
 do {
+Line 4853 
 do {
     /* .* is not anchored unless DOTALL is set and it isn't in brackets that
     are or may be referenced. */
-    else if ((op == OP_TYPESTAR || op == OP_TYPEMINSTAR) &&
+    else if ((op == OP_TYPESTAR || op == OP_TYPEMINSTAR ||
+              op == OP_TYPEPOSSTAR) &&
              (*options & PCRE_DOTALL) != 0)
       {
       if (scode[1] != OP_ANY || (bracket_map & backref_map) != 0) return FALSE;
-Line 3869 
 is_startline(const uschar *code, unsigne
+Line 4899 
 is_startline(const uschar *code, unsigne
    unsigned int backref_map)
  {
  do {
-    const uschar *scode = first_significant_code(code + 1+LINK_SIZE, NULL, 0,
+    const uschar *scode = first_significant_code(code + _pcre_OP_lengths[*code],
-      FALSE);
+      NULL, 0, FALSE);
     register int op = *scode;
+    /* Non-capturing brackets */
+    if (op == OP_BRA)
+      {
+      if (!is_startline(scode, bracket_map, backref_map)) return FALSE;
+      }
     /* Capturing brackets */
-    if (op > OP_BRA)
+    else if (op == OP_CBRA)
       {
-      int new_map;
+      int n = GET2(scode, 1+LINK_SIZE);
-      op -= OP_BRA;
+      int new_map = bracket_map | ((n < 32)? (1 << n) : 1);
-      if (op > EXTRACT_BASIC_MAX) op = GET2(scode, 2+LINK_SIZE);
-      new_map = bracket_map | ((op < 32)? (1 << op) : 1);
       if (!is_startline(scode, new_map, backref_map)) return FALSE;
       }
     /* Other brackets */
-    else if (op == OP_BRA || op == OP_ASSERT || op == OP_ONCE || op == OP_COND)
+    else if (op == OP_ASSERT || op == OP_ONCE || op == OP_COND)
       { if (!is_startline(scode, bracket_map, backref_map)) return FALSE; }
     /* .* means "start at start or after \n" if it isn't in brackets that
     may be referenced. */
-    else if (op == OP_TYPESTAR || op == OP_TYPEMINSTAR)
+    else if (op == OP_TYPESTAR || op == OP_TYPEMINSTAR || op == OP_TYPEPOSSTAR)
       {
       if (scode[1] != OP_ANY || (bracket_map & backref_map) != 0) return FALSE;
       }
-Line 3941 
 do {
+Line 4976 
 do {
       first_significant_code(code + 1+LINK_SIZE, options, PCRE_CASELESS, TRUE);
     register int op = *scode;
-    if (op >= OP_BRA) op = OP_BRA;
     switch(op)
       {
       default:
       return -1;
       case OP_BRA:
+      case OP_CBRA:
       case OP_ASSERT:
       case OP_ONCE:
       case OP_COND:
-Line 3964 
 do {
+Line 4998 
 do {
       case OP_CHARNC:
       case OP_PLUS:
       case OP_MINPLUS:
+      case OP_POSPLUS:
       if (!inassert) return -1;
       if (c < 0)
         {
-Line 4012 
 return pcre_compile2(pattern, options, N
+Line 5047 
 return pcre_compile2(pattern, options, N
  }
  PCRE_DATA_SCOPE pcre *
  pcre_compile2(const char *pattern, int options, int *errorcodeptr,
    const char **errorptr, int *erroroffset, const unsigned char *tables)
  {
  real_pcre *re;
- int length = 1 + LINK_SIZE;      /* For initial BRA plus length */
+ int length = 1;  /* For final END opcode */
- int c, firstbyte, reqbyte, newline;
+ int firstbyte, reqbyte, newline;
- int bracount = 0;
- int branch_extra = 0;
- int branch_newextra;
- int item_count = -1;
- int name_count = 0;
- int max_name_size = 0;
- int lastitemlength = 0;
  int errorcode = 0;
  #ifdef SUPPORT_UTF8
  BOOL utf8;
- BOOL class_utf8;
  #endif
- BOOL inescq = FALSE;
- BOOL capturing;
- unsigned int brastackptr = 0;
  size_t size;
  uschar *code;
  const uschar *codestart;
  const uschar *ptr;
  compile_data compile_block;
  compile_data *cd = &compile_block;
- int brastack[BRASTACK_SIZE];
- uschar bralenstack[BRASTACK_SIZE];
+ /* This space is used for "compiling" into during the first phase, when we are
+ computing the amount of memory that is needed. Compiled items are thrown away
+ as soon as possible, so that a fairly large buffer should be sufficient for
+ this purpose. The same space is used in the second phase for remembering where
+ to fill in forward references to subpatterns. */
+ uschar cworkspace[COMPILE_WORK_SIZE];
+ /* Set this early so that early errors get offset 0. */
+ ptr = (const uschar *)pattern;
  /* We can't pass back an error message if errorptr is NULL; I guess the best we
  can do is just return NULL, but we can set a code value if there is a code
-Line 4075 
 if (utf8 && (options & PCRE_NO_UTF8_CHEC
+Line 5109 
 if (utf8 && (options & PCRE_NO_UTF8_CHEC
       (*erroroffset = _pcre_valid_utf8((uschar *)pattern, -1)) >= 0)
    {
    errorcode = ERR44;
-   goto PCRE_EARLY_ERROR_RETURN;
+   goto PCRE_UTF8_ERROR_RETURN;
    }
  #else
  if ((options & PCRE_UTF8) != 0)
-Line 4099 
 cd->fcc = tables + fcc_offset;
+Line 5133 
 cd->fcc = tables + fcc_offset;
  cd->cbits = tables + cbits_offset;
  cd->ctypes = tables + ctypes_offset;
- /* Handle different types of newline. The two bits give four cases. The current
+ /* Handle different types of newline. The three bits give seven cases. The
- code allows for one- or two-byte sequences. */
+ current code allows for fixed one- or two-byte sequences, plus "any". */
- switch (options & PCRE_NEWLINE_CRLF)
+ switch (options & (PCRE_NEWLINE_CRLF | PCRE_NEWLINE_ANY))
    {
-   default:              newline = NEWLINE; break;   /* Compile-time default */
+   case 0: newline = NEWLINE; break;   /* Compile-time default */
    case PCRE_NEWLINE_CR: newline = '\r'; break;
    case PCRE_NEWLINE_LF: newline = '\n'; break;
    case PCRE_NEWLINE_CR+
         PCRE_NEWLINE_LF: newline = ('\r' << 8) | '\n'; break;
+   case PCRE_NEWLINE_ANY: newline = -1; break;
+   default: errorcode = ERR56; goto PCRE_EARLY_ERROR_RETURN;
    }
- if (newline > 255)
+ if (newline < 0)
    {
-   cd->nllen = 2;
+   cd->nltype = NLTYPE_ANY;
-   cd->nl[0] = (newline >> 8) & 255;
-   cd->nl[1] = newline & 255;
    }
  else
    {
-   cd->nllen = 1;
+   cd->nltype = NLTYPE_FIXED;
-   cd->nl[0] = newline;
+   if (newline > 255)
+     {
+     cd->nllen = 2;
+     cd->nl[0] = (newline >> 8) & 255;
+     cd->nl[1] = newline & 255;
+     }
+   else
+     {
+     cd->nllen = 1;
+     cd->nl[0] = newline;
+     }
    }
- /* Maximum back reference and backref bitmap. This is updated for numeric
+ /* Maximum back reference and backref bitmap. The bitmap records up to 31 back
- references during the first pass, but for named references during the actual
+ references to help in deciding whether (.*) can be treated as anchored or not.
- compile pass. The bitmap records up to 31 back references to help in deciding
+ */
- whether (.*) can be treated as anchored or not. */
  cd->top_backref = 0;
  cd->backref_map = 0;
-Line 4136 
 cd->backref_map = 0;
+Line 5179 
 cd->backref_map = 0;
  DPRINTF(("------------------------------------------------------------------\n"));
  DPRINTF(("%s\n", pattern));
- /* The first thing to do is to make a pass over the pattern to compute the
+ /* Pretend to compile the pattern while actually just accumulating the length
- amount of store required to hold the compiled code. This does not have to be
+ of memory required. This behaviour is triggered by passing a non-NULL final
- perfect as long as errors are overestimates. At the same time we can detect any
+ argument to compile_regex(). We pass a block of workspace (cworkspace) for it
- flag settings right at the start, and extract them. Make an attempt to correct
+ to compile parts of the pattern into; the compiled code is discarded when it is
- for any counted white space if an "extended" flag setting appears late in the
+ no longer needed, so hopefully this workspace will never overflow, though there
- pattern. We can't be so clever for #-comments. */
+ is a test for its doing so. */
- ptr = (const uschar *)(pattern - 1);
- while ((c = *(++ptr)) != 0)
-   {
-   int min, max;
-   int class_optcount;
-   int bracket_length;
-   int duplength;
-   /* If we are inside a \Q...\E sequence, all chars are literal */
+ cd->bracount = 0;
+ cd->names_found = 0;
+ cd->name_entry_size = 0;
+ cd->name_table = NULL;
+ cd->start_workspace = cworkspace;
+ cd->start_code = cworkspace;
+ cd->hwm = cworkspace;
+ cd->start_pattern = (const uschar *)pattern;
+ cd->end_pattern = (const uschar *)(pattern + strlen(pattern));
+ cd->req_varyopt = 0;
+ cd->nopartial = FALSE;
+ cd->external_options = options;
-   if (inescq)
+ /* Now do the pre-compile. On error, errorcode will be set non-zero, so we
-     {
+ don't need to look at the result of the function here. The initial options have
-     if ((options & PCRE_AUTO_CALLOUT) != 0) length += 2 + 2*LINK_SIZE;
+ been put into the cd block so that they can be changed if an option setting is
-     goto NORMAL_CHAR;
+ found within the regex right at the beginning. Bringing initial option settings
-     }
+ outside can help speed up starting point checks. */
-   /* Otherwise, first check for ignored whitespace and comments */
+ code = cworkspace;
+ *code = OP_BRA;
+ (void)compile_regex(cd->external_options, cd->external_options & PCRE_IMS,
+   &code, &ptr, &errorcode, FALSE, 0, &firstbyte, &reqbyte, NULL, cd, &length);
+ if (errorcode != 0) goto PCRE_EARLY_ERROR_RETURN;
-   if ((options & PCRE_EXTENDED) != 0)
+ DPRINTF(("end pre-compile: length=%d workspace=%d\n", length,
-     {
+   cd->hwm - cworkspace));
-     if ((cd->ctypes[c] & ctype_space) != 0) continue;
-     if (c == '#')
-       {
-       while (*(++ptr) != 0) if (IS_NEWLINE(ptr)) break;
-       if (*ptr != 0)
-         {
-         ptr += cd->nllen - 1;
-         continue;
-         }
-       break;    /* End loop at end of pattern */
-       }
-     }
-   item_count++;    /* Is zero for the first non-comment item */
+ if (length > MAX_PATTERN_SIZE)
+   {
+   errorcode = ERR20;
+   goto PCRE_EARLY_ERROR_RETURN;
+   }
-   /* Allow space for auto callout before every item except quantifiers. */
+ /* Compute the size of data block needed and get it, either from malloc or
+ externally provided function. Integer overflow should no longer be possible
+ because nowadays we limit the maximum value of cd->names_found and
+ cd->name_entry_size. */
-   if ((options & PCRE_AUTO_CALLOUT) != 0 &&
+ size = length + sizeof(real_pcre) + cd->names_found * (cd->name_entry_size + 3);
-        c != '*' && c != '+' && c != '?' &&
+ re = (real_pcre *)(pcre_malloc)(size);
-        (c != '{' || !is_counted_repeat(ptr + 1)))
-     length += 2 + 2*LINK_SIZE;
-   switch(c)
+ if (re == NULL)
-     {
+   {
-     /* A backslashed item may be an escaped data character or it may be a
+   errorcode = ERR21;
-     character type. */
+   goto PCRE_EARLY_ERROR_RETURN;
+   }
-     case '\\':
+ /* Put in the magic number, and save the sizes, initial options, and character
-     c = check_escape(&ptr, &errorcode, bracount, options, FALSE);
+ table pointer. NULL is used for the default character tables. The nullpad field
-     if (errorcode != 0) goto PCRE_ERROR_RETURN;
+ is at the end; it's there to help in the case when a regex compiled on a system
+ with 4-byte pointers is run on another with 8-byte pointers. */
-     lastitemlength = 1;     /* Default length of last item for repeats */
+ re->magic_number = MAGIC_NUMBER;
+ re->size = size;
+ re->options = cd->external_options;
+ re->dummy1 = 0;
+ re->first_byte = 0;
+ re->req_byte = 0;
+ re->name_table_offset = sizeof(real_pcre);
+ re->name_entry_size = cd->name_entry_size;
+ re->name_count = cd->names_found;
+ re->ref_count = 0;
+ re->tables = (tables == _pcre_default_tables)? NULL : tables;
+ re->nullpad = NULL;
-     if (c >= 0)             /* Data character */
+ /* The starting points of the name/number translation table and of the code are
-       {
+ passed around in the compile data block. The start/end pattern and initial
-       length += 2;          /* For a one-byte character */
+ options are already set from the pre-compile phase, as is the name_entry_size
+ field. Reset the bracket count and the names_found field. Also reset the hwm
+ field; this time it's used for remembering forward references to subpatterns.
+ */
- #ifdef SUPPORT_UTF8
+ cd->bracount = 0;
-       if (utf8 && c > 127)
+ cd->names_found = 0;
-         {
+ cd->name_table = (uschar *)re + re->name_table_offset;
-         int i;
+ codestart = cd->name_table + re->name_entry_size * re->name_count;
-         for (i = 0; i < _pcre_utf8_table1_size; i++)
+ cd->start_code = codestart;
-           if (c <= _pcre_utf8_table1[i]) break;
+ cd->hwm = cworkspace;
-         length += i;
+ cd->req_varyopt = 0;
-         lastitemlength += i;
+ cd->nopartial = FALSE;
-         }
- #endif
-       continue;
+ /* Set up a starting, non-extracting bracket, then compile the expression. On
-       }
+ error, errorcode will be set non-zero, so we don't need to look at the result
+ of the function here. */
-     /* If \Q, enter "literal" mode */
+ ptr = (const uschar *)pattern;
+ code = (uschar *)codestart;
+ *code = OP_BRA;
+ (void)compile_regex(re->options, re->options & PCRE_IMS, &code, &ptr,
+   &errorcode, FALSE, 0, &firstbyte, &reqbyte, NULL, cd, NULL);
+ re->top_bracket = cd->bracount;
+ re->top_backref = cd->top_backref;
-     if (-c == ESC_Q)
+ if (cd->nopartial) re->options |= PCRE_NOPARTIAL;
-       {
-       inescq = TRUE;
-       continue;
-       }
-     /* \X is supported only if Unicode property support is compiled */
+ /* If not reached end of pattern on success, there's an excess bracket. */
- #ifndef SUPPORT_UCP
+ if (errorcode == 0 && *ptr != 0) errorcode = ERR22;
-     if (-c == ESC_X)
-       {
-       errorcode = ERR45;
-       goto PCRE_ERROR_RETURN;
-       }
- #endif
-     /* \P and \p are for Unicode properties, but only when the support has
+ /* Fill in the terminating state and check for disastrous overflow, but
-     been compiled. Each item needs 3 bytes. */
+ if debugging, leave the test till after things are printed out. */
-     else if (-c == ESC_P || -c == ESC_p)
+ *code++ = OP_END;
-       {
- #ifdef SUPPORT_UCP
+ #ifndef DEBUG
-       BOOL negated;
+ if (code - codestart > length) errorcode = ERR23;
-       BOOL pdata;
-       length += 3;
-       lastitemlength = 3;
-       if (get_ucp(&ptr, &negated, &pdata, &errorcode) < 0)
-         goto PCRE_ERROR_RETURN;
-       continue;
- #else
-       errorcode = ERR45;
-       goto PCRE_ERROR_RETURN;
  #endif
-       }
-     /* Other escapes need one byte */
+ /* Fill in any forward references that are required. */
-     length++;
+ while (errorcode == 0 && cd->hwm > cworkspace)
+   {
+   int offset, recno;
+   const uschar *groupptr;
+   cd->hwm -= LINK_SIZE;
+   offset = GET(cd->hwm, 0);
+   recno = GET(codestart, offset);
+   groupptr = find_bracket(codestart, (re->options & PCRE_UTF8) != 0, recno);
+   if (groupptr == NULL) errorcode = ERR53;
+     else PUT(((uschar *)codestart), offset, groupptr - codestart);
+   }
-     /* A back reference needs an additional 2 bytes, plus either one or 5
+ /* Give an error if there's back reference to a non-existent capturing
-     bytes for a repeat. We also need to keep the value of the highest
+ subpattern. */
-     back reference. */
-     if (c <= -ESC_REF)
+ if (errorcode == 0 && re->top_backref > re->top_bracket) errorcode = ERR15;
-       {
-       int refnum = -c - ESC_REF;
-       cd->backref_map |= (refnum < 32)? (1 << refnum) : 1;
-       if (refnum > cd->top_backref)
-         cd->top_backref = refnum;
-       length += 2;   /* For single back reference */
-       if (ptr[1] == '{' && is_counted_repeat(ptr+2))
-         {
-         ptr = read_repeat_counts(ptr+2, &min, &max, &errorcode);
-         if (errorcode != 0) goto PCRE_ERROR_RETURN;
-         if ((min == 0 && (max == 1 || max == -1)) ||
-           (min == 1 && max == -1))
-             length++;
-         else length += 5;
-         if (ptr[1] == '?') ptr++;
-         }
-       }
-     continue;
-     case '^':     /* Single-byte metacharacters */
-     case '.':
-     case '$':
-     length++;
-     lastitemlength = 1;
-     continue;
-     case '*':            /* These repeats won't be after brackets; */
-     case '+':            /* those are handled separately */
-     case '?':
-     length++;
-     goto POSESSIVE;      /* A few lines below */
-     /* This covers the cases of braced repeats after a single char, metachar,
-     class, or back reference. */
-     case '{':
-     if (!is_counted_repeat(ptr+1)) goto NORMAL_CHAR;
-     ptr = read_repeat_counts(ptr+1, &min, &max, &errorcode);
-     if (errorcode != 0) goto PCRE_ERROR_RETURN;
-     /* These special cases just insert one extra opcode */
-     if ((min == 0 && (max == 1 || max == -1)) ||
-       (min == 1 && max == -1))
-         length++;
-     /* These cases might insert additional copies of a preceding character. */
-     else
-       {
-       if (min != 1)
-         {
-         length -= lastitemlength;   /* Uncount the original char or metachar */
-         if (min > 0) length += 3 + lastitemlength;
-         }
-       length += lastitemlength + ((max > 0)? 3 : 1);
-       }
-     if (ptr[1] == '?') ptr++;      /* Needs no extra length */
-     POSESSIVE:                     /* Test for possessive quantifier */
-     if (ptr[1] == '+')
-       {
-       ptr++;
-       length += 2 + 2*LINK_SIZE;   /* Allow for atomic brackets */
-       }
-     continue;
-     /* An alternation contains an offset to the next branch or ket. If any ims
-     options changed in the previous branch(es), and/or if we are in a
-     lookbehind assertion, extra space will be needed at the start of the
-     branch. This is handled by branch_extra. */
-     case '|':
-     length += 1 + LINK_SIZE + branch_extra;
-     continue;
-     /* A character class uses 33 characters provided that all the character
-     values are less than 256. Otherwise, it uses a bit map for low valued
-     characters, and individual items for others. Don't worry about character
-     types that aren't allowed in classes - they'll get picked up during the
-     compile. A character class that contains only one single-byte character
-     uses 2 or 3 bytes, depending on whether it is negated or not. Notice this
-     where we can. (In UTF-8 mode we can do this only for chars < 128.) */
-     case '[':
-     if (*(++ptr) == '^')
-       {
-       class_optcount = 10;  /* Greater than one */
-       ptr++;
-       }
-     else class_optcount = 0;
- #ifdef SUPPORT_UTF8
-     class_utf8 = FALSE;
- #endif
-     /* Written as a "do" so that an initial ']' is taken as data */
-     if (*ptr != 0) do
-       {
-       /* Inside \Q...\E everything is literal except \E */
-       if (inescq)
-         {
-         if (*ptr != '\\' || ptr[1] != 'E') goto GET_ONE_CHARACTER;
-         inescq = FALSE;
-         ptr += 1;
-         continue;
-         }
-       /* Outside \Q...\E, check for escapes */
-       if (*ptr == '\\')
-         {
-         c = check_escape(&ptr, &errorcode, bracount, options, TRUE);
-         if (errorcode != 0) goto PCRE_ERROR_RETURN;
-         /* \b is backspace inside a class; \X is literal */
-         if (-c == ESC_b) c = '\b';
-         else if (-c == ESC_X) c = 'X';
-         /* \Q enters quoting mode */
-         else if (-c == ESC_Q)
-           {
-           inescq = TRUE;
-           continue;
-           }
-         /* Handle escapes that turn into characters */
-         if (c >= 0) goto NON_SPECIAL_CHARACTER;
-         /* Escapes that are meta-things. The normal ones just affect the
-         bit map, but Unicode properties require an XCLASS extended item. */
-         else
-           {
-           class_optcount = 10;         /* \d, \s etc; make sure > 1 */
- #ifdef SUPPORT_UTF8
-           if (-c == ESC_p || -c == ESC_P)
-             {
-             if (!class_utf8)
-               {
-               class_utf8 = TRUE;
-               length += LINK_SIZE + 2;
-               }
-             length += 3;
-             }
- #endif
-           }
-         }
-       /* Check the syntax for POSIX stuff. The bits we actually handle are
-       checked during the real compile phase. */
-       else if (*ptr == '[' &&
-                 (ptr[1] == ':' || ptr[1] == '.' || ptr[1] == '=') &&
-                 check_posix_syntax(ptr, &ptr, cd))
-         {
-         ptr++;
-         class_optcount = 10;    /* Make sure > 1 */
-         }
-       /* Anything else increments the possible optimization count. We have to
-       detect ranges here so that we can compute the number of extra ranges for
-       caseless wide characters when UCP support is available. If there are wide
-       characters, we are going to have to use an XCLASS, even for single
-       characters. */
-       else
-         {
-         int d;
-         GET_ONE_CHARACTER:
- #ifdef SUPPORT_UTF8
-         if (utf8)
-           {
-           int extra = 0;
-           GETCHARLEN(c, ptr, extra);
-           ptr += extra;
-           }
-         else c = *ptr;
- #else
-         c = *ptr;
- #endif
-         /* Come here from handling \ above when it escapes to a char value */
-         NON_SPECIAL_CHARACTER:
-         class_optcount++;
-         d = -1;
-         if (ptr[1] == '-')
-           {
-           uschar const *hyptr = ptr++;
-           if (ptr[1] == '\\')
-             {
-             ptr++;
-             d = check_escape(&ptr, &errorcode, bracount, options, TRUE);
-             if (errorcode != 0) goto PCRE_ERROR_RETURN;
-             if (-d == ESC_b) d = '\b';        /* backspace */
-             else if (-d == ESC_X) d = 'X';    /* literal X in a class */
-             }
-           else if (ptr[1] != 0 && ptr[1] != ']')
-             {
-             ptr++;
- #ifdef SUPPORT_UTF8
-             if (utf8)
-               {
-               int extra = 0;
-               GETCHARLEN(d, ptr, extra);
-               ptr += extra;
-               }
-             else
- #endif
-             d = *ptr;
-             }
-           if (d < 0) ptr = hyptr;      /* go back to hyphen as data */
-           }
-         /* If d >= 0 we have a range. In UTF-8 mode, if the end is > 255, or >
-for caseless matching, we will need to use an XCLASS. */
-         if (d >= 0)
-           {
-           class_optcount = 10;     /* Ensure > 1 */
-           if (d < c)
-             {
-             errorcode = ERR8;
-             goto PCRE_ERROR_RETURN;
-             }
- #ifdef SUPPORT_UTF8
-           if (utf8 && (d > 255 || ((options & PCRE_CASELESS) != 0 && d > 127)))
-             {
-             uschar buffer[6];
-             if (!class_utf8)         /* Allow for XCLASS overhead */
-               {
-               class_utf8 = TRUE;
-               length += LINK_SIZE + 2;
-               }
- #ifdef SUPPORT_UCP
-             /* If we have UCP support, find out how many extra ranges are
-             needed to map the other case of characters within this range. We
-             have to mimic the range optimization here, because extending the
-             range upwards might push d over a boundary that makes is use
-             another byte in the UTF-8 representation. */
-             if ((options & PCRE_CASELESS) != 0)
-               {
-               int occ, ocd;
-               int cc = c;
-               int origd = d;
-               while (get_othercase_range(&cc, origd, &occ, &ocd))
-                 {
-                 if (occ >= c && ocd <= d) continue;   /* Skip embedded */
-                 if (occ < c  && ocd >= c - 1)  /* Extend the basic range */
-                   {                            /* if there is overlap,   */
-                   c = occ;                     /* noting that if occ < c */
-                   continue;                    /* we can't have ocd > d  */
-                   }                            /* because a subrange is  */
-                 if (ocd > d && occ <= d + 1)   /* always shorter than    */
-                   {                            /* the basic range.       */
-                   d = ocd;
-                   continue;
-                   }
-                 /* An extra item is needed */
-                 length += 1 + _pcre_ord2utf8(occ, buffer) +
-                   ((occ == ocd)? 0 : _pcre_ord2utf8(ocd, buffer));
-                 }
-               }
  #endif  /* SUPPORT_UCP *