regexec.c: pull array lookup out of loop

[perl5.git] / regexec.c

/*    regexec.c
 */

/*
 * 	One Ring to rule them all, One Ring to find them
 &
 *     [p.v of _The Lord of the Rings_, opening poem]
 *     [p.50 of _The Lord of the Rings_, I/iii: "The Shadow of the Past"]
 *     [p.254 of _The Lord of the Rings_, II/ii: "The Council of Elrond"]
 */

/* This file contains functions for executing a regular expression.  See
 * also regcomp.c which funnily enough, contains functions for compiling
 * a regular expression.
 *
 * This file is also copied at build time to ext/re/re_exec.c, where
 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
 * This causes the main functions to be compiled under new names and with
 * debugging support added, which makes "use re 'debug'" work.
 */

/* NOTE: this is derived from Henry Spencer's regexp code, and should not
 * confused with the original package (see point 3 below).  Thanks, Henry!
 */

/* Additional note: this code is very heavily munged from Henry's version
 * in places.  In some spots I've traded clarity for efficiency, so don't
 * blame Henry for some of the lack of readability.
 */

/* The names of the functions have been changed from regcomp and
 * regexec to  pregcomp and pregexec in order to avoid conflicts
 * with the POSIX routines of the same names.
*/

#ifdef PERL_EXT_RE_BUILD
#include "re_top.h"
#endif

/*
 * pregcomp and pregexec -- regsub and regerror are not used in perl
 *
 *	Copyright (c) 1986 by University of Toronto.
 *	Written by Henry Spencer.  Not derived from licensed software.
 *
 *	Permission is granted to anyone to use this software for any
 *	purpose on any computer system, and to redistribute it freely,
 *	subject to the following restrictions:
 *
 *	1. The author is not responsible for the consequences of use of
 *		this software, no matter how awful, even if they arise
 *		from defects in it.
 *
 *	2. The origin of this software must not be misrepresented, either
 *		by explicit claim or by omission.
 *
 *	3. Altered versions must be plainly marked as such, and must not
 *		be misrepresented as being the original software.
 *
 ****    Alterations to Henry's code are...
 ****
 ****    Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
 ****    2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
 ****    by Larry Wall and others
 ****
 ****    You may distribute under the terms of either the GNU General Public
 ****    License or the Artistic License, as specified in the README file.
 *
 * Beware that some of this code is subtly aware of the way operator
 * precedence is structured in regular expressions.  Serious changes in
 * regular-expression syntax might require a total rethink.
 */
#include "EXTERN.h"
#define PERL_IN_REGEXEC_C
#include "perl.h"

#ifdef PERL_IN_XSUB_RE
#  include "re_comp.h"
#else
#  include "regcomp.h"
#endif

#define RF_tainted	1		/* tainted information used? */
#define RF_warned	2		/* warned about big count? */

#define RF_utf8		8		/* Pattern contains multibyte chars? */

#define UTF_PATTERN ((PL_reg_flags & RF_utf8) != 0)

#define RS_init		1		/* eval environment created */
#define RS_set		2		/* replsv value is set */

#ifndef STATIC
#define	STATIC	static
#endif

/* Valid for non-utf8 strings only: avoids the reginclass call if there are no
 * complications: i.e., if everything matchable is straight forward in the
 * bitmap */
#define REGINCLASS(prog,p,c)  (ANYOF_FLAGS(p) ? reginclass(prog,p,c,0,0)   \
					      : ANYOF_BITMAP_TEST(p,*(c)))

/*
 * Forwards.
 */

#define CHR_SVLEN(sv) (utf8_target ? sv_len_utf8(sv) : SvCUR(sv))
#define CHR_DIST(a,b) (PL_reg_match_utf8 ? utf8_distance(a,b) : a - b)

#define HOPc(pos,off) \
	(char *)(PL_reg_match_utf8 \
	    ? reghop3((U8*)pos, off, (U8*)(off >= 0 ? PL_regeol : PL_bostr)) \
	    : (U8*)(pos + off))
#define HOPBACKc(pos, off) \
	(char*)(PL_reg_match_utf8\
	    ? reghopmaybe3((U8*)pos, -off, (U8*)PL_bostr) \
	    : (pos - off >= PL_bostr)		\
		? (U8*)pos - off		\
		: NULL)

#define HOP3(pos,off,lim) (PL_reg_match_utf8 ? reghop3((U8*)(pos), off, (U8*)(lim)) : (U8*)(pos + off))
#define HOP3c(pos,off,lim) ((char*)HOP3(pos,off,lim))

/* these are unrolled below in the CCC_TRY_XXX defined */
#define LOAD_UTF8_CHARCLASS(class,str) STMT_START { \
    if (!CAT2(PL_utf8_,class)) { bool ok; ENTER; save_re_context(); ok=CAT2(is_utf8_,class)((const U8*)str); assert(ok); LEAVE; } } STMT_END

/* Doesn't do an assert to verify that is correct */
#define LOAD_UTF8_CHARCLASS_NO_CHECK(class) STMT_START { \
    if (!CAT2(PL_utf8_,class)) { bool ok; ENTER; save_re_context(); ok=CAT2(is_utf8_,class)((const U8*)" "); LEAVE; } } STMT_END

#define LOAD_UTF8_CHARCLASS_ALNUM() LOAD_UTF8_CHARCLASS(alnum,"a")
#define LOAD_UTF8_CHARCLASS_DIGIT() LOAD_UTF8_CHARCLASS(digit,"0")
#define LOAD_UTF8_CHARCLASS_SPACE() LOAD_UTF8_CHARCLASS(space," ")

#define LOAD_UTF8_CHARCLASS_GCB()  /* Grapheme cluster boundaries */        \
	LOAD_UTF8_CHARCLASS(X_begin, " ");                                  \
	LOAD_UTF8_CHARCLASS(X_non_hangul, "A");                             \
	/* These are utf8 constants, and not utf-ebcdic constants, so the   \
	    * assert should likely and hopefully fail on an EBCDIC machine */ \
	LOAD_UTF8_CHARCLASS(X_extend, "\xcc\x80"); /* U+0300 */             \
									    \
	/* No asserts are done for these, in case called on an early        \
	    * Unicode version in which they map to nothing */               \
	LOAD_UTF8_CHARCLASS_NO_CHECK(X_prepend);/* U+0E40 "\xe0\xb9\x80" */ \
	LOAD_UTF8_CHARCLASS_NO_CHECK(X_L);	    /* U+1100 "\xe1\x84\x80" */ \
	LOAD_UTF8_CHARCLASS_NO_CHECK(X_LV);     /* U+AC00 "\xea\xb0\x80" */ \
	LOAD_UTF8_CHARCLASS_NO_CHECK(X_LVT);    /* U+AC01 "\xea\xb0\x81" */ \
	LOAD_UTF8_CHARCLASS_NO_CHECK(X_LV_LVT_V);/* U+AC01 "\xea\xb0\x81" */\
	LOAD_UTF8_CHARCLASS_NO_CHECK(X_T);      /* U+11A8 "\xe1\x86\xa8" */ \
	LOAD_UTF8_CHARCLASS_NO_CHECK(X_V)       /* U+1160 "\xe1\x85\xa0" */  

/* 
   We dont use PERL_LEGACY_UNICODE_CHARCLASS_MAPPINGS as the direct test
   so that it is possible to override the option here without having to 
   rebuild the entire core. as we are required to do if we change regcomp.h
   which is where PERL_LEGACY_UNICODE_CHARCLASS_MAPPINGS is defined.
*/
#if PERL_LEGACY_UNICODE_CHARCLASS_MAPPINGS
#define BROKEN_UNICODE_CHARCLASS_MAPPINGS
#endif

#ifdef BROKEN_UNICODE_CHARCLASS_MAPPINGS
#define LOAD_UTF8_CHARCLASS_PERL_WORD()   LOAD_UTF8_CHARCLASS_ALNUM()
#define LOAD_UTF8_CHARCLASS_PERL_SPACE()  LOAD_UTF8_CHARCLASS_SPACE()
#define LOAD_UTF8_CHARCLASS_POSIX_DIGIT() LOAD_UTF8_CHARCLASS_DIGIT()
#define RE_utf8_perl_word   PL_utf8_alnum
#define RE_utf8_perl_space  PL_utf8_space
#define RE_utf8_posix_digit PL_utf8_digit
#define perl_word  alnum
#define perl_space space
#define posix_digit digit
#else
#define LOAD_UTF8_CHARCLASS_PERL_WORD()   LOAD_UTF8_CHARCLASS(perl_word,"a")
#define LOAD_UTF8_CHARCLASS_PERL_SPACE()  LOAD_UTF8_CHARCLASS(perl_space," ")
#define LOAD_UTF8_CHARCLASS_POSIX_DIGIT() LOAD_UTF8_CHARCLASS(posix_digit,"0")
#define RE_utf8_perl_word   PL_utf8_perl_word
#define RE_utf8_perl_space  PL_utf8_perl_space
#define RE_utf8_posix_digit PL_utf8_posix_digit
#endif


#define _CCC_TRY_AFF_COMMON(NAME,NAMEL,CLASS,STR,LCFUNC_utf8,FUNC)          \
        case NAMEL:                                                         \
            PL_reg_flags |= RF_tainted;                                     \
            /* FALL THROUGH */                                              \
        case NAME:                                                          \
            if (!nextchr)                                                   \
                sayNO;                                                      \
            if (utf8_target && UTF8_IS_CONTINUED(nextchr)) {                \
                if (!CAT2(PL_utf8_,CLASS)) {                                \
                    bool ok;                                                \
                    ENTER;                                                  \
                    save_re_context();                                      \
                    ok=CAT2(is_utf8_,CLASS)((const U8*)STR);                \
                    assert(ok);                                             \
                    LEAVE;                                                  \
                }                                                           \
                if (!(OP(scan) == NAME                                      \
                    ? cBOOL(swash_fetch(CAT2(PL_utf8_,CLASS), (U8*)locinput, utf8_target))  \
                    : LCFUNC_utf8((U8*)locinput)))                          \
                {                                                           \
                    sayNO;                                                  \
                }                                                           \
                locinput += PL_utf8skip[nextchr];                           \
                nextchr = UCHARAT(locinput);                                \
                break;                                                      \
            }                                                               \
	    /* Drops through to the macro that calls this one */

#define CCC_TRY_AFF(NAME,NAMEL,CLASS,STR,LCFUNC_utf8,FUNC,LCFUNC)           \
    _CCC_TRY_AFF_COMMON(NAME,NAMEL,CLASS,STR,LCFUNC_utf8,FUNC)              \
            if (!(OP(scan) == NAME ? FUNC(nextchr) : LCFUNC(nextchr)))      \
                sayNO;                                                      \
            nextchr = UCHARAT(++locinput);                                  \
            break

/* Almost identical to the above, but has a case for a node that matches chars
 * between 128 and 255 using Unicode (latin1) semantics. */
#define CCC_TRY_AFF_U(NAME,NAMEL,CLASS,STR,LCFUNC_utf8,FUNCU,LCFUNC)         \
    _CCC_TRY_AFF_COMMON(NAME,NAMEL,CLASS,STR,LCFUNC_utf8,FUNC)               \
            if (!(OP(scan) == NAMEL ? LCFUNC(nextchr) : (FUNCU(nextchr) && (isASCII(nextchr) || (FLAGS(scan) & USE_UNI))))) \
                sayNO;                                                       \
            nextchr = UCHARAT(++locinput);                                   \
            break

#define _CCC_TRY_NEG_COMMON(NAME,NAMEL,CLASS,STR,LCFUNC_utf8,FUNC)           \
        case NAMEL:                                                          \
            PL_reg_flags |= RF_tainted;                                      \
            /* FALL THROUGH */                                               \
        case NAME :                                                          \
            if (!nextchr && locinput >= PL_regeol)                           \
                sayNO;                                                       \
            if (utf8_target && UTF8_IS_CONTINUED(nextchr)) {                 \
                if (!CAT2(PL_utf8_,CLASS)) {                                 \
                    bool ok;                                                 \
                    ENTER;                                                   \
                    save_re_context();                                       \
                    ok=CAT2(is_utf8_,CLASS)((const U8*)STR);                 \
                    assert(ok);                                              \
                    LEAVE;                                                   \
                }                                                            \
                if ((OP(scan) == NAME                                        \
                    ? cBOOL(swash_fetch(CAT2(PL_utf8_,CLASS), (U8*)locinput, utf8_target))  \
                    : LCFUNC_utf8((U8*)locinput)))                           \
                {                                                            \
                    sayNO;                                                   \
                }                                                            \
                locinput += PL_utf8skip[nextchr];                            \
                nextchr = UCHARAT(locinput);                                 \
                break;                                                       \
            }

#define CCC_TRY_NEG(NAME,NAMEL,CLASS,STR,LCFUNC_utf8,FUNC,LCFUNC)            \
    _CCC_TRY_NEG_COMMON(NAME,NAMEL,CLASS,STR,LCFUNC_utf8,FUNC)               \
            if ((OP(scan) == NAME ? FUNC(nextchr) : LCFUNC(nextchr)))        \
                sayNO;                                                       \
            nextchr = UCHARAT(++locinput);                                   \
            break


#define CCC_TRY_NEG_U(NAME,NAMEL,CLASS,STR,LCFUNC_utf8,FUNCU,LCFUNC)         \
    _CCC_TRY_NEG_COMMON(NAME,NAMEL,CLASS,STR,LCFUNC_utf8,FUNCU)              \
            if ((OP(scan) == NAMEL ? LCFUNC(nextchr) : (FUNCU(nextchr) && (isASCII(nextchr) || (FLAGS(scan) & USE_UNI))))) \
                sayNO;                                                       \
            nextchr = UCHARAT(++locinput);                                   \
            break



/* TODO: Combine JUMPABLE and HAS_TEXT to cache OP(rn) */

/* for use after a quantifier and before an EXACT-like node -- japhy */
/* it would be nice to rework regcomp.sym to generate this stuff. sigh
 *
 * NOTE that *nothing* that affects backtracking should be in here, specifically
 * VERBS must NOT be included. JUMPABLE is used to determine  if we can ignore a
 * node that is in between two EXACT like nodes when ascertaining what the required
 * "follow" character is. This should probably be moved to regex compile time
 * although it may be done at run time beause of the REF possibility - more
 * investigation required. -- demerphq
*/
#define JUMPABLE(rn) (      \
    OP(rn) == OPEN ||       \
    (OP(rn) == CLOSE && (!cur_eval || cur_eval->u.eval.close_paren != ARG(rn))) || \
    OP(rn) == EVAL ||   \
    OP(rn) == SUSPEND || OP(rn) == IFMATCH || \
    OP(rn) == PLUS || OP(rn) == MINMOD || \
    OP(rn) == KEEPS || \
    (PL_regkind[OP(rn)] == CURLY && ARG1(rn) > 0) \
)
#define IS_EXACT(rn) (PL_regkind[OP(rn)] == EXACT)

#define HAS_TEXT(rn) ( IS_EXACT(rn) || PL_regkind[OP(rn)] == REF )

#if 0 
/* Currently these are only used when PL_regkind[OP(rn)] == EXACT so
   we don't need this definition. */
#define IS_TEXT(rn)   ( OP(rn)==EXACT   || OP(rn)==REF   || OP(rn)==NREF   )
#define IS_TEXTF(rn)  ( OP(rn)==EXACTF  || OP(rn)==REFF  || OP(rn)==NREFF  )
#define IS_TEXTFL(rn) ( OP(rn)==EXACTFL || OP(rn)==REFFL || OP(rn)==NREFFL )

#else
/* ... so we use this as its faster. */
#define IS_TEXT(rn)   ( OP(rn)==EXACT   )
#define IS_TEXTF(rn)  ( OP(rn)==EXACTF  )
#define IS_TEXTFL(rn) ( OP(rn)==EXACTFL )

#endif

/*
  Search for mandatory following text node; for lookahead, the text must
  follow but for lookbehind (rn->flags != 0) we skip to the next step.
*/
#define FIND_NEXT_IMPT(rn) STMT_START { \
    while (JUMPABLE(rn)) { \
	const OPCODE type = OP(rn); \
	if (type == SUSPEND || PL_regkind[type] == CURLY) \
	    rn = NEXTOPER(NEXTOPER(rn)); \
	else if (type == PLUS) \
	    rn = NEXTOPER(rn); \
	else if (type == IFMATCH) \
	    rn = (rn->flags == 0) ? NEXTOPER(NEXTOPER(rn)) : rn + ARG(rn); \
	else rn += NEXT_OFF(rn); \
    } \
} STMT_END 


static void restore_pos(pTHX_ void *arg);

#define REGCP_PAREN_ELEMS 4
#define REGCP_OTHER_ELEMS 5
#define REGCP_FRAME_ELEMS 1
/* REGCP_FRAME_ELEMS are not part of the REGCP_OTHER_ELEMS and
 * are needed for the regexp context stack bookkeeping. */

STATIC CHECKPOINT
S_regcppush(pTHX_ I32 parenfloor)
{
    dVAR;
    const int retval = PL_savestack_ix;
    const int paren_elems_to_push = (PL_regsize - parenfloor) * REGCP_PAREN_ELEMS;
    const UV total_elems = paren_elems_to_push + REGCP_OTHER_ELEMS;
    const UV elems_shifted = total_elems << SAVE_TIGHT_SHIFT;
    int p;
    GET_RE_DEBUG_FLAGS_DECL;

    if (paren_elems_to_push < 0)
	Perl_croak(aTHX_ "panic: paren_elems_to_push < 0");

    if ((elems_shifted >> SAVE_TIGHT_SHIFT) != total_elems)
	Perl_croak(aTHX_ "panic: paren_elems_to_push offset %"UVuf
		   " out of range (%lu-%ld)",
		   total_elems, (unsigned long)PL_regsize, (long)parenfloor);

    SSGROW(total_elems + REGCP_FRAME_ELEMS);
    
    for (p = PL_regsize; p > parenfloor; p--) {
/* REGCP_PARENS_ELEMS are pushed per pairs of parentheses. */
	SSPUSHINT(PL_regoffs[p].end);
	SSPUSHINT(PL_regoffs[p].start);
	SSPUSHPTR(PL_reg_start_tmp[p]);
	SSPUSHINT(p);
	DEBUG_BUFFERS_r(PerlIO_printf(Perl_debug_log,
	  "     saving \\%"UVuf" %"IVdf"(%"IVdf")..%"IVdf"\n",
		      (UV)p, (IV)PL_regoffs[p].start,
		      (IV)(PL_reg_start_tmp[p] - PL_bostr),
		      (IV)PL_regoffs[p].end
	));
    }
/* REGCP_OTHER_ELEMS are pushed in any case, parentheses or no. */
    SSPUSHPTR(PL_regoffs);
    SSPUSHINT(PL_regsize);
    SSPUSHINT(*PL_reglastparen);
    SSPUSHINT(*PL_reglastcloseparen);
    SSPUSHPTR(PL_reginput);
    SSPUSHUV(SAVEt_REGCONTEXT | elems_shifted); /* Magic cookie. */

    return retval;
}

/* These are needed since we do not localize EVAL nodes: */
#define REGCP_SET(cp)                                           \
    DEBUG_STATE_r(                                              \
            PerlIO_printf(Perl_debug_log,		        \
	        "  Setting an EVAL scope, savestack=%"IVdf"\n",	\
	        (IV)PL_savestack_ix));                          \
    cp = PL_savestack_ix

#define REGCP_UNWIND(cp)                                        \
    DEBUG_STATE_r(                                              \
        if (cp != PL_savestack_ix) 		                \
    	    PerlIO_printf(Perl_debug_log,		        \
		"  Clearing an EVAL scope, savestack=%"IVdf"..%"IVdf"\n", \
	        (IV)(cp), (IV)PL_savestack_ix));                \
    regcpblow(cp)

STATIC char *
S_regcppop(pTHX_ const regexp *rex)
{
    dVAR;
    UV i;
    char *input;
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_REGCPPOP;

    /* Pop REGCP_OTHER_ELEMS before the parentheses loop starts. */
    i = SSPOPUV;
    assert((i & SAVE_MASK) == SAVEt_REGCONTEXT); /* Check that the magic cookie is there. */
    i >>= SAVE_TIGHT_SHIFT; /* Parentheses elements to pop. */
    input = (char *) SSPOPPTR;
    *PL_reglastcloseparen = SSPOPINT;
    *PL_reglastparen = SSPOPINT;
    PL_regsize = SSPOPINT;
    PL_regoffs=(regexp_paren_pair *) SSPOPPTR;

    i -= REGCP_OTHER_ELEMS;
    /* Now restore the parentheses context. */
    for ( ; i > 0; i -= REGCP_PAREN_ELEMS) {
	I32 tmps;
	U32 paren = (U32)SSPOPINT;
	PL_reg_start_tmp[paren] = (char *) SSPOPPTR;
	PL_regoffs[paren].start = SSPOPINT;
	tmps = SSPOPINT;
	if (paren <= *PL_reglastparen)
	    PL_regoffs[paren].end = tmps;
	DEBUG_BUFFERS_r(
	    PerlIO_printf(Perl_debug_log,
			  "     restoring \\%"UVuf" to %"IVdf"(%"IVdf")..%"IVdf"%s\n",
			  (UV)paren, (IV)PL_regoffs[paren].start,
			  (IV)(PL_reg_start_tmp[paren] - PL_bostr),
			  (IV)PL_regoffs[paren].end,
			  (paren > *PL_reglastparen ? "(no)" : ""));
	);
    }
    DEBUG_BUFFERS_r(
	if (*PL_reglastparen + 1 <= rex->nparens) {
	    PerlIO_printf(Perl_debug_log,
			  "     restoring \\%"IVdf"..\\%"IVdf" to undef\n",
			  (IV)(*PL_reglastparen + 1), (IV)rex->nparens);
	}
    );
#if 1
    /* It would seem that the similar code in regtry()
     * already takes care of this, and in fact it is in
     * a better location to since this code can #if 0-ed out
     * but the code in regtry() is needed or otherwise tests
     * requiring null fields (pat.t#187 and split.t#{13,14}
     * (as of patchlevel 7877)  will fail.  Then again,
     * this code seems to be necessary or otherwise
     * this erroneously leaves $1 defined: "1" =~ /^(?:(\d)x)?\d$/
     * --jhi updated by dapm */
    for (i = *PL_reglastparen + 1; i <= rex->nparens; i++) {
	if (i > PL_regsize)
	    PL_regoffs[i].start = -1;
	PL_regoffs[i].end = -1;
    }
#endif
    return input;
}

#define regcpblow(cp) LEAVE_SCOPE(cp)	/* Ignores regcppush()ed data. */

/*
 * pregexec and friends
 */

#ifndef PERL_IN_XSUB_RE
/*
 - pregexec - match a regexp against a string
 */
I32
Perl_pregexec(pTHX_ REGEXP * const prog, char* stringarg, register char *strend,
	 char *strbeg, I32 minend, SV *screamer, U32 nosave)
/* strend: pointer to null at end of string */
/* strbeg: real beginning of string */
/* minend: end of match must be >=minend after stringarg. */
/* nosave: For optimizations. */
{
    PERL_ARGS_ASSERT_PREGEXEC;

    return
	regexec_flags(prog, stringarg, strend, strbeg, minend, screamer, NULL,
		      nosave ? 0 : REXEC_COPY_STR);
}
#endif

/*
 * Need to implement the following flags for reg_anch:
 *
 * USE_INTUIT_NOML		- Useful to call re_intuit_start() first
 * USE_INTUIT_ML
 * INTUIT_AUTORITATIVE_NOML	- Can trust a positive answer
 * INTUIT_AUTORITATIVE_ML
 * INTUIT_ONCE_NOML		- Intuit can match in one location only.
 * INTUIT_ONCE_ML
 *
 * Another flag for this function: SECOND_TIME (so that float substrs
 * with giant delta may be not rechecked).
 */

/* Assumptions: if ANCH_GPOS, then strpos is anchored. XXXX Check GPOS logic */

/* If SCREAM, then SvPVX_const(sv) should be compatible with strpos and strend.
   Otherwise, only SvCUR(sv) is used to get strbeg. */

/* XXXX We assume that strpos is strbeg unless sv. */

/* XXXX Some places assume that there is a fixed substring.
	An update may be needed if optimizer marks as "INTUITable"
	RExen without fixed substrings.  Similarly, it is assumed that
	lengths of all the strings are no more than minlen, thus they
	cannot come from lookahead.
	(Or minlen should take into account lookahead.) 
  NOTE: Some of this comment is not correct. minlen does now take account
  of lookahead/behind. Further research is required. -- demerphq

*/

/* A failure to find a constant substring means that there is no need to make
   an expensive call to REx engine, thus we celebrate a failure.  Similarly,
   finding a substring too deep into the string means that less calls to
   regtry() should be needed.

   REx compiler's optimizer found 4 possible hints:
	a) Anchored substring;
	b) Fixed substring;
	c) Whether we are anchored (beginning-of-line or \G);
	d) First node (of those at offset 0) which may distingush positions;
   We use a)b)d) and multiline-part of c), and try to find a position in the
   string which does not contradict any of them.
 */

/* Most of decisions we do here should have been done at compile time.
   The nodes of the REx which we used for the search should have been
   deleted from the finite automaton. */

char *
Perl_re_intuit_start(pTHX_ REGEXP * const rx, SV *sv, char *strpos,
		     char *strend, const U32 flags, re_scream_pos_data *data)
{
    dVAR;
    struct regexp *const prog = (struct regexp *)SvANY(rx);
    register I32 start_shift = 0;
    /* Should be nonnegative! */
    register I32 end_shift   = 0;
    register char *s;
    register SV *check;
    char *strbeg;
    char *t;
    const bool utf8_target = (sv && SvUTF8(sv)) ? 1 : 0; /* if no sv we have to assume bytes */
    I32 ml_anch;
    register char *other_last = NULL;	/* other substr checked before this */
    char *check_at = NULL;		/* check substr found at this pos */
    const I32 multiline = prog->extflags & RXf_PMf_MULTILINE;
    RXi_GET_DECL(prog,progi);
#ifdef DEBUGGING
    const char * const i_strpos = strpos;
#endif
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_RE_INTUIT_START;

    RX_MATCH_UTF8_set(rx,utf8_target);

    if (RX_UTF8(rx)) {
	PL_reg_flags |= RF_utf8;
    }
    DEBUG_EXECUTE_r( 
        debug_start_match(rx, utf8_target, strpos, strend,
            sv ? "Guessing start of match in sv for"
               : "Guessing start of match in string for");
	      );

    /* CHR_DIST() would be more correct here but it makes things slow. */
    if (prog->minlen > strend - strpos) {
	DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log,
			      "String too short... [re_intuit_start]\n"));
	goto fail;
    }
                
    strbeg = (sv && SvPOK(sv)) ? strend - SvCUR(sv) : strpos;
    PL_regeol = strend;
    if (utf8_target) {
	if (!prog->check_utf8 && prog->check_substr)
	    to_utf8_substr(prog);
	check = prog->check_utf8;
    } else {
	if (!prog->check_substr && prog->check_utf8)
	    to_byte_substr(prog);
	check = prog->check_substr;
    }
    if (check == &PL_sv_undef) {
	DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log,
		"Non-utf8 string cannot match utf8 check string\n"));
	goto fail;
    }
    if (prog->extflags & RXf_ANCH) {	/* Match at beg-of-str or after \n */
	ml_anch = !( (prog->extflags & RXf_ANCH_SINGLE)
		     || ( (prog->extflags & RXf_ANCH_BOL)
			  && !multiline ) );	/* Check after \n? */

	if (!ml_anch) {
	  if ( !(prog->extflags & RXf_ANCH_GPOS) /* Checked by the caller */
		&& !(prog->intflags & PREGf_IMPLICIT) /* not a real BOL */
	       /* SvCUR is not set on references: SvRV and SvPVX_const overlap */
	       && sv && !SvROK(sv)
	       && (strpos != strbeg)) {
	      DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "Not at start...\n"));
	      goto fail;
	  }
	  if (prog->check_offset_min == prog->check_offset_max &&
	      !(prog->extflags & RXf_CANY_SEEN)) {
	    /* Substring at constant offset from beg-of-str... */
	    I32 slen;

	    s = HOP3c(strpos, prog->check_offset_min, strend);
	    
	    if (SvTAIL(check)) {
		slen = SvCUR(check);	/* >= 1 */

		if ( strend - s > slen || strend - s < slen - 1
		     || (strend - s == slen && strend[-1] != '\n')) {
		    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "String too long...\n"));
		    goto fail_finish;
		}
		/* Now should match s[0..slen-2] */
		slen--;
		if (slen && (*SvPVX_const(check) != *s
			     || (slen > 1
				 && memNE(SvPVX_const(check), s, slen)))) {
		  report_neq:
		    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "String not equal...\n"));
		    goto fail_finish;
		}
	    }
	    else if (*SvPVX_const(check) != *s
		     || ((slen = SvCUR(check)) > 1
			 && memNE(SvPVX_const(check), s, slen)))
		goto report_neq;
	    check_at = s;
	    goto success_at_start;
	  }
	}
	/* Match is anchored, but substr is not anchored wrt beg-of-str. */
	s = strpos;
	start_shift = prog->check_offset_min; /* okay to underestimate on CC */
	end_shift = prog->check_end_shift;
	
	if (!ml_anch) {
	    const I32 end = prog->check_offset_max + CHR_SVLEN(check)
					 - (SvTAIL(check) != 0);
	    const I32 eshift = CHR_DIST((U8*)strend, (U8*)s) - end;

	    if (end_shift < eshift)
		end_shift = eshift;
	}
    }
    else {				/* Can match at random position */
	ml_anch = 0;
	s = strpos;
	start_shift = prog->check_offset_min;  /* okay to underestimate on CC */
	end_shift = prog->check_end_shift;
	
	/* end shift should be non negative here */
    }

#ifdef QDEBUGGING	/* 7/99: reports of failure (with the older version) */
    if (end_shift < 0)
	Perl_croak(aTHX_ "panic: end_shift: %"IVdf" pattern:\n%s\n ",
		   (IV)end_shift, RX_PRECOMP(prog));
#endif

  restart:
    /* Find a possible match in the region s..strend by looking for
       the "check" substring in the region corrected by start/end_shift. */
    
    {
        I32 srch_start_shift = start_shift;
        I32 srch_end_shift = end_shift;
        if (srch_start_shift < 0 && strbeg - s > srch_start_shift) {
	    srch_end_shift -= ((strbeg - s) - srch_start_shift); 
	    srch_start_shift = strbeg - s;
	}
    DEBUG_OPTIMISE_MORE_r({
        PerlIO_printf(Perl_debug_log, "Check offset min: %"IVdf" Start shift: %"IVdf" End shift %"IVdf" Real End Shift: %"IVdf"\n",
            (IV)prog->check_offset_min,
            (IV)srch_start_shift,
            (IV)srch_end_shift, 
            (IV)prog->check_end_shift);
    });       
        
    if (flags & REXEC_SCREAM) {
	I32 p = -1;			/* Internal iterator of scream. */
	I32 * const pp = data ? data->scream_pos : &p;

	if (PL_screamfirst[BmRARE(check)] >= 0
	    || ( BmRARE(check) == '\n'
		 && (BmPREVIOUS(check) == SvCUR(check) - 1)
		 && SvTAIL(check) ))
	    s = screaminstr(sv, check,
			    srch_start_shift + (s - strbeg), srch_end_shift, pp, 0);
	else
	    goto fail_finish;
	/* we may be pointing at the wrong string */
	if (s && RXp_MATCH_COPIED(prog))
	    s = strbeg + (s - SvPVX_const(sv));
	if (data)
	    *data->scream_olds = s;
    }
    else {
        U8* start_point;
        U8* end_point;
        if (prog->extflags & RXf_CANY_SEEN) {
            start_point= (U8*)(s + srch_start_shift);
            end_point= (U8*)(strend - srch_end_shift);
        } else {
	    start_point= HOP3(s, srch_start_shift, srch_start_shift < 0 ? strbeg : strend);
            end_point= HOP3(strend, -srch_end_shift, strbeg);
	}
	DEBUG_OPTIMISE_MORE_r({
            PerlIO_printf(Perl_debug_log, "fbm_instr len=%d str=<%.*s>\n", 
                (int)(end_point - start_point),
                (int)(end_point - start_point) > 20 ? 20 : (int)(end_point - start_point), 
                start_point);
        });

	s = fbm_instr( start_point, end_point,
		      check, multiline ? FBMrf_MULTILINE : 0);
    }
    }
    /* Update the count-of-usability, remove useless subpatterns,
	unshift s.  */

    DEBUG_EXECUTE_r({
        RE_PV_QUOTED_DECL(quoted, utf8_target, PERL_DEBUG_PAD_ZERO(0),
            SvPVX_const(check), RE_SV_DUMPLEN(check), 30);
        PerlIO_printf(Perl_debug_log, "%s %s substr %s%s%s",
			  (s ? "Found" : "Did not find"),
	    (check == (utf8_target ? prog->anchored_utf8 : prog->anchored_substr)
	        ? "anchored" : "floating"),
	    quoted,
	    RE_SV_TAIL(check),
	    (s ? " at offset " : "...\n") ); 
    });

    if (!s)
	goto fail_finish;
    /* Finish the diagnostic message */
    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "%ld...\n", (long)(s - i_strpos)) );

    /* XXX dmq: first branch is for positive lookbehind...
       Our check string is offset from the beginning of the pattern.
       So we need to do any stclass tests offset forward from that 
       point. I think. :-(
     */
    
        
    
    check_at=s;
     

    /* Got a candidate.  Check MBOL anchoring, and the *other* substr.
       Start with the other substr.
       XXXX no SCREAM optimization yet - and a very coarse implementation
       XXXX /ttx+/ results in anchored="ttx", floating="x".  floating will
		*always* match.  Probably should be marked during compile...
       Probably it is right to do no SCREAM here...
     */

    if (utf8_target ? (prog->float_utf8 && prog->anchored_utf8)
                : (prog->float_substr && prog->anchored_substr)) 
    {
	/* Take into account the "other" substring. */
	/* XXXX May be hopelessly wrong for UTF... */
	if (!other_last)
	    other_last = strpos;
	if (check == (utf8_target ? prog->float_utf8 : prog->float_substr)) {
	  do_other_anchored:
	    {
		char * const last = HOP3c(s, -start_shift, strbeg);
		char *last1, *last2;
		char * const saved_s = s;
		SV* must;

		t = s - prog->check_offset_max;
		if (s - strpos > prog->check_offset_max  /* signed-corrected t > strpos */
		    && (!utf8_target
			|| ((t = (char*)reghopmaybe3((U8*)s, -(prog->check_offset_max), (U8*)strpos))
			    && t > strpos)))
		    NOOP;
		else
		    t = strpos;
		t = HOP3c(t, prog->anchored_offset, strend);
		if (t < other_last)	/* These positions already checked */
		    t = other_last;
		last2 = last1 = HOP3c(strend, -prog->minlen, strbeg);
		if (last < last1)
		    last1 = last;
                /* XXXX It is not documented what units *_offsets are in.  
                   We assume bytes, but this is clearly wrong. 
                   Meaning this code needs to be carefully reviewed for errors.
                   dmq.
                  */
 
		/* On end-of-str: see comment below. */
		must = utf8_target ? prog->anchored_utf8 : prog->anchored_substr;
		if (must == &PL_sv_undef) {
		    s = (char*)NULL;
		    DEBUG_r(must = prog->anchored_utf8);	/* for debug */
		}
		else
		    s = fbm_instr(
			(unsigned char*)t,
			HOP3(HOP3(last1, prog->anchored_offset, strend)
				+ SvCUR(must), -(SvTAIL(must)!=0), strbeg),
			must,
			multiline ? FBMrf_MULTILINE : 0
		    );
                DEBUG_EXECUTE_r({
                    RE_PV_QUOTED_DECL(quoted, utf8_target, PERL_DEBUG_PAD_ZERO(0),
                        SvPVX_const(must), RE_SV_DUMPLEN(must), 30);
                    PerlIO_printf(Perl_debug_log, "%s anchored substr %s%s",
			(s ? "Found" : "Contradicts"),
                        quoted, RE_SV_TAIL(must));
                });		    
		
			    
		if (!s) {
		    if (last1 >= last2) {
			DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log,
						", giving up...\n"));
			goto fail_finish;
		    }
		    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log,
			", trying floating at offset %ld...\n",
			(long)(HOP3c(saved_s, 1, strend) - i_strpos)));
		    other_last = HOP3c(last1, prog->anchored_offset+1, strend);
		    s = HOP3c(last, 1, strend);
		    goto restart;
		}
		else {
		    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, " at offset %ld...\n",
			  (long)(s - i_strpos)));
		    t = HOP3c(s, -prog->anchored_offset, strbeg);
		    other_last = HOP3c(s, 1, strend);
		    s = saved_s;
		    if (t == strpos)
			goto try_at_start;
		    goto try_at_offset;
		}
	    }
	}
	else {		/* Take into account the floating substring. */
	    char *last, *last1;
	    char * const saved_s = s;
	    SV* must;

	    t = HOP3c(s, -start_shift, strbeg);
	    last1 = last =
		HOP3c(strend, -prog->minlen + prog->float_min_offset, strbeg);
	    if (CHR_DIST((U8*)last, (U8*)t) > prog->float_max_offset)
		last = HOP3c(t, prog->float_max_offset, strend);
	    s = HOP3c(t, prog->float_min_offset, strend);
	    if (s < other_last)
		s = other_last;
 /* XXXX It is not documented what units *_offsets are in.  Assume bytes.  */
	    must = utf8_target ? prog->float_utf8 : prog->float_substr;
	    /* fbm_instr() takes into account exact value of end-of-str
	       if the check is SvTAIL(ed).  Since false positives are OK,
	       and end-of-str is not later than strend we are OK. */
	    if (must == &PL_sv_undef) {
		s = (char*)NULL;
		DEBUG_r(must = prog->float_utf8);	/* for debug message */
	    }
	    else
		s = fbm_instr((unsigned char*)s,
			      (unsigned char*)last + SvCUR(must)
				  - (SvTAIL(must)!=0),
			      must, multiline ? FBMrf_MULTILINE : 0);
	    DEBUG_EXECUTE_r({
	        RE_PV_QUOTED_DECL(quoted, utf8_target, PERL_DEBUG_PAD_ZERO(0),
	            SvPVX_const(must), RE_SV_DUMPLEN(must), 30);
	        PerlIO_printf(Perl_debug_log, "%s floating substr %s%s",
		    (s ? "Found" : "Contradicts"),
		    quoted, RE_SV_TAIL(must));
            });
	    if (!s) {
		if (last1 == last) {
		    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log,
					    ", giving up...\n"));
		    goto fail_finish;
		}
		DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log,
		    ", trying anchored starting at offset %ld...\n",
		    (long)(saved_s + 1 - i_strpos)));
		other_last = last;
		s = HOP3c(t, 1, strend);
		goto restart;
	    }
	    else {
		DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, " at offset %ld...\n",
		      (long)(s - i_strpos)));
		other_last = s; /* Fix this later. --Hugo */
		s = saved_s;
		if (t == strpos)
		    goto try_at_start;
		goto try_at_offset;
	    }
	}
    }

    
    t= (char*)HOP3( s, -prog->check_offset_max, (prog->check_offset_max<0) ? strend : strpos);
        
    DEBUG_OPTIMISE_MORE_r(
        PerlIO_printf(Perl_debug_log, 
            "Check offset min:%"IVdf" max:%"IVdf" S:%"IVdf" t:%"IVdf" D:%"IVdf" end:%"IVdf"\n",
            (IV)prog->check_offset_min,
            (IV)prog->check_offset_max,
            (IV)(s-strpos),
            (IV)(t-strpos),
            (IV)(t-s),
            (IV)(strend-strpos)
        )
    );

    if (s - strpos > prog->check_offset_max  /* signed-corrected t > strpos */
        && (!utf8_target
	    || ((t = (char*)reghopmaybe3((U8*)s, -prog->check_offset_max, (U8*) ((prog->check_offset_max<0) ? strend : strpos)))
		 && t > strpos))) 
    {
	/* Fixed substring is found far enough so that the match
	   cannot start at strpos. */
      try_at_offset:
	if (ml_anch && t[-1] != '\n') {
	    /* Eventually fbm_*() should handle this, but often
	       anchored_offset is not 0, so this check will not be wasted. */
	    /* XXXX In the code below we prefer to look for "^" even in
	       presence of anchored substrings.  And we search even
	       beyond the found float position.  These pessimizations
	       are historical artefacts only.  */
	  find_anchor:
	    while (t < strend - prog->minlen) {
		if (*t == '\n') {
		    if (t < check_at - prog->check_offset_min) {
			if (utf8_target ? prog->anchored_utf8 : prog->anchored_substr) {
			    /* Since we moved from the found position,
			       we definitely contradict the found anchored
			       substr.  Due to the above check we do not
			       contradict "check" substr.
			       Thus we can arrive here only if check substr
			       is float.  Redo checking for "other"=="fixed".
			     */
			    strpos = t + 1;			
			    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "Found /%s^%s/m at offset %ld, rescanning for anchored from offset %ld...\n",
				PL_colors[0], PL_colors[1], (long)(strpos - i_strpos), (long)(strpos - i_strpos + prog->anchored_offset)));
			    goto do_other_anchored;
			}
			/* We don't contradict the found floating substring. */
			/* XXXX Why not check for STCLASS? */
			s = t + 1;
			DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "Found /%s^%s/m at offset %ld...\n",
			    PL_colors[0], PL_colors[1], (long)(s - i_strpos)));
			goto set_useful;
		    }
		    /* Position contradicts check-string */
		    /* XXXX probably better to look for check-string
		       than for "\n", so one should lower the limit for t? */
		    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "Found /%s^%s/m, restarting lookup for check-string at offset %ld...\n",
			PL_colors[0], PL_colors[1], (long)(t + 1 - i_strpos)));
		    other_last = strpos = s = t + 1;
		    goto restart;
		}
		t++;
	    }
	    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "Did not find /%s^%s/m...\n",
			PL_colors[0], PL_colors[1]));
	    goto fail_finish;
	}
	else {
	    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "Starting position does not contradict /%s^%s/m...\n",
			PL_colors[0], PL_colors[1]));
	}
	s = t;
      set_useful:
	++BmUSEFUL(utf8_target ? prog->check_utf8 : prog->check_substr);	/* hooray/5 */
    }
    else {
	/* The found string does not prohibit matching at strpos,
	   - no optimization of calling REx engine can be performed,
	   unless it was an MBOL and we are not after MBOL,
	   or a future STCLASS check will fail this. */
      try_at_start:
	/* Even in this situation we may use MBOL flag if strpos is offset
	   wrt the start of the string. */
	if (ml_anch && sv && !SvROK(sv)	/* See prev comment on SvROK */
	    && (strpos != strbeg) && strpos[-1] != '\n'
	    /* May be due to an implicit anchor of m{.*foo}  */
	    && !(prog->intflags & PREGf_IMPLICIT))
	{
	    t = strpos;
	    goto find_anchor;
	}
	DEBUG_EXECUTE_r( if (ml_anch)
	    PerlIO_printf(Perl_debug_log, "Position at offset %ld does not contradict /%s^%s/m...\n",
			  (long)(strpos - i_strpos), PL_colors[0], PL_colors[1]);
	);
      success_at_start:
	if (!(prog->intflags & PREGf_NAUGHTY)	/* XXXX If strpos moved? */
	    && (utf8_target ? (
		prog->check_utf8		/* Could be deleted already */
		&& --BmUSEFUL(prog->check_utf8) < 0
		&& (prog->check_utf8 == prog->float_utf8)
	    ) : (
		prog->check_substr		/* Could be deleted already */
		&& --BmUSEFUL(prog->check_substr) < 0
		&& (prog->check_substr == prog->float_substr)
	    )))
	{
	    /* If flags & SOMETHING - do not do it many times on the same match */
	    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "... Disabling check substring...\n"));
	    /* XXX Does the destruction order has to change with utf8_target? */
	    SvREFCNT_dec(utf8_target ? prog->check_utf8 : prog->check_substr);
	    SvREFCNT_dec(utf8_target ? prog->check_substr : prog->check_utf8);
	    prog->check_substr = prog->check_utf8 = NULL;	/* disable */
	    prog->float_substr = prog->float_utf8 = NULL;	/* clear */
	    check = NULL;			/* abort */
	    s = strpos;
	    /* XXXX If the check string was an implicit check MBOL, then we need to unset the relevent flag
		    see http://bugs.activestate.com/show_bug.cgi?id=87173 */
	    if (prog->intflags & PREGf_IMPLICIT)
		prog->extflags &= ~RXf_ANCH_MBOL;
	    /* XXXX This is a remnant of the old implementation.  It
	            looks wasteful, since now INTUIT can use many
	            other heuristics. */
	    prog->extflags &= ~RXf_USE_INTUIT;
	    /* XXXX What other flags might need to be cleared in this branch? */
	}
	else
	    s = strpos;
    }

    /* Last resort... */
    /* XXXX BmUSEFUL already changed, maybe multiple change is meaningful... */
    /* trie stclasses are too expensive to use here, we are better off to
       leave it to regmatch itself */
    if (progi->regstclass && PL_regkind[OP(progi->regstclass)]!=TRIE) {
	/* minlen == 0 is possible if regstclass is \b or \B,
	   and the fixed substr is ''$.
	   Since minlen is already taken into account, s+1 is before strend;
	   accidentally, minlen >= 1 guaranties no false positives at s + 1
	   even for \b or \B.  But (minlen? 1 : 0) below assumes that
	   regstclass does not come from lookahead...  */
	/* If regstclass takes bytelength more than 1: If charlength==1, OK.
	   This leaves EXACTF only, which is dealt with in find_byclass().  */
        const U8* const str = (U8*)STRING(progi->regstclass);
        const int cl_l = (PL_regkind[OP(progi->regstclass)] == EXACT
		    ? CHR_DIST(str+STR_LEN(progi->regstclass), str)
		    : 1);
	char * endpos;
	if (prog->anchored_substr || prog->anchored_utf8 || ml_anch)
            endpos= HOP3c(s, (prog->minlen ? cl_l : 0), strend);
        else if (prog->float_substr || prog->float_utf8)
	    endpos= HOP3c(HOP3c(check_at, -start_shift, strbeg), cl_l, strend);
        else 
            endpos= strend;
		    
        DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "start_shift: %"IVdf" check_at: %"IVdf" s: %"IVdf" endpos: %"IVdf"\n",
				      (IV)start_shift, (IV)(check_at - strbeg), (IV)(s - strbeg), (IV)(endpos - strbeg)));
	
	t = s;
        s = find_byclass(prog, progi->regstclass, s, endpos, NULL);
	if (!s) {
#ifdef DEBUGGING
	    const char *what = NULL;
#endif
	    if (endpos == strend) {
		DEBUG_EXECUTE_r( PerlIO_printf(Perl_debug_log,
				"Could not match STCLASS...\n") );
		goto fail;
	    }
	    DEBUG_EXECUTE_r( PerlIO_printf(Perl_debug_log,
				   "This position contradicts STCLASS...\n") );
	    if ((prog->extflags & RXf_ANCH) && !ml_anch)
		goto fail;
	    /* Contradict one of substrings */
	    if (prog->anchored_substr || prog->anchored_utf8) {
		if ((utf8_target ? prog->anchored_utf8 : prog->anchored_substr) == check) {
		    DEBUG_EXECUTE_r( what = "anchored" );
		  hop_and_restart:
		    s = HOP3c(t, 1, strend);
		    if (s + start_shift + end_shift > strend) {
			/* XXXX Should be taken into account earlier? */
			DEBUG_EXECUTE_r( PerlIO_printf(Perl_debug_log,
					       "Could not match STCLASS...\n") );
			goto fail;
		    }
		    if (!check)
			goto giveup;
		    DEBUG_EXECUTE_r( PerlIO_printf(Perl_debug_log,
				"Looking for %s substr starting at offset %ld...\n",
				 what, (long)(s + start_shift - i_strpos)) );
		    goto restart;
		}
		/* Have both, check_string is floating */
		if (t + start_shift >= check_at) /* Contradicts floating=check */
		    goto retry_floating_check;
		/* Recheck anchored substring, but not floating... */
		s = check_at;
		if (!check)
		    goto giveup;
		DEBUG_EXECUTE_r( PerlIO_printf(Perl_debug_log,
			  "Looking for anchored substr starting at offset %ld...\n",
			  (long)(other_last - i_strpos)) );
		goto do_other_anchored;
	    }
	    /* Another way we could have checked stclass at the
               current position only: */
	    if (ml_anch) {
		s = t = t + 1;
		if (!check)
		    goto giveup;
		DEBUG_EXECUTE_r( PerlIO_printf(Perl_debug_log,
			  "Looking for /%s^%s/m starting at offset %ld...\n",
			  PL_colors[0], PL_colors[1], (long)(t - i_strpos)) );
		goto try_at_offset;
	    }
	    if (!(utf8_target ? prog->float_utf8 : prog->float_substr))	/* Could have been deleted */
		goto fail;
	    /* Check is floating subtring. */
	  retry_floating_check:
	    t = check_at - start_shift;
	    DEBUG_EXECUTE_r( what = "floating" );
	    goto hop_and_restart;
	}
	if (t != s) {
            DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log,
			"By STCLASS: moving %ld --> %ld\n",
                                  (long)(t - i_strpos), (long)(s - i_strpos))
                   );
        }
        else {
            DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log,
                                  "Does not contradict STCLASS...\n"); 
                   );
        }
    }
  giveup:
    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "%s%s:%s match at offset %ld\n",
			  PL_colors[4], (check ? "Guessed" : "Giving up"),
			  PL_colors[5], (long)(s - i_strpos)) );
    return s;

  fail_finish:				/* Substring not found */
    if (prog->check_substr || prog->check_utf8)		/* could be removed already */
	BmUSEFUL(utf8_target ? prog->check_utf8 : prog->check_substr) += 5; /* hooray */
  fail:
    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "%sMatch rejected by optimizer%s\n",
			  PL_colors[4], PL_colors[5]));
    return NULL;
}

#define DECL_TRIE_TYPE(scan) \
    const enum { trie_plain, trie_utf8, trie_utf8_fold, trie_latin_utf8_fold } \
		    trie_type = (scan->flags != EXACT) \
		              ? (utf8_target ? trie_utf8_fold : (UTF_PATTERN ? trie_latin_utf8_fold : trie_plain)) \
                              : (utf8_target ? trie_utf8 : trie_plain)

#define REXEC_TRIE_READ_CHAR(trie_type, trie, widecharmap, uc, uscan, len,  \
uvc, charid, foldlen, foldbuf, uniflags) STMT_START {                       \
    switch (trie_type) {                                                    \
    case trie_utf8_fold:                                                    \
	if ( foldlen>0 ) {                                                  \
	    uvc = utf8n_to_uvuni( uscan, UTF8_MAXLEN, &len, uniflags ); \
	    foldlen -= len;                                                 \
	    uscan += len;                                                   \
	    len=0;                                                          \
	} else {                                                            \
	    uvc = utf8n_to_uvuni( (U8*)uc, UTF8_MAXLEN, &len, uniflags ); \
	    uvc = to_uni_fold( uvc, foldbuf, &foldlen );                    \
	    foldlen -= UNISKIP( uvc );                                      \
	    uscan = foldbuf + UNISKIP( uvc );                               \
	}                                                                   \
	break;                                                              \
    case trie_latin_utf8_fold:                                              \
	if ( foldlen>0 ) {                                                  \
	    uvc = utf8n_to_uvuni( uscan, UTF8_MAXLEN, &len, uniflags );     \
	    foldlen -= len;                                                 \
	    uscan += len;                                                   \
	    len=0;                                                          \
	} else {                                                            \
	    len = 1;                                                        \
	    uvc = to_uni_fold( *(U8*)uc, foldbuf, &foldlen );               \
	    foldlen -= UNISKIP( uvc );                                      \
	    uscan = foldbuf + UNISKIP( uvc );                               \
	}                                                                   \
	break;                                                              \
    case trie_utf8:                                                         \
	uvc = utf8n_to_uvuni( (U8*)uc, UTF8_MAXLEN, &len, uniflags );       \
	break;                                                              \
    case trie_plain:                                                        \
	uvc = (UV)*uc;                                                      \
	len = 1;                                                            \
    }                                                                       \
    if (uvc < 256) {                                                        \
	charid = trie->charmap[ uvc ];                                      \
    }                                                                       \
    else {                                                                  \
	charid = 0;                                                         \
	if (widecharmap) {                                                  \
	    SV** const svpp = hv_fetch(widecharmap,                         \
			(char*)&uvc, sizeof(UV), 0);                        \
	    if (svpp)                                                       \
		charid = (U16)SvIV(*svpp);                                  \
	}                                                                   \
    }                                                                       \
} STMT_END

#define REXEC_FBC_EXACTISH_CHECK(CoNd)                 \
{                                                      \
    char *my_strend= (char *)strend;                   \
    if ( (CoNd)                                        \
	 && (ln == len ||                              \
	     foldEQ_utf8(s, &my_strend, 0,  utf8_target,   \
			m, NULL, ln, cBOOL(UTF_PATTERN)))      \
	 && (!reginfo || regtry(reginfo, &s)) )        \
	goto got_it;                                   \
    else {                                             \
	 U8 foldbuf[UTF8_MAXBYTES_CASE+1];             \
	 uvchr_to_utf8(tmpbuf, c);                     \
	 f = to_utf8_fold(tmpbuf, foldbuf, &foldlen);  \
	 if ( f != c                                   \
	      && (f == c1 || f == c2)                  \
	      && (ln == len ||                         \
	        foldEQ_utf8(s, &my_strend, 0,  utf8_target,\
			      m, NULL, ln, cBOOL(UTF_PATTERN)))\
	      && (!reginfo || regtry(reginfo, &s)) )   \
	      goto got_it;                             \
    }                                                  \
}                                                      \
s += len

#define REXEC_FBC_EXACTISH_SCAN(CoNd)                     \
STMT_START {                                              \
    while (s <= e) {                                      \
	if ( (CoNd)                                       \
	     && (ln == 1 || (OP(c) == EXACTF             \
			      ? foldEQ(s, m, ln)           \
			      : foldEQ_locale(s, m, ln)))  \
	     && (!reginfo || regtry(reginfo, &s)) )        \
	    goto got_it;                                  \
	s++;                                              \
    }                                                     \
} STMT_END

#define REXEC_FBC_UTF8_SCAN(CoDe)                     \
STMT_START {                                          \
    while (s + (uskip = UTF8SKIP(s)) <= strend) {     \
	CoDe                                          \
	s += uskip;                                   \
    }                                                 \
} STMT_END

#define REXEC_FBC_SCAN(CoDe)                          \
STMT_START {                                          \
    while (s < strend) {                              \
	CoDe                                          \
	s++;                                          \
    }                                                 \
} STMT_END

#define REXEC_FBC_UTF8_CLASS_SCAN(CoNd)               \
REXEC_FBC_UTF8_SCAN(                                  \
    if (CoNd) {                                       \
	if (tmp && (!reginfo || regtry(reginfo, &s)))  \
	    goto got_it;                              \
	else                                          \
	    tmp = doevery;                            \
    }                                                 \
    else                                              \
	tmp = 1;                                      \
)

#define REXEC_FBC_CLASS_SCAN(CoNd)                    \
REXEC_FBC_SCAN(                                       \
    if (CoNd) {                                       \
	if (tmp && (!reginfo || regtry(reginfo, &s)))  \
	    goto got_it;                              \
	else                                          \
	    tmp = doevery;                            \
    }                                                 \
    else                                              \
	tmp = 1;                                      \
)

#define REXEC_FBC_TRYIT               \
if ((!reginfo || regtry(reginfo, &s))) \
    goto got_it

#define REXEC_FBC_CSCAN(CoNdUtF8,CoNd)                         \
    if (utf8_target) {                                             \
	REXEC_FBC_UTF8_CLASS_SCAN(CoNdUtF8);                   \
    }                                                          \
    else {                                                     \
	REXEC_FBC_CLASS_SCAN(CoNd);                            \
    }                                                          \
    break
    
#define REXEC_FBC_CSCAN_PRELOAD(UtFpReLoAd,CoNdUtF8,CoNd)      \
    if (utf8_target) {                                             \
	UtFpReLoAd;                                            \
	REXEC_FBC_UTF8_CLASS_SCAN(CoNdUtF8);                   \
    }                                                          \
    else {                                                     \
	REXEC_FBC_CLASS_SCAN(CoNd);                            \
    }                                                          \
    break

#define REXEC_FBC_CSCAN_TAINT(CoNdUtF8,CoNd)                   \
    PL_reg_flags |= RF_tainted;                                \
    if (utf8_target) {                                             \
	REXEC_FBC_UTF8_CLASS_SCAN(CoNdUtF8);                   \
    }                                                          \
    else {                                                     \
	REXEC_FBC_CLASS_SCAN(CoNd);                            \
    }                                                          \
    break

#define DUMP_EXEC_POS(li,s,doutf8) \
    dump_exec_pos(li,s,(PL_regeol),(PL_bostr),(PL_reg_starttry),doutf8)

/* We know what class REx starts with.  Try to find this position... */
/* if reginfo is NULL, its a dryrun */
/* annoyingly all the vars in this routine have different names from their counterparts
   in regmatch. /grrr */

STATIC char *
S_find_byclass(pTHX_ regexp * prog, const regnode *c, char *s, 
    const char *strend, regmatch_info *reginfo)
{
	dVAR;
	const I32 doevery = (prog->intflags & PREGf_SKIP) == 0;
	char *m;
	STRLEN ln;
	STRLEN lnc;
	register STRLEN uskip;
	unsigned int c1;
	unsigned int c2;
	char *e;
	register I32 tmp = 1;	/* Scratch variable? */
	register const bool utf8_target = PL_reg_match_utf8;
        RXi_GET_DECL(prog,progi);

	PERL_ARGS_ASSERT_FIND_BYCLASS;
        
	/* We know what class it must start with. */
	switch (OP(c)) {
	case ANYOF:
	    if (utf8_target) {
		 REXEC_FBC_UTF8_CLASS_SCAN((ANYOF_FLAGS(c) & ANYOF_NONBITMAP) ||
			  !UTF8_IS_INVARIANT((U8)s[0]) ?
			  reginclass(prog, c, (U8*)s, 0, utf8_target) :
			  REGINCLASS(prog, c, (U8*)s));
	    }
	    else {
		 while (s < strend) {
		      STRLEN skip = 1;

		      if (REGINCLASS(prog, c, (U8*)s) ||
			  (ANYOF_FOLD_SHARP_S(c, s, strend) &&
			   /* The assignment of 2 is intentional:
			    * for the folded sharp s, the skip is 2. */
			   (skip = SHARP_S_SKIP))) {
			   if (tmp && (!reginfo || regtry(reginfo, &s)))
				goto got_it;
			   else
				tmp = doevery;
		      }
		      else 
			   tmp = 1;
		      s += skip;
		 }
	    }
	    break;
	case CANY:
	    REXEC_FBC_SCAN(
	        if (tmp && (!reginfo || regtry(reginfo, &s)))
		    goto got_it;
		else
		    tmp = doevery;
	    );
	    break;
	case EXACTF:
	    m   = STRING(c);
	    ln  = STR_LEN(c);	/* length to match in octets/bytes */
	    lnc = (I32) ln;	/* length to match in characters */
	    if (UTF_PATTERN) {
	        STRLEN ulen1, ulen2;
		U8 *sm = (U8 *) m;
		U8 tmpbuf1[UTF8_MAXBYTES_CASE+1];
		U8 tmpbuf2[UTF8_MAXBYTES_CASE+1];
		/* used by commented-out code below */
		/*const U32 uniflags = UTF8_ALLOW_DEFAULT;*/
		
                /* XXX: Since the node will be case folded at compile
                   time this logic is a little odd, although im not 
                   sure that its actually wrong. --dmq */
                   
		c1 = to_utf8_lower((U8*)m, tmpbuf1, &ulen1);
		c2 = to_utf8_upper((U8*)m, tmpbuf2, &ulen2);

		/* XXX: This is kinda strange. to_utf8_XYZ returns the 
                   codepoint of the first character in the converted
                   form, yet originally we did the extra step. 
                   No tests fail by commenting this code out however
                   so Ive left it out. -- dmq.
                   
		c1 = utf8n_to_uvchr(tmpbuf1, UTF8_MAXBYTES_CASE, 
				    0, uniflags);
		c2 = utf8n_to_uvchr(tmpbuf2, UTF8_MAXBYTES_CASE,
				    0, uniflags);
                */
                
		lnc = 0;
		while (sm < ((U8 *) m + ln)) {
		    lnc++;
		    sm += UTF8SKIP(sm);
		}
	    }
	    else {
		c1 = *(U8*)m;
		c2 = PL_fold[c1];
	    }
	    goto do_exactf;
	case EXACTFL:
	    m   = STRING(c);
	    ln  = STR_LEN(c);
	    lnc = (I32) ln;
	    c1 = *(U8*)m;
	    c2 = PL_fold_locale[c1];
	  do_exactf:
	    e = HOP3c(strend, -((I32)lnc), s);

	    if (!reginfo && e < s)
		e = s;			/* Due to minlen logic of intuit() */

	    /* The idea in the EXACTF* cases is to first find the
	     * first character of the EXACTF* node and then, if
	     * necessary, case-insensitively compare the full
	     * text of the node.  The c1 and c2 are the first
	     * characters (though in Unicode it gets a bit
	     * more complicated because there are more cases
	     * than just upper and lower: one needs to use
	     * the so-called folding case for case-insensitive
	     * matching (called "loose matching" in Unicode).
	     * foldEQ_utf8() will do just that. */

	    if (utf8_target || UTF_PATTERN) {
	        UV c, f;
	        U8 tmpbuf [UTF8_MAXBYTES+1];
		STRLEN len = 1;
		STRLEN foldlen;
		const U32 uniflags = UTF8_ALLOW_DEFAULT;
		if (c1 == c2) {
		    /* Upper and lower of 1st char are equal -
		     * probably not a "letter". */
		    while (s <= e) {
		        if (utf8_target) {
		            c = utf8n_to_uvchr((U8*)s, UTF8_MAXBYTES, &len,
					   uniflags);
                        } else {
                            c = *((U8*)s);
                        }					  
			REXEC_FBC_EXACTISH_CHECK(c == c1);
		    }
		}
		else {
		    while (s <= e) {
		        if (utf8_target) {
		            c = utf8n_to_uvchr((U8*)s, UTF8_MAXBYTES, &len,
					   uniflags);
                        } else {
                            c = *((U8*)s);
                        }

			/* Handle some of the three Greek sigmas cases.
			 * Note that not all the possible combinations
			 * are handled here: some of them are handled
			 * by the standard folding rules, and some of
			 * them (the character class or ANYOF cases)
			 * are handled during compiletime in
			 * regexec.c:S_regclass(). */
			if (c == (UV)UNICODE_GREEK_CAPITAL_LETTER_SIGMA ||
			    c == (UV)UNICODE_GREEK_SMALL_LETTER_FINAL_SIGMA)
			    c = (UV)UNICODE_GREEK_SMALL_LETTER_SIGMA;

			REXEC_FBC_EXACTISH_CHECK(c == c1 || c == c2);
		    }
		}
	    }
	    else {
	        /* Neither pattern nor string are UTF8 */
		if (c1 == c2)
		    REXEC_FBC_EXACTISH_SCAN(*(U8*)s == c1);
		else
		    REXEC_FBC_EXACTISH_SCAN(*(U8*)s == c1 || *(U8*)s == c2);
	    }
	    break;
	case BOUNDL:
	    PL_reg_flags |= RF_tainted;
	    /* FALL THROUGH */
	case BOUND:
	    if (utf8_target) {
		if (s == PL_bostr)
		    tmp = '\n';
		else {
		    U8 * const r = reghop3((U8*)s, -1, (U8*)PL_bostr);
		    tmp = utf8n_to_uvchr(r, UTF8SKIP(r), 0, UTF8_ALLOW_DEFAULT);
		}
		tmp = ((OP(c) == BOUND ?
			isALNUM_uni(tmp) : isALNUM_LC_uvchr(UNI_TO_NATIVE(tmp))) != 0);
		LOAD_UTF8_CHARCLASS_ALNUM();
		REXEC_FBC_UTF8_SCAN(
		    if (tmp == !(OP(c) == BOUND ?
				 cBOOL(swash_fetch(PL_utf8_alnum, (U8*)s, utf8_target)) :
				 isALNUM_LC_utf8((U8*)s)))
		    {
			tmp = !tmp;
			REXEC_FBC_TRYIT;
		}
		);
	    }
            else {  /* Not utf8 */
		tmp = (s != PL_bostr) ? UCHARAT(s - 1) : '\n';
                tmp = cBOOL((OP(c) == BOUNDL)
                            ? isALNUM_LC(tmp)
                            : (isWORDCHAR_L1(tmp)
                               && (isASCII(tmp) || (FLAGS(c) & USE_UNI))));
		REXEC_FBC_SCAN(
		    if (tmp ==
                        !((OP(c) == BOUNDL)
                          ? isALNUM_LC(*s)
                          : (isWORDCHAR_L1((U8) *s)
                             && (isASCII((U8) *s) || (FLAGS(c) & USE_UNI)))))
		    {
			tmp = !tmp;
			REXEC_FBC_TRYIT;
		}
		);
	    }
	    if ((!prog->minlen && tmp) && (!reginfo || regtry(reginfo, &s)))
		goto got_it;
	    break;
	case NBOUNDL:
	    PL_reg_flags |= RF_tainted;
	    /* FALL THROUGH */
	case NBOUND:
	    if (utf8_target) {
		if (s == PL_bostr)
		    tmp = '\n';
		else {
		    U8 * const r = reghop3((U8*)s, -1, (U8*)PL_bostr);
		    tmp = utf8n_to_uvchr(r, UTF8SKIP(r), 0, UTF8_ALLOW_DEFAULT);
		}
		tmp = ((OP(c) == NBOUND ?
			isALNUM_uni(tmp) : isALNUM_LC_uvchr(UNI_TO_NATIVE(tmp))) != 0);
		LOAD_UTF8_CHARCLASS_ALNUM();
		REXEC_FBC_UTF8_SCAN(
		    if (tmp == !(OP(c) == NBOUND ?
				 cBOOL(swash_fetch(PL_utf8_alnum, (U8*)s, utf8_target)) :
				 isALNUM_LC_utf8((U8*)s)))
			tmp = !tmp;
		    else REXEC_FBC_TRYIT;
		);
	    }
	    else {
		tmp = (s != PL_bostr) ? UCHARAT(s - 1) : '\n';
                tmp = cBOOL((OP(c) == NBOUNDL)
                            ? isALNUM_LC(tmp)
                            : (isWORDCHAR_L1(tmp)
                               && (isASCII(tmp) || (FLAGS(c) & USE_UNI))));
		REXEC_FBC_SCAN(
		    if (tmp == ! cBOOL(
                            (OP(c) == NBOUNDL)
                            ? isALNUM_LC(*s)
                            : (isWORDCHAR_L1((U8) *s)
                               && (isASCII((U8) *s) || (FLAGS(c) & USE_UNI)))))
                    {
			tmp = !tmp;
                    }
		    else REXEC_FBC_TRYIT;
		);
	    }
	    if ((!prog->minlen && !tmp) && (!reginfo || regtry(reginfo, &s)))
		goto got_it;
	    break;
	case ALNUM:
	    REXEC_FBC_CSCAN_PRELOAD(
		LOAD_UTF8_CHARCLASS_PERL_WORD(),
		swash_fetch(RE_utf8_perl_word, (U8*)s, utf8_target),
                (FLAGS(c) & USE_UNI) ? isWORDCHAR_L1((U8) *s) : isALNUM(*s)
	    );
	case ALNUML:
	    REXEC_FBC_CSCAN_TAINT(
		isALNUM_LC_utf8((U8*)s),
		isALNUM_LC(*s)
	    );
	case NALNUM:
	    REXEC_FBC_CSCAN_PRELOAD(
		LOAD_UTF8_CHARCLASS_PERL_WORD(),
		!swash_fetch(RE_utf8_perl_word, (U8*)s, utf8_target),
                ! ((FLAGS(c) & USE_UNI) ? isWORDCHAR_L1((U8) *s) : isALNUM(*s))
	    );
	case NALNUML:
	    REXEC_FBC_CSCAN_TAINT(
		!isALNUM_LC_utf8((U8*)s),
		!isALNUM_LC(*s)
	    );
	case SPACE:
	    REXEC_FBC_CSCAN_PRELOAD(
		LOAD_UTF8_CHARCLASS_PERL_SPACE(),
		*s == ' ' || swash_fetch(RE_utf8_perl_space,(U8*)s, utf8_target),
                isSPACE_L1((U8) *s) && (isASCII((U8) *s) || (FLAGS(c) & USE_UNI))
	    );
	case SPACEL:
	    REXEC_FBC_CSCAN_TAINT(
		*s == ' ' || isSPACE_LC_utf8((U8*)s),
		isSPACE_LC(*s)
	    );
	case NSPACE:
	    REXEC_FBC_CSCAN_PRELOAD(
		LOAD_UTF8_CHARCLASS_PERL_SPACE(),
		!(*s == ' ' || swash_fetch(RE_utf8_perl_space,(U8*)s, utf8_target)),
                !(isSPACE_L1((U8) *s) && (isASCII((U8) *s) || (FLAGS(c) & USE_UNI)))
	    );
	case NSPACEL:
	    REXEC_FBC_CSCAN_TAINT(
		!(*s == ' ' || isSPACE_LC_utf8((U8*)s)),
		!isSPACE_LC(*s)
	    );
	case DIGIT:
	    REXEC_FBC_CSCAN_PRELOAD(
		LOAD_UTF8_CHARCLASS_POSIX_DIGIT(),
		swash_fetch(RE_utf8_posix_digit,(U8*)s, utf8_target),
		isDIGIT(*s)
	    );
	case DIGITL:
	    REXEC_FBC_CSCAN_TAINT(
		isDIGIT_LC_utf8((U8*)s),
		isDIGIT_LC(*s)
	    );
	case NDIGIT:
	    REXEC_FBC_CSCAN_PRELOAD(
		LOAD_UTF8_CHARCLASS_POSIX_DIGIT(),
		!swash_fetch(RE_utf8_posix_digit,(U8*)s, utf8_target),
		!isDIGIT(*s)
	    );
	case NDIGITL:
	    REXEC_FBC_CSCAN_TAINT(
		!isDIGIT_LC_utf8((U8*)s),
		!isDIGIT_LC(*s)
	    );
	case LNBREAK:
	    REXEC_FBC_CSCAN(
		is_LNBREAK_utf8(s),
		is_LNBREAK_latin1(s)
	    );
	case VERTWS:
	    REXEC_FBC_CSCAN(
		is_VERTWS_utf8(s),
		is_VERTWS_latin1(s)
	    );
	case NVERTWS:
	    REXEC_FBC_CSCAN(
		!is_VERTWS_utf8(s),
		!is_VERTWS_latin1(s)
	    );
	case HORIZWS:
	    REXEC_FBC_CSCAN(
		is_HORIZWS_utf8(s),
		is_HORIZWS_latin1(s)
	    );
	case NHORIZWS:
	    REXEC_FBC_CSCAN(
		!is_HORIZWS_utf8(s),
		!is_HORIZWS_latin1(s)
	    );	    
	case AHOCORASICKC:
	case AHOCORASICK: 
	    {
	        DECL_TRIE_TYPE(c);
                /* what trie are we using right now */
        	reg_ac_data *aho
        	    = (reg_ac_data*)progi->data->data[ ARG( c ) ];
        	reg_trie_data *trie
		    = (reg_trie_data*)progi->data->data[ aho->trie ];
		HV *widecharmap = MUTABLE_HV(progi->data->data[ aho->trie + 1 ]);

		const char *last_start = strend - trie->minlen;
#ifdef DEBUGGING
		const char *real_start = s;
#endif
		STRLEN maxlen = trie->maxlen;
		SV *sv_points;
		U8 **points; /* map of where we were in the input string
		                when reading a given char. For ASCII this
		                is unnecessary overhead as the relationship
		                is always 1:1, but for Unicode, especially
		                case folded Unicode this is not true. */
		U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
		U8 *bitmap=NULL;


                GET_RE_DEBUG_FLAGS_DECL;

                /* We can't just allocate points here. We need to wrap it in
                 * an SV so it gets freed properly if there is a croak while
                 * running the match */
                ENTER;
	        SAVETMPS;
                sv_points=newSV(maxlen * sizeof(U8 *));
                SvCUR_set(sv_points,
                    maxlen * sizeof(U8 *));
                SvPOK_on(sv_points);
                sv_2mortal(sv_points);
                points=(U8**)SvPV_nolen(sv_points );
                if ( trie_type != trie_utf8_fold 
                     && (trie->bitmap || OP(c)==AHOCORASICKC) ) 
                {
                    if (trie->bitmap) 
                        bitmap=(U8*)trie->bitmap;
                    else
                        bitmap=(U8*)ANYOF_BITMAP(c);
                }
                /* this is the Aho-Corasick algorithm modified a touch
                   to include special handling for long "unknown char" 
                   sequences. The basic idea being that we use AC as long
                   as we are dealing with a possible matching char, when
                   we encounter an unknown char (and we have not encountered
                   an accepting state) we scan forward until we find a legal 
                   starting char. 
                   AC matching is basically that of trie matching, except
                   that when we encounter a failing transition, we fall back
                   to the current states "fail state", and try the current char 
                   again, a process we repeat until we reach the root state, 
                   state 1, or a legal transition. If we fail on the root state 
                   then we can either terminate if we have reached an accepting 
                   state previously, or restart the entire process from the beginning 
                   if we have not.

                 */
                while (s <= last_start) {
                    const U32 uniflags = UTF8_ALLOW_DEFAULT;
                    U8 *uc = (U8*)s;
                    U16 charid = 0;
                    U32 base = 1;
                    U32 state = 1;
                    UV uvc = 0;
                    STRLEN len = 0;
                    STRLEN foldlen = 0;
                    U8 *uscan = (U8*)NULL;
                    U8 *leftmost = NULL;
#ifdef DEBUGGING                    
                    U32 accepted_word= 0;
#endif
                    U32 pointpos = 0;

                    while ( state && uc <= (U8*)strend ) {
                        int failed=0;
                        U32 word = aho->states[ state ].wordnum;

                        if( state==1 ) {
                            if ( bitmap ) {
                                DEBUG_TRIE_EXECUTE_r(
                                    if ( uc <= (U8*)last_start && !BITMAP_TEST(bitmap,*uc) ) {
                                        dump_exec_pos( (char *)uc, c, strend, real_start, 
                                            (char *)uc, utf8_target );
                                        PerlIO_printf( Perl_debug_log,
                                            " Scanning for legal start char...\n");
                                    }
                                );
				if (utf8_target) {
				    while ( uc <= (U8*)last_start && !BITMAP_TEST(bitmap,*uc) ) {
					uc += UTF8SKIP(uc);
				    }
				} else {
				    while ( uc <= (U8*)last_start  && !BITMAP_TEST(bitmap,*uc) ) {
					uc++;
				    }
				}
                                s= (char *)uc;
                            }
                            if (uc >(U8*)last_start) break;
                        }
                                            
                        if ( word ) {
                            U8 *lpos= points[ (pointpos - trie->wordinfo[word].len) % maxlen ];
                            if (!leftmost || lpos < leftmost) {
                                DEBUG_r(accepted_word=word);
                                leftmost= lpos;
                            }
                            if (base==0) break;
                            
                        }
                        points[pointpos++ % maxlen]= uc;
			REXEC_TRIE_READ_CHAR(trie_type, trie, widecharmap, uc,
					     uscan, len, uvc, charid, foldlen,
					     foldbuf, uniflags);
                        DEBUG_TRIE_EXECUTE_r({
                            dump_exec_pos( (char *)uc, c, strend, real_start, 
                                s,   utf8_target );
                            PerlIO_printf(Perl_debug_log,
                                " Charid:%3u CP:%4"UVxf" ",
                                 charid, uvc);
                        });

                        do {
#ifdef DEBUGGING
                            word = aho->states[ state ].wordnum;
#endif
                            base = aho->states[ state ].trans.base;

                            DEBUG_TRIE_EXECUTE_r({
                                if (failed) 
                                    dump_exec_pos( (char *)uc, c, strend, real_start, 
                                        s,   utf8_target );
                                PerlIO_printf( Perl_debug_log,
                                    "%sState: %4"UVxf", word=%"UVxf,
                                    failed ? " Fail transition to " : "",
                                    (UV)state, (UV)word);
                            });
                            if ( base ) {
                                U32 tmp;
				I32 offset;
                                if (charid &&
				     ( ((offset = base + charid
					- 1 - trie->uniquecharcount)) >= 0)
                                     && ((U32)offset < trie->lasttrans)
                                     && trie->trans[offset].check == state
                                     && (tmp=trie->trans[offset].next))
                                {
                                    DEBUG_TRIE_EXECUTE_r(
                                        PerlIO_printf( Perl_debug_log," - legal\n"));
                                    state = tmp;
                                    break;
                                }
                                else {
                                    DEBUG_TRIE_EXECUTE_r(
                                        PerlIO_printf( Perl_debug_log," - fail\n"));
                                    failed = 1;
                                    state = aho->fail[state];
                                }
                            }
                            else {
                                /* we must be accepting here */
                                DEBUG_TRIE_EXECUTE_r(
                                        PerlIO_printf( Perl_debug_log," - accepting\n"));
                                failed = 1;
                                break;
                            }
                        } while(state);
                        uc += len;
                        if (failed) {
                            if (leftmost)
                                break;
                            if (!state) state = 1;
                        }
                    }
                    if ( aho->states[ state ].wordnum ) {
                        U8 *lpos = points[ (pointpos - trie->wordinfo[aho->states[ state ].wordnum].len) % maxlen ];
                        if (!leftmost || lpos < leftmost) {
                            DEBUG_r(accepted_word=aho->states[ state ].wordnum);
                            leftmost = lpos;
                        }
                    }
                    if (leftmost) {
                        s = (char*)leftmost;
                        DEBUG_TRIE_EXECUTE_r({
                            PerlIO_printf( 
                                Perl_debug_log,"Matches word #%"UVxf" at position %"IVdf". Trying full pattern...\n",
                                (UV)accepted_word, (IV)(s - real_start)
                            );
                        });
                        if (!reginfo || regtry(reginfo, &s)) {
                            FREETMPS;
		            LEAVE;
                            goto got_it;
                        }
                        s = HOPc(s,1);
                        DEBUG_TRIE_EXECUTE_r({
                            PerlIO_printf( Perl_debug_log,"Pattern failed. Looking for new start point...\n");
                        });
                    } else {
                        DEBUG_TRIE_EXECUTE_r(
                            PerlIO_printf( Perl_debug_log,"No match.\n"));
                        break;
                    }
                }
                FREETMPS;
                LEAVE;
	    }
	    break;
	default:
	    Perl_croak(aTHX_ "panic: unknown regstclass %d", (int)OP(c));
	    break;
	}
	return 0;
      got_it:
	return s;
}


/*
 - regexec_flags - match a regexp against a string
 */
I32
Perl_regexec_flags(pTHX_ REGEXP * const rx, char *stringarg, register char *strend,
	      char *strbeg, I32 minend, SV *sv, void *data, U32 flags)
/* strend: pointer to null at end of string */
/* strbeg: real beginning of string */
/* minend: end of match must be >=minend after stringarg. */
/* data: May be used for some additional optimizations. 
         Currently its only used, with a U32 cast, for transmitting 
         the ganch offset when doing a /g match. This will change */
/* nosave: For optimizations. */
{
    dVAR;
    struct regexp *const prog = (struct regexp *)SvANY(rx);
    /*register*/ char *s;
    register regnode *c;
    /*register*/ char *startpos = stringarg;
    I32 minlen;		/* must match at least this many chars */
    I32 dontbother = 0;	/* how many characters not to try at end */
    I32 end_shift = 0;			/* Same for the end. */		/* CC */
    I32 scream_pos = -1;		/* Internal iterator of scream. */
    char *scream_olds = NULL;
    const bool utf8_target = cBOOL(DO_UTF8(sv));
    I32 multiline;
    RXi_GET_DECL(prog,progi);
    regmatch_info reginfo;  /* create some info to pass to regtry etc */
    regexp_paren_pair *swap = NULL;
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_REGEXEC_FLAGS;
    PERL_UNUSED_ARG(data);

    /* Be paranoid... */
    if (prog == NULL || startpos == NULL) {
	Perl_croak(aTHX_ "NULL regexp parameter");
	return 0;
    }

    multiline = prog->extflags & RXf_PMf_MULTILINE;
    reginfo.prog = rx;	 /* Yes, sorry that this is confusing.  */

    RX_MATCH_UTF8_set(rx, utf8_target);
    DEBUG_EXECUTE_r( 
        debug_start_match(rx, utf8_target, startpos, strend,
        "Matching");
    );

    minlen = prog->minlen;
    
    if (strend - startpos < (minlen+(prog->check_offset_min<0?prog->check_offset_min:0))) {
        DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log,
			      "String too short [regexec_flags]...\n"));
	goto phooey;
    }

    
    /* Check validity of program. */
    if (UCHARAT(progi->program) != REG_MAGIC) {
	Perl_croak(aTHX_ "corrupted regexp program");
    }

    PL_reg_flags = 0;
    PL_reg_eval_set = 0;
    PL_reg_maxiter = 0;

    if (RX_UTF8(rx))
	PL_reg_flags |= RF_utf8;

    /* Mark beginning of line for ^ and lookbehind. */
    reginfo.bol = startpos; /* XXX not used ??? */
    PL_bostr  = strbeg;
    reginfo.sv = sv;

    /* Mark end of line for $ (and such) */
    PL_regeol = strend;

    /* see how far we have to get to not match where we matched before */
    reginfo.till = startpos+minend;

    /* If there is a "must appear" string, look for it. */
    s = startpos;

    if (prog->extflags & RXf_GPOS_SEEN) { /* Need to set reginfo->ganch */
	MAGIC *mg;
	if (flags & REXEC_IGNOREPOS){	/* Means: check only at start */
	    reginfo.ganch = startpos + prog->gofs;
	    DEBUG_GPOS_r(PerlIO_printf(Perl_debug_log,
	      "GPOS IGNOREPOS: reginfo.ganch = startpos + %"UVxf"\n",(UV)prog->gofs));
	} else if (sv && SvTYPE(sv) >= SVt_PVMG
		  && SvMAGIC(sv)
		  && (mg = mg_find(sv, PERL_MAGIC_regex_global))
		  && mg->mg_len >= 0) {
	    reginfo.ganch = strbeg + mg->mg_len;	/* Defined pos() */
	    DEBUG_GPOS_r(PerlIO_printf(Perl_debug_log,
		"GPOS MAGIC: reginfo.ganch = strbeg + %"IVdf"\n",(IV)mg->mg_len));

	    if (prog->extflags & RXf_ANCH_GPOS) {
	        if (s > reginfo.ganch)
		    goto phooey;
		s = reginfo.ganch - prog->gofs;
	        DEBUG_GPOS_r(PerlIO_printf(Perl_debug_log,
		     "GPOS ANCH_GPOS: s = ganch - %"UVxf"\n",(UV)prog->gofs));
		if (s < strbeg)
		    goto phooey;
	    }
	}
	else if (data) {
	    reginfo.ganch = strbeg + PTR2UV(data);
            DEBUG_GPOS_r(PerlIO_printf(Perl_debug_log,
		 "GPOS DATA: reginfo.ganch= strbeg + %"UVxf"\n",PTR2UV(data)));

	} else {				/* pos() not defined */
	    reginfo.ganch = strbeg;
            DEBUG_GPOS_r(PerlIO_printf(Perl_debug_log,
		 "GPOS: reginfo.ganch = strbeg\n"));
	}
    }
    if (PL_curpm && (PM_GETRE(PL_curpm) == rx)) {
        /* We have to be careful. If the previous successful match
           was from this regex we don't want a subsequent partially
           successful match to clobber the old results.
           So when we detect this possibility we add a swap buffer
           to the re, and switch the buffer each match. If we fail
           we switch it back, otherwise we leave it swapped.
        */
        swap = prog->offs;
        /* do we need a save destructor here for eval dies? */
        Newxz(prog->offs, (prog->nparens + 1), regexp_paren_pair);
    }
    if (!(flags & REXEC_CHECKED) && (prog->check_substr != NULL || prog->check_utf8 != NULL)) {
	re_scream_pos_data d;

	d.scream_olds = &scream_olds;
	d.scream_pos = &scream_pos;
	s = re_intuit_start(rx, sv, s, strend, flags, &d);
	if (!s) {
	    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "Not present...\n"));
	    goto phooey;	/* not present */
	}
    }



    /* Simplest case:  anchored match need be tried only once. */
    /*  [unless only anchor is BOL and multiline is set] */
    if (prog->extflags & (RXf_ANCH & ~RXf_ANCH_GPOS)) {
	if (s == startpos && regtry(&reginfo, &startpos))
	    goto got_it;
	else if (multiline || (prog->intflags & PREGf_IMPLICIT)
		 || (prog->extflags & RXf_ANCH_MBOL)) /* XXXX SBOL? */
	{
	    char *end;

	    if (minlen)
		dontbother = minlen - 1;
	    end = HOP3c(strend, -dontbother, strbeg) - 1;
	    /* for multiline we only have to try after newlines */
	    if (prog->check_substr || prog->check_utf8) {
                /* because of the goto we can not easily reuse the macros for bifurcating the
                   unicode/non-unicode match modes here like we do elsewhere - demerphq */
                if (utf8_target) {
                    if (s == startpos)
                        goto after_try_utf8;
                    while (1) {
                        if (regtry(&reginfo, &s)) {
                            goto got_it;
                        }
                      after_try_utf8:
                        if (s > end) {
                            goto phooey;
                        }
                        if (prog->extflags & RXf_USE_INTUIT) {
                            s = re_intuit_start(rx, sv, s + UTF8SKIP(s), strend, flags, NULL);
                            if (!s) {
                                goto phooey;
                            }
                        }
                        else {
                            s += UTF8SKIP(s);
                        }
                    }
                } /* end search for check string in unicode */
                else {
                    if (s == startpos) {
                        goto after_try_latin;
                    }
                    while (1) {
                        if (regtry(&reginfo, &s)) {
                            goto got_it;
                        }
                      after_try_latin:
                        if (s > end) {
                            goto phooey;
                        }
                        if (prog->extflags & RXf_USE_INTUIT) {
                            s = re_intuit_start(rx, sv, s + 1, strend, flags, NULL);
                            if (!s) {
                                goto phooey;
                            }
                        }
                        else {
                            s++;
                        }
                    }
                } /* end search for check string in latin*/
	    } /* end search for check string */
	    else { /* search for newline */
		if (s > startpos) {
                    /*XXX: The s-- is almost definitely wrong here under unicode - demeprhq*/
		    s--;
		}
                /* We can use a more efficient search as newlines are the same in unicode as they are in latin */
		while (s < end) {
		    if (*s++ == '\n') {	/* don't need PL_utf8skip here */
			if (regtry(&reginfo, &s))
			    goto got_it;
		    }
		}
	    } /* end search for newline */
	} /* end anchored/multiline check string search */
	goto phooey;
    } else if (RXf_GPOS_CHECK == (prog->extflags & RXf_GPOS_CHECK)) 
    {
        /* the warning about reginfo.ganch being used without intialization
           is bogus -- we set it above, when prog->extflags & RXf_GPOS_SEEN 
           and we only enter this block when the same bit is set. */
        char *tmp_s = reginfo.ganch - prog->gofs;

	if (tmp_s >= strbeg && regtry(&reginfo, &tmp_s))
	    goto got_it;
	goto phooey;
    }

    /* Messy cases:  unanchored match. */
    if ((prog->anchored_substr || prog->anchored_utf8) && prog->intflags & PREGf_SKIP) {
	/* we have /x+whatever/ */
	/* it must be a one character string (XXXX Except UTF_PATTERN?) */
	char ch;
#ifdef DEBUGGING
	int did_match = 0;
#endif
	if (!(utf8_target ? prog->anchored_utf8 : prog->anchored_substr))
	    utf8_target ? to_utf8_substr(prog) : to_byte_substr(prog);
	ch = SvPVX_const(utf8_target ? prog->anchored_utf8 : prog->anchored_substr)[0];

	if (utf8_target) {
	    REXEC_FBC_SCAN(
		if (*s == ch) {
		    DEBUG_EXECUTE_r( did_match = 1 );
		    if (regtry(&reginfo, &s)) goto got_it;
		    s += UTF8SKIP(s);
		    while (s < strend && *s == ch)
			s += UTF8SKIP(s);
		}
	    );
	}
	else {
	    REXEC_FBC_SCAN(
		if (*s == ch) {
		    DEBUG_EXECUTE_r( did_match = 1 );
		    if (regtry(&reginfo, &s)) goto got_it;
		    s++;
		    while (s < strend && *s == ch)
			s++;
		}
	    );
	}
	DEBUG_EXECUTE_r(if (!did_match)
		PerlIO_printf(Perl_debug_log,
                                  "Did not find anchored character...\n")
               );
    }
    else if (prog->anchored_substr != NULL
	      || prog->anchored_utf8 != NULL
	      || ((prog->float_substr != NULL || prog->float_utf8 != NULL)
		  && prog->float_max_offset < strend - s)) {
	SV *must;
	I32 back_max;
	I32 back_min;
	char *last;
	char *last1;		/* Last position checked before */
#ifdef DEBUGGING
	int did_match = 0;
#endif
	if (prog->anchored_substr || prog->anchored_utf8) {
	    if (!(utf8_target ? prog->anchored_utf8 : prog->anchored_substr))
		utf8_target ? to_utf8_substr(prog) : to_byte_substr(prog);
	    must = utf8_target ? prog->anchored_utf8 : prog->anchored_substr;
	    back_max = back_min = prog->anchored_offset;
	} else {
	    if (!(utf8_target ? prog->float_utf8 : prog->float_substr))
		utf8_target ? to_utf8_substr(prog) : to_byte_substr(prog);
	    must = utf8_target ? prog->float_utf8 : prog->float_substr;
	    back_max = prog->float_max_offset;
	    back_min = prog->float_min_offset;
	}
	
	    
	if (must == &PL_sv_undef)
	    /* could not downgrade utf8 check substring, so must fail */
	    goto phooey;

        if (back_min<0) {
	    last = strend;
	} else {
            last = HOP3c(strend,	/* Cannot start after this */
        	  -(I32)(CHR_SVLEN(must)
        		 - (SvTAIL(must) != 0) + back_min), strbeg);
        }
	if (s > PL_bostr)
	    last1 = HOPc(s, -1);
	else
	    last1 = s - 1;	/* bogus */

	/* XXXX check_substr already used to find "s", can optimize if
	   check_substr==must. */
	scream_pos = -1;
	dontbother = end_shift;
	strend = HOPc(strend, -dontbother);
	while ( (s <= last) &&
		((flags & REXEC_SCREAM)
		 ? (s = screaminstr(sv, must, HOP3c(s, back_min, (back_min<0 ? strbeg : strend)) - strbeg,
				    end_shift, &scream_pos, 0))
		 : (s = fbm_instr((unsigned char*)HOP3(s, back_min, (back_min<0 ? strbeg : strend)),
				  (unsigned char*)strend, must,
				  multiline ? FBMrf_MULTILINE : 0))) ) {
	    /* we may be pointing at the wrong string */
	    if ((flags & REXEC_SCREAM) && RXp_MATCH_COPIED(prog))
		s = strbeg + (s - SvPVX_const(sv));
	    DEBUG_EXECUTE_r( did_match = 1 );
	    if (HOPc(s, -back_max) > last1) {
		last1 = HOPc(s, -back_min);
		s = HOPc(s, -back_max);
	    }
	    else {
		char * const t = (last1 >= PL_bostr) ? HOPc(last1, 1) : last1 + 1;

		last1 = HOPc(s, -back_min);
		s = t;
	    }
	    if (utf8_target) {
		while (s <= last1) {
		    if (regtry(&reginfo, &s))
			goto got_it;
		    s += UTF8SKIP(s);
		}
	    }
	    else {
		while (s <= last1) {
		    if (regtry(&reginfo, &s))
			goto got_it;
		    s++;
		}
	    }
	}
	DEBUG_EXECUTE_r(if (!did_match) {
            RE_PV_QUOTED_DECL(quoted, utf8_target, PERL_DEBUG_PAD_ZERO(0),
                SvPVX_const(must), RE_SV_DUMPLEN(must), 30);
            PerlIO_printf(Perl_debug_log, "Did not find %s substr %s%s...\n",
			      ((must == prog->anchored_substr || must == prog->anchored_utf8)
			       ? "anchored" : "floating"),
                quoted, RE_SV_TAIL(must));
        });		    
	goto phooey;
    }
    else if ( (c = progi->regstclass) ) {
	if (minlen) {
	    const OPCODE op = OP(progi->regstclass);
	    /* don't bother with what can't match */
	    if (PL_regkind[op] != EXACT && op != CANY && PL_regkind[op] != TRIE)
	        strend = HOPc(strend, -(minlen - 1));
	}
	DEBUG_EXECUTE_r({
	    SV * const prop = sv_newmortal();
	    regprop(prog, prop, c);
	    {
		RE_PV_QUOTED_DECL(quoted,utf8_target,PERL_DEBUG_PAD_ZERO(1),
		    s,strend-s,60);
		PerlIO_printf(Perl_debug_log,
		    "Matching stclass %.*s against %s (%d bytes)\n",
		    (int)SvCUR(prop), SvPVX_const(prop),
		     quoted, (int)(strend - s));
	    }
	});
        if (find_byclass(prog, c, s, strend, &reginfo))
	    goto got_it;
	DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "Contradicts stclass... [regexec_flags]\n"));
    }
    else {
	dontbother = 0;
	if (prog->float_substr != NULL || prog->float_utf8 != NULL) {
	    /* Trim the end. */
	    char *last;
	    SV* float_real;

	    if (!(utf8_target ? prog->float_utf8 : prog->float_substr))
		utf8_target ? to_utf8_substr(prog) : to_byte_substr(prog);
	    float_real = utf8_target ? prog->float_utf8 : prog->float_substr;

	    if (flags & REXEC_SCREAM) {
		last = screaminstr(sv, float_real, s - strbeg,
				   end_shift, &scream_pos, 1); /* last one */
		if (!last)
		    last = scream_olds; /* Only one occurrence. */
		/* we may be pointing at the wrong string */
		else if (RXp_MATCH_COPIED(prog))
		    s = strbeg + (s - SvPVX_const(sv));
	    }
	    else {
		STRLEN len;
                const char * const little = SvPV_const(float_real, len);

		if (SvTAIL(float_real)) {
		    if (memEQ(strend - len + 1, little, len - 1))
			last = strend - len + 1;
		    else if (!multiline)
			last = memEQ(strend - len, little, len)
			    ? strend - len : NULL;
		    else
			goto find_last;
		} else {
		  find_last:
		    if (len)
			last = rninstr(s, strend, little, little + len);
		    else
			last = strend;	/* matching "$" */
		}
	    }
	    if (last == NULL) {
		DEBUG_EXECUTE_r(
		    PerlIO_printf(Perl_debug_log,
			"%sCan't trim the tail, match fails (should not happen)%s\n",
	                PL_colors[4], PL_colors[5]));
		goto phooey; /* Should not happen! */
	    }
	    dontbother = strend - last + prog->float_min_offset;
	}
	if (minlen && (dontbother < minlen))
	    dontbother = minlen - 1;
	strend -= dontbother; 		   /* this one's always in bytes! */
	/* We don't know much -- general case. */
	if (utf8_target) {
	    for (;;) {
		if (regtry(&reginfo, &s))
		    goto got_it;
		if (s >= strend)
		    break;
		s += UTF8SKIP(s);
	    };
	}
	else {
	    do {
		if (regtry(&reginfo, &s))
		    goto got_it;
	    } while (s++ < strend);
	}
    }

    /* Failure. */
    goto phooey;

got_it:
    Safefree(swap);
    RX_MATCH_TAINTED_set(rx, PL_reg_flags & RF_tainted);

    if (PL_reg_eval_set)
	restore_pos(aTHX_ prog);
    if (RXp_PAREN_NAMES(prog)) 
        (void)hv_iterinit(RXp_PAREN_NAMES(prog));

    /* make sure $`, $&, $', and $digit will work later */
    if ( !(flags & REXEC_NOT_FIRST) ) {
	RX_MATCH_COPY_FREE(rx);
	if (flags & REXEC_COPY_STR) {
	    const I32 i = PL_regeol - startpos + (stringarg - strbeg);
#ifdef PERL_OLD_COPY_ON_WRITE
	    if ((SvIsCOW(sv)
		 || (SvFLAGS(sv) & CAN_COW_MASK) == CAN_COW_FLAGS)) {
		if (DEBUG_C_TEST) {
		    PerlIO_printf(Perl_debug_log,
				  "Copy on write: regexp capture, type %d\n",
				  (int) SvTYPE(sv));
		}
		prog->saved_copy = sv_setsv_cow(prog->saved_copy, sv);
		prog->subbeg = (char *)SvPVX_const(prog->saved_copy);
		assert (SvPOKp(prog->saved_copy));
	    } else
#endif
	    {
		RX_MATCH_COPIED_on(rx);
		s = savepvn(strbeg, i);
		prog->subbeg = s;
	    }
	    prog->sublen = i;
	}
	else {
	    prog->subbeg = strbeg;
	    prog->sublen = PL_regeol - strbeg;	/* strend may have been modified */
	}
    }

    return 1;

phooey:
    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "%sMatch failed%s\n",
			  PL_colors[4], PL_colors[5]));
    if (PL_reg_eval_set)
	restore_pos(aTHX_ prog);
    if (swap) {
        /* we failed :-( roll it back */
        Safefree(prog->offs);
        prog->offs = swap;
    }

    return 0;
}


/*
 - regtry - try match at specific point
 */
STATIC I32			/* 0 failure, 1 success */
S_regtry(pTHX_ regmatch_info *reginfo, char **startpos)
{
    dVAR;
    CHECKPOINT lastcp;
    REGEXP *const rx = reginfo->prog;
    regexp *const prog = (struct regexp *)SvANY(rx);
    RXi_GET_DECL(prog,progi);
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_REGTRY;

    reginfo->cutpoint=NULL;

    if ((prog->extflags & RXf_EVAL_SEEN) && !PL_reg_eval_set) {
	MAGIC *mg;

	PL_reg_eval_set = RS_init;
	DEBUG_EXECUTE_r(DEBUG_s(
	    PerlIO_printf(Perl_debug_log, "  setting stack tmpbase at %"IVdf"\n",
			  (IV)(PL_stack_sp - PL_stack_base));
	    ));
	SAVESTACK_CXPOS();
	cxstack[cxstack_ix].blk_oldsp = PL_stack_sp - PL_stack_base;
	/* Otherwise OP_NEXTSTATE will free whatever on stack now.  */
	SAVETMPS;
	/* Apparently this is not needed, judging by wantarray. */
	/* SAVEI8(cxstack[cxstack_ix].blk_gimme);
	   cxstack[cxstack_ix].blk_gimme = G_SCALAR; */

	if (reginfo->sv) {
	    /* Make $_ available to executed code. */
	    if (reginfo->sv != DEFSV) {
		SAVE_DEFSV;
		DEFSV_set(reginfo->sv);
	    }
	
	    if (!(SvTYPE(reginfo->sv) >= SVt_PVMG && SvMAGIC(reginfo->sv)
		  && (mg = mg_find(reginfo->sv, PERL_MAGIC_regex_global)))) {
		/* prepare for quick setting of pos */
#ifdef PERL_OLD_COPY_ON_WRITE
		if (SvIsCOW(reginfo->sv))
		    sv_force_normal_flags(reginfo->sv, 0);
#endif
		mg = sv_magicext(reginfo->sv, NULL, PERL_MAGIC_regex_global,
				 &PL_vtbl_mglob, NULL, 0);
		mg->mg_len = -1;
	    }
	    PL_reg_magic    = mg;
	    PL_reg_oldpos   = mg->mg_len;
	    SAVEDESTRUCTOR_X(restore_pos, prog);
        }
        if (!PL_reg_curpm) {
	    Newxz(PL_reg_curpm, 1, PMOP);
#ifdef USE_ITHREADS
            {
		SV* const repointer = &PL_sv_undef;
                /* this regexp is also owned by the new PL_reg_curpm, which
		   will try to free it.  */
                av_push(PL_regex_padav, repointer);
                PL_reg_curpm->op_pmoffset = av_len(PL_regex_padav);
                PL_regex_pad = AvARRAY(PL_regex_padav);
            }
#endif      
        }
#ifdef USE_ITHREADS
	/* It seems that non-ithreads works both with and without this code.
	   So for efficiency reasons it seems best not to have the code
	   compiled when it is not needed.  */
	/* This is safe against NULLs: */
	ReREFCNT_dec(PM_GETRE(PL_reg_curpm));
	/* PM_reg_curpm owns a reference to this regexp.  */
	ReREFCNT_inc(rx);
#endif
	PM_SETRE(PL_reg_curpm, rx);
	PL_reg_oldcurpm = PL_curpm;
	PL_curpm = PL_reg_curpm;
	if (RXp_MATCH_COPIED(prog)) {
	    /*  Here is a serious problem: we cannot rewrite subbeg,
		since it may be needed if this match fails.  Thus
		$` inside (?{}) could fail... */
	    PL_reg_oldsaved = prog->subbeg;
	    PL_reg_oldsavedlen = prog->sublen;
#ifdef PERL_OLD_COPY_ON_WRITE
	    PL_nrs = prog->saved_copy;
#endif
	    RXp_MATCH_COPIED_off(prog);
	}
	else
	    PL_reg_oldsaved = NULL;
	prog->subbeg = PL_bostr;
	prog->sublen = PL_regeol - PL_bostr; /* strend may have been modified */
    }
    DEBUG_EXECUTE_r(PL_reg_starttry = *startpos);
    prog->offs[0].start = *startpos - PL_bostr;
    PL_reginput = *startpos;
    PL_reglastparen = &prog->lastparen;
    PL_reglastcloseparen = &prog->lastcloseparen;
    prog->lastparen = 0;
    prog->lastcloseparen = 0;
    PL_regsize = 0;
    PL_regoffs = prog->offs;
    if (PL_reg_start_tmpl <= prog->nparens) {
	PL_reg_start_tmpl = prog->nparens*3/2 + 3;
        if(PL_reg_start_tmp)
            Renew(PL_reg_start_tmp, PL_reg_start_tmpl, char*);
        else
            Newx(PL_reg_start_tmp, PL_reg_start_tmpl, char*);
    }

    /* XXXX What this code is doing here?!!!  There should be no need
       to do this again and again, PL_reglastparen should take care of
       this!  --ilya*/

    /* Tests pat.t#187 and split.t#{13,14} seem to depend on this code.
     * Actually, the code in regcppop() (which Ilya may be meaning by
     * PL_reglastparen), is not needed at all by the test suite
     * (op/regexp, op/pat, op/split), but that code is needed otherwise
     * this erroneously leaves $1 defined: "1" =~ /^(?:(\d)x)?\d$/
     * Meanwhile, this code *is* needed for the
     * above-mentioned test suite tests to succeed.  The common theme
     * on those tests seems to be returning null fields from matches.
     * --jhi updated by dapm */
#if 1
    if (prog->nparens) {
	regexp_paren_pair *pp = PL_regoffs;
	register I32 i;
	for (i = prog->nparens; i > (I32)*PL_reglastparen; i--) {
	    ++pp;
	    pp->start = -1;
	    pp->end = -1;
	}
    }
#endif
    REGCP_SET(lastcp);
    if (regmatch(reginfo, progi->program + 1)) {
	PL_regoffs[0].end = PL_reginput - PL_bostr;
	return 1;
    }
    if (reginfo->cutpoint)
        *startpos= reginfo->cutpoint;
    REGCP_UNWIND(lastcp);
    return 0;
}


#define sayYES goto yes
#define sayNO goto no
#define sayNO_SILENT goto no_silent

/* we dont use STMT_START/END here because it leads to 
   "unreachable code" warnings, which are bogus, but distracting. */
#define CACHEsayNO \
    if (ST.cache_mask) \
       PL_reg_poscache[ST.cache_offset] |= ST.cache_mask; \
    sayNO

/* this is used to determine how far from the left messages like
   'failed...' are printed. It should be set such that messages 
   are inline with the regop output that created them.
*/
#define REPORT_CODE_OFF 32


#define CHRTEST_UNINIT -1001 /* c1/c2 haven't been calculated yet */
#define CHRTEST_VOID   -1000 /* the c1/c2 "next char" test should be skipped */

#define SLAB_FIRST(s) (&(s)->states[0])
#define SLAB_LAST(s)  (&(s)->states[PERL_REGMATCH_SLAB_SLOTS-1])

/* grab a new slab and return the first slot in it */

STATIC regmatch_state *
S_push_slab(pTHX)
{
#if PERL_VERSION < 9 && !defined(PERL_CORE)
    dMY_CXT;
#endif
    regmatch_slab *s = PL_regmatch_slab->next;
    if (!s) {
	Newx(s, 1, regmatch_slab);
	s->prev = PL_regmatch_slab;
	s->next = NULL;
	PL_regmatch_slab->next = s;
    }
    PL_regmatch_slab = s;
    return SLAB_FIRST(s);
}


/* push a new state then goto it */

#define PUSH_STATE_GOTO(state, node) \
    scan = node; \
    st->resume_state = state; \
    goto push_state;

/* push a new state with success backtracking, then goto it */

#define PUSH_YES_STATE_GOTO(state, node) \
    scan = node; \
    st->resume_state = state; \
    goto push_yes_state;



/*

regmatch() - main matching routine

This is basically one big switch statement in a loop. We execute an op,
set 'next' to point the next op, and continue. If we come to a point which
we may need to backtrack to on failure such as (A|B|C), we push a
backtrack state onto the backtrack stack. On failure, we pop the top
state, and re-enter the loop at the state indicated. If there are no more
states to pop, we return failure.

Sometimes we also need to backtrack on success; for example /A+/, where
after successfully matching one A, we need to go back and try to
match another one; similarly for lookahead assertions: if the assertion
completes successfully, we backtrack to the state just before the assertion
and then carry on.  In these cases, the pushed state is marked as
'backtrack on success too'. This marking is in fact done by a chain of
pointers, each pointing to the previous 'yes' state. On success, we pop to
the nearest yes state, discarding any intermediate failure-only states.
Sometimes a yes state is pushed just to force some cleanup code to be
called at the end of a successful match or submatch; e.g. (??{$re}) uses
it to free the inner regex.

Note that failure backtracking rewinds the cursor position, while
success backtracking leaves it alone.

A pattern is complete when the END op is executed, while a subpattern
such as (?=foo) is complete when the SUCCESS op is executed. Both of these
ops trigger the "pop to last yes state if any, otherwise return true"
behaviour.

A common convention in this function is to use A and B to refer to the two
subpatterns (or to the first nodes thereof) in patterns like /A*B/: so A is
the subpattern to be matched possibly multiple times, while B is the entire
rest of the pattern. Variable and state names reflect this convention.

The states in the main switch are the union of ops and failure/success of
substates associated with with that op.  For example, IFMATCH is the op
that does lookahead assertions /(?=A)B/ and so the IFMATCH state means
'execute IFMATCH'; while IFMATCH_A is a state saying that we have just
successfully matched A and IFMATCH_A_fail is a state saying that we have
just failed to match A. Resume states always come in pairs. The backtrack
state we push is marked as 'IFMATCH_A', but when that is popped, we resume
at IFMATCH_A or IFMATCH_A_fail, depending on whether we are backtracking
on success or failure.

The struct that holds a backtracking state is actually a big union, with
one variant for each major type of op. The variable st points to the
top-most backtrack struct. To make the code clearer, within each
block of code we #define ST to alias the relevant union.

Here's a concrete example of a (vastly oversimplified) IFMATCH
implementation:

    switch (state) {
    ....

#define ST st->u.ifmatch

    case IFMATCH: // we are executing the IFMATCH op, (?=A)B
	ST.foo = ...; // some state we wish to save
	...
	// push a yes backtrack state with a resume value of
	// IFMATCH_A/IFMATCH_A_fail, then continue execution at the
	// first node of A:
	PUSH_YES_STATE_GOTO(IFMATCH_A, A);
	// NOTREACHED

    case IFMATCH_A: // we have successfully executed A; now continue with B
	next = B;
	bar = ST.foo; // do something with the preserved value
	break;

    case IFMATCH_A_fail: // A failed, so the assertion failed
	...;   // do some housekeeping, then ...
	sayNO; // propagate the failure

#undef ST

    ...
    }

For any old-timers reading this who are familiar with the old recursive
approach, the code above is equivalent to:

    case IFMATCH: // we are executing the IFMATCH op, (?=A)B
    {
	int foo = ...
	...
	if (regmatch(A)) {
	    next = B;
	    bar = foo;
	    break;
	}
	...;   // do some housekeeping, then ...
	sayNO; // propagate the failure
    }

The topmost backtrack state, pointed to by st, is usually free. If you
want to claim it, populate any ST.foo fields in it with values you wish to
save, then do one of

	PUSH_STATE_GOTO(resume_state, node);
	PUSH_YES_STATE_GOTO(resume_state, node);

which sets that backtrack state's resume value to 'resume_state', pushes a
new free entry to the top of the backtrack stack, then goes to 'node'.
On backtracking, the free slot is popped, and the saved state becomes the
new free state. An ST.foo field in this new top state can be temporarily
accessed to retrieve values, but once the main loop is re-entered, it
becomes available for reuse.

Note that the depth of the backtrack stack constantly increases during the
left-to-right execution of the pattern, rather than going up and down with
the pattern nesting. For example the stack is at its maximum at Z at the
end of the pattern, rather than at X in the following:

    /(((X)+)+)+....(Y)+....Z/

The only exceptions to this are lookahead/behind assertions and the cut,
(?>A), which pop all the backtrack states associated with A before
continuing.
 
Bascktrack state structs are allocated in slabs of about 4K in size.
PL_regmatch_state and st always point to the currently active state,
and PL_regmatch_slab points to the slab currently containing
PL_regmatch_state.  The first time regmatch() is called, the first slab is
allocated, and is never freed until interpreter destruction. When the slab
is full, a new one is allocated and chained to the end. At exit from
regmatch(), slabs allocated since entry are freed.

*/
 

#define DEBUG_STATE_pp(pp)				    \
    DEBUG_STATE_r({					    \
	DUMP_EXEC_POS(locinput, scan, utf8_target);		    \
	PerlIO_printf(Perl_debug_log,			    \
	    "    %*s"pp" %s%s%s%s%s\n",			    \
	    depth*2, "",				    \
	    PL_reg_name[st->resume_state],                     \
	    ((st==yes_state||st==mark_state) ? "[" : ""),   \
	    ((st==yes_state) ? "Y" : ""),                   \
	    ((st==mark_state) ? "M" : ""),                  \
	    ((st==yes_state||st==mark_state) ? "]" : "")    \
	);                                                  \
    });


#define REG_NODE_NUM(x) ((x) ? (int)((x)-prog) : -1)

#ifdef DEBUGGING

STATIC void
S_debug_start_match(pTHX_ const REGEXP *prog, const bool utf8_target,
    const char *start, const char *end, const char *blurb)
{
    const bool utf8_pat = RX_UTF8(prog) ? 1 : 0;

    PERL_ARGS_ASSERT_DEBUG_START_MATCH;

    if (!PL_colorset)   
            reginitcolors();    
    {
        RE_PV_QUOTED_DECL(s0, utf8_pat, PERL_DEBUG_PAD_ZERO(0), 
            RX_PRECOMP_const(prog), RX_PRELEN(prog), 60);   
        
        RE_PV_QUOTED_DECL(s1, utf8_target, PERL_DEBUG_PAD_ZERO(1),
            start, end - start, 60); 
        
        PerlIO_printf(Perl_debug_log, 
            "%s%s REx%s %s against %s\n", 
		       PL_colors[4], blurb, PL_colors[5], s0, s1); 
        
        if (utf8_target||utf8_pat)
            PerlIO_printf(Perl_debug_log, "UTF-8 %s%s%s...\n",
                utf8_pat ? "pattern" : "",
                utf8_pat && utf8_target ? " and " : "",
                utf8_target ? "string" : ""
            ); 
    }
}

STATIC void
S_dump_exec_pos(pTHX_ const char *locinput, 
                      const regnode *scan, 
                      const char *loc_regeol, 
                      const char *loc_bostr, 
                      const char *loc_reg_starttry,
                      const bool utf8_target)
{
    const int docolor = *PL_colors[0] || *PL_colors[2] || *PL_colors[4];
    const int taill = (docolor ? 10 : 7); /* 3 chars for "> <" */
    int l = (loc_regeol - locinput) > taill ? taill : (loc_regeol - locinput);
    /* The part of the string before starttry has one color
       (pref0_len chars), between starttry and current
       position another one (pref_len - pref0_len chars),
       after the current position the third one.
       We assume that pref0_len <= pref_len, otherwise we
       decrease pref0_len.  */
    int pref_len = (locinput - loc_bostr) > (5 + taill) - l
	? (5 + taill) - l : locinput - loc_bostr;
    int pref0_len;

    PERL_ARGS_ASSERT_DUMP_EXEC_POS;

    while (utf8_target && UTF8_IS_CONTINUATION(*(U8*)(locinput - pref_len)))
	pref_len++;
    pref0_len = pref_len  - (locinput - loc_reg_starttry);
    if (l + pref_len < (5 + taill) && l < loc_regeol - locinput)
	l = ( loc_regeol - locinput > (5 + taill) - pref_len
	      ? (5 + taill) - pref_len : loc_regeol - locinput);
    while (utf8_target && UTF8_IS_CONTINUATION(*(U8*)(locinput + l)))
	l--;
    if (pref0_len < 0)
	pref0_len = 0;
    if (pref0_len > pref_len)
	pref0_len = pref_len;
    {
	const int is_uni = (utf8_target && OP(scan) != CANY) ? 1 : 0;

	RE_PV_COLOR_DECL(s0,len0,is_uni,PERL_DEBUG_PAD(0),
	    (locinput - pref_len),pref0_len, 60, 4, 5);
	
	RE_PV_COLOR_DECL(s1,len1,is_uni,PERL_DEBUG_PAD(1),
		    (locinput - pref_len + pref0_len),
		    pref_len - pref0_len, 60, 2, 3);
	
	RE_PV_COLOR_DECL(s2,len2,is_uni,PERL_DEBUG_PAD(2),
		    locinput, loc_regeol - locinput, 10, 0, 1);

	const STRLEN tlen=len0+len1+len2;
	PerlIO_printf(Perl_debug_log,
		    "%4"IVdf" <%.*s%.*s%s%.*s>%*s|",
		    (IV)(locinput - loc_bostr),
		    len0, s0,
		    len1, s1,
		    (docolor ? "" : "> <"),
		    len2, s2,
		    (int)(tlen > 19 ? 0 :  19 - tlen),
		    "");
    }
}

#endif

/* reg_check_named_buff_matched()
 * Checks to see if a named buffer has matched. The data array of 
 * buffer numbers corresponding to the buffer is expected to reside
 * in the regexp->data->data array in the slot stored in the ARG() of
 * node involved. Note that this routine doesn't actually care about the
 * name, that information is not preserved from compilation to execution.
 * Returns the index of the leftmost defined buffer with the given name
 * or 0 if non of the buffers matched.
 */
STATIC I32
S_reg_check_named_buff_matched(pTHX_ const regexp *rex, const regnode *scan)
{
    I32 n;
    RXi_GET_DECL(rex,rexi);
    SV *sv_dat= MUTABLE_SV(rexi->data->data[ ARG( scan ) ]);
    I32 *nums=(I32*)SvPVX(sv_dat);

    PERL_ARGS_ASSERT_REG_CHECK_NAMED_BUFF_MATCHED;

    for ( n=0; n<SvIVX(sv_dat); n++ ) {
        if ((I32)*PL_reglastparen >= nums[n] &&
            PL_regoffs[nums[n]].end != -1)
        {
            return nums[n];
        }
    }
    return 0;
}


/* free all slabs above current one  - called during LEAVE_SCOPE */

STATIC void
S_clear_backtrack_stack(pTHX_ void *p)
{
    regmatch_slab *s = PL_regmatch_slab->next;
    PERL_UNUSED_ARG(p);

    if (!s)
	return;
    PL_regmatch_slab->next = NULL;
    while (s) {
	regmatch_slab * const osl = s;
	s = s->next;
	Safefree(osl);
    }
}


#define SETREX(Re1,Re2) \
    if (PL_reg_eval_set) PM_SETRE((PL_reg_curpm), (Re2)); \
    Re1 = (Re2)

STATIC I32			/* 0 failure, 1 success */
S_regmatch(pTHX_ regmatch_info *reginfo, regnode *prog)
{
#if PERL_VERSION < 9 && !defined(PERL_CORE)
    dMY_CXT;
#endif
    dVAR;
    register const bool utf8_target = PL_reg_match_utf8;
    const U32 uniflags = UTF8_ALLOW_DEFAULT;
    REGEXP *rex_sv = reginfo->prog;
    regexp *rex = (struct regexp *)SvANY(rex_sv);
    RXi_GET_DECL(rex,rexi);
    I32	oldsave;
    /* the current state. This is a cached copy of PL_regmatch_state */
    register regmatch_state *st;
    /* cache heavy used fields of st in registers */
    register regnode *scan;
    register regnode *next;
    register U32 n = 0;	/* general value; init to avoid compiler warning */
    register I32 ln = 0; /* len or last;  init to avoid compiler warning */
    register char *locinput = PL_reginput;
    register I32 nextchr;   /* is always set to UCHARAT(locinput) */

    bool result = 0;	    /* return value of S_regmatch */
    int depth = 0;	    /* depth of backtrack stack */
    U32 nochange_depth = 0; /* depth of GOSUB recursion with nochange */
    const U32 max_nochange_depth =
        (3 * rex->nparens > MAX_RECURSE_EVAL_NOCHANGE_DEPTH) ?
        3 * rex->nparens : MAX_RECURSE_EVAL_NOCHANGE_DEPTH;
    regmatch_state *yes_state = NULL; /* state to pop to on success of
							    subpattern */
    /* mark_state piggy backs on the yes_state logic so that when we unwind 
       the stack on success we can update the mark_state as we go */
    regmatch_state *mark_state = NULL; /* last mark state we have seen */
    regmatch_state *cur_eval = NULL; /* most recent EVAL_AB state */
    struct regmatch_state  *cur_curlyx = NULL; /* most recent curlyx */
    U32 state_num;
    bool no_final = 0;      /* prevent failure from backtracking? */
    bool do_cutgroup = 0;   /* no_final only until next branch/trie entry */
    char *startpoint = PL_reginput;
    SV *popmark = NULL;     /* are we looking for a mark? */
    SV *sv_commit = NULL;   /* last mark name seen in failure */
    SV *sv_yes_mark = NULL; /* last mark name we have seen 
                               during a successfull match */
    U32 lastopen = 0;       /* last open we saw */
    bool has_cutgroup = RX_HAS_CUTGROUP(rex) ? 1 : 0;   
    SV* const oreplsv = GvSV(PL_replgv);
    /* these three flags are set by various ops to signal information to
     * the very next op. They have a useful lifetime of exactly one loop
     * iteration, and are not preserved or restored by state pushes/pops
     */
    bool sw = 0;	    /* the condition value in (?(cond)a|b) */
    bool minmod = 0;	    /* the next "{n,m}" is a "{n,m}?" */
    int logical = 0;	    /* the following EVAL is:
				0: (?{...})
				1: (?(?{...})X|Y)
				2: (??{...})
			       or the following IFMATCH/UNLESSM is:
			        false: plain (?=foo)
				true:  used as a condition: (?(?=foo))
			    */
#ifdef DEBUGGING
    GET_RE_DEBUG_FLAGS_DECL;
#endif

    PERL_ARGS_ASSERT_REGMATCH;

    DEBUG_OPTIMISE_r( DEBUG_EXECUTE_r({
	    PerlIO_printf(Perl_debug_log,"regmatch start\n");
    }));
    /* on first ever call to regmatch, allocate first slab */
    if (!PL_regmatch_slab) {
	Newx(PL_regmatch_slab, 1, regmatch_slab);
	PL_regmatch_slab->prev = NULL;
	PL_regmatch_slab->next = NULL;
	PL_regmatch_state = SLAB_FIRST(PL_regmatch_slab);
    }

    oldsave = PL_savestack_ix;
    SAVEDESTRUCTOR_X(S_clear_backtrack_stack, NULL);
    SAVEVPTR(PL_regmatch_slab);
    SAVEVPTR(PL_regmatch_state);

    /* grab next free state slot */
    st = ++PL_regmatch_state;
    if (st >  SLAB_LAST(PL_regmatch_slab))
	st = PL_regmatch_state = S_push_slab(aTHX);

    /* Note that nextchr is a byte even in UTF */
    nextchr = UCHARAT(locinput);
    scan = prog;
    while (scan != NULL) {

        DEBUG_EXECUTE_r( {
	    SV * const prop = sv_newmortal();
	    regnode *rnext=regnext(scan);
	    DUMP_EXEC_POS( locinput, scan, utf8_target );
	    regprop(rex, prop, scan);
            
	    PerlIO_printf(Perl_debug_log,
		    "%3"IVdf":%*s%s(%"IVdf")\n",
		    (IV)(scan - rexi->program), depth*2, "",
		    SvPVX_const(prop),
		    (PL_regkind[OP(scan)] == END || !rnext) ? 
		        0 : (IV)(rnext - rexi->program));
	});

	next = scan + NEXT_OFF(scan);
	if (next == scan)
	    next = NULL;
	state_num = OP(scan);

      reenter_switch:

	assert(PL_reglastparen == &rex->lastparen);
	assert(PL_reglastcloseparen == &rex->lastcloseparen);
	assert(PL_regoffs == rex->offs);

	switch (state_num) {
	case BOL:
	    if (locinput == PL_bostr)
	    {
		/* reginfo->till = reginfo->bol; */
		break;
	    }
	    sayNO;
	case MBOL:
	    if (locinput == PL_bostr ||
		((nextchr || locinput < PL_regeol) && locinput[-1] == '\n'))
	    {
		break;
	    }
	    sayNO;
	case SBOL:
	    if (locinput == PL_bostr)
		break;
	    sayNO;
	case GPOS:
	    if (locinput == reginfo->ganch)
		break;
	    sayNO;

	case KEEPS:
	    /* update the startpoint */
	    st->u.keeper.val = PL_regoffs[0].start;
	    PL_reginput = locinput;
	    PL_regoffs[0].start = locinput - PL_bostr;
	    PUSH_STATE_GOTO(KEEPS_next, next);
	    /*NOT-REACHED*/
	case KEEPS_next_fail:
	    /* rollback the start point change */
	    PL_regoffs[0].start = st->u.keeper.val;
	    sayNO_SILENT;
	    /*NOT-REACHED*/
	case EOL:
		goto seol;
	case MEOL:
	    if ((nextchr || locinput < PL_regeol) && nextchr != '\n')
		sayNO;
	    break;
	case SEOL:
	  seol:
	    if ((nextchr || locinput < PL_regeol) && nextchr != '\n')
		sayNO;
	    if (PL_regeol - locinput > 1)
		sayNO;
	    break;
	case EOS:
	    if (PL_regeol != locinput)
		sayNO;
	    break;
	case SANY:
	    if (!nextchr && locinput >= PL_regeol)
		sayNO;
 	    if (utf8_target) {
	        locinput += PL_utf8skip[nextchr];
		if (locinput > PL_regeol)
 		    sayNO;
 		nextchr = UCHARAT(locinput);
 	    }
 	    else
 		nextchr = UCHARAT(++locinput);
	    break;
	case CANY:
	    if (!nextchr && locinput >= PL_regeol)
		sayNO;
	    nextchr = UCHARAT(++locinput);
	    break;
	case REG_ANY:
	    if ((!nextchr && locinput >= PL_regeol) || nextchr == '\n')
		sayNO;
	    if (utf8_target) {
		locinput += PL_utf8skip[nextchr];
		if (locinput > PL_regeol)
		    sayNO;
		nextchr = UCHARAT(locinput);
	    }
	    else
		nextchr = UCHARAT(++locinput);
	    break;

#undef  ST
#define ST st->u.trie
        case TRIEC:
            /* In this case the charclass data is available inline so
               we can fail fast without a lot of extra overhead. 
             */
            if (scan->flags == EXACT || !utf8_target) {
                if(!ANYOF_BITMAP_TEST(scan, *locinput)) {
                    DEBUG_EXECUTE_r(
                        PerlIO_printf(Perl_debug_log,
                    	          "%*s  %sfailed to match trie start class...%s\n",
                    	          REPORT_CODE_OFF+depth*2, "", PL_colors[4], PL_colors[5])
                    );
                    sayNO_SILENT;
                    /* NOTREACHED */
                }        	        
            }
            /* FALL THROUGH */
	case TRIE:
	    /* the basic plan of execution of the trie is:
	     * At the beginning, run though all the states, and
	     * find the longest-matching word. Also remember the position
	     * of the shortest matching word. For example, this pattern:
	     *    1  2 3 4    5
	     *    ab|a|x|abcd|abc
	     * when matched against the string "abcde", will generate
	     * accept states for all words except 3, with the longest
	     * matching word being 4, and the shortest being 1 (with
	     * the position being after char 1 of the string).
	     *
	     * Then for each matching word, in word order (i.e. 1,2,4,5),
	     * we run the remainder of the pattern; on each try setting
	     * the current position to the character following the word,
	     * returning to try the next word on failure.
	     *
	     * We avoid having to build a list of words at runtime by
	     * using a compile-time structure, wordinfo[].prev, which
	     * gives, for each word, the previous accepting word (if any).
	     * In the case above it would contain the mappings 1->2, 2->0,
	     * 3->0, 4->5, 5->1.  We can use this table to generate, from
	     * the longest word (4 above), a list of all words, by
	     * following the list of prev pointers; this gives us the
	     * unordered list 4,5,1,2. Then given the current word we have
	     * just tried, we can go through the list and find the
	     * next-biggest word to try (so if we just failed on word 2,
	     * the next in the list is 4).
	     *
	     * Since at runtime we don't record the matching position in
	     * the string for each word, we have to work that out for
	     * each word we're about to process. The wordinfo table holds
	     * the character length of each word; given that we recorded
	     * at the start: the position of the shortest word and its
	     * length in chars, we just need to move the pointer the
	     * difference between the two char lengths. Depending on
	     * Unicode status and folding, that's cheap or expensive.
	     *
	     * This algorithm is optimised for the case where are only a
	     * small number of accept states, i.e. 0,1, or maybe 2.
	     * With lots of accepts states, and having to try all of them,
	     * it becomes quadratic on number of accept states to find all
	     * the next words.
	     */

	    {
                /* what type of TRIE am I? (utf8 makes this contextual) */
                DECL_TRIE_TYPE(scan);

                /* what trie are we using right now */
		reg_trie_data * const trie
        	    = (reg_trie_data*)rexi->data->data[ ARG( scan ) ];
		HV * widecharmap = MUTABLE_HV(rexi->data->data[ ARG( scan ) + 1 ]);
                U32 state = trie->startstate;

        	if (trie->bitmap && trie_type != trie_utf8_fold &&
        	    !TRIE_BITMAP_TEST(trie,*locinput)
        	) {
        	    if (trie->states[ state ].wordnum) {
        	         DEBUG_EXECUTE_r(
                            PerlIO_printf(Perl_debug_log,
                        	          "%*s  %smatched empty string...%s\n",
                        	          REPORT_CODE_OFF+depth*2, "", PL_colors[4], PL_colors[5])
                        );
			if (!trie->jump)
			    break;
        	    } else {
        	        DEBUG_EXECUTE_r(
                            PerlIO_printf(Perl_debug_log,
                        	          "%*s  %sfailed to match trie start class...%s\n",
                        	          REPORT_CODE_OFF+depth*2, "", PL_colors[4], PL_colors[5])
                        );
        	        sayNO_SILENT;
        	   }
                }

            { 
		U8 *uc = ( U8* )locinput;

		STRLEN len = 0;
		STRLEN foldlen = 0;
		U8 *uscan = (U8*)NULL;
		U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
		U32 charcount = 0; /* how many input chars we have matched */
		U32 accepted = 0; /* have we seen any accepting states? */

		ST.B = next;
		ST.jump = trie->jump;
		ST.me = scan;
		ST.firstpos = NULL;
		ST.longfold = FALSE; /* char longer if folded => it's harder */
		ST.nextword = 0;

		/* fully traverse the TRIE; note the position of the
		   shortest accept state and the wordnum of the longest
		   accept state */

		while ( state && uc <= (U8*)PL_regeol ) {
                    U32 base = trie->states[ state ].trans.base;
                    UV uvc = 0;
                    U16 charid = 0;
		    U16 wordnum;
                    wordnum = trie->states[ state ].wordnum;

		    if (wordnum) { /* it's an accept state */
			if (!accepted) {
			    accepted = 1;
			    /* record first match position */
			    if (ST.longfold) {
				ST.firstpos = (U8*)locinput;
				ST.firstchars = 0;
			    }
			    else {
				ST.firstpos = uc;
				ST.firstchars = charcount;
			    }
			}
			if (!ST.nextword || wordnum < ST.nextword)
			    ST.nextword = wordnum;
			ST.topword = wordnum;
		    }

		    DEBUG_TRIE_EXECUTE_r({
		                DUMP_EXEC_POS( (char *)uc, scan, utf8_target );
			        PerlIO_printf( Perl_debug_log,
			            "%*s  %sState: %4"UVxf" Accepted: %c ",
			            2+depth * 2, "", PL_colors[4],
			            (UV)state, (accepted ? 'Y' : 'N'));
		    });

		    /* read a char and goto next state */
		    if ( base ) {
			I32 offset;
			REXEC_TRIE_READ_CHAR(trie_type, trie, widecharmap, uc,
					     uscan, len, uvc, charid, foldlen,
					     foldbuf, uniflags);
			charcount++;
			if (foldlen>0)
			    ST.longfold = TRUE;
			if (charid &&
			     ( ((offset =
			      base + charid - 1 - trie->uniquecharcount)) >= 0)

			     && ((U32)offset < trie->lasttrans)
			     && trie->trans[offset].check == state)
			{
			    state = trie->trans[offset].next;
			}
			else {
			    state = 0;
			}
			uc += len;

		    }
		    else {
			state = 0;
		    }
		    DEBUG_TRIE_EXECUTE_r(
		        PerlIO_printf( Perl_debug_log,
		            "Charid:%3x CP:%4"UVxf" After State: %4"UVxf"%s\n",
		            charid, uvc, (UV)state, PL_colors[5] );
		    );
		}
		if (!accepted)
		   sayNO;

		/* calculate total number of accept states */
		{
		    U16 w = ST.topword;
		    accepted = 0;
		    while (w) {
			w = trie->wordinfo[w].prev;
			accepted++;
		    }
		    ST.accepted = accepted;
		}

		DEBUG_EXECUTE_r(
		    PerlIO_printf( Perl_debug_log,
			"%*s  %sgot %"IVdf" possible matches%s\n",
			REPORT_CODE_OFF + depth * 2, "",
			PL_colors[4], (IV)ST.accepted, PL_colors[5] );
		);
		goto trie_first_try; /* jump into the fail handler */
	    }}
	    /* NOTREACHED */

	case TRIE_next_fail: /* we failed - try next alternative */
            if ( ST.jump) {
                REGCP_UNWIND(ST.cp);
	        for (n = *PL_reglastparen; n > ST.lastparen; n--)
		    PL_regoffs[n].end = -1;
	        *PL_reglastparen = n;
	    }
	    if (!--ST.accepted) {
	        DEBUG_EXECUTE_r({
		    PerlIO_printf( Perl_debug_log,
			"%*s  %sTRIE failed...%s\n",
			REPORT_CODE_OFF+depth*2, "", 
			PL_colors[4],
			PL_colors[5] );
		});
		sayNO_SILENT;
	    }
	    {
		/* Find next-highest word to process.  Note that this code
		 * is O(N^2) per trie run (O(N) per branch), so keep tight */
		register U16 min = 0;
		register U16 word;
		r