regcomp.c: Use new complement union

[perl5.git] / regcomp.c

/*    regcomp.c
 */

/*
 * 'A fair jaw-cracker dwarf-language must be.'            --Samwise Gamgee
 *
 *     [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
 */

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

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

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

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

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

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

 *
 * Beware that some of this code is subtly aware of the way operator
 * precedence is structured in regular expressions.  Serious changes in
 * regular-expression syntax might require a total rethink.
 */
#include "EXTERN.h"
#define PERL_IN_REGCOMP_C
#include "perl.h"

#ifndef PERL_IN_XSUB_RE
#  include "INTERN.h"
#endif

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

#include "dquote_static.c"

#ifdef op
#undef op
#endif /* op */

#ifdef MSDOS
#  if defined(BUGGY_MSC6)
 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
#    pragma optimize("a",off)
 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
#    pragma optimize("w",on )
#  endif /* BUGGY_MSC6 */
#endif /* MSDOS */

#ifndef STATIC
#define	STATIC	static
#endif

typedef struct RExC_state_t {
    U32		flags;			/* are we folding, multilining? */
    char	*precomp;		/* uncompiled string. */
    REGEXP	*rx_sv;			/* The SV that is the regexp. */
    regexp	*rx;                    /* perl core regexp structure */
    regexp_internal	*rxi;           /* internal data for regexp object pprivate field */        
    char	*start;			/* Start of input for compile */
    char	*end;			/* End of input for compile */
    char	*parse;			/* Input-scan pointer. */
    I32		whilem_seen;		/* number of WHILEM in this expr */
    regnode	*emit_start;		/* Start of emitted-code area */
    regnode	*emit_bound;		/* First regnode outside of the allocated space */
    regnode	*emit;			/* Code-emit pointer; &regdummy = don't = compiling */
    I32		naughty;		/* How bad is this pattern? */
    I32		sawback;		/* Did we see \1, ...? */
    U32		seen;
    I32		size;			/* Code size. */
    I32		npar;			/* Capture buffer count, (OPEN). */
    I32		cpar;			/* Capture buffer count, (CLOSE). */
    I32		nestroot;		/* root parens we are in - used by accept */
    I32		extralen;
    I32		seen_zerolen;
    I32		seen_evals;
    regnode	**open_parens;		/* pointers to open parens */
    regnode	**close_parens;		/* pointers to close parens */
    regnode	*opend;			/* END node in program */
    I32		utf8;		/* whether the pattern is utf8 or not */
    I32		orig_utf8;	/* whether the pattern was originally in utf8 */
				/* XXX use this for future optimisation of case
				 * where pattern must be upgraded to utf8. */
    I32		uni_semantics;	/* If a d charset modifier should use unicode
				   rules, even if the pattern is not in
				   utf8 */
    HV		*paren_names;		/* Paren names */
    
    regnode	**recurse;		/* Recurse regops */
    I32		recurse_count;		/* Number of recurse regops */
    I32		in_lookbehind;
    I32		contains_locale;
    I32		override_recoding;
#if ADD_TO_REGEXEC
    char 	*starttry;		/* -Dr: where regtry was called. */
#define RExC_starttry	(pRExC_state->starttry)
#endif
#ifdef DEBUGGING
    const char  *lastparse;
    I32         lastnum;
    AV          *paren_name_list;       /* idx -> name */
#define RExC_lastparse	(pRExC_state->lastparse)
#define RExC_lastnum	(pRExC_state->lastnum)
#define RExC_paren_name_list    (pRExC_state->paren_name_list)
#endif
} RExC_state_t;

#define RExC_flags	(pRExC_state->flags)
#define RExC_precomp	(pRExC_state->precomp)
#define RExC_rx_sv	(pRExC_state->rx_sv)
#define RExC_rx		(pRExC_state->rx)
#define RExC_rxi	(pRExC_state->rxi)
#define RExC_start	(pRExC_state->start)
#define RExC_end	(pRExC_state->end)
#define RExC_parse	(pRExC_state->parse)
#define RExC_whilem_seen	(pRExC_state->whilem_seen)
#ifdef RE_TRACK_PATTERN_OFFSETS
#define RExC_offsets	(pRExC_state->rxi->u.offsets) /* I am not like the others */
#endif
#define RExC_emit	(pRExC_state->emit)
#define RExC_emit_start	(pRExC_state->emit_start)
#define RExC_emit_bound	(pRExC_state->emit_bound)
#define RExC_naughty	(pRExC_state->naughty)
#define RExC_sawback	(pRExC_state->sawback)
#define RExC_seen	(pRExC_state->seen)
#define RExC_size	(pRExC_state->size)
#define RExC_npar	(pRExC_state->npar)
#define RExC_nestroot   (pRExC_state->nestroot)
#define RExC_extralen	(pRExC_state->extralen)
#define RExC_seen_zerolen	(pRExC_state->seen_zerolen)
#define RExC_seen_evals	(pRExC_state->seen_evals)
#define RExC_utf8	(pRExC_state->utf8)
#define RExC_uni_semantics	(pRExC_state->uni_semantics)
#define RExC_orig_utf8	(pRExC_state->orig_utf8)
#define RExC_open_parens	(pRExC_state->open_parens)
#define RExC_close_parens	(pRExC_state->close_parens)
#define RExC_opend	(pRExC_state->opend)
#define RExC_paren_names	(pRExC_state->paren_names)
#define RExC_recurse	(pRExC_state->recurse)
#define RExC_recurse_count	(pRExC_state->recurse_count)
#define RExC_in_lookbehind	(pRExC_state->in_lookbehind)
#define RExC_contains_locale	(pRExC_state->contains_locale)
#define RExC_override_recoding	(pRExC_state->override_recoding)


#define	ISMULT1(c)	((c) == '*' || (c) == '+' || (c) == '?')
#define	ISMULT2(s)	((*s) == '*' || (*s) == '+' || (*s) == '?' || \
	((*s) == '{' && regcurly(s)))

#ifdef SPSTART
#undef SPSTART		/* dratted cpp namespace... */
#endif
/*
 * Flags to be passed up and down.
 */
#define	WORST		0	/* Worst case. */
#define	HASWIDTH	0x01	/* Known to match non-null strings. */

/* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
 * character, and if utf8, must be invariant.  Note that this is not the same thing as REGNODE_SIMPLE */
#define	SIMPLE		0x02
#define	SPSTART		0x04	/* Starts with * or +. */
#define TRYAGAIN	0x08	/* Weeded out a declaration. */
#define POSTPONED	0x10    /* (?1),(?&name), (??{...}) or similar */

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

/* whether trie related optimizations are enabled */
#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
#define TRIE_STUDY_OPT
#define FULL_TRIE_STUDY
#define TRIE_STCLASS
#endif



#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
#define PBITVAL(paren) (1 << ((paren) & 7))
#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))

/* If not already in utf8, do a longjmp back to the beginning */
#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
#define REQUIRE_UTF8	STMT_START {                                       \
                                     if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
                        } STMT_END

/* About scan_data_t.

  During optimisation we recurse through the regexp program performing
  various inplace (keyhole style) optimisations. In addition study_chunk
  and scan_commit populate this data structure with information about
  what strings MUST appear in the pattern. We look for the longest 
  string that must appear at a fixed location, and we look for the
  longest string that may appear at a floating location. So for instance
  in the pattern:
  
    /FOO[xX]A.*B[xX]BAR/
    
  Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
  strings (because they follow a .* construct). study_chunk will identify
  both FOO and BAR as being the longest fixed and floating strings respectively.
  
  The strings can be composites, for instance
  
     /(f)(o)(o)/
     
  will result in a composite fixed substring 'foo'.
  
  For each string some basic information is maintained:
  
  - offset or min_offset
    This is the position the string must appear at, or not before.
    It also implicitly (when combined with minlenp) tells us how many
    characters must match before the string we are searching for.
    Likewise when combined with minlenp and the length of the string it
    tells us how many characters must appear after the string we have 
    found.
  
  - max_offset
    Only used for floating strings. This is the rightmost point that
    the string can appear at. If set to I32 max it indicates that the
    string can occur infinitely far to the right.
  
  - minlenp
    A pointer to the minimum length of the pattern that the string 
    was found inside. This is important as in the case of positive 
    lookahead or positive lookbehind we can have multiple patterns 
    involved. Consider
    
    /(?=FOO).*F/
    
    The minimum length of the pattern overall is 3, the minimum length
    of the lookahead part is 3, but the minimum length of the part that
    will actually match is 1. So 'FOO's minimum length is 3, but the 
    minimum length for the F is 1. This is important as the minimum length
    is used to determine offsets in front of and behind the string being 
    looked for.  Since strings can be composites this is the length of the
    pattern at the time it was committed with a scan_commit. Note that
    the length is calculated by study_chunk, so that the minimum lengths
    are not known until the full pattern has been compiled, thus the 
    pointer to the value.
  
  - lookbehind
  
    In the case of lookbehind the string being searched for can be
    offset past the start point of the final matching string. 
    If this value was just blithely removed from the min_offset it would
    invalidate some of the calculations for how many chars must match
    before or after (as they are derived from min_offset and minlen and
    the length of the string being searched for). 
    When the final pattern is compiled and the data is moved from the
    scan_data_t structure into the regexp structure the information
    about lookbehind is factored in, with the information that would 
    have been lost precalculated in the end_shift field for the 
    associated string.

  The fields pos_min and pos_delta are used to store the minimum offset
  and the delta to the maximum offset at the current point in the pattern.    

*/

typedef struct scan_data_t {
    /*I32 len_min;      unused */
    /*I32 len_delta;    unused */
    I32 pos_min;
    I32 pos_delta;
    SV *last_found;
    I32 last_end;	    /* min value, <0 unless valid. */
    I32 last_start_min;
    I32 last_start_max;
    SV **longest;	    /* Either &l_fixed, or &l_float. */
    SV *longest_fixed;      /* longest fixed string found in pattern */
    I32 offset_fixed;       /* offset where it starts */
    I32 *minlen_fixed;      /* pointer to the minlen relevant to the string */
    I32 lookbehind_fixed;   /* is the position of the string modfied by LB */
    SV *longest_float;      /* longest floating string found in pattern */
    I32 offset_float_min;   /* earliest point in string it can appear */
    I32 offset_float_max;   /* latest point in string it can appear */
    I32 *minlen_float;      /* pointer to the minlen relevant to the string */
    I32 lookbehind_float;   /* is the position of the string modified by LB */
    I32 flags;
    I32 whilem_c;
    I32 *last_closep;
    struct regnode_charclass_class *start_class;
} scan_data_t;

/*
 * Forward declarations for pregcomp()'s friends.
 */

static const scan_data_t zero_scan_data =
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};

#define SF_BEFORE_EOL		(SF_BEFORE_SEOL|SF_BEFORE_MEOL)
#define SF_BEFORE_SEOL		0x0001
#define SF_BEFORE_MEOL		0x0002
#define SF_FIX_BEFORE_EOL	(SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
#define SF_FL_BEFORE_EOL	(SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)

#ifdef NO_UNARY_PLUS
#  define SF_FIX_SHIFT_EOL	(0+2)
#  define SF_FL_SHIFT_EOL		(0+4)
#else
#  define SF_FIX_SHIFT_EOL	(+2)
#  define SF_FL_SHIFT_EOL		(+4)
#endif

#define SF_FIX_BEFORE_SEOL	(SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
#define SF_FIX_BEFORE_MEOL	(SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)

#define SF_FL_BEFORE_SEOL	(SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
#define SF_FL_BEFORE_MEOL	(SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
#define SF_IS_INF		0x0040
#define SF_HAS_PAR		0x0080
#define SF_IN_PAR		0x0100
#define SF_HAS_EVAL		0x0200
#define SCF_DO_SUBSTR		0x0400
#define SCF_DO_STCLASS_AND	0x0800
#define SCF_DO_STCLASS_OR	0x1000
#define SCF_DO_STCLASS		(SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
#define SCF_WHILEM_VISITED_POS	0x2000

#define SCF_TRIE_RESTUDY        0x4000 /* Do restudy? */
#define SCF_SEEN_ACCEPT         0x8000 

#define UTF cBOOL(RExC_utf8)
#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
#define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)

#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)

#define OOB_UNICODE		12345678
#define OOB_NAMEDCLASS		-1

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


/* length of regex to show in messages that don't mark a position within */
#define RegexLengthToShowInErrorMessages 127

/*
 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
 * op/pragma/warn/regcomp.
 */
#define MARKER1 "<-- HERE"    /* marker as it appears in the description */
#define MARKER2 " <-- HERE "  /* marker as it appears within the regex */

#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"

/*
 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
 * arg. Show regex, up to a maximum length. If it's too long, chop and add
 * "...".
 */
#define _FAIL(code) STMT_START {					\
    const char *ellipses = "";						\
    IV len = RExC_end - RExC_precomp;					\
									\
    if (!SIZE_ONLY)							\
	SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv);			\
    if (len > RegexLengthToShowInErrorMessages) {			\
	/* chop 10 shorter than the max, to ensure meaning of "..." */	\
	len = RegexLengthToShowInErrorMessages - 10;			\
	ellipses = "...";						\
    }									\
    code;                                                               \
} STMT_END

#define	FAIL(msg) _FAIL(			    \
    Perl_croak(aTHX_ "%s in regex m/%.*s%s/",	    \
	    msg, (int)len, RExC_precomp, ellipses))

#define	FAIL2(msg,arg) _FAIL(			    \
    Perl_croak(aTHX_ msg " in regex m/%.*s%s/",	    \
	    arg, (int)len, RExC_precomp, ellipses))

/*
 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
 */
#define	Simple_vFAIL(m) STMT_START {					\
    const IV offset = RExC_parse - RExC_precomp;			\
    Perl_croak(aTHX_ "%s" REPORT_LOCATION,				\
	    m, (int)offset, RExC_precomp, RExC_precomp + offset);	\
} STMT_END

/*
 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
 */
#define	vFAIL(m) STMT_START {				\
    if (!SIZE_ONLY)					\
	SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv);	\
    Simple_vFAIL(m);					\
} STMT_END

/*
 * Like Simple_vFAIL(), but accepts two arguments.
 */
#define	Simple_vFAIL2(m,a1) STMT_START {			\
    const IV offset = RExC_parse - RExC_precomp;			\
    S_re_croak2(aTHX_ m, REPORT_LOCATION, a1,			\
	    (int)offset, RExC_precomp, RExC_precomp + offset);	\
} STMT_END

/*
 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
 */
#define	vFAIL2(m,a1) STMT_START {			\
    if (!SIZE_ONLY)					\
	SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv);	\
    Simple_vFAIL2(m, a1);				\
} STMT_END


/*
 * Like Simple_vFAIL(), but accepts three arguments.
 */
#define	Simple_vFAIL3(m, a1, a2) STMT_START {			\
    const IV offset = RExC_parse - RExC_precomp;		\
    S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2,		\
	    (int)offset, RExC_precomp, RExC_precomp + offset);	\
} STMT_END

/*
 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
 */
#define	vFAIL3(m,a1,a2) STMT_START {			\
    if (!SIZE_ONLY)					\
	SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv);	\
    Simple_vFAIL3(m, a1, a2);				\
} STMT_END

/*
 * Like Simple_vFAIL(), but accepts four arguments.
 */
#define	Simple_vFAIL4(m, a1, a2, a3) STMT_START {		\
    const IV offset = RExC_parse - RExC_precomp;		\
    S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3,		\
	    (int)offset, RExC_precomp, RExC_precomp + offset);	\
} STMT_END

#define	ckWARNreg(loc,m) STMT_START {					\
    const IV offset = loc - RExC_precomp;				\
    Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION,	\
	    (int)offset, RExC_precomp, RExC_precomp + offset);		\
} STMT_END

#define	ckWARNregdep(loc,m) STMT_START {				\
    const IV offset = loc - RExC_precomp;				\
    Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP),	\
	    m REPORT_LOCATION,						\
	    (int)offset, RExC_precomp, RExC_precomp + offset);		\
} STMT_END

#define	ckWARN2regdep(loc,m, a1) STMT_START {				\
    const IV offset = loc - RExC_precomp;				\
    Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP),	\
	    m REPORT_LOCATION,						\
	    a1, (int)offset, RExC_precomp, RExC_precomp + offset);	\
} STMT_END

#define	ckWARN2reg(loc, m, a1) STMT_START {				\
    const IV offset = loc - RExC_precomp;				\
    Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION,	\
	    a1, (int)offset, RExC_precomp, RExC_precomp + offset);	\
} STMT_END

#define	vWARN3(loc, m, a1, a2) STMT_START {				\
    const IV offset = loc - RExC_precomp;				\
    Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION,		\
	    a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset);	\
} STMT_END

#define	ckWARN3reg(loc, m, a1, a2) STMT_START {				\
    const IV offset = loc - RExC_precomp;				\
    Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION,	\
	    a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset);	\
} STMT_END

#define	vWARN4(loc, m, a1, a2, a3) STMT_START {				\
    const IV offset = loc - RExC_precomp;				\
    Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION,		\
	    a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
} STMT_END

#define	ckWARN4reg(loc, m, a1, a2, a3) STMT_START {			\
    const IV offset = loc - RExC_precomp;				\
    Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION,	\
	    a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
} STMT_END

#define	vWARN5(loc, m, a1, a2, a3, a4) STMT_START {			\
    const IV offset = loc - RExC_precomp;				\
    Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION,		\
	    a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
} STMT_END


/* Allow for side effects in s */
#define REGC(c,s) STMT_START {			\
    if (!SIZE_ONLY) *(s) = (c); else (void)(s);	\
} STMT_END

/* Macros for recording node offsets.   20001227 mjd@plover.com 
 * Nodes are numbered 1, 2, 3, 4.  Node #n's position is recorded in
 * element 2*n-1 of the array.  Element #2n holds the byte length node #n.
 * Element 0 holds the number n.
 * Position is 1 indexed.
 */
#ifndef RE_TRACK_PATTERN_OFFSETS
#define Set_Node_Offset_To_R(node,byte)
#define Set_Node_Offset(node,byte)
#define Set_Cur_Node_Offset
#define Set_Node_Length_To_R(node,len)
#define Set_Node_Length(node,len)
#define Set_Node_Cur_Length(node)
#define Node_Offset(n) 
#define Node_Length(n) 
#define Set_Node_Offset_Length(node,offset,len)
#define ProgLen(ri) ri->u.proglen
#define SetProgLen(ri,x) ri->u.proglen = x
#else
#define ProgLen(ri) ri->u.offsets[0]
#define SetProgLen(ri,x) ri->u.offsets[0] = x
#define Set_Node_Offset_To_R(node,byte) STMT_START {			\
    if (! SIZE_ONLY) {							\
	MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n",		\
		    __LINE__, (int)(node), (int)(byte)));		\
	if((node) < 0) {						\
	    Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
	} else {							\
	    RExC_offsets[2*(node)-1] = (byte);				\
	}								\
    }									\
} STMT_END

#define Set_Node_Offset(node,byte) \
    Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)

#define Set_Node_Length_To_R(node,len) STMT_START {			\
    if (! SIZE_ONLY) {							\
	MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n",		\
		__LINE__, (int)(node), (int)(len)));			\
	if((node) < 0) {						\
	    Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
	} else {							\
	    RExC_offsets[2*(node)] = (len);				\
	}								\
    }									\
} STMT_END

#define Set_Node_Length(node,len) \
    Set_Node_Length_To_R((node)-RExC_emit_start, len)
#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
#define Set_Node_Cur_Length(node) \
    Set_Node_Length(node, RExC_parse - parse_start)

/* Get offsets and lengths */
#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])

#define Set_Node_Offset_Length(node,offset,len) STMT_START {	\
    Set_Node_Offset_To_R((node)-RExC_emit_start, (offset));	\
    Set_Node_Length_To_R((node)-RExC_emit_start, (len));	\
} STMT_END
#endif

#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
#define EXPERIMENTAL_INPLACESCAN
#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/

#define DEBUG_STUDYDATA(str,data,depth)                              \
DEBUG_OPTIMISE_MORE_r(if(data){                                      \
    PerlIO_printf(Perl_debug_log,                                    \
        "%*s" str "Pos:%"IVdf"/%"IVdf                                \
        " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s",       \
        (int)(depth)*2, "",                                          \
        (IV)((data)->pos_min),                                       \
        (IV)((data)->pos_delta),                                     \
        (UV)((data)->flags),                                         \
        (IV)((data)->whilem_c),                                      \
        (IV)((data)->last_closep ? *((data)->last_closep) : -1),     \
        is_inf ? "INF " : ""                                         \
    );                                                               \
    if ((data)->last_found)                                          \
        PerlIO_printf(Perl_debug_log,                                \
            "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
            " %sFloat: '%s' @ %"IVdf"/%"IVdf"",                      \
            SvPVX_const((data)->last_found),                         \
            (IV)((data)->last_end),                                  \
            (IV)((data)->last_start_min),                            \
            (IV)((data)->last_start_max),                            \
            ((data)->longest &&                                      \
             (data)->longest==&((data)->longest_fixed)) ? "*" : "",  \
            SvPVX_const((data)->longest_fixed),                      \
            (IV)((data)->offset_fixed),                              \
            ((data)->longest &&                                      \
             (data)->longest==&((data)->longest_float)) ? "*" : "",  \
            SvPVX_const((data)->longest_float),                      \
            (IV)((data)->offset_float_min),                          \
            (IV)((data)->offset_float_max)                           \
        );                                                           \
    PerlIO_printf(Perl_debug_log,"\n");                              \
});

static void clear_re(pTHX_ void *r);

/* Mark that we cannot extend a found fixed substring at this point.
   Update the longest found anchored substring and the longest found
   floating substrings if needed. */

STATIC void
S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
{
    const STRLEN l = CHR_SVLEN(data->last_found);
    const STRLEN old_l = CHR_SVLEN(*data->longest);
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_SCAN_COMMIT;

    if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
	SvSetMagicSV(*data->longest, data->last_found);
	if (*data->longest == data->longest_fixed) {
	    data->offset_fixed = l ? data->last_start_min : data->pos_min;
	    if (data->flags & SF_BEFORE_EOL)
		data->flags
		    |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
	    else
		data->flags &= ~SF_FIX_BEFORE_EOL;
	    data->minlen_fixed=minlenp;
	    data->lookbehind_fixed=0;
	}
	else { /* *data->longest == data->longest_float */
	    data->offset_float_min = l ? data->last_start_min : data->pos_min;
	    data->offset_float_max = (l
				      ? data->last_start_max
				      : data->pos_min + data->pos_delta);
	    if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
		data->offset_float_max = I32_MAX;
	    if (data->flags & SF_BEFORE_EOL)
		data->flags
		    |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
	    else
		data->flags &= ~SF_FL_BEFORE_EOL;
            data->minlen_float=minlenp;
            data->lookbehind_float=0;
	}
    }
    SvCUR_set(data->last_found, 0);
    {
	SV * const sv = data->last_found;
	if (SvUTF8(sv) && SvMAGICAL(sv)) {
	    MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
	    if (mg)
		mg->mg_len = 0;
	}
    }
    data->last_end = -1;
    data->flags &= ~SF_BEFORE_EOL;
    DEBUG_STUDYDATA("commit: ",data,0);
}

/* Can match anything (initialization) */
STATIC void
S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
{
    PERL_ARGS_ASSERT_CL_ANYTHING;

    ANYOF_BITMAP_SETALL(cl);
    cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
		|ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;

    /* If any portion of the regex is to operate under locale rules,
     * initialization includes it.  The reason this isn't done for all regexes
     * is that the optimizer was written under the assumption that locale was
     * all-or-nothing.  Given the complexity and lack of documentation in the
     * optimizer, and that there are inadequate test cases for locale, so many
     * parts of it may not work properly, it is safest to avoid locale unless
     * necessary. */
    if (RExC_contains_locale) {
	ANYOF_CLASS_SETALL(cl);	    /* /l uses class */
	cl->flags |= ANYOF_LOCALE;
    }
    else {
	ANYOF_CLASS_ZERO(cl);	    /* Only /l uses class now */
    }
}

/* Can match anything (initialization) */
STATIC int
S_cl_is_anything(const struct regnode_charclass_class *cl)
{
    int value;

    PERL_ARGS_ASSERT_CL_IS_ANYTHING;

    for (value = 0; value <= ANYOF_MAX; value += 2)
	if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
	    return 1;
    if (!(cl->flags & ANYOF_UNICODE_ALL))
	return 0;
    if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
	return 0;
    return 1;
}

/* Can match anything (initialization) */
STATIC void
S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
{
    PERL_ARGS_ASSERT_CL_INIT;

    Zero(cl, 1, struct regnode_charclass_class);
    cl->type = ANYOF;
    cl_anything(pRExC_state, cl);
    ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
}

/* These two functions currently do the exact same thing */
#define cl_init_zero		S_cl_init

/* 'AND' a given class with another one.  Can create false positives.  'cl'
 * should not be inverted.  'and_with->flags & ANYOF_CLASS' should be 0 if
 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
STATIC void
S_cl_and(struct regnode_charclass_class *cl,
	const struct regnode_charclass_class *and_with)
{
    PERL_ARGS_ASSERT_CL_AND;

    assert(and_with->type == ANYOF);

    /* I (khw) am not sure all these restrictions are necessary XXX */
    if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
	&& !(ANYOF_CLASS_TEST_ANY_SET(cl))
	&& (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
	&& !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
	&& !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
	int i;

	if (and_with->flags & ANYOF_INVERT)
	    for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
		cl->bitmap[i] &= ~and_with->bitmap[i];
	else
	    for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
		cl->bitmap[i] &= and_with->bitmap[i];
    } /* XXXX: logic is complicated otherwise, leave it along for a moment. */

    if (and_with->flags & ANYOF_INVERT) {

        /* Here, the and'ed node is inverted.  Get the AND of the flags that
         * aren't affected by the inversion.  Those that are affected are
         * handled individually below */
	U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
	cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
	cl->flags |= affected_flags;

        /* We currently don't know how to deal with things that aren't in the
         * bitmap, but we know that the intersection is no greater than what
         * is already in cl, so let there be false positives that get sorted
         * out after the synthetic start class succeeds, and the node is
         * matched for real. */

        /* The inversion of these two flags indicate that the resulting
         * intersection doesn't have them */
	if (and_with->flags & ANYOF_UNICODE_ALL) {
	    cl->flags &= ~ANYOF_UNICODE_ALL;
	}
	if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
	    cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
	}
    }
    else {   /* and'd node is not inverted */
	U8 outside_bitmap_but_not_utf8; /* Temp variable */

	if (! ANYOF_NONBITMAP(and_with)) {

            /* Here 'and_with' doesn't match anything outside the bitmap
             * (except possibly ANYOF_UNICODE_ALL), which means the
             * intersection can't either, except for ANYOF_UNICODE_ALL, in
             * which case we don't know what the intersection is, but it's no
             * greater than what cl already has, so can just leave it alone,
             * with possible false positives */
            if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
                ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
		cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
            }
	}
	else if (! ANYOF_NONBITMAP(cl)) {

	    /* Here, 'and_with' does match something outside the bitmap, and cl
	     * doesn't have a list of things to match outside the bitmap.  If
             * cl can match all code points above 255, the intersection will
             * be those above-255 code points that 'and_with' matches.  If cl
             * can't match all Unicode code points, it means that it can't
             * match anything outside the bitmap (since the 'if' that got us
             * into this block tested for that), so we leave the bitmap empty.
             */
	    if (cl->flags & ANYOF_UNICODE_ALL) {
		ARG_SET(cl, ARG(and_with));

                /* and_with's ARG may match things that don't require UTF8.
                 * And now cl's will too, in spite of this being an 'and'.  See
                 * the comments below about the kludge */
		cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
	    }
	}
	else {
            /* Here, both 'and_with' and cl match something outside the
             * bitmap.  Currently we do not do the intersection, so just match
             * whatever cl had at the beginning.  */
	}


        /* Take the intersection of the two sets of flags.  However, the
         * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'.  This is a
         * kludge around the fact that this flag is not treated like the others
         * which are initialized in cl_anything().  The way the optimizer works
         * is that the synthetic start class (SSC) is initialized to match
         * anything, and then the first time a real node is encountered, its
         * values are AND'd with the SSC's with the result being the values of
         * the real node.  However, there are paths through the optimizer where
         * the AND never gets called, so those initialized bits are set
         * inappropriately, which is not usually a big deal, as they just cause
         * false positives in the SSC, which will just mean a probably
         * imperceptible slow down in execution.  However this bit has a
         * higher false positive consequence in that it can cause utf8.pm,
         * utf8_heavy.pl ... to be loaded when not necessary, which is a much
         * bigger slowdown and also causes significant extra memory to be used.
         * In order to prevent this, the code now takes a different tack.  The
         * bit isn't set unless some part of the regular expression needs it,
         * but once set it won't get cleared.  This means that these extra
         * modules won't get loaded unless there was some path through the
         * pattern that would have required them anyway, and  so any false
         * positives that occur by not ANDing them out when they could be
         * aren't as severe as they would be if we treated this bit like all
         * the others */
        outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
                                      & ANYOF_NONBITMAP_NON_UTF8;
	cl->flags &= and_with->flags;
	cl->flags |= outside_bitmap_but_not_utf8;
    }
}

/* 'OR' a given class with another one.  Can create false positives.  'cl'
 * should not be inverted.  'or_with->flags & ANYOF_CLASS' should be 0 if
 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
STATIC void
S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
{
    PERL_ARGS_ASSERT_CL_OR;

    if (or_with->flags & ANYOF_INVERT) {

        /* Here, the or'd node is to be inverted.  This means we take the
         * complement of everything not in the bitmap, but currently we don't
         * know what that is, so give up and match anything */
	if (ANYOF_NONBITMAP(or_with)) {
	    cl_anything(pRExC_state, cl);
	}
	/* We do not use
	 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
	 *   <= (B1 | !B2) | (CL1 | !CL2)
	 * which is wasteful if CL2 is small, but we ignore CL2:
	 *   (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
	 * XXXX Can we handle case-fold?  Unclear:
	 *   (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
	 *   (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
	 */
	else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
	     && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
	     && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
	    int i;

	    for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
		cl->bitmap[i] |= ~or_with->bitmap[i];
	} /* XXXX: logic is complicated otherwise */
	else {
	    cl_anything(pRExC_state, cl);
	}

        /* And, we can just take the union of the flags that aren't affected
         * by the inversion */
	cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;

        /* For the remaining flags:
            ANYOF_UNICODE_ALL and inverted means to not match anything above
                    255, which means that the union with cl should just be
                    what cl has in it, so can ignore this flag
            ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
                    is 127-255 to match them, but then invert that, so the
                    union with cl should just be what cl has in it, so can
                    ignore this flag
         */
    } else {    /* 'or_with' is not inverted */
	/* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
	if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
	     && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
		 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
	    int i;

	    /* OR char bitmap and class bitmap separately */
	    for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
		cl->bitmap[i] |= or_with->bitmap[i];
	    if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
		for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
		    cl->classflags[i] |= or_with->classflags[i];
		cl->flags |= ANYOF_CLASS;
	    }
	}
	else { /* XXXX: logic is complicated, leave it along for a moment. */
	    cl_anything(pRExC_state, cl);
	}

	if (ANYOF_NONBITMAP(or_with)) {

	    /* Use the added node's outside-the-bit-map match if there isn't a
	     * conflict.  If there is a conflict (both nodes match something
	     * outside the bitmap, but what they match outside is not the same
	     * pointer, and hence not easily compared until XXX we extend
	     * inversion lists this far), give up and allow the start class to
	     * match everything outside the bitmap.  If that stuff is all above
	     * 255, can just set UNICODE_ALL, otherwise caould be anything. */
	    if (! ANYOF_NONBITMAP(cl)) {
		ARG_SET(cl, ARG(or_with));
	    }
	    else if (ARG(cl) != ARG(or_with)) {

		if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
		    cl_anything(pRExC_state, cl);
		}
		else {
		    cl->flags |= ANYOF_UNICODE_ALL;
		}
	    }
	}

        /* Take the union */
	cl->flags |= or_with->flags;
    }
}

#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
#define TRIE_LIST_CUR(state)  ( TRIE_LIST_ITEM( state, 0 ).forid )
#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
#define TRIE_LIST_USED(idx)  ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )


#ifdef DEBUGGING
/*
   dump_trie(trie,widecharmap,revcharmap)
   dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
   dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)

   These routines dump out a trie in a somewhat readable format.
   The _interim_ variants are used for debugging the interim
   tables that are used to generate the final compressed
   representation which is what dump_trie expects.

   Part of the reason for their existence is to provide a form
   of documentation as to how the different representations function.

*/

/*
  Dumps the final compressed table form of the trie to Perl_debug_log.
  Used for debugging make_trie().
*/

STATIC void
S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
	    AV *revcharmap, U32 depth)
{
    U32 state;
    SV *sv=sv_newmortal();
    int colwidth= widecharmap ? 6 : 4;
    U16 word;
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_DUMP_TRIE;

    PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
        (int)depth * 2 + 2,"",
        "Match","Base","Ofs" );

    for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
	SV ** const tmp = av_fetch( revcharmap, state, 0);
        if ( tmp ) {
            PerlIO_printf( Perl_debug_log, "%*s", 
                colwidth,
                pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth, 
	                    PL_colors[0], PL_colors[1],
	                    (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
	                    PERL_PV_ESCAPE_FIRSTCHAR 
                ) 
            );
        }
    }
    PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
        (int)depth * 2 + 2,"");

    for( state = 0 ; state < trie->uniquecharcount ; state++ )
        PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
    PerlIO_printf( Perl_debug_log, "\n");

    for( state = 1 ; state < trie->statecount ; state++ ) {
	const U32 base = trie->states[ state ].trans.base;

        PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);

        if ( trie->states[ state ].wordnum ) {
            PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
        } else {
            PerlIO_printf( Perl_debug_log, "%6s", "" );
        }

        PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );

        if ( base ) {
            U32 ofs = 0;

            while( ( base + ofs  < trie->uniquecharcount ) ||
                   ( base + ofs - trie->uniquecharcount < trie->lasttrans
                     && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
                    ofs++;

            PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);

            for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
                if ( ( base + ofs >= trie->uniquecharcount ) &&
                     ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
                     trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
                {
                   PerlIO_printf( Perl_debug_log, "%*"UVXf,
                    colwidth,
                    (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
                } else {
                    PerlIO_printf( Perl_debug_log, "%*s",colwidth,"   ." );
                }
            }

            PerlIO_printf( Perl_debug_log, "]");

        }
        PerlIO_printf( Perl_debug_log, "\n" );
    }
    PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
    for (word=1; word <= trie->wordcount; word++) {
	PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
	    (int)word, (int)(trie->wordinfo[word].prev),
	    (int)(trie->wordinfo[word].len));
    }
    PerlIO_printf(Perl_debug_log, "\n" );
}    
/*
  Dumps a fully constructed but uncompressed trie in list form.
  List tries normally only are used for construction when the number of 
  possible chars (trie->uniquecharcount) is very high.
  Used for debugging make_trie().
*/
STATIC void
S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
			 HV *widecharmap, AV *revcharmap, U32 next_alloc,
			 U32 depth)
{
    U32 state;
    SV *sv=sv_newmortal();
    int colwidth= widecharmap ? 6 : 4;
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;

    /* print out the table precompression.  */
    PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
        (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
        "------:-----+-----------------\n" );
    
    for( state=1 ; state < next_alloc ; state ++ ) {
        U16 charid;
    
        PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
            (int)depth * 2 + 2,"", (UV)state  );
        if ( ! trie->states[ state ].wordnum ) {
            PerlIO_printf( Perl_debug_log, "%5s| ","");
        } else {
            PerlIO_printf( Perl_debug_log, "W%4x| ",
                trie->states[ state ].wordnum
            );
        }
        for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
	    SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
	    if ( tmp ) {
                PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
                    colwidth,
                    pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth, 
	                    PL_colors[0], PL_colors[1],
	                    (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
	                    PERL_PV_ESCAPE_FIRSTCHAR 
                    ) ,
                    TRIE_LIST_ITEM(state,charid).forid,
                    (UV)TRIE_LIST_ITEM(state,charid).newstate
                );
                if (!(charid % 10)) 
                    PerlIO_printf(Perl_debug_log, "\n%*s| ",
                        (int)((depth * 2) + 14), "");
            }
        }
        PerlIO_printf( Perl_debug_log, "\n");
    }
}    

/*
  Dumps a fully constructed but uncompressed trie in table form.
  This is the normal DFA style state transition table, with a few 
  twists to facilitate compression later. 
  Used for debugging make_trie().
*/
STATIC void
S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
			  HV *widecharmap, AV *revcharmap, U32 next_alloc,
			  U32 depth)
{
    U32 state;
    U16 charid;
    SV *sv=sv_newmortal();
    int colwidth= widecharmap ? 6 : 4;
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
    
    /*
       print out the table precompression so that we can do a visual check
       that they are identical.
     */
    
    PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );

    for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
	SV ** const tmp = av_fetch( revcharmap, charid, 0);
        if ( tmp ) {
            PerlIO_printf( Perl_debug_log, "%*s", 
                colwidth,
                pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth, 
	                    PL_colors[0], PL_colors[1],
	                    (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
	                    PERL_PV_ESCAPE_FIRSTCHAR 
                ) 
            );
        }
    }

    PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );

    for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
        PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
    }

    PerlIO_printf( Perl_debug_log, "\n" );

    for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {

        PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ", 
            (int)depth * 2 + 2,"",
            (UV)TRIE_NODENUM( state ) );

        for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
            UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
            if (v)
                PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
            else
                PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
        }
        if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
            PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
        } else {
            PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
            trie->states[ TRIE_NODENUM( state ) ].wordnum );
        }
    }
}

#endif


/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
  startbranch: the first branch in the whole branch sequence
  first      : start branch of sequence of branch-exact nodes.
	       May be the same as startbranch
  last       : Thing following the last branch.
	       May be the same as tail.
  tail       : item following the branch sequence
  count      : words in the sequence
  flags      : currently the OP() type we will be building one of /EXACT(|F|Fl)/
  depth      : indent depth

Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.

A trie is an N'ary tree where the branches are determined by digital
decomposition of the key. IE, at the root node you look up the 1st character and
follow that branch repeat until you find the end of the branches. Nodes can be
marked as "accepting" meaning they represent a complete word. Eg:

  /he|she|his|hers/

would convert into the following structure. Numbers represent states, letters
following numbers represent valid transitions on the letter from that state, if
the number is in square brackets it represents an accepting state, otherwise it
will be in parenthesis.

      +-h->+-e->[3]-+-r->(8)-+-s->[9]
      |    |
      |   (2)
      |    |
     (1)   +-i->(6)-+-s->[7]
      |
      +-s->(3)-+-h->(4)-+-e->[5]

      Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)

This shows that when matching against the string 'hers' we will begin at state 1
read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
is also accepting. Thus we know that we can match both 'he' and 'hers' with a
single traverse. We store a mapping from accepting to state to which word was
matched, and then when we have multiple possibilities we try to complete the
rest of the regex in the order in which they occured in the alternation.

The only prior NFA like behaviour that would be changed by the TRIE support is
the silent ignoring of duplicate alternations which are of the form:

 / (DUPE|DUPE) X? (?{ ... }) Y /x

Thus EVAL blocks following a trie may be called a different number of times with
and without the optimisation. With the optimisations dupes will be silently
ignored. This inconsistent behaviour of EVAL type nodes is well established as
the following demonstrates:

 'words'=~/(word|word|word)(?{ print $1 })[xyz]/

which prints out 'word' three times, but

 'words'=~/(word|word|word)(?{ print $1 })S/

which doesnt print it out at all. This is due to other optimisations kicking in.

Example of what happens on a structural level:

The regexp /(ac|ad|ab)+/ will produce the following debug output:

   1: CURLYM[1] {1,32767}(18)
   5:   BRANCH(8)
   6:     EXACT <ac>(16)
   8:   BRANCH(11)
   9:     EXACT <ad>(16)
  11:   BRANCH(14)
  12:     EXACT <ab>(16)
  16:   SUCCEED(0)
  17:   NOTHING(18)
  18: END(0)

This would be optimizable with startbranch=5, first=5, last=16, tail=16
and should turn into:

   1: CURLYM[1] {1,32767}(18)
   5:   TRIE(16)
	[Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
	  <ac>
	  <ad>
	  <ab>
  16:   SUCCEED(0)
  17:   NOTHING(18)
  18: END(0)

Cases where tail != last would be like /(?foo|bar)baz/:

   1: BRANCH(4)
   2:   EXACT <foo>(8)
   4: BRANCH(7)
   5:   EXACT <bar>(8)
   7: TAIL(8)
   8: EXACT <baz>(10)
  10: END(0)

which would be optimizable with startbranch=1, first=1, last=7, tail=8
and would end up looking like:

    1: TRIE(8)
      [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
	<foo>
	<bar>
   7: TAIL(8)
   8: EXACT <baz>(10)
  10: END(0)

    d = uvuni_to_utf8_flags(d, uv, 0);

is the recommended Unicode-aware way of saying

    *(d++) = uv;
*/

#define TRIE_STORE_REVCHAR                                                 \
    STMT_START {                                                           \
	if (UTF) {							   \
	    SV *zlopp = newSV(2);					   \
	    unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp);	   \
	    unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
	    SvCUR_set(zlopp, kapow - flrbbbbb);				   \
	    SvPOK_on(zlopp);						   \
	    SvUTF8_on(zlopp);						   \
	    av_push(revcharmap, zlopp);					   \
	} else {							   \
	    char ooooff = (char)uvc;					   	   \
	    av_push(revcharmap, newSVpvn(&ooooff, 1));			   \
	}								   \
        } STMT_END

#define TRIE_READ_CHAR STMT_START {                                           \
    wordlen++;                                                                \
    if ( UTF ) {                                                              \
	if ( folder ) {                                                       \
	    if ( foldlen > 0 ) {                                              \
	       uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags );     \
	       foldlen -= len;                                                \
	       scan += len;                                                   \
	       len = 0;                                                       \
	    } else {                                                          \
		len = UTF8SKIP(uc);\
		uvc = to_utf8_fold( uc, foldbuf, &foldlen);                   \
		foldlen -= UNISKIP( uvc );                                    \
		scan = foldbuf + UNISKIP( uvc );                              \
	    }                                                                 \
	} else {                                                              \
	    uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
	}                                                                     \
    } else {                                                                  \
	uvc = (U32)*uc;                                                       \
	len = 1;                                                              \
    }                                                                         \
} STMT_END



#define TRIE_LIST_PUSH(state,fid,ns) STMT_START {               \
    if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) {    \
	U32 ging = TRIE_LIST_LEN( state ) *= 2;                 \
	Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
    }                                                           \
    TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid;     \
    TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns;   \
    TRIE_LIST_CUR( state )++;                                   \
} STMT_END

#define TRIE_LIST_NEW(state) STMT_START {                       \
    Newxz( trie->states[ state ].trans.list,               \
	4, reg_trie_trans_le );                                 \
     TRIE_LIST_CUR( state ) = 1;                                \
     TRIE_LIST_LEN( state ) = 4;                                \
} STMT_END

#define TRIE_HANDLE_WORD(state) STMT_START {                    \
    U16 dupe= trie->states[ state ].wordnum;                    \
    regnode * const noper_next = regnext( noper );              \
                                                                \
    DEBUG_r({                                                   \
        /* store the word for dumping */                        \
        SV* tmp;                                                \
        if (OP(noper) != NOTHING)                               \
            tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF);	\
        else                                                    \
            tmp = newSVpvn_utf8( "", 0, UTF );			\
        av_push( trie_words, tmp );                             \
    });                                                         \
                                                                \
    curword++;                                                  \
    trie->wordinfo[curword].prev   = 0;                         \
    trie->wordinfo[curword].len    = wordlen;                   \
    trie->wordinfo[curword].accept = state;                     \
                                                                \
    if ( noper_next < tail ) {                                  \
        if (!trie->jump)                                        \
            trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
        trie->jump[curword] = (U16)(noper_next - convert);      \
        if (!jumper)                                            \
            jumper = noper_next;                                \
        if (!nextbranch)                                        \
            nextbranch= regnext(cur);                           \
    }                                                           \
                                                                \
    if ( dupe ) {                                               \
        /* It's a dupe. Pre-insert into the wordinfo[].prev   */\
        /* chain, so that when the bits of chain are later    */\
        /* linked together, the dups appear in the chain      */\
	trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
	trie->wordinfo[dupe].prev = curword;                    \
    } else {                                                    \
        /* we haven't inserted this word yet.                */ \
        trie->states[ state ].wordnum = curword;                \
    }                                                           \
} STMT_END


#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special)		\
     ( ( base + charid >=  ucharcount					\
         && base + charid < ubound					\
         && state == trie->trans[ base - ucharcount + charid ].check	\
         && trie->trans[ base - ucharcount + charid ].next )		\
           ? trie->trans[ base - ucharcount + charid ].next		\
           : ( state==1 ? special : 0 )					\
      )

#define MADE_TRIE       1
#define MADE_JUMP_TRIE  2
#define MADE_EXACT_TRIE 4

STATIC I32
S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
{
    dVAR;
    /* first pass, loop through and scan words */
    reg_trie_data *trie;
    HV *widecharmap = NULL;
    AV *revcharmap = newAV();
    regnode *cur;
    const U32 uniflags = UTF8_ALLOW_DEFAULT;
    STRLEN len = 0;
    UV uvc = 0;
    U16 curword = 0;
    U32 next_alloc = 0;
    regnode *jumper = NULL;
    regnode *nextbranch = NULL;
    regnode *convert = NULL;
    U32 *prev_states; /* temp array mapping each state to previous one */
    /* we just use folder as a flag in utf8 */
    const U8 * folder = NULL;

#ifdef DEBUGGING
    const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
    AV *trie_words = NULL;
    /* along with revcharmap, this only used during construction but both are
     * useful during debugging so we store them in the struct when debugging.
     */
#else
    const U32 data_slot = add_data( pRExC_state, 2, "tu" );
    STRLEN trie_charcount=0;
#endif
    SV *re_trie_maxbuff;
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_MAKE_TRIE;
#ifndef DEBUGGING
    PERL_UNUSED_ARG(depth);
#endif

    switch (flags) {
	case EXACT: break;
	case EXACTFA:
	case EXACTFU: folder = PL_fold_latin1; break;
	case EXACTF:  folder = PL_fold; break;
	case EXACTFL: folder = PL_fold_locale; break;
        default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u", (unsigned) flags );
    }

    trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
    trie->refcount = 1;
    trie->startstate = 1;
    trie->wordcount = word_count;
    RExC_rxi->data->data[ data_slot ] = (void*)trie;
    trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
    if (!(UTF && folder))
	trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
    trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
                       trie->wordcount+1, sizeof(reg_trie_wordinfo));

    DEBUG_r({
        trie_words = newAV();
    });

    re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
    if (!SvIOK(re_trie_maxbuff)) {
        sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
    }
    DEBUG_OPTIMISE_r({
                PerlIO_printf( Perl_debug_log,
                  "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
                  (int)depth * 2 + 2, "", 
                  REG_NODE_NUM(startbranch),REG_NODE_NUM(first), 
                  REG_NODE_NUM(last), REG_NODE_NUM(tail),
                  (int)depth);
    });
   
   /* Find the node we are going to overwrite */
    if ( first == startbranch && OP( last ) != BRANCH ) {
        /* whole branch chain */
        convert = first;
    } else {
        /* branch sub-chain */
        convert = NEXTOPER( first );
    }
        
    /*  -- First loop and Setup --

       We first traverse the branches and scan each word to determine if it
       contains widechars, and how many unique chars there are, this is
       important as we have to build a table with at least as many columns as we
       have unique chars.

       We use an array of integers to represent the character codes 0..255
       (trie->charmap) and we use a an HV* to store Unicode characters. We use the
       native representation of the character value as the key and IV's for the
       coded index.

       *TODO* If we keep track of how many times each character is used we can
       remap the columns so that the table compression later on is more
       efficient in terms of memory by ensuring the most common value is in the
       middle and the least common are on the outside.  IMO this would be better
       than a most to least common mapping as theres a decent chance the most
       common letter will share a node with the least common, meaning the node
       will not be compressible. With a middle is most common approach the worst
       case is when we have the least common nodes twice.

     */

    for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
        regnode * const noper = NEXTOPER( cur );
        const U8 *uc = (U8*)STRING( noper );
        const U8 * const e  = uc + STR_LEN( noper );
        STRLEN foldlen = 0;
        U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
        const U8 *scan = (U8*)NULL;
        U32 wordlen      = 0;         /* required init */
        STRLEN chars = 0;
        bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/

        if (OP(noper) == NOTHING) {
            trie->minlen= 0;
            continue;
        }
        if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
            TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
                                          regardless of encoding */

        for ( ; uc < e ; uc += len ) {
            TRIE_CHARCOUNT(trie)++;
            TRIE_READ_CHAR;
            chars++;
            if ( uvc < 256 ) {
                if ( !trie->charmap[ uvc ] ) {
                    trie->charmap[ uvc ]=( ++trie->uniquecharcount );
                    if ( folder )
                        trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
                    TRIE_STORE_REVCHAR;
                }
                if ( set_bit ) {
		    /* store the codepoint in the bitmap, and its folded
		     * equivalent. */
                    TRIE_BITMAP_SET(trie,uvc);

		    /* store the folded codepoint */
		    if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);

		    if ( !UTF ) {
			/* store first byte of utf8 representation of
			   variant codepoints */
			if (! UNI_IS_INVARIANT(uvc)) {
			    TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
			}
		    }
                    set_bit = 0; /* We've done our bit :-) */
                }
            } else {
                SV** svpp;
                if ( !widecharmap )
                    widecharmap = newHV();

                svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );

                if ( !svpp )
                    Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );

                if ( !SvTRUE( *svpp ) ) {
                    sv_setiv( *svpp, ++trie->uniquecharcount );
                    TRIE_STORE_REVCHAR;
                }
            }
        }
        if( cur == first ) {
            trie->minlen=chars;
            trie->maxlen=chars;
        } else if (chars < trie->minlen) {
            trie->minlen=chars;
        } else if (chars > trie->maxlen) {
            trie->maxlen=chars;
        }

    } /* end first pass */
    DEBUG_TRIE_COMPILE_r(
        PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
                (int)depth * 2 + 2,"",
                ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
		(int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
		(int)trie->minlen, (int)trie->maxlen )
    );

    /*
        We now know what we are dealing with in terms of unique chars and
        string sizes so we can calculate how much memory a naive
        representation using a flat table  will take. If it's over a reasonable
        limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
        conservative but potentially much slower representation using an array
        of lists.

        At the end we convert both representations into the same compressed
        form that will be used in regexec.c for matching with. The latter
        is a form that cannot be used to construct with but has memory
        properties similar to the list form and access properties similar
        to the table form making it both suitable for fast searches and
        small enough that its feasable to store for the duration of a program.

        See the comment in the code where the compressed table is produced
        inplace from the flat tabe representation for an explanation of how
        the compression works.

    */


    Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
    prev_states[1] = 0;

    if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
        /*
            Second Pass -- Array Of Lists Representation

            Each state will be represented by a list of charid:state records
            (reg_trie_trans_le) the first such element holds the CUR and LEN
            points of the allocated array. (See defines above).

            We build the initial structure using the lists, and then convert
            it into the compressed table form which allows faster lookups
            (but cant be modified once converted).
        */

        STRLEN transcount = 1;

        DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log, 
            "%*sCompiling trie using list compiler\n",
            (int)depth * 2 + 2, ""));

	trie->states = (reg_trie_state *)
	    PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
				  sizeof(reg_trie_state) );
        TRIE_LIST_NEW(1);
        next_alloc = 2;

        for ( cur = first ; cur < last ; cur = regnext( cur ) ) {

	    regnode * const noper = NEXTOPER( cur );
	    U8 *uc           = (U8*)STRING( noper );
	    const U8 * const e = uc + STR_LEN( noper );
	    U32 state        = 1;         /* required init */
	    U16 charid       = 0;         /* sanity init */
	    U8 *scan         = (U8*)NULL; /* sanity init */
	    STRLEN foldlen   = 0;         /* required init */
            U32 wordlen      = 0;         /* required init */
	    U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];

            if (OP(noper) != NOTHING) {
                for ( ; uc < e ; uc += len ) {

                    TRIE_READ_CHAR;

                    if ( uvc < 256 ) {
                        charid = trie->charmap[ uvc ];
		    } else {
                        SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
                        if ( !svpp ) {
                            charid = 0;
                        } else {
                            charid=(U16)SvIV( *svpp );
                        }
		    }
                    /* charid is now 0 if we dont know the char read, or nonzero if we do */
                    if ( charid ) {

                        U16 check;
                        U32 newstate = 0;

                        charid--;
                        if ( !trie->states[ state ].trans.list ) {
                            TRIE_LIST_NEW( state );
			}
                        for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
                            if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
                                newstate = TRIE_LIST_ITEM( state, check ).newstate;
                                break;
                            }
                        }
                        if ( ! newstate ) {
                            newstate = next_alloc++;
			    prev_states[newstate] = state;
                            TRIE_LIST_PUSH( state, charid, newstate );
                            transcount++;
                        }
                        state = newstate;
                    } else {
                        Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
		    }
		}
	    }
            TRIE_HANDLE_WORD(state);

        } /* end second pass */

        /* next alloc is the NEXT state to be allocated */
        trie->statecount = next_alloc; 
        trie->states = (reg_trie_state *)
	    PerlMemShared_realloc( trie->states,
				   next_alloc
				   * sizeof(reg_trie_state) );

        /* and now dump it out before we compress it */
        DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
							 revcharmap, next_alloc,
							 depth+1)
        );

        trie->trans = (reg_trie_trans *)
	    PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
        {
            U32 state;
            U32 tp = 0;
            U32 zp = 0;


            for( state=1 ; state < next_alloc ; state ++ ) {
                U32 base=0;

                /*
                DEBUG_TRIE_COMPILE_MORE_r(
                    PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
                );
                */

                if (trie->states[state].trans.list) {
                    U16 minid=TRIE_LIST_ITEM( state, 1).forid;
                    U16 maxid=minid;
		    U16 idx;

                    for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
			const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
			if ( forid < minid ) {
			    minid=forid;
			} else if ( forid > maxid ) {
			    maxid=forid;
			}
                    }
                    if ( transcount < tp + maxid - minid + 1) {
                        transcount *= 2;
			trie->trans = (reg_trie_trans *)
			    PerlMemShared_realloc( trie->trans,
						     transcount
						     * sizeof(reg_trie_trans) );
                        Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
                    }
                    base = trie->uniquecharcount + tp - minid;
                    if ( maxid == minid ) {
                        U32 set = 0;
                        for ( ; zp < tp ; zp++ ) {
                            if ( ! trie->trans[ zp ].next ) {
                                base = trie->uniquecharcount + zp - minid;
                                trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
                                trie->trans[ zp ].check = state;
                                set = 1;
                                break;
                            }
                        }
                        if ( !set ) {
                            trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
                            trie->trans[ tp ].check = state;
                            tp++;
                            zp = tp;
                        }
                    } else {
                        for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
                            const U32 tid = base -  trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
                            trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
                            trie->trans[ tid ].check = state;
                        }
                        tp += ( maxid - minid + 1 );
                    }
                    Safefree(trie->states[ state ].trans.list);
                }
                /*
                DEBUG_TRIE_COMPILE_MORE_r(
                    PerlIO_printf( Perl_debug_log, " base: %d\n",base);
                );
                */
                trie->states[ state ].trans.base=base;
            }
            trie->lasttrans = tp + 1;
        }
    } else {
        /*
           Second Pass -- Flat Table Representation.

           we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
           We know that we will need Charcount+1 trans at most to store the data
           (one row per char at worst case) So we preallocate both structures
           assuming worst case.

           We then construct the trie using only the .next slots of the entry
           structs.

           We use the .check field of the first entry of the node temporarily to
           make compression both faster and easier by keeping track of how many non
           zero fields are in the node.

           Since trans are numbered from 1 any 0 pointer in the table is a FAIL
           transition.

           There are two terms at use here: state as a TRIE_NODEIDX() which is a
           number representing the first entry of the node, and state as a
           TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
           TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
           are 2 entrys per node. eg:

             A B       A B
          1. 2 4    1. 3 7
          2. 0 3    3. 0 5
          3. 0 0    5. 0 0
          4. 0 0    7. 0 0

           The table is internally in the right hand, idx form. However as we also
           have to deal with the states array which is indexed by nodenum we have to
           use TRIE_NODENUM() to convert.

        */
        DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log, 
            "%*sCompiling trie using table compiler\n",
            (int)depth * 2 + 2, ""));

	trie->trans = (reg_trie_trans *)
	    PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
				  * trie->uniquecharcount + 1,
				  sizeof(reg_trie_trans) );
        trie->states = (reg_trie_state *)
	    PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
				  sizeof(reg_trie_state) );
        next_alloc = trie->uniquecharcount + 1;


        for ( cur = first ; cur < last ; cur = regnext( cur ) ) {

	    regnode * const noper   = NEXTOPER( cur );
	    const U8 *uc     = (U8*)STRING( noper );
	    const U8 * const e = uc + STR_LEN( noper );

            U32 state        = 1;         /* required init */

            U16 charid       = 0;         /* sanity init */
            U32 accept_state = 0;         /* sanity init */
            U8 *scan         = (U8*)NULL; /* sanity init */

            STRLEN foldlen   = 0;         /* required init */
            U32 wordlen      = 0;         /* required init */
            U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];

            if ( OP(noper) != NOTHING ) {
                for ( ; uc < e ; uc += len ) {

                    TRIE_READ_CHAR;

                    if ( uvc < 256 ) {
                        charid = trie->charmap[ uvc ];
                    } else {
                        SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
                        charid = svpp ? (U16)SvIV(*svpp) : 0;
                    }
                    if ( charid ) {
                        charid--;
                        if ( !trie->trans[ state + charid ].next ) {
                            trie->trans[ state + charid ].next = next_alloc;
                            trie->trans[ state ].check++;
			    prev_states[TRIE_NODENUM(next_alloc)]
				    = TRIE_NODENUM(state);
                            next_alloc += trie->uniquecharcount;
                        }
                        state = trie->trans[ state + charid ].next;
                    } else {
                        Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
                    }
                    /* charid is now 0 if we dont know the char read, or nonzero if we do */
                }
            }
            accept_state = TRIE_NODENUM( state );
            TRIE_HANDLE_WORD(accept_state);

        } /* end second pass */

        /* and now dump it out before we compress it */
        DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
							  revcharmap,
							  next_alloc, depth+1));

        {
        /*
           * Inplace compress the table.*

           For sparse data sets the table constructed by the trie algorithm will
           be mostly 0/FAIL transitions or to put it another way mostly empty.
           (Note that leaf nodes will not contain any transitions.)

           This algorithm compresses the tables by eliminating most such
           transitions, at the cost of a modest bit of extra work during lookup:

           - Each states[] entry contains a .base field which indicates the
           index in the state[] array wheres its transition data is stored.

           - If .base is 0 there are no valid transitions from that node.

           - If .base is nonzero then charid is added to it to find an entry in
           the trans array.

           -If trans[states[state].base+charid].check!=state then the
           transition is taken to be a 0/Fail transition. Thus if there are fail
           transitions at the front of the node then the .base offset will point
           somewhere inside the previous nodes data (or maybe even into a node
           even earlier), but the .check field determines if the transition is
           valid.

           XXX - wrong maybe?
           The following process inplace converts the table to the compressed
           table: We first do not compress the root node 1,and mark all its
           .check pointers as 1 and set its .base pointer as 1 as well. This
           allows us to do a DFA construction from the compressed table later,
           and ensures that any .base pointers we calculate later are greater
           than 0.

           - We set 'pos' to indicate the first entry of the second node.

           - We then iterate over the columns of the node, finding the first and
           last used entry at l and m. We then copy l..m into pos..(pos+m-l),
           and set the .check pointers accordingly, and advance pos
           appropriately and repreat for the next node. Note that when we copy
           the next pointers we have to convert them from the original
           NODEIDX form to NODENUM form as the former is not valid post
           compression.

           - If a node has no transitions used we mark its base as 0 and do not
           advance the pos pointer.

           - If a node only has one transition we use a second pointer into the
           structure to fill in allocated fail transitions from other states.
           This pointer is independent of the main pointer and scans forward
           looking for null transitions that are allocated to a state. When it
           finds one it writes the single transition into the "hole".  If the
           pointer doesnt find one the single transition is appended as normal.

           - Once compressed we can Renew/realloc the structures to release the
           excess space.

           See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
           specifically Fig 3.47 and the associated pseudocode.

           demq
        */
        const U32 laststate = TRIE_NODENUM( next_alloc );
	U32 state, charid;
        U32 pos = 0, zp=0;
        trie->statecount = laststate;

        for ( state = 1 ; state < laststate ; state++ ) {
            U8 flag = 0;
	    const U32 stateidx = TRIE_NODEIDX( state );
	    const U32 o_used = trie->trans[ stateidx ].check;
	    U32 used = trie->trans[ stateidx ].check;
            trie->trans[ stateidx ].check = 0;

            for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
                if ( flag || trie->trans[ stateidx + charid ].next ) {
                    if ( trie->trans[ stateidx + charid ].next ) {
                        if (o_used == 1) {
                            for ( ; zp < pos ; zp++ ) {
                                if ( ! trie->trans[ zp ].next ) {
                                    break;
                                }
                            }
                            trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
                            trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
                            trie->trans[ zp ].check = state;
                            if ( ++zp > pos ) pos = zp;
                            break;
                        }
                        used--;
                    }
                    if ( !flag ) {
                        flag = 1;
                        trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
                    }
                    trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
                    trie->trans[ pos ].check = state;
                    pos++;
                }
            }
        }
        trie->lasttrans = pos + 1;
        trie->states = (reg_trie_state *)
	    PerlMemShared_realloc( trie->states, laststate
				   * sizeof(reg_trie_state) );
        DEBUG_TRIE_COMPILE_MORE_r(
                PerlIO_printf( Perl_debug_log,
		    "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
		    (int)depth * 2 + 2,"",
                    (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
		    (IV)next_alloc,
		    (IV)pos,
                    ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
            );

        } /* end table compress */
    }
    DEBUG_TRIE_COMPILE_MORE_r(
            PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
                (int)depth * 2 + 2, "",
                (UV)trie->statecount,
                (UV)trie->lasttrans)
    );
    /* resize the trans array to remove unused space */
    trie->trans = (reg_trie_trans *)
	PerlMemShared_realloc( trie->trans, trie->lasttrans
			       * sizeof(reg_trie_trans) );

    {   /* Modify the program and insert the new TRIE node */ 
        U8 nodetype =(U8)(flags & 0xFF);
        char *str=NULL;
        
#ifdef DEBUGGING
        regnode *optimize = NULL;
#ifdef RE_TRACK_PATTERN_OFFSETS

        U32 mjd_offset = 0;
        U32 mjd_nodelen = 0;
#endif /* RE_TRACK_PATTERN_OFFSETS */
#endif /* DEBUGGING */
        /*
           This means we convert either the first branch or the first Exact,
           depending on whether the thing following (in 'last') is a branch
           or not and whther first is the startbranch (ie is it a sub part of
           the alternation or is it the whole thing.)
           Assuming its a sub part we convert the EXACT otherwise we convert
           the whole branch sequence, including the first.
         */
        /* Find the node we are going to overwrite */
        if ( first != startbranch || OP( last ) == BRANCH ) {
            /* branch sub-chain */
            NEXT_OFF( first ) = (U16)(last - first);
#ifdef RE_TRACK_PATTERN_OFFSETS
            DEBUG_r({
                mjd_offset= Node_Offset((convert));
                mjd_nodelen= Node_Length((convert));
            });
#endif
            /* whole branch chain */
        }
#ifdef RE_TRACK_PATTERN_OFFSETS
        else {
            DEBUG_r({
                const  regnode *nop = NEXTOPER( convert );
                mjd_offset= Node_Offset((nop));
                mjd_nodelen= Node_Length((nop));
            });
        }
        DEBUG_OPTIMISE_r(
            PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
                (int)depth * 2 + 2, "",
                (UV)mjd_offset, (UV)mjd_nodelen)
        );
#endif
        /* But first we check to see if there is a common prefix we can 
           split out as an EXACT and put in front of the TRIE node.  */
        trie->startstate= 1;
        if ( trie->bitmap && !widecharmap && !trie->jump  ) {
            U32 state;
            for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
                U32 ofs = 0;
                I32 idx = -1;
                U32 count = 0;
                const U32 base = trie->states[ state ].trans.base;

                if ( trie->states[state].wordnum )
                        count = 1;

                for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
                    if ( ( base + ofs >= trie->uniquecharcount ) &&
                         ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
                         trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
                    {
                        if ( ++count > 1 ) {
                            SV **tmp = av_fetch( revcharmap, ofs, 0);
			    const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
                            if ( state == 1 ) break;
                            if ( count == 2 ) {
                                Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
                                DEBUG_OPTIMISE_r(
                                    PerlIO_printf(Perl_debug_log,
					"%*sNew Start State=%"UVuf" Class: [",
                                        (int)depth * 2 + 2, "",
                                        (UV)state));
				if (idx >= 0) {
				    SV ** const tmp = av_fetch( revcharmap, idx, 0);
				    const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );

                                    TRIE_BITMAP_SET(trie,*ch);
                                    if ( folder )
                                        TRIE_BITMAP_SET(trie, folder[ *ch ]);
                                    DEBUG_OPTIMISE_r(
                                        PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
                                    );
				}
			    }
			    TRIE_BITMAP_SET(trie,*ch);
			    if ( folder )
				TRIE_BITMAP_SET(trie,folder[ *ch ]);
			    DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
			}
                        idx = ofs;
		    }
                }
                if ( count == 1 ) {
                    SV **tmp = av_fetch( revcharmap, idx, 0);
                    STRLEN len;
                    char *ch = SvPV( *tmp, len );
                    DEBUG_OPTIMISE_r({
                        SV *sv=sv_newmortal();
                        PerlIO_printf( Perl_debug_log,
			    "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
                            (int)depth * 2 + 2, "",
                            (UV)state, (UV)idx, 
                            pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6, 
	                        PL_colors[0], PL_colors[1],
	                        (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
	                        PERL_PV_ESCAPE_FIRSTCHAR 
                            )
                        );
                    });
                    if ( state==1 ) {
                        OP( convert ) = nodetype;
                        str=STRING(convert);
                        STR_LEN(convert)=0;
                    }
                    STR_LEN(convert) += len;
                    while (len--)
                        *str++ = *ch++;
		} else {
#ifdef DEBUGGING	    
		    if (state>1)
			DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
#endif
		    break;
		}
	    }
	    trie->prefixlen = (state-1);
            if (str) {
                regnode *n = convert+NODE_SZ_STR(convert);
                NEXT_OFF(convert) = NODE_SZ_STR(convert);
                trie->startstate = state;
                trie->minlen -= (state - 1);
                trie->maxlen -= (state - 1);
#ifdef DEBUGGING
               /* At least the UNICOS C compiler choked on this
                * being argument to DEBUG_r(), so let's just have
                * it right here. */
               if (
#ifdef PERL_EXT_RE_BUILD
                   1
#else
                   DEBUG_r_TEST
#endif
                   ) {
                   regnode *fix = convert;
                   U32 word = trie->wordcount;
                   mjd_nodelen++;
                   Set_Node_Offset_Length(convert, mjd_offset, state - 1);
                   while( ++fix < n ) {
                       Set_Node_Offset_Length(fix, 0, 0);
                   }
                   while (word--) {
                       SV ** const tmp = av_fetch( trie_words, word, 0 );
                       if (tmp) {
                           if ( STR_LEN(convert) <= SvCUR(*tmp) )
                               sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
                           else
                               sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
                       }
                   }
               }
#endif
                if (trie->maxlen) {
                    convert = n;
		} else {
                    NEXT_OFF(convert) = (U16)(tail - convert);
                    DEBUG_r(optimize= n);
                }
            }
        }
        if (!jumper) 
            jumper = last; 
        if ( trie->maxlen ) {
	    NEXT_OFF( convert ) = (U16)(tail - convert);
	    ARG_SET( convert, data_slot );
	    /* Store the offset to the first unabsorbed branch in 
	       jump[0], which is otherwise unused by the jump logic. 
	       We use this when dumping a trie and during optimisation. */
	    if (trie->jump) 
	        trie->jump[0] = (U16)(nextbranch - convert);
            
            /* If the start state is not accepting (meaning there is no empty string/NOTHING)
	     *   and there is a bitmap
	     *   and the first "jump target" node we found leaves enough room
	     * then convert the TRIE node into a TRIEC node, with the bitmap
	     * embedded inline in the opcode - this is hypothetically faster.
	     */
            if ( !trie->states[trie->startstate].wordnum
		 && trie->bitmap
		 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
            {
                OP( convert ) = TRIEC;
                Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
                PerlMemShared_free(trie->bitmap);
                trie->bitmap= NULL;
            } else 
                OP( convert ) = TRIE;

            /* store the type in the flags */
            convert->flags = nodetype;
            DEBUG_r({
            optimize = convert 
                      + NODE_STEP_REGNODE 
                      + regarglen[ OP( convert ) ];
            });
            /* XXX We really should free up the resource in trie now, 
                   as we won't use them - (which resources?) dmq */
        }
        /* needed for dumping*/
        DEBUG_r(if (optimize) {
            regnode *opt = convert;

            while ( ++opt < optimize) {
                Set_Node_Offset_Length(opt,0,0);
            }
            /* 
                Try to clean up some of the debris left after the 
                optimisation.
             */
            while( optimize < jumper ) {
                mjd_nodelen += Node_Length((optimize));
                OP( optimize ) = OPTIMIZED;
                Set_Node_Offset_Length(optimize,0,0);
                optimize++;
            }
            Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
        });
    } /* end node insert */

    /*  Finish populating the prev field of the wordinfo array.  Walk back
     *  from each accept state until we find another accept state, and if
     *  so, point the first word's .prev field at the second word. If the
     *  second already has a .prev field set, stop now. This will be the
     *  case either if we've already processed that word's accept state,
     *  or that state had multiple words, and the overspill words were
     *  already linked up earlier.
     */
    {
	U16 word;
	U32 state;
	U16 prev;

	for (word=1; word <= trie->wordcount; word++) {
	    prev = 0;
	    if (trie->wordinfo[word].prev)
		continue;
	    state = trie->wordinfo[word].accept;
	    while (state) {
		state = prev_states[state];
		if (!state)
		    break;
		prev = trie->states[state].wordnum;
		if (prev)
		    break;
	    }
	    trie->wordinfo[word].prev = prev;
	}
	Safefree(prev_states);
    }


    /* and now dump out the compressed format */
    DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));

    RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
#ifdef DEBUGGING
    RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
    RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
#else
    SvREFCNT_dec(revcharmap);
#endif
    return trie->jump 
           ? MADE_JUMP_TRIE 
           : trie->startstate>1 
             ? MADE_EXACT_TRIE 
             : MADE_TRIE;
}

STATIC void
S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source,  regnode *stclass, U32 depth)
{
/* The Trie is constructed and compressed now so we can build a fail array if it's needed

   This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
   "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
   ISBN 0-201-10088-6

   We find the fail state for each state in the trie, this state is the longest proper
   suffix of the current state's 'word' that is also a proper prefix of another word in our
   trie. State 1 represents the word '' and is thus the default fail state. This allows
   the DFA not to have to restart after its tried and failed a word at a given point, it
   simply continues as though it had been matching the other word in the first place.
   Consider
      'abcdgu'=~/abcdefg|cdgu/
   When we get to 'd' we are still matching the first word, we would encounter 'g' which would
   fail, which would bring us to the state representing 'd' in the second word where we would
   try 'g' and succeed, proceeding to match 'cdgu'.
 */
 /* add a fail transition */
    const U32 trie_offset = ARG(source);
    reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
    U32 *q;
    const U32 ucharcount = trie->uniquecharcount;
    const U32 numstates = trie->statecount;
    const U32 ubound = trie->lasttrans + ucharcount;
    U32 q_read = 0;
    U32 q_write = 0;
    U32 charid;
    U32 base = trie->states[ 1 ].trans.base;
    U32 *fail;
    reg_ac_data *aho;
    const U32 data_slot = add_data( pRExC_state, 1, "T" );
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
#ifndef DEBUGGING
    PERL_UNUSED_ARG(depth);
#endif


    ARG_SET( stclass, data_slot );
    aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
    RExC_rxi->data->data[ data_slot ] = (void*)aho;
    aho->trie=trie_offset;
    aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
    Copy( trie->states, aho->states, numstates, reg_trie_state );
    Newxz( q, numstates, U32);
    aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
    aho->refcount = 1;
    fail = aho->fail;
    /* initialize fail[0..1] to be 1 so that we always have
       a valid final fail state */
    fail[ 0 ] = fail[ 1 ] = 1;

    for ( charid = 0; charid < ucharcount ; charid++ ) {
	const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
	if ( newstate ) {
            q[ q_write ] = newstate;
            /* set to point at the root */
            fail[ q[ q_write++ ] ]=1;
        }
    }
    while ( q_read < q_write) {
	const U32 cur = q[ q_read++ % numstates ];
        base = trie->states[ cur ].trans.base;

        for ( charid = 0 ; charid < ucharcount ; charid++ ) {
	    const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
	    if (ch_state) {
                U32 fail_state = cur;
                U32 fail_base;
                do {
                    fail_state = fail[ fail_state ];
                    fail_base = aho->states[ fail_state ].trans.base;
                } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );

                fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
                fail[ ch_state ] = fail_state;
                if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
                {
                        aho->states[ ch_state ].wordnum =  aho->states[ fail_state ].wordnum;
                }
                q[ q_write++ % numstates] = ch_state;
            }
        }
    }
    /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
       when we fail in state 1, this allows us to use the
       charclass scan to find a valid start char. This is based on the principle
       that theres a good chance the string being searched contains lots of stuff
       that cant be a start char.
     */
    fail[ 0 ] = fail[ 1 ] = 0;
    DEBUG_TRIE_COMPILE_r({
        PerlIO_printf(Perl_debug_log,
		      "%*sStclass Failtable (%"UVuf" states): 0", 
		      (int)(depth * 2), "", (UV)numstates
        );
        for( q_read=1; q_read<numstates; q_read++ ) {
            PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
        }
        PerlIO_printf(Perl_debug_log, "\n");
    });
    Safefree(q);
    /*RExC_seen |= REG_SEEN_TRIEDFA;*/
}


/*
 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
 * These need to be revisited when a newer toolchain becomes available.
 */
#if defined(__sparc64__) && defined(__GNUC__)
#   if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
#       undef  SPARC64_GCC_WORKAROUND
#       define SPARC64_GCC_WORKAROUND 1
#   endif
#endif

#define DEBUG_PEEP(str,scan,depth) \
    DEBUG_OPTIMISE_r({if (scan){ \
       SV * const mysv=sv_newmortal(); \
       regnode *Next = regnext(scan); \
       regprop(RExC_rx, mysv, scan); \
       PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
       (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
       Next ? (REG_NODE_NUM(Next)) : 0 ); \
   }});


/* The below joins as many adjacent EXACTish nodes as possible into a single
 * one, and looks for problematic sequences of characters whose folds vs.
 * non-folds have sufficiently different lengths, that the optimizer would be
 * fooled into rejecting legitimate matches of them, and the trie construction
 * code can't cope with them.  The joining is only done if:
 * 1) there is room in the current conglomerated node to entirely contain the
 *    next one.
 * 2) they are the exact same node type
 *
 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
 * these get optimized out
 *
 * If there are problematic code sequences, *min_subtract is set to the delta
 * that the minimum size of the node can be less than its actual size.  And,
 * the node type of the result is changed to reflect that it contains these
 * sequences.
 *
 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
 * and contains LATIN SMALL LETTER SHARP S
 *
 * This is as good a place as any to discuss the design of handling these
 * problematic sequences.  It's been wrong in Perl for a very long time.  There
 * are three code points in Unicode whose folded lengths differ so much from
 * the un-folded lengths that it causes problems for the optimizer and trie
 * construction.  Why only these are problematic, and not others where lengths
 * also differ is something I (khw) do not understand.  New versions of Unicode
 * might add more such code points.  Hopefully the logic in fold_grind.t that
 * figures out what to test (in part by verifying that each size-combination
 * gets tested) will catch any that do come along, so they can be added to the
 * special handling below.  The chances of new ones are actually rather small,
 * as most, if not all, of the world's scripts that have casefolding have
 * already been encoded by Unicode.  Also, a number of Unicode's decisions were
 * made to allow compatibility with pre-existing standards, and almost all of
 * those have already been dealt with.  These would otherwise be the most
 * likely candidates for generating further tricky sequences.  In other words,
 * Unicode by itself is unlikely to add new ones unless it is for compatibility
 * with pre-existing standards, and there aren't many of those left.
 *
 * The previous designs for dealing with these involved assigning a special
 * node for them.  This approach doesn't work, as evidenced by this example:
 *      "\xDFs" =~ /s\xDF/ui    # Used to fail before these patches
 * Both these fold to "sss", but if the pattern is parsed to create a node of
 * that would match just the \xDF, it won't be able to handle the case where a
 * successful match would have to cross the node's boundary.  The new approach
 * that hopefully generally solves the problem generates an EXACTFU_SS node
 * that is "sss".
 *
 * There are a number of components to the approach (a lot of work for just
 * three code points!):
 * 1)   This routine examines each EXACTFish node that could contain the
 *      problematic sequences.  It returns in *min_subtract how much to
 *      subtract from the the actual length of the string to get a real minimum
 *      for one that could match it.  This number is usually 0 except for the
 *      problematic sequences.  This delta is used by the caller to adjust the
 *      min length of the match, and the delta between min and max, so that the
 *      optimizer doesn't reject these possibilities based on size constraints.
 * 2)   These sequences are not currently correctly handled by the trie code
 *      either, so it changes the joined node type to ops that are not handled
 *      by trie's, those new ops being EXACTFU_SS and EXACTFU_NO_TRIE.
 * 3)   This is sufficient for the two Greek sequences (described below), but
 *      the one involving the Sharp s (\xDF) needs more.  The node type
 *      EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
 *      sequence in it.  For non-UTF-8 patterns and strings, this is the only
 *      case where there is a possible fold length change.  That means that a
 *      regular EXACTFU node without UTF-8 involvement doesn't have to concern
 *      itself with length changes, and so can be processed faster.  regexec.c
 *      takes advantage of this.  Generally, an EXACTFish node that is in UTF-8
 *      is pre-folded by regcomp.c.  This saves effort in regex matching.
 *      However, probably mostly for historical reasons, the pre-folding isn't
 *      done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
 *      nodes, as what they fold to isn't known until runtime.)  The fold
 *      possibilities for the non-UTF8 patterns are quite simple, except for
 *      the sharp s.  All the ones that don't involve a UTF-8 target string
 *      are members of a fold-pair, and arrays are set up for all of them
 *      that quickly find the other member of the pair.  It might actually
 *      be faster to pre-fold these, but it isn't currently done, except for
 *      the sharp s.  Code elsewhere in this file makes sure that it gets
 *      folded to 'ss', even if the pattern isn't UTF-8.  This avoids the
 *      issues described in the next item.
 * 4)   A problem remains for the sharp s in EXACTF nodes.  Whether it matches
 *      'ss' or not is not knowable at compile time.  It will match iff the
 *      target string is in UTF-8, unlike the EXACTFU nodes, where it always
 *      matches; and the EXACTFL and EXACTFA nodes where it never does.  Thus
 *      it can't be folded to "ss" at compile time, unlike EXACTFU does as
 *      described in item 3).  An assumption that the optimizer part of
 *      regexec.c (probably unwittingly) makes is that a character in the
 *      pattern corresponds to at most a single character in the target string.
 *      (And I do mean character, and not byte here, unlike other parts of the
 *      documentation that have never been updated to account for multibyte
 *      Unicode.)  This assumption is wrong only in this case, as all other
 *      cases are either 1-1 folds when no UTF-8 is involved; or is true by
 *      virtue of having this file pre-fold UTF-8 patterns.   I'm
 *      reluctant to try to change this assumption, so instead the code punts.
 *      This routine examines EXACTF nodes for the sharp s, and returns a
 *      boolean indicating whether or not the node is an EXACTF node that
 *      contains a sharp s.  When it is true, the caller sets a flag that later
 *      causes the optimizer in this file to not set values for the floating
 *      and fixed string lengths, and thus avoids the optimizer code in
 *      regexec.c that makes the invalid assumption.  Thus, there is no
 *      optimization based on string lengths for EXACTF nodes that contain the
 *      sharp s.  This only happens for /id rules (which means the pattern
 *      isn't in UTF-8).
 */

#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
    if (PL_regkind[OP(scan)] == EXACT) \
        join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)

STATIC U32
S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, UV *min_subtract, bool *has_exactf_sharp_s, U32 flags,regnode *val, U32 depth) {
    /* Merge several consecutive EXACTish nodes into one. */
    regnode *n = regnext(scan);
    U32 stringok = 1;
    regnode *next = scan + NODE_SZ_STR(scan);
    U32 merged = 0;
    U32 stopnow = 0;
#ifdef DEBUGGING
    regnode *stop = scan;
    GET_RE_DEBUG_FLAGS_DECL;
#else
    PERL_UNUSED_ARG(depth);
#endif

    PERL_ARGS_ASSERT_JOIN_EXACT;
#ifndef EXPERIMENTAL_INPLACESCAN
    PERL_UNUSED_ARG(flags);
    PERL_UNUSED_ARG(val);
#endif
    DEBUG_PEEP("join",scan,depth);

    /* Look through the subsequent nodes in the chain.  Skip NOTHING, merge
     * EXACT ones that are mergeable to the current one. */
    while (n
           && (PL_regkind[OP(n)] == NOTHING
               || (stringok && OP(n) == OP(scan)))
           && NEXT_OFF(n)
           && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
    {
        
        if (OP(n) == TAIL || n > next)
            stringok = 0;
        if (PL_regkind[OP(n)] == NOTHING) {
            DEBUG_PEEP("skip:",n,depth);
            NEXT_OFF(scan) += NEXT_OFF(n);
            next = n + NODE_STEP_REGNODE;
#ifdef DEBUGGING
            if (stringok)
                stop = n;
#endif
            n = regnext(n);
        }
        else if (stringok) {
            const unsigned int oldl = STR_LEN(scan);
            regnode * const nnext = regnext(n);

            if (oldl + STR_LEN(n) > U8_MAX)
                break;
            
            DEBUG_PEEP("merg",n,depth);
            merged++;

            NEXT_OFF(scan) += NEXT_OFF(n);
            STR_LEN(scan) += STR_LEN(n);
            next = n + NODE_SZ_STR(n);
            /* Now we can overwrite *n : */
            Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
#ifdef DEBUGGING
            stop = next - 1;
#endif
            n = nnext;
            if (stopnow) break;
        }

#ifdef EXPERIMENTAL_INPLACESCAN
	if (flags && !NEXT_OFF(n)) {
	    DEBUG_PEEP("atch", val, depth);
	    if (reg_off_by_arg[OP(n)]) {
		ARG_SET(n, val - n);
	    }
	    else {
		NEXT_OFF(n) = val - n;
	    }
	    stopnow = 1;
	}
#endif
    }

    *min_subtract = 0;
    *has_exactf_sharp_s = FALSE;

    /* Here, all the adjacent mergeable EXACTish nodes have been merged.  We
     * can now analyze for sequences of problematic code points.  (Prior to
     * this final joining, sequences could have been split over boundaries, and
     * hence missed).  The sequences only happen in folding, hence for any
     * non-EXACT EXACTish node */
    if (OP(scan) != EXACT) {
        U8 *s;
        U8 * s0 = (U8*) STRING(scan);
        U8 * const s_end = s0 + STR_LEN(scan);

	/* The below is perhaps overboard, but this allows us to save a test
	 * each time through the loop at the expense of a mask.  This is
	 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
	 * single bit.  On ASCII they are 32 apart; on EBCDIC, they are 64.
	 * This uses an exclusive 'or' to find that bit and then inverts it to
	 * form a mask, with just a single 0, in the bit position where 'S' and
	 * 's' differ. */
	const U8 S_or_s_mask = ~ ('S' ^ 's');
	const U8 s_masked = 's' & S_or_s_mask;

	/* One pass is made over the node's string looking for all the
	 * possibilities.  to avoid some tests in the loop, there are two main
	 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
	 * non-UTF-8 */
	if (UTF) {

	    /* There are two problematic Greek code points in Unicode
	     * casefolding
	     *
	     * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
	     * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
	     *
	     * which casefold to
	     *
	     * Unicode                      UTF-8
	     *
	     * U+03B9 U+0308 U+0301         0xCE 0xB9 0xCC 0x88 0xCC 0x81
	     * U+03C5 U+0308 U+0301         0xCF 0x85 0xCC 0x88 0xCC 0x81
             *
	     * This means that in case-insensitive matching (or "loose
	     * matching", as Unicode calls it), an EXACTF of length six (the
	     * UTF-8 encoded byte length of the above casefolded versions) can
	     * match a target string of length two (the byte length of UTF-8
	     * encoded U+0390 or U+03B0).  This would rather mess up the
	     * minimum length computation.  (there are other code points that
	     * also fold to these two sequences, but the delta is smaller)
	     *
	     * If these sequences are found, the minimum length is decreased by
	     * four (six minus two).
	     *
	     * Similarly, 'ss' may match the single char and byte LATIN SMALL
	     * LETTER SHARP S.  We decrease the min length by 1 for each
	     * occurrence of 'ss' found */

#ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
#	    define U390_first_byte 0xb4
	    const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
#	    define U3B0_first_byte 0xb5
	    const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
#else
#	    define U390_first_byte 0xce
	    const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
#	    define U3B0_first_byte 0xcf
	    const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
#endif
	    const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
						 yields a net of 0 */
	    /* Examine the string for one of the problematic sequences */
	    for (s = s0;
		 s < s_end - 1; /* Can stop 1 before the end, as minimum length
				 * sequence we are looking for is 2 */
		 s += UTF8SKIP(s))
	    {

		/* Look for the first byte in each problematic sequence */
		switch (*s) {
		    /* We don't have to worry about other things that fold to
		     * 's' (such as the long s, U+017F), as all above-latin1
		     * code points have been pre-folded */
		    case 's':
		    case 'S':

                        /* Current character is an 's' or 'S'.  If next one is
                         * as well, we have the dreaded sequence */
			if (((*(s+1) & S_or_s_mask) == s_masked)
			    /* These two node types don't have special handling
			     * for 'ss' */
			    && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
			{
			    *min_subtract += 1;
			    OP(scan) = EXACTFU_SS;
			    s++;    /* No need to look at this character again */
			}
			break;

		    case U390_first_byte:
			if (s_end - s >= len

			    /* The 1's are because are skipping comparing the
			     * first byte */
			    && memEQ(s + 1, U390_tail, len - 1))
			{
			    goto greek_sequence;
			}
			break;

		    case U3B0_first_byte:
			if (! (s_end - s >= len
			       && memEQ(s + 1, U3B0_tail, len - 1)))
			{
			    break;
			}
		      greek_sequence:
			*min_subtract += 4;

			/* This can't currently be handled by trie's, so change
			 * the node type to indicate this.  If EXACTFA and
			 * EXACTFL were ever to be handled by trie's, this
			 * would have to be changed.  If this node has already
			 * been changed to EXACTFU_SS in this loop, leave it as
			 * is.  (I (khw) think it doesn't matter in regexec.c
			 * for UTF patterns, but no need to change it */
			if (OP(scan) == EXACTFU) {
			    OP(scan) = EXACTFU_NO_TRIE;
			}
			s += 6;	/* We already know what this sequence is.  Skip
				   the rest of it */
			break;
		}
	    }
	}
	else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {

	    /* Here, the pattern is not UTF-8.  We need to look only for the
	     * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
	     * in the final position.  Otherwise we can stop looking 1 byte
	     * earlier because have to find both the first and second 's' */
	    const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;

	    for (s = s0; s < upper; s++) {
		switch (*s) {
		    case 'S':
		    case 's':
			if (s_end - s > 1
			    && ((*(s+1) & S_or_s_mask) == s_masked))
			{
			    *min_subtract += 1;

			    /* EXACTF nodes need to know that the minimum
			     * length changed so that a sharp s in the string
			     * can match this ss in the pattern, but they
			     * remain EXACTF nodes, as they are not trie'able,
			     * so don't have to invent a new node type to
			     * exclude them from the trie code */
			    if (OP(scan) != EXACTF) {
				OP(scan) = EXACTFU_SS;
			    }
			    s++;
			}
			break;
		    case LATIN_SMALL_LETTER_SHARP_S:
			if (OP(scan) == EXACTF) {
			    *has_exactf_sharp_s = TRUE;
			}
			break;
		}
	    }
	}
    }

#ifdef DEBUGGING
    /* Allow dumping but overwriting the collection of skipped
     * ops and/or strings with fake optimized ops */
    n = scan + NODE_SZ_STR(scan);
    while (n <= stop) {
	OP(n) = OPTIMIZED;
	FLAGS(n) = 0;
	NEXT_OFF(n) = 0;
        n++;
    }
#endif
    DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
    return stopnow;
}

/* REx optimizer.  Converts nodes into quicker variants "in place".
   Finds fixed substrings.  */

/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
   to the position after last scanned or to NULL. */

#define INIT_AND_WITHP \
    assert(!and_withp); \
    Newx(and_withp,1,struct regnode_charclass_class); \
    SAVEFREEPV(and_withp)

/* this is a chain of data about sub patterns we are processing that
   need to be handled separately/specially in study_chunk. Its so
   we can simulate recursion without losing state.  */
struct scan_frame;
typedef struct scan_frame {
    regnode *last;  /* last node to process in this frame */
    regnode *next;  /* next node to process when last is reached */
    struct scan_frame *prev; /*previous frame*/
    I32 stop; /* what stopparen do we use */
} scan_frame;


#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)

#define CASE_SYNST_FNC(nAmE)                                       \
case nAmE:                                                         \
    if (flags & SCF_DO_STCLASS_AND) {                              \
	    for (value = 0; value < 256; value++)                  \
		if (!is_ ## nAmE ## _cp(value))                       \
		    ANYOF_BITMAP_CLEAR(data->start_class, value);  \
    }                                                              \
    else {                                                         \
	    for (value = 0; value < 256; value++)                  \
		if (is_ ## nAmE ## _cp(value))                        \
		    ANYOF_BITMAP_SET(data->start_class, value);	   \
    }                                                              \
    break;                                                         \
case N ## nAmE:                                                    \
    if (flags & SCF_DO_STCLASS_AND) {                              \
	    for (value = 0; value < 256; value++)                   \
		if (is_ ## nAmE ## _cp(value))                         \
		    ANYOF_BITMAP_CLEAR(data->start_class, value);   \
    }                                                               \
    else {                                                          \
	    for (value = 0; value < 256; value++)                   \
		if (!is_ ## nAmE ## _cp(value))                        \
		    ANYOF_BITMAP_SET(data->start_class, value);	    \
    }                                                               \
    break



STATIC I32
S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
                        I32 *minlenp, I32 *deltap,
			regnode *last,
			scan_data_t *data,
			I32 stopparen,
			U8* recursed,
			struct regnode_charclass_class *and_withp,
			U32 flags, U32 depth)
			/* scanp: Start here (read-write). */
			/* deltap: Write maxlen-minlen here. */
			/* last: Stop before this one. */
			/* data: string data about the pattern */
			/* stopparen: treat close N as END */
			/* recursed: which subroutines have we recursed into */
			/* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
{
    dVAR;
    I32 min = 0, pars = 0, code;
    regnode *scan = *scanp, *next;
    I32 delta = 0;
    int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
    int is_inf_internal = 0;		/* The studied chunk is infinite */
    I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
    scan_data_t data_fake;
    SV *re_trie_maxbuff = NULL;
    regnode *first_non_open = scan;
    I32 stopmin = I32_MAX;
    scan_frame *frame = NULL;
    GET_RE_DEBUG_FLAGS_DECL;

    PERL_ARGS_ASSERT_STUDY_CHUNK;

#ifdef DEBUGGING
    StructCopy(&zero_scan_data, &data_fake, scan_data_t);
#endif

    if ( depth == 0 ) {
        while (first_non_open && OP(first_non_open) == OPEN)
            first_non_open=regnext(first_non_open);
    }


  fake_study_recurse:
    while ( scan && OP(scan) != END && scan < last ){
        UV min_subtract = 0;    /* How much to subtract from the minimum node
                                   length to get a real minimum (because the
                                   folded version may be shorter) */
	bool has_exactf_sharp_s = FALSE;
	/* Peephole optimizer: */
	DEBUG_STUDYDATA("Peep:", data,depth);
	DEBUG_PEEP("Peep",scan,depth);

        /* Its not clear to khw or hv why this is done here, and not in the
         * clauses that deal with EXACT nodes.  khw's guess is that it's
         * because of a previous design */
        JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);

	/* Follow the next-chain of the current node and optimize
	   away all the NOTHINGs from it.  */
	if (OP(scan) != CURLYX) {
	    const int max = (reg_off_by_arg[OP(scan)]
		       ? I32_MAX
		       /* I32 may be smaller than U16 on CRAYs! */
		       : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
	    int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
	    int noff;
	    regnode *n = scan;

	    /* Skip NOTHING and LONGJMP. */
	    while ((n = regnext(n))
		   && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
		       || ((OP(n) == LONGJMP) && (noff = ARG(n))))
		   && off + noff < max)
		off += noff;
	    if (reg_off_by_arg[OP(scan)])
		ARG(scan) = off;
	    else
		NEXT_OFF(scan) = off;
	}



	/* The principal pseudo-switch.  Cannot be a switch, since we
	   look into several different things.  */
	if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
		   || OP(scan) == IFTHEN) {
	    next = regnext(scan);
	    code = OP(scan);
	    /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */

	    if (OP(next) == code || code == IFTHEN) {
	        /* NOTE - There is similar code to this block below for handling
	           TRIE nodes on a re-study.  If you change stuff here check there
	           too. */
		I32 max1 = 0, min1 = I32_MAX, num = 0;
		struct regnode_charclass_class accum;
		regnode * const startbranch=scan;

		if (flags & SCF_DO_SUBSTR)
		    SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
		if (flags & SCF_DO_STCLASS)
		    cl_init_zero(pRExC_state, &accum);

		while (OP(scan) == code) {
		    I32 deltanext, minnext, f = 0, fake;
		    struct regnode_charclass_class this_class;

		    num++;
		    data_fake.flags = 0;
		    if (data) {
			data_fake.whilem_c = data->whilem_c;
			data_fake.last_closep = data->last_closep;
		    }
		    else
			data_fake.last_closep = &fake;

		    data_fake.pos_delta = delta;
		    next = regnext(scan);
		    scan = NEXTOPER(scan);
		    if (code != BRANCH)
			scan = NEXTOPER(scan);
		    if (flags & SCF_DO_STCLASS) {
			cl_init(pRExC_state, &this_class);
			data_fake.start_class = &this_class;
			f = SCF_DO_STCLASS_AND;
		    }
		    if (flags & SCF_WHILEM_VISITED_POS)
			f |= SCF_WHILEM_VISITED_POS;

		    /* we suppose the run is continuous, last=next...*/
		    minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
					  next, &data_fake,
					  stopparen, recursed, NULL, f,depth+1);
		    if (min1 > minnext)
			min1 = minnext;
		    if (max1 < minnext + deltanext)
			max1 = minnext + deltanext;
		    if (deltanext == I32_MAX)
			is_inf = is_inf_internal = 1;
		    scan = next;
		    if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
			pars++;
	            if (data_fake.flags & SCF_SEEN_ACCEPT) {
	                if ( stopmin > minnext) 
	                    stopmin = min + min1;
	                flags &= ~SCF_DO_SUBSTR;
	                if (data)
	                    data->flags |= SCF_SEEN_ACCEPT;
	            }
		    if (data) {
			if (data_fake.flags & SF_HAS_EVAL)
			    data->flags |= SF_HAS_EVAL;
			data->whilem_c = data_fake.whilem_c;
		    }
		    if (flags & SCF_DO_STCLASS)
			cl_or(pRExC_state, &accum, &this_class);
		}
		if (code == IFTHEN && num < 2) /* Empty ELSE branch */
		    min1 = 0;
		if (flags & SCF_DO_SUBSTR) {
		    data->pos_min += min1;
		    data->pos_delta += max1 - min1;
		    if (max1 != min1 || is_inf)
			data->longest = &(data->longest_float);
		}
		min += min1;
		delta += max1 - min1;
		if (flags & SCF_DO_STCLASS_OR) {
		    cl_or(pRExC_state, data->start_class, &accum);
		    if (min1) {
			cl_and(data->start_class, and_withp);
			flags &= ~SCF_DO_STCLASS;
		    }
		}
		else if (flags & SCF_DO_STCLASS_AND) {
		    if (min1) {
			cl_and(data->start_class, &accum);
			flags &= ~SCF_DO_STCLASS;
		    }
		    else {
			/* Switch to OR mode: cache the old value of
			 * data->start_class */
			INIT_AND_WITHP;
			StructCopy(data->start_class, and_withp,
				   struct regnode_charclass_class);
			flags &= ~SCF_DO_STCLASS_AND;
			StructCopy(&accum, data->start_class,
				   struct regnode_charclass_class);
			flags |= SCF_DO_STCLASS_OR;
			data->start_class->flags |= ANYOF_EOS;
		    }
		}

                if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
		/* demq.

		   Assuming this was/is a branch we are dealing with: 'scan' now
		   points at the item that follows the branch sequence, whatever
		   it is. We now start at the beginning of the sequence and look
		   for subsequences of

		   BRANCH->EXACT=>x1
		   BRANCH->EXACT=>x2
		   tail

		   which would be constructed from a pattern like /A|LIST|OF|WORDS/

		   If we can find such a subsequence we need to turn the first
		   element into a trie and then add the subsequent branch exact
		   strings to the trie.

		   We have two cases

		     1. patterns where the whole set of branches can be converted. 

		     2. patterns where only a subset can be converted.

		   In case 1 we can replace the whole set with a single regop
		   for the trie. In case 2 we need to keep the start and end
		   branches so

		     'BRANCH EXACT; BRANCH EXACT; BRANCH X'
		     becomes BRANCH TRIE; BRANCH X;

		  There is an additional case, that being where there is a 
		  common prefix, which gets split out into an EXACT like node
		  preceding the TRIE node.

		  If x(1..n)==tail then we can do a simple trie, if not we make
		  a "jump" trie, such that when we match the appropriate word
		  we "jump" to the appropriate tail node. Essentially we turn
		  a nested if into a case structure of sorts.

		*/

		    int made=0;
		    if (!re_trie_maxbuff) {
			re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
			if (!SvIOK(re_trie_maxbuff))
			    sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
		    }
                    if ( SvIV(re_trie_maxbuff)>=0  ) {
                        regnode *cur;
                        regnode *first = (regnode *)NULL;
                        regnode *last = (regnode *)NULL;
                        regnode *tail = scan;
                        U8 optype = 0;
                        U32 count=0;

#ifdef DEBUGGING
                        SV * const mysv = sv_newmortal();       /* for dumping */
#endif
                        /* var tail is used because there may be a TAIL
                           regop in the way. Ie, the exacts will point to the
                           thing following the TAIL, but the last branch will
                           point at the TAIL. So we advance tail. If we
                           have nested (?:) we may have to move through several
                           tails.
                         */

                        while ( OP( tail ) == TAIL ) {
                            /* this is the TAIL generated by (?:) */
                            tail = regnext( tail );
                        }

                        
                        DEBUG_OPTIMISE_r({
                            regprop(RExC_rx, mysv, tail );
                            PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
                                (int)depth * 2 + 2, "", 
                                "Looking for TRIE'able sequences. Tail node is: ", 
                                SvPV_nolen_const( mysv )
                            );
                        });
                        
                        /*

                           step through the branches, cur represents each
                           branch, noper is the first thing to be matched
                           as part of that branch and noper_next is the
                           regnext() of that node. if noper is an EXACT
                           and noper_next is the same as scan (our current
                           position in the regex) then the EXACT branch is
                           a possible optimization target. Once we have
                           two or more consecutive such branches we can
                           create a trie of the EXACT's contents and stich
                           it in place. If the sequence represents all of
                           the branches we eliminate the whole thing and
                           replace it with a single TRIE. If it is a
                           subsequence then we need to stitch it in. This
                           means the first branch has to remain, and needs
                           to be repointed at the item on the branch chain
                           following the last branch optimized. This could
                           be either a BRANCH, in which case the
                           subsequence is internal, or it could be the
                           item following the branch sequence in which
                           case the subsequence is at the end.

                        */

                        /* dont use tail as the end marker for this traverse */
                        for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
                            regnode * const noper = NEXTOPER( cur );
#if defined(DEBUGGING) || defined(NOJUMPTRIE)
                            regnode * const noper_next = regnext( noper );
#endif

                            DEBUG_OPTIMISE_r({
                                regprop(RExC_rx, mysv, cur);
                                PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
                                   (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );

                                regprop(RExC_rx, mysv, noper);
                                PerlIO_printf( Perl_debug_log, " -> %s",
                                    SvPV_nolen_const(mysv));

                                if ( noper_next ) {
                                  regprop(RExC_rx, mysv, noper_next );
                                  PerlIO_printf( Perl_debug_log,"\t=> %s\t",
                                    SvPV_nolen_const(mysv));
                                }
                                PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
                                   REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
                            });
                            if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
                                         : PL_regkind[ OP( noper ) ] == EXACT )
                                  || OP(noper) == NOTHING )
#ifdef NOJUMPTRIE
                                  && noper_next == tail
#endif
                                  && count < U16_MAX)
                            {
                                count++;
                                if ( !first || optype == NOTHING ) {
                                    if (!first) first = cur;
                                    optype = OP( noper );
                                } else {
                                    last = cur;
                                }
                            } else {
/* 
    Currently the trie logic handles case insensitive matching properly only
    when the pattern is UTF-8 and the node is EXACTFU (thus forcing unicode
    semantics).

    If/when this is fixed the following define can be swapped
    in below to fully enable trie logic.

#define TRIE_TYPE_IS_SAFE 1

Note that join_exact() assumes that the other types of EXACTFish nodes are not
used in tries, so that would have to be updated if this changed

*/
#define TRIE_TYPE_IS_SAFE ((UTF && optype == EXACTFU) || optype==EXACT)

                                if ( last && TRIE_TYPE_IS_SAFE ) {
                                    make_trie( pRExC_state, 
                                            startbranch, first, cur, tail, count, 
                                            optype, depth+1 );
                                }
                                if ( PL_regkind[ OP( noper ) ] == EXACT
#ifdef NOJUMPTRIE
                                     && noper_next == tail
#endif
                                ){
                                    count = 1;
                                    first = cur;
                                    optype = OP( noper );
                                } else {
                                    count = 0;
                                    first = NULL;
                                    optype = 0;
                                }
                                last = NULL;
                            }
                        }
                        DEBUG_OPTIMISE_r({
                            regprop(RExC_rx, mysv, cur);
                            PerlIO_printf( Perl_debug_log,
                              "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
                              "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));

                        });
                        
                        if ( last && TRIE_TYPE_IS_SAFE ) {
                            made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
#ifdef TRIE_STUDY_OPT
                            if ( ((made == MADE_EXACT_TRIE && 
                                 startbranch == first) 
                                 || ( first_non_open == first )) && 
                                 depth==0 ) {
                                flags |= SCF_TRIE_RESTUDY;
                                if ( startbranch == first 
                                     && scan == tail ) 
                                {
                                    RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
                                }
                            }
#endif
                        }
                    }
                    
                } /* do trie */
                
	    }
	    else if ( code == BRANCHJ ) {  /* single branch is optimized. */
		scan = NEXTOPER(NEXTOPER(scan));
	    } else			/* single branch is optimized. */
		scan = NEXTOPER(scan);
	    continue;
	} else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
	    scan_frame *newframe = NULL;
	    I32 paren;
	    regnode *start;
	    regnode *end;

	    if (OP(scan) != SUSPEND) {
	    /* set the pointer */
	        if (OP(scan) == GOSUB) {
	            paren = ARG(scan);
	            RExC_recurse[ARG2L(scan)] = scan;
                    start = RExC_open_parens[paren-1];
                    end   = RExC_close_parens[paren-1];
                } else {
                    paren = 0;
                    start = RExC_rxi->program + 1;
                    end   = RExC_opend;
                }
                if (!recursed) {
                    Newxz(recursed, (((RExC_npar)>>3) +1), U8);
                    SAVEFREEPV(recursed);
                }
                if (!PAREN_TEST(recursed,paren+1)) {
		    PAREN_SET(recursed,paren+1);
                    Newx(newframe,1,scan_frame);
                } else {
                    if (flags & SCF_DO_SUBSTR) {
                        SCAN_COMMIT(pRExC_state,data,minlenp);
                        data->longest = &(data->longest_float);
                    }
                    is_inf = is_inf_internal = 1;
                    if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
                        cl_anything(pRExC_state, data->start_class);
                    flags &= ~SCF_DO_STCLASS;
	        }
            } else {
	        Newx(newframe,1,scan_frame);
	        paren = stopparen;
	        start = scan+2;
	        end = regnext(scan);
	    }
	    if (newframe) {
                assert(start);
                assert(end);
	        SAVEFREEPV(newframe);
	        newframe->next = regnext(scan);
	        newframe->last = last;
	        newframe->stop = stopparen;
	        newframe->prev = frame;

	        frame = newframe;
	        scan =  start;
	        stopparen = paren;
	        last = end;

	        continue;
	    }
	}
	else if (OP(scan) == EXACT) {
	    I32 l = STR_LEN(scan);
	    UV uc;
	    if (UTF) {
		const U8 * const s = (U8*)STRING(scan);
		l = utf8_length(s, s + l);
		uc = utf8_to_uvchr(s, NULL);
	    } else {
		uc = *((U8*)STRING(scan));
	    }
	    min += l;
	    if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
		/* The code below prefers earlier match for fixed
		   offset, later match for variable offset.  */
		if (data->last_end == -1) { /* Update the start info. */
		    data->last_start_min = data->pos_min;
 		    data->last_start_max = is_inf
 			? I32_MAX : data->pos_min + data->pos_delta;
		}
		sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
		if (UTF)
		    SvUTF8_on(data->last_found);
		{
		    SV * const sv = data->last_found;
		    MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
			mg_find(sv, PERL_MAGIC_utf8) : NULL;
		    if (mg && mg->mg_len >= 0)
			mg->mg_len += utf8_length((U8*)STRING(scan),
						  (U8*)STRING(scan)+STR_LEN(scan));
		}
		data->last_end = data->pos_min + l;
		data->pos_min += l; /* As in the first entry. */
		data->flags &= ~SF_BEFORE_EOL;
	    }
	    if (flags & SCF_DO_STCLASS_AND) {
		/* Check whether it is compatible with what we know already! */
		int compat = 1;


		/* If compatible, we or it in below.  It is compatible if is
		 * in the bitmp and either 1) its bit or its fold is set, or 2)
		 * it's for a locale.  Even if there isn't unicode semantics
		 * here, at runtime there may be because of matching against a
		 * utf8 string, so accept a possible false positive for
		 * latin1-range folds */
		if (uc >= 0x100 ||
		    (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
		    && !ANYOF_BITMAP_TEST(data->start_class, uc)
		    && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
			|| !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
                    )
		{
		    compat = 0;
		}
		ANYOF_CLASS_ZERO(data->start_class);
		ANYOF_BITMAP_ZERO(data->start_class);
		if (compat)
		    ANYOF_BITMAP_SET(data->start_class, uc);
		else if (uc >= 0x100) {
		    int i;

		    /* Some Unicode code points fold to the Latin1 range; as
		     * XXX temporary code, instead of figuring out if this is
		     * one, just assume it is and set all the start class bits
		     * that could be some such above 255 code point's fold
		     * which will generate fals positives.  As the code
		     * elsewhere that does compute the fold settles down, it
		     * can be extracted out and re-used here */
		    for (i = 0; i < 256; i++){
			if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
			    ANYOF_BITMAP_SET(data->start_class, i);
			}
		    }
		}
		data->start_class->flags &= ~ANYOF_EOS;
		if (uc < 0x100)
		  data->start_class->flags &= ~ANYOF_UNICODE_ALL;
	    }
	    else if (flags & SCF_DO_STCLASS_OR) {
		/* false positive possible if the class is case-folded */
		if (uc < 0x100)
		    ANYOF_BITMAP_SET(data->start_class, uc);
		else
		    data->start_class->flags |= ANYOF_UNICODE_ALL;
		data->start_class->flags &= ~ANYOF_EOS;
		cl_and(data->start_class, and_withp);
	    }
	    flags &= ~SCF_DO_STCLASS;
	}
	else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
	    I32 l = STR_LEN(scan);
	    UV uc = *((U8*)STRING(scan));

	    /* Search for fixed substrings supports EXACT only. */
	    if (flags & SCF_DO_SUBSTR) {
		assert(data);
		SCAN_COMMIT(pRExC_state, data, minlenp);
	    }
	    if (UTF) {
		const U8 * const s = (U8 *)STRING(scan);
		l = utf8_length(s, s + l);
		uc = utf8_to_uvchr(s, NULL);
	    }
	    else if (has_exactf_sharp_s) {
		RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
	    }
	    min += l - min_subtract;
            if (min < 0) {
                min = 0;
            }
            delta += min_subtract;
	    if (flags & SCF_DO_SUBSTR) {
		data->pos_min += l - min_subtract;
		if (data->pos_min < 0) {
                    data->pos_min = 0;
                }
                data->pos_delta += min_subtract;
		if (min_subtract) {
		    data->longest = &(data->longest_float);
		}
	    }
	    if (flags & SCF_DO_STCLASS_AND) {
		/* Check whether it is compatible with what we know already! */
		int compat = 1;
		if (uc >= 0x100 ||
		 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
		  && !ANYOF_BITMAP_TEST(data->start_class, uc)
		  && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
		{
		    compat = 0;
		}
		ANYOF_CLASS_ZERO(data->start_class);
		ANYOF_BITMAP_ZERO(data->start_class);
		if (compat) {
		    ANYOF_BITMAP_SET(data->start_class, uc);
		    data->start_class->flags &= ~ANYOF_EOS;
		    data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
		    if (OP(scan) == EXACTFL) {
			/* XXX This set is probably no longer necessary, and
			 * probably wrong as LOCALE now is on in the initial
			 * state */
			data->start_class->flags |= ANYOF_LOCALE;
		    }
		    else {

			/* Also set the other member of the fold pair.  In case
			 * that unicode semantics is called for at runtime, use
			 * the full latin1 fold.  (Can't do this for locale,
			 * because not known until runtime) */
			ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);

                        /* All other (EXACTFL handled above) folds except under
                         * /iaa that include s, S, and sharp_s also may include
                         * the others */
			if (OP(scan) != EXACTFA) {
			    if (uc == 's' || uc == 'S') {
				ANYOF_BITMAP_SET(data->start_class,
					         LATIN_SMALL_LETTER_SHARP_S);
			    }
			    else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
				ANYOF_BITMAP_SET(data->start_class, 's');
				ANYOF_BITMAP_SET(data->start_class, 'S');
			    }
			}
		    }
		}
		else if (uc >= 0x100) {
		    int i;
		    for (i = 0; i < 256; i++){
			if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
			    ANYOF_BITMAP_SET(data->start_class, i);
			}
		    }
		}
	    }
	    else if (flags & SCF_DO_STCLASS_OR) {
		if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
		    /* false positive possible if the class is case-folded.
		       Assume that the locale settings are the same... */
		    if (uc < 0x100) {
			ANYOF_BITMAP_SET(data->start_class, uc);
                        if (OP(scan) != EXACTFL) {

                            /* And set the other member of the fold pair, but
                             * can't do that in locale because not known until
                             * run-time */
                            ANYOF_BITMAP_SET(data->start_class,
					     PL_fold_latin1[uc]);

			    /* All folds except under /iaa that include s, S,
			     * and sharp_s also may include the others */
			    if (OP(scan) != EXACTFA) {
				if (uc == 's' || uc == 'S') {
				    ANYOF_BITMAP_SET(data->start_class,
					           LATIN_SMALL_LETTER_SHARP_S);
				}
				else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
				    ANYOF_BITMAP_SET(data->start_class, 's');
				    ANYOF_BITMAP_SET(data->start_class, 'S');
				}
			    }
                        }
		    }
		    data->start_class->flags &= ~ANYOF_EOS;
		}
		cl_and(data->start_class, and_withp);
	    }
	    flags &= ~SCF_DO_STCLASS;
	}
	else if (REGNODE_VARIES(OP(scan))) {
	    I32 mincount, maxcount, minnext, deltanext, fl = 0;
	    I32 f = flags, pos_before = 0;
	    regnode * const oscan = scan;
	    struct regnode_charclass_class this_class;
	    struct regnode_charclass_class *oclass = NULL;
	    I32 next_is_eval = 0;

	    switch (PL_regkind[OP(scan)]) {
	    case WHILEM:		/* End of (?:...)* . */
		scan = NEXTOPER(scan);
		goto finish;
	    case PLUS:
		if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
		    next = NEXTOPER(scan);
		    if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
			mincount = 1;
			maxcount = REG_INFTY;
			next = regnext(scan);
			scan = NEXTOPER(scan);
			goto do_curly;