/*
** 2012-11-13
**
** The author disclaims copyright to this source code.  In place of
** a legal notice, here is a blessing:
**
**    May you do good and not evil.
**    May you find forgiveness for yourself and forgive others.
**    May you share freely, never taking more than you give.
**
******************************************************************************
**
** The code in this file implements a compact but reasonably
** efficient regular-expression matcher for posix extended regular
** expressions against UTF8 text.
**
** This file is an SQLite extension.  It registers a single function
** named "regexp(A,B)" where A is the regular expression and B is the
** string to be matched.  By registering this function, SQLite will also
** then implement the "B regexp A" operator.  Note that with the function
** the regular expression comes first, but with the operator it comes
** second.
**
**  The following regular expression syntax is supported:
**
**     X*      zero or more occurrences of X
**     X+      one or more occurrences of X
**     X?      zero or one occurrences of X
**     X{p,q}  between p and q occurrences of X
**     (X)     match X
**     X|Y     X or Y
**     ^X      X occurring at the beginning of the string
**     X$      X occurring at the end of the string
**     .       Match any single character
**     \c      Character c where c is one of \{}()[]|*+?.
**     \c      C-language escapes for c in afnrtv.  ex: \t or \n
**     \uXXXX  Where XXXX is exactly 4 hex digits, unicode value XXXX
**     \xXX    Where XX is exactly 2 hex digits, unicode value XX
**     [abc]   Any single character from the set abc
**     [^abc]  Any single character not in the set abc
**     [a-z]   Any single character in the range a-z
**     [^a-z]  Any single character not in the range a-z
**     \b      Word boundary
**     \w      Word character.  [A-Za-z0-9_]
**     \W      Non-word character
**     \d      Digit
**     \D      Non-digit
**     \s      Whitespace character
**     \S      Non-whitespace character
**
** A nondeterministic finite automaton (NFA) is used for matching, so the
** performance is bounded by O(N*M) where N is the size of the regular
** expression and M is the size of the input string.  The matcher never
** exhibits exponential behavior.  Note that the X{p,q} operator expands
** to p copies of X following by q-p copies of X? and that the size of the
** regular expression in the O(N*M) performance bound is computed after
** this expansion.
*/
#include <string.h>
#include <stdlib.h>
#include "sqlite3ext.h"
SQLITE_EXTENSION_INIT1

/*
** The following #defines change the names of some functions implemented in
** this file to prevent name collisions with C-library functions of the
** same name.
*/
#define re_match   sqlite3re_match
#define re_compile sqlite3re_compile
#define re_free    sqlite3re_free

/* The end-of-input character */
#define RE_EOF            0    /* End of input */

/* The NFA is implemented as sequence of opcodes taken from the following
** set.  Each opcode has a single integer argument.
*/
#define RE_OP_MATCH       1    /* Match the one character in the argument */
#define RE_OP_ANY         2    /* Match any one character.  (Implements ".") */
#define RE_OP_ANYSTAR     3    /* Special optimized version of .* */
#define RE_OP_FORK        4    /* Continue to both next and opcode at iArg */
#define RE_OP_GOTO        5    /* Jump to opcode at iArg */
#define RE_OP_ACCEPT      6    /* Halt and indicate a successful match */
#define RE_OP_CC_INC      7    /* Beginning of a [...] character class */
#define RE_OP_CC_EXC      8    /* Beginning of a [^...] character class */
#define RE_OP_CC_VALUE    9    /* Single value in a character class */
#define RE_OP_CC_RANGE   10    /* Range of values in a character class */
#define RE_OP_WORD       11    /* Perl word character [A-Za-z0-9_] */
#define RE_OP_NOTWORD    12    /* Not a perl word character */
#define RE_OP_DIGIT      13    /* digit:  [0-9] */
#define RE_OP_NOTDIGIT   14    /* Not a digit */
#define RE_OP_SPACE      15    /* space:  [ \t\n\r\v\f] */
#define RE_OP_NOTSPACE   16    /* Not a digit */
#define RE_OP_BOUNDARY   17    /* Boundary between word and non-word */

/* Each opcode is a "state" in the NFA */
typedef unsigned short ReStateNumber;

/* Because this is an NFA and not a DFA, multiple states can be active at
** once.  An instance of the following object records all active states in
** the NFA.  The implementation is optimized for the common case where the
** number of actives states is small.
*/
typedef struct ReStateSet {
  unsigned nState;            /* Number of current states */
  ReStateNumber *aState;      /* Current states */
} ReStateSet;

/* An input string read one character at a time.
*/
typedef struct ReInput ReInput;
struct ReInput {
  const unsigned char *z;  /* All text */
  int i;                   /* Next byte to read */
  int mx;                  /* EOF when i>=mx */
};

/* A compiled NFA (or an NFA that is in the process of being compiled) is
** an instance of the following object.
*/
typedef struct ReCompiled ReCompiled;
struct ReCompiled {
  ReInput sIn;                /* Regular expression text */
  const char *zErr;           /* Error message to return */
  char *aOp;                  /* Operators for the virtual machine */
  int *aArg;                  /* Arguments to each operator */
  unsigned (*xNextChar)(ReInput*);  /* Next character function */
  unsigned char zInit[12];    /* Initial text to match */
  int nInit;                  /* Number of characters in zInit */
  unsigned nState;            /* Number of entries in aOp[] and aArg[] */
  unsigned nAlloc;            /* Slots allocated for aOp[] and aArg[] */
};

/* Add a state to the given state set if it is not already there */
static void re_add_state(ReStateSet *pSet, int newState){
  unsigned i;
  for(i=0; i<pSet->nState; i++) if( pSet->aState[i]==newState ) return;
  pSet->aState[pSet->nState++] = (ReStateNumber)newState;
}

/* Extract the next unicode character from *pzIn and return it.  Advance
** *pzIn to the first byte past the end of the character returned.  To
** be clear:  this routine converts utf8 to unicode.  This routine is 
** optimized for the common case where the next character is a single byte.
*/
static unsigned re_next_char(ReInput *p){
  unsigned c;
  if( p->i>=p->mx ) return 0;
  c = p->z[p->i++];
  if( c>=0x80 ){
    if( (c&0xe0)==0xc0 && p->i<p->mx && (p->z[p->i]&0xc0)==0x80 ){
      c = (c&0x1f)<<6 | (p->z[p->i++]&0x3f);
      if( c<0x80 ) c = 0xfffd;
    }else if( (c&0xf0)==0xe0 && p->i+1<p->mx && (p->z[p->i]&0xc0)==0x80
           && (p->z[p->i+1]&0xc0)==0x80 ){
      c = (c&0x0f)<<12 | ((p->z[p->i]&0x3f)<<6) | (p->z[p->i+1]&0x3f);
      p->i += 2;
      if( c<=0x7ff || (c>=0xd800 && c<=0xdfff) ) c = 0xfffd;
    }else if( (c&0xf8)==0xf0 && p->i+3<p->mx && (p->z[p->i]&0xc0)==0x80
           && (p->z[p->i+1]&0xc0)==0x80 && (p->z[p->i+2]&0xc0)==0x80 ){
      c = (c&0x07)<<18 | ((p->z[p->i]&0x3f)<<12) | ((p->z[p->i+1]&0x3f)<<6)
                       | (p->z[p->i+2]&0x3f);
      p->i += 3;
      if( c<=0xffff || c>0x10ffff ) c = 0xfffd;
    }else{
      c = 0xfffd;
    }
  }
  return c;
}
static unsigned re_next_char_nocase(ReInput *p){
  unsigned c = re_next_char(p);
  if( c>='A' && c<='Z' ) c += 'a' - 'A';
  return c;
}

/* Return true if c is a perl "word" character:  [A-Za-z0-9_] */
static int re_word_char(int c){
  return (c>='0' && c<='9') || (c>='a' && c<='z')
      || (c>='A' && c<='Z') || c=='_';
}

/* Return true if c is a "digit" character:  [0-9] */
static int re_digit_char(int c){
  return (c>='0' && c<='9');
}

/* Return true if c is a perl "space" character:  [ \t\r\n\v\f] */
static int re_space_char(int c){
  return c==' ' || c=='\t' || c=='\n' || c=='\r' || c=='\v' || c=='\f';
}

/* Run a compiled regular expression on the zero-terminated input
** string zIn[].  Return true on a match and false if there is no match.
*/
static int re_match(ReCompiled *pRe, const unsigned char *zIn, int nIn){
  ReStateSet aStateSet[2], *pThis, *pNext;
  ReStateNumber aSpace[100];
  ReStateNumber *pToFree;
  unsigned int i = 0;
  unsigned int iSwap = 0;
  int c = RE_EOF+1;
  int cPrev = 0;
  int rc = 0;
  ReInput in;

  in.z = zIn;
  in.i = 0;
  in.mx = nIn>=0 ? nIn : (int)strlen((char const*)zIn);

  /* Look for the initial prefix match, if there is one. */
  if( pRe->nInit ){
    unsigned char x = pRe->zInit[0];
    while( in.i+pRe->nInit<=in.mx 
     && (zIn[in.i]!=x ||
         strncmp((const char*)zIn+in.i, (const char*)pRe->zInit, pRe->nInit)!=0)
    ){
      in.i++;
    }
    if( in.i+pRe->nInit>in.mx ) return 0;
  }

  if( pRe->nState<=(sizeof(aSpace)/(sizeof(aSpace[0])*2)) ){
    pToFree = 0;
    aStateSet[0].aState = aSpace;
  }else{
    pToFree = sqlite3_malloc64( sizeof(ReStateNumber)*2*pRe->nState );
    if( pToFree==0 ) return -1;
    aStateSet[0].aState = pToFree;
  }
  aStateSet[1].aState = &aStateSet[0].aState[pRe->nState];
  pNext = &aStateSet[1];
  pNext->nState = 0;
  re_add_state(pNext, 0);
  while( c!=RE_EOF && pNext->nState>0 ){
    cPrev = c;
    c = pRe->xNextChar(&in);
    pThis = pNext;
    pNext = &aStateSet[iSwap];
    iSwap = 1 - iSwap;
    pNext->nState = 0;
    for(i=0; i<pThis->nState; i++){
      int x = pThis->aState[i];
      switch( pRe->aOp[x] ){
        case RE_OP_MATCH: {
          if( pRe->aArg[x]==c ) re_add_state(pNext, x+1);
          break;
        }
        case RE_OP_ANY: {
          if( c!=0 ) re_add_state(pNext, x+1);
          break;
        }
        case RE_OP_WORD: {
          if( re_word_char(c) ) re_add_state(pNext, x+1);
          break;
        }
        case RE_OP_NOTWORD: {
          if( !re_word_char(c) && c!=0 ) re_add_state(pNext, x+1);
          break;
        }
        case RE_OP_DIGIT: {
          if( re_digit_char(c) ) re_add_state(pNext, x+1);
          break;
        }
        case RE_OP_NOTDIGIT: {
          if( !re_digit_char(c) && c!=0 ) re_add_state(pNext, x+1);
          break;
        }
        case RE_OP_SPACE: {
          if( re_space_char(c) ) re_add_state(pNext, x+1);
          break;
        }
        case RE_OP_NOTSPACE: {
          if( !re_space_char(c) && c!=0 ) re_add_state(pNext, x+1);
          break;
        }
        case RE_OP_BOUNDARY: {
          if( re_word_char(c)!=re_word_char(cPrev) ) re_add_state(pThis, x+1);
          break;
        }
        case RE_OP_ANYSTAR: {
          re_add_state(pNext, x);
          re_add_state(pThis, x+1);
          break;
        }
        case RE_OP_FORK: {
          re_add_state(pThis, x+pRe->aArg[x]);
          re_add_state(pThis, x+1);
          break;
        }
        case RE_OP_GOTO: {
          re_add_state(pThis, x+pRe->aArg[x]);
          break;
        }
        case RE_OP_ACCEPT: {
          rc = 1;
          goto re_match_end;
        }
        case RE_OP_CC_EXC: {
          if( c==0 ) break;
          /* fall-through */ goto re_op_cc_inc;
        }
        case RE_OP_CC_INC: re_op_cc_inc: {
          int j = 1;
          int n = pRe->aArg[x];
          int hit = 0;
          for(j=1; j>0 && j<n; j++){
            if( pRe->aOp[x+j]==RE_OP_CC_VALUE ){
              if( pRe->aArg[x+j]==c ){
                hit = 1;
                j = -1;
              }
            }else{
              if( pRe->aArg[x+j]<=c && pRe->aArg[x+j+1]>=c ){
                hit = 1;
                j = -1;
              }else{
                j++;
              }
            }
          }
          if( pRe->aOp[x]==RE_OP_CC_EXC ) hit = !hit;
          if( hit ) re_add_state(pNext, x+n);
          break;
        }
      }
    }
  }
  for(i=0; i<pNext->nState; i++){
    if( pRe->aOp[pNext->aState[i]]==RE_OP_ACCEPT ){ rc = 1; break; }
  }
re_match_end:
  sqlite3_free(pToFree);
  return rc;
}

/* Resize the opcode and argument arrays for an RE under construction.
*/
static int re_resize(ReCompiled *p, int N){
  char *aOp;
  int *aArg;
  aOp = sqlite3_realloc64(p->aOp, N*sizeof(p->aOp[0]));
  if( aOp==0 ) return 1;
  p->aOp = aOp;
  aArg = sqlite3_realloc64(p->aArg, N*sizeof(p->aArg[0]));
  if( aArg==0 ) return 1;
  p->aArg = aArg;
  p->nAlloc = N;
  return 0;
}

/* Insert a new opcode and argument into an RE under construction.  The
** insertion point is just prior to existing opcode iBefore.
*/
static int re_insert(ReCompiled *p, int iBefore, int op, int arg){
  int i;
  if( p->nAlloc<=p->nState && re_resize(p, p->nAlloc*2) ) return 0;
  for(i=p->nState; i>iBefore; i--){
    p->aOp[i] = p->aOp[i-1];
    p->aArg[i] = p->aArg[i-1];
  }
  p->nState++;
  p->aOp[iBefore] = (char)op;
  p->aArg[iBefore] = arg;
  return iBefore;
}

/* Append a new opcode and argument to the end of the RE under construction.
*/
static int re_append(ReCompiled *p, int op, int arg){
  return re_insert(p, p->nState, op, arg);
}

/* Make a copy of N opcodes starting at iStart onto the end of the RE
** under construction.
*/
static void re_copy(ReCompiled *p, int iStart, int N){
  if( p->nState+N>=p->nAlloc && re_resize(p, p->nAlloc*2+N) ) return;
  memcpy(&p->aOp[p->nState], &p->aOp[iStart], N*sizeof(p->aOp[0]));
  memcpy(&p->aArg[p->nState], &p->aArg[iStart], N*sizeof(p->aArg[0]));
  p->nState += N;
}

/* Return true if c is a hexadecimal digit character:  [0-9a-fA-F]
** If c is a hex digit, also set *pV = (*pV)*16 + valueof(c).  If
** c is not a hex digit *pV is unchanged.
*/
static int re_hex(int c, int *pV){
  if( c>='0' && c<='9' ){
    c -= '0';
  }else if( c>='a' && c<='f' ){
    c -= 'a' - 10;
  }else if( c>='A' && c<='F' ){
    c -= 'A' - 10;
  }else{
    return 0;
  }
  *pV = (*pV)*16 + (c & 0xff);
  return 1;
}

/* A backslash character has been seen, read the next character and
** return its interpretation.
*/
static unsigned re_esc_char(ReCompiled *p){
  static const char zEsc[] = "afnrtv\\()*.+?[$^{|}]";
  static const char zTrans[] = "\a\f\n\r\t\v";
  int i, v = 0;
  char c;
  if( p->sIn.i>=p->sIn.mx ) return 0;
  c = p->sIn.z[p->sIn.i];
  if( c=='u' && p->sIn.i+4<p->sIn.mx ){
    const unsigned char *zIn = p->sIn.z + p->sIn.i;
    if( re_hex(zIn[1],&v)
     && re_hex(zIn[2],&v)
     && re_hex(zIn[3],&v)
     && re_hex(zIn[4],&v)
    ){
      p->sIn.i += 5;
      return v;
    }
  }
  if( c=='x' && p->sIn.i+2<p->sIn.mx ){
    const unsigned char *zIn = p->sIn.z + p->sIn.i;
    if( re_hex(zIn[1],&v)
     && re_hex(zIn[2],&v)
    ){
      p->sIn.i += 3;
      return v;
    }
  }
  for(i=0; zEsc[i] && zEsc[i]!=c; i++){}
  if( zEsc[i] ){
    if( i<6 ) c = zTrans[i];
    p->sIn.i++;
  }else{
    p->zErr = "unknown \\ escape";
  }
  return c;
}

/* Forward declaration */
static const char *re_subcompile_string(ReCompiled*);

/* Peek at the next byte of input */
static unsigned char rePeek(ReCompiled *p){
  return p->sIn.i<p->sIn.mx ? p->sIn.z[p->sIn.i] : 0;
}

/* Compile RE text into a sequence of opcodes.  Continue up to the
** first unmatched ")" character, then return.  If an error is found,
** return a pointer to the error message string.
*/
static const char *re_subcompile_re(ReCompiled *p){
  const char *zErr;
  int iStart, iEnd, iGoto;
  iStart = p->nState;
  zErr = re_subcompile_string(p);
  if( zErr ) return zErr;
  while( rePeek(p)=='|' ){
    iEnd = p->nState;
    re_insert(p, iStart, RE_OP_FORK, iEnd + 2 - iStart);
    iGoto = re_append(p, RE_OP_GOTO, 0);
    p->sIn.i++;
    zErr = re_subcompile_string(p);
    if( zErr ) return zErr;
    p->aArg[iGoto] = p->nState - iGoto;
  }
  return 0;
}

/* Compile an element of regular expression text (anything that can be
** an operand to the "|" operator).  Return NULL on success or a pointer
** to the error message if there is a problem.
*/
static const char *re_subcompile_string(ReCompiled *p){
  int iPrev = -1;
  int iStart;
  unsigned c;
  const char *zErr;
  while( (c = p->xNextChar(&p->sIn))!=0 ){
    iStart = p->nState;
    switch( c ){
      case '|':
      case '$':
      case ')': {
        p->sIn.i--;
        return 0;
      }
      case '(': {
        zErr = re_subcompile_re(p);
        if( zErr ) return zErr;
        if( rePeek(p)!=')' ) return "unmatched '('";
        p->sIn.i++;
        break;
      }
      case '.': {
        if( rePeek(p)=='*' ){
          re_append(p, RE_OP_ANYSTAR, 0);
          p->sIn.i++;
        }else{
          re_append(p, RE_OP_ANY, 0);
        }
        break;
      }
      case '*': {
        if( iPrev<0 ) return "'*' without operand";
        re_insert(p, iPrev, RE_OP_GOTO, p->nState - iPrev + 1);
        re_append(p, RE_OP_FORK, iPrev - p->nState + 1);
        break;
      }
      case '+': {
        if( iPrev<0 ) return "'+' without operand";
        re_append(p, RE_OP_FORK, iPrev - p->nState);
        break;
      }
      case '?': {
        if( iPrev<0 ) return "'?' without operand";
        re_insert(p, iPrev, RE_OP_FORK, p->nState - iPrev+1);
        break;
      }
      case '{': {
        int m = 0, n = 0;
        int sz, j;
        if( iPrev<0 ) return "'{m,n}' without operand";
        while( (c=rePeek(p))>='0' && c<='9' ){ m = m*10 + c - '0'; p->sIn.i++; }
        n = m;
        if( c==',' ){
          p->sIn.i++;
          n = 0;
          while( (c=rePeek(p))>='0' && c<='9' ){ n = n*10 + c-'0'; p->sIn.i++; }
        }
        if( c!='}' ) return "unmatched '{'";
        if( n>0 && n<m ) return "n less than m in '{m,n}'";
        p->sIn.i++;
        sz = p->nState - iPrev;
        if( m==0 ){
          if( n==0 ) return "both m and n are zero in '{m,n}'";
          re_insert(p, iPrev, RE_OP_FORK, sz+1);
          n--;
        }else{
          for(j=1; j<m; j++) re_copy(p, iPrev, sz);
        }
        for(j=m; j<n; j++){
          re_append(p, RE_OP_FORK, sz+1);
          re_copy(p, iPrev, sz);
        }
        if( n==0 && m>0 ){
          re_append(p, RE_OP_FORK, -sz);
        }
        break;
      }
      case '[': {
        int iFirst = p->nState;
        if( rePeek(p)=='^' ){
          re_append(p, RE_OP_CC_EXC, 0);
          p->sIn.i++;
        }else{
          re_append(p, RE_OP_CC_INC, 0);
        }
        while( (c = p->xNextChar(&p->sIn))!=0 ){
          if( c=='[' && rePeek(p)==':' ){
            return "POSIX character classes not supported";
          }
          if( c=='\\' ) c = re_esc_char(p);
          if( rePeek(p)=='-' ){
            re_append(p, RE_OP_CC_RANGE, c);
            p->sIn.i++;
            c = p->xNextChar(&p->sIn);
            if( c=='\\' ) c = re_esc_char(p);
            re_append(p, RE_OP_CC_RANGE, c);
          }else{
            re_append(p, RE_OP_CC_VALUE, c);
          }
          if( rePeek(p)==']' ){ p->sIn.i++; break; }
        }
        if( c==0 ) return "unclosed '['";
        p->aArg[iFirst] = p->nState - iFirst;
        break;
      }
      case '\\': {
        int specialOp = 0;
        switch( rePeek(p) ){
          case 'b': specialOp = RE_OP_BOUNDARY;   break;
          case 'd': specialOp = RE_OP_DIGIT;      break;
          case 'D': specialOp = RE_OP_NOTDIGIT;   break;
          case 's': specialOp = RE_OP_SPACE;      break;
          case 'S': specialOp = RE_OP_NOTSPACE;   break;
          case 'w': specialOp = RE_OP_WORD;       break;
          case 'W': specialOp = RE_OP_NOTWORD;    break;
        }
        if( specialOp ){
          p->sIn.i++;
          re_append(p, specialOp, 0);
        }else{
          c = re_esc_char(p);
          re_append(p, RE_OP_MATCH, c);
        }
        break;
      }
      default: {
        re_append(p, RE_OP_MATCH, c);
        break;
      }
    }
    iPrev = iStart;
  }
  return 0;
}

/* Free and reclaim all the memory used by a previously compiled
** regular expression.  Applications should invoke this routine once
** for every call to re_compile() to avoid memory leaks.
*/
static void re_free(ReCompiled *pRe){
  if( pRe ){
    sqlite3_free(pRe->aOp);
    sqlite3_free(pRe->aArg);
    sqlite3_free(pRe);
  }
}

/*
** Compile a textual regular expression in zIn[] into a compiled regular
** expression suitable for us by re_match() and return a pointer to the
** compiled regular expression in *ppRe.  Return NULL on success or an
** error message if something goes wrong.
*/
static const char *re_compile(ReCompiled **ppRe, const char *zIn, int noCase){
  ReCompiled *pRe;
  const char *zErr;
  int i, j;

  *ppRe = 0;
  pRe = sqlite3_malloc( sizeof(*pRe) );
  if( pRe==0 ){
    return "out of memory";
  }
  memset(pRe, 0, sizeof(*pRe));
  pRe->xNextChar = noCase ? re_next_char_nocase : re_next_char;
  if( re_resize(pRe, 30) ){
    re_free(pRe);
    return "out of memory";
  }
  if( zIn[0]=='^' ){
    zIn++;
  }else{
    re_append(pRe, RE_OP_ANYSTAR, 0);
  }
  pRe->sIn.z = (unsigned char*)zIn;
  pRe->sIn.i = 0;
  pRe->sIn.mx = (int)strlen(zIn);
  zErr = re_subcompile_re(pRe);
  if( zErr ){
    re_free(pRe);
    return zErr;
  }
  if( rePeek(pRe)=='$' && pRe->sIn.i+1>=pRe->sIn.mx ){
    re_append(pRe, RE_OP_MATCH, RE_EOF);
    re_append(pRe, RE_OP_ACCEPT, 0);
    *ppRe = pRe;
  }else if( pRe->sIn.i>=pRe->sIn.mx ){
    re_append(pRe, RE_OP_ACCEPT, 0);
    *ppRe = pRe;
  }else{
    re_free(pRe);
    return "unrecognized character";
  }

  /* The following is a performance optimization.  If the regex begins with
  ** ".*" (if the input regex lacks an initial "^") and afterwards there are
  ** one or more matching characters, enter those matching characters into
  ** zInit[].  The re_match() routine can then search ahead in the input 
  ** string looking for the initial match without having to run the whole
  ** regex engine over the string.  Do not worry able trying to match
  ** unicode characters beyond plane 0 - those are very rare and this is
  ** just an optimization. */
  if( pRe->aOp[0]==RE_OP_ANYSTAR && !noCase ){
    for(j=0, i=1; j<(int)sizeof(pRe->zInit)-2 && pRe->aOp[i]==RE_OP_MATCH; i++){
      unsigned x = pRe->aArg[i];
      if( x<=127 ){
        pRe->zInit[j++] = (unsigned char)x;
      }else if( x<=0xfff ){
        pRe->zInit[j++] = (unsigned char)(0xc0 | (x>>6));
        pRe->zInit[j++] = 0x80 | (x&0x3f);
      }else if( x<=0xffff ){
        pRe->zInit[j++] = (unsigned char)(0xd0 | (x>>12));
        pRe->zInit[j++] = 0x80 | ((x>>6)&0x3f);
        pRe->zInit[j++] = 0x80 | (x&0x3f);
      }else{
        break;
      }
    }
    if( j>0 && pRe->zInit[j-1]==0 ) j--;
    pRe->nInit = j;
  }
  return pRe->zErr;
}

/*
** Implementation of the regexp() SQL function.  This function implements
** the build-in REGEXP operator.  The first argument to the function is the
** pattern and the second argument is the string.  So, the SQL statements:
**
**       A REGEXP B
**
** is implemented as regexp(B,A).
*/
static void re_sql_func(
  sqlite3_context *context,
  int argc,
  sqlite3_value **argv
){
  ReCompiled *pRe;          /* Compiled regular expression */
  const char *zPattern;     /* The regular expression */
  const unsigned char *zStr;/* String being searched */
  const char *zErr;         /* Compile error message */
  int setAux = 0;           /* True to invoke sqlite3_set_auxdata() */

  (void)argc;  /* Unused */
  pRe = sqlite3_get_auxdata(context, 0);
  if( pRe==0 ){
    zPattern = (const char*)sqlite3_value_text(argv[0]);
    if( zPattern==0 ) return;
    zErr = re_compile(&pRe, zPattern, sqlite3_user_data(context)!=0);
    if( zErr ){
      re_free(pRe);
      sqlite3_result_error(context, zErr, -1);
      return;
    }
    if( pRe==0 ){
      sqlite3_result_error_nomem(context);
      return;
    }
    setAux = 1;
  }
  zStr = (const unsigned char*)sqlite3_value_text(argv[1]);
  if( zStr!=0 ){
    sqlite3_result_int(context, re_match(pRe, zStr, -1));
  }
  if( setAux ){
    sqlite3_set_auxdata(context, 0, pRe, (void(*)(void*))re_free);
  }
}

/*
** Invoke this routine to register the regexp() function with the
** SQLite database connection.
*/
#ifdef _WIN32
__declspec(dllexport)
#endif
int sqlite3_regexp_init(
  sqlite3 *db, 
  char **pzErrMsg, 
  const sqlite3_api_routines *pApi
){
  int rc = SQLITE_OK;
  SQLITE_EXTENSION_INIT2(pApi);
  (void)pzErrMsg;  /* Unused */
  rc = sqlite3_create_function(db, "regexp", 2, 
                            SQLITE_UTF8|SQLITE_INNOCUOUS|SQLITE_DETERMINISTIC,
                            0, re_sql_func, 0, 0);
  if( rc==SQLITE_OK ){
    /* The regexpi(PATTERN,STRING) function is a case-insensitive version
    ** of regexp(PATTERN,STRING). */
    rc = sqlite3_create_function(db, "regexpi", 2,
                            SQLITE_UTF8|SQLITE_INNOCUOUS|SQLITE_DETERMINISTIC,
                            (void*)db, re_sql_func, 0, 0);
  }
  return rc;
}