regexec: swap ganch setting and gofs offsetting

[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

/* At least one required character in the target string is expressible only in
 * UTF-8. */
static const char* const non_utf8_target_but_utf8_required
                = "Can't match, because target string needs to be in UTF-8\n";

#define NON_UTF8_TARGET_BUT_UTF8_REQUIRED(target) STMT_START { \
    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "%s", non_utf8_target_but_utf8_required));\
    goto target; \
} STMT_END

/*
 * 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

#include "inline_invlist.c"
#include "unicode_constants.h"

#define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)

#ifndef STATIC
#define	STATIC	static
#endif

/* Valid for non-utf8 strings: 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)   \
					      : ANYOF_BITMAP_TEST(p,*(c)))

/*
 * Forwards.
 */

#define CHR_SVLEN(sv) (utf8_target ? sv_len_utf8(sv) : SvCUR(sv))
#define CHR_DIST(a,b) (reginfo->is_utf8_target ? utf8_distance(a,b) : a - b)

#define HOPc(pos,off) \
	(char *)(reginfo->is_utf8_target \
	    ? reghop3((U8*)pos, off, \
                    (U8*)(off >= 0 ? reginfo->strend : reginfo->strbeg)) \
	    : (U8*)(pos + off))
#define HOPBACKc(pos, off) \
	(char*)(reginfo->is_utf8_target \
	    ? reghopmaybe3((U8*)pos, -off, (U8*)(reginfo->strbeg)) \
	    : (pos - off >= reginfo->strbeg)	\
		? (U8*)pos - off		\
		: NULL)

#define HOP3(pos,off,lim) (reginfo->is_utf8_target  ? reghop3((U8*)(pos), off, (U8*)(lim)) : (U8*)(pos + off))
#define HOP3c(pos,off,lim) ((char*)HOP3(pos,off,lim))


#define NEXTCHR_EOS -10 /* nextchr has fallen off the end */
#define NEXTCHR_IS_EOS (nextchr < 0)

#define SET_nextchr \
    nextchr = ((locinput < reginfo->strend) ? UCHARAT(locinput) : NEXTCHR_EOS)

#define SET_locinput(p) \
    locinput = (p);  \
    SET_nextchr


#define LOAD_UTF8_CHARCLASS(swash_ptr, property_name) STMT_START {            \
        if (!swash_ptr) {                                                     \
            U8 flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;                       \
            swash_ptr = _core_swash_init("utf8", property_name, &PL_sv_undef, \
                                         1, 0, NULL, &flags);                 \
            assert(swash_ptr);                                                \
        }                                                                     \
    } STMT_END

/* If in debug mode, we test that a known character properly matches */
#ifdef DEBUGGING
#   define LOAD_UTF8_CHARCLASS_DEBUG_TEST(swash_ptr,                          \
                                          property_name,                      \
                                          utf8_char_in_property)              \
        LOAD_UTF8_CHARCLASS(swash_ptr, property_name);                        \
        assert(swash_fetch(swash_ptr, (U8 *) utf8_char_in_property, TRUE));
#else
#   define LOAD_UTF8_CHARCLASS_DEBUG_TEST(swash_ptr,                          \
                                          property_name,                      \
                                          utf8_char_in_property)              \
        LOAD_UTF8_CHARCLASS(swash_ptr, property_name)
#endif

#define LOAD_UTF8_CHARCLASS_ALNUM() LOAD_UTF8_CHARCLASS_DEBUG_TEST(           \
                                        PL_utf8_swash_ptrs[_CC_WORDCHAR],     \
                                        swash_property_names[_CC_WORDCHAR],   \
                                        GREEK_SMALL_LETTER_IOTA_UTF8)

#define LOAD_UTF8_CHARCLASS_GCB()  /* Grapheme cluster boundaries */          \
    STMT_START {                                                              \
	LOAD_UTF8_CHARCLASS_DEBUG_TEST(PL_utf8_X_regular_begin,               \
                                       "_X_regular_begin",                    \
                                       GREEK_SMALL_LETTER_IOTA_UTF8);         \
	LOAD_UTF8_CHARCLASS_DEBUG_TEST(PL_utf8_X_extend,                      \
                                       "_X_extend",                           \
                                       COMBINING_GRAVE_ACCENT_UTF8);          \
    } STMT_END

#define PLACEHOLDER	/* Something for the preprocessor to grab onto */
/* 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)==EXACTFU || OP(rn)==EXACTFU_SS || OP(rn)==EXACTFU_TRICKYFOLD || OP(rn)==EXACTFA || 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_TEXTFU(rn)  ( OP(rn)==EXACTFU || OP(rn)==EXACTFU_SS || OP(rn)==EXACTFU_TRICKYFOLD || OP(rn) == EXACTFA)
#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 

/* These constants are for finding GCB=LV and GCB=LVT in the CLUMP regnode.
 * These are for the pre-composed Hangul syllables, which are all in a
 * contiguous block and arranged there in such a way so as to facilitate
 * alorithmic determination of their characteristics.  As such, they don't need
 * a swash, but can be determined by simple arithmetic.  Almost all are
 * GCB=LVT, but every 28th one is a GCB=LV */
#define SBASE 0xAC00    /* Start of block */
#define SCount 11172    /* Length of block */
#define TCount 28

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

static void S_setup_eval_state(pTHX_ regmatch_info *const reginfo);
static void S_cleanup_regmatch_info_aux(pTHX_ void *arg);
static regmatch_state * S_push_slab(pTHX);

#define REGCP_PAREN_ELEMS 3
#define REGCP_OTHER_ELEMS 3
#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_ const regexp *rex, I32 parenfloor, U32 maxopenparen)
{
    dVAR;
    const int retval = PL_savestack_ix;
    const int paren_elems_to_push =
                (maxopenparen - parenfloor) * REGCP_PAREN_ELEMS;
    const UV total_elems = paren_elems_to_push + REGCP_OTHER_ELEMS;
    const UV elems_shifted = total_elems << SAVE_TIGHT_SHIFT;
    I32 p;
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_REGCPPUSH;

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

    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)maxopenparen,
                   (long)parenfloor);

    SSGROW(total_elems + REGCP_FRAME_ELEMS);
    
    DEBUG_BUFFERS_r(
	if ((int)maxopenparen > (int)parenfloor)
	    PerlIO_printf(Perl_debug_log,
		"rex=0x%"UVxf" offs=0x%"UVxf": saving capture indices:\n",
		PTR2UV(rex),
		PTR2UV(rex->offs)
	    );
    );
    for (p = parenfloor+1; p <= (I32)maxopenparen;  p++) {
/* REGCP_PARENS_ELEMS are pushed per pairs of parentheses. */
	SSPUSHINT(rex->offs[p].end);
	SSPUSHINT(rex->offs[p].start);
	SSPUSHINT(rex->offs[p].start_tmp);
	DEBUG_BUFFERS_r(PerlIO_printf(Perl_debug_log,
	    "    \\%"UVuf": %"IVdf"(%"IVdf")..%"IVdf"\n",
	    (UV)p,
	    (IV)rex->offs[p].start,
	    (IV)rex->offs[p].start_tmp,
	    (IV)rex->offs[p].end
	));
    }
/* REGCP_OTHER_ELEMS are pushed in any case, parentheses or no. */
    SSPUSHINT(maxopenparen);
    SSPUSHINT(rex->lastparen);
    SSPUSHINT(rex->lastcloseparen);
    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)

#define UNWIND_PAREN(lp, lcp)               \
    for (n = rex->lastparen; n > lp; n--)   \
        rex->offs[n].end = -1;              \
    rex->lastparen = n;                     \
    rex->lastcloseparen = lcp;


STATIC void
S_regcppop(pTHX_ regexp *rex, U32 *maxopenparen_p)
{
    dVAR;
    UV i;
    U32 paren;
    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. */
    rex->lastcloseparen = SSPOPINT;
    rex->lastparen = SSPOPINT;
    *maxopenparen_p = SSPOPINT;

    i -= REGCP_OTHER_ELEMS;
    /* Now restore the parentheses context. */
    DEBUG_BUFFERS_r(
	if (i || rex->lastparen + 1 <= rex->nparens)
	    PerlIO_printf(Perl_debug_log,
		"rex=0x%"UVxf" offs=0x%"UVxf": restoring capture indices to:\n",
		PTR2UV(rex),
		PTR2UV(rex->offs)
	    );
    );
    paren = *maxopenparen_p;
    for ( ; i > 0; i -= REGCP_PAREN_ELEMS) {
	I32 tmps;
	rex->offs[paren].start_tmp = SSPOPINT;
	rex->offs[paren].start = SSPOPINT;
	tmps = SSPOPINT;
	if (paren <= rex->lastparen)
	    rex->offs[paren].end = tmps;
	DEBUG_BUFFERS_r( PerlIO_printf(Perl_debug_log,
	    "    \\%"UVuf": %"IVdf"(%"IVdf")..%"IVdf"%s\n",
	    (UV)paren,
	    (IV)rex->offs[paren].start,
	    (IV)rex->offs[paren].start_tmp,
	    (IV)rex->offs[paren].end,
	    (paren > rex->lastparen ? "(skipped)" : ""));
	);
	paren--;
    }
#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 = rex->lastparen + 1; i <= rex->nparens; i++) {
	if (i > *maxopenparen_p)
	    rex->offs[i].start = -1;
	rex->offs[i].end = -1;
	DEBUG_BUFFERS_r( PerlIO_printf(Perl_debug_log,
	    "    \\%"UVuf": %s   ..-1 undeffing\n",
	    (UV)i,
	    (i > *maxopenparen_p) ? "-1" : "  "
	));
    }
#endif
}

/* restore the parens and associated vars at savestack position ix,
 * but without popping the stack */

STATIC void
S_regcp_restore(pTHX_ regexp *rex, I32 ix, U32 *maxopenparen_p)
{
    I32 tmpix = PL_savestack_ix;
    PL_savestack_ix = ix;
    regcppop(rex, maxopenparen_p);
    PL_savestack_ix = tmpix;
}

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

STATIC bool
S_isFOO_lc(pTHX_ const U8 classnum, const U8 character)
{
    /* Returns a boolean as to whether or not 'character' is a member of the
     * Posix character class given by 'classnum' that should be equivalent to a
     * value in the typedef '_char_class_number'.
     *
     * Ideally this could be replaced by a just an array of function pointers
     * to the C library functions that implement the macros this calls.
     * However, to compile, the precise function signatures are required, and
     * these may vary from platform to to platform.  To avoid having to figure
     * out what those all are on each platform, I (khw) am using this method,
     * which adds an extra layer of function call overhead (unless the C
     * optimizer strips it away).  But we don't particularly care about
     * performance with locales anyway. */

    switch ((_char_class_number) classnum) {
        case _CC_ENUM_ALPHANUMERIC: return isALPHANUMERIC_LC(character);
        case _CC_ENUM_ALPHA:     return isALPHA_LC(character);
        case _CC_ENUM_ASCII:     return isASCII_LC(character);
        case _CC_ENUM_BLANK:     return isBLANK_LC(character);
        case _CC_ENUM_CASED:     return isLOWER_LC(character)
                                        || isUPPER_LC(character);
        case _CC_ENUM_CNTRL:     return isCNTRL_LC(character);
        case _CC_ENUM_DIGIT:     return isDIGIT_LC(character);
        case _CC_ENUM_GRAPH:     return isGRAPH_LC(character);
        case _CC_ENUM_LOWER:     return isLOWER_LC(character);
        case _CC_ENUM_PRINT:     return isPRINT_LC(character);
        case _CC_ENUM_PSXSPC:    return isPSXSPC_LC(character);
        case _CC_ENUM_PUNCT:     return isPUNCT_LC(character);
        case _CC_ENUM_SPACE:     return isSPACE_LC(character);
        case _CC_ENUM_UPPER:     return isUPPER_LC(character);
        case _CC_ENUM_WORDCHAR:  return isWORDCHAR_LC(character);
        case _CC_ENUM_XDIGIT:    return isXDIGIT_LC(character);
        default:    /* VERTSPACE should never occur in locales */
            Perl_croak(aTHX_ "panic: isFOO_lc() has an unexpected character class '%d'", classnum);
    }

    assert(0); /* NOTREACHED */
    return FALSE;
}

STATIC bool
S_isFOO_utf8_lc(pTHX_ const U8 classnum, const U8* character)
{
    /* Returns a boolean as to whether or not the (well-formed) UTF-8-encoded
     * 'character' is a member of the Posix character class given by 'classnum'
     * that should be equivalent to a value in the typedef
     * '_char_class_number'.
     *
     * This just calls isFOO_lc on the code point for the character if it is in
     * the range 0-255.  Outside that range, all characters avoid Unicode
     * rules, ignoring any locale.  So use the Unicode function if this class
     * requires a swash, and use the Unicode macro otherwise. */

    PERL_ARGS_ASSERT_ISFOO_UTF8_LC;

    if (UTF8_IS_INVARIANT(*character)) {
        return isFOO_lc(classnum, *character);
    }
    else if (UTF8_IS_DOWNGRADEABLE_START(*character)) {
        return isFOO_lc(classnum,
                        TWO_BYTE_UTF8_TO_UNI(*character, *(character + 1)));
    }

    if (classnum < _FIRST_NON_SWASH_CC) {

        /* Initialize the swash unless done already */
        if (! PL_utf8_swash_ptrs[classnum]) {
            U8 flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
            PL_utf8_swash_ptrs[classnum] = _core_swash_init("utf8",
                swash_property_names[classnum], &PL_sv_undef, 1, 0, NULL, &flags);
        }

        return cBOOL(swash_fetch(PL_utf8_swash_ptrs[classnum], (U8 *)
                                 character,
                                 TRUE /* is UTF */ ));
    }

    switch ((_char_class_number) classnum) {
        case _CC_ENUM_SPACE:
        case _CC_ENUM_PSXSPC:    return is_XPERLSPACE_high(character);

        case _CC_ENUM_BLANK:     return is_HORIZWS_high(character);
        case _CC_ENUM_XDIGIT:    return is_XDIGIT_high(character);
        case _CC_ENUM_VERTSPACE: return is_VERTWS_high(character);
        default:                 return 0;  /* Things like CNTRL are always
                                               below 256 */
    }

    assert(0); /* NOTREACHED */
    return FALSE;
}

/*
 * pregexec and friends
 */

#ifndef PERL_IN_XSUB_RE
/*
 - pregexec - match a regexp against a string
 */
I32
Perl_pregexec(pTHX_ REGEXP * const prog, char* stringarg, char *strend,
	 char *strbeg, I32 minend, SV *screamer, U32 nosave)
/* stringarg: the point in the string at which to begin matching */
/* strend:    pointer to null at end of string */
/* strbeg:    real beginning of string */
/* minend:    end of match must be >= minend bytes after stringarg. */
/* screamer:  SV being matched: only used for utf8 flag, pos() etc; string
 *            itself is accessed via the pointers above */
/* 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).
 */

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

/* 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 fewer 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 distinguish 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. */

/* args:
 * rx:     the regex to match against
 * sv:     the SV being matched: only used for utf8 flag; the string
 *         itself is accessed via the pointers below. Note that on
 *         something like an overloaded SV, SvPOK(sv) may be false
 *         and the string pointers may point to something unrelated to
 *         the SV itself.
 * strbeg: real beginning of string
 * strpos: the point in the string at which to begin matching
 * strend: pointer to the byte following the last char of the string
 * flags   currently unused; set to 0
 * data:   currently unused; set to NULL
 */

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

    PERL_ARGS_ASSERT_RE_INTUIT_START;
    PERL_UNUSED_ARG(flags);
    PERL_UNUSED_ARG(data);

    /* 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;
    }

    reginfo->is_utf8_target = cBOOL(utf8_target);
    reginfo->info_aux = NULL;
    reginfo->strbeg = strbeg;
    reginfo->strend = strend;
    reginfo->is_utf8_pat = cBOOL(RX_UTF8(rx));
    reginfo->intuit = 1;
    /* not actually used within intuit, but zero for safety anyway */
    reginfo->poscache_maxiter = 0;

    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) {
	    if (! to_byte_substr(prog)) {
                NON_UTF8_TARGET_BUT_UTF8_REQUIRED(fail);
            }
        }
	check = prog->check_substr;
    }
    if ((prog->extflags & RXf_ANCH)	/* Match at beg-of-str or after \n */
	 && !(prog->extflags & RXf_ANCH_GPOS)) /* \G isn't a BOS or \n */
    {
        ml_anch = !( (prog->extflags & RXf_ANCH_SINGLE)
		     || ( (prog->extflags & RXf_ANCH_BOL)
			  && !multiline ) );	/* Check after \n? */

	if (!ml_anch) {
	  if (    !(prog->intflags & PREGf_IMPLICIT) /* not a real BOL */
	       && (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)
              && ! multiline)   /* /m can cause \n's to match that aren't
                                   accounted for in the string max length.
                                   See [perl #115242] */
          {
	    /* 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;
        U8* start_point;
        U8* end_point;
        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 (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 && (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 relevant 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-ish only, which are dealt with in find_byclass().  */
        const U8* const str = (U8*)STRING(progi->regstclass);
        /* XXX this value could be pre-computed */
        const int cl_l = (PL_regkind[OP(progi->regstclass)] == EXACT
		    ?  (reginfo->is_utf8_pat
                        ? utf8_distance(str + STR_LEN(progi->regstclass), str)
                        : STR_LEN(progi->regstclass))
		    : 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;
		    
        if (checked_upto < s)
           checked_upto = s;
        DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "start_shift: %"IVdf" check_at: %"IVdf" s: %"IVdf" endpos: %"IVdf" checked_upto: %"IVdf"\n",
                                      (IV)start_shift, (IV)(check_at - strbeg), (IV)(s - strbeg), (IV)(endpos - strbeg), (IV)(checked_upto- strbeg)));

	t = s;
        s = find_byclass(prog, progi->regstclass, checked_upto, endpos,
                            reginfo);
	if (s) {
	    checked_upto = s;
	} else {
#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;
	    checked_upto = HOPBACKc(endpos, start_shift);
	    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "start_shift: %"IVdf" check_at: %"IVdf" endpos: %"IVdf" checked_upto: %"IVdf"\n",
                                      (IV)start_shift, (IV)(check_at - strbeg), (IV)(endpos - strbeg), (IV)(checked_upto- strbeg)));
	    /* 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 substring. */
	  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 : trie_plain) \
                              : (utf8_target ? trie_utf8_fold : trie_latin_utf8_fold))

#define REXEC_TRIE_READ_CHAR(trie_type, trie, widecharmap, uc, uscan, len, uvc, charid, foldlen, foldbuf, uniflags) \
STMT_START {                               \
    STRLEN skiplen;                                                                 \
    switch (trie_type) {                                                            \
    case trie_utf8_fold:                                                            \
        if ( foldlen>0 ) {                                                          \
            uvc = utf8n_to_uvuni( (const U8*) uscan, UTF8_MAXLEN, &len, uniflags ); \
            foldlen -= len;                                                         \
            uscan += len;                                                           \
            len=0;                                                                  \
        } else {                                                                    \
            uvc = to_utf8_fold( (const U8*) uc, foldbuf, &foldlen );                \
            len = UTF8SKIP(uc);                                                     \
            skiplen = UNISKIP( uvc );                                               \
            foldlen -= skiplen;                                                     \
            uscan = foldbuf + skiplen;                                              \
        }                                                                           \
        break;                                                                      \
    case trie_latin_utf8_fold:                                                      \
        if ( foldlen>0 ) {                                                          \
            uvc = utf8n_to_uvuni( (const U8*) uscan, UTF8_MAXLEN, &len, uniflags ); \
            foldlen -= len;                                                         \
            uscan += len;                                                           \
            len=0;                                                                  \
        } else {                                                                    \
            len = 1;                                                                \
            uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, FOLD_FLAGS_FULL);   \
            skiplen = UNISKIP( uvc );                                               \
            foldlen -= skiplen;                                                     \
            uscan = foldbuf + skiplen;                                              \
        }                                                                           \
        break;                                                                      \
    case trie_utf8:                                                                 \
        uvc = utf8n_to_uvuni( (const 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_SCAN(CoNd)                     \
STMT_START {                                              \
    while (s <= e) {                                      \
	if ( (CoNd)                                       \
	     && (ln == 1 || folder(s, pat_string, ln))    \
	     && (reginfo->intuit || regtry(reginfo, &s)) )\
	    goto got_it;                                  \
	s++;                                              \
    }                                                     \
} STMT_END

#define REXEC_FBC_UTF8_SCAN(CoDe)                     \
STMT_START {                                          \
    while (s < strend) {                              \
	CoDe                                          \
	s += UTF8SKIP(s);                             \
    }                                                 \
} 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->intuit || 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->intuit || regtry(reginfo, &s)))  \
	    goto got_it;                              \
	else                                          \
	    tmp = doevery;                            \
    }                                                 \
    else                                              \
	tmp = 1;                                      \
)

#define REXEC_FBC_TRYIT               \
if ((reginfo->intuit || 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);                            \
    }
    
#define DUMP_EXEC_POS(li,s,doutf8) \
    dump_exec_pos(li,s,(reginfo->strend),(reginfo->strbeg), \
                startpos, doutf8)


#define UTF8_NOLOAD(TEST_NON_UTF8, IF_SUCCESS, IF_FAIL) \
	tmp = (s != reginfo->strbeg) ? UCHARAT(s - 1) : '\n';                  \
	tmp = TEST_NON_UTF8(tmp);                                              \
	REXEC_FBC_UTF8_SCAN(                                                   \
	    if (tmp == ! TEST_NON_UTF8((U8) *s)) { \
		tmp = !tmp;                                                    \
		IF_SUCCESS;                                                    \
	    }                                                                  \
	    else {                                                             \
		IF_FAIL;                                                       \
	    }                                                                  \
	);                                                                     \

#define UTF8_LOAD(TeSt1_UtF8, TeSt2_UtF8, IF_SUCCESS, IF_FAIL) \
	if (s == reginfo->strbeg) {                                            \
	    tmp = '\n';                                                        \
	}                                                                      \
	else {                                                                 \
	    U8 * const r = reghop3((U8*)s, -1, (U8*)reginfo->strbeg);          \
	    tmp = utf8n_to_uvchr(r, UTF8SKIP(r), 0, UTF8_ALLOW_DEFAULT);       \
	}                                                                      \
	tmp = TeSt1_UtF8;                                                      \
	LOAD_UTF8_CHARCLASS_ALNUM();                                                                \
	REXEC_FBC_UTF8_SCAN(                                                   \
	    if (tmp == ! (TeSt2_UtF8)) { \
		tmp = !tmp;                                                    \
		IF_SUCCESS;                                                    \
	    }                                                                  \
	    else {                                                             \
		IF_FAIL;                                                       \
	    }                                                                  \
	);                                                                     \

/* The only difference between the BOUND and NBOUND cases is that
 * REXEC_FBC_TRYIT is called when matched in BOUND, and when non-matched in
 * NBOUND.  This is accomplished by passing it in either the if or else clause,
 * with the other one being empty */
#define FBC_BOUND(TEST_NON_UTF8, TEST1_UTF8, TEST2_UTF8) \
    FBC_BOUND_COMMON(UTF8_LOAD(TEST1_UTF8, TEST2_UTF8, REXEC_FBC_TRYIT, PLACEHOLDER), TEST_NON_UTF8, REXEC_FBC_TRYIT, PLACEHOLDER)

#define FBC_BOUND_NOLOAD(TEST_NON_UTF8, TEST1_UTF8, TEST2_UTF8) \
    FBC_BOUND_COMMON(UTF8_NOLOAD(TEST_NON_UTF8, REXEC_FBC_TRYIT, PLACEHOLDER), TEST_NON_UTF8, REXEC_FBC_TRYIT, PLACEHOLDER)

#define FBC_NBOUND(TEST_NON_UTF8, TEST1_UTF8, TEST2_UTF8) \
    FBC_BOUND_COMMON(UTF8_LOAD(TEST1_UTF8, TEST2_UTF8, PLACEHOLDER, REXEC_FBC_TRYIT), TEST_NON_UTF8, PLACEHOLDER, REXEC_FBC_TRYIT)

#define FBC_NBOUND_NOLOAD(TEST_NON_UTF8, TEST1_UTF8, TEST2_UTF8) \
    FBC_BOUND_COMMON(UTF8_NOLOAD(TEST_NON_UTF8, PLACEHOLDER, REXEC_FBC_TRYIT), TEST_NON_UTF8, PLACEHOLDER, REXEC_FBC_TRYIT)


/* Common to the BOUND and NBOUND cases.  Unfortunately the UTF8 tests need to
 * be passed in completely with the variable name being tested, which isn't
 * such a clean interface, but this is easier to read than it was before.  We
 * are looking for the boundary (or non-boundary between a word and non-word
 * character.  The utf8 and non-utf8 cases have the same logic, but the details
 * must be different.  Find the "wordness" of the character just prior to this
 * one, and compare it with the wordness of this one.  If they differ, we have
 * a boundary.  At the beginning of the string, pretend that the previous
 * character was a new-line */
#define FBC_BOUND_COMMON(UTF8_CODE, TEST_NON_UTF8, IF_SUCCESS, IF_FAIL) \
    if (utf8_target) {                                                         \
		UTF8_CODE \
    }                                                                          \
    else {  /* Not utf8 */                                                     \
	tmp = (s != reginfo->strbeg) ? UCHARAT(s - 1) : '\n';                  \
	tmp = TEST_NON_UTF8(tmp);                                              \
	REXEC_FBC_SCAN(                                                        \
	    if (tmp == ! TEST_NON_UTF8((U8) *s)) {                             \
		tmp = !tmp;                                                    \
		IF_SUCCESS;                                                    \
	    }                                                                  \
	    else {                                                             \
		IF_FAIL;                                                       \
	    }                                                                  \
	);                                                                     \
    }                                                                          \
    if ((!prog->minlen && tmp) && (reginfo->intuit || regtry(reginfo, &s)))           \
	goto got_it;

/* We know what class REx starts with.  Try to find this position... */
/* if reginfo->intuit, 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 *pat_string;   /* The pattern's exactish string */
    char *pat_end;	    /* ptr to end char of pat_string */
    re_fold_t folder;	/* Function for computing non-utf8 folds */
    const U8 *fold_array;   /* array for folding ords < 256 */
    STRLEN ln;
    STRLEN lnc;
    U8 c1;
    U8 c2;
    char *e;
    I32 tmp = 1;	/* Scratch variable? */
    const bool utf8_target = reginfo->is_utf8_target;
    UV utf8_fold_flags = 0;
    const bool is_utf8_pat = reginfo->is_utf8_pat;
    bool to_complement = FALSE; /* Invert the result?  Taking the xor of this
                                   with a result inverts that result, as 0^1 =
                                   1 and 1^1 = 0 */
    _char_class_number classnum;

    RXi_GET_DECL(prog,progi);

    PERL_ARGS_ASSERT_FIND_BYCLASS;

    /* We know what class it must start with. */
    switch (OP(c)) {
    case ANYOF:
    case ANYOF_SYNTHETIC:
    case ANYOF_WARN_SUPER:
        if (utf8_target) {
            REXEC_FBC_UTF8_CLASS_SCAN(
                      reginclass(prog, c, (U8*)s, utf8_target));
        }
        else {
            REXEC_FBC_CLASS_SCAN(REGINCLASS(prog, c, (U8*)s));
        }
        break;
    case CANY:
        REXEC_FBC_SCAN(
            if (tmp && (reginfo->intuit || regtry(reginfo, &s)))
                goto got_it;
            else
                tmp = doevery;
        );
        break;

    case EXACTFA:
        if (is_utf8_pat || utf8_target) {
            utf8_fold_flags = FOLDEQ_UTF8_NOMIX_ASCII;
            goto do_exactf_utf8;
        }
        fold_array = PL_fold_latin1;    /* Latin1 folds are not affected by */
        folder = foldEQ_latin1;	        /* /a, except the sharp s one which */
        goto do_exactf_non_utf8;	/* isn't dealt with by these */

    case EXACTF:
        if (utf8_target) {

            /* regcomp.c already folded this if pattern is in UTF-8 */
            utf8_fold_flags = 0;
            goto do_exactf_utf8;
        }
        fold_array = PL_fold;
        folder = foldEQ;
        goto do_exactf_non_utf8;

    case EXACTFL:
        if (is_utf8_pat || utf8_target) {
            utf8_fold_flags = FOLDEQ_UTF8_LOCALE;
            goto do_exactf_utf8;
        }
        fold_array = PL_fold_locale;
        folder = foldEQ_locale;
        goto do_exactf_non_utf8;

    case EXACTFU_SS:
        if (is_utf8_pat) {
            utf8_fold_flags = FOLDEQ_S2_ALREADY_FOLDED;
        }
        goto do_exactf_utf8;

    case EXACTFU_TRICKYFOLD:
    case EXACTFU:
        if (is_utf8_pat || utf8_target) {
            utf8_fold_flags = is_utf8_pat ? FOLDEQ_S2_ALREADY_FOLDED : 0;
            goto do_exactf_utf8;
        }

        /* Any 'ss' in the pattern should have been replaced by regcomp,
         * so we don't have to worry here about this single special case
         * in the Latin1 range */
        fold_array = PL_fold_latin1;
        folder = foldEQ_latin1;

        /* FALL THROUGH */

    do_exactf_non_utf8: /* Neither pattern nor string are UTF8, and there
                           are no glitches with fold-length differences
                           between the target string and pattern */

        /* The idea in the non-utf8 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.  c1 is the
         * first character.  c2 is its fold.  This logic will not work for
         * Unicode semantics and the german sharp ss, which hence should
         * not be compiled into a node that gets here. */
        pat_string = STRING(c);
        ln  = STR_LEN(c);	/* length to match in octets/bytes */

        /* We know that we have to match at least 'ln' bytes (which is the
         * same as characters, since not utf8).  If we have to match 3
         * characters, and there are only 2 availabe, we know without
         * trying that it will fail; so don't start a match past the
         * required minimum number from the far end */
        e = HOP3c(strend, -((I32)ln), s);

        if (reginfo->intuit && e < s) {
            e = s;			/* Due to minlen logic of intuit() */
        }

        c1 = *pat_string;
        c2 = fold_array[c1];
        if (c1 == c2) { /* If char and fold are the same */
            REXEC_FBC_EXACTISH_SCAN(*(U8*)s == c1);
        }
        else {
            REXEC_FBC_EXACTISH_SCAN(*(U8*)s == c1 || *(U8*)s == c2);
        }
        break;

    do_exactf_utf8:
    {
        unsigned expansion;

        /* If one of the operands is in utf8, we can't use the simpler folding
         * above, due to the fact that many different characters can have the
         * same fold, or portion of a fold, or different- length fold */
        pat_string = STRING(c);
        ln  = STR_LEN(c);	/* length to match in octets/bytes */
        pat_end = pat_string + ln;
        lnc = is_utf8_pat       /* length to match in characters */
                ? utf8_length((U8 *) pat_string, (U8 *) pat_end)
                : ln;

        /* We have 'lnc' characters to match in the pattern, but because of
         * multi-character folding, each character in the target can match
         * up to 3 characters (Unicode guarantees it will never exceed
         * this) if it is utf8-encoded; and up to 2 if not (based on the
         * fact that the Latin 1 folds are already determined, and the
         * only multi-char fold in that range is the sharp-s folding to
         * 'ss'.  Thus, a pattern character can match as little as 1/3 of a
         * string character.  Adjust lnc accordingly, rounding up, so that
         * if we need to match at least 4+1/3 chars, that really is 5. */
        expansion = (utf8_target) ? UTF8_MAX_FOLD_CHAR_EXPAND : 2;
        lnc = (lnc + expansion - 1) / expansion;

        /* As in the non-UTF8 case, if we have to match 3 characters, and
         * only 2 are left, it's guaranteed to fail, so don't start a
         * match that would require us to go beyond the end of the string
         */
        e = HOP3c(strend, -((I32)lnc), s);

        if (reginfo->intuit && e < s) {
            e = s;			/* Due to minlen logic of intuit() */
        }

        /* XXX Note that we could recalculate e to stop the loop earlier,
         * as the worst case expansion above will rarely be met, and as we
         * go along we would usually find that e moves further to the left.
         * This would happen only after we reached the point in the loop
         * where if there were no expansion we should fail.  Unclear if
         * worth the expense */

        while (s <= e) {
            char *my_strend= (char *)strend;
            if (foldEQ_utf8_flags(s, &my_strend, 0,  utf8_target,
                  pat_string, NULL, ln, is_utf8_pat, utf8_fold_flags)
                && (reginfo->intuit || regtry(reginfo, &s)) )
            {
                goto got_it;
            }
            s += (utf8_target) ? UTF8SKIP(s) : 1;
        }
        break;
    }
    case BOUNDL:
        RXp_MATCH_TAINTED_on(prog);
        FBC_BOUND(isWORDCHAR_LC,
                  isWORDCHAR_LC_uvchr(UNI_TO_NATIVE(tmp)),
                  isWORDCHAR_LC_utf8((U8*)s));
        break;
    case NBOUNDL:
        RXp_MATCH_TAINTED_on(prog);
        FBC_NBOUND(isWORDCHAR_LC,
                   isWORDCHAR_LC_uvchr(UNI_TO_NATIVE(tmp)),
                   isWORDCHAR_LC_utf8((U8*)s));
        break;
    case BOUND:
        FBC_BOUND(isWORDCHAR,
                  isWORDCHAR_uni(tmp),
                  cBOOL(swash_fetch(PL_utf8_swash_ptrs[_CC_WORDCHAR], (U8*)s, utf8_target)));
        break;
    case BOUNDA:
        FBC_BOUND_NOLOAD(isWORDCHAR_A,
                         isWORDCHAR_A(tmp),
                         isWORDCHAR_A((U8*)s));
        break;
    case NBOUND:
        FBC_NBOUND(isWORDCHAR,
                   isWORDCHAR_uni(tmp),
                   cBOOL(swash_fetch(PL_utf8_swash_ptrs[_CC_WORDCHAR], (U8*)s, utf8_target)));
        break;
    case NBOUNDA:
        FBC_NBOUND_NOLOAD(isWORDCHAR_A,
                          isWORDCHAR_A(tmp),
                          isWORDCHAR_A((U8*)s));
        break;
    case BOUNDU:
        FBC_BOUND(isWORDCHAR_L1,
                  isWORDCHAR_uni(tmp),
                  cBOOL(swash_fetch(PL_utf8_swash_ptrs[_CC_WORDCHAR], (U8*)s, utf8_target)));
        break;
    case NBOUNDU:
        FBC_NBOUND(isWORDCHAR_L1,
                   isWORDCHAR_uni(tmp),
                   cBOOL(swash_fetch(PL_utf8_swash_ptrs[_CC_WORDCHAR], (U8*)s, utf8_target)));
        break;
    case LNBREAK:
        REXEC_FBC_CSCAN(is_LNBREAK_utf8_safe(s, strend),
                        is_LNBREAK_latin1_safe(s, strend)
        );
        break;

    /* The argument to all the POSIX node types is the class number to pass to
     * _generic_isCC() to build a mask for searching in PL_charclass[] */

    case NPOSIXL:
        to_complement = 1;
        /* FALLTHROUGH */

    case POSIXL:
        RXp_MATCH_TAINTED_on(prog);
        REXEC_FBC_CSCAN(to_complement ^ cBOOL(isFOO_utf8_lc(FLAGS(c), (U8 *) s)),
                        to_complement ^ cBOOL(isFOO_lc(FLAGS(c), *s)));
        break;

    case NPOSIXD:
        to_complement = 1;
        /* FALLTHROUGH */

    case POSIXD:
        if (utf8_target) {
            goto posix_utf8;
        }
        goto posixa;

    case NPOSIXA:
        if (utf8_target) {
            /* The complement of something that matches only ASCII matches all
             * UTF-8 variant code points, plus everything in ASCII that isn't
             * in the class */
            REXEC_FBC_UTF8_CLASS_SCAN(! UTF8_IS_INVARIANT(*s)
                                      || ! _generic_isCC_A(*s, FLAGS(c)));
            break;
        }

        to_complement = 1;
        /* FALLTHROUGH */

    case POSIXA:
      posixa:
        /* Don't need to worry about utf8, as it can match only a single
         * byte invariant character. */
        REXEC_FBC_CLASS_SCAN(
                        to_complement ^ cBOOL(_generic_isCC_A(*s, FLAGS(c))));
        break;

    case NPOSIXU:
        to_complement = 1;
        /* FALLTHROUGH */

    case POSIXU:
        if (! utf8_target) {
            REXEC_FBC_CLASS_SCAN(to_complement ^ cBOOL(_generic_isCC(*s,
                                                                    FLAGS(c))));
        }
        else {

      posix_utf8:
            classnum = (_char_class_number) FLAGS(c);
            if (classnum < _FIRST_NON_SWASH_CC) {
                while (s < strend) {

                    /* We avoid loading in the swash as long as possible, but
                     * should we have to, we jump to a separate loop.  This
                     * extra 'if' statement is what keeps this code from being
                     * just a call to REXEC_FBC_UTF8_CLASS_SCAN() */
                    if (UTF8_IS_ABOVE_LATIN1(*s)) {
                        goto found_above_latin1;
                    }
                    if ((UTF8_IS_INVARIANT(*s)
                         && to_complement ^ cBOOL(_generic_isCC((U8) *s,
                                                                classnum)))
                        || (UTF8_IS_DOWNGRADEABLE_START(*s)
                            && to_complement ^ cBOOL(
                                _generic_isCC(TWO_BYTE_UTF8_TO_UNI(*s, *(s + 1)),
                                              classnum))))
                    {
                        if (tmp && (reginfo->intuit || regtry(reginfo, &s)))
                            goto got_it;
                        else {
                            tmp = doevery;
                        }
                    }
                    else {
                        tmp = 1;
                    }
                    s += UTF8SKIP(s);
                }
            }
            else switch (classnum) {    /* These classes are implemented as
                                           macros */
                case _CC_ENUM_SPACE: /* XXX would require separate code if we
                                        revert the change of \v matching this */
                    /* FALL THROUGH */

                case _CC_ENUM_PSXSPC:
                    REXEC_FBC_UTF8_CLASS_SCAN(
                                        to_complement ^ cBOOL(isSPACE_utf8(s)));
                    break;

                case _CC_ENUM_BLANK:
                    REXEC_FBC_UTF8_CLASS_SCAN(
                                        to_complement ^ cBOOL(isBLANK_utf8(s)));
                    break;

                case _CC_ENUM_XDIGIT:
                    REXEC_FBC_UTF8_CLASS_SCAN(
                                       to_complement ^ cBOOL(isXDIGIT_utf8(s)));
                    break;

                case _CC_ENUM_VERTSPACE:
                    REXEC_FBC_UTF8_CLASS_SCAN(
                                       to_complement ^ cBOOL(isVERTWS_utf8(s)));
                    break;

                case _CC_ENUM_CNTRL:
                    REXEC_FBC_UTF8_CLASS_SCAN(
                                        to_complement ^ cBOOL(isCNTRL_utf8(s)));
                    break;

                default:
                    Perl_croak(aTHX_ "panic: find_byclass() node %d='%s' has an unexpected character class '%d'", OP(c), PL_reg_name[OP(c)], classnum);
                    assert(0); /* NOTREACHED */
            }
        }
        break;

      found_above_latin1:   /* Here we have to load a swash to get the result
                               for the current code point */
        if (! PL_utf8_swash_ptrs[classnum]) {
            U8 flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
            PL_utf8_swash_ptrs[classnum] =
                    _core_swash_init("utf8", swash_property_names[classnum],
                                     &PL_sv_undef, 1, 0, NULL, &flags);
        }

        /* This is a copy of the loop above for swash classes, though using the
         * FBC macro instead of being expanded out.  Since we've loaded the
         * swash, we don't have to check for that each time through the loop */
        REXEC_FBC_UTF8_CLASS_SCAN(
                to_complement ^ cBOOL(_generic_utf8(
                                      classnum,
                                      s,
                                      swash_fetch(PL_utf8_swash_ptrs[classnum],
                                                  (U8 *) s, TRUE))));
        break;

    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;
                    if (foldlen || uc < (U8*)strend) {
                        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);
                        });
                    }
                    else {
                        len = 0;
                        charid = 0;
                    }


                    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->intuit || 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;
}

/* set RX_SAVED_COPY, RX_SUBBEG etc.
 * flags have same meanings as with regexec_flags() */

static void
S_reg_set_capture_string(pTHX_ REGEXP * const rx,
                            char *strbeg,
                            char *strend,
                            SV *sv,
                            U32 flags,
                            bool utf8_target)
{
    struct regexp *const prog = ReANY(rx);

    if (flags & REXEC_COPY_STR) {
#ifdef PERL_ANY_COW
        if (SvCANCOW(sv)) {
            if (DEBUG_C_TEST) {
                PerlIO_printf(Perl_debug_log,
                              "Copy on write: regexp capture, type %d\n",
                              (int) SvTYPE(sv));
            }
            /* Create a new COW SV to share the match string and store
             * in saved_copy, unless the current COW SV in saved_copy
             * is valid and suitable for our purpose */
            if ((   prog->saved_copy
                 && SvIsCOW(prog->saved_copy)
                 && SvPOKp(prog->saved_copy)
                 && SvIsCOW(sv)
                 && SvPOKp(sv)
                 && SvPVX(sv) == SvPVX(prog->saved_copy)))
            {
                /* just reuse saved_copy SV */
                if (RXp_MATCH_COPIED(prog)) {
                    Safefree(prog->subbeg);
                    RXp_MATCH_COPIED_off(prog);
                }
            }
            else {
                /* create new COW SV to share string */
                RX_MATCH_COPY_FREE(rx);
                prog->saved_copy = sv_setsv_cow(prog->saved_copy, sv);
            }
            prog->subbeg = (char *)SvPVX_const(prog->saved_copy);
            assert (SvPOKp(prog->saved_copy));
            prog->sublen  = strend - strbeg;
            prog->suboffset = 0;
            prog->subcoffset = 0;
        } else
#endif
        {
            I32 min = 0;
            I32 max = strend - strbeg;
            I32 sublen;

            if (    (flags & REXEC_COPY_SKIP_POST)
                && !(prog->extflags & RXf_PMf_KEEPCOPY) /* //p */
                && !(PL_sawampersand & SAWAMPERSAND_RIGHT)
            ) { /* don't copy $' part of string */
                U32 n = 0;
                max = -1;
                /* calculate the right-most part of the string covered
                 * by a capture. Due to look-ahead, this may be to
                 * the right of $&, so we have to scan all captures */
                while (n <= prog->lastparen) {
                    if (prog->offs[n].end > max)
                        max = prog->offs[n].end;
                    n++;
                }
                if (max == -1)
                    max = (PL_sawampersand & SAWAMPERSAND_LEFT)
                            ? prog->offs[0].start
                            : 0;
                assert(max >= 0 && max <= strend - strbeg);
            }

            if (    (flags & REXEC_COPY_SKIP_PRE)
                && !(prog->extflags & RXf_PMf_KEEPCOPY) /* //p */
                && !(PL_sawampersand & SAWAMPERSAND_LEFT)
            ) { /* don't copy $` part of string */
                U32 n = 0;
                min = max;
                /* calculate the left-most part of the string covered
                 * by a capture. Due to look-behind, this may be to
                 * the left of $&, so we have to scan all captures */
                while (min && n <= prog->lastparen) {
                    if (   prog->offs[n].start != -1
                        && prog->offs[n].start < min)
                    {
                        min = prog->offs[n].start;
                    }
                    n++;
                }
                if ((PL_sawampersand & SAWAMPERSAND_RIGHT)
                    && min >  prog->offs[0].end
                )
                    min = prog->offs[0].end;

            }

            assert(min >= 0 && min <= max && min <= strend - strbeg);
            sublen = max - min;

            if (RX_MATCH_COPIED(rx)) {
                if (sublen > prog->sublen)
                    prog->subbeg =
                            (char*)saferealloc(prog->subbeg, sublen+1);
            }
            else
                prog->subbeg = (char*)safemalloc(sublen+1);
            Copy(strbeg + min, prog->subbeg, sublen, char);
            prog->subbeg[sublen] = '\0';
            prog->suboffset = min;
            prog->sublen = sublen;
            RX_MATCH_COPIED_on(rx);
        }
        prog->subcoffset = prog->suboffset;
        if (prog->suboffset && utf8_target) {
            /* Convert byte offset to chars.
             * XXX ideally should only compute this if @-/@+
             * has been seen, a la PL_sawampersand ??? */

            /* If there's a direct correspondence between the
             * string which we're matching and the original SV,
             * then we can use the utf8 len cache associated with
             * the SV. In particular, it means that under //g,
             * sv_pos_b2u() will use the previously cached
             * position to speed up working out the new length of
             * subcoffset, rather than counting from the start of
             * the string each time. This stops
             *   $x = "\x{100}" x 1E6; 1 while $x =~ /(.)/g;
             * from going quadratic */
            if (SvPOKp(sv) && SvPVX(sv) == strbeg)
                sv_pos_b2u(sv, &(prog->subcoffset));
            else
                prog->subcoffset = utf8_length((U8*)strbeg,
                                    (U8*)(strbeg+prog->suboffset));
        }
    }
    else {
        RX_MATCH_COPY_FREE(rx);
        prog->subbeg = strbeg;
        prog->suboffset = 0;
        prog->subcoffset = 0;
        prog->sublen = strend - strbeg;
    }
}




/*
 - regexec_flags - match a regexp against a string
 */
I32
Perl_regexec_flags(pTHX_ REGEXP * const rx, char *stringarg, char *strend,
	      char *strbeg, I32 minend, SV *sv, void *data, U32 flags)
/* stringarg: the point in the string at which to begin matching */
/* strend:    pointer to null at end of string */
/* strbeg:    real beginning of string */
/* minend:    end of match must be >= minend bytes after stringarg. */
/* sv:        SV being matched: only used for utf8 flag, pos() etc; string
 *            itself is accessed via the pointers above */
/* data:      May be used for some additional optimizations.
              Currently unused. */
/* nosave:    For optimizations. */

{
    dVAR;
    struct regexp *const prog = ReANY(rx);
    char *s;
    regnode *c;
    char *startpos;
    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 */
    const bool utf8_target = cBOOL(DO_UTF8(sv));
    I32 multiline;
    RXi_GET_DECL(prog,progi);
    regmatch_info reginfo_buf;  /* create some info to pass to regtry etc */
    regmatch_info *const reginfo = &reginfo_buf;
    regexp_paren_pair *swap = NULL;
    I32 oldsave;
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_REGEXEC_FLAGS;
    PERL_UNUSED_ARG(data);

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

    DEBUG_EXECUTE_r(
        debug_start_match(rx, utf8_target, stringarg, strend,
        "Matching");
    );

    startpos = stringarg;

    if (prog->extflags & RXf_GPOS_SEEN) {
        MAGIC *mg;

        /* set reginfo->ganch, the position where \G can match */

        reginfo->ganch =
            (flags & REXEC_IGNOREPOS)
            ? stringarg /* use start pos rather than pos() */
            : (sv && (mg = mg_find_mglob(sv)) && mg->mg_len >= 0)
            ? strbeg + mg->mg_len /* Defined pos() */
            : strbeg; /* pos() not defined; use start of string */

        DEBUG_GPOS_r(PerlIO_printf(Perl_debug_log,
            "GPOS ganch set to strbeg[%"IVdf"]\n", reginfo->ganch - strbeg));

        /* in the presence of \G, we may need to start looking earlier in
         * the string than the suggested start point of stringarg:
         * if gofs->prog is set, then that's a known, fixed minimum
         * offset, such as
         * /..\G/:   gofs = 2
         * /ab|c\G/: gofs = 1
         * or if the minimum offset isn't known, then we have to go back
         * to the start of the string, e.g. /w+\G/
         */
        if (prog->gofs) {
            if (startpos - prog->gofs < strbeg)
                startpos = strbeg;
            else
                startpos -= prog->gofs;
        }
        else if (prog->extflags & RXf_GPOS_FLOAT)
            startpos = strbeg;
    }

    minlen = prog->minlen;
    if ((startpos + minlen) > strend || startpos < strbeg) {
        DEBUG_r(PerlIO_printf(Perl_debug_log,
                    "Regex match can't succeed, so not even tried\n"));
        return 0;
    }

    /* at the end of this function, we'll do a LEAVE_SCOPE(oldsave),
     * which will call destuctors to reset PL_regmatch_state, free higher
     * PL_regmatch_slabs, and clean up regmatch_info_aux and
     * regmatch_info_aux_eval */

    oldsave = PL_savestack_ix;

    s = startpos;

    if ((prog->extflags & RXf_USE_INTUIT)
        && !(flags & REXEC_CHECKED))
    {
	s = re_intuit_start(rx, sv, strbeg, startpos, strend,
                                    flags, NULL);
	if (!s)
	    return 0;

	if (prog->extflags & RXf_CHECK_ALL) {
            /* we can match based purely on the result of INTUIT.
             * Set up captures etc just for $& and $-[0]
             * (an intuit-only match wont have $1,$2,..) */
            assert(!prog->nparens);

            /* s/// doesn't like it if $& is earlier than where we asked it to
             * start searching (which can happen on something like /.\G/) */
            if (       (flags & REXEC_FAIL_ON_UNDERFLOW)
                    && (s < stringarg))
            {
                /* this should only be possible under \G */
                assert(prog->extflags & RXf_GPOS_SEEN);
                DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log,
                    "matched, but failing for REXEC_FAIL_ON_UNDERFLOW\n"));
                goto phooey;
            }

            /* match via INTUIT shouldn't have any captures.
             * Let @-, @+, $^N know */
            prog->lastparen = prog->lastcloseparen = 0;
            RX_MATCH_UTF8_set(rx, utf8_target);
            prog->offs[0].start = s - strbeg;
            prog->offs[0].end = utf8_target
                ? (char*)utf8_hop((U8*)s, prog->minlenret) - strbeg
                : s - strbeg + prog->minlenret;
            if ( !(flags & REXEC_NOT_FIRST) )
                S_reg_set_capture_string(aTHX_ rx,
                                        strbeg, strend,
                                        sv, flags, utf8_target);

	    return 1;
        }
    }

    multiline = prog->extflags & RXf_PMf_MULTILINE;
    
    if (strend - s < (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");
    }

    RX_MATCH_TAINTED_off(rx);

    reginfo->prog = rx;	 /* Yes, sorry that this is confusing.  */
    reginfo->intuit = 0;
    reginfo->is_utf8_target = cBOOL(utf8_target);
    reginfo->is_utf8_pat = cBOOL(RX_UTF8(rx));
    reginfo->warned = FALSE;
    reginfo->strbeg  = strbeg;
    reginfo->sv = sv;
    reginfo->poscache_maxiter = 0; /* not yet started a countdown */
    reginfo->strend = strend;
    /* see how far we have to get to not match where we matched before */
    reginfo->till = stringarg + minend;

    if (prog->extflags & RXf_EVAL_SEEN && SvPADTMP(sv) && !IS_PADGV(sv)) {
        /* SAVEFREESV, not sv_mortalcopy, as this SV must last until after
           S_cleanup_regmatch_info_aux has executed (registered by
           SAVEDESTRUCTOR_X below).  S_cleanup_regmatch_info_aux modifies
           magic belonging to this SV.
           Not newSVsv, either, as it does not COW.
        */
        reginfo->sv = newSV(0);
        sv_setsv(reginfo->sv, sv);
        SAVEFREESV(reginfo->sv);
    }

    /* reserve next 2 or 3 slots in PL_regmatch_state:
     * slot N+0: may currently be in use: skip it
     * slot N+1: use for regmatch_info_aux struct
     * slot N+2: use for regmatch_info_aux_eval struct if we have (?{})'s
     * slot N+3: ready for use by regmatch()
     */

    {
        regmatch_state *old_regmatch_state;
        regmatch_slab  *old_regmatch_slab;
        int i, max = (prog->extflags & RXf_EVAL_SEEN) ? 2 : 1;

        /* on first ever match, 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);
        }

        old_regmatch_state = PL_regmatch_state;
        old_regmatch_slab  = PL_regmatch_slab;

        for (i=0; i <= max; i++) {
            if (i == 1)
                reginfo->info_aux = &(PL_regmatch_state->u.info_aux);
            else if (i ==2)
                reginfo->info_aux_eval =
                reginfo->info_aux->info_aux_eval =
                            &(PL_regmatch_state->u.info_aux_eval);

            if (++PL_regmatch_state >  SLAB_LAST(PL_regmatch_slab))
                PL_regmatch_state = S_push_slab(aTHX);
        }

        /* note initial PL_regmatch_state position; at end of match we'll
         * pop back to there and free any higher slabs */

        reginfo->info_aux->old_regmatch_state = old_regmatch_state;
        reginfo->info_aux->old_regmatch_slab  = old_regmatch_slab;
        reginfo->info_aux->poscache = NULL;

        SAVEDESTRUCTOR_X(S_cleanup_regmatch_info_aux, reginfo->info_aux);

        if ((prog->extflags & RXf_EVAL_SEEN))
            S_setup_eval_state(aTHX_ reginfo);
        else
            reginfo->info_aux_eval = reginfo->info_aux->info_aux_eval = NULL;
    }

    /* If there is a "must appear" string, look for it. */

    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);
	DEBUG_BUFFERS_r(PerlIO_printf(Perl_debug_log,
	    "rex=0x%"UVxf" saving  offs: orig=0x%"UVxf" new=0x%"UVxf"\n",
	    PTR2UV(prog),
	    PTR2UV(swap),
	    PTR2UV(prog->offs)
	));
    }

    /* 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, &s))
	    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, strbeg,
                                    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, strbeg,
                                        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) { /* note it could be possible to match at the end of the string */
		    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 initialization
           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 (s <= tmp_s && 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 is_utf8_pat?) */
	char ch;
#ifdef DEBUGGING
	int did_match = 0;
#endif
	if (utf8_target) {
            if (! prog->anchored_utf8) {
                to_utf8_substr(prog);
            }
            ch = SvPVX_const(prog->anchored_utf8)[0];
	    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 {
            if (! prog->anchored_substr) {
                if (! to_byte_substr(prog)) {
                    NON_UTF8_TARGET_BUT_UTF8_REQUIRED(phooey);
                }
            }
            ch = SvPVX_const(prog->anchored_substr)[0];
	    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) {
                if (! prog->anchored_utf8) {
                    to_utf8_substr(prog);
                }
                must = prog->anchored_utf8;
            }
            else {
                if (! prog->anchored_substr) {
                    if (! to_byte_substr(prog)) {
                        NON_UTF8_TARGET_BUT_UTF8_REQUIRED(phooey);
                    }
                }
                must = prog->anchored_substr;
            }
	    back_max = back_min = prog->anchored_offset;
	} else {
	    if (utf8_target) {
                if (! prog->float_utf8) {
                    to_utf8_substr(prog);
                }
                must = prog->float_utf8;
            }
            else {
                if (! prog->float_substr) {
                    if (! to_byte_substr(prog)) {
                        NON_UTF8_TARGET_BUT_UTF8_REQUIRED(phooey);
                    }
                }
                must = prog->float_substr;
            }
	    back_max = prog->float_max_offset;
	    back_min = prog->float_min_offset;
	}
	    
        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 > reginfo->strbeg)
	    last1 = HOPc(s, -1);
	else
	    last1 = s - 1;	/* bogus */

	/* XXXX check_substr already used to find "s", can optimize if
	   check_substr==must. */
	dontbother = end_shift;
	strend = HOPc(strend, -dontbother);
	while ( (s <= last) &&
		(s = fbm_instr((unsigned char*)HOP3(s, back_min, (back_min<0 ? strbeg : strend)),
				  (unsigned char*)strend, must,
				  multiline ? FBMrf_MULTILINE : 0)) ) {
	    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 >= reginfo->strbeg)
                                    ? 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;
                    if (s >= last1) {
                        s++; /* to break out of outer loop */
                        break;
                    }
                    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= NULL;
	    SV* float_real;
	    STRLEN len;
	    const char *little;

	    if (utf8_target) {
                if (! prog->float_utf8) {
                    to_utf8_substr(prog);
                }
                float_real = prog->float_utf8;
            }
            else {
                if (! prog->float_substr) {
                    if (! to_byte_substr(prog)) {
                        NON_UTF8_TARGET_BUT_UTF8_REQUIRED(phooey);
                    }
                }
                float_real = prog->float_substr;
            }

            little = SvPV_const(float_real, len);
	    if (SvTAIL(float_real)) {
                    /* This means that float_real contains an artificial \n on
                     * the end due to the presence of something like this:
                     * /foo$/ where we can match both "foo" and "foo\n" at the
                     * end of the string.  So we have to compare the end of the
                     * string first against the float_real without the \n and
                     * then against the full float_real with the string.  We
                     * have to watch out for cases where the string might be
                     * smaller than the float_real or the float_real without
                     * the \n. */
		    char *checkpos= strend - len;
		    DEBUG_OPTIMISE_r(
			PerlIO_printf(Perl_debug_log,
			    "%sChecking for float_real.%s\n",
			    PL_colors[4], PL_colors[5]));
		    if (checkpos + 1 < strbeg) {
                        /* can't match, even if we remove the trailing \n
                         * string is too short to match */
			DEBUG_EXECUTE_r(
			    PerlIO_printf(Perl_debug_log,
				"%sString shorter than required trailing substring, cannot match.%s\n",
				PL_colors[4], PL_colors[5]));
			goto phooey;
		    } else if (memEQ(checkpos + 1, little, len - 1)) {
                        /* can match, the end of the string matches without the
                         * "\n" */
			last = checkpos + 1;
		    } else if (checkpos < strbeg) {
                        /* cant match, string is too short when the "\n" is
                         * included */
			DEBUG_EXECUTE_r(
			    PerlIO_printf(Perl_debug_log,
				"%sString does not contain required trailing substring, cannot match.%s\n",
				PL_colors[4], PL_colors[5]));
			goto phooey;
		    } else if (!multiline) {
                        /* non multiline match, so compare with the "\n" at the
                         * end of the string */
			if (memEQ(checkpos, little, len)) {
			    last= checkpos;
			} else {
			    DEBUG_EXECUTE_r(
				PerlIO_printf(Perl_debug_log,
				    "%sString does not contain required trailing substring, cannot match.%s\n",
				    PL_colors[4], PL_colors[5]));
			    goto phooey;
			}
		    } else {
                        /* multiline match, so we have to search for a place
                         * where the full string is located */
			goto find_last;
		    }
	    } else {
		  find_last:
		    if (len)
			last = rninstr(s, strend, little, little + len);
		    else
			last = strend;	/* matching "$" */
	    }
	    if (!last) {
                /* at one point this block contained a comment which was
                 * probably incorrect, which said that this was a "should not
                 * happen" case.  Even if it was true when it was written I am
                 * pretty sure it is not anymore, so I have removed the comment
                 * and replaced it with this one. Yves */
		DEBUG_EXECUTE_r(
		    PerlIO_printf(Perl_debug_log,
			"String does not contain required substring, cannot match.\n"
	            ));
		goto phooey;
	    }
	    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:
    /* s/// doesn't like it if $& is earlier than where we asked it to
     * start searching (which can happen on something like /.\G/) */
    if (       (flags & REXEC_FAIL_ON_UNDERFLOW)
            && (prog->offs[0].start < stringarg - strbeg))
    {
        /* this should only be possible under \G */
        assert(prog->extflags & RXf_GPOS_SEEN);
        DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log,
            "matched, but failing for REXEC_FAIL_ON_UNDERFLOW\n"));
        goto phooey;
    }

    DEBUG_BUFFERS_r(
	if (swap)
	    PerlIO_printf(Perl_debug_log,
		"rex=0x%"UVxf" freeing offs: 0x%"UVxf"\n",
		PTR2UV(prog),
		PTR2UV(swap)
	    );
    );
    Safefree(swap);

    /* clean up; this will trigger destructors that will free all slabs
     * above the current one, and cleanup the regmatch_info_aux
     * and regmatch_info_aux_eval sructs */

    LEAVE_SCOPE(oldsave);

    if (RXp_PAREN_NAMES(prog)) 
        (void)hv_iterinit(RXp_PAREN_NAMES(prog));

    RX_MATCH_UTF8_set(rx, utf8_target);

    /* make sure $`, $&, $', and $digit will work later */
    if ( !(flags & REXEC_NOT_FIRST) )
        S_reg_set_capture_string(aTHX_ rx,
                                    strbeg, reginfo->strend,
                                    sv, flags, utf8_target);

    return 1;

phooey:
    DEBUG_EXECUTE_r(PerlIO_printf(Perl_debug_log, "%sMatch failed%s\n",
			  PL_colors[4], PL_colors[5]));

    /* clean up; this will trigger destructors that will free all slabs
     * above the current one, and cleanup the regmatch_info_aux
     * and regmatch_info_aux_eval sructs */

    LEAVE_SCOPE(oldsave);

    if (swap) {
        /* we failed :-( roll it back */
	DEBUG_BUFFERS_r(PerlIO_printf(Perl_debug_log,
	    "rex=0x%"UVxf" rolling back offs: freeing=0x%"UVxf" restoring=0x%"UVxf"\n",
	    PTR2UV(prog),
	    PTR2UV(prog->offs),
	    PTR2UV(swap)
	));
        Safefree(prog->offs);
        prog->offs = swap;
    }
    return 0;
}


/* Set which rex is pointed to by PL_reg_curpm, handling ref counting.
 * Do inc before dec, in case old and new rex are the same */
#define SET_reg_curpm(Re2) \
    if (reginfo->info_aux_eval) {                   \
	(void)ReREFCNT_inc(Re2);		    \
	ReREFCNT_dec(PM_GETRE(PL_reg_curpm));	    \
	PM_SETRE((PL_reg_curpm), (Re2));	    \
    }


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

    PERL_ARGS_ASSERT_REGTRY;

    reginfo->cutpoint=NULL;

    prog->offs[0].start = *startposp - reginfo->strbeg;
    prog->lastparen = 0;
    prog->lastcloseparen = 0;

    /* XXXX What this code is doing here?!!!  There should be no need
       to do this again and again, prog->lastparen 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
     * prog->lastparen), 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 = prog->offs;
	I32 i;
	for (i = prog->nparens; i > (I32)prog->lastparen; i--) {
	    ++pp;
	    pp->start = -1;
	    pp->end = -1;
	}
    }
#endif
    REGCP_SET(lastcp);
    result = regmatch(reginfo, *startposp, progi->program + 1);
    if (result != -1) {
	prog->offs[0].end = result;
	return 1;
    }
    if (reginfo->cutpoint)
        *startposp= 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) \
       reginfo->info_aux->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 CHRTEST_NOT_A_CP_1 -999
#define CHRTEST_NOT_A_CP_2 -998

/* 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, input) \
    pushinput = input; \
    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, input) \
    pushinput = input; \
    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, newinput);
	// 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, newinput);
	PUSH_YES_STATE_GOTO(resume_state, node, newinput);

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.
 
Backtrack 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)rex->lastparen >= nums[n] &&
            rex->offs[nums[n]].end != -1)
        {
            return nums[n];
        }
    }
    return 0;
}


static bool
S_setup_EXACTISH_ST_c1_c2(pTHX_ const regnode * const text_node, int *c1p,
        U8* c1_utf8, int *c2p, U8* c2_utf8, regmatch_info *reginfo)
{
    /* This function determines if there are one or two characters that match
     * the first character of the passed-in EXACTish node <text_node>, and if
     * so, returns them in the passed-in pointers.
     *
     * If it determines that no possible character in the target string can
     * match, it returns FALSE; otherwise TRUE.  (The FALSE situation occurs if
     * the first character in <text_node> requires UTF-8 to represent, and the
     * target string isn't in UTF-8.)
     *
     * If there are more than two characters that could match the beginning of
     * <text_node>, or if more context is required to determine a match or not,
     * it sets both *<c1p> and *<c2p> to CHRTEST_VOID.
     *
     * The motiviation behind this function is to allow the caller to set up
     * tight loops for matching.  If <text_node> is of type EXACT, there is
     * only one possible character that can match its first character, and so
     * the situation is quite simple.  But things get much more complicated if
     * folding is involved.  It may be that the first character of an EXACTFish
     * node doesn't participate in any possible fold, e.g., punctuation, so it
     * can be matched only by itself.  The vast majority of characters that are
     * in folds match just two things, their lower and upper-case equivalents.
     * But not all are like that; some have multiple possible matches, or match
     * sequences of more than one character.  This function sorts all that out.
     *
     * Consider the patterns A*B or A*?B where A and B are arbitrary.  In a
     * loop of trying to match A*, we know we can't exit where the thing
     * following it isn't a B.  And something can't be a B unless it is the
     * beginning of B.  By putting a quick test for that beginning in a tight
     * loop, we can rule out things that can't possibly be B without having to
     * break out of the loop, thus avoiding work.  Similarly, if A is a single
     * character, we can make a tight loop matching A*, using the outputs of
     * this function.
     *
     * If the target string to match isn't in UTF-8, and there aren't
     * complications which require CHRTEST_VOID, *<c1p> and *<c2p> are set to
     * the one or two possible octets (which are characters in this situation)
     * that can match.  In all cases, if there is only one character that can
     * match, *<c1p> and *<c2p> will be identical.
     *
     * If the target string is in UTF-8, the buffers pointed to by <c1_utf8>
     * and <c2_utf8> will contain the one or two UTF-8 sequences of bytes that
     * can match the beginning of <text_node>.  They should be declared with at
     * least length UTF8_MAXBYTES+1.  (If the target string isn't in UTF-8, it is
     * undefined what these contain.)  If one or both of the buffers are
     * invariant under UTF-8, *<c1p>, and *<c2p> will also be set to the
     * corresponding invariant.  If variant, the corresponding *<c1p> and/or
     * *<c2p> will be set to a negative number(s) that shouldn't match any code
     * point (unless inappropriately coerced to unsigned).   *<c1p> will equal
     * *<c2p> if and only if <c1_utf8> and <c2_utf8> are the same. */

    const bool utf8_target = reginfo->is_utf8_target;

    UV c1 = CHRTEST_NOT_A_CP_1;
    UV c2 = CHRTEST_NOT_A_CP_2;
    bool use_chrtest_void = FALSE;
    const bool is_utf8_pat = reginfo->is_utf8_pat;

    /* Used when we have both utf8 input and utf8 output, to avoid converting
     * to/from code points */
    bool utf8_has_been_setup = FALSE;

    dVAR;

    U8 *pat = (U8*)STRING(text_node);

    if (OP(text_node) == EXACT) {

        /* In an exact node, only one thing can be matched, that first
         * character.  If both the pat and the target are UTF-8, we can just
         * copy the input to the output, avoiding finding the code point of
         * that character */
        if (!is_utf8_pat) {
            c2 = c1 = *pat;
        }
        else if (utf8_target) {
            Copy(pat, c1_utf8, UTF8SKIP(pat), U8);
            Copy(pat, c2_utf8, UTF8SKIP(pat), U8);
            utf8_has_been_setup = TRUE;
        }
        else {
            c2 = c1 = valid_utf8_to_uvchr(pat, NULL);
        }
    }
    else /* an EXACTFish node */
         if ((is_utf8_pat
                    && is_MULTI_CHAR_FOLD_utf8_safe(pat,
                                                    pat + STR_LEN(text_node)))
             || (!is_utf8_pat
                    && is_MULTI_CHAR_FOLD_latin1_safe(pat,
                                                    pat + STR_LEN(text_node))))
    {
        /* Multi-character folds require more context to sort out.  Also
         * PL_utf8_foldclosures used below doesn't handle them, so have to be
         * handled outside this routine */
        use_chrtest_void = TRUE;
    }
    else { /* an EXACTFish node which doesn't begin with a multi-char fold */
        c1 = is_utf8_pat ? valid_utf8_to_uvchr(pat, NULL) : *pat;
        if (c1 > 256) {
            /* Load the folds hash, if not already done */
            SV** listp;
            if (! PL_utf8_foldclosures) {
                if (! PL_utf8_tofold) {
                    U8 dummy[UTF8_MAXBYTES+1];

                    /* Force loading this by folding an above-Latin1 char */
                    to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
                    assert(PL_utf8_tofold); /* Verify that worked */
                }
                PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
            }

            /* The fold closures data structure is a hash with the keys being
             * the UTF-8 of every character that is folded to, like 'k', and
             * the values each an array of all code points that fold to its
             * key.  e.g. [ 'k', 'K', KELVIN_SIGN ].  Multi-character folds are
             * not included */
            if ((! (listp = hv_fetch(PL_utf8_foldclosures,
                                     (char *) pat,
                                     UTF8SKIP(pat),
                                     FALSE))))
            {
                /* Not found in the hash, therefore there are no folds
                 * containing it, so there is only a single character that
                 * could match */
                c2 = c1;
            }
            else {  /* Does participate in folds */
                AV* list = (AV*) *listp;
                if (av_len(list) != 1) {

                    /* If there aren't exactly two folds to this, it is outside
                     * the scope of this function */
                    use_chrtest_void = TRUE;
                }
                else {  /* There are two.  Get them */
                    SV** c_p = av_fetch(list, 0, FALSE);
                    if (c_p == NULL) {
                        Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
                    }
                    c1 = SvUV(*c_p);

                    c_p = av_fetch(list, 1, FALSE);
                    if (c_p == NULL) {
                        Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
                    }
                    c2 = SvUV(*c_p);

                    /* Folds that cross the 255/256 boundary are forbidden if
                     * EXACTFL, or EXACTFA and one is ASCIII.  Since the
                     * pattern character is above 256, and its only other match
                     * is below 256, the only legal match will be to itself.
                     * We have thrown away the original, so have to compute
                     * which is the one above 255 */
                    if ((c1 < 256) != (c2 < 256)) {
                        if (OP(text_node) == EXACTFL
                            || (OP(text_node) == EXACTFA
                                && (isASCII(c1) || isASCII(c2))))
                        {
                            if (c1 < 256) {
                                c1 = c2;
                            }
                            else {
                                c2 = c1;
                            }
                        }
                    }
                }
            }
        }
        else /* Here, c1 is < 255 */
             if (utf8_target
                 && HAS_NONLATIN1_FOLD_CLOSURE(c1)
                 && OP(text_node) != EXACTFL
                 && (OP(text_node) != EXACTFA || ! isASCII(c1)))
        {
            /* Here, there could be something above Latin1 in the target which
             * folds to this character in the pattern.  All such cases except
             * LATIN SMALL LETTER Y WITH DIAERESIS have more than two characters
             * involved in their folds, so are outside the scope of this
             * function */
            if (UNLIKELY(c1 == LATIN_SMALL_LETTER_Y_WITH_DIAERESIS)) {
                c2 = LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS;
            }
            else {
                use_chrtest_void = TRUE;
            }
        }
        else { /* Here nothing above Latin1 can fold to the pattern character */
            switch (OP(text_node)) {

                case EXACTFL:   /* /l rules */
                    c2 = PL_fold_locale[c1];
                    break;

                case EXACTF:
                    if (! utf8_target) {    /* /d rules */
                        c2 = PL_fold[c1];
                        break;
                    }
                    /* FALLTHROUGH */
                    /* /u rules for all these.  This happens to work for
                     * EXACTFA as nothing in Latin1 folds to ASCII */
                case EXACTFA:
                case EXACTFU_TRICKYFOLD:
                case EXACTFU_SS:
                case EXACTFU:
                    c2 = PL_fold_latin1[c1];
                    break;

		default:
                    Perl_croak(aTHX_ "panic: Unexpected op %u", OP(text_node));
                    assert(0); /* NOTREACHED */
            }
        }
    }

    /* Here have figured things out.  Set up the returns */
    if (use_chrtest_void) {
        *c2p = *c1p = CHRTEST_VOID;
    }
    else if (utf8_target) {
        if (! utf8_has_been_setup) {    /* Don't have the utf8; must get it */
            uvchr_to_utf8(c1_utf8, c1);
            uvchr_to_utf8(c2_utf8, c2);
        }