backup.lo bitvec.lo btmutex.lo btree.lo build.lo \
         callback.lo complete.lo ctime.lo date.lo dbstat.lo delete.lo \
         expr.lo fault.lo fkey.lo \
         fts3.lo fts3_aux.lo fts3_expr.lo fts3_hash.lo fts3_icu.lo \
         fts3_porter.lo fts3_snippet.lo fts3_tokenizer.lo fts3_tokenizer1.lo \
         fts3_tokenize_vtab.lo \
         fts3_unicode.lo fts3_unicode2.lo fts3_write.lo \

         func.lo global.lo hash.lo \
         icu.lo insert.lo journal.lo legacy.lo loadext.lo \
         main.lo malloc.lo mem0.lo mem1.lo mem2.lo mem3.lo mem5.lo \
         memjournal.lo \
         mutex.lo mutex_noop.lo mutex_unix.lo mutex_w32.lo \
         notify.lo opcodes.lo os.lo os_unix.lo os_win.lo \
         pager.lo parse.lo pcache.lo pcache1.lo pragma.lo prepare.lo printf.lo \
................................................................................
#
TESTSRC += \
  $(TOP)/ext/misc/amatch.c \
  $(TOP)/ext/misc/closure.c \
  $(TOP)/ext/misc/eval.c \
  $(TOP)/ext/misc/fileio.c \
  $(TOP)/ext/misc/fuzzer.c \

  $(TOP)/ext/misc/ieee754.c \
  $(TOP)/ext/misc/nextchar.c \
  $(TOP)/ext/misc/percentile.c \
  $(TOP)/ext/misc/regexp.c \
  $(TOP)/ext/misc/spellfix.c \
  $(TOP)/ext/misc/totype.c \
  $(TOP)/ext/misc/wholenumber.c
................................................................................

fts3_write.lo:	$(TOP)/ext/fts3/fts3_write.c $(HDR) $(EXTHDR)
	$(LTCOMPILE) -DSQLITE_CORE -c $(TOP)/ext/fts3/fts3_write.c

rtree.lo:	$(TOP)/ext/rtree/rtree.c $(HDR) $(EXTHDR)
	$(LTCOMPILE) -DSQLITE_CORE -c $(TOP)/ext/rtree/rtree.c







































# Rules to build the 'testfixture' application.
#
# If using the amalgamation, use sqlite3.c directly to build the test
# fixture.  Otherwise link against libsqlite3.la.  (This distinction is
# necessary because the test fixture requires non-API symbols which are
# hidden when the library is built via the amalgamation).

         backup.lo bitvec.lo btmutex.lo btree.lo build.lo \
         callback.lo complete.lo ctime.lo date.lo dbstat.lo delete.lo \
         expr.lo fault.lo fkey.lo \
         fts3.lo fts3_aux.lo fts3_expr.lo fts3_hash.lo fts3_icu.lo \
         fts3_porter.lo fts3_snippet.lo fts3_tokenizer.lo fts3_tokenizer1.lo \
         fts3_tokenize_vtab.lo \
         fts3_unicode.lo fts3_unicode2.lo fts3_write.lo \
	 fts5.lo \
         func.lo global.lo hash.lo \
         icu.lo insert.lo journal.lo legacy.lo loadext.lo \
         main.lo malloc.lo mem0.lo mem1.lo mem2.lo mem3.lo mem5.lo \
         memjournal.lo \
         mutex.lo mutex_noop.lo mutex_unix.lo mutex_w32.lo \
         notify.lo opcodes.lo os.lo os_unix.lo os_win.lo \
         pager.lo parse.lo pcache.lo pcache1.lo pragma.lo prepare.lo printf.lo \
................................................................................
#
TESTSRC += \
  $(TOP)/ext/misc/amatch.c \
  $(TOP)/ext/misc/closure.c \
  $(TOP)/ext/misc/eval.c \
  $(TOP)/ext/misc/fileio.c \
  $(TOP)/ext/misc/fuzzer.c \
  $(TOP)/ext/fts5/fts5_tcl.c \
  $(TOP)/ext/misc/ieee754.c \
  $(TOP)/ext/misc/nextchar.c \
  $(TOP)/ext/misc/percentile.c \
  $(TOP)/ext/misc/regexp.c \
  $(TOP)/ext/misc/spellfix.c \
  $(TOP)/ext/misc/totype.c \
  $(TOP)/ext/misc/wholenumber.c
................................................................................

fts3_write.lo:	$(TOP)/ext/fts3/fts3_write.c $(HDR) $(EXTHDR)
	$(LTCOMPILE) -DSQLITE_CORE -c $(TOP)/ext/fts3/fts3_write.c

rtree.lo:	$(TOP)/ext/rtree/rtree.c $(HDR) $(EXTHDR)
	$(LTCOMPILE) -DSQLITE_CORE -c $(TOP)/ext/rtree/rtree.c

# FTS5 things
#
FTS5_SRC = \
   $(TOP)/ext/fts5/fts5.h \
   $(TOP)/ext/fts5/fts5Int.h \
   $(TOP)/ext/fts5/fts5_aux.c \
   $(TOP)/ext/fts5/fts5_buffer.c \
   $(TOP)/ext/fts5/fts5_main.c \
   $(TOP)/ext/fts5/fts5_config.c \
   $(TOP)/ext/fts5/fts5_expr.c \
   $(TOP)/ext/fts5/fts5_hash.c \
   $(TOP)/ext/fts5/fts5_index.c \
   fts5parse.c fts5parse.h \
   $(TOP)/ext/fts5/fts5_storage.c \
   $(TOP)/ext/fts5/fts5_tokenize.c \
   $(TOP)/ext/fts5/fts5_unicode2.c \
   $(TOP)/ext/fts5/fts5_varint.c \
   $(TOP)/ext/fts5/fts5_vocab.c  \

fts5parse.c:	$(TOP)/ext/fts5/fts5parse.y lemon 
	cp $(TOP)/ext/fts5/fts5parse.y .
	rm -f fts5parse.h
	./lemon $(OPTS) fts5parse.y
	mv fts5parse.c fts5parse.c.orig
	echo "#ifdef SQLITE_ENABLE_FTS5" > fts5parse.c
	cat fts5parse.c.orig | sed 's/yy/fts5yy/g' | sed 's/YY/fts5YY/g' \
		| sed 's/TOKEN/FTS5TOKEN/g' >> fts5parse.c
	echo "#endif /* SQLITE_ENABLE_FTS5 */" >> fts5parse.c

fts5parse.h: fts5parse.c

fts5.c: $(FTS5_SRC)
	$(TCLSH_CMD) $(TOP)/ext/fts5/tool/mkfts5c.tcl

fts5.lo:	fts5.c $(HDR) $(EXTHDR)
	$(LTCOMPILE) -DSQLITE_CORE -c fts5.c


# Rules to build the 'testfixture' application.
#
# If using the amalgamation, use sqlite3.c directly to build the test
# fixture.  Otherwise link against libsqlite3.la.  (This distinction is
# necessary because the test fixture requires non-API symbols which are
# hidden when the library is built via the amalgamation).

LIBOBJS0 = vdbe.lo parse.lo alter.lo analyze.lo attach.lo auth.lo \
         backup.lo bitvec.lo btmutex.lo btree.lo build.lo \
         callback.lo complete.lo ctime.lo date.lo dbstat.lo delete.lo \
         expr.lo fault.lo fkey.lo \
         fts3.lo fts3_aux.lo fts3_expr.lo fts3_hash.lo fts3_icu.lo \
         fts3_porter.lo fts3_snippet.lo fts3_tokenizer.lo fts3_tokenizer1.lo \
         fts3_tokenize_vtab.lo fts3_unicode.lo fts3_unicode2.lo fts3_write.lo \

         func.lo global.lo hash.lo \
         icu.lo insert.lo journal.lo legacy.lo loadext.lo \
         main.lo malloc.lo mem0.lo mem1.lo mem2.lo mem3.lo mem5.lo \
         memjournal.lo \
         mutex.lo mutex_noop.lo mutex_unix.lo mutex_w32.lo \
         notify.lo opcodes.lo os.lo os_unix.lo os_win.lo \
         pager.lo pcache.lo pcache1.lo pragma.lo prepare.lo printf.lo \
................................................................................
#
TESTEXT = \
  $(TOP)\ext\misc\amatch.c \
  $(TOP)\ext\misc\closure.c \
  $(TOP)\ext\misc\eval.c \
  $(TOP)\ext\misc\fileio.c \
  $(TOP)\ext\misc\fuzzer.c \


  $(TOP)\ext\misc\ieee754.c \
  $(TOP)\ext\misc\nextchar.c \
  $(TOP)\ext\misc\percentile.c \
  $(TOP)\ext\misc\regexp.c \
  $(TOP)\ext\misc\spellfix.c \
  $(TOP)\ext\misc\totype.c \
  $(TOP)\ext\misc\wholenumber.c
................................................................................

fts3_write.lo:	$(TOP)\ext\fts3\fts3_write.c $(HDR) $(EXTHDR)
	$(LTCOMPILE) $(CORE_COMPILE_OPTS) $(NO_WARN) -DSQLITE_CORE -c $(TOP)\ext\fts3\fts3_write.c

rtree.lo:	$(TOP)\ext\rtree\rtree.c $(HDR) $(EXTHDR)
	$(LTCOMPILE) $(CORE_COMPILE_OPTS) $(NO_WARN) -DSQLITE_CORE -c $(TOP)\ext\rtree\rtree.c











































# Rules to build the 'testfixture' application.
#
# If using the amalgamation, use sqlite3.c directly to build the test
# fixture.  Otherwise link against libsqlite3.lib.  (This distinction is
# necessary because the test fixture requires non-API symbols which are
# hidden when the library is built via the amalgamation).
................................................................................
	del /Q sqlite3.exe sqlite3.dll sqlite3.def 2>NUL
	del /Q sqlite3.c sqlite3-*.c 2>NUL
	del /Q sqlite3rc.h 2>NUL
	del /Q shell.c sqlite3ext.h 2>NUL
	del /Q sqlite3_analyzer.exe sqlite3_analyzer.c 2>NUL
	del /Q sqlite-*-output.vsix 2>NUL
	del /Q fuzzershell.exe fuzzcheck.exe sqldiff.exe 2>NUL


# Dynamic link library section.
#
dll: sqlite3.dll

sqlite3.def: libsqlite3.lib
	echo EXPORTS > sqlite3.def
	dumpbin /all libsqlite3.lib \
		| $(NAWK) "/ 1 _?sqlite3_/ { sub(/^.* _?/,\"\");print }" \
		| sort >> sqlite3.def

sqlite3.dll: $(LIBOBJ) $(LIBRESOBJS) $(CORE_LINK_DEP)
	$(LD) $(LDFLAGS) $(LTLINKOPTS) $(LTLIBPATHS) /DLL $(CORE_LINK_OPTS) /OUT:$@ $(LIBOBJ) $(LIBRESOBJS) $(LTLIBS) $(TLIBS)

LIBOBJS0 = vdbe.lo parse.lo alter.lo analyze.lo attach.lo auth.lo \
         backup.lo bitvec.lo btmutex.lo btree.lo build.lo \
         callback.lo complete.lo ctime.lo date.lo dbstat.lo delete.lo \
         expr.lo fault.lo fkey.lo \
         fts3.lo fts3_aux.lo fts3_expr.lo fts3_hash.lo fts3_icu.lo \
         fts3_porter.lo fts3_snippet.lo fts3_tokenizer.lo fts3_tokenizer1.lo \
         fts3_tokenize_vtab.lo fts3_unicode.lo fts3_unicode2.lo fts3_write.lo \
         fts5.lo \
         func.lo global.lo hash.lo \
         icu.lo insert.lo journal.lo legacy.lo loadext.lo \
         main.lo malloc.lo mem0.lo mem1.lo mem2.lo mem3.lo mem5.lo \
         memjournal.lo \
         mutex.lo mutex_noop.lo mutex_unix.lo mutex_w32.lo \
         notify.lo opcodes.lo os.lo os_unix.lo os_win.lo \
         pager.lo pcache.lo pcache1.lo pragma.lo prepare.lo printf.lo \
................................................................................
#
TESTEXT = \
  $(TOP)\ext\misc\amatch.c \
  $(TOP)\ext\misc\closure.c \
  $(TOP)\ext\misc\eval.c \
  $(TOP)\ext\misc\fileio.c \
  $(TOP)\ext\misc\fuzzer.c \
  fts5.c \
  $(TOP)\ext\fts5\fts5_tcl.c \
  $(TOP)\ext\misc\ieee754.c \
  $(TOP)\ext\misc\nextchar.c \
  $(TOP)\ext\misc\percentile.c \
  $(TOP)\ext\misc\regexp.c \
  $(TOP)\ext\misc\spellfix.c \
  $(TOP)\ext\misc\totype.c \
  $(TOP)\ext\misc\wholenumber.c
................................................................................

fts3_write.lo:	$(TOP)\ext\fts3\fts3_write.c $(HDR) $(EXTHDR)
	$(LTCOMPILE) $(CORE_COMPILE_OPTS) $(NO_WARN) -DSQLITE_CORE -c $(TOP)\ext\fts3\fts3_write.c

rtree.lo:	$(TOP)\ext\rtree\rtree.c $(HDR) $(EXTHDR)
	$(LTCOMPILE) $(CORE_COMPILE_OPTS) $(NO_WARN) -DSQLITE_CORE -c $(TOP)\ext\rtree\rtree.c

# FTS5 things
#
FTS5_SRC = \
   $(TOP)\ext\fts5\fts5.h \
   $(TOP)\ext\fts5\fts5Int.h \
   $(TOP)\ext\fts5\fts5_aux.c \
   $(TOP)\ext\fts5\fts5_buffer.c \
   $(TOP)\ext\fts5\fts5_main.c \
   $(TOP)\ext\fts5\fts5_config.c \
   $(TOP)\ext\fts5\fts5_expr.c \
   $(TOP)\ext\fts5\fts5_hash.c \
   $(TOP)\ext\fts5\fts5_index.c \
   fts5parse.c fts5parse.h \
   $(TOP)\ext\fts5\fts5_storage.c \
   $(TOP)\ext\fts5\fts5_tokenize.c \
   $(TOP)\ext\fts5\fts5_unicode2.c \
   $(TOP)\ext\fts5\fts5_varint.c \
   $(TOP)\ext\fts5\fts5_vocab.c

fts5parse.c:	$(TOP)\ext\fts5\fts5parse.y lemon.exe
	copy $(TOP)\ext\fts5\fts5parse.y .
	del /Q fts5parse.h 2>NUL
	.\lemon.exe $(REQ_FEATURE_FLAGS) $(OPT_FEATURE_FLAGS) $(OPTS) fts5parse.y
	move fts5parse.c fts5parse.c.orig
	echo #ifdef SQLITE_ENABLE_FTS5 > $@
	type fts5parse.c.orig \
		| $(NAWK) "/.*/ { gsub(/yy/,\"fts5yy\");print }" \
		| $(NAWK) "/.*/ { gsub(/YY/,\"fts5YY\");print }" \
		| $(NAWK) "/.*/ { gsub(/TOKEN/,\"FTS5TOKEN\");print }" >> $@
	echo #endif /* SQLITE_ENABLE_FTS5 */ >> $@

fts5parse.h: fts5parse.c

fts5.c: $(FTS5_SRC)
	$(TCLSH_CMD) $(TOP)\ext\fts5\tool\mkfts5c.tcl

fts5.lo:	fts5.c $(HDR) $(EXTHDR)
	$(LTCOMPILE) $(NO_WARN) -DSQLITE_ENABLE_FTS5 -c fts5.c

fts5.dll:	fts5.lo
	$(LD) $(LDFLAGS) $(LTLINKOPTS) $(LTLIBPATHS) /DLL /OUT:$@ fts5.lo

# Rules to build the 'testfixture' application.
#
# If using the amalgamation, use sqlite3.c directly to build the test
# fixture.  Otherwise link against libsqlite3.lib.  (This distinction is
# necessary because the test fixture requires non-API symbols which are
# hidden when the library is built via the amalgamation).
................................................................................
	del /Q sqlite3.exe sqlite3.dll sqlite3.def 2>NUL
	del /Q sqlite3.c sqlite3-*.c 2>NUL
	del /Q sqlite3rc.h 2>NUL
	del /Q shell.c sqlite3ext.h 2>NUL
	del /Q sqlite3_analyzer.exe sqlite3_analyzer.c 2>NUL
	del /Q sqlite-*-output.vsix 2>NUL
	del /Q fuzzershell.exe fuzzcheck.exe sqldiff.exe 2>NUL
	del /Q fts5.c fts5parse.* 2>NUL

# Dynamic link library section.
#
dll: sqlite3.dll

sqlite3.def: libsqlite3.lib
	echo EXPORTS > sqlite3.def
	dumpbin /all libsqlite3.lib \
		| $(NAWK) "/ 1 _?sqlite3_/ { sub(/^.* _?/,\"\");print }" \
		| sort >> sqlite3.def

sqlite3.dll: $(LIBOBJ) $(LIBRESOBJS) $(CORE_LINK_DEP)
	$(LD) $(LDFLAGS) $(LTLINKOPTS) $(LTLIBPATHS) /DLL $(CORE_LINK_OPTS) /OUT:$@ $(LIBOBJ) $(LIBRESOBJS) $(LTLIBS) $(TLIBS)

#
# Parameter $zName must be a path to the file UnicodeData.txt. This command
# reads the file and returns a list of mappings required to remove all
# diacritical marks from a unicode string. Each mapping is itself a list
# consisting of two elements - the unicode codepoint and the single ASCII
# character that it should be replaced with, or an empty string if the 
# codepoint should simply be removed from the input. Examples:
#
#   { 224 a  }     (replace codepoint 224 to "a")
#   { 769 "" }     (remove codepoint 769 from input)
#
# Mappings are only returned for non-upper case codepoints. It is assumed
# that the input has already been folded to lower case.
#
proc rd_load_unicodedata_text {zName} {
  global tl_lookup_table

  set fd [open $zName]
  set lField {
    code
    character_name
    general_category
    canonical_combining_classes
    bidirectional_category
    character_decomposition_mapping
    decimal_digit_value
    digit_value
    numeric_value
    mirrored
    unicode_1_name
    iso10646_comment_field
    uppercase_mapping
    lowercase_mapping
    titlecase_mapping
  }
  set lRet [list]

  while { ![eof $fd] } {
    set line [gets $fd]
    if {$line == ""} continue

    set fields [split $line ";"]
    if {[llength $fields] != [llength $lField]} { error "parse error: $line" }
    foreach $lField $fields {}
    if { [llength $character_decomposition_mapping]!=2
      || [string is xdigit [lindex $character_decomposition_mapping 0]]==0
    } {
      continue
    }

    set iCode  [expr "0x$code"]
    set iAscii [expr "0x[lindex $character_decomposition_mapping 0]"]
    set iDia   [expr "0x[lindex $character_decomposition_mapping 1]"]

    if {[info exists tl_lookup_table($iCode)]} continue

    if { ($iAscii >= 97 && $iAscii <= 122)
      || ($iAscii >= 65 && $iAscii <= 90)
    } {
      lappend lRet [list $iCode [string tolower [format %c $iAscii]]]
      set dia($iDia) 1
    }
  }

  foreach d [array names dia] {
    lappend lRet [list $d ""]
  }
  set lRet [lsort -integer -index 0 $lRet]

  close $fd
  set lRet
}



proc print_rd {map} {
  global tl_lookup_table
  set aChar [list]
  set lRange [list]

  set nRange 1
................................................................................
  puts "** If the argument is a codepoint corresponding to a lowercase letter"
  puts "** in the ASCII range with a diacritic added, return the codepoint"
  puts "** of the ASCII letter only. For example, if passed 235 - \"LATIN"
  puts "** SMALL LETTER E WITH DIAERESIS\" - return 65 (\"LATIN SMALL LETTER"
  puts "** E\"). The resuls of passing a codepoint that corresponds to an"
  puts "** uppercase letter are undefined."
  puts "*/"
  puts "static int remove_diacritic(int c)\{"
  puts "  unsigned short aDia\[\] = \{"
  puts -nonewline "        0, "
  set i 1
  foreach r $lRange {
    foreach {iCode nRange} $r {}
    if {($i % 8)==0} {puts "" ; puts -nonewline "    " }
    incr i
................................................................................
  puts "      (mask1 & (1 << (c-$iFirst-32)));"
  puts "\}"
}


#-------------------------------------------------------------------------

# Parameter $zName must be a path to the file UnicodeData.txt. This command
# reads the file and returns a list of codepoints (integers). The list
# contains all codepoints in the UnicodeData.txt assigned to any "General
# Category" that is not a "Letter" or "Number".
#
proc an_load_unicodedata_text {zName} {
  set fd [open $zName]
  set lField {
    code
    character_name
    general_category
    canonical_combining_classes
    bidirectional_category
    character_decomposition_mapping
    decimal_digit_value
    digit_value
    numeric_value
    mirrored
    unicode_1_name
    iso10646_comment_field
    uppercase_mapping
    lowercase_mapping
    titlecase_mapping
  }
  set lRet [list]

  while { ![eof $fd] } {
    set line [gets $fd]
    if {$line == ""} continue

    set fields [split $line ";"]
    if {[llength $fields] != [llength $lField]} { error "parse error: $line" }
    foreach $lField $fields {}

    set iCode [expr "0x$code"]
    set bAlnum [expr {
         [lsearch {L N} [string range $general_category 0 0]] >= 0
      || $general_category=="Co"
    }]

    if { !$bAlnum } { lappend lRet $iCode }
  }

  close $fd
  set lRet
}

proc an_load_separator_ranges {} {
  global unicodedata.txt
  set lSep [an_load_unicodedata_text ${unicodedata.txt}]
  unset -nocomplain iFirst 
  unset -nocomplain nRange 
  set lRange [list]
  foreach sep $lSep {
................................................................................
  }]
  puts "  return 0;"
  puts "\}"
}

#-------------------------------------------------------------------------

proc tl_load_casefolding_txt {zName} {
  global tl_lookup_table

  set fd [open $zName]
  while { ![eof $fd] } {
    set line [gets $fd]
    if {[string range $line 0 0] == "#"} continue
    if {$line == ""} continue

    foreach x {a b c d} {unset -nocomplain $x}
    foreach {a b c d} [split $line ";"] {}

    set a2 [list]
    set c2 [list]
    foreach elem $a { lappend a2 [expr "0x[string trim $elem]"] }
    foreach elem $c { lappend c2 [expr "0x[string trim $elem]"] }
    set b [string trim $b]
    set d [string trim $d]

    if {$b=="C" || $b=="S"} { set tl_lookup_table($a2) $c2 }
  }
}

proc tl_create_records {} {
  global tl_lookup_table

  set iFirst ""
  set nOff 0
  set nRange 0
  set nIncr 0
................................................................................
    if {[tl_print_table_entry toggle $entry $liOff]} { 
      lappend lHigh $entry 
    } 
  }
  tl_print_table_footer toggle
  tl_print_ioff_table $liOff

  puts {
  int ret = c;

  assert( c>=0 );
  assert( sizeof(unsigned short)==2 && sizeof(unsigned char)==1 );

  if( c<128 ){
    if( c>='A' && c<='Z' ) ret = c + ('a' - 'A');
  }else if( c<65536 ){

    int iHi = sizeof(aEntry)/sizeof(aEntry[0]) - 1;
    int iLo = 0;
    int iRes = -1;


    while( iHi>=iLo ){
      int iTest = (iHi + iLo) / 2;
      int cmp = (c - aEntry[iTest].iCode);
      if( cmp>=0 ){
        iRes = iTest;
        iLo = iTest+1;
      }else{
        iHi = iTest-1;
      }
    }
    assert( iRes<0 || c>=aEntry[iRes].iCode );

    if( iRes>=0 ){

      const struct TableEntry *p = &aEntry[iRes];
      if( c<(p->iCode + p->nRange) && 0==(0x01 & p->flags & (p->iCode ^ c)) ){
        ret = (c + (aiOff[p->flags>>1])) & 0x0000FFFF;
        assert( ret>0 );
      }
    }

    if( bRemoveDiacritic ) ret = remove_diacritic(ret);
  }
  }

  foreach entry $lHigh {
    tl_print_if_entry $entry
  }

  puts ""
  puts "  return ret;"
................................................................................
*/

/*
** DO NOT EDIT THIS MACHINE GENERATED FILE.
*/
  }]
  puts ""



  puts "#ifndef SQLITE_DISABLE_FTS3_UNICODE"
  puts "#if defined(SQLITE_ENABLE_FTS3) || defined(SQLITE_ENABLE_FTS4)"

  puts ""
  puts "#include <assert.h>"
  puts ""
}

proc print_test_main {} {
  puts ""
................................................................................
  puts "\}"
}

# Proces the command line arguments. Exit early if they are not to
# our liking.
#
proc usage {} {
  puts -nonewline stderr "Usage: $::argv0 ?-test? "
  puts            stderr "<CaseFolding.txt file> <UnicodeData.txt file>"
  exit 1
}
if {[llength $argv]!=2 && [llength $argv]!=3} usage
if {[llength $argv]==3 && [lindex $argv 0]!="-test"} usage
set unicodedata.txt [lindex $argv end]
set casefolding.txt [lindex $argv end-1]


set generate_test_code [expr {[llength $argv]==3}]


















print_fileheader

# Print the isalnum() function to stdout.
#
set lRange [an_load_separator_ranges]
print_isalnum sqlite3FtsUnicodeIsalnum $lRange

# Leave a gap between the two generated C functions.
#
puts ""
puts ""

# Load the fold data. This is used by the [rd_XXX] commands
................................................................................
# as well as [print_fold].
tl_load_casefolding_txt ${casefolding.txt}

set mappings [rd_load_unicodedata_text ${unicodedata.txt}]
print_rd $mappings
puts ""
puts ""
print_isdiacritic sqlite3FtsUnicodeIsdiacritic $mappings
puts ""
puts ""

# Print the fold() function to stdout.
#
print_fold sqlite3FtsUnicodeFold

# Print the test routines and main() function to stdout, if -test 
# was specified.
#
if {$::generate_test_code} {
  print_test_isalnum sqlite3FtsUnicodeIsalnum $lRange
  print_fold_test sqlite3FtsUnicodeFold $mappings
  print_test_main 
}




puts "#endif /* defined(SQLITE_ENABLE_FTS3) || defined(SQLITE_ENABLE_FTS4) */"
puts "#endif /* !defined(SQLITE_DISABLE_FTS3_UNICODE) */"

source [file join [file dirname [info script]] parseunicode.tcl]

proc print_rd {map} {
  global tl_lookup_table
  set aChar [list]
  set lRange [list]

  set nRange 1
................................................................................
  puts "** If the argument is a codepoint corresponding to a lowercase letter"
  puts "** in the ASCII range with a diacritic added, return the codepoint"
  puts "** of the ASCII letter only. For example, if passed 235 - \"LATIN"
  puts "** SMALL LETTER E WITH DIAERESIS\" - return 65 (\"LATIN SMALL LETTER"
  puts "** E\"). The resuls of passing a codepoint that corresponds to an"
  puts "** uppercase letter are undefined."
  puts "*/"
  puts "static int ${::remove_diacritic}(int c)\{"
  puts "  unsigned short aDia\[\] = \{"
  puts -nonewline "        0, "
  set i 1
  foreach r $lRange {
    foreach {iCode nRange} $r {}
    if {($i % 8)==0} {puts "" ; puts -nonewline "    " }
    incr i
................................................................................
  puts "      (mask1 & (1 << (c-$iFirst-32)));"
  puts "\}"
}


#-------------------------------------------------------------------------
















































proc an_load_separator_ranges {} {
  global unicodedata.txt
  set lSep [an_load_unicodedata_text ${unicodedata.txt}]
  unset -nocomplain iFirst 
  unset -nocomplain nRange 
  set lRange [list]
  foreach sep $lSep {
................................................................................
  }]
  puts "  return 0;"
  puts "\}"
}

#-------------------------------------------------------------------------
























proc tl_create_records {} {
  global tl_lookup_table

  set iFirst ""
  set nOff 0
  set nRange 0
  set nIncr 0
................................................................................
    if {[tl_print_table_entry toggle $entry $liOff]} { 
      lappend lHigh $entry 
    } 
  }
  tl_print_table_footer toggle
  tl_print_ioff_table $liOff

  puts [subst -nocommands {
  int ret = c;


  assert( sizeof(unsigned short)==2 && sizeof(unsigned char)==1 );

  if( c<128 ){
    if( c>='A' && c<='Z' ) ret = c + ('a' - 'A');
  }else if( c<65536 ){
    const struct TableEntry *p;
    int iHi = sizeof(aEntry)/sizeof(aEntry[0]) - 1;
    int iLo = 0;
    int iRes = -1;

    assert( c>aEntry[0].iCode );
    while( iHi>=iLo ){
      int iTest = (iHi + iLo) / 2;
      int cmp = (c - aEntry[iTest].iCode);
      if( cmp>=0 ){
        iRes = iTest;
        iLo = iTest+1;
      }else{
        iHi = iTest-1;
      }
    }



    assert( iRes>=0 && c>=aEntry[iRes].iCode );
    p = &aEntry[iRes];
    if( c<(p->iCode + p->nRange) && 0==(0x01 & p->flags & (p->iCode ^ c)) ){
      ret = (c + (aiOff[p->flags>>1])) & 0x0000FFFF;
      assert( ret>0 );
    }


    if( bRemoveDiacritic ) ret = ${::remove_diacritic}(ret);
  }
  }]

  foreach entry $lHigh {
    tl_print_if_entry $entry
  }

  puts ""
  puts "  return ret;"
................................................................................
*/

/*
** DO NOT EDIT THIS MACHINE GENERATED FILE.
*/
  }]
  puts ""
  if {$::generate_fts5_code} {
    puts "#if defined(SQLITE_ENABLE_FTS5)"
  } else {
    puts "#ifndef SQLITE_DISABLE_FTS3_UNICODE"
    puts "#if defined(SQLITE_ENABLE_FTS3) || defined(SQLITE_ENABLE_FTS4)"
  }
  puts ""
  puts "#include <assert.h>"
  puts ""
}

proc print_test_main {} {
  puts ""
................................................................................
  puts "\}"
}

# Proces the command line arguments. Exit early if they are not to
# our liking.
#
proc usage {} {
  puts -nonewline stderr "Usage: $::argv0 ?-test? ?-fts5? "
  puts            stderr "<CaseFolding.txt file> <UnicodeData.txt file>"
  exit 1
}
if {[llength $argv]<2} usage

set unicodedata.txt [lindex $argv end]
set casefolding.txt [lindex $argv end-1]

set remove_diacritic remove_diacritic
set generate_test_code 0
set generate_fts5_code 0
set function_prefix "sqlite3Fts"
for {set i 0} {$i < [llength $argv]-2} {incr i} {
  switch -- [lindex $argv $i] {
    -test {
      set generate_test_code 1
    }
    -fts5 {
      set function_prefix sqlite3Fts5
      set generate_fts5_code 1
      set remove_diacritic fts5_remove_diacritic
    }
    default {
      usage
    }
  }
}

print_fileheader

# Print the isalnum() function to stdout.
#
set lRange [an_load_separator_ranges]
print_isalnum ${function_prefix}UnicodeIsalnum $lRange

# Leave a gap between the two generated C functions.
#
puts ""
puts ""

# Load the fold data. This is used by the [rd_XXX] commands
................................................................................
# as well as [print_fold].
tl_load_casefolding_txt ${casefolding.txt}

set mappings [rd_load_unicodedata_text ${unicodedata.txt}]
print_rd $mappings
puts ""
puts ""
print_isdiacritic ${function_prefix}UnicodeIsdiacritic $mappings
puts ""
puts ""

# Print the fold() function to stdout.
#
print_fold ${function_prefix}UnicodeFold

# Print the test routines and main() function to stdout, if -test 
# was specified.
#
if {$::generate_test_code} {
  print_test_isalnum ${function_prefix}UnicodeIsalnum $lRange
  print_fold_test ${function_prefix}UnicodeFold $mappings
  print_test_main 
}

if {$generate_fts5_code} {
  puts "#endif /* defined(SQLITE_ENABLE_FTS5) */"
} else {
  puts "#endif /* defined(SQLITE_ENABLE_FTS3) || defined(SQLITE_ENABLE_FTS4) */"
  puts "#endif /* !defined(SQLITE_DISABLE_FTS3_UNICODE) */"
}

#--------------------------------------------------------------------------
# Parameter $zName must be a path to the file UnicodeData.txt. This command
# reads the file and returns a list of mappings required to remove all
# diacritical marks from a unicode string. Each mapping is itself a list
# consisting of two elements - the unicode codepoint and the single ASCII
# character that it should be replaced with, or an empty string if the 
# codepoint should simply be removed from the input. Examples:
#
#   { 224 a  }     (replace codepoint 224 to "a")
#   { 769 "" }     (remove codepoint 769 from input)
#
# Mappings are only returned for non-upper case codepoints. It is assumed
# that the input has already been folded to lower case.
#
proc rd_load_unicodedata_text {zName} {
  global tl_lookup_table

  set fd [open $zName]
  set lField {
    code
    character_name
    general_category
    canonical_combining_classes
    bidirectional_category
    character_decomposition_mapping
    decimal_digit_value
    digit_value
    numeric_value
    mirrored
    unicode_1_name
    iso10646_comment_field
    uppercase_mapping
    lowercase_mapping
    titlecase_mapping
  }
  set lRet [list]

  while { ![eof $fd] } {
    set line [gets $fd]
    if {$line == ""} continue

    set fields [split $line ";"]
    if {[llength $fields] != [llength $lField]} { error "parse error: $line" }
    foreach $lField $fields {}
    if { [llength $character_decomposition_mapping]!=2
      || [string is xdigit [lindex $character_decomposition_mapping 0]]==0
    } {
      continue
    }

    set iCode  [expr "0x$code"]
    set iAscii [expr "0x[lindex $character_decomposition_mapping 0]"]
    set iDia   [expr "0x[lindex $character_decomposition_mapping 1]"]

    if {[info exists tl_lookup_table($iCode)]} continue

    if { ($iAscii >= 97 && $iAscii <= 122)
      || ($iAscii >= 65 && $iAscii <= 90)
    } {
      lappend lRet [list $iCode [string tolower [format %c $iAscii]]]
      set dia($iDia) 1
    }
  }

  foreach d [array names dia] {
    lappend lRet [list $d ""]
  }
  set lRet [lsort -integer -index 0 $lRet]

  close $fd
  set lRet
}

#-------------------------------------------------------------------------
# Parameter $zName must be a path to the file UnicodeData.txt. This command
# reads the file and returns a list of codepoints (integers). The list
# contains all codepoints in the UnicodeData.txt assigned to any "General
# Category" that is not a "Letter" or "Number".
#
proc an_load_unicodedata_text {zName} {
  set fd [open $zName]
  set lField {
    code
    character_name
    general_category
    canonical_combining_classes
    bidirectional_category
    character_decomposition_mapping
    decimal_digit_value
    digit_value
    numeric_value
    mirrored
    unicode_1_name
    iso10646_comment_field
    uppercase_mapping
    lowercase_mapping
    titlecase_mapping
  }
  set lRet [list]

  while { ![eof $fd] } {
    set line [gets $fd]
    if {$line == ""} continue

    set fields [split $line ";"]
    if {[llength $fields] != [llength $lField]} { error "parse error: $line" }
    foreach $lField $fields {}

    set iCode [expr "0x$code"]
    set bAlnum [expr {
         [lsearch {L N} [string range $general_category 0 0]] >= 0
      || $general_category=="Co"
    }]

    if { !$bAlnum } { lappend lRet $iCode }
  }

  close $fd
  set lRet
}

proc tl_load_casefolding_txt {zName} {
  global tl_lookup_table

  set fd [open $zName]
  while { ![eof $fd] } {
    set line [gets $fd]
    if {[string range $line 0 0] == "#"} continue
    if {$line == ""} continue

    foreach x {a b c d} {unset -nocomplain $x}
    foreach {a b c d} [split $line ";"] {}

    set a2 [list]
    set c2 [list]
    foreach elem $a { lappend a2 [expr "0x[string trim $elem]"] }
    foreach elem $c { lappend c2 [expr "0x[string trim $elem]"] }
    set b [string trim $b]
    set d [string trim $d]

    if {$b=="C" || $b=="S"} { set tl_lookup_table($a2) $c2 }
  }
}

#
# 2014 August 24
#
# The author disclaims copyright to this source code.  In place of
# a legal notice, here is a blessing:
#
#    May you do good and not evil.
#    May you find forgiveness for yourself and forgive others.
#    May you share freely, never taking more than you give.
#
#--------------------------------------------------------------------------
#
# This script extracts the documentation for the API used by fts5 auxiliary 
# functions from header file fts5.h. It outputs html text on stdout that
# is included in the documentation on the web.
# 

set ::fts5_docs_output ""
if {[info commands hd_putsnl]==""} {
  if {[llength $argv]>0} { set ::extract_api_docs_mode [lindex $argv 0] }
  proc output {text} {
    puts $text
  }
} else {
  proc output {text} {
    append ::fts5_docs_output "$text\n"
  }
}
if {[info exists ::extract_api_docs_mode]==0} {set ::extract_api_docs_mode api}


set input_file [file join [file dir [info script]] fts5.h]
set fd [open $input_file]
set data [read $fd]
close $fd


# Argument $data is the entire text of the fts5.h file. This function 
# extracts the definition of the Fts5ExtensionApi structure from it and
# returns a key/value list of structure member names and definitions. i.e.
#
#   iVersion {int iVersion} xUserData {void *(*xUserData)(Fts5Context*)} ...
#
proc get_struct_members {data} {

  # Extract the structure definition from the fts5.h file.
  regexp "struct Fts5ExtensionApi {(.*?)};" $data -> defn

  # Remove all comments from the structure definition
  regsub -all {/[*].*?[*]/} $defn {} defn2

  set res [list]
  foreach member [split $defn2 {;}] {

    set member [string trim $member]
    if {$member!=""} { 
      catch { set name [lindex $member end] }
      regexp {.*?[(][*]([^)]*)[)]} $member -> name
      lappend res $name $member
    }
  }

  set res
}

proc get_struct_docs {data names} {
  # Extract the structure definition from the fts5.h file.
  regexp {EXTENSION API FUNCTIONS(.*?)[*]/} $data -> docs

  set current_doc    ""
  set current_header ""

  foreach line [split $docs "\n"] {
    regsub {[*]*} $line {} line
    if {[regexp {^  } $line]} {
      append current_doc "$line\n"
    } elseif {[string trim $line]==""} {
      if {$current_header!=""} { append current_doc "\n" }
    } else {
      if {$current_doc != ""} {
        lappend res $current_header $current_doc
        set current_doc ""
      }
      set subject n/a
      regexp {^ *([[:alpha:]]*)} $line -> subject
      if {[lsearch $names $subject]>=0} {
        set current_header $subject
      } else {
        set current_header [string trim $line]
      }
    }
  }

  if {$current_doc != ""} {
    lappend res $current_header $current_doc
  }

  set res
}

proc get_tokenizer_docs {data} {
  regexp {(xCreate:.*?)[*]/} $data -> docs

  set res "<dl>\n"
  foreach line [split [string trim $docs] "\n"] {
    regexp {[*][*](.*)} $line -> line
    if {[regexp {^ ?x.*:} $line]} {
      append res "<dt><b>$line</b></dt><dd><p style=margin-top:0>\n"
      continue
    }
    if {[string trim $line] == ""} {
      append res "<p>\n"
    } else {
      append res "$line\n"
    }
  }
  append res "</dl>\n"

  set res
}

proc get_api_docs {data} {
  # Initialize global array M as a map from Fts5StructureApi member name
  # to member definition. i.e.
  #
  #   iVersion  -> {int iVersion}
  #   xUserData -> {void *(*xUserData)(Fts5Context*)}
  #   ...
  #
  array set M [get_struct_members $data]
  
  # Initialize global list D as a map from section name to documentation
  # text. Most (all?) section names are structure member names.
  #
  set D [get_struct_docs $data [array names M]]
  
  foreach {sub docs} $D {
    if {[info exists M($sub)]} {
      set hdr $M($sub)
      set link " id=$sub"
    } else {
      set link ""
    }

    output "<hr color=#eeeee style=\"margin:1em 8.4ex 0 8.4ex;\"$link>"
    set style "padding-left:6ex;font-size:1.4em;display:block"
    output "<h style=\"$style\"><pre>$hdr</pre></h>"
  
    set mode ""
    set bEmpty 1
    foreach line [split [string trim $docs] "\n"] {
      if {[string trim $line]==""} {
        if {$mode != ""} {output "</$mode>"}
        set mode ""
      } elseif {$mode == ""} {
        if {[regexp {^     } $line]} {
          set mode codeblock
        } else {
          set mode p
        }
        output "<$mode>"
      }
      output $line
    }
    if {$mode != ""} {output "</$mode>"}
  }
}

proc get_fts5_struct {data start end} {
  set res ""
  set bOut 0
  foreach line [split $data "\n"] {
    if {$bOut==0} {
      if {[regexp $start $line]} {
        set bOut 1
      }
    }

    if {$bOut} {
      append res "$line\n"
    }

    if {$bOut} {
      if {[regexp $end $line]} {
        set bOut 0
      }
    }
  }

  set map [list /* <i>/* */ */</i>]
  string map $map $res
}

proc main {data} {
  switch $::extract_api_docs_mode {
    fts5_api {
      output [get_fts5_struct $data "typedef struct fts5_api" "^\};"]
    }

    fts5_tokenizer {
      output [get_fts5_struct $data "typedef struct Fts5Tokenizer" "^\};"]
    }

    fts5_extension {
      output [get_fts5_struct $data "typedef.*Fts5ExtensionApi" "^.;"]
    }

    Fts5ExtensionApi {
      set struct [get_fts5_struct $data "^struct Fts5ExtensionApi" "^.;"]
      set map [list]
      foreach {k v} [get_struct_members $data] {
        if {[string match x* $k]==0} continue
        lappend map $k "<a href=#$k>$k</a>"
      }
      output [string map $map $struct]
    }

    api {
      get_api_docs $data
    }

    tokenizer_api {
      output [get_tokenizer_docs $data]
    }

    default {
    }
  }
}
main $data

set ::fts5_docs_output

/*
** 2014 May 31
**
** The author disclaims copyright to this source code.  In place of
** a legal notice, here is a blessing:
**
**    May you do good and not evil.
**    May you find forgiveness for yourself and forgive others.
**    May you share freely, never taking more than you give.
**
******************************************************************************
**
** Interfaces to extend FTS5. Using the interfaces defined in this file, 
** FTS5 may be extended with:
**
**     * custom tokenizers, and
**     * custom auxiliary functions.
*/


#ifndef _FTS5_H
#define _FTS5_H

#include "sqlite3.h"

/*************************************************************************
** CUSTOM AUXILIARY FUNCTIONS
**
** Virtual table implementations may overload SQL functions by implementing
** the sqlite3_module.xFindFunction() method.
*/

typedef struct Fts5ExtensionApi Fts5ExtensionApi;
typedef struct Fts5Context Fts5Context;

typedef void (*fts5_extension_function)(
  const Fts5ExtensionApi *pApi,   /* API offered by current FTS version */
  Fts5Context *pFts,              /* First arg to pass to pApi functions */
  sqlite3_context *pCtx,          /* Context for returning result/error */
  int nVal,                       /* Number of values in apVal[] array */
  sqlite3_value **apVal           /* Array of trailing arguments */
);

/*
** EXTENSION API FUNCTIONS
**
** xUserData(pFts):
**   Return a copy of the context pointer the extension function was 
**   registered with.
**
** xColumnTotalSize(pFts, iCol, pnToken):
**   If parameter iCol is less than zero, set output variable *pnToken
**   to the total number of tokens in the FTS5 table. Or, if iCol is
**   non-negative but less than the number of columns in the table, return
**   the total number of tokens in column iCol, considering all rows in 
**   the FTS5 table.
**
**   If parameter iCol is greater than or equal to the number of columns
**   in the table, SQLITE_RANGE is returned. Or, if an error occurs (e.g.
**   an OOM condition or IO error), an appropriate SQLite error code is 
**   returned.
**
** xColumnCount:
**   Returns the number of columns in the FTS5 table.
**
** xColumnSize:
**   Reports the size in tokens of a column value from the current row.
**
** xColumnText:
**   This function attempts to retrieve the text of column iCol of the
**   current document. If successful, (*pz) is set to point to a buffer
**   containing the text in utf-8 encoding, (*pn) is set to the size in bytes
**   (not characters) of the buffer and SQLITE_OK is returned. Otherwise,
**   if an error occurs, an SQLite error code is returned and the final values
**   of (*pz) and (*pn) are undefined.
**
** xPhraseCount:
**   Returns the number of phrases in the current query expression.
**
** xPhraseSize:
**   Returns the number of tokens in phrase iPhrase of the query. Phrases
**   are numbered starting from zero.
**
** xInstCount:
**   Set *pnInst to the total number of occurrences of all phrases within
**   the query within the current row. Return SQLITE_OK if successful, or
**   an error code (i.e. SQLITE_NOMEM) if an error occurs.
**
** xInst:
**   Query for the details of phrase match iIdx within the current row.
**   Phrase matches are numbered starting from zero, so the iIdx argument
**   should be greater than or equal to zero and smaller than the value
**   output by xInstCount().
**
**   Returns SQLITE_OK if successful, or an error code (i.e. SQLITE_NOMEM) 
**   if an error occurs.
**
** xRowid:
**   Returns the rowid of the current row.
**
** xTokenize:
**   Tokenize text using the tokenizer belonging to the FTS5 table.
**
** xQueryPhrase(pFts5, iPhrase, pUserData, xCallback):
**   This API function is used to query the FTS table for phrase iPhrase
**   of the current query. Specifically, a query equivalent to:
**
**       ... FROM ftstable WHERE ftstable MATCH $p ORDER BY rowid
**
**   with $p set to a phrase equivalent to the phrase iPhrase of the
**   current query is executed. For each row visited, the callback function
**   passed as the fourth argument is invoked. The context and API objects 
**   passed to the callback function may be used to access the properties of
**   each matched row. Invoking Api.xUserData() returns a copy of the pointer
**   passed as the third argument to pUserData.
**
**   If the callback function returns any value other than SQLITE_OK, the
**   query is abandoned and the xQueryPhrase function returns immediately.
**   If the returned value is SQLITE_DONE, xQueryPhrase returns SQLITE_OK.
**   Otherwise, the error code is propagated upwards.
**
**   If the query runs to completion without incident, SQLITE_OK is returned.
**   Or, if some error occurs before the query completes or is aborted by
**   the callback, an SQLite error code is returned.
**
**
** xSetAuxdata(pFts5, pAux, xDelete)
**
**   Save the pointer passed as the second argument as the extension functions 
**   "auxiliary data". The pointer may then be retrieved by the current or any
**   future invocation of the same fts5 extension function made as part of
**   of the same MATCH query using the xGetAuxdata() API.
**
**   Each extension function is allocated a single auxiliary data slot for
**   each FTS query (MATCH expression). If the extension function is invoked 
**   more than once for a single FTS query, then all invocations share a 
**   single auxiliary data context.
**
**   If there is already an auxiliary data pointer when this function is
**   invoked, then it is replaced by the new pointer. If an xDelete callback
**   was specified along with the original pointer, it is invoked at this
**   point.
**
**   The xDelete callback, if one is specified, is also invoked on the
**   auxiliary data pointer after the FTS5 query has finished.
**
**   If an error (e.g. an OOM condition) occurs within this function, an
**   the auxiliary data is set to NULL and an error code returned. If the
**   xDelete parameter was not NULL, it is invoked on the auxiliary data
**   pointer before returning.
**
**
** xGetAuxdata(pFts5, bClear)
**
**   Returns the current auxiliary data pointer for the fts5 extension 
**   function. See the xSetAuxdata() method for details.
**
**   If the bClear argument is non-zero, then the auxiliary data is cleared
**   (set to NULL) before this function returns. In this case the xDelete,
**   if any, is not invoked.
**
**
** xRowCount(pFts5, pnRow)
**
**   This function is used to retrieve the total number of rows in the table.
**   In other words, the same value that would be returned by:
**
**        SELECT count(*) FROM ftstable;
*/
struct Fts5ExtensionApi {
  int iVersion;                   /* Currently always set to 1 */

  void *(*xUserData)(Fts5Context*);

  int (*xColumnCount)(Fts5Context*);
  int (*xRowCount)(Fts5Context*, sqlite3_int64 *pnRow);
  int (*xColumnTotalSize)(Fts5Context*, int iCol, sqlite3_int64 *pnToken);

  int (*xTokenize)(Fts5Context*, 
    const char *pText, int nText, /* Text to tokenize */
    void *pCtx,                   /* Context passed to xToken() */
    int (*xToken)(void*, const char*, int, int, int)       /* Callback */
  );

  int (*xPhraseCount)(Fts5Context*);
  int (*xPhraseSize)(Fts5Context*, int iPhrase);

  int (*xInstCount)(Fts5Context*, int *pnInst);
  int (*xInst)(Fts5Context*, int iIdx, int *piPhrase, int *piCol, int *piOff);

  sqlite3_int64 (*xRowid)(Fts5Context*);
  int (*xColumnText)(Fts5Context*, int iCol, const char **pz, int *pn);
  int (*xColumnSize)(Fts5Context*, int iCol, int *pnToken);

  int (*xQueryPhrase)(Fts5Context*, int iPhrase, void *pUserData,
    int(*)(const Fts5ExtensionApi*,Fts5Context*,void*)
  );
  int (*xSetAuxdata)(Fts5Context*, void *pAux, void(*xDelete)(void*));
  void *(*xGetAuxdata)(Fts5Context*, int bClear);
};

/* 
** CUSTOM AUXILIARY FUNCTIONS
*************************************************************************/

/*************************************************************************
** CUSTOM TOKENIZERS
**
** Applications may also register custom tokenizer types. A tokenizer 
** is registered by providing fts5 with a populated instance of the 
** following structure. All structure methods must be defined, setting
** any member of the fts5_tokenizer struct to NULL leads to undefined
** behaviour. The structure methods are expected to function as follows:
**
** xCreate:
**   This function is used to allocate and inititalize a tokenizer instance.
**   A tokenizer instance is required to actually tokenize text.
**
**   The first argument passed to this function is a copy of the (void*)
**   pointer provided by the application when the fts5_tokenizer object
**   was registered with FTS5 (the third argument to xCreateTokenizer()). 
**   The second and third arguments are an array of nul-terminated strings
**   containing the tokenizer arguments, if any, specified following the
**   tokenizer name as part of the CREATE VIRTUAL TABLE statement used
**   to create the FTS5 table.
**
**   The final argument is an output variable. If successful, (*ppOut) 
**   should be set to point to the new tokenizer handle and SQLITE_OK
**   returned. If an error occurs, some value other than SQLITE_OK should
**   be returned. In this case, fts5 assumes that the final value of *ppOut 
**   is undefined.
**
** xDelete:
**   This function is invoked to delete a tokenizer handle previously
**   allocated using xCreate(). Fts5 guarantees that this function will
**   be invoked exactly once for each successful call to xCreate().
**
** xTokenize:
**   This function is expected to tokenize the nText byte string indicated 
**   by argument pText. pText may not be nul-terminated. The first argument
**   passed to this function is a pointer to an Fts5Tokenizer object returned 
**   by an earlier call to xCreate().
**
**   For each token in the input string, the supplied callback xToken() must
**   be invoked. The first argument to it should be a copy of the pointer
**   passed as the second argument to xTokenize(). The next two arguments
**   are a pointer to a buffer containing the token text, and the size of
**   the token in bytes. The 4th and 5th arguments are the byte offsets of
**   the first byte of and first byte immediately following the text from 
**   which the token is derived within the input.
**
**   FTS5 assumes the xToken() callback is invoked for each token in the 
**   order that they occur within the input text.
**
**   If an xToken() callback returns any value other than SQLITE_OK, then
**   the tokenization should be abandoned and the xTokenize() method should
**   immediately return a copy of the xToken() return value. Or, if the
**   input buffer is exhausted, xTokenize() should return SQLITE_OK. Finally,
**   if an error occurs with the xTokenize() implementation itself, it
**   may abandon the tokenization and return any error code other than
**   SQLITE_OK or SQLITE_DONE.
**
*/
typedef struct Fts5Tokenizer Fts5Tokenizer;
typedef struct fts5_tokenizer fts5_tokenizer;
struct fts5_tokenizer {
  int (*xCreate)(void*, const char **azArg, int nArg, Fts5Tokenizer **ppOut);
  void (*xDelete)(Fts5Tokenizer*);
  int (*xTokenize)(Fts5Tokenizer*, 
      void *pCtx,
      const char *pText, int nText, 
      int (*xToken)(
        void *pCtx,         /* Copy of 2nd argument to xTokenize() */
        const char *pToken, /* Pointer to buffer containing token */
        int nToken,         /* Size of token in bytes */
        int iStart,         /* Byte offset of token within input text */
        int iEnd            /* Byte offset of end of token within input text */
      )
  );
};

/*
** END OF CUSTOM TOKENIZERS
*************************************************************************/

/*************************************************************************
** FTS5 EXTENSION REGISTRATION API
*/
typedef struct fts5_api fts5_api;
struct fts5_api {
  int iVersion;                   /* Currently always set to 1 */

  /* Create a new tokenizer */
  int (*xCreateTokenizer)(
    fts5_api *pApi,
    const char *zName,
    void *pContext,
    fts5_tokenizer *pTokenizer,
    void (*xDestroy)(void*)
  );

  /* Find an existing tokenizer */
  int (*xFindTokenizer)(
    fts5_api *pApi,
    const char *zName,
    void **ppContext,
    fts5_tokenizer *pTokenizer
  );

  /* Create a new auxiliary function */
  int (*xCreateFunction)(
    fts5_api *pApi,
    const char *zName,
    void *pContext,
    fts5_extension_function xFunction,
    void (*xDestroy)(void*)
  );
};

/*
** END OF REGISTRATION API
*************************************************************************/

#endif /* _FTS5_H */

/*
** 2014 May 31
**
** The author disclaims copyright to this source code.  In place of
** a legal notice, here is a blessing:
**
**    May you do good and not evil.
**    May you find forgiveness for yourself and forgive others.
**    May you share freely, never taking more than you give.
**
******************************************************************************
**
*/
#ifndef _FTS5INT_H
#define _FTS5INT_H

#ifdef SQLITE_ENABLE_FTS5

#include "fts5.h"
#include "sqlite3ext.h"
SQLITE_EXTENSION_INIT1

#include <string.h>
#include <assert.h>

#ifndef SQLITE_AMALGAMATION

typedef unsigned char  u8;
typedef unsigned int   u32;
typedef unsigned short u16;
typedef sqlite3_int64 i64;
typedef sqlite3_uint64 u64;

#define ArraySize(x) (sizeof(x) / sizeof(x[0]))

#define testcase(x)
#define ALWAYS(x) 1
#define NEVER(x) 0

#define MIN(x,y) (((x) < (y)) ? (x) : (y))

#endif


/*
** Maximum number of prefix indexes on single FTS5 table. This must be
** less than 32. If it is set to anything large than that, an #error
** directive in fts5_index.c will cause the build to fail.
*/
#define FTS5_MAX_PREFIX_INDEXES 31

#define FTS5_DEFAULT_NEARDIST 10
#define FTS5_DEFAULT_RANK     "bm25"

/* Name of rank and rowid columns */
#define FTS5_RANK_NAME "rank"
#define FTS5_ROWID_NAME "rowid"

#ifdef SQLITE_DEBUG
# define FTS5_CORRUPT sqlite3Fts5Corrupt()
int sqlite3Fts5Corrupt(void);
#else
# define FTS5_CORRUPT SQLITE_CORRUPT_VTAB
#endif

/*
** The assert_nc() macro is similar to the assert() macro, except that it
** is used for assert() conditions that are true only if it can be 
** guranteed that the database is not corrupt.
*/
#ifdef SQLITE_DEBUG
extern int sqlite3_fts5_may_be_corrupt;
# define assert_nc(x) assert(sqlite3_fts5_may_be_corrupt || (x))
#else
# define assert_nc(x) assert(x)
#endif

typedef struct Fts5Global Fts5Global;

/**************************************************************************
** Interface to code in fts5_config.c. fts5_config.c contains contains code
** to parse the arguments passed to the CREATE VIRTUAL TABLE statement.
*/

typedef struct Fts5Config Fts5Config;

/*
** An instance of the following structure encodes all information that can
** be gleaned from the CREATE VIRTUAL TABLE statement.
**
** And all information loaded from the %_config table.
**
** nAutomerge:
**   The minimum number of segments that an auto-merge operation should
**   attempt to merge together. A value of 1 sets the object to use the 
**   compile time default. Zero disables auto-merge altogether.
**
** zContent:
**
** zContentRowid:
**   The value of the content_rowid= option, if one was specified. Or 
**   the string "rowid" otherwise. This text is not quoted - if it is
**   used as part of an SQL statement it needs to be quoted appropriately.
**
** zContentExprlist:
**
** pzErrmsg:
**   This exists in order to allow the fts5_index.c module to return a 
**   decent error message if it encounters a file-format version it does
**   not understand.
**
** bColumnsize:
**   True if the %_docsize table is created.
**
*/
struct Fts5Config {
  sqlite3 *db;                    /* Database handle */
  char *zDb;                      /* Database holding FTS index (e.g. "main") */
  char *zName;                    /* Name of FTS index */
  int nCol;                       /* Number of columns */
  char **azCol;                   /* Column names */
  u8 *abUnindexed;                /* True for unindexed columns */
  int nPrefix;                    /* Number of prefix indexes */
  int *aPrefix;                   /* Sizes in bytes of nPrefix prefix indexes */
  int eContent;                   /* An FTS5_CONTENT value */
  char *zContent;                 /* content table */ 
  char *zContentRowid;            /* "content_rowid=" option value */ 
  int bColumnsize;                /* "columnsize=" option value (dflt==1) */
  char *zContentExprlist;
  Fts5Tokenizer *pTok;
  fts5_tokenizer *pTokApi;

  /* Values loaded from the %_config table */
  int iCookie;                    /* Incremented when %_config is modified */
  int pgsz;                       /* Approximate page size used in %_data */
  int nAutomerge;                 /* 'automerge' setting */
  int nCrisisMerge;               /* Maximum allowed segments per level */
  char *zRank;                    /* Name of rank function */
  char *zRankArgs;                /* Arguments to rank function */

  /* If non-NULL, points to sqlite3_vtab.base.zErrmsg. Often NULL. */
  char **pzErrmsg;
};

/* Current expected value of %_config table 'version' field */
#define FTS5_CURRENT_VERSION 2

#define FTS5_CONTENT_NORMAL   0
#define FTS5_CONTENT_NONE     1
#define FTS5_CONTENT_EXTERNAL 2




int sqlite3Fts5ConfigParse(
    Fts5Global*, sqlite3*, int, const char **, Fts5Config**, char**
);
void sqlite3Fts5ConfigFree(Fts5Config*);

int sqlite3Fts5ConfigDeclareVtab(Fts5Config *pConfig);

int sqlite3Fts5Tokenize(
  Fts5Config *pConfig,            /* FTS5 Configuration object */
  const char *pText, int nText,   /* Text to tokenize */
  void *pCtx,                     /* Context passed to xToken() */
  int (*xToken)(void*, const char*, int, int, int)    /* Callback */
);

void sqlite3Fts5Dequote(char *z);

/* Load the contents of the %_config table */
int sqlite3Fts5ConfigLoad(Fts5Config*, int);

/* Set the value of a single config attribute */
int sqlite3Fts5ConfigSetValue(Fts5Config*, const char*, sqlite3_value*, int*);

int sqlite3Fts5ConfigParseRank(const char*, char**, char**);

/*
** End of interface to code in fts5_config.c.
**************************************************************************/

/**************************************************************************
** Interface to code in fts5_buffer.c.
*/

/*
** Buffer object for the incremental building of string data.
*/
typedef struct Fts5Buffer Fts5Buffer;
struct Fts5Buffer {
  u8 *p;
  int n;
  int nSpace;
};

int sqlite3Fts5BufferGrow(int*, Fts5Buffer*, int);
void sqlite3Fts5BufferAppendVarint(int*, Fts5Buffer*, i64);
void sqlite3Fts5BufferAppendBlob(int*, Fts5Buffer*, int, const u8*);
void sqlite3Fts5BufferAppendString(int *, Fts5Buffer*, const char*);
void sqlite3Fts5BufferFree(Fts5Buffer*);
void sqlite3Fts5BufferZero(Fts5Buffer*);
void sqlite3Fts5BufferSet(int*, Fts5Buffer*, int, const u8*);
void sqlite3Fts5BufferAppendPrintf(int *, Fts5Buffer*, char *zFmt, ...);
void sqlite3Fts5BufferAppend32(int*, Fts5Buffer*, int);

char *sqlite3Fts5Mprintf(int *pRc, const char *zFmt, ...);

#define fts5BufferZero(x)             sqlite3Fts5BufferZero(x)
#define fts5BufferGrow(a,b,c)         sqlite3Fts5BufferGrow(a,b,c)
#define fts5BufferAppendVarint(a,b,c) sqlite3Fts5BufferAppendVarint(a,b,c)
#define fts5BufferFree(a)             sqlite3Fts5BufferFree(a)
#define fts5BufferAppendBlob(a,b,c,d) sqlite3Fts5BufferAppendBlob(a,b,c,d)
#define fts5BufferSet(a,b,c,d)        sqlite3Fts5BufferSet(a,b,c,d)
#define fts5BufferAppend32(a,b,c)     sqlite3Fts5BufferAppend32(a,b,c)

/* Write and decode big-endian 32-bit integer values */
void sqlite3Fts5Put32(u8*, int);
int sqlite3Fts5Get32(const u8*);

#define FTS5_POS2COLUMN(iPos) (int)(iPos >> 32)
#define FTS5_POS2OFFSET(iPos) (int)(iPos & 0xFFFFFFFF)

typedef struct Fts5PoslistReader Fts5PoslistReader;
struct Fts5PoslistReader {
  /* Variables used only by sqlite3Fts5PoslistIterXXX() functions. */
  int iCol;                       /* If (iCol>=0), this column only */
  const u8 *a;                    /* Position list to iterate through */
  int n;                          /* Size of buffer at a[] in bytes */
  int i;                          /* Current offset in a[] */

  /* Output variables */
  int bEof;                       /* Set to true at EOF */
  i64 iPos;                       /* (iCol<<32) + iPos */
};
int sqlite3Fts5PoslistReaderInit(
  int iCol,                       /* If (iCol>=0), this column only */
  const u8 *a, int n,             /* Poslist buffer to iterate through */
  Fts5PoslistReader *pIter        /* Iterator object to initialize */
);
int sqlite3Fts5PoslistReaderNext(Fts5PoslistReader*);

typedef struct Fts5PoslistWriter Fts5PoslistWriter;
struct Fts5PoslistWriter {
  i64 iPrev;
};
int sqlite3Fts5PoslistWriterAppend(Fts5Buffer*, Fts5PoslistWriter*, i64);

int sqlite3Fts5PoslistNext64(
  const u8 *a, int n,             /* Buffer containing poslist */
  int *pi,                        /* IN/OUT: Offset within a[] */
  i64 *piOff                      /* IN/OUT: Current offset */
);

/* Malloc utility */
void *sqlite3Fts5MallocZero(int *pRc, int nByte);
char *sqlite3Fts5Strndup(int *pRc, const char *pIn, int nIn);

/* Character set tests (like isspace(), isalpha() etc.) */
int sqlite3Fts5IsBareword(char t);

/*
** End of interface to code in fts5_buffer.c.
**************************************************************************/

/**************************************************************************
** Interface to code in fts5_index.c. fts5_index.c contains contains code
** to access the data stored in the %_data table.
*/

typedef struct Fts5Index Fts5Index;
typedef struct Fts5IndexIter Fts5IndexIter;

/*
** Values used as part of the flags argument passed to IndexQuery().
*/
#define FTS5INDEX_QUERY_PREFIX     0x0001   /* Prefix query */
#define FTS5INDEX_QUERY_DESC       0x0002   /* Docs in descending rowid order */
#define FTS5INDEX_QUERY_TEST_NOIDX 0x0004   /* Do not use prefix index */
#define FTS5INDEX_QUERY_SCAN       0x0008   /* Scan query (fts5vocab) */

/*
** Create/destroy an Fts5Index object.
*/
int sqlite3Fts5IndexOpen(Fts5Config *pConfig, int bCreate, Fts5Index**, char**);
int sqlite3Fts5IndexClose(Fts5Index *p);

/*
** for(
**   pIter = sqlite3Fts5IndexQuery(p, "token", 5, 0);
**   0==sqlite3Fts5IterEof(pIter);
**   sqlite3Fts5IterNext(pIter)
** ){
**   i64 iRowid = sqlite3Fts5IterRowid(pIter);
** }
*/

/*
** Open a new iterator to iterate though all docids that match the 
** specified token or token prefix.
*/
int sqlite3Fts5IndexQuery(
  Fts5Index *p,                   /* FTS index to query */
  const char *pToken, int nToken, /* Token (or prefix) to query for */
  int flags,                      /* Mask of FTS5INDEX_QUERY_X flags */
  Fts5IndexIter **ppIter
);

/*
** The various operations on open token or token prefix iterators opened
** using sqlite3Fts5IndexQuery().
*/
int sqlite3Fts5IterEof(Fts5IndexIter*);
int sqlite3Fts5IterNext(Fts5IndexIter*);
int sqlite3Fts5IterNextFrom(Fts5IndexIter*, i64 iMatch);
i64 sqlite3Fts5IterRowid(Fts5IndexIter*);
int sqlite3Fts5IterPoslist(Fts5IndexIter*, const u8 **pp, int *pn, i64 *pi);
int sqlite3Fts5IterPoslistBuffer(Fts5IndexIter *pIter, Fts5Buffer *pBuf);

/*
** Close an iterator opened by sqlite3Fts5IndexQuery().
*/
void sqlite3Fts5IterClose(Fts5IndexIter*);

/*
** This interface is used by the fts5vocab module.
*/
const char *sqlite3Fts5IterTerm(Fts5IndexIter*, int*);
int sqlite3Fts5IterNextScan(Fts5IndexIter*);


/*
** Insert or remove data to or from the index. Each time a document is 
** added to or removed from the index, this function is called one or more
** times.
**
** For an insert, it must be called once for each token in the new document.
** If the operation is a delete, it must be called (at least) once for each
** unique token in the document with an iCol value less than zero. The iPos
** argument is ignored for a delete.
*/
int sqlite3Fts5IndexWrite(
  Fts5Index *p,                   /* Index to write to */
  int iCol,                       /* Column token appears in (-ve -> delete) */
  int iPos,                       /* Position of token within column */
  const char *pToken, int nToken  /* Token to add or remove to or from index */
);

/*
** Indicate that subsequent calls to sqlite3Fts5IndexWrite() pertain to
** document iDocid.
*/
int sqlite3Fts5IndexBeginWrite(
  Fts5Index *p,                   /* Index to write to */
  i64 iDocid                      /* Docid to add or remove data from */
);

/*
** Flush any data stored in the in-memory hash tables to the database.
** If the bCommit flag is true, also close any open blob handles.
*/
int sqlite3Fts5IndexSync(Fts5Index *p, int bCommit);

/*
** Discard any data stored in the in-memory hash tables. Do not write it
** to the database. Additionally, assume that the contents of the %_data
** table may have changed on disk. So any in-memory caches of %_data 
** records must be invalidated.
*/
int sqlite3Fts5IndexRollback(Fts5Index *p);

/*
** Retrieve and clear the current error code, respectively.
*/
int sqlite3Fts5IndexErrcode(Fts5Index*);
void sqlite3Fts5IndexReset(Fts5Index*);

/*
** Get or set the "averages" record.
*/
int sqlite3Fts5IndexGetAverages(Fts5Index *p, Fts5Buffer *pBuf);
int sqlite3Fts5IndexSetAverages(Fts5Index *p, const u8*, int);

/*
** Functions called by the storage module as part of integrity-check.
*/
u64 sqlite3Fts5IndexCksum(Fts5Config*,i64,int,int,const char*,int);
int sqlite3Fts5IndexIntegrityCheck(Fts5Index*, u64 cksum);

/* 
** Called during virtual module initialization to register UDF 
** fts5_decode() with SQLite 
*/
int sqlite3Fts5IndexInit(sqlite3*);

int sqlite3Fts5IndexSetCookie(Fts5Index*, int);

/*
** Return the total number of entries read from the %_data table by 
** this connection since it was created.
*/
int sqlite3Fts5IndexReads(Fts5Index *p);

int sqlite3Fts5IndexReinit(Fts5Index *p);
int sqlite3Fts5IndexOptimize(Fts5Index *p);
int sqlite3Fts5IndexMerge(Fts5Index *p, int nMerge);

int sqlite3Fts5IndexLoadConfig(Fts5Index *p);

/*
** End of interface to code in fts5_index.c.
**************************************************************************/

/**************************************************************************
** Interface to code in fts5_varint.c. 
*/
int sqlite3Fts5GetVarint32(const unsigned char *p, u32 *v);
int sqlite3Fts5GetVarintLen(u32 iVal);
u8 sqlite3Fts5GetVarint(const unsigned char*, u64*);
int sqlite3Fts5PutVarint(unsigned char *p, u64 v);

#define fts5GetVarint32(a,b) sqlite3Fts5GetVarint32(a,(u32*)&b)
#define fts5GetVarint    sqlite3Fts5GetVarint

/*
** End of interface to code in fts5_varint.c.
**************************************************************************/


/**************************************************************************
** Interface to code in fts5.c. 
*/

int sqlite3Fts5GetTokenizer(
  Fts5Global*, 
  const char **azArg,
  int nArg,
  Fts5Tokenizer**,
  fts5_tokenizer**,
  char **pzErr
);

Fts5Index *sqlite3Fts5IndexFromCsrid(Fts5Global*, i64, int*);

/*
** End of interface to code in fts5.c.
**************************************************************************/

/**************************************************************************
** Interface to code in fts5_hash.c. 
*/
typedef struct Fts5Hash Fts5Hash;

/*
** Create a hash table, free a hash table.
*/
int sqlite3Fts5HashNew(Fts5Hash**, int *pnSize);
void sqlite3Fts5HashFree(Fts5Hash*);

int sqlite3Fts5HashWrite(
  Fts5Hash*,
  i64 iRowid,                     /* Rowid for this entry */
  int iCol,                       /* Column token appears in (-ve -> delete) */
  int iPos,                       /* Position of token within column */
  char bByte,
  const char *pToken, int nToken  /* Token to add or remove to or from index */
);

/*
** Empty (but do not delete) a hash table.
*/
void sqlite3Fts5HashClear(Fts5Hash*);

int sqlite3Fts5HashQuery(
  Fts5Hash*,                      /* Hash table to query */
  const char *pTerm, int nTerm,   /* Query term */
  const u8 **ppDoclist,           /* OUT: Pointer to doclist for pTerm */
  int *pnDoclist                  /* OUT: Size of doclist in bytes */
);

int sqlite3Fts5HashScanInit(
  Fts5Hash*,                      /* Hash table to query */
  const char *pTerm, int nTerm    /* Query prefix */
);
void sqlite3Fts5HashScanNext(Fts5Hash*);
int sqlite3Fts5HashScanEof(Fts5Hash*);
void sqlite3Fts5HashScanEntry(Fts5Hash *,
  const char **pzTerm,            /* OUT: term (nul-terminated) */
  const u8 **ppDoclist,           /* OUT: pointer to doclist */
  int *pnDoclist                  /* OUT: size of doclist in bytes */
);


/*
** End of interface to code in fts5_hash.c.
**************************************************************************/

/**************************************************************************
** Interface to code in fts5_storage.c. fts5_storage.c contains contains 
** code to access the data stored in the %_content and %_docsize tables.
*/

#define FTS5_STMT_SCAN_ASC  0     /* SELECT rowid, * FROM ... ORDER BY 1 ASC */
#define FTS5_STMT_SCAN_DESC 1     /* SELECT rowid, * FROM ... ORDER BY 1 DESC */
#define FTS5_STMT_LOOKUP    2     /* SELECT rowid, * FROM ... WHERE rowid=? */

typedef struct Fts5Storage Fts5Storage;

int sqlite3Fts5StorageOpen(Fts5Config*, Fts5Index*, int, Fts5Storage**, char**);
int sqlite3Fts5StorageClose(Fts5Storage *p);
int sqlite3Fts5StorageRename(Fts5Storage*, const char *zName);

int sqlite3Fts5DropAll(Fts5Config*);
int sqlite3Fts5CreateTable(Fts5Config*, const char*, const char*, int, char **);

int sqlite3Fts5StorageDelete(Fts5Storage *p, i64);
int sqlite3Fts5StorageInsert(Fts5Storage *p, sqlite3_value **apVal, int, i64*);

int sqlite3Fts5StorageIntegrity(Fts5Storage *p);

int sqlite3Fts5StorageStmt(Fts5Storage *p, int eStmt, sqlite3_stmt**, char**);
void sqlite3Fts5StorageStmtRelease(Fts5Storage *p, int eStmt, sqlite3_stmt*);

int sqlite3Fts5StorageDocsize(Fts5Storage *p, i64 iRowid, int *aCol);
int sqlite3Fts5StorageSize(Fts5Storage *p, int iCol, i64 *pnAvg);
int sqlite3Fts5StorageRowCount(Fts5Storage *p, i64 *pnRow);

int sqlite3Fts5StorageSync(Fts5Storage *p, int bCommit);
int sqlite3Fts5StorageRollback(Fts5Storage *p);

int sqlite3Fts5StorageConfigValue(
    Fts5Storage *p, const char*, sqlite3_value*, int
);

int sqlite3Fts5StorageSpecialDelete(Fts5Storage *p, i64 iDel, sqlite3_value**);

int sqlite3Fts5StorageDeleteAll(Fts5Storage *p);
int sqlite3Fts5StorageRebuild(Fts5Storage *p);
int sqlite3Fts5StorageOptimize(Fts5Storage *p);
int sqlite3Fts5StorageMerge(Fts5Storage *p, int nMerge);

/*
** End of interface to code in fts5_storage.c.
**************************************************************************/


/**************************************************************************
** Interface to code in fts5_expr.c. 
*/
typedef struct Fts5Expr Fts5Expr;
typedef struct Fts5ExprNode Fts5ExprNode;
typedef struct Fts5Parse Fts5Parse;
typedef struct Fts5Token Fts5Token;
typedef struct Fts5ExprPhrase Fts5ExprPhrase;
typedef struct Fts5ExprNearset Fts5ExprNearset;
typedef struct Fts5ExprColset Fts5ExprColset;

struct Fts5Token {
  const char *p;                  /* Token text (not NULL terminated) */
  int n;                          /* Size of buffer p in bytes */
};

/* Parse a MATCH expression. */
int sqlite3Fts5ExprNew(
  Fts5Config *pConfig, 
  const char *zExpr,
  Fts5Expr **ppNew, 
  char **pzErr
);

/*
** for(rc = sqlite3Fts5ExprFirst(pExpr, pIdx, bDesc);
**     rc==SQLITE_OK && 0==sqlite3Fts5ExprEof(pExpr);
**     rc = sqlite3Fts5ExprNext(pExpr)
** ){
**   // The document with rowid iRowid matches the expression!
**   i64 iRowid = sqlite3Fts5ExprRowid(pExpr);
** }
*/
int sqlite3Fts5ExprFirst(Fts5Expr*, Fts5Index *pIdx, i64 iMin, int bDesc);
int sqlite3Fts5ExprNext(Fts5Expr*, i64 iMax);
int sqlite3Fts5ExprEof(Fts5Expr*);
i64 sqlite3Fts5ExprRowid(Fts5Expr*);

void sqlite3Fts5ExprFree(Fts5Expr*);

/* Called during startup to register a UDF with SQLite */
int sqlite3Fts5ExprInit(Fts5Global*, sqlite3*);

int sqlite3Fts5ExprPhraseCount(Fts5Expr*);
int sqlite3Fts5ExprPhraseSize(Fts5Expr*, int iPhrase);
int sqlite3Fts5ExprPoslist(Fts5Expr*, int, const u8 **);

int sqlite3Fts5ExprPhraseExpr(Fts5Config*, Fts5Expr*, int, Fts5Expr**);

/*******************************************
** The fts5_expr.c API above this point is used by the other hand-written
** C code in this module. The interfaces below this point are called by
** the parser code in fts5parse.y.  */

void sqlite3Fts5ParseError(Fts5Parse *pParse, const char *zFmt, ...);

Fts5ExprNode *sqlite3Fts5ParseNode(
  Fts5Parse *pParse,
  int eType,
  Fts5ExprNode *pLeft,
  Fts5ExprNode *pRight,
  Fts5ExprNearset *pNear
);

Fts5ExprPhrase *sqlite3Fts5ParseTerm(
  Fts5Parse *pParse, 
  Fts5ExprPhrase *pPhrase, 
  Fts5Token *pToken,
  int bPrefix
);

Fts5ExprNearset *sqlite3Fts5ParseNearset(
  Fts5Parse*, 
  Fts5ExprNearset*,
  Fts5ExprPhrase* 
);

Fts5ExprColset *sqlite3Fts5ParseColset(
  Fts5Parse*, 
  Fts5ExprColset*, 
  Fts5Token *
);

void sqlite3Fts5ParsePhraseFree(Fts5ExprPhrase*);
void sqlite3Fts5ParseNearsetFree(Fts5ExprNearset*);
void sqlite3Fts5ParseNodeFree(Fts5ExprNode*);

void sqlite3Fts5ParseSetDistance(Fts5Parse*, Fts5ExprNearset*, Fts5Token*);
void sqlite3Fts5ParseSetColset(Fts5Parse*, Fts5ExprNearset*, Fts5ExprColset*);
void sqlite3Fts5ParseFinished(Fts5Parse *pParse, Fts5ExprNode *p);
void sqlite3Fts5ParseNear(Fts5Parse *pParse, Fts5Token*);

/*
** End of interface to code in fts5_expr.c.
**************************************************************************/



/**************************************************************************
** Interface to code in fts5_aux.c. 
*/

int sqlite3Fts5AuxInit(fts5_api*);
/*
** End of interface to code in fts5_aux.c.
**************************************************************************/

/**************************************************************************
** Interface to code in fts5_tokenizer.c. 
*/

int sqlite3Fts5TokenizerInit(fts5_api*);
/*
** End of interface to code in fts5_tokenizer.c.
**************************************************************************/

/**************************************************************************
** Interface to code in fts5_sorter.c. 
*/
typedef struct Fts5Sorter Fts5Sorter;

int sqlite3Fts5SorterNew(Fts5Expr *pExpr, Fts5Sorter **pp);

/*
** End of interface to code in fts5_sorter.c.
**************************************************************************/

/**************************************************************************
** Interface to code in fts5_vocab.c. 
*/

int sqlite3Fts5VocabInit(Fts5Global*, sqlite3*);

/*
** End of interface to code in fts5_vocab.c.
**************************************************************************/


/**************************************************************************
** Interface to automatically generated code in fts5_unicode2.c. 
*/
int sqlite3Fts5UnicodeIsalnum(int c);
int sqlite3Fts5UnicodeIsdiacritic(int c);
int sqlite3Fts5UnicodeFold(int c, int bRemoveDiacritic);
/*
** End of interface to code in fts5_unicode2.c.
**************************************************************************/

#endif
#endif

/*
** 2014 May 31
**
** The author disclaims copyright to this source code.  In place of
** a legal notice, here is a blessing:
**
**    May you do good and not evil.
**    May you find forgiveness for yourself and forgive others.
**    May you share freely, never taking more than you give.
**
******************************************************************************
*/

#ifdef SQLITE_ENABLE_FTS5

#include "fts5Int.h"
#include <math.h>

/*
** Object used to iterate through all "coalesced phrase instances" in 
** a single column of the current row. If the phrase instances in the
** column being considered do not overlap, this object simply iterates
** through them. Or, if they do overlap (share one or more tokens in
** common), each set of overlapping instances is treated as a single
** match. See documentation for the highlight() auxiliary function for
** details.
**
** Usage is:
**
**   for(rc = fts5CInstIterNext(pApi, pFts, iCol, &iter);
**      (rc==SQLITE_OK && 0==fts5CInstIterEof(&iter);
**      rc = fts5CInstIterNext(&iter)
**   ){
**     printf("instance starts at %d, ends at %d\n", iter.iStart, iter.iEnd);
**   }
**
*/
typedef struct CInstIter CInstIter;
struct CInstIter {
  const Fts5ExtensionApi *pApi;   /* API offered by current FTS version */
  Fts5Context *pFts;              /* First arg to pass to pApi functions */
  int iCol;                       /* Column to search */
  int iInst;                      /* Next phrase instance index */
  int nInst;                      /* Total number of phrase instances */

  /* Output variables */
  int iStart;                     /* First token in coalesced phrase instance */
  int iEnd;                       /* Last token in coalesced phrase instance */
};

/*
** Advance the iterator to the next coalesced phrase instance. Return
** an SQLite error code if an error occurs, or SQLITE_OK otherwise.
*/
static int fts5CInstIterNext(CInstIter *pIter){
  int rc = SQLITE_OK;
  pIter->iStart = -1;
  pIter->iEnd = -1;

  while( rc==SQLITE_OK && pIter->iInst<pIter->nInst ){
    int ip; int ic; int io;
    rc = pIter->pApi->xInst(pIter->pFts, pIter->iInst, &ip, &ic, &io);
    if( rc==SQLITE_OK ){
      if( ic==pIter->iCol ){
        int iEnd = io - 1 + pIter->pApi->xPhraseSize(pIter->pFts, ip);
        if( pIter->iStart<0 ){
          pIter->iStart = io;
          pIter->iEnd = iEnd;
        }else if( io<=pIter->iEnd ){
          if( iEnd>pIter->iEnd ) pIter->iEnd = iEnd;
        }else{
          break;
        }
      }
      pIter->iInst++;
    }
  }

  return rc;
}

/*
** Initialize the iterator object indicated by the final parameter to 
** iterate through coalesced phrase instances in column iCol.
*/
static int fts5CInstIterInit(
  const Fts5ExtensionApi *pApi,
  Fts5Context *pFts,
  int iCol,
  CInstIter *pIter
){
  int rc;

  memset(pIter, 0, sizeof(CInstIter));
  pIter->pApi = pApi;
  pIter->pFts = pFts;
  pIter->iCol = iCol;
  rc = pApi->xInstCount(pFts, &pIter->nInst);

  if( rc==SQLITE_OK ){
    rc = fts5CInstIterNext(pIter);
  }

  return rc;
}



/*************************************************************************
** Start of highlight() implementation.
*/
typedef struct HighlightContext HighlightContext;
struct HighlightContext {
  CInstIter iter;                 /* Coalesced Instance Iterator */
  int iPos;                       /* Current token offset in zIn[] */
  int iRangeStart;                /* First token to include */
  int iRangeEnd;                  /* If non-zero, last token to include */
  const char *zOpen;              /* Opening highlight */
  const char *zClose;             /* Closing highlight */
  const char *zIn;                /* Input text */
  int nIn;                        /* Size of input text in bytes */
  int iOff;                       /* Current offset within zIn[] */
  char *zOut;                     /* Output value */
};

/*
** Append text to the HighlightContext output string - p->zOut. Argument
** z points to a buffer containing n bytes of text to append. If n is 
** negative, everything up until the first '\0' is appended to the output.
**
** If *pRc is set to any value other than SQLITE_OK when this function is 
** called, it is a no-op. If an error (i.e. an OOM condition) is encountered, 
** *pRc is set to an error code before returning. 
*/
static void fts5HighlightAppend(
  int *pRc, 
  HighlightContext *p, 
  const char *z, int n
){
  if( *pRc==SQLITE_OK ){
    if( n<0 ) n = strlen(z);
    p->zOut = sqlite3_mprintf("%z%.*s", p->zOut, n, z);
    if( p->zOut==0 ) *pRc = SQLITE_NOMEM;
  }
}

/*
** Tokenizer callback used by implementation of highlight() function.
*/
static int fts5HighlightCb(
  void *pContext,                 /* Pointer to HighlightContext object */
  const char *pToken,             /* Buffer containing token */
  int nToken,                     /* Size of token in bytes */
  int iStartOff,                  /* Start offset of token */
  int iEndOff                     /* End offset of token */
){
  HighlightContext *p = (HighlightContext*)pContext;
  int rc = SQLITE_OK;
  int iPos = p->iPos++;

  if( p->iRangeEnd>0 ){
    if( iPos<p->iRangeStart || iPos>p->iRangeEnd ) return SQLITE_OK;
    if( p->iRangeStart && iPos==p->iRangeStart ) p->iOff = iStartOff;
  }

  if( iPos==p->iter.iStart ){
    fts5HighlightAppend(&rc, p, &p->zIn[p->iOff], iStartOff - p->iOff);
    fts5HighlightAppend(&rc, p, p->zOpen, -1);
    p->iOff = iStartOff;
  }

  if( iPos==p->iter.iEnd ){
    if( p->iRangeEnd && p->iter.iStart<p->iRangeStart ){
      fts5HighlightAppend(&rc, p, p->zOpen, -1);
    }
    fts5HighlightAppend(&rc, p, &p->zIn[p->iOff], iEndOff - p->iOff);
    fts5HighlightAppend(&rc, p, p->zClose, -1);
    p->iOff = iEndOff;
    if( rc==SQLITE_OK ){
      rc = fts5CInstIterNext(&p->iter);
    }
  }

  if( p->iRangeEnd>0 && iPos==p->iRangeEnd ){
    fts5HighlightAppend(&rc, p, &p->zIn[p->iOff], iEndOff - p->iOff);
    p->iOff = iEndOff;
    if( iPos<p->iter.iEnd ){
      fts5HighlightAppend(&rc, p, p->zClose, -1);
    }
  }

  return rc;
}

/*
** Implementation of highlight() function.
*/
static void fts5HighlightFunction(
  const Fts5ExtensionApi *pApi,   /* API offered by current FTS version */
  Fts5Context *pFts,              /* First arg to pass to pApi functions */
  sqlite3_context *pCtx,          /* Context for returning result/error */
  int nVal,                       /* Number of values in apVal[] array */
  sqlite3_value **apVal           /* Array of trailing arguments */
){
  HighlightContext ctx;
  int rc;
  int iCol;

  if( nVal!=3 ){
    const char *zErr = "wrong number of arguments to function highlight()";
    sqlite3_result_error(pCtx, zErr, -1);
    return;
  }

  iCol = sqlite3_value_int(apVal[0]);
  memset(&ctx, 0, sizeof(HighlightContext));
  ctx.zOpen = (const char*)sqlite3_value_text(apVal[1]);
  ctx.zClose = (const char*)sqlite3_value_text(apVal[2]);
  rc = pApi->xColumnText(pFts, iCol, &ctx.zIn, &ctx.nIn);

  if( ctx.zIn ){
    if( rc==SQLITE_OK ){
      rc = fts5CInstIterInit(pApi, pFts, iCol, &ctx.iter);
    }

    if( rc==SQLITE_OK ){
      rc = pApi->xTokenize(pFts, ctx.zIn, ctx.nIn, (void*)&ctx,fts5HighlightCb);
    }
    fts5HighlightAppend(&rc, &ctx, &ctx.zIn[ctx.iOff], ctx.nIn - ctx.iOff);

    if( rc==SQLITE_OK ){
      sqlite3_result_text(pCtx, (const char*)ctx.zOut, -1, SQLITE_TRANSIENT);
    }
    sqlite3_free(ctx.zOut);
  }
  if( rc!=SQLITE_OK ){
    sqlite3_result_error_code(pCtx, rc);
  }
}
/*
** End of highlight() implementation.
**************************************************************************/

/*
** Implementation of snippet() function.
*/
static void fts5SnippetFunction(
  const Fts5ExtensionApi *pApi,   /* API offered by current FTS version */
  Fts5Context *pFts,              /* First arg to pass to pApi functions */
  sqlite3_context *pCtx,          /* Context for returning result/error */
  int nVal,                       /* Number of values in apVal[] array */
  sqlite3_value **apVal           /* Array of trailing arguments */
){
  HighlightContext ctx;
  int rc = SQLITE_OK;             /* Return code */
  int iCol;                       /* 1st argument to snippet() */
  const char *zEllips;            /* 4th argument to snippet() */
  int nToken;                     /* 5th argument to snippet() */
  int nInst;                      /* Number of instance matches this row */
  int i;                          /* Used to iterate through instances */
  int nPhrase;                    /* Number of phrases in query */
  unsigned char *aSeen;           /* Array of "seen instance" flags */
  int iBestCol;                   /* Column containing best snippet */
  int iBestStart = 0;             /* First token of best snippet */
  int iBestLast;                  /* Last token of best snippet */
  int nBestScore = 0;             /* Score of best snippet */
  int nColSize;                   /* Total size of iBestCol in tokens */

  if( nVal!=5 ){
    const char *zErr = "wrong number of arguments to function snippet()";
    sqlite3_result_error(pCtx, zErr, -1);
    return;
  }

  memset(&ctx, 0, sizeof(HighlightContext));
  iCol = sqlite3_value_int(apVal[0]);
  ctx.zOpen = (const char*)sqlite3_value_text(apVal[1]);
  ctx.zClose = (const char*)sqlite3_value_text(apVal[2]);
  zEllips = (const char*)sqlite3_value_text(apVal[3]);
  nToken = sqlite3_value_int(apVal[4]);
  iBestLast = nToken-1;

  iBestCol = (iCol>=0 ? iCol : 0);
  nPhrase = pApi->xPhraseCount(pFts);
  aSeen = sqlite3_malloc(nPhrase);
  if( aSeen==0 ){
    rc = SQLITE_NOMEM;
  }

  if( rc==SQLITE_OK ){
    rc = pApi->xInstCount(pFts, &nInst);
  }
  for(i=0; rc==SQLITE_OK && i<nInst; i++){
    int ip, iSnippetCol, iStart;
    memset(aSeen, 0, nPhrase);
    rc = pApi->xInst(pFts, i, &ip, &iSnippetCol, &iStart);
    if( rc==SQLITE_OK && (iCol<0 || iSnippetCol==iCol) ){
      int nScore = 1000;
      int iLast = iStart - 1 + pApi->xPhraseSize(pFts, ip);
      int j;
      aSeen[ip] = 1;

      for(j=i+1; rc==SQLITE_OK && j<nInst; j++){
        int ic; int io; int iFinal;
        rc = pApi->xInst(pFts, j, &ip, &ic, &io);
        iFinal = io + pApi->xPhraseSize(pFts, ip) - 1;
        if( rc==SQLITE_OK && ic==iSnippetCol && iLast<iStart+nToken ){
          nScore += aSeen[ip] ? 1000 : 1;
          aSeen[ip] = 1;
          if( iFinal>iLast ) iLast = iFinal;
        }
      }

      if( rc==SQLITE_OK && nScore>nBestScore ){
        iBestCol = iSnippetCol;
        iBestStart = iStart;
        iBestLast = iLast;
        nBestScore = nScore;
      }
    }
  }

  if( rc==SQLITE_OK ){
    rc = pApi->xColumnSize(pFts, iBestCol, &nColSize);
  }
  if( rc==SQLITE_OK ){
    rc = pApi->xColumnText(pFts, iBestCol, &ctx.zIn, &ctx.nIn);
  }
  if( ctx.zIn ){
    if( rc==SQLITE_OK ){
      rc = fts5CInstIterInit(pApi, pFts, iBestCol, &ctx.iter);
    }

    if( (iBestStart+nToken-1)>iBestLast ){
      iBestStart -= (iBestStart+nToken-1-iBestLast) / 2;
    }
    if( iBestStart+nToken>nColSize ){
      iBestStart = nColSize - nToken;
    }
    if( iBestStart<0 ) iBestStart = 0;

    ctx.iRangeStart = iBestStart;
    ctx.iRangeEnd = iBestStart + nToken - 1;

    if( iBestStart>0 ){
      fts5HighlightAppend(&rc, &ctx, zEllips, -1);
    }
    if( rc==SQLITE_OK ){
      rc = pApi->xTokenize(pFts, ctx.zIn, ctx.nIn, (void*)&ctx,fts5HighlightCb);
    }
    if( ctx.iRangeEnd>=(nColSize-1) ){
      fts5HighlightAppend(&rc, &ctx, &ctx.zIn[ctx.iOff], ctx.nIn - ctx.iOff);
    }else{
      fts5HighlightAppend(&rc, &ctx, zEllips, -1);
    }

    if( rc==SQLITE_OK ){
      sqlite3_result_text(pCtx, (const char*)ctx.zOut, -1, SQLITE_TRANSIENT);
    }else{
      sqlite3_result_error_code(pCtx, rc);
    }
    sqlite3_free(ctx.zOut);
  }
  sqlite3_free(aSeen);
}

/************************************************************************/

/*
** The first time the bm25() function is called for a query, an instance
** of the following structure is allocated and populated.
*/
typedef struct Fts5Bm25Data Fts5Bm25Data;
struct Fts5Bm25Data {
  int nPhrase;                    /* Number of phrases in query */
  double avgdl;                   /* Average number of tokens in each row */
  double *aIDF;                   /* IDF for each phrase */
  double *aFreq;                  /* Array used to calculate phrase freq. */
};

/*
** Callback used by fts5Bm25GetData() to count the number of rows in the
** table matched by each individual phrase within the query.
*/
static int fts5CountCb(
  const Fts5ExtensionApi *pApi, 
  Fts5Context *pFts,
  void *pUserData                 /* Pointer to sqlite3_int64 variable */
){
  sqlite3_int64 *pn = (sqlite3_int64*)pUserData;
  (*pn)++;
  return SQLITE_OK;
}

/*
** Set *ppData to point to the Fts5Bm25Data object for the current query. 
** If the object has not already been allocated, allocate and populate it
** now.
*/
static int fts5Bm25GetData(
  const Fts5ExtensionApi *pApi, 
  Fts5Context *pFts,
  Fts5Bm25Data **ppData           /* OUT: bm25-data object for this query */
){
  int rc = SQLITE_OK;             /* Return code */
  Fts5Bm25Data *p;                /* Object to return */

  p = pApi->xGetAuxdata(pFts, 0);
  if( p==0 ){
    int nPhrase;                  /* Number of phrases in query */
    sqlite3_int64 nRow;           /* Number of rows in table */
    sqlite3_int64 nToken;         /* Number of tokens in table */
    int nByte;                    /* Bytes of space to allocate */
    int i;

    /* Allocate the Fts5Bm25Data object */
    nPhrase = pApi->xPhraseCount(pFts);
    nByte = sizeof(Fts5Bm25Data) + nPhrase*2*sizeof(double);
    p = (Fts5Bm25Data*)sqlite3_malloc(nByte);
    if( p==0 ){
      rc = SQLITE_NOMEM;
    }else{
      memset(p, 0, nByte);
      p->nPhrase = nPhrase;
      p->aIDF = (double*)&p[1];
      p->aFreq = &p->aIDF[nPhrase];
    }

    /* Calculate the average document length for this FTS5 table */
    if( rc==SQLITE_OK ) rc = pApi->xRowCount(pFts, &nRow);
    if( rc==SQLITE_OK ) rc = pApi->xColumnTotalSize(pFts, -1, &nToken);
    if( rc==SQLITE_OK ) p->avgdl = (double)nToken  / (double)nRow;

    /* Calculate an IDF for each phrase in the query */
    for(i=0; rc==SQLITE_OK && i<nPhrase; i++){
      sqlite3_int64 nHit = 0;
      rc = pApi->xQueryPhrase(pFts, i, (void*)&nHit, fts5CountCb);
      if( rc==SQLITE_OK ){
        /* Calculate the IDF (Inverse Document Frequency) for phrase i.
        ** This is done using the standard BM25 formula as found on wikipedia:
        **
        **   IDF = log( (N - nHit + 0.5) / (nHit + 0.5) )
        **
        ** where "N" is the total number of documents in the set and nHit
        ** is the number that contain at least one instance of the phrase
        ** under consideration.
        **
        ** The problem with this is that if (N < 2*nHit), the IDF is 
        ** negative. Which is undesirable. So the mimimum allowable IDF is
        ** (1e-6) - roughly the same as a term that appears in just over
        ** half of set of 5,000,000 documents.  */
        double idf = log( (nRow - nHit + 0.5) / (nHit + 0.5) );
        if( idf<=0.0 ) idf = 1e-6;
        p->aIDF[i] = idf;
      }
    }

    if( rc!=SQLITE_OK ){
      sqlite3_free(p);
    }else{
      rc = pApi->xSetAuxdata(pFts, p, sqlite3_free);
    }
    if( rc!=SQLITE_OK ) p = 0;
  }
  *ppData = p;
  return rc;
}

/*
** Implementation of bm25() function.
*/
static void fts5Bm25Function(
  const Fts5ExtensionApi *pApi,   /* API offered by current FTS version */
  Fts5Context *pFts,              /* First arg to pass to pApi functions */
  sqlite3_context *pCtx,          /* Context for returning result/error */
  int nVal,                       /* Number of values in apVal[] array */
  sqlite3_value **apVal           /* Array of trailing arguments */
){
  const double k1 = 1.2;          /* Constant "k1" from BM25 formula */
  const double b = 0.75;          /* Constant "b" from BM25 formula */
  int rc = SQLITE_OK;             /* Error code */
  double score = 0.0;             /* SQL function return value */
  Fts5Bm25Data *pData;            /* Values allocated/calculated once only */
  int i;                          /* Iterator variable */
  int nInst;                      /* Value returned by xInstCount() */
  double D;                       /* Total number of tokens in row */
  double *aFreq;                  /* Array of phrase freq. for current row */

  /* Calculate the phrase frequency (symbol "f(qi,D)" in the documentation)
  ** for each phrase in the query for the current row. */
  rc = fts5Bm25GetData(pApi, pFts, &pData);
  if( rc==SQLITE_OK ){
    aFreq = pData->aFreq;
    memset(aFreq, 0, sizeof(double) * pData->nPhrase);
    rc = pApi->xInstCount(pFts, &nInst);
  }
  for(i=0; rc==SQLITE_OK && i<nInst; i++){
    int ip; int ic; int io;
    rc = pApi->xInst(pFts, i, &ip, &ic, &io);
    if( rc==SQLITE_OK ){
      double w = (nVal > ic) ? sqlite3_value_double(apVal[ic]) : 1.0;
      aFreq[ip] += w;
    }
  }

  /* Figure out the total size of the current row in tokens. */
  if( rc==SQLITE_OK ){
    int nTok;
    rc = pApi->xColumnSize(pFts, -1, &nTok);
    D = (double)nTok;
  }

  /* Determine the BM25 score for the current row. */
  for(i=0; rc==SQLITE_OK && i<pData->nPhrase; i++){
    score += pData->aIDF[i] * (
      ( aFreq[i] * (k1 + 1.0) ) / 
      ( aFreq[i] + k1 * (1 - b + b * D / pData->avgdl) )
    );
  }
  
  /* If no error has occurred, return the calculated score. Otherwise,
  ** throw an SQL exception.  */
  if( rc==SQLITE_OK ){
    sqlite3_result_double(pCtx, -1.0 * score);
  }else{
    sqlite3_result_error_code(pCtx, rc);
  }
}

int sqlite3Fts5AuxInit(fts5_api *pApi){
  struct Builtin {
    const char *zFunc;            /* Function name (nul-terminated) */
    void *pUserData;              /* User-data pointer */
    fts5_extension_function xFunc;/* Callback function */
    void (*xDestroy)(void*);      /* Destructor function */
  } aBuiltin [] = {
    { "snippet",   0, fts5SnippetFunction, 0 },
    { "highlight", 0, fts5HighlightFunction, 0 },
    { "bm25",      0, fts5Bm25Function,    0 },
  };
  int rc = SQLITE_OK;             /* Return code */
  int i;                          /* To iterate through builtin functions */

  for(i=0; rc==SQLITE_OK && i<sizeof(aBuiltin)/sizeof(aBuiltin[0]); i++){
    rc = pApi->xCreateFunction(pApi,
        aBuiltin[i].zFunc,
        aBuiltin[i].pUserData,
        aBuiltin[i].xFunc,
        aBuiltin[i].xDestroy
    );
  }

  return rc;
}
#endif /* SQLITE_ENABLE_FTS5 */

/*
** 2014 May 31
**
** The author disclaims copyright to this source code.  In place of
** a legal notice, here is a blessing:
**
**    May you do good and not evil.
**    May you find forgiveness for yourself and forgive others.
**    May you share freely, never taking more than you give.
**
******************************************************************************
*/


#ifdef SQLITE_ENABLE_FTS5

#include "fts5Int.h"

int sqlite3Fts5BufferGrow(int *pRc, Fts5Buffer *pBuf, int nByte){
  /* A no-op if an error has already occurred */
  if( *pRc ) return 1;

  if( (pBuf->n + nByte) > pBuf->nSpace ){
    u8 *pNew;
    int nNew = pBuf->nSpace ? pBuf->nSpace*2 : 64;
    while( nNew<(pBuf->n + nByte) ){
      nNew = nNew * 2;
    }
    pNew = sqlite3_realloc(pBuf->p, nNew);
    if( pNew==0 ){
      *pRc = SQLITE_NOMEM;
      return 1;
    }else{
      pBuf->nSpace = nNew;
      pBuf->p = pNew;
    }
  }
  return 0;
}

/*
** Encode value iVal as an SQLite varint and append it to the buffer object
** pBuf. If an OOM error occurs, set the error code in p.
*/
void sqlite3Fts5BufferAppendVarint(int *pRc, Fts5Buffer *pBuf, i64 iVal){
  if( sqlite3Fts5BufferGrow(pRc, pBuf, 9) ) return;
  pBuf->n += sqlite3Fts5PutVarint(&pBuf->p[pBuf->n], iVal);
}

void sqlite3Fts5Put32(u8 *aBuf, int iVal){
  aBuf[0] = (iVal>>24) & 0x00FF;
  aBuf[1] = (iVal>>16) & 0x00FF;
  aBuf[2] = (iVal>> 8) & 0x00FF;
  aBuf[3] = (iVal>> 0) & 0x00FF;
}

int sqlite3Fts5Get32(const u8 *aBuf){
  return (aBuf[0] << 24) + (aBuf[1] << 16) + (aBuf[2] << 8) + aBuf[3];
}

void sqlite3Fts5BufferAppend32(int *pRc, Fts5Buffer *pBuf, int iVal){
  if( sqlite3Fts5BufferGrow(pRc, pBuf, 4) ) return;
  sqlite3Fts5Put32(&pBuf->p[pBuf->n], iVal);
  pBuf->n += 4;
}

/*
** Append buffer nData/pData to buffer pBuf. If an OOM error occurs, set 
** the error code in p. If an error has already occurred when this function
** is called, it is a no-op.
*/
void sqlite3Fts5BufferAppendBlob(
  int *pRc,
  Fts5Buffer *pBuf, 
  int nData, 
  const u8 *pData
){
  assert( *pRc || nData>=0 );
  if( sqlite3Fts5BufferGrow(pRc, pBuf, nData) ) return;
  memcpy(&pBuf->p[pBuf->n], pData, nData);
  pBuf->n += nData;
}

/*
** Append the nul-terminated string zStr to the buffer pBuf. This function
** ensures that the byte following the buffer data is set to 0x00, even 
** though this byte is not included in the pBuf->n count.
*/
void sqlite3Fts5BufferAppendString(
  int *pRc,
  Fts5Buffer *pBuf, 
  const char *zStr
){
  int nStr = strlen(zStr);
  sqlite3Fts5BufferAppendBlob(pRc, pBuf, nStr+1, (const u8*)zStr);
  pBuf->n--;
}

/*
** Argument zFmt is a printf() style format string. This function performs
** the printf() style processing, then appends the results to buffer pBuf.
**
** Like sqlite3Fts5BufferAppendString(), this function ensures that the byte 
** following the buffer data is set to 0x00, even though this byte is not
** included in the pBuf->n count.
*/ 
void sqlite3Fts5BufferAppendPrintf(
  int *pRc,
  Fts5Buffer *pBuf, 
  char *zFmt, ...
){
  if( *pRc==SQLITE_OK ){
    char *zTmp;
    va_list ap;
    va_start(ap, zFmt);
    zTmp = sqlite3_vmprintf(zFmt, ap);
    va_end(ap);

    if( zTmp==0 ){
      *pRc = SQLITE_NOMEM;
    }else{
      sqlite3Fts5BufferAppendString(pRc, pBuf, zTmp);
      sqlite3_free(zTmp);
    }
  }
}

char *sqlite3Fts5Mprintf(int *pRc, const char *zFmt, ...){
  char *zRet = 0;
  if( *pRc==SQLITE_OK ){
    va_list ap;
    va_start(ap, zFmt);
    zRet = sqlite3_vmprintf(zFmt, ap);
    va_end(ap);
    if( zRet==0 ){
      *pRc = SQLITE_NOMEM; 
    }
  }
  return zRet;
}
 

/*
** Free any buffer allocated by pBuf. Zero the structure before returning.
*/
void sqlite3Fts5BufferFree(Fts5Buffer *pBuf){
  sqlite3_free(pBuf->p);
  memset(pBuf, 0, sizeof(Fts5Buffer));
}

/*
** Zero the contents of the buffer object. But do not free the associated 
** memory allocation.
*/
void sqlite3Fts5BufferZero(Fts5Buffer *pBuf){
  pBuf->n = 0;
}

/*
** Set the buffer to contain nData/pData. If an OOM error occurs, leave an
** the error code in p. If an error has already occurred when this function
** is called, it is a no-op.
*/
void sqlite3Fts5BufferSet(
  int *pRc,
  Fts5Buffer *pBuf, 
  int nData, 
  const u8 *pData
){
  pBuf->n = 0;
  sqlite3Fts5BufferAppendBlob(pRc, pBuf, nData, pData);
}

int sqlite3Fts5PoslistNext64(
  const u8 *a, int n,             /* Buffer containing poslist */
  int *pi,                        /* IN/OUT: Offset within a[] */
  i64 *piOff                      /* IN/OUT: Current offset */
){
  int i = *pi;
  if( i>=n ){
    /* EOF */
    *piOff = -1;
    return 1;  
  }else{
    i64 iOff = *piOff;
    int iVal;
    i += fts5GetVarint32(&a[i], iVal);
    if( iVal==1 ){
      i += fts5GetVarint32(&a[i], iVal);
      iOff = ((i64)iVal) << 32;
      i += fts5GetVarint32(&a[i], iVal);
    }
    *piOff = iOff + (iVal-2);
    *pi = i;
    return 0;
  }
}


/*
** Advance the iterator object passed as the only argument. Return true
** if the iterator reaches EOF, or false otherwise.
*/
int sqlite3Fts5PoslistReaderNext(Fts5PoslistReader *pIter){
  if( sqlite3Fts5PoslistNext64(pIter->a, pIter->n, &pIter->i, &pIter->iPos) 
   || (pIter->iCol>=0 && (pIter->iPos >> 32) > pIter->iCol)
  ){
    pIter->bEof = 1;
  }
  return pIter->bEof;
}

int sqlite3Fts5PoslistReaderInit(
  int iCol,                       /* If (iCol>=0), this column only */
  const u8 *a, int n,             /* Poslist buffer to iterate through */
  Fts5PoslistReader *pIter        /* Iterator object to initialize */
){
  memset(pIter, 0, sizeof(*pIter));
  pIter->a = a;
  pIter->n = n;
  pIter->iCol = iCol;
  do {
    sqlite3Fts5PoslistReaderNext(pIter);
  }while( pIter->bEof==0 && (pIter->iPos >> 32)<iCol );
  return pIter->bEof;
}

int sqlite3Fts5PoslistWriterAppend(
  Fts5Buffer *pBuf, 
  Fts5PoslistWriter *pWriter,
  i64 iPos
){
  static const i64 colmask = ((i64)(0x7FFFFFFF)) << 32;
  int rc = SQLITE_OK;
  if( (iPos & colmask) != (pWriter->iPrev & colmask) ){
    fts5BufferAppendVarint(&rc, pBuf, 1);
    fts5BufferAppendVarint(&rc, pBuf, (iPos >> 32));
    pWriter->iPrev = (iPos & colmask);
  }
  fts5BufferAppendVarint(&rc, pBuf, (iPos - pWriter->iPrev) + 2);
  pWriter->iPrev = iPos;
  return rc;
}

void *sqlite3Fts5MallocZero(int *pRc, int nByte){
  void *pRet = 0;
  if( *pRc==SQLITE_OK ){
    pRet = sqlite3_malloc(nByte);
    if( pRet==0 && nByte>0 ){
      *pRc = SQLITE_NOMEM;
    }else{
      memset(pRet, 0, nByte);
    }
  }
  return pRet;
}

/*
** Return a nul-terminated copy of the string indicated by pIn. If nIn
** is non-negative, then it is the length of the string in bytes. Otherwise,
** the length of the string is determined using strlen().
**
** It is the responsibility of the caller to eventually free the returned
** buffer using sqlite3_free(). If an OOM error occurs, NULL is returned. 
*/
char *sqlite3Fts5Strndup(int *pRc, const char *pIn, int nIn){
  char *zRet = 0;
  if( *pRc==SQLITE_OK ){
    if( nIn<0 ){
      nIn = strlen(pIn);
    }
    zRet = (char*)sqlite3_malloc(nIn+1);
    if( zRet ){
      memcpy(zRet, pIn, nIn);
      zRet[nIn] = '\0';
    }else{
      *pRc = SQLITE_NOMEM;
    }
  }
  return zRet;
}


/*
** Return true if character 't' may be part of an FTS5 bareword, or false
** otherwise. Characters that may be part of barewords:
**
**   * All non-ASCII characters,
**   * The 52 upper and lower case ASCII characters, and
**   * The 10 integer ASCII characters.
**   * The underscore character "_" (0x5F).
*/
int sqlite3Fts5IsBareword(char t){
  u8 aBareword[128] = {
    0, 0, 0, 0, 0, 0, 0, 0,    0, 0, 0, 0, 0, 0, 0, 0,   /* 0x00 .. 0x0F */
    0, 0, 0, 0, 0, 0, 0, 0,    0, 0, 0, 0, 0, 0, 0, 0,   /* 0x10 .. 0x1F */
    0, 0, 0, 0, 0, 0, 0, 0,    0, 0, 0, 0, 0, 0, 0, 0,   /* 0x20 .. 0x2F */
    1, 1, 1, 1, 1, 1, 1, 1,    1, 1, 0, 0, 0, 0, 0, 0,   /* 0x30 .. 0x3F */
    0, 1, 1, 1, 1, 1, 1, 1,    1, 1, 1, 1, 1, 1, 1, 1,   /* 0x40 .. 0x4F */
    1, 1, 1, 1, 1, 1, 1, 1,    1, 1, 1, 0, 0, 0, 0, 1,   /* 0x50 .. 0x5F */
    0, 1, 1, 1, 1, 1, 1, 1,    1, 1, 1, 1, 1, 1, 1, 1,   /* 0x60 .. 0x6F */
    1, 1, 1, 1, 1, 1, 1, 1,    1, 1, 1, 0, 0, 0, 0, 0    /* 0x70 .. 0x7F */
  };

  return (t & 0x80) || aBareword[(int)t];
}

#endif /* SQLITE_ENABLE_FTS5 */

/*
** 2014 Jun 09
**
** The author disclaims copyright to this source code.  In place of
** a legal notice, here is a blessing:
**
**    May you do good and not evil.
**    May you find forgiveness for yourself and forgive others.
**    May you share freely, never taking more than you give.
**
******************************************************************************
**
** This is an SQLite module implementing full-text search.
*/

#ifdef SQLITE_ENABLE_FTS5


#include "fts5Int.h"

#define FTS5_DEFAULT_PAGE_SIZE   1000
#define FTS5_DEFAULT_AUTOMERGE      4
#define FTS5_DEFAULT_CRISISMERGE   16

/* Maximum allowed page size */
#define FTS5_MAX_PAGE_SIZE (128*1024)

static int fts5_iswhitespace(char x){
  return (x==' ');
}

static int fts5_isopenquote(char x){
  return (x=='"' || x=='\'' || x=='[' || x=='`');
}

/*
** Argument pIn points to a character that is part of a nul-terminated 
** string. Return a pointer to the first character following *pIn in 
** the string that is not a white-space character.
*/
static const char *fts5ConfigSkipWhitespace(const char *pIn){
  const char *p = pIn;
  if( p ){
    while( fts5_iswhitespace(*p) ){ p++; }
  }
  return p;
}

/*
** Argument pIn points to a character that is part of a nul-terminated 
** string. Return a pointer to the first character following *pIn in 
** the string that is not a "bareword" character.
*/
static const char *fts5ConfigSkipBareword(const char *pIn){
  const char *p = pIn;
  while ( sqlite3Fts5IsBareword(*p) ) p++;
  if( p==pIn ) p = 0;
  return p;
}

static int fts5_isdigit(char a){
  return (a>='0' && a<='9');
}



static const char *fts5ConfigSkipLiteral(const char *pIn){
  const char *p = pIn;
  switch( *p ){
    case 'n': case 'N':
      if( sqlite3_strnicmp("null", p, 4)==0 ){
        p = &p[4];
      }else{
        p = 0;
      }
      break;

    case 'x': case 'X':
      p++;
      if( *p=='\'' ){
        p++;
        while( (*p>='a' && *p<='f') 
            || (*p>='A' && *p<='F') 
            || (*p>='0' && *p<='9') 
            ){
          p++;
        }
        if( *p=='\'' && 0==((p-pIn)%2) ){
          p++;
        }else{
          p = 0;
        }
      }else{
        p = 0;
      }
      break;

    case '\'':
      p++;
      while( p ){
        if( *p=='\'' ){
          p++;
          if( *p!='\'' ) break;
        }
        p++;
        if( *p==0 ) p = 0;
      }
      break;

    default:
      /* maybe a number */
      if( *p=='+' || *p=='-' ) p++;
      while( fts5_isdigit(*p) ) p++;

      /* At this point, if the literal was an integer, the parse is 
      ** finished. Or, if it is a floating point value, it may continue
      ** with either a decimal point or an 'E' character. */
      if( *p=='.' && fts5_isdigit(p[1]) ){
        p += 2;
        while( fts5_isdigit(*p) ) p++;
      }
      if( p==pIn ) p = 0;

      break;
  }

  return p;
}

/*
** The first character of the string pointed to by argument z is guaranteed
** to be an open-quote character (see function fts5_isopenquote()).
**
** This function searches for the corresponding close-quote character within
** the string and, if found, dequotes the string in place and adds a new
** nul-terminator byte.
**
** If the close-quote is found, the value returned is the byte offset of
** the character immediately following it. Or, if the close-quote is not 
** found, -1 is returned. If -1 is returned, the buffer is left in an 
** undefined state.
*/
static int fts5Dequote(char *z){
  char q;
  int iIn = 1;
  int iOut = 0;
  q = z[0];

  /* Set stack variable q to the close-quote character */
  assert( q=='[' || q=='\'' || q=='"' || q=='`' );
  if( q=='[' ) q = ']';  

  while( ALWAYS(z[iIn]) ){
    if( z[iIn]==q ){
      if( z[iIn+1]!=q ){
        /* Character iIn was the close quote. */
        iIn++;
        break;
      }else{
        /* Character iIn and iIn+1 form an escaped quote character. Skip
        ** the input cursor past both and copy a single quote character 
        ** to the output buffer. */
        iIn += 2;
        z[iOut++] = q;
      }
    }else{
      z[iOut++] = z[iIn++];
    }
  }

  z[iOut] = '\0';
  return iIn;
}

/*
** Convert an SQL-style quoted string into a normal string by removing
** the quote characters.  The conversion is done in-place.  If the
** input does not begin with a quote character, then this routine
** is a no-op.
**
** Examples:
**
**     "abc"   becomes   abc
**     'xyz'   becomes   xyz
**     [pqr]   becomes   pqr
**     `mno`   becomes   mno
*/
void sqlite3Fts5Dequote(char *z){
  char quote;                     /* Quote character (if any ) */

  assert( 0==fts5_iswhitespace(z[0]) );
  quote = z[0];
  if( quote=='[' || quote=='\'' || quote=='"' || quote=='`' ){
    fts5Dequote(z);
  }
}

/*
** Parse a "special" CREATE VIRTUAL TABLE directive and update
** configuration object pConfig as appropriate.
**
** If successful, object pConfig is updated and SQLITE_OK returned. If
** an error occurs, an SQLite error code is returned and an error message
** may be left in *pzErr. It is the responsibility of the caller to
** eventually free any such error message using sqlite3_free().
*/
static int fts5ConfigParseSpecial(
  Fts5Global *pGlobal,
  Fts5Config *pConfig,            /* Configuration object to update */
  const char *zCmd,               /* Special command to parse */
  const char *zArg,               /* Argument to parse */
  char **pzErr                    /* OUT: Error message */
){
  int rc = SQLITE_OK;
  int nCmd = strlen(zCmd);
  if( sqlite3_strnicmp("prefix", zCmd, nCmd)==0 ){
    const int nByte = sizeof(int) * FTS5_MAX_PREFIX_INDEXES;
    const char *p;
    if( pConfig->aPrefix ){
      *pzErr = sqlite3_mprintf("multiple prefix=... directives");
      rc = SQLITE_ERROR;
    }else{
      pConfig->aPrefix = sqlite3Fts5MallocZero(&rc, nByte);
    }
    p = zArg;
    while( rc==SQLITE_OK && p[0] ){
      int nPre = 0;
      while( p[0]==' ' ) p++;
      while( p[0]>='0' && p[0]<='9' && nPre<1000 ){
        nPre = nPre*10 + (p[0] - '0');
        p++;
      }
      while( p[0]==' ' ) p++;
      if( p[0]==',' ){
        p++;
      }else if( p[0] ){
        *pzErr = sqlite3_mprintf("malformed prefix=... directive");
        rc = SQLITE_ERROR;
      }
      if( rc==SQLITE_OK && (nPre==0 || nPre>=1000) ){
        *pzErr = sqlite3_mprintf("prefix length out of range: %d", nPre);
        rc = SQLITE_ERROR;
      }
      pConfig->aPrefix[pConfig->nPrefix] = nPre;
      pConfig->nPrefix++;
    }
    return rc;
  }

  if( sqlite3_strnicmp("tokenize", zCmd, nCmd)==0 ){
    const char *p = (const char*)zArg;
    int nArg = strlen(zArg) + 1;
    char **azArg = sqlite3Fts5MallocZero(&rc, sizeof(char*) * nArg);
    char *pDel = sqlite3Fts5MallocZero(&rc, nArg * 2);
    char *pSpace = pDel;

    if( azArg && pSpace ){
      if( pConfig->pTok ){
        *pzErr = sqlite3_mprintf("multiple tokenize=... directives");
        rc = SQLITE_ERROR;
      }else{
        for(nArg=0; p && *p; nArg++){
          const char *p2 = fts5ConfigSkipWhitespace(p);
          if( *p2=='\'' ){
            p = fts5ConfigSkipLiteral(p2);
          }else{
            p = fts5ConfigSkipBareword(p2);
          }
          if( p ){
            memcpy(pSpace, p2, p-p2);
            azArg[nArg] = pSpace;
            sqlite3Fts5Dequote(pSpace);
            pSpace += (p - p2) + 1;
            p = fts5ConfigSkipWhitespace(p);
          }
        }
        if( p==0 ){
          *pzErr = sqlite3_mprintf("parse error in tokenize directive");
          rc = SQLITE_ERROR;
        }else{
          rc = sqlite3Fts5GetTokenizer(pGlobal, 
              (const char**)azArg, nArg, &pConfig->pTok, &pConfig->pTokApi,
              pzErr
          );
        }
      }
    }

    sqlite3_free(azArg);
    sqlite3_free(pDel);
    return rc;
  }

  if( sqlite3_strnicmp("content", zCmd, nCmd)==0 ){
    if( pConfig->eContent!=FTS5_CONTENT_NORMAL ){
      *pzErr = sqlite3_mprintf("multiple content=... directives");
      rc = SQLITE_ERROR;
    }else{
      if( zArg[0] ){
        pConfig->eContent = FTS5_CONTENT_EXTERNAL;
        pConfig->zContent = sqlite3Fts5Mprintf(&rc, "%Q.%Q", pConfig->zDb,zArg);
      }else{
        pConfig->eContent = FTS5_CONTENT_NONE;
      }
    }
    return rc;
  }

  if( sqlite3_strnicmp("content_rowid", zCmd, nCmd)==0 ){
    if( pConfig->zContentRowid ){
      *pzErr = sqlite3_mprintf("multiple content_rowid=... directives");
      rc = SQLITE_ERROR;
    }else{
      pConfig->zContentRowid = sqlite3Fts5Strndup(&rc, zArg, -1);
    }
    return rc;
  }

  if( sqlite3_strnicmp("columnsize", zCmd, nCmd)==0 ){
    if( (zArg[0]!='0' && zArg[0]!='1') || zArg[1]!='\0' ){
      *pzErr = sqlite3_mprintf("malformed columnsize=... directive");
      rc = SQLITE_ERROR;
    }else{
      pConfig->bColumnsize = (zArg[0]=='1');
    }
    return rc;
  }

  *pzErr = sqlite3_mprintf("unrecognized option: \"%.*s\"", nCmd, zCmd);
  return SQLITE_ERROR;
}

/*
** Allocate an instance of the default tokenizer ("simple") at 
** Fts5Config.pTokenizer. Return SQLITE_OK if successful, or an SQLite error
** code if an error occurs.
*/
static int fts5ConfigDefaultTokenizer(Fts5Global *pGlobal, Fts5Config *pConfig){
  assert( pConfig->pTok==0 && pConfig->pTokApi==0 );
  return sqlite3Fts5GetTokenizer(
      pGlobal, 0, 0, &pConfig->pTok, &pConfig->pTokApi, 0
  );
}

/*
** Gobble up the first bareword or quoted word from the input buffer zIn.
** Return a pointer to the character immediately following the last in
** the gobbled word if successful, or a NULL pointer otherwise (failed
** to find close-quote character).
**
** Before returning, set pzOut to point to a new buffer containing a
** nul-terminated, dequoted copy of the gobbled word. If the word was
** quoted, *pbQuoted is also set to 1 before returning.
**
** If *pRc is other than SQLITE_OK when this function is called, it is
** a no-op (NULL is returned). Otherwise, if an OOM occurs within this
** function, *pRc is set to SQLITE_NOMEM before returning. *pRc is *not*
** set if a parse error (failed to find close quote) occurs.
*/
static const char *fts5ConfigGobbleWord(
  int *pRc,                       /* IN/OUT: Error code */
  const char *zIn,                /* Buffer to gobble string/bareword from */
  char **pzOut,                   /* OUT: malloc'd buffer containing str/bw */
  int *pbQuoted                   /* OUT: Set to true if dequoting required */
){
  const char *zRet = 0;

  int nIn = strlen(zIn);
  char *zOut = sqlite3_malloc(nIn+1);

  assert( *pRc==SQLITE_OK );
  *pbQuoted = 0;
  *pzOut = 0;

  if( zOut==0 ){
    *pRc = SQLITE_NOMEM;
  }else{
    memcpy(zOut, zIn, nIn+1);
    if( fts5_isopenquote(zOut[0]) ){
      int ii = fts5Dequote(zOut);
      zRet = &zIn[ii];
      *pbQuoted = 1;
    }else{
      zRet = fts5ConfigSkipBareword(zIn);
      zOut[zRet-zIn] = '\0';
    }
  }

  if( zRet==0 ){
    sqlite3_free(zOut);
  }else{
    *pzOut = zOut;
  }

  return zRet;
}

static int fts5ConfigParseColumn(
  Fts5Config *p, 
  char *zCol, 
  char *zArg, 
  char **pzErr
){
  int rc = SQLITE_OK;
  if( 0==sqlite3_stricmp(zCol, FTS5_RANK_NAME) 
   || 0==sqlite3_stricmp(zCol, FTS5_ROWID_NAME) 
  ){
    *pzErr = sqlite3_mprintf("reserved fts5 column name: %s", zCol);
    rc = SQLITE_ERROR;
  }else if( zArg ){
    if( 0==sqlite3_stricmp(zArg, "unindexed") ){
      p->abUnindexed[p->nCol] = 1;
    }else{
      *pzErr = sqlite3_mprintf("unrecognized column option: %s", zArg);
      rc = SQLITE_ERROR;
    }
  }

  p->azCol[p->nCol++] = zCol;
  return rc;
}

/*
** Populate the Fts5Config.zContentExprlist string.
*/
static int fts5ConfigMakeExprlist(Fts5Config *p){
  int i;
  int rc = SQLITE_OK;
  Fts5Buffer buf = {0, 0, 0};

  sqlite3Fts5BufferAppendPrintf(&rc, &buf, "T.%Q", p->zContentRowid);
  if( p->eContent!=FTS5_CONTENT_NONE ){
    for(i=0; i<p->nCol; i++){
      if( p->eContent==FTS5_CONTENT_EXTERNAL ){
        sqlite3Fts5BufferAppendPrintf(&rc, &buf, ", T.%Q", p->azCol[i]);
      }else{
        sqlite3Fts5BufferAppendPrintf(&rc, &buf, ", T.c%d", i);
      }
    }
  }

  assert( p->zContentExprlist==0 );
  p->zContentExprlist = (char*)buf.p;
  return rc;
}

/*
** Arguments nArg/azArg contain the string arguments passed to the xCreate
** or xConnect method of the virtual table. This function attempts to 
** allocate an instance of Fts5Config containing the results of parsing
** those arguments.
**
** If successful, SQLITE_OK is returned and *ppOut is set to point to the
** new Fts5Config object. If an error occurs, an SQLite error code is 
** returned, *ppOut is set to NULL and an error message may be left in
** *pzErr. It is the responsibility of the caller to eventually free any 
** such error message using sqlite3_free().
*/
int sqlite3Fts5ConfigParse(
  Fts5Global *pGlobal,
  sqlite3 *db,
  int nArg,                       /* Number of arguments */
  const char **azArg,             /* Array of nArg CREATE VIRTUAL TABLE args */
  Fts5Config **ppOut,             /* OUT: Results of parse */
  char **pzErr                    /* OUT: Error message */
){
  int rc = SQLITE_OK;             /* Return code */
  Fts5Config *pRet;               /* New object to return */
  int i;
  int nByte;

  *ppOut = pRet = (Fts5Config*)sqlite3_malloc(sizeof(Fts5Config));
  if( pRet==0 ) return SQLITE_NOMEM;
  memset(pRet, 0, sizeof(Fts5Config));
  pRet->db = db;
  pRet->iCookie = -1;

  nByte = nArg * (sizeof(char*) + sizeof(u8));
  pRet->azCol = (char**)sqlite3Fts5MallocZero(&rc, nByte);
  pRet->abUnindexed = (u8*)&pRet->azCol[nArg];
  pRet->zDb = sqlite3Fts5Strndup(&rc, azArg[1], -1);
  pRet->zName = sqlite3Fts5Strndup(&rc, azArg[2], -1);
  pRet->bColumnsize = 1;
  if( rc==SQLITE_OK && sqlite3_stricmp(pRet->zName, FTS5_RANK_NAME)==0 ){
    *pzErr = sqlite3_mprintf("reserved fts5 table name: %s", pRet->zName);
    rc = SQLITE_ERROR;
  }

  for(i=3; rc==SQLITE_OK && i<nArg; i++){
    const char *zOrig = azArg[i];
    const char *z;
    char *zOne = 0;
    char *zTwo = 0;
    int bOption = 0;
    int bMustBeCol = 0;

    z = fts5ConfigGobbleWord(&rc, zOrig, &zOne, &bMustBeCol);
    z = fts5ConfigSkipWhitespace(z);
    if( z && *z=='=' ){
      bOption = 1;
      z++;
      if( bMustBeCol ) z = 0;
    }
    z = fts5ConfigSkipWhitespace(z);
    if( z && z[0] ){
      int bDummy;
      z = fts5ConfigGobbleWord(&rc, z, &zTwo, &bDummy);
      if( z && z[0] ) z = 0;
    }

    if( rc==SQLITE_OK ){
      if( z==0 ){
        *pzErr = sqlite3_mprintf("parse error in \"%s\"", zOrig);
        rc = SQLITE_ERROR;
      }else{
        if( bOption ){
          rc = fts5ConfigParseSpecial(pGlobal, pRet, zOne, zTwo?zTwo:"", pzErr);
        }else{
          rc = fts5ConfigParseColumn(pRet, zOne, zTwo, pzErr);
          zOne = 0;
        }
      }
    }

    sqlite3_free(zOne);
    sqlite3_free(zTwo);
  }

  /* If a tokenizer= option was successfully parsed, the tokenizer has
  ** already been allocated. Otherwise, allocate an instance of the default
  ** tokenizer (unicode61) now.  */
  if( rc==SQLITE_OK && pRet->pTok==0 ){
    rc = fts5ConfigDefaultTokenizer(pGlobal, pRet);
  }

  /* If no zContent option was specified, fill in the default values. */
  if( rc==SQLITE_OK && pRet->zContent==0 ){
    const char *zTail = 0;
    assert( pRet->eContent==FTS5_CONTENT_NORMAL 
         || pRet->eContent==FTS5_CONTENT_NONE 
    );
    if( pRet->eContent==FTS5_CONTENT_NORMAL ){
      zTail = "content";
    }else if( pRet->bColumnsize ){
      zTail = "docsize";
    }

    if( zTail ){
      pRet->zContent = sqlite3Fts5Mprintf(
          &rc, "%Q.'%q_%s'", pRet->zDb, pRet->zName, zTail
      );
    }
  }

  if( rc==SQLITE_OK && pRet->zContentRowid==0 ){
    pRet->zContentRowid = sqlite3Fts5Strndup(&rc, "rowid", -1);
  }

  /* Formulate the zContentExprlist text */
  if( rc==SQLITE_OK ){
    rc = fts5ConfigMakeExprlist(pRet);
  }

  if( rc!=SQLITE_OK ){
    sqlite3Fts5ConfigFree(pRet);
    *ppOut = 0;
  }
  return rc;
}

/*
** Free the configuration object passed as the only argument.
*/
void sqlite3Fts5ConfigFree(Fts5Config *pConfig){
  if( pConfig ){
    int i;
    if( pConfig->pTok ){
      pConfig->pTokApi->xDelete(pConfig->pTok);
    }
    sqlite3_free(pConfig->zDb);
    sqlite3_free(pConfig->zName);
    for(i=0; i<pConfig->nCol; i++){
      sqlite3_free(pConfig->azCol[i]);
    }
    sqlite3_free(pConfig->azCol);
    sqlite3_free(pConfig->aPrefix);
    sqlite3_free(pConfig->zRank);
    sqlite3_free(pConfig->zRankArgs);
    sqlite3_free(pConfig->zContent);
    sqlite3_free(pConfig->zContentRowid);
    sqlite3_free(pConfig->zContentExprlist);
    sqlite3_free(pConfig);
  }
}

/*
** Call sqlite3_declare_vtab() based on the contents of the configuration
** object passed as the only argument. Return SQLITE_OK if successful, or
** an SQLite error code if an error occurs.
*/
int sqlite3Fts5ConfigDeclareVtab(Fts5Config *pConfig){
  int i;
  int rc = SQLITE_OK;
  char *zSql;

  zSql = sqlite3Fts5Mprintf(&rc, "CREATE TABLE x(");
  for(i=0; zSql && i<pConfig->nCol; i++){
    const char *zSep = (i==0?"":", ");
    zSql = sqlite3Fts5Mprintf(&rc, "%z%s%Q", zSql, zSep, pConfig->azCol[i]);
  }
  zSql = sqlite3Fts5Mprintf(&rc, "%z, %Q HIDDEN, %s HIDDEN)", 
      zSql, pConfig->zName, FTS5_RANK_NAME
  );

  assert( zSql || rc==SQLITE_NOMEM );
  if( zSql ){
    rc = sqlite3_declare_vtab(pConfig->db, zSql);
    sqlite3_free(zSql);
  }
  
  return rc;
}

/*
** Tokenize the text passed via the second and third arguments.
**
** The callback is invoked once for each token in the input text. The
** arguments passed to it are, in order:
**
**     void *pCtx          // Copy of 4th argument to sqlite3Fts5Tokenize()
**     const char *pToken  // Pointer to buffer containing token
**     int nToken          // Size of token in bytes
**     int iStart          // Byte offset of start of token within input text
**     int iEnd            // Byte offset of end of token within input text
**     int iPos            // Position of token in input (first token is 0)
**
** If the callback returns a non-zero value the tokenization is abandoned
** and no further callbacks are issued. 
**
** This function returns SQLITE_OK if successful or an SQLite error code
** if an error occurs. If the tokenization was abandoned early because
** the callback returned SQLITE_DONE, this is not an error and this function
** still returns SQLITE_OK. Or, if the tokenization was abandoned early
** because the callback returned another non-zero value, it is assumed
** to be an SQLite error code and returned to the caller.
*/
int sqlite3Fts5Tokenize(
  Fts5Config *pConfig,            /* FTS5 Configuration object */
  const char *pText, int nText,   /* Text to tokenize */
  void *pCtx,                     /* Context passed to xToken() */
  int (*xToken)(void*, const char*, int, int, int)    /* Callback */
){
  if( pText==0 ) return SQLITE_OK;
  return pConfig->pTokApi->xTokenize(pConfig->pTok, pCtx, pText, nText, xToken);
}

/*
** Argument pIn points to the first character in what is expected to be
** a comma-separated list of SQL literals followed by a ')' character.
** If it actually is this, return a pointer to the ')'. Otherwise, return
** NULL to indicate a parse error.
*/
static const char *fts5ConfigSkipArgs(const char *pIn){
  const char *p = pIn;
  
  while( 1 ){
    p = fts5ConfigSkipWhitespace(p);
    p = fts5ConfigSkipLiteral(p);
    p = fts5ConfigSkipWhitespace(p);
    if( p==0 || *p==')' ) break;
    if( *p!=',' ){
      p = 0;
      break;
    }
    p++;
  }

  return p;
}

/*
** Parameter zIn contains a rank() function specification. The format of 
** this is:
**
**   + Bareword (function name)
**   + Open parenthesis - "("
**   + Zero or more SQL literals in a comma separated list
**   + Close parenthesis - ")"
*/
int sqlite3Fts5ConfigParseRank(
  const char *zIn,                /* Input string */
  char **pzRank,                  /* OUT: Rank function name */
  char **pzRankArgs               /* OUT: Rank function arguments */
){
  const char *p = zIn;
  const char *pRank;
  char *zRank = 0;
  char *zRankArgs = 0;
  int rc = SQLITE_OK;

  *pzRank = 0;
  *pzRankArgs = 0;

  p = fts5ConfigSkipWhitespace(p);
  pRank = p;
  p = fts5ConfigSkipBareword(p);

  if( p ){
    zRank = sqlite3Fts5MallocZero(&rc, 1 + p - pRank);
    if( zRank ) memcpy(zRank, pRank, p-pRank);
  }else{
    rc = SQLITE_ERROR;
  }

  if( rc==SQLITE_OK ){
    p = fts5ConfigSkipWhitespace(p);
    if( *p!='(' ) rc = SQLITE_ERROR;
    p++;
  }
  if( rc==SQLITE_OK ){
    const char *pArgs; 
    p = fts5ConfigSkipWhitespace(p);
    pArgs = p;
    if( *p!=')' ){
      p = fts5ConfigSkipArgs(p);
      if( p==0 ){
        rc = SQLITE_ERROR;
      }else{
        zRankArgs = sqlite3Fts5MallocZero(&rc, 1 + p - pArgs);
        if( zRankArgs ) memcpy(zRankArgs, pArgs, p-pArgs);
      }
    }
  }

  if( rc!=SQLITE_OK ){
    sqlite3_free(zRank);
    assert( zRankArgs==0 );
  }else{
    *pzRank = zRank;
    *pzRankArgs = zRankArgs;
  }
  return rc;
}

int sqlite3Fts5ConfigSetValue(
  Fts5Config *pConfig, 
  const char *zKey, 
  sqlite3_value *pVal,
  int *pbBadkey
){
  int rc = SQLITE_OK;

  if( 0==sqlite3_stricmp(zKey, "pgsz") ){
    int pgsz = 0;
    if( SQLITE_INTEGER==sqlite3_value_numeric_type(pVal) ){
      pgsz = sqlite3_value_int(pVal);
    }
    if( pgsz<=0 || pgsz>FTS5_MAX_PAGE_SIZE ){
      *pbBadkey = 1;
    }else{
      pConfig->pgsz = pgsz;
    }
  }

  else if( 0==sqlite3_stricmp(zKey, "automerge") ){
    int nAutomerge = -1;
    if( SQLITE_INTEGER==sqlite3_value_numeric_type(pVal) ){
      nAutomerge = sqlite3_value_int(pVal);
    }
    if( nAutomerge<0 || nAutomerge>64 ){
      *pbBadkey = 1;
    }else{
      if( nAutomerge==1 ) nAutomerge = FTS5_DEFAULT_AUTOMERGE;
      pConfig->nAutomerge = nAutomerge;
    }
  }

  else if( 0==sqlite3_stricmp(zKey, "crisismerge") ){
    int nCrisisMerge = -1;
    if( SQLITE_INTEGER==sqlite3_value_numeric_type(pVal) ){
      nCrisisMerge = sqlite3_value_int(pVal);
    }
    if( nCrisisMerge<0 ){
      *pbBadkey = 1;
    }else{
      if( nCrisisMerge<=1 ) nCrisisMerge = FTS5_DEFAULT_CRISISMERGE;
      pConfig->nCrisisMerge = nCrisisMerge;
    }
  }

  else if( 0==sqlite3_stricmp(zKey, "rank") ){
    const char *zIn = (const char*)sqlite3_value_text(pVal);
    char *zRank;
    char *zRankArgs;
    rc = sqlite3Fts5ConfigParseRank(zIn, &zRank, &zRankArgs);
    if( rc==SQLITE_OK ){
      sqlite3_free(pConfig->zRank);
      sqlite3_free(pConfig->zRankArgs);
      pConfig->zRank = zRank;
      pConfig->zRankArgs = zRankArgs;
    }else if( rc==SQLITE_ERROR ){
      rc = SQLITE_OK;
      *pbBadkey = 1;
    }
  }else{
    *pbBadkey = 1;
  }
  return rc;
}

/*
** Load the contents of the %_config table into memory.
*/
int sqlite3Fts5ConfigLoad(Fts5Config *pConfig, int iCookie){
  const char *zSelect = "SELECT k, v FROM %Q.'%q_config'";
  char *zSql;
  sqlite3_stmt *p = 0;
  int rc = SQLITE_OK;
  int iVersion = 0;

  /* Set default values */
  pConfig->pgsz = FTS5_DEFAULT_PAGE_SIZE;
  pConfig->nAutomerge = FTS5_DEFAULT_AUTOMERGE;
  pConfig->nCrisisMerge = FTS5_DEFAULT_CRISISMERGE;

  zSql = sqlite3Fts5Mprintf(&rc, zSelect, pConfig->zDb, pConfig->zName);
  if( zSql ){
    rc = sqlite3_prepare_v2(pConfig->db, zSql, -1, &p, 0);
    sqlite3_free(zSql);
  }

  assert( rc==SQLITE_OK || p==0 );
  if( rc==SQLITE_OK ){
    while( SQLITE_ROW==sqlite3_step(p) ){
      const char *zK = (const char*)sqlite3_column_text(p, 0);
      sqlite3_value *pVal = sqlite3_column_value(p, 1);
      if( 0==sqlite3_stricmp(zK, "version") ){
        iVersion = sqlite3_value_int(pVal);
      }else{
        int bDummy = 0;
        sqlite3Fts5ConfigSetValue(pConfig, zK, pVal, &bDummy);
      }
    }
    rc = sqlite3_finalize(p);
  }
  
  if( rc==SQLITE_OK && iVersion!=FTS5_CURRENT_VERSION ){
    rc = SQLITE_ERROR;
    if( pConfig->pzErrmsg ){
      assert( 0==*pConfig->pzErrmsg );
      *pConfig->pzErrmsg = sqlite3_mprintf(
          "invalid fts5 file format (found %d, expected %d) - run 'rebuild'",
          iVersion, FTS5_CURRENT_VERSION
      );
    }
  }

  if( rc==SQLITE_OK ){
    pConfig->iCookie = iCookie;
  }
  return rc;
}

#endif /* SQLITE_ENABLE_FTS5 */

/*
** 2014 May 31
**
** The author disclaims copyright to this source code.  In place of
** a legal notice, here is a blessing:
**
**    May you do good and not evil.
**    May you find forgiveness for yourself and forgive others.
**    May you share freely, never taking more than you give.
**
******************************************************************************
**
*/

#ifdef SQLITE_ENABLE_FTS5


#include "fts5Int.h"
#include "fts5parse.h"

/*
** All token types in the generated fts5parse.h file are greater than 0.
*/
#define FTS5_EOF 0

typedef struct Fts5ExprTerm Fts5ExprTerm;

/*
** Functions generated by lemon from fts5parse.y.
*/
void *sqlite3Fts5ParserAlloc(void *(*mallocProc)(u64));
void sqlite3Fts5ParserFree(void*, void (*freeProc)(void*));
void sqlite3Fts5Parser(void*, int, Fts5Token, Fts5Parse*);

struct Fts5Expr {
  Fts5Index *pIndex;
  Fts5ExprNode *pRoot;
  int bDesc;                      /* Iterate in descending docid order */
  int nPhrase;                    /* Number of phrases in expression */
  Fts5ExprPhrase **apExprPhrase;  /* Pointers to phrase objects */
};

/*
** eType:
**   Expression node type. Always one of:
**
**       FTS5_AND                 (nChild, apChild valid)
**       FTS5_OR                  (nChild, apChild valid)
**       FTS5_NOT                 (nChild, apChild valid)
**       FTS5_STRING              (pNear valid)
**       FTS5_TERM                (pNear valid)
*/
struct Fts5ExprNode {
  int eType;                      /* Node type */
  int bEof;                       /* True at EOF */
  int bNomatch;                   /* True if entry is not a match */

  i64 iRowid;                     /* Current rowid */
  Fts5ExprNearset *pNear;         /* For FTS5_STRING - cluster of phrases */

  /* Child nodes. For a NOT node, this array always contains 2 entries. For 
  ** AND or OR nodes, it contains 2 or more entries.  */
  int nChild;                     /* Number of child nodes */
  Fts5ExprNode *apChild[0];       /* Array of child nodes */
};

#define Fts5NodeIsString(p) ((p)->eType==FTS5_TERM || (p)->eType==FTS5_STRING)

/*
** An instance of the following structure represents a single search term
** or term prefix.
*/
struct Fts5ExprTerm {
  int bPrefix;                    /* True for a prefix term */
  char *zTerm;                    /* nul-terminated term */
  Fts5IndexIter *pIter;           /* Iterator for this term */
};

/*
** A phrase. One or more terms that must appear in a contiguous sequence
** within a document for it to match.
*/
struct Fts5ExprPhrase {
  Fts5ExprNode *pNode;            /* FTS5_STRING node this phrase is part of */
  Fts5Buffer poslist;             /* Current position list */
  int nTerm;                      /* Number of entries in aTerm[] */
  Fts5ExprTerm aTerm[0];          /* Terms that make up this phrase */
};

/*
** If a NEAR() clump may only match a specific set of columns, then
** Fts5ExprNearset.pColset points to an object of the following type.
** Each entry in the aiCol[] array
*/
struct Fts5ExprColset {
  int nCol;
  int aiCol[1];
};

/*
** One or more phrases that must appear within a certain token distance of
** each other within each matching document.
*/
struct Fts5ExprNearset {
  int nNear;                      /* NEAR parameter */
  Fts5ExprColset *pColset;        /* Columns to search (NULL -> all columns) */
  int nPhrase;                    /* Number of entries in aPhrase[] array */
  Fts5ExprPhrase *apPhrase[0];    /* Array of phrase pointers */
};


/*
** Parse context.
*/
struct Fts5Parse {
  Fts5Config *pConfig;
  char *zErr;
  int rc;
  int nPhrase;                    /* Size of apPhrase array */
  Fts5ExprPhrase **apPhrase;      /* Array of all phrases */
  Fts5ExprNode *pExpr;            /* Result of a successful parse */
};

void sqlite3Fts5ParseError(Fts5Parse *pParse, const char *zFmt, ...){
  va_list ap;
  va_start(ap, zFmt);
  if( pParse->rc==SQLITE_OK ){
    pParse->zErr = sqlite3_vmprintf(zFmt, ap);
    pParse->rc = SQLITE_ERROR;
  }
  va_end(ap);
}

static int fts5ExprIsspace(char t){
  return t==' ' || t=='\t' || t=='\n' || t=='\r';
}

/*
** Read the first token from the nul-terminated string at *pz.
*/
static int fts5ExprGetToken(
  Fts5Parse *pParse, 
  const char **pz,                /* IN/OUT: Pointer into buffer */
  Fts5Token *pToken
){
  const char *z = *pz;
  int tok;

  /* Skip past any whitespace */
  while( fts5ExprIsspace(*z) ) z++;

  pToken->p = z;
  pToken->n = 1;
  switch( *z ){
    case '(':  tok = FTS5_LP;    break;
    case ')':  tok = FTS5_RP;    break;
    case '{':  tok = FTS5_LCP;   break;
    case '}':  tok = FTS5_RCP;   break;
    case ':':  tok = FTS5_COLON; break;
    case ',':  tok = FTS5_COMMA; break;
    case '+':  tok = FTS5_PLUS;  break;
    case '*':  tok = FTS5_STAR;  break;
    case '\0': tok = FTS5_EOF;   break;

    case '"': {
      const char *z2;
      tok = FTS5_STRING;

      for(z2=&z[1]; 1; z2++){
        if( z2[0]=='"' ){
          z2++;
          if( z2[0]!='"' ) break;
        }
        if( z2[0]=='\0' ){
          sqlite3Fts5ParseError(pParse, "unterminated string");
          return FTS5_EOF;
        }
      }
      pToken->n = (z2 - z);
      break;
    }

    default: {
      const char *z2;
      tok = FTS5_STRING;
      for(z2=&z[1]; sqlite3Fts5IsBareword(*z2); z2++);
      pToken->n = (z2 - z);
      if( pToken->n==2 && memcmp(pToken->p, "OR", 2)==0 )  tok = FTS5_OR;
      if( pToken->n==3 && memcmp(pToken->p, "NOT", 3)==0 ) tok = FTS5_NOT;
      if( pToken->n==3 && memcmp(pToken->p, "AND", 3)==0 ) tok = FTS5_AND;
      break;
    }
  }

  *pz = &pToken->p[pToken->n];
  return tok;
}

static void *fts5ParseAlloc(u64 t){ return sqlite3_malloc((int)t); }
static void fts5ParseFree(void *p){ sqlite3_free(p); }

int sqlite3Fts5ExprNew(
  Fts5Config *pConfig,            /* FTS5 Configuration */
  const char *zExpr,              /* Expression text */
  Fts5Expr **ppNew, 
  char **pzErr
){
  Fts5Parse sParse;
  Fts5Token token;
  const char *z = zExpr;
  int t;                          /* Next token type */
  void *pEngine;
  Fts5Expr *pNew;

  *ppNew = 0;
  *pzErr = 0;
  memset(&sParse, 0, sizeof(sParse));
  pEngine = sqlite3Fts5ParserAlloc(fts5ParseAlloc);
  if( pEngine==0 ){ return SQLITE_NOMEM; }
  sParse.pConfig = pConfig;

  do {
    t = fts5ExprGetToken(&sParse, &z, &token);
    sqlite3Fts5Parser(pEngine, t, token, &sParse);
  }while( sParse.rc==SQLITE_OK && t!=FTS5_EOF );
  sqlite3Fts5ParserFree(pEngine, fts5ParseFree);

  assert( sParse.rc!=SQLITE_OK || sParse.zErr==0 );
  if( sParse.rc==SQLITE_OK ){
    *ppNew = pNew = sqlite3_malloc(sizeof(Fts5Expr));
    if( pNew==0 ){
      sParse.rc = SQLITE_NOMEM;
      sqlite3Fts5ParseNodeFree(sParse.pExpr);
    }else{
      pNew->pRoot = sParse.pExpr;
      pNew->pIndex = 0;
      pNew->apExprPhrase = sParse.apPhrase;
      pNew->nPhrase = sParse.nPhrase;
      sParse.apPhrase = 0;
    }
  }

  sqlite3_free(sParse.apPhrase);
  *pzErr = sParse.zErr;
  return sParse.rc;
}

/*
** Create a new FTS5 expression by cloning phrase iPhrase of the
** expression passed as the second argument.
*/
int sqlite3Fts5ExprPhraseExpr(
  Fts5Config *pConfig,
  Fts5Expr *pExpr, 
  int iPhrase, 
  Fts5Expr **ppNew
){
  int rc = SQLITE_OK;             /* Return code */
  Fts5ExprPhrase *pOrig;          /* The phrase extracted from pExpr */
  Fts5ExprPhrase *pCopy;          /* Copy of pOrig */
  Fts5Expr *pNew = 0;             /* Expression to return via *ppNew */

  pOrig = pExpr->apExprPhrase[iPhrase];
  pCopy = (Fts5ExprPhrase*)sqlite3Fts5MallocZero(&rc, 
      sizeof(Fts5ExprPhrase) + sizeof(Fts5ExprTerm) * pOrig->nTerm
  );
  if( pCopy ){
    int i;                          /* Used to iterate through phrase terms */
    Fts5ExprPhrase **apPhrase;
    Fts5ExprNode *pNode;
    Fts5ExprNearset *pNear;

    pNew = (Fts5Expr*)sqlite3Fts5MallocZero(&rc, sizeof(Fts5Expr));
    apPhrase = (Fts5ExprPhrase**)sqlite3Fts5MallocZero(&rc, 
        sizeof(Fts5ExprPhrase*)
    );
    pNode = (Fts5ExprNode*)sqlite3Fts5MallocZero(&rc, sizeof(Fts5ExprNode));
    pNear = (Fts5ExprNearset*)sqlite3Fts5MallocZero(&rc, 
        sizeof(Fts5ExprNearset) + sizeof(Fts5ExprPhrase*)
    );

    for(i=0; i<pOrig->nTerm; i++){
      pCopy->aTerm[i].zTerm = sqlite3Fts5Strndup(&rc, pOrig->aTerm[i].zTerm,-1);
      pCopy->aTerm[i].bPrefix = pOrig->aTerm[i].bPrefix;
    }

    if( rc==SQLITE_OK ){
      /* All the allocations succeeded. Put the expression object together. */
      pNew->pIndex = pExpr->pIndex;
      pNew->pRoot = pNode;
      pNew->nPhrase = 1;
      pNew->apExprPhrase = apPhrase;
      pNew->apExprPhrase[0] = pCopy;

      pNode->eType = (pOrig->nTerm==1 ? FTS5_TERM : FTS5_STRING);
      pNode->pNear = pNear;

      pNear->nPhrase = 1;
      pNear->apPhrase[0] = pCopy;

      pCopy->nTerm = pOrig->nTerm;
      pCopy->pNode = pNode;
    }else{
      /* At least one allocation failed. Free them all. */
      for(i=0; i<pOrig->nTerm; i++){
        sqlite3_free(pCopy->aTerm[i].zTerm);
      }
      sqlite3_free(pCopy);
      sqlite3_free(pNear);
      sqlite3_free(pNode);
      sqlite3_free(apPhrase);
      sqlite3_free(pNew);
      pNew = 0;
    }
  }

  *ppNew = pNew;
  return rc;
}

/*
** Free the expression node object passed as the only argument.
*/
void sqlite3Fts5ParseNodeFree(Fts5ExprNode *p){
  if( p ){
    int i;
    for(i=0; i<p->nChild; i++){
      sqlite3Fts5ParseNodeFree(p->apChild[i]);
    }
    sqlite3Fts5ParseNearsetFree(p->pNear);
    sqlite3_free(p);
  }
}

/*
** Free the expression object passed as the only argument.
*/
void sqlite3Fts5ExprFree(Fts5Expr *p){
  if( p ){
    sqlite3Fts5ParseNodeFree(p->pRoot);
    sqlite3_free(p->apExprPhrase);
    sqlite3_free(p);
  }
}

static int fts5ExprColsetTest(Fts5ExprColset *pColset, int iCol){
  int i;
  for(i=0; i<pColset->nCol; i++){
    if( pColset->aiCol[i]==iCol ) return 1;
  }
  return 0;
}

/*
** All individual term iterators in pPhrase are guaranteed to be valid and
** pointing to the same rowid when this function is called. This function 
** checks if the current rowid really is a match, and if so populates
** the pPhrase->poslist buffer accordingly. Output parameter *pbMatch
** is set to true if this is really a match, or false otherwise.
**
** SQLITE_OK is returned if an error occurs, or an SQLite error code 
** otherwise. It is not considered an error code if the current rowid is 
** not a match.
*/
static int fts5ExprPhraseIsMatch(
  Fts5Expr *pExpr,                /* Expression pPhrase belongs to */
  Fts5ExprColset *pColset,        /* Restrict matches to these columns */
  Fts5ExprPhrase *pPhrase,        /* Phrase object to initialize */
  int *pbMatch                    /* OUT: Set to true if really a match */
){
  Fts5PoslistWriter writer = {0};
  Fts5PoslistReader aStatic[4];
  Fts5PoslistReader *aIter = aStatic;
  int i;
  int rc = SQLITE_OK;
  int iCol = -1;
  
  if( pColset && pColset->nCol==1 ){
    iCol = pColset->aiCol[0];
    pColset = 0;
  }

  fts5BufferZero(&pPhrase->poslist);

  /* If the aStatic[] array is not large enough, allocate a large array
  ** using sqlite3_malloc(). This approach could be improved upon. */
  if( pPhrase->nTerm>(sizeof(aStatic) / sizeof(aStatic[0])) ){
    int nByte = sizeof(Fts5PoslistReader) * pPhrase->nTerm;
    aIter = (Fts5PoslistReader*)sqlite3_malloc(nByte);
    if( !aIter ) return SQLITE_NOMEM;
  }

  /* Initialize a term iterator for each term in the phrase */
  for(i=0; i<pPhrase->nTerm; i++){
    i64 dummy;
    int n;
    const u8 *a;
    rc = sqlite3Fts5IterPoslist(pPhrase->aTerm[i].pIter, &a, &n, &dummy);
    if( rc || sqlite3Fts5PoslistReaderInit(iCol, a, n, &aIter[i]) ){
      goto ismatch_out;
    }
  }

  while( 1 ){
    int bMatch;
    i64 iPos = aIter[0].iPos;
    do {
      bMatch = 1;
      for(i=0; i<pPhrase->nTerm; i++){
        Fts5PoslistReader *pPos = &aIter[i];
        i64 iAdj = iPos + i;
        if( pPos->iPos!=iAdj ){
          bMatch = 0;
          while( pPos->iPos<iAdj ){
            if( sqlite3Fts5PoslistReaderNext(pPos) ) goto ismatch_out;
          }
          if( pPos->iPos>iAdj ) iPos = pPos->iPos-i;
        }
      }
    }while( bMatch==0 );

    if( pColset==0 || fts5ExprColsetTest(pColset, FTS5_POS2COLUMN(iPos)) ){
      /* Append position iPos to the output */
      rc = sqlite3Fts5PoslistWriterAppend(&pPhrase->poslist, &writer, iPos);
      if( rc!=SQLITE_OK ) goto ismatch_out;
    }

    for(i=0; i<pPhrase->nTerm; i++){
      if( sqlite3Fts5PoslistReaderNext(&aIter[i]) ) goto ismatch_out;
    }
  }

 ismatch_out:
  *pbMatch = (pPhrase->poslist.n>0);
  if( aIter!=aStatic ) sqlite3_free(aIter);
  return rc;
}

typedef struct Fts5LookaheadReader Fts5LookaheadReader;
struct Fts5LookaheadReader {
  const u8 *a;                    /* Buffer containing position list */
  int n;                          /* Size of buffer a[] in bytes */
  int i;                          /* Current offset in position list */
  i64 iPos;                       /* Current position */
  i64 iLookahead;                 /* Next position */
};

#define FTS5_LOOKAHEAD_EOF (((i64)1) << 62)

static int fts5LookaheadReaderNext(Fts5LookaheadReader *p){
  p->iPos = p->iLookahead;
  if( sqlite3Fts5PoslistNext64(p->a, p->n, &p->i, &p->iLookahead) ){
    p->iLookahead = FTS5_LOOKAHEAD_EOF;
  }
  return (p->iPos==FTS5_LOOKAHEAD_EOF);
}

static int fts5LookaheadReaderInit(
  const u8 *a, int n,             /* Buffer to read position list from */
  Fts5LookaheadReader *p          /* Iterator object to initialize */
){
  memset(p, 0, sizeof(Fts5LookaheadReader));
  p->a = a;
  p->n = n;
  fts5LookaheadReaderNext(p);
  return fts5LookaheadReaderNext(p);
}

#if 0
static int fts5LookaheadReaderEof(Fts5LookaheadReader *p){
  return (p->iPos==FTS5_LOOKAHEAD_EOF);
}
#endif

typedef struct Fts5NearTrimmer Fts5NearTrimmer;
struct Fts5NearTrimmer {
  Fts5LookaheadReader reader;     /* Input iterator */
  Fts5PoslistWriter writer;       /* Writer context */
  Fts5Buffer *pOut;               /* Output poslist */
};

/*
** The near-set object passed as the first argument contains more than
** one phrase. All phrases currently point to the same row. The
** Fts5ExprPhrase.poslist buffers are populated accordingly. This function
** tests if the current row contains instances of each phrase sufficiently
** close together to meet the NEAR constraint. Non-zero is returned if it
** does, or zero otherwise.
**
** If in/out parameter (*pRc) is set to other than SQLITE_OK when this
** function is called, it is a no-op. Or, if an error (e.g. SQLITE_NOMEM)
** occurs within this function (*pRc) is set accordingly before returning.
** The return value is undefined in both these cases.
** 
** If no error occurs and non-zero (a match) is returned, the position-list
** of each phrase object is edited to contain only those entries that
** meet the constraint before returning.
*/
static int fts5ExprNearIsMatch(int *pRc, Fts5ExprNearset *pNear){
  Fts5NearTrimmer aStatic[4];
  Fts5NearTrimmer *a = aStatic;
  Fts5ExprPhrase **apPhrase = pNear->apPhrase;

  int i;
  int rc = *pRc;
  int bMatch;

  assert( pNear->nPhrase>1 );

  /* If the aStatic[] array is not large enough, allocate a large array
  ** using sqlite3_malloc(). This approach could be improved upon. */
  if( pNear->nPhrase>(sizeof(aStatic) / sizeof(aStatic[0])) ){
    int nByte = sizeof(Fts5NearTrimmer) * pNear->nPhrase;
    a = (Fts5NearTrimmer*)sqlite3Fts5MallocZero(&rc, nByte);
  }else{
    memset(aStatic, 0, sizeof(aStatic));
  }
  if( rc!=SQLITE_OK ){
    *pRc = rc;
    return 0;
  }

  /* Initialize a lookahead iterator for each phrase. After passing the
  ** buffer and buffer size to the lookaside-reader init function, zero
  ** the phrase poslist buffer. The new poslist for the phrase (containing
  ** the same entries as the original with some entries removed on account 
  ** of the NEAR constraint) is written over the original even as it is
  ** being read. This is safe as the entries for the new poslist are a
  ** subset of the old, so it is not possible for data yet to be read to
  ** be overwritten.  */
  for(i=0; i<pNear->nPhrase; i++){
    Fts5Buffer *pPoslist = &apPhrase[i]->poslist;
    fts5LookaheadReaderInit(pPoslist->p, pPoslist->n, &a[i].reader);
    pPoslist->n = 0;
    a[i].pOut = pPoslist;
  }

  while( 1 ){
    int iAdv;
    i64 iMin;
    i64 iMax;

    /* This block advances the phrase iterators until they point to a set of
    ** entries that together comprise a match.  */
    iMax = a[0].reader.iPos;
    do {
      bMatch = 1;
      for(i=0; i<pNear->nPhrase; i++){
        Fts5LookaheadReader *pPos = &a[i].reader;
        iMin = iMax - pNear->apPhrase[i]->nTerm - pNear->nNear;
        if( pPos->iPos<iMin || pPos->iPos>iMax ){
          bMatch = 0;
          while( pPos->iPos<iMin ){
            if( fts5LookaheadReaderNext(pPos) ) goto ismatch_out;
          }
          if( pPos->iPos>iMax ) iMax = pPos->iPos;
        }
      }
    }while( bMatch==0 );

    /* Add an entry to each output position list */
    for(i=0; i<pNear->nPhrase; i++){
      i64 iPos = a[i].reader.iPos;
      Fts5PoslistWriter *pWriter = &a[i].writer;
      if( a[i].pOut->n==0 || iPos!=pWriter->iPrev ){
        sqlite3Fts5PoslistWriterAppend(a[i].pOut, pWriter, iPos);
      }
    }

    iAdv = 0;
    iMin = a[0].reader.iLookahead;
    for(i=0; i<pNear->nPhrase; i++){
      if( a[i].reader.iLookahead < iMin ){
        iMin = a[i].reader.iLookahead;
        iAdv = i;
      }
    }
    if( fts5LookaheadReaderNext(&a[iAdv].reader) ) goto ismatch_out;
  }

  ismatch_out: {
    int bRet = a[0].pOut->n>0;
    *pRc = rc;
    if( a!=aStatic ) sqlite3_free(a);
    return bRet;
  }
}

/*
** Advance the first term iterator in the first phrase of pNear. Set output
** variable *pbEof to true if it reaches EOF or if an error occurs.
**
** Return SQLITE_OK if successful, or an SQLite error code if an error
** occurs.
*/
static int fts5ExprNearAdvanceFirst(
  Fts5Expr *pExpr,                /* Expression pPhrase belongs to */
  Fts5ExprNode *pNode,            /* FTS5_STRING or FTS5_TERM node */
  int bFromValid,
  i64 iFrom 
){
  Fts5IndexIter *pIter = pNode->pNear->apPhrase[0]->aTerm[0].pIter;
  int rc;

  assert( Fts5NodeIsString(pNode) );
  if( bFromValid ){
    rc = sqlite3Fts5IterNextFrom(pIter, iFrom);
  }else{
    rc = sqlite3Fts5IterNext(pIter);
  }

  pNode->bEof = (rc || sqlite3Fts5IterEof(pIter));
  return rc;
}

/*
** Advance iterator pIter until it points to a value equal to or laster
** than the initial value of *piLast. If this means the iterator points
** to a value laster than *piLast, update *piLast to the new lastest value.
**
** If the iterator reaches EOF, set *pbEof to true before returning. If
** an error occurs, set *pRc to an error code. If either *pbEof or *pRc
** are set, return a non-zero value. Otherwise, return zero.
*/
static int fts5ExprAdvanceto(
  Fts5IndexIter *pIter,           /* Iterator to advance */
  int bDesc,                      /* True if iterator is "rowid DESC" */
  i64 *piLast,                    /* IN/OUT: Lastest rowid seen so far */
  int *pRc,                       /* OUT: Error code */
  int *pbEof                      /* OUT: Set to true if EOF */
){
  i64 iLast = *piLast;
  i64 iRowid;

  iRowid = sqlite3Fts5IterRowid(pIter);
  if( (bDesc==0 && iLast>iRowid) || (bDesc && iLast<iRowid) ){
    int rc = sqlite3Fts5IterNextFrom(pIter, iLast);
    if( rc || sqlite3Fts5IterEof(pIter) ){
      *pRc = rc;
      *pbEof = 1;
      return 1;
    }
    iRowid = sqlite3Fts5IterRowid(pIter);
    assert( (bDesc==0 && iRowid>=iLast) || (bDesc==1 && iRowid<=iLast) );
  }
  *piLast = iRowid;

  return 0;
}

/*
** IN/OUT parameter (*pa) points to a position list n bytes in size. If
** the position list contains entries for column iCol, then (*pa) is set
** to point to the sub-position-list for that column and the number of
** bytes in it returned. Or, if the argument position list does not
** contain any entries for column iCol, return 0.
*/
static int fts5ExprExtractCol(
  const u8 **pa,                  /* IN/OUT: Pointer to poslist */
  int n,                          /* IN: Size of poslist in bytes */
  int iCol                        /* Column to extract from poslist */
){
  int iCurrent = 0;
  const u8 *p = *pa;
  const u8 *pEnd = &p[n];         /* One byte past end of position list */
  u8 prev = 0;

  while( iCol!=iCurrent ){
    /* Advance pointer p until it points to pEnd or an 0x01 byte that is
    ** not part of a varint */
    while( (prev & 0x80) || *p!=0x01 ){
      prev = *p++;
      if( p==pEnd ) return 0;
    }
    *pa = p++;
    p += fts5GetVarint32(p, iCurrent);
  }

  /* Advance pointer p until it points to pEnd or an 0x01 byte that is
  ** not part of a varint */
  assert( (prev & 0x80)==0 );
  while( p<pEnd && ((prev & 0x80) || *p!=0x01) ){
    prev = *p++;
  }
  return p - (*pa);
}

static int fts5ExprExtractColset (
  Fts5ExprColset *pColset,        /* Colset to filter on */
  const u8 *pPos, int nPos,       /* Position list */
  Fts5Buffer *pBuf                /* Output buffer */
){
  int rc = SQLITE_OK;
  int i;

  fts5BufferZero(pBuf);
  for(i=0; i<pColset->nCol; i++){
    const u8 *pSub = pPos;
    int nSub = fts5ExprExtractCol(&pSub, nPos, pColset->aiCol[i]);
    if( nSub ){
      fts5BufferAppendBlob(&rc, pBuf, nSub, pSub);
    }
  }
  return rc;
}

static int fts5ExprNearTest(
  int *pRc,
  Fts5Expr *pExpr,                /* Expression that pNear is a part of */
  Fts5ExprNode *pNode             /* The "NEAR" node (FTS5_STRING) */
){
  Fts5ExprNearset *pNear = pNode->pNear;
  int rc = *pRc;
  int i;

  /* Check that each phrase in the nearset matches the current row.
  ** Populate the pPhrase->poslist buffers at the same time. If any
  ** phrase is not a match, break out of the loop early.  */
  for(i=0; rc==SQLITE_OK && i<pNear->nPhrase; i++){
    Fts5ExprPhrase *pPhrase = pNear->apPhrase[i];
    if( pPhrase->nTerm>1 || pNear->pColset ){
      int bMatch = 0;
      rc = fts5ExprPhraseIsMatch(pExpr, pNear->pColset, pPhrase, &bMatch);
      if( bMatch==0 ) break;
    }else{
      rc = sqlite3Fts5IterPoslistBuffer(
          pPhrase->aTerm[0].pIter, &pPhrase->poslist
      );
    }
  }

  *pRc = rc;
  if( i==pNear->nPhrase && (i==1 || fts5ExprNearIsMatch(pRc, pNear)) ){
    return 1;
  }

  return 0;
}

static int fts5ExprTokenTest(
  Fts5Expr *pExpr,                /* Expression that pNear is a part of */
  Fts5ExprNode *pNode             /* The "NEAR" node (FTS5_TERM) */
){
  /* As this "NEAR" object is actually a single phrase that consists 
  ** of a single term only, grab pointers into the poslist managed by the
  ** fts5_index.c iterator object. This is much faster than synthesizing 
  ** a new poslist the way we have to for more complicated phrase or NEAR
  ** expressions.  */
  Fts5ExprNearset *pNear = pNode->pNear;
  Fts5ExprPhrase *pPhrase = pNear->apPhrase[0];
  Fts5IndexIter *pIter = pPhrase->aTerm[0].pIter;
  Fts5ExprColset *pColset = pNear->pColset;
  const u8 *pPos;
  int nPos;
  int rc;

  assert( pNode->eType==FTS5_TERM );
  assert( pNear->nPhrase==1 && pPhrase->nTerm==1 );

  rc = sqlite3Fts5IterPoslist(pIter, &pPos, &nPos, &pNode->iRowid);

  /* If the term may match any column, then this must be a match. 
  ** Return immediately in this case. Otherwise, try to find the
  ** part of the poslist that corresponds to the required column.
  ** If it can be found, return. If it cannot, the next iteration
  ** of the loop will test the next rowid in the database for this
  ** term.  */
  if( pColset==0 ){
    assert( pPhrase->poslist.nSpace==0 );
    pPhrase->poslist.p = (u8*)pPos;
    pPhrase->poslist.n = nPos;
  }else if( pColset->nCol==1 ){
    assert( pPhrase->poslist.nSpace==0 );
    pPhrase->poslist.n = fts5ExprExtractCol(&pPos, nPos, pColset->aiCol[0]);
    pPhrase->poslist.p = (u8*)pPos;
  }else if( rc==SQLITE_OK ){
    rc = fts5ExprExtractColset(pColset, pPos, nPos, &pPhrase->poslist);
  }

  pNode->bNomatch = (pPhrase->poslist.n==0);
  return rc;
}

/*
** All individual term iterators in pNear are guaranteed to be valid when
** this function is called. This function checks if all term iterators
** point to the same rowid, and if not, advances them until they do.
** If an EOF is reached before this happens, *pbEof is set to true before
** returning.
**
** SQLITE_OK is returned if an error occurs, or an SQLite error code 
** otherwise. It is not considered an error code if an iterator reaches
** EOF.
*/
static int fts5ExprNearNextMatch(
  Fts5Expr *pExpr,                /* Expression pPhrase belongs to */
  Fts5ExprNode *pNode
){
  Fts5ExprNearset *pNear = pNode->pNear;
  Fts5ExprPhrase *pLeft = pNear->apPhrase[0];
  int rc = SQLITE_OK;
  i64 iLast;                      /* Lastest rowid any iterator points to */
  int i, j;                       /* Phrase and token index, respectively */
  int bMatch;                     /* True if all terms are at the same rowid */

  assert( pNear->nPhrase>1 || pNear->apPhrase[0]->nTerm>1 );

  /* Initialize iLast, the "lastest" rowid any iterator points to. If the
  ** iterator skips through rowids in the default ascending order, this means
  ** the maximum rowid. Or, if the iterator is "ORDER BY rowid DESC", then it
  ** means the minimum rowid.  */
  iLast = sqlite3Fts5IterRowid(pLeft->aTerm[0].pIter);

  do {
    bMatch = 1;
    for(i=0; i<pNear->nPhrase; i++){
      Fts5ExprPhrase *pPhrase = pNear->apPhrase[i];
      for(j=0; j<pPhrase->nTerm; j++){
        Fts5IndexIter *pIter = pPhrase->aTerm[j].pIter;
        i64 iRowid = sqlite3Fts5IterRowid(pIter);
        if( iRowid!=iLast ) bMatch = 0;
        if( fts5ExprAdvanceto(pIter, pExpr->bDesc, &iLast,&rc,&pNode->bEof) ){
          return rc;
        }
      }
    }
  }while( bMatch==0 );

  pNode->bNomatch = (0==fts5ExprNearTest(&rc, pExpr, pNode));
  pNode->iRowid = iLast;

  return rc;
}

/*
** Initialize all term iterators in the pNear object. If any term is found
** to match no documents at all, set *pbEof to true and return immediately,
** without initializing any further iterators.
*/
static int fts5ExprNearInitAll(
  Fts5Expr *pExpr,
  Fts5ExprNode *pNode
){
  Fts5ExprNearset *pNear = pNode->pNear;
  Fts5ExprTerm *pTerm;
  Fts5ExprPhrase *pPhrase;
  int i, j;
  int rc = SQLITE_OK;

  for(i=0; rc==SQLITE_OK && i<pNear->nPhrase; i++){
    pPhrase = pNear->apPhrase[i];
    for(j=0; j<pPhrase->nTerm; j++){
      pTerm = &pPhrase->aTerm[j];
      if( pTerm->pIter ){
        sqlite3Fts5IterClose(pTerm->pIter);
        pTerm->pIter = 0;
      }
      rc = sqlite3Fts5IndexQuery(
          pExpr->pIndex, pTerm->zTerm, strlen(pTerm->zTerm),
          (pTerm->bPrefix ? FTS5INDEX_QUERY_PREFIX : 0) |
          (pExpr->bDesc ? FTS5INDEX_QUERY_DESC : 0),
          &pTerm->pIter
      );
      assert( rc==SQLITE_OK || pTerm->pIter==0 );
      if( pTerm->pIter==0 || sqlite3Fts5IterEof(pTerm->pIter) ){
        pNode->bEof = 1;
        break;
      }
    }
  }

  return rc;
}

/* fts5ExprNodeNext() calls fts5ExprNodeNextMatch(). And vice-versa. */
static int fts5ExprNodeNextMatch(Fts5Expr*, Fts5ExprNode*);


/*
** If pExpr is an ASC iterator, this function returns a value with the
** same sign as:
**
**   (iLhs - iRhs)
**
** Otherwise, if this is a DESC iterator, the opposite is returned:
**
**   (iRhs - iLhs)
*/
static int fts5RowidCmp(
  Fts5Expr *pExpr,
  i64 iLhs,
  i64 iRhs
){
  assert( pExpr->bDesc==0 || pExpr->bDesc==1 );
  if( pExpr->bDesc==0 ){
    if( iLhs<iRhs ) return -1;
    return (iLhs > iRhs);
  }else{
    if( iLhs>iRhs ) return -1;
    return (iLhs < iRhs);
  }
}

static void fts5ExprSetEof(Fts5ExprNode *pNode){
  int i;
  pNode->bEof = 1;
  for(i=0; i<pNode->nChild; i++){
    fts5ExprSetEof(pNode->apChild[i]);
  }
}

static void fts5ExprNodeZeroPoslist(Fts5ExprNode *pNode){
  if( pNode->eType==FTS5_STRING || pNode->eType==FTS5_TERM ){
    Fts5ExprNearset *pNear = pNode->pNear;
    int i;
    for(i=0; i<pNear->nPhrase; i++){
      Fts5ExprPhrase *pPhrase = pNear->apPhrase[i];
      pPhrase->poslist.n = 0;
    }
  }else{
    int i;
    for(i=0; i<pNode->nChild; i++){
      fts5ExprNodeZeroPoslist(pNode->apChild[i]);
    }
  }
}


static int fts5ExprNodeNext(Fts5Expr*, Fts5ExprNode*, int, i64);

/*
** Argument pNode is an FTS5_AND node.
*/
static int fts5ExprAndNextRowid(
  Fts5Expr *pExpr,                /* Expression pPhrase belongs to */
  Fts5ExprNode *pAnd              /* FTS5_AND node to advance */
){
  int iChild;
  i64 iLast = pAnd->iRowid;
  int rc = SQLITE_OK;
  int bMatch;

  assert( pAnd->bEof==0 );
  do {
    pAnd->bNomatch = 0;
    bMatch = 1;
    for(iChild=0; iChild<pAnd->nChild; iChild++){
      Fts5ExprNode *pChild = pAnd->apChild[iChild];
      if( 0 && pChild->eType==FTS5_STRING ){
        /* TODO */
      }else{
        int cmp = fts5RowidCmp(pExpr, iLast, pChild->iRowid);
        if( cmp>0 ){
          /* Advance pChild until it points to iLast or laster */
          rc = fts5ExprNodeNext(pExpr, pChild, 1, iLast);
          if( rc!=SQLITE_OK ) return rc;
        }
      }

      /* If the child node is now at EOF, so is the parent AND node. Otherwise,
      ** the child node is guaranteed to have advanced at least as far as
      ** rowid iLast. So if it is not at exactly iLast, pChild->iRowid is the
      ** new lastest rowid seen so far.  */
      assert( pChild->bEof || fts5RowidCmp(pExpr, iLast, pChild->iRowid)<=0 );
      if( pChild->bEof ){
        fts5ExprSetEof(pAnd);
        bMatch = 1;
        break;
      }else if( iLast!=pChild->iRowid ){
        bMatch = 0;
        iLast = pChild->iRowid;
      }

      if( pChild->bNomatch ){
        pAnd->bNomatch = 1;
      }
    }
  }while( bMatch==0 );

  if( pAnd->bNomatch && pAnd!=pExpr->pRoot ){
    fts5ExprNodeZeroPoslist(pAnd);
  }
  pAnd->iRowid = iLast;
  return SQLITE_OK;
}


/*
** Compare the values currently indicated by the two nodes as follows:
**
**    res = (*p1) - (*p2)
**
** Nodes that point to values that come later in the iteration order are
** considered to be larger. Nodes at EOF are the largest of all.
**
** This means that if the iteration order is ASC, then numerically larger
** rowids are considered larger. Or if it is the default DESC, numerically
** smaller rowids are larger.
*/
static int fts5NodeCompare(
  Fts5Expr *pExpr,
  Fts5ExprNode *p1, 
  Fts5ExprNode *p2
){
  if( p2->bEof ) return -1;
  if( p1->bEof ) return +1;
  return fts5RowidCmp(pExpr, p1->iRowid, p2->iRowid);
}

/*
** Advance node iterator pNode, part of expression pExpr. If argument
** bFromValid is zero, then pNode is advanced exactly once. Or, if argument
** bFromValid is non-zero, then pNode is advanced until it is at or past
** rowid value iFrom. Whether "past" means "less than" or "greater than"
** depends on whether this is an ASC or DESC iterator.
*/
static int fts5ExprNodeNext(
  Fts5Expr *pExpr, 
  Fts5ExprNode *pNode,
  int bFromValid,
  i64 iFrom
){
  int rc = SQLITE_OK;

  if( pNode->bEof==0 ){
    switch( pNode->eType ){
      case FTS5_STRING: {
        rc = fts5ExprNearAdvanceFirst(pExpr, pNode, bFromValid, iFrom);
        break;
      };

      case FTS5_TERM: {
        rc = fts5ExprNearAdvanceFirst(pExpr, pNode, bFromValid, iFrom);
        if( pNode->bEof==0 ){
          assert( rc==SQLITE_OK );
          rc = fts5ExprTokenTest(pExpr, pNode);
        }
        return rc;
      };

      case FTS5_AND: {
        Fts5ExprNode *pLeft = pNode->apChild[0];
        rc = fts5ExprNodeNext(pExpr, pLeft, bFromValid, iFrom);
        break;
      }

      case FTS5_OR: {
        int i;
        i64 iLast = pNode->iRowid;

        for(i=0; rc==SQLITE_OK && i<pNode->nChild; i++){
          Fts5ExprNode *p1 = pNode->apChild[i];
          assert( p1->bEof || fts5RowidCmp(pExpr, p1->iRowid, iLast)>=0 );
          if( p1->bEof==0 ){
            if( (p1->iRowid==iLast) 
             || (bFromValid && fts5RowidCmp(pExpr, p1->iRowid, iFrom)<0)
            ){
              rc = fts5ExprNodeNext(pExpr, p1, bFromValid, iFrom);
            }
          }
        }

        break;
      }

      default: assert( pNode->eType==FTS5_NOT ); {
        assert( pNode->nChild==2 );
        rc = fts5ExprNodeNext(pExpr, pNode->apChild[0], bFromValid, iFrom);
        break;
      }
    }

    if( rc==SQLITE_OK ){
      rc = fts5ExprNodeNextMatch(pExpr, pNode);
    }
  }

  /* Assert that if bFromValid was true, either:
  **
  **   a) an error occurred, or
  **   b) the node is now at EOF, or
  **   c) the node is now at or past rowid iFrom.
  */
  assert( bFromValid==0 
      || rc!=SQLITE_OK                                                  /* a */
      || pNode->bEof                                                    /* b */
      || pNode->iRowid==iFrom || pExpr->bDesc==(pNode->iRowid<iFrom)    /* c */
  );

  return rc;
}


/*
** If pNode currently points to a match, this function returns SQLITE_OK
** without modifying it. Otherwise, pNode is advanced until it does point
** to a match or EOF is reached.
*/
static int fts5ExprNodeNextMatch(
  Fts5Expr *pExpr,                /* Expression of which pNode is a part */
  Fts5ExprNode *pNode             /* Expression node to test */
){
  int rc = SQLITE_OK;
  if( pNode->bEof==0 ){
    switch( pNode->eType ){

      case FTS5_STRING: {
        /* Advance the iterators until they all point to the same rowid */
        rc = fts5ExprNearNextMatch(pExpr, pNode);
        break;
      }

      case FTS5_TERM: {
        rc = fts5ExprTokenTest(pExpr, pNode);
        break;
      }

      case FTS5_AND: {
        rc = fts5ExprAndNextRowid(pExpr, pNode);
        break;
      }

      case FTS5_OR: {
        Fts5ExprNode *pNext = pNode->apChild[0];
        int i;

        for(i=1; i<pNode->nChild; i++){
          Fts5ExprNode *pChild = pNode->apChild[i];
          int cmp = fts5NodeCompare(pExpr, pNext, pChild);
          if( cmp>0 || (cmp==0 && pChild->bNomatch==0) ){
            pNext = pChild;
          }
        }
        pNode->iRowid = pNext->iRowid;
        pNode->bEof = pNext->bEof;
        pNode->bNomatch = pNext->bNomatch;
        break;
      }

      default: assert( pNode->eType==FTS5_NOT ); {
        Fts5ExprNode *p1 = pNode->apChild[0];
        Fts5ExprNode *p2 = pNode->apChild[1];
        assert( pNode->nChild==2 );

        while( rc==SQLITE_OK && p1->bEof==0 ){
          int cmp = fts5NodeCompare(pExpr, p1, p2);
          if( cmp>0 ){
            rc = fts5ExprNodeNext(pExpr, p2, 1, p1->iRowid);
            cmp = fts5NodeCompare(pExpr, p1, p2);
          }
          assert( rc!=SQLITE_OK || cmp<=0 );
          if( cmp || p2->bNomatch ) break;
          rc = fts5ExprNodeNext(pExpr, p1, 0, 0);
        }
        pNode->bEof = p1->bEof;
        pNode->iRowid = p1->iRowid;
        break;
      }
    }
  }
  return rc;
}

 
/*
** Set node pNode, which is part of expression pExpr, to point to the first
** match. If there are no matches, set the Node.bEof flag to indicate EOF.
**
** Return an SQLite error code if an error occurs, or SQLITE_OK otherwise.
** It is not an error if there are no matches.
*/
static int fts5ExprNodeFirst(Fts5Expr *pExpr, Fts5ExprNode *pNode){
  int rc = SQLITE_OK;
  pNode->bEof = 0;

  if( Fts5NodeIsString(pNode) ){
    /* Initialize all term iterators in the NEAR object. */
    rc = fts5ExprNearInitAll(pExpr, pNode);
  }else{
    int i;
    for(i=0; i<pNode->nChild && rc==SQLITE_OK; i++){
      rc = fts5ExprNodeFirst(pExpr, pNode->apChild[i]);
    }
    pNode->iRowid = pNode->apChild[0]->iRowid;
  }

  if( rc==SQLITE_OK ){
    rc = fts5ExprNodeNextMatch(pExpr, pNode);
  }
  return rc;
}


/*
** Begin iterating through the set of documents in index pIdx matched by
** the MATCH expression passed as the first argument. If the "bDesc" 
** parameter is passed a non-zero value, iteration is in descending rowid 
** order. Or, if it is zero, in ascending order.
**
** If iterating in ascending rowid order (bDesc==0), the first document
** visited is that with the smallest rowid that is larger than or equal
** to parameter iFirst. Or, if iterating in ascending order (bDesc==1),
** then the first document visited must have a rowid smaller than or
** equal to iFirst.
**
** Return SQLITE_OK if successful, or an SQLite error code otherwise. It
** is not considered an error if the query does not match any documents.
*/
int sqlite3Fts5ExprFirst(Fts5Expr *p, Fts5Index *pIdx, i64 iFirst, int bDesc){
  Fts5ExprNode *pRoot = p->pRoot;
  int rc = SQLITE_OK;
  if( pRoot ){
    p->pIndex = pIdx;
    p->bDesc = bDesc;
    rc = fts5ExprNodeFirst(p, pRoot);

    /* If not at EOF but the current rowid occurs earlier than iFirst in
    ** the iteration order, move to document iFirst or later. */
    if( pRoot->bEof==0 && fts5RowidCmp(p, pRoot->iRowid, iFirst)<0 ){
      rc = fts5ExprNodeNext(p, pRoot, 1, iFirst);
    }

    /* If the iterator is not at a real match, skip forward until it is. */
    while( pRoot->bNomatch && rc==SQLITE_OK && pRoot->bEof==0 ){
      rc = fts5ExprNodeNext(p, pRoot, 0, 0);
    }
  }
  return rc;
}

/*
** Move to the next document 
**
** Return SQLITE_OK if successful, or an SQLite error code otherwise. It
** is not considered an error if the query does not match any documents.
*/
int sqlite3Fts5ExprNext(Fts5Expr *p, i64 iLast){
  int rc;
  Fts5ExprNode *pRoot = p->pRoot;
  do {
    rc = fts5ExprNodeNext(p, pRoot, 0, 0);
  }while( pRoot->bNomatch && pRoot->bEof==0 && rc==SQLITE_OK );
  if( fts5RowidCmp(p, pRoot->iRowid, iLast)>0 ){
    pRoot->bEof = 1;
  }
  return rc;
}

int sqlite3Fts5ExprEof(Fts5Expr *p){
  return (p->pRoot==0 || p->pRoot->bEof);
}

i64 sqlite3Fts5ExprRowid(Fts5Expr *p){
  return p->pRoot->iRowid;
}

static int fts5ParseStringFromToken(Fts5Token *pToken, char **pz){
  int rc = SQLITE_OK;
  *pz = sqlite3Fts5Strndup(&rc, pToken->p, pToken->n);
  return rc;
}

/*
** Free the phrase object passed as the only argument.
*/
static void fts5ExprPhraseFree(Fts5ExprPhrase *pPhrase){
  if( pPhrase ){
    int i;
    for(i=0; i<pPhrase->nTerm; i++){
      Fts5ExprTerm *pTerm = &pPhrase->aTerm[i];
      sqlite3_free(pTerm->zTerm);
      if( pTerm->pIter ){
        sqlite3Fts5IterClose(pTerm->pIter);
      }
    }
    if( pPhrase->poslist.nSpace>0 ) fts5BufferFree(&pPhrase->poslist);
    sqlite3_free(pPhrase);
  }
}

/*
** If argument pNear is NULL, then a new Fts5ExprNearset object is allocated
** and populated with pPhrase. Or, if pNear is not NULL, phrase pPhrase is
** appended to it and the results returned.
**
** If an OOM error occurs, both the pNear and pPhrase objects are freed and
** NULL returned.
*/
Fts5ExprNearset *sqlite3Fts5ParseNearset(
  Fts5Parse *pParse,              /* Parse context */
  Fts5ExprNearset *pNear,         /* Existing nearset, or NULL */
  Fts5ExprPhrase *pPhrase         /* Recently parsed phrase */
){
  const int SZALLOC = 8;
  Fts5ExprNearset *pRet = 0;

  if( pParse->rc==SQLITE_OK ){
    if( pPhrase==0 ){
      return pNear;
    }
    if( pNear==0 ){
      int nByte = sizeof(Fts5ExprNearset) + SZALLOC * sizeof(Fts5ExprPhrase*);
      pRet = sqlite3_malloc(nByte);
      if( pRet==0 ){
        pParse->rc = SQLITE_NOMEM;
      }else{
        memset(pRet, 0, nByte);
      }
    }else if( (pNear->nPhrase % SZALLOC)==0 ){
      int nNew = pNear->nPhrase + SZALLOC;
      int nByte = sizeof(Fts5ExprNearset) + nNew * sizeof(Fts5ExprPhrase*);

      pRet = (Fts5ExprNearset*)sqlite3_realloc(pNear, nByte);
      if( pRet==0 ){
        pParse->rc = SQLITE_NOMEM;
      }
    }else{
      pRet = pNear;
    }
  }

  if( pRet==0 ){
    assert( pParse->rc!=SQLITE_OK );
    sqlite3Fts5ParseNearsetFree(pNear);
    sqlite3Fts5ParsePhraseFree(pPhrase);
  }else{
    pRet->apPhrase[pRet->nPhrase++] = pPhrase;
  }
  return pRet;
}

typedef struct TokenCtx TokenCtx;
struct TokenCtx {
  Fts5ExprPhrase *pPhrase;
};

/*
** Callback for tokenizing terms used by ParseTerm().
*/
static int fts5ParseTokenize(
  void *pContext,                 /* Pointer to Fts5InsertCtx object */
  const char *pToken,             /* Buffer containing token */
  int nToken,                     /* Size of token in bytes */
  int iStart,                     /* Start offset of token */
  int iEnd                        /* End offset of token */
){
  int rc = SQLITE_OK;
  const int SZALLOC = 8;
  TokenCtx *pCtx = (TokenCtx*)pContext;
  Fts5ExprPhrase *pPhrase = pCtx->pPhrase;
  Fts5ExprTerm *pTerm;

  if( pPhrase==0 || (pPhrase->nTerm % SZALLOC)==0 ){
    Fts5ExprPhrase *pNew;
    int nNew = SZALLOC + (pPhrase ? pPhrase->nTerm : 0);

    pNew = (Fts5ExprPhrase*)sqlite3_realloc(pPhrase, 
        sizeof(Fts5ExprPhrase) + sizeof(Fts5ExprTerm) * nNew
    );
    if( pNew==0 ) return SQLITE_NOMEM;
    if( pPhrase==0 ) memset(pNew, 0, sizeof(Fts5ExprPhrase));
    pCtx->pPhrase = pPhrase = pNew;
    pNew->nTerm = nNew - SZALLOC;
  }

  pTerm = &pPhrase->aTerm[pPhrase->nTerm++];
  memset(pTerm, 0, sizeof(Fts5ExprTerm));
  pTerm->zTerm = sqlite3Fts5Strndup(&rc, pToken, nToken);

  return rc;
}


/*
** Free the phrase object passed as the only argument.
*/
void sqlite3Fts5ParsePhraseFree(Fts5ExprPhrase *pPhrase){
  fts5ExprPhraseFree(pPhrase);
}

/*
** Free the phrase object passed as the second argument.
*/
void sqlite3Fts5ParseNearsetFree(Fts5ExprNearset *pNear){
  if( pNear ){
    int i;
    for(i=0; i<pNear->nPhrase; i++){
      fts5ExprPhraseFree(pNear->apPhrase[i]);
    }
    sqlite3_free(pNear->pColset);
    sqlite3_free(pNear);
  }
}

void sqlite3Fts5ParseFinished(Fts5Parse *pParse, Fts5ExprNode *p){
  assert( pParse->pExpr==0 );
  pParse->pExpr = p;
}

/*
** This function is called by the parser to process a string token. The
** string may or may not be quoted. In any case it is tokenized and a
** phrase object consisting of all tokens returned.
*/
Fts5ExprPhrase *sqlite3Fts5ParseTerm(
  Fts5Parse *pParse,              /* Parse context */
  Fts5ExprPhrase *pAppend,        /* Phrase to append to */
  Fts5Token *pToken,              /* String to tokenize */
  int bPrefix                     /* True if there is a trailing "*" */
){
  Fts5Config *pConfig = pParse->pConfig;
  TokenCtx sCtx;                  /* Context object passed to callback */
  int rc;                         /* Tokenize return code */
  char *z = 0;

  memset(&sCtx, 0, sizeof(TokenCtx));
  sCtx.pPhrase = pAppend;

  rc = fts5ParseStringFromToken(pToken, &z);
  if( rc==SQLITE_OK ){
    sqlite3Fts5Dequote(z);
    rc = sqlite3Fts5Tokenize(pConfig, z, strlen(z), &sCtx, fts5ParseTokenize);
  }
  sqlite3_free(z);
  if( rc ){
    pParse->rc = rc;
    fts5ExprPhraseFree(sCtx.pPhrase);
    sCtx.pPhrase = 0;
  }else if( sCtx.pPhrase ){

    if( pAppend==0 ){
      if( (pParse->nPhrase % 8)==0 ){
        int nByte = sizeof(Fts5ExprPhrase*) * (pParse->nPhrase + 8);
        Fts5ExprPhrase **apNew;
        apNew = (Fts5ExprPhrase**)sqlite3_realloc(pParse->apPhrase, nByte);
        if( apNew==0 ){
          pParse->rc = SQLITE_NOMEM;
          fts5ExprPhraseFree(sCtx.pPhrase);
          return 0;
        }
        pParse->apPhrase = apNew;
      }
      pParse->nPhrase++;
    }

    pParse->apPhrase[pParse->nPhrase-1] = sCtx.pPhrase;
    assert( sCtx.pPhrase->nTerm>0 );
    sCtx.pPhrase->aTerm[sCtx.pPhrase->nTerm-1].bPrefix = bPrefix;
  }

  return sCtx.pPhrase;
}

/*
** Token pTok has appeared in a MATCH expression where the NEAR operator
** is expected. If token pTok does not contain "NEAR", store an error
** in the pParse object.
*/
void sqlite3Fts5ParseNear(Fts5Parse *pParse, Fts5Token *pTok){
  if( pTok->n!=4 || memcmp("NEAR", pTok->p, 4) ){
    sqlite3Fts5ParseError(
        pParse, "fts5: syntax error near \"%.*s\"", pTok->n, pTok->p
    );
  }
}

void sqlite3Fts5ParseSetDistance(
  Fts5Parse *pParse, 
  Fts5ExprNearset *pNear,
  Fts5Token *p
){
  int nNear = 0;
  int i;
  if( p->n ){
    for(i=0; i<p->n; i++){
      char c = (char)p->p[i];
      if( c<'0' || c>'9' ){
        sqlite3Fts5ParseError(
            pParse, "expected integer, got \"%.*s\"", p->n, p->p
        );
        return;
      }
      nNear = nNear * 10 + (p->p[i] - '0');
    }
  }else{
    nNear = FTS5_DEFAULT_NEARDIST;
  }
  pNear->nNear = nNear;
}

/*
** The second argument passed to this function may be NULL, or it may be
** an existing Fts5ExprColset object. This function returns a pointer to
** a new colset object containing the contents of (p) with new value column
** number iCol appended. 
**
** If an OOM error occurs, store an error code in pParse and return NULL.
** The old colset object (if any) is not freed in this case.
*/
static Fts5ExprColset *fts5ParseColset(
  Fts5Parse *pParse,              /* Store SQLITE_NOMEM here if required */
  Fts5ExprColset *p,              /* Existing colset object */
  int iCol                        /* New column to add to colset object */
){
  int nCol = p ? p->nCol : 0;     /* Num. columns already in colset object */
  Fts5ExprColset *pNew;           /* New colset object to return */

  assert( pParse->rc==SQLITE_OK );
  assert( iCol>=0 && iCol<pParse->pConfig->nCol );

  pNew = sqlite3_realloc(p, sizeof(Fts5ExprColset) + sizeof(int)*nCol);
  if( pNew==0 ){
    pParse->rc = SQLITE_NOMEM;
  }else{
    int *aiCol = pNew->aiCol;
    int i, j;
    for(i=0; i<nCol; i++){
      if( aiCol[i]==iCol ) return pNew;
      if( aiCol[i]>iCol ) break;
    }
    for(j=nCol; j>i; j--){
      aiCol[j] = aiCol[j-1];
    }
    aiCol[i] = iCol;
    pNew->nCol = nCol+1;

#ifndef NDEBUG
    /* Check that the array is in order and contains no duplicate entries. */
    for(i=1; i<pNew->nCol; i++) assert( pNew->aiCol[i]>pNew->aiCol[i-1] );
#endif
  }

  return pNew;
}

Fts5ExprColset *sqlite3Fts5ParseColset(
  Fts5Parse *pParse,              /* Store SQLITE_NOMEM here if required */
  Fts5ExprColset *pColset,        /* Existing colset object */
  Fts5Token *p
){
  Fts5ExprColset *pRet = 0;
  int iCol;
  char *z;                        /* Dequoted copy of token p */

  z = sqlite3Fts5Strndup(&pParse->rc, p->p, p->n);
  if( pParse->rc==SQLITE_OK ){
    Fts5Config *pConfig = pParse->pConfig;
    sqlite3Fts5Dequote(z);
    for(iCol=0; iCol<pConfig->nCol; iCol++){
      if( 0==sqlite3_stricmp(pConfig->azCol[iCol], z) ) break;
    }
    if( iCol==pConfig->nCol ){
      sqlite3Fts5ParseError(pParse, "no such column: %s", z);
    }else{
      pRet = fts5ParseColset(pParse, pColset, iCol);
    }
    sqlite3_free(z);
  }

  if( pRet==0 ){
    assert( pParse->rc!=SQLITE_OK );
    sqlite3_free(pColset);
  }

  return pRet;
}

void sqlite3Fts5ParseSetColset(
  Fts5Parse *pParse, 
  Fts5ExprNearset *pNear, 
  Fts5ExprColset *pColset 
){
  if( pNear ){
    pNear->pColset = pColset;
  }else{
    sqlite3_free(pColset);
  }
}

static void fts5ExprAddChildren(Fts5ExprNode *p, Fts5ExprNode *pSub){
  if( p->eType!=FTS5_NOT && pSub->eType==p->eType ){
    int nByte = sizeof(Fts5ExprNode*) * pSub->nChild;
    memcpy(&p->apChild[p->nChild], pSub->apChild, nByte);
    p->nChild += pSub->nChild;
    sqlite3_free(pSub);
  }else{
    p->apChild[p->nChild++] = pSub;
  }
}

/*
** Allocate and return a new expression object. If anything goes wrong (i.e.
** OOM error), leave an error code in pParse and return NULL.
*/
Fts5ExprNode *sqlite3Fts5ParseNode(
  Fts5Parse *pParse,              /* Parse context */
  int eType,                      /* FTS5_STRING, AND, OR or NOT */
  Fts5ExprNode *pLeft,            /* Left hand child expression */
  Fts5ExprNode *pRight,           /* Right hand child expression */
  Fts5ExprNearset *pNear          /* For STRING expressions, the near cluster */
){
  Fts5ExprNode *pRet = 0;

  if( pParse->rc==SQLITE_OK ){
    int nChild = 0;               /* Number of children of returned node */
    int nByte;                    /* Bytes of space to allocate for this node */
 
    assert( (eType!=FTS5_STRING && !pNear)
         || (eType==FTS5_STRING && !pLeft && !pRight)
    );
    if( eType==FTS5_STRING && pNear==0 ) return 0;
    if( eType!=FTS5_STRING && pLeft==0 ) return pRight;
    if( eType!=FTS5_STRING && pRight==0 ) return pLeft;

    if( eType==FTS5_NOT ){
      nChild = 2;
    }else if( eType==FTS5_AND || eType==FTS5_OR ){
      nChild = 2;
      if( pLeft->eType==eType ) nChild += pLeft->nChild-1;
      if( pRight->eType==eType ) nChild += pRight->nChild-1;
    }

    nByte = sizeof(Fts5ExprNode) + sizeof(Fts5ExprNode*)*nChild;
    pRet = (Fts5ExprNode*)sqlite3Fts5MallocZero(&pParse->rc, nByte);

    if( pRet ){
      pRet->eType = eType;
      pRet->pNear = pNear;
      if( eType==FTS5_STRING ){
        int iPhrase;
        for(iPhrase=0; iPhrase<pNear->nPhrase; iPhrase++){
          pNear->apPhrase[iPhrase]->pNode = pRet;
        }
        if( pNear->nPhrase==1 && pNear->apPhrase[0]->nTerm==1 ){
          pRet->eType = FTS5_TERM;
        }
      }else{
        fts5ExprAddChildren(pRet, pLeft);
        fts5ExprAddChildren(pRet, pRight);
      }
    }
  }

  if( pRet==0 ){
    assert( pParse->rc!=SQLITE_OK );
    sqlite3Fts5ParseNodeFree(pLeft);
    sqlite3Fts5ParseNodeFree(pRight);
    sqlite3Fts5ParseNearsetFree(pNear);
  }
  return pRet;
}

static char *fts5ExprTermPrint(Fts5ExprTerm *pTerm){
  char *zQuoted = sqlite3_malloc(strlen(pTerm->zTerm) * 2 + 3 + 2);
  if( zQuoted ){
    int i = 0;
    char *zIn = pTerm->zTerm;
    zQuoted[i++] = '"';
    while( *zIn ){
      if( *zIn=='"' ) zQuoted[i++] = '"';
      zQuoted[i++] = *zIn++;
    }
    zQuoted[i++] = '"';
    if( pTerm->bPrefix ){
      zQuoted[i++] = ' ';
      zQuoted[i++] = '*';
    }
    zQuoted[i++] = '\0';
  }
  return zQuoted;
}

static char *fts5PrintfAppend(char *zApp, const char *zFmt, ...){
  char *zNew;
  va_list ap;
  va_start(ap, zFmt);
  zNew = sqlite3_vmprintf(zFmt, ap);
  va_end(ap);
  if( zApp && zNew ){
    char *zNew2 = sqlite3_mprintf("%s%s", zApp, zNew);
    sqlite3_free(zNew);
    zNew = zNew2;
  }
  sqlite3_free(zApp);
  return zNew;
}

/*
** Compose a tcl-readable representation of expression pExpr. Return a 
** pointer to a buffer containing that representation. It is the 
** responsibility of the caller to at some point free the buffer using 
** sqlite3_free().
*/
static char *fts5ExprPrintTcl(
  Fts5Config *pConfig, 
  const char *zNearsetCmd,
  Fts5ExprNode *pExpr
){
  char *zRet = 0;
  if( pExpr->eType==FTS5_STRING || pExpr->eType==FTS5_TERM ){
    Fts5ExprNearset *pNear = pExpr->pNear;
    int i; 
    int iTerm;

    zRet = fts5PrintfAppend(zRet, "%s ", zNearsetCmd);
    if( zRet==0 ) return 0;
    if( pNear->pColset ){
      int *aiCol = pNear->pColset->aiCol;
      int nCol = pNear->pColset->nCol;
      if( nCol==1 ){
        zRet = fts5PrintfAppend(zRet, "-col %d ", aiCol[0]);
      }else{
        zRet = fts5PrintfAppend(zRet, "-col {%d", aiCol[0]);
        for(i=1; i<pNear->pColset->nCol; i++){
          zRet = fts5PrintfAppend(zRet, " %d", aiCol[i]);
        }
        zRet = fts5PrintfAppend(zRet, "} ");
      }
      if( zRet==0 ) return 0;
    }

    if( pNear->nPhrase>1 ){
      zRet = fts5PrintfAppend(zRet, "-near %d ", pNear->nNear);
      if( zRet==0 ) return 0;
    }

    zRet = fts5PrintfAppend(zRet, "--");
    if( zRet==0 ) return 0;

    for(i=0; i<pNear->nPhrase; i++){
      Fts5ExprPhrase *pPhrase = pNear->apPhrase[i];

      zRet = fts5PrintfAppend(zRet, " {");
      for(iTerm=0; zRet && iTerm<pPhrase->nTerm; iTerm++){
        char *zTerm = pPhrase->aTerm[iTerm].zTerm;
        zRet = fts5PrintfAppend(zRet, "%s%s", iTerm==0?"":" ", zTerm);
      }

      if( zRet ) zRet = fts5PrintfAppend(zRet, "}");
      if( zRet==0 ) return 0;
    }

  }else{
    char const *zOp = 0;
    int i;
    switch( pExpr->eType ){
      case FTS5_AND: zOp = "AND"; break;
      case FTS5_NOT: zOp = "NOT"; break;
      default: 
        assert( pExpr->eType==FTS5_OR );
        zOp = "OR"; 
        break;
    }

    zRet = sqlite3_mprintf("%s", zOp);
    for(i=0; zRet && i<pExpr->nChild; i++){
      char *z = fts5ExprPrintTcl(pConfig, zNearsetCmd, pExpr->apChild[i]);
      if( !z ){
        sqlite3_free(zRet);
        zRet = 0;
      }else{
        zRet = fts5PrintfAppend(zRet, " [%z]", z);
      }
    }
  }

  return zRet;
}

static char *fts5ExprPrint(Fts5Config *pConfig, Fts5ExprNode *pExpr){
  char *zRet = 0;
  if( pExpr->eType==FTS5_STRING || pExpr->eType==FTS5_TERM ){
    Fts5ExprNearset *pNear = pExpr->pNear;
    int i; 
    int iTerm;

    if( pNear->pColset ){
      int iCol = pNear->pColset->aiCol[0];
      zRet = fts5PrintfAppend(zRet, "%s : ", pConfig->azCol[iCol]);
      if( zRet==0 ) return 0;
    }

    if( pNear->nPhrase>1 ){
      zRet = fts5PrintfAppend(zRet, "NEAR(");
      if( zRet==0 ) return 0;
    }

    for(i=0; i<pNear->nPhrase; i++){
      Fts5ExprPhrase *pPhrase = pNear->apPhrase[i];
      if( i!=0 ){
        zRet = fts5PrintfAppend(zRet, " ");
        if( zRet==0 ) return 0;
      }
      for(iTerm=0; iTerm<pPhrase->nTerm; iTerm++){
        char *zTerm = fts5ExprTermPrint(&pPhrase->aTerm[iTerm]);
        if( zTerm ){
          zRet = fts5PrintfAppend(zRet, "%s%s", iTerm==0?"":" + ", zTerm);
          sqlite3_free(zTerm);
        }
        if( zTerm==0 || zRet==0 ){
          sqlite3_free(zRet);
          return 0;
        }
      }
    }

    if( pNear->nPhrase>1 ){
      zRet = fts5PrintfAppend(zRet, ", %d)", pNear->nNear);
      if( zRet==0 ) return 0;
    }

  }else{
    char const *zOp = 0;
    int i;

    switch( pExpr->eType ){
      case FTS5_AND: zOp = " AND "; break;
      case FTS5_NOT: zOp = " NOT "; break;
      default:  
        assert( pExpr->eType==FTS5_OR );
        zOp = " OR "; 
        break;
    }

    for(i=0; i<pExpr->nChild; i++){
      char *z = fts5ExprPrint(pConfig, pExpr->apChild[i]);
      if( z==0 ){
        sqlite3_free(zRet);
        zRet = 0;
      }else{
        int e = pExpr->apChild[i]->eType;
        int b = (e!=FTS5_STRING && e!=FTS5_TERM);
        zRet = fts5PrintfAppend(zRet, "%s%s%z%s", 
            (i==0 ? "" : zOp),
            (b?"(":""), z, (b?")":"")
        );
      }
      if( zRet==0 ) break;
    }
  }

  return zRet;
}

/*
** The implementation of user-defined scalar functions fts5_expr() (bTcl==0)
** and fts5_expr_tcl() (bTcl!=0).
*/
static void fts5ExprFunction(
  sqlite3_context *pCtx,          /* Function call context */
  int nArg,                       /* Number of args */
  sqlite3_value **apVal,          /* Function arguments */
  int bTcl
){
  Fts5Global *pGlobal = (Fts5Global*)sqlite3_user_data(pCtx);
  sqlite3 *db = sqlite3_context_db_handle(pCtx);
  const char *zExpr = 0;
  char *zErr = 0;
  Fts5Expr *pExpr = 0;
  int rc;
  int i;

  const char **azConfig;          /* Array of arguments for Fts5Config */
  const char *zNearsetCmd = "nearset";
  int nConfig;                    /* Size of azConfig[] */
  Fts5Config *pConfig = 0;
  int iArg = 1;

  if( bTcl && nArg>1 ){
    zNearsetCmd = (const char*)sqlite3_value_text(apVal[1]);
    iArg = 2;
  }

  nConfig = 3 + (nArg-iArg);
  azConfig = (const char**)sqlite3_malloc(sizeof(char*) * nConfig);
  if( azConfig==0 ){
    sqlite3_result_error_nomem(pCtx);
    return;
  }
  azConfig[0] = 0;
  azConfig[1] = "main";
  azConfig[2] = "tbl";
  for(i=3; iArg<nArg; iArg++){
    azConfig[i++] = (const char*)sqlite3_value_text(apVal[iArg]);
  }

  zExpr = (const char*)sqlite3_value_text(apVal[0]);

  rc = sqlite3Fts5ConfigParse(pGlobal, db, nConfig, azConfig, &pConfig, &zErr);
  if( rc==SQLITE_OK ){
    rc = sqlite3Fts5ExprNew(pConfig, zExpr, &pExpr, &zErr);
  }
  if( rc==SQLITE_OK ){
    char *zText;
    if( pExpr->pRoot==0 ){
      zText = sqlite3_mprintf("");
    }else if( bTcl ){
      zText = fts5ExprPrintTcl(pConfig, zNearsetCmd, pExpr->pRoot);
    }else{
      zText = fts5ExprPrint(pConfig, pExpr->pRoot);
    }
    if( zText==0 ){
      rc = SQLITE_NOMEM;
    }else{
      sqlite3_result_text(pCtx, zText, -1, SQLITE_TRANSIENT);
      sqlite3_free(zText);
    }
  }

  if( rc!=SQLITE_OK ){
    if( zErr ){
      sqlite3_result_error(pCtx, zErr, -1);
      sqlite3_free(zErr);
    }else{
      sqlite3_result_error_code(pCtx, rc);
    }
  }
  sqlite3_free((void *)azConfig);
  sqlite3Fts5ConfigFree(pConfig);
  sqlite3Fts5ExprFree(pExpr);
}

static void fts5ExprFunctionHr(
  sqlite3_context *pCtx,          /* Function call context */
  int nArg,                       /* Number of args */
  sqlite3_value **apVal           /* Function arguments */
){
  fts5ExprFunction(pCtx, nArg, apVal, 0);
}
static void fts5ExprFunctionTcl(
  sqlite3_context *pCtx,          /* Function call context */
  int nArg,                       /* Number of args */
  sqlite3_value **apVal           /* Function arguments */
){
  fts5ExprFunction(pCtx, nArg, apVal, 1);
}

/*
** The implementation of an SQLite user-defined-function that accepts a
** single integer as an argument. If the integer is an alpha-numeric 
** unicode code point, 1 is returned. Otherwise 0.
*/
static void fts5ExprIsAlnum(
  sqlite3_context *pCtx,          /* Function call context */
  int nArg,                       /* Number of args */
  sqlite3_value **apVal           /* Function arguments */
){
  int iCode;
  if( nArg!=1 ){
    sqlite3_result_error(pCtx, 
        "wrong number of arguments to function fts5_isalnum", -1
    );
    return;
  }
  iCode = sqlite3_value_int(apVal[0]);
  sqlite3_result_int(pCtx, sqlite3Fts5UnicodeIsalnum(iCode));
}

static void fts5ExprFold(
  sqlite3_context *pCtx,          /* Function call context */
  int nArg,                       /* Number of args */
  sqlite3_value **apVal           /* Function arguments */
){
  if( nArg!=1 && nArg!=2 ){
    sqlite3_result_error(pCtx, 
        "wrong number of arguments to function fts5_fold", -1
    );
  }else{
    int iCode;
    int bRemoveDiacritics = 0;
    iCode = sqlite3_value_int(apVal[0]);
    if( nArg==2 ) bRemoveDiacritics = sqlite3_value_int(apVal[1]);
    sqlite3_result_int(pCtx, sqlite3Fts5UnicodeFold(iCode, bRemoveDiacritics));
  }
}

/*
** This is called during initialization to register the fts5_expr() scalar
** UDF with the SQLite handle passed as the only argument.
*/
int sqlite3Fts5ExprInit(Fts5Global *pGlobal, sqlite3 *db){
  struct Fts5ExprFunc {
    const char *z;
    void (*x)(sqlite3_context*,int,sqlite3_value**);
  } aFunc[] = {
    { "fts5_expr",     fts5ExprFunctionHr },
    { "fts5_expr_tcl", fts5ExprFunctionTcl },
    { "fts5_isalnum",  fts5ExprIsAlnum },
    { "fts5_fold",     fts5ExprFold },
  };
  int i;
  int rc = SQLITE_OK;
  void *pCtx = (void*)pGlobal;

  for(i=0; rc==SQLITE_OK && i<(sizeof(aFunc) / sizeof(aFunc[0])); i++){
    struct Fts5ExprFunc *p = &aFunc[i];
    rc = sqlite3_create_function(db, p->z, -1, SQLITE_UTF8, pCtx, p->x, 0, 0);
  }

  return rc;
}

/*
** Return the number of phrases in expression pExpr.
*/
int sqlite3Fts5ExprPhraseCount(Fts5Expr *pExpr){
  return pExpr->nPhrase;
}

/*
** Return the number of terms in the iPhrase'th phrase in pExpr.
*/
int sqlite3Fts5ExprPhraseSize(Fts5Expr *pExpr, int iPhrase){
  if( iPhrase<0 || iPhrase>=pExpr->nPhrase ) return 0;
  return pExpr->apExprPhrase[iPhrase]->nTerm;
}

/*
** This function is used to access the current position list for phrase
** iPhrase.
*/
int sqlite3Fts5ExprPoslist(Fts5Expr *pExpr, int iPhrase, const u8 **pa){
  int nRet;
  Fts5ExprPhrase *pPhrase = pExpr->apExprPhrase[iPhrase];
  Fts5ExprNode *pNode = pPhrase->pNode;
  if( pNode->bEof==0 && pNode->iRowid==pExpr->pRoot->iRowid ){
    *pa = pPhrase->poslist.p;
    nRet = pPhrase->poslist.n;
  }else{
    *pa = 0;
    nRet = 0;
  }
  return nRet;
}

#endif /* SQLITE_ENABLE_FTS5 */

/*
** 2014 August 11
**
** The author disclaims copyright to this source code.  In place of
** a legal notice, here is a blessing:
**
**    May you do good and not evil.
**    May you find forgiveness for yourself and forgive others.
**    May you share freely, never taking more than you give.
**
******************************************************************************
**
*/

#ifdef SQLITE_ENABLE_FTS5


#include "fts5Int.h"

typedef struct Fts5HashEntry Fts5HashEntry;

/*
** This file contains the implementation of an in-memory hash table used
** to accumuluate "term -> doclist" content before it is flused to a level-0
** segment.
*/


struct Fts5Hash {
  int *pnByte;                    /* Pointer to bytes counter */
  int nEntry;                     /* Number of entries currently in hash */
  int nSlot;                      /* Size of aSlot[] array */
  Fts5HashEntry *pScan;           /* Current ordered scan item */
  Fts5HashEntry **aSlot;          /* Array of hash slots */
};

/*
** Each entry in the hash table is represented by an object of the 
** following type. Each object, its key (zKey[]) and its current data
** are stored in a single memory allocation. The position list data 
** immediately follows the key data in memory.
**
** The data that follows the key is in a similar, but not identical format
** to the doclist data stored in the database. It is:
**
**   * Rowid, as a varint
**   * Position list, without 0x00 terminator.
**   * Size of previous position list and rowid, as a 4 byte
**     big-endian integer.
**
** iRowidOff:
**   Offset of last rowid written to data area. Relative to first byte of
**   structure.
**
** nData:
**   Bytes of data written since iRowidOff.
*/
struct Fts5HashEntry {
  Fts5HashEntry *pHashNext;       /* Next hash entry with same hash-key */
  Fts5HashEntry *pScanNext;       /* Next entry in sorted order */
  
  int nAlloc;                     /* Total size of allocation */
  int iSzPoslist;                 /* Offset of space for 4-byte poslist size */
  int nData;                      /* Total bytes of data (incl. structure) */
  u8 bDel;                        /* Set delete-flag @ iSzPoslist */

  int iCol;                       /* Column of last value written */
  int iPos;                       /* Position of last value written */
  i64 iRowid;                     /* Rowid of last value written */
  char zKey[0];                   /* Nul-terminated entry key */
};

/*
** Allocate a new hash table.
*/
int sqlite3Fts5HashNew(Fts5Hash **ppNew, int *pnByte){
  int rc = SQLITE_OK;
  Fts5Hash *pNew;

  *ppNew = pNew = (Fts5Hash*)sqlite3_malloc(sizeof(Fts5Hash));
  if( pNew==0 ){
    rc = SQLITE_NOMEM;
  }else{
    int nByte;
    memset(pNew, 0, sizeof(Fts5Hash));
    pNew->pnByte = pnByte;

    pNew->nSlot = 1024;
    nByte = sizeof(Fts5HashEntry*) * pNew->nSlot;
    pNew->aSlot = (Fts5HashEntry**)sqlite3_malloc(nByte);
    if( pNew->aSlot==0 ){
      sqlite3_free(pNew);
      *ppNew = 0;
      rc = SQLITE_NOMEM;
    }else{
      memset(pNew->aSlot, 0, nByte);
    }
  }
  return rc;
}

/*
** Free a hash table object.
*/
void sqlite3Fts5HashFree(Fts5Hash *pHash){
  if( pHash ){
    sqlite3Fts5HashClear(pHash);
    sqlite3_free(pHash->aSlot);
    sqlite3_free(pHash);
  }
}

/*
** Empty (but do not delete) a hash table.
*/
void sqlite3Fts5HashClear(Fts5Hash *pHash){
  int i;
  for(i=0; i<pHash->nSlot; i++){
    Fts5HashEntry *pNext;
    Fts5HashEntry *pSlot;
    for(pSlot=pHash->aSlot[i]; pSlot; pSlot=pNext){
      pNext = pSlot->pHashNext;
      sqlite3_free(pSlot);
    }
  }
  memset(pHash->aSlot, 0, pHash->nSlot * sizeof(Fts5HashEntry*));
  pHash->nEntry = 0;
}

static unsigned int fts5HashKey(int nSlot, const char *p, int n){
  int i;
  unsigned int h = 13;
  for(i=n-1; i>=0; i--){
    h = (h << 3) ^ h ^ p[i];
  }
  return (h % nSlot);
}

static unsigned int fts5HashKey2(int nSlot, char b, const char *p, int n){
  int i;
  unsigned int h = 13;
  for(i=n-1; i>=0; i--){
    h = (h << 3) ^ h ^ p[i];
  }
  h = (h << 3) ^ h ^ b;
  return (h % nSlot);
}

/*
** Resize the hash table by doubling the number of slots.
*/
static int fts5HashResize(Fts5Hash *pHash){
  int nNew = pHash->nSlot*2;
  int i;
  Fts5HashEntry **apNew;
  Fts5HashEntry **apOld = pHash->aSlot;

  apNew = (Fts5HashEntry**)sqlite3_malloc(nNew*sizeof(Fts5HashEntry*));
  if( !apNew ) return SQLITE_NOMEM;
  memset(apNew, 0, nNew*sizeof(Fts5HashEntry*));

  for(i=0; i<pHash->nSlot; i++){
    while( apOld[i] ){
      int iHash;
      Fts5HashEntry *p = apOld[i];
      apOld[i] = p->pHashNext;
      iHash = fts5HashKey(nNew, p->zKey, strlen(p->zKey));
      p->pHashNext = apNew[iHash];
      apNew[iHash] = p;
    }
  }

  sqlite3_free(apOld);
  pHash->nSlot = nNew;
  pHash->aSlot = apNew;
  return SQLITE_OK;
}

static void fts5HashAddPoslistSize(Fts5HashEntry *p){
  if( p->iSzPoslist ){
    u8 *pPtr = (u8*)p;
    int nSz = (p->nData - p->iSzPoslist - 1);         /* Size in bytes */
    int nPos = nSz*2 + p->bDel;                       /* Value of nPos field */

    assert( p->bDel==0 || p->bDel==1 );
    if( nPos<=127 ){
      pPtr[p->iSzPoslist] = nPos;
    }else{
      int nByte = sqlite3Fts5GetVarintLen((u32)nPos);
      memmove(&pPtr[p->iSzPoslist + nByte], &pPtr[p->iSzPoslist + 1], nSz);
      sqlite3Fts5PutVarint(&pPtr[p->iSzPoslist], nPos);
      p->nData += (nByte-1);
    }
    p->bDel = 0;
    p->iSzPoslist = 0;
  }
}

int sqlite3Fts5HashWrite(
  Fts5Hash *pHash,
  i64 iRowid,                     /* Rowid for this entry */
  int iCol,                       /* Column token appears in (-ve -> delete) */
  int iPos,                       /* Position of token within column */
  char bByte,                     /* First byte of token */
  const char *pToken, int nToken  /* Token to add or remove to or from index */
){
  unsigned int iHash = fts5HashKey2(pHash->nSlot, bByte, pToken, nToken);
  Fts5HashEntry *p;
  u8 *pPtr;
  int nIncr = 0;                  /* Amount to increment (*pHash->pnByte) by */

  /* Attempt to locate an existing hash entry */
  for(p=pHash->aSlot[iHash]; p; p=p->pHashNext){
    if( p->zKey[0]==bByte 
     && memcmp(&p->zKey[1], pToken, nToken)==0 
     && p->zKey[nToken+1]==0 
    ){
      break;
    }
  }

  /* If an existing hash entry cannot be found, create a new one. */
  if( p==0 ){
    int nByte = sizeof(Fts5HashEntry) + (nToken+1) + 1 + 64;
    if( nByte<128 ) nByte = 128;

    if( (pHash->nEntry*2)>=pHash->nSlot ){
      int rc = fts5HashResize(pHash);
      if( rc!=SQLITE_OK ) return rc;
      iHash = fts5HashKey2(pHash->nSlot, bByte, pToken, nToken);
    }

    p = (Fts5HashEntry*)sqlite3_malloc(nByte);
    if( !p ) return SQLITE_NOMEM;
    memset(p, 0, sizeof(Fts5HashEntry));
    p->nAlloc = nByte;
    p->zKey[0] = bByte;
    memcpy(&p->zKey[1], pToken, nToken);
    assert( iHash==fts5HashKey(pHash->nSlot, p->zKey, nToken+1) );
    p->zKey[nToken+1] = '\0';
    p->nData = nToken+1 + 1 + sizeof(Fts5HashEntry);
    p->nData += sqlite3Fts5PutVarint(&((u8*)p)[p->nData], iRowid);
    p->iSzPoslist = p->nData;
    p->nData += 1;
    p->iRowid = iRowid;
    p->pHashNext = pHash->aSlot[iHash];
    pHash->aSlot[iHash] = p;
    pHash->nEntry++;
    nIncr += p->nData;
  }

  /* Check there is enough space to append a new entry. Worst case scenario
  ** is:
  **
  **     + 9 bytes for a new rowid,
  **     + 4 byte reserved for the "poslist size" varint.
  **     + 1 byte for a "new column" byte,
  **     + 3 bytes for a new column number (16-bit max) as a varint,
  **     + 5 bytes for the new position offset (32-bit max).
  */
  if( (p->nAlloc - p->nData) < (9 + 4 + 1 + 3 + 5) ){
    int nNew = p->nAlloc * 2;
    Fts5HashEntry *pNew;
    Fts5HashEntry **pp;
    pNew = (Fts5HashEntry*)sqlite3_realloc(p, nNew);
    if( pNew==0 ) return SQLITE_NOMEM;
    pNew->nAlloc = nNew;
    for(pp=&pHash->aSlot[iHash]; *pp!=p; pp=&(*pp)->pHashNext);
    *pp = pNew;
    p = pNew;
  }
  pPtr = (u8*)p;
  nIncr -= p->nData;

  /* If this is a new rowid, append the 4-byte size field for the previous
  ** entry, and the new rowid for this entry.  */
  if( iRowid!=p->iRowid ){
    fts5HashAddPoslistSize(p);
    p->nData += sqlite3Fts5PutVarint(&pPtr[p->nData], iRowid - p->iRowid);
    p->iSzPoslist = p->nData;
    p->nData += 1;
    p->iCol = 0;
    p->iPos = 0;
    p->iRowid = iRowid;
  }

  if( iCol>=0 ){
    /* Append a new column value, if necessary */
    assert( iCol>=p->iCol );
    if( iCol!=p->iCol ){
      pPtr[p->nData++] = 0x01;
      p->nData += sqlite3Fts5PutVarint(&pPtr[p->nData], iCol);
      p->iCol = iCol;
      p->iPos = 0;
    }

    /* Append the new position offset */
    p->nData += sqlite3Fts5PutVarint(&pPtr[p->nData], iPos - p->iPos + 2);
    p->iPos = iPos;
  }else{
    /* This is a delete. Set the delete flag. */
    p->bDel = 1;
  }
  nIncr += p->nData;

  *pHash->pnByte += nIncr;
  return SQLITE_OK;
}


/*
** Arguments pLeft and pRight point to linked-lists of hash-entry objects,
** each sorted in key order. This function merges the two lists into a
** single list and returns a pointer to its first element.
*/
static Fts5HashEntry *fts5HashEntryMerge(
  Fts5HashEntry *pLeft,
  Fts5HashEntry *pRight
){
  Fts5HashEntry *p1 = pLeft;
  Fts5HashEntry *p2 = pRight;
  Fts5HashEntry *pRet = 0;
  Fts5HashEntry **ppOut = &pRet;

  while( p1 || p2 ){
    if( p1==0 ){
      *ppOut = p2;
      p2 = 0;
    }else if( p2==0 ){
      *ppOut = p1;
      p1 = 0;
    }else{
      int i = 0;
      while( p1->zKey[i]==p2->zKey[i] ) i++;

      if( ((u8)p1->zKey[i])>((u8)p2->zKey[i]) ){
        /* p2 is smaller */
        *ppOut = p2;
        ppOut = &p2->pScanNext;
        p2 = p2->pScanNext;
      }else{
        /* p1 is smaller */
        *ppOut = p1;
        ppOut = &p1->pScanNext;
        p1 = p1->pScanNext;
      }
      *ppOut = 0;
    }
  }

  return pRet;
}

/*
** Extract all tokens from hash table iHash and link them into a list
** in sorted order. The hash table is cleared before returning. It is
** the responsibility of the caller to free the elements of the returned
** list.
*/
static int fts5HashEntrySort(
  Fts5Hash *pHash, 
  const char *pTerm, int nTerm,   /* Query prefix, if any */
  Fts5HashEntry **ppSorted
){
  const int nMergeSlot = 32;
  Fts5HashEntry **ap;
  Fts5HashEntry *pList;
  int iSlot;
  int i;

  *ppSorted = 0;
  ap = sqlite3_malloc(sizeof(Fts5HashEntry*) * nMergeSlot);
  if( !ap ) return SQLITE_NOMEM;
  memset(ap, 0, sizeof(Fts5HashEntry*) * nMergeSlot);

  for(iSlot=0; iSlot<pHash->nSlot; iSlot++){
    Fts5HashEntry *pIter;
    for(pIter=pHash->aSlot[iSlot]; pIter; pIter=pIter->pHashNext){
      if( pTerm==0 || 0==memcmp(pIter->zKey, pTerm, nTerm) ){
        Fts5HashEntry *pEntry = pIter;
        pEntry->pScanNext = 0;
        for(i=0; ap[i]; i++){
          pEntry = fts5HashEntryMerge(pEntry, ap[i]);
          ap[i] = 0;
        }
        ap[i] = pEntry;
      }
    }
  }

  pList = 0;
  for(i=0; i<nMergeSlot; i++){
    pList = fts5HashEntryMerge(pList, ap[i]);
  }

  pHash->nEntry = 0;
  sqlite3_free(ap);
  *ppSorted = pList;
  return SQLITE_OK;
}

/*
** Query the hash table for a doclist associated with term pTerm/nTerm.
*/
int sqlite3Fts5HashQuery(
  Fts5Hash *pHash,                /* Hash table to query */
  const char *pTerm, int nTerm,   /* Query term */
  const u8 **ppDoclist,           /* OUT: Pointer to doclist for pTerm */
  int *pnDoclist                  /* OUT: Size of doclist in bytes */
){
  unsigned int iHash = fts5HashKey(pHash->nSlot, pTerm, nTerm);
  Fts5HashEntry *p;

  for(p=pHash->aSlot[iHash]; p; p=p->pHashNext){
    if( memcmp(p->zKey, pTerm, nTerm)==0 && p->zKey[nTerm]==0 ) break;
  }

  if( p ){
    fts5HashAddPoslistSize(p);
    *ppDoclist = (const u8*)&p->zKey[nTerm+1];
    *pnDoclist = p->nData - (sizeof(*p) + nTerm + 1);
  }else{
    *ppDoclist = 0;
    *pnDoclist = 0;
  }

  return SQLITE_OK;
}

int sqlite3Fts5HashScanInit(
  Fts5Hash *p,                    /* Hash table to query */
  const char *pTerm, int nTerm    /* Query prefix */
){
  return fts5HashEntrySort(p, pTerm, nTerm, &p->pScan);
}

void sqlite3Fts5HashScanNext(Fts5Hash *p){
  assert( !sqlite3Fts5HashScanEof(p) );
  p->pScan = p->pScan->pScanNext;
}

int sqlite3Fts5HashScanEof(Fts5Hash *p){
  return (p->pScan==0);
}

void sqlite3Fts5HashScanEntry(
  Fts5Hash *pHash,
  const char **pzTerm,            /* OUT: term (nul-terminated) */
  const u8 **ppDoclist,           /* OUT: pointer to doclist */
  int *pnDoclist                  /* OUT: size of doclist in bytes */
){
  Fts5HashEntry *p;
  if( (p = pHash->pScan) ){
    int nTerm = strlen(p->zKey);
    fts5HashAddPoslistSize(p);
    *pzTerm = p->zKey;
    *ppDoclist = (const u8*)&p->zKey[nTerm+1];
    *pnDoclist = p->nData - (sizeof(*p) + nTerm + 1);
  }else{
    *pzTerm = 0;
    *ppDoclist = 0;
    *pnDoclist = 0;
  }
}

#endif /* SQLITE_ENABLE_FTS5 */

/*
** 2014 May 31
**
** The author disclaims copyright to this source code.  In place of
** a legal notice, here is a blessing:
**
**    May you do good and not evil.
**    May you find forgiveness for yourself and forgive others.
**    May you share freely, never taking more than you give.
**
******************************************************************************
**
** Low level access to the FTS index stored in the database file. The 
** routines in this file file implement all read and write access to the
** %_data table. Other parts of the system access this functionality via
** the interface defined in fts5Int.h.
*/

#ifdef SQLITE_ENABLE_FTS5

#include "fts5Int.h"

/*
** Overview:
**
** The %_data table contains all the FTS indexes for an FTS5 virtual table.
** As well as the main term index, there may be up to 31 prefix indexes.
** The format is similar to FTS3/4, except that:
**
**   * all segment b-tree leaf data is stored in fixed size page records 
**     (e.g. 1000 bytes). A single doclist may span multiple pages. Care is 
**     taken to ensure it is possible to iterate in either direction through 
**     the entries in a doclist, or to seek to a specific entry within a 
**     doclist, without loading it into memory.
**
**   * large doclists that span many pages have associated "doclist index"
**     records that contain a copy of the first docid on each page spanned by
**     the doclist. This is used to speed up seek operations, and merges of
**     large doclists with very small doclists.
**
**   * extra fields in the "structure record" record the state of ongoing
**     incremental merge operations.
**
*/


#define FTS5_OPT_WORK_UNIT  1000  /* Number of leaf pages per optimize step */
#define FTS5_WORK_UNIT      64    /* Number of leaf pages in unit of work */

#define FTS5_MIN_DLIDX_SIZE 4     /* Add dlidx if this many empty pages */

#define FTS5_MAIN_PREFIX '0'

#if FTS5_MAX_PREFIX_INDEXES > 31
# error "FTS5_MAX_PREFIX_INDEXES is too large"
#endif

/*
** Details:
**
** The %_data table managed by this module,
**
**     CREATE TABLE %_data(id INTEGER PRIMARY KEY, block BLOB);
**
** , contains the following 5 types of records. See the comments surrounding
** the FTS5_*_ROWID macros below for a description of how %_data rowids are 
** assigned to each fo them.
**
** 1. Structure Records:
**
**   The set of segments that make up an index - the index structure - are
**   recorded in a single record within the %_data table. The record consists
**   of a single 32-bit configuration cookie value followed by a list of 
**   SQLite varints. If the FTS table features more than one index (because
**   there are one or more prefix indexes), it is guaranteed that all share
**   the same cookie value.
**
**   Immediately following the configuration cookie, the record begins with
**   three varints:
**
**     + number of levels,
**     + total number of segments on all levels,
**     + value of write counter.
**
**   Then, for each level from 0 to nMax:
**
**     + number of input segments in ongoing merge.
**     + total number of segments in level.
**     + for each segment from oldest to newest:
**         + segment id (always > 0)
**         + b-tree height (1 -> root is leaf, 2 -> root is parent of leaf etc.)
**         + first leaf page number (often 1, always greater than 0)
**         + final leaf page number
**
** 2. The Averages Record:
**
**   A single record within the %_data table. The data is a list of varints.
**   The first value is the number of rows in the index. Then, for each column
**   from left to right, the total number of tokens in the column for all 
**   rows of the table.
**
** 3. Segment leaves:
**
**   TERM DOCLIST FORMAT:
**
**     Most of each segment leaf is taken up by term/doclist data. The 
**     general format of the term/doclist data is:
**
**         varint : size of first term
**         blob:    first term data
**         doclist: first doclist
**         zero-or-more {
**           varint:  number of bytes in common with previous term
**           varint:  number of bytes of new term data (nNew)
**           blob:    nNew bytes of new term data
**           doclist: next doclist
**         }
**
**     doclist format:
**
**         varint:  first rowid
**         poslist: first poslist
**         zero-or-more {
**           varint:  rowid delta (always > 0)
**           poslist: next poslist
**         }
**         0x00 byte
**
**     poslist format:
**
**         varint: size of poslist in bytes multiplied by 2, not including
**                 this field. Plus 1 if this entry carries the "delete" flag.
**         collist: collist for column 0
**         zero-or-more {
**           0x01 byte
**           varint: column number (I)
**           collist: collist for column I
**         }
**
**     collist format:
**
**         varint: first offset + 2
**         zero-or-more {
**           varint: offset delta + 2
**         }
**
**   PAGINATION
**
**     The format described above is only accurate if the entire term/doclist
**     data fits on a single leaf page. If this is not the case, the format
**     is changed in two ways:
**
**       + if the first rowid on a page occurs before the first term, it
**         is stored as a literal value:
**
**             varint:  first rowid
**
**       + the first term on each page is stored in the same way as the
**         very first term of the segment:
**
**             varint : size of first term
**             blob:    first term data
**
**     Each leaf page begins with:
**
**       + 2-byte unsigned containing offset to first rowid (or 0).
**       + 2-byte unsigned containing offset to first term (or 0).
**
**   Followed by term/doclist data.
**
** 4. Segment interior nodes:
**
**   The interior nodes turn the list of leaves into a b+tree. 
**
**   Each interior node begins with a varint - the page number of the left
**   most child node. Following this, for each leaf page except the first,
**   the interior nodes contain:
**
**     a) If the leaf page contains at least one term, then a term-prefix that
**        is greater than all previous terms, and less than or equal to the
**        first term on the leaf page.
**
**     b) If the leaf page no terms, a record indicating how many consecutive
**        leaves contain no terms, and whether or not there is an associated
**        by-rowid index record.
**
**   By definition, there is never more than one type (b) record in a row.
**   Type (b) records only ever appear on height=1 pages - immediate parents
**   of leaves. Only type (a) records are pushed to higher levels.
**
**   Term format:
**
**     * Number of bytes in common with previous term plus 2, as a varint.
**     * Number of bytes of new term data, as a varint.
**     * new term data.
**
**   No-term format:
**
**     * either an 0x00 or 0x01 byte. If the value 0x01 is used, then there 
**       is an associated index-by-rowid record.
**     * the number of zero-term leaves as a varint.
**
** 5. Segment doclist indexes:
**
**   Doclist indexes are themselves b-trees, however they usually consist of
**   a single leaf record only. The format of each doclist index leaf page 
**   is:
**
**     * Flags byte. Bits are:
**         0x01: Clear if leaf is also the root page, otherwise set.
**
**     * Page number of fts index leaf page. As a varint.
**
**     * First docid on page indicated by previous field. As a varint.
**
**     * A list of varints, one for each subsequent termless page. A 
**       positive delta if the termless page contains at least one docid, 
**       or an 0x00 byte otherwise.
**
**   Internal doclist index nodes are:
**
**     * Flags byte. Bits are:
**         0x01: Clear for root page, otherwise set.
**
**     * Page number of first child page. As a varint.
**
**     * Copy of first docid on page indicated by previous field. As a varint.
**
**     * A list of delta-encoded varints - the first docid on each subsequent
**       child page. 
**
*/

/*
** Rowids for the averages and structure records in the %_data table.
*/
#define FTS5_AVERAGES_ROWID     1    /* Rowid used for the averages record */
#define FTS5_STRUCTURE_ROWID   10    /* The structure record */

/*
** Macros determining the rowids used by segment nodes. All nodes in all
** segments for all indexes (the regular FTS index and any prefix indexes)
** are stored in the %_data table with large positive rowids.
**
** The %_data table may contain up to (1<<FTS5_SEGMENT_INDEX_BITS) 
** indexes - one regular term index and zero or more prefix indexes.
**
** Each segment in an index has a unique id greater than zero.
**
** Each node in a segment b-tree is assigned a "page number" that is unique
** within nodes of its height within the segment (leaf nodes have a height 
** of 0, parents 1, etc.). Page numbers are allocated sequentially so that
** a nodes page number is always one more than its left sibling.
**
** The rowid for a node is then found using the FTS5_SEGMENT_ROWID() macro
** below. The FTS5_SEGMENT_*_BITS macros define the number of bits used
** to encode the three FTS5_SEGMENT_ROWID() arguments. This module returns
** SQLITE_FULL and fails the current operation if they ever prove too small.
*/
#define FTS5_DATA_ID_B     16     /* Max seg id number 65535 */
#define FTS5_DATA_DLI_B     1     /* Doclist-index flag (1 bit) */
#define FTS5_DATA_HEIGHT_B  5     /* Max b-tree height of 32 */
#define FTS5_DATA_PAGE_B   31     /* Max page number of 2147483648 */

#define fts5_dri(segid, dlidx, height, pgno) (                                 \
 ((i64)(segid)  << (FTS5_DATA_PAGE_B+FTS5_DATA_HEIGHT_B+FTS5_DATA_DLI_B)) +    \
 ((i64)(dlidx)  << (FTS5_DATA_PAGE_B + FTS5_DATA_HEIGHT_B)) +                  \
 ((i64)(height) << (FTS5_DATA_PAGE_B)) +                                       \
 ((i64)(pgno))                                                                 \
)

#define FTS5_SEGMENT_ROWID(segid, height, pgno) fts5_dri(segid, 0, height, pgno)
#define FTS5_DLIDX_ROWID(segid, height, pgno)   fts5_dri(segid, 1, height, pgno)

/*
** Maximum segments permitted in a single index 
*/
#define FTS5_MAX_SEGMENT 2000

#ifdef SQLITE_DEBUG
int sqlite3Fts5Corrupt() { return SQLITE_CORRUPT_VTAB; }
#endif


/*
** Each time a blob is read from the %_data table, it is padded with this
** many zero bytes. This makes it easier to decode the various record formats
** without overreading if the records are corrupt.
*/
#define FTS5_DATA_ZERO_PADDING 8

typedef struct Fts5BtreeIter Fts5BtreeIter;
typedef struct Fts5BtreeIterLevel Fts5BtreeIterLevel;
typedef struct Fts5Data Fts5Data;
typedef struct Fts5DlidxIter Fts5DlidxIter;
typedef struct Fts5DlidxLvl Fts5DlidxLvl;
typedef struct Fts5DlidxWriter Fts5DlidxWriter;
typedef struct Fts5MultiSegIter Fts5MultiSegIter;
typedef struct Fts5NodeIter Fts5NodeIter;
typedef struct Fts5PageWriter Fts5PageWriter;
typedef struct Fts5SegIter Fts5SegIter;
typedef struct Fts5DoclistIter Fts5DoclistIter;
typedef struct Fts5SegWriter Fts5SegWriter;
typedef struct Fts5Structure Fts5Structure;
typedef struct Fts5StructureLevel Fts5StructureLevel;
typedef struct Fts5StructureSegment Fts5StructureSegment;

struct Fts5Data {
  u8 *p;                          /* Pointer to buffer containing record */
  int n;                          /* Size of record in bytes */
};

/*
** One object per %_data table.
*/
struct Fts5Index {
  Fts5Config *pConfig;            /* Virtual table configuration */
  char *zDataTbl;                 /* Name of %_data table */
  int nWorkUnit;                  /* Leaf pages in a "unit" of work */

  /*
  ** Variables related to the accumulation of tokens and doclists within the
  ** in-memory hash tables before they are flushed to disk.
  */
  Fts5Hash *pHash;                /* Hash table for in-memory data */
  int nMaxPendingData;            /* Max pending data before flush to disk */
  int nPendingData;               /* Current bytes of pending data */
  i64 iWriteRowid;                /* Rowid for current doc being written */
  Fts5Buffer scratch;

  /* Error state. */
  int rc;                         /* Current error code */

  /* State used by the fts5DataXXX() functions. */
  sqlite3_blob *pReader;          /* RO incr-blob open on %_data table */
  sqlite3_stmt *pWriter;          /* "INSERT ... %_data VALUES(?,?)" */
  sqlite3_stmt *pDeleter;         /* "DELETE FROM %_data ... id>=? AND id<=?" */
  int nRead;                      /* Total number of blocks read */
};

struct Fts5DoclistIter {
  u8 *a;
  int n;
  int i;

  /* Output variables. aPoslist==0 at EOF */
  i64 iRowid;
  u8 *aPoslist;
  int nPoslist;
};

/*
** Each iterator used by external modules is an instance of this type.
*/
struct Fts5IndexIter {
  Fts5Index *pIndex;
  Fts5Structure *pStruct;
  Fts5MultiSegIter *pMulti;
  Fts5Buffer poslist;             /* Buffer containing current poslist */
};

/*
** The contents of the "structure" record for each index are represented
** using an Fts5Structure record in memory. Which uses instances of the 
** other Fts5StructureXXX types as components.
*/
struct Fts5StructureSegment {
  int iSegid;                     /* Segment id */
  int nHeight;                    /* Height of segment b-tree */
  int pgnoFirst;                  /* First leaf page number in segment */
  int pgnoLast;                   /* Last leaf page number in segment */
};
struct Fts5StructureLevel {
  int nMerge;                     /* Number of segments in incr-merge */
  int nSeg;                       /* Total number of segments on level */
  Fts5StructureSegment *aSeg;     /* Array of segments. aSeg[0] is oldest. */
};
struct Fts5Structure {
  u64 nWriteCounter;              /* Total leaves written to level 0 */
  int nSegment;                   /* Total segments in this structure */
  int nLevel;                     /* Number of levels in this index */
  Fts5StructureLevel aLevel[0];   /* Array of nLevel level objects */
};

/*
** An object of type Fts5SegWriter is used to write to segments.
*/
struct Fts5PageWriter {
  int pgno;                       /* Page number for this page */
  Fts5Buffer buf;                 /* Buffer containing page data */
  Fts5Buffer term;                /* Buffer containing previous term on page */
};
struct Fts5DlidxWriter {
  int pgno;                       /* Page number for this page */
  int bPrevValid;                 /* True if iPrev is valid */
  i64 iPrev;                      /* Previous docid value written to page */
  Fts5Buffer buf;                 /* Buffer containing page data */
};
struct Fts5SegWriter {
  int iSegid;                     /* Segid to write to */
  int nWriter;                    /* Number of entries in aWriter */
  Fts5PageWriter *aWriter;        /* Array of PageWriter objects */
  i64 iPrevRowid;                 /* Previous docid written to current leaf */
  u8 bFirstRowidInDoclist;        /* True if next rowid is first in doclist */
  u8 bFirstRowidInPage;           /* True if next rowid is first in page */
  u8 bFirstTermInPage;            /* True if next term will be first in leaf */
  int nLeafWritten;               /* Number of leaf pages written */
  int nEmpty;                     /* Number of contiguous term-less nodes */

  int nDlidx;                     /* Allocated size of aDlidx[] array */
  Fts5DlidxWriter *aDlidx;        /* Array of Fts5DlidxWriter objects */
};

/*
** Object for iterating through the merged results of one or more segments,
** visiting each term/docid pair in the merged data.
**
** nSeg is always a power of two greater than or equal to the number of
** segments that this object is merging data from. Both the aSeg[] and
** aFirst[] arrays are sized at nSeg entries. The aSeg[] array is padded
** with zeroed objects - these are handled as if they were iterators opened
** on empty segments.
**
** The results of comparing segments aSeg[N] and aSeg[N+1], where N is an
** even number, is stored in aFirst[(nSeg+N)/2]. The "result" of the 
** comparison in this context is the index of the iterator that currently
** points to the smaller term/rowid combination. Iterators at EOF are
** considered to be greater than all other iterators.
**
** aFirst[1] contains the index in aSeg[] of the iterator that points to
** the smallest key overall. aFirst[0] is unused. 
*/

typedef struct Fts5CResult Fts5CResult;
struct Fts5CResult {
  u16 iFirst;                     /* aSeg[] index of firstest iterator */
  u8 bTermEq;                     /* True if the terms are equal */
};

struct Fts5MultiSegIter {
  int nSeg;                       /* Size of aSeg[] array */
  int bRev;                       /* True to iterate in reverse order */
  int bSkipEmpty;                 /* True to skip deleted entries */
  Fts5SegIter *aSeg;              /* Array of segment iterators */
  Fts5CResult *aFirst;            /* Current merge state (see above) */
};

/*
** Object for iterating through a single segment, visiting each term/docid
** pair in the segment.
**
** pSeg:
**   The segment to iterate through.
**
** iLeafPgno:
**   Current leaf page number within segment.
**
** iLeafOffset:
**   Byte offset within the current leaf that is the first byte of the 
**   position list data (one byte passed the position-list size field).
**   rowid field of the current entry. Usually this is the size field of the
**   position list data. The exception is if the rowid for the current entry 
**   is the last thing on the leaf page.
**
** pLeaf:
**   Buffer containing current leaf page data. Set to NULL at EOF.
**
** iTermLeafPgno, iTermLeafOffset:
**   Leaf page number containing the last term read from the segment. And
**   the offset immediately following the term data.
**
** flags:
**   Mask of FTS5_SEGITER_XXX values. Interpreted as follows:
**
**   FTS5_SEGITER_ONETERM:
**     If set, set the iterator to point to EOF after the current doclist 
**     has been exhausted. Do not proceed to the next term in the segment.
**
**   FTS5_SEGITER_REVERSE:
**     This flag is only ever set if FTS5_SEGITER_ONETERM is also set. If
**     it is set, iterate through docids in descending order instead of the
**     default ascending order.
**
** iRowidOffset/nRowidOffset/aRowidOffset:
**     These are used if the FTS5_SEGITER_REVERSE flag is set.
**
**     For each rowid on the page corresponding to the current term, the
**     corresponding aRowidOffset[] entry is set to the byte offset of the
**     start of the "position-list-size" field within the page.
*/
struct Fts5SegIter {
  Fts5StructureSegment *pSeg;     /* Segment to iterate through */
  int flags;                      /* Mask of configuration flags */
  int iLeafPgno;                  /* Current leaf page number */
  Fts5Data *pLeaf;                /* Current leaf data */
  Fts5Data *pNextLeaf;            /* Leaf page (iLeafPgno+1) */
  int iLeafOffset;                /* Byte offset within current leaf */

  /* The page and offset from which the current term was read. The offset 
  ** is the offset of the first rowid in the current doclist.  */
  int iTermLeafPgno;
  int iTermLeafOffset;

  /* The following are only used if the FTS5_SEGITER_REVERSE flag is set. */
  int iRowidOffset;               /* Current entry in aRowidOffset[] */
  int nRowidOffset;               /* Allocated size of aRowidOffset[] array */
  int *aRowidOffset;              /* Array of offset to rowid fields */

  Fts5DlidxIter *pDlidx;          /* If there is a doclist-index */

  /* Variables populated based on current entry. */
  Fts5Buffer term;                /* Current term */
  i64 iRowid;                     /* Current rowid */
  int nPos;                       /* Number of bytes in current position list */
  int bDel;                       /* True if the delete flag is set */
};

#define FTS5_SEGITER_ONETERM 0x01
#define FTS5_SEGITER_REVERSE 0x02


/*
** Object for iterating through the conents of a single internal node in 
** memory.
*/
struct Fts5NodeIter {
  /* Internal. Set and managed by fts5NodeIterXXX() functions. Except, 
  ** the EOF test for the iterator is (Fts5NodeIter.aData==0).  */
  const u8 *aData;
  int nData;
  int iOff;

  /* Output variables */
  Fts5Buffer term;
  int nEmpty;
  int iChild;
  int bDlidx;
};

/*
** An instance of the following type is used to iterate through the contents
** of a doclist-index record.
**
** pData:
**   Record containing the doclist-index data.
**
** bEof:
**   Set to true once iterator has reached EOF.
**
** iOff:
**   Set to the current offset within record pData.
*/
struct Fts5DlidxLvl {
  Fts5Data *pData;              /* Data for current page of this level */
  int iOff;                     /* Current offset into pData */
  int bEof;                     /* At EOF already */
  int iFirstOff;                /* Used by reverse iterators */

  /* Output variables */
  int iLeafPgno;                /* Page number of current leaf page */
  i64 iRowid;                   /* First rowid on leaf iLeafPgno */
};
struct Fts5DlidxIter {
  int nLvl;
  int iSegid;
  Fts5DlidxLvl aLvl[1];
};



/*
** An Fts5BtreeIter object is used to iterate through all entries in the
** b-tree hierarchy belonging to a single fts5 segment. In this case the
** "b-tree hierarchy" is all b-tree nodes except leaves. Each entry in the
** b-tree hierarchy consists of the following:
**
**   iLeaf:  The page number of the leaf page the entry points to.
**
**   term:   A split-key that all terms on leaf page $iLeaf must be greater
**           than or equal to. The "term" associated with the first b-tree
**           hierarchy entry (the one that points to leaf page 1) is always 
**           an empty string.
**
**   nEmpty: The number of empty (termless) leaf pages that immediately
**           following iLeaf.
**
** The Fts5BtreeIter object is only used as part of the integrity-check code.
*/
struct Fts5BtreeIterLevel {
  Fts5NodeIter s;                 /* Iterator for the current node */
  Fts5Data *pData;                /* Data for the current node */
};
struct Fts5BtreeIter {
  Fts5Index *p;                   /* FTS5 backend object */
  Fts5StructureSegment *pSeg;     /* Iterate through this segment's b-tree */
  int nLvl;                       /* Size of aLvl[] array */
  Fts5BtreeIterLevel *aLvl;       /* Level for each tier of b-tree */

  /* Output variables */
  Fts5Buffer term;                /* Current term */
  int iLeaf;                      /* Leaf containing terms >= current term */
  int nEmpty;                     /* Number of "empty" leaves following iLeaf */
  int bEof;                       /* Set to true at EOF */
  int bDlidx;                     /* True if there exists a dlidx */
};


static void fts5PutU16(u8 *aOut, u16 iVal){
  aOut[0] = (iVal>>8);
  aOut[1] = (iVal&0xFF);
}

static u16 fts5GetU16(const u8 *aIn){
  return ((u16)aIn[0] << 8) + aIn[1];
} 

/*
** Allocate and return a buffer at least nByte bytes in size.
**
** If an OOM error is encountered, return NULL and set the error code in
** the Fts5Index handle passed as the first argument.
*/
static void *fts5IdxMalloc(Fts5Index *p, int nByte){
  return sqlite3Fts5MallocZero(&p->rc, nByte);
}

/*
** Compare the contents of the pLeft buffer with the pRight/nRight blob.
**
** Return -ve if pLeft is smaller than pRight, 0 if they are equal or
** +ve if pRight is smaller than pLeft. In other words:
**
**     res = *pLeft - *pRight
*/
static int fts5BufferCompareBlob(
  Fts5Buffer *pLeft,              /* Left hand side of comparison */
  const u8 *pRight, int nRight    /* Right hand side of comparison */
){
  int nCmp = MIN(pLeft->n, nRight);
  int res = memcmp(pLeft->p, pRight, nCmp);
  return (res==0 ? (pLeft->n - nRight) : res);
}


/*
** Compare the contents of the two buffers using memcmp(). If one buffer
** is a prefix of the other, it is considered the lesser.
**
** Return -ve if pLeft is smaller than pRight, 0 if they are equal or
** +ve if pRight is smaller than pLeft. In other words:
**
**     res = *pLeft - *pRight
*/
static int fts5BufferCompare(Fts5Buffer *pLeft, Fts5Buffer *pRight){
  int nCmp = MIN(pLeft->n, pRight->n);
  int res = memcmp(pLeft->p, pRight->p, nCmp);
  return (res==0 ? (pLeft->n - pRight->n) : res);
}

#ifdef SQLITE_DEBUG
static int fts5BlobCompare(
  const u8 *pLeft, int nLeft, 
  const u8 *pRight, int nRight
){
  int nCmp = MIN(nLeft, nRight);
  int res = memcmp(pLeft, pRight, nCmp);
  return (res==0 ? (nLeft - nRight) : res);
}
#endif


/*
** Close the read-only blob handle, if it is open.
*/
static void fts5CloseReader(Fts5Index *p){
  if( p->pReader ){
    sqlite3_blob *pReader = p->pReader;
    p->pReader = 0;
    sqlite3_blob_close(pReader);
  }
}

static Fts5Data *fts5DataReadOrBuffer(
  Fts5Index *p, 
  Fts5Buffer *pBuf, 
  i64 iRowid
){
  Fts5Data *pRet = 0;
  if( p->rc==SQLITE_OK ){
    int rc = SQLITE_OK;

    if( p->pReader ){
      /* This call may return SQLITE_ABORT if there has been a savepoint
      ** rollback since it was last used. In this case a new blob handle
      ** is required.  */
      sqlite3_blob *pBlob = p->pReader;
      p->pReader = 0;
      rc = sqlite3_blob_reopen(pBlob, iRowid);
      assert( p->pReader==0 );
      p->pReader = pBlob;
      if( rc!=SQLITE_OK ){
        fts5CloseReader(p);
      }
      if( rc==SQLITE_ABORT ) rc = SQLITE_OK;
    }

    /* If the blob handle is not yet open, open and seek it. Otherwise, use
    ** the blob_reopen() API to reseek the existing blob handle.  */
    if( p->pReader==0 && rc==SQLITE_OK ){
      Fts5Config *pConfig = p->pConfig;
      rc = sqlite3_blob_open(pConfig->db, 
          pConfig->zDb, p->zDataTbl, "block", iRowid, 0, &p->pReader
      );
    }

    /* If either of the sqlite3_blob_open() or sqlite3_blob_reopen() calls
    ** above returned SQLITE_ERROR, return SQLITE_CORRUPT_VTAB instead.
    ** All the reasons those functions might return SQLITE_ERROR - missing
    ** table, missing row, non-blob/text in block column - indicate 
    ** backing store corruption.  */
    if( rc==SQLITE_ERROR ) rc = FTS5_CORRUPT;

    if( rc==SQLITE_OK ){
      u8 *aOut;                   /* Read blob data into this buffer */
      int nByte = sqlite3_blob_bytes(p->pReader);
      if( pBuf ){
        fts5BufferZero(pBuf);
        fts5BufferGrow(&rc, pBuf, nByte);
        aOut = pBuf->p;
        pBuf->n = nByte;
      }else{
        int nSpace = nByte + FTS5_DATA_ZERO_PADDING;
        pRet = (Fts5Data*)sqlite3Fts5MallocZero(&rc, nSpace+sizeof(Fts5Data));
        if( pRet ){
          pRet->n = nByte;
          aOut = pRet->p = (u8*)&pRet[1];
        }
      }

      if( rc==SQLITE_OK ){
        rc = sqlite3_blob_read(p->pReader, aOut, nByte, 0);
      }
      if( rc!=SQLITE_OK ){
        sqlite3_free(pRet);
        pRet = 0;
      }
    }
    p->rc = rc;
    p->nRead++;
  }

  return pRet;
}

/*
** Retrieve a record from the %_data table.
**
** If an error occurs, NULL is returned and an error left in the 
** Fts5Index object.
*/
static Fts5Data *fts5DataRead(Fts5Index *p, i64 iRowid){
  Fts5Data *pRet = fts5DataReadOrBuffer(p, 0, iRowid);
  assert( (pRet==0)==(p->rc!=SQLITE_OK) );
  return pRet;
}

/*
** Read a record from the %_data table into the buffer supplied as the
** second argument.
**
** If an error occurs, an error is left in the Fts5Index object. If an
** error has already occurred when this function is called, it is a 
** no-op.
*/
static void fts5DataBuffer(Fts5Index *p, Fts5Buffer *pBuf, i64 iRowid){
  (void)fts5DataReadOrBuffer(p, pBuf, iRowid);
}

/*
** Release a reference to data record returned by an earlier call to
** fts5DataRead().
*/
static void fts5DataRelease(Fts5Data *pData){
  sqlite3_free(pData);
}

/*
** INSERT OR REPLACE a record into the %_data table.
*/
static void fts5DataWrite(Fts5Index *p, i64 iRowid, const u8 *pData, int nData){
  if( p->rc!=SQLITE_OK ) return;

  if( p->pWriter==0 ){
    int rc = SQLITE_OK;
    Fts5Config *pConfig = p->pConfig;
    char *zSql = sqlite3Fts5Mprintf(&rc,
        "REPLACE INTO '%q'.%Q(id, block) VALUES(?,?)", pConfig->zDb, p->zDataTbl
    );
    if( zSql ){
      rc = sqlite3_prepare_v2(pConfig->db, zSql, -1, &p->pWriter, 0);
      sqlite3_free(zSql);
    }
    if( rc!=SQLITE_OK ){
      p->rc = rc;
      return;
    }
  }

  sqlite3_bind_int64(p->pWriter, 1, iRowid);
  sqlite3_bind_blob(p->pWriter, 2, pData, nData, SQLITE_STATIC);
  sqlite3_step(p->pWriter);
  p->rc = sqlite3_reset(p->pWriter);
}

/*
** Execute the following SQL:
**
**     DELETE FROM %_data WHERE id BETWEEN $iFirst AND $iLast
*/
static void fts5DataDelete(Fts5Index *p, i64 iFirst, i64 iLast){
  if( p->rc!=SQLITE_OK ) return;

  if( p->pDeleter==0 ){
    int rc;
    Fts5Config *pConfig = p->pConfig;
    char *zSql = sqlite3_mprintf(
        "DELETE FROM '%q'.%Q WHERE id>=? AND id<=?", pConfig->zDb, p->zDataTbl
    );
    if( zSql==0 ){
      rc = SQLITE_NOMEM;
    }else{
      rc = sqlite3_prepare_v2(pConfig->db, zSql, -1, &p->pDeleter, 0);
      sqlite3_free(zSql);
    }
    if( rc!=SQLITE_OK ){
      p->rc = rc;
      return;
    }
  }

  sqlite3_bind_int64(p->pDeleter, 1, iFirst);
  sqlite3_bind_int64(p->pDeleter, 2, iLast);
  sqlite3_step(p->pDeleter);
  p->rc = sqlite3_reset(p->pDeleter);
}

/*
** Remove all records associated with segment iSegid.
*/
static void fts5DataRemoveSegment(Fts5Index *p, int iSegid){
  i64 iFirst = FTS5_SEGMENT_ROWID(iSegid, 0, 0);
  i64 iLast = FTS5_SEGMENT_ROWID(iSegid+1, 0, 0)-1;
  fts5DataDelete(p, iFirst, iLast);
}

/*
** Release a reference to an Fts5Structure object returned by an earlier 
** call to fts5StructureRead() or fts5StructureDecode().
*/
static void fts5StructureRelease(Fts5Structure *pStruct){
  if( pStruct ){
    int i;
    for(i=0; i<pStruct->nLevel; i++){
      sqlite3_free(pStruct->aLevel[i].aSeg);
    }
    sqlite3_free(pStruct);
  }
}

/*
** Deserialize and return the structure record currently stored in serialized
** form within buffer pData/nData.
**
** The Fts5Structure.aLevel[] and each Fts5StructureLevel.aSeg[] array
** are over-allocated by one slot. This allows the structure contents
** to be more easily edited.
**
** If an error occurs, *ppOut is set to NULL and an SQLite error code
** returned. Otherwise, *ppOut is set to point to the new object and
** SQLITE_OK returned.
*/
static int fts5StructureDecode(
  const u8 *pData,                /* Buffer containing serialized structure */
  int nData,                      /* Size of buffer pData in bytes */
  int *piCookie,                  /* Configuration cookie value */
  Fts5Structure **ppOut           /* OUT: Deserialized object */
){
  int rc = SQLITE_OK;
  int i = 0;
  int iLvl;
  int nLevel = 0;
  int nSegment = 0;
  int nByte;                      /* Bytes of space to allocate at pRet */
  Fts5Structure *pRet = 0;        /* Structure object to return */

  /* Grab the cookie value */
  if( piCookie ) *piCookie = sqlite3Fts5Get32(pData);
  i = 4;

  /* Read the total number of levels and segments from the start of the
  ** structure record.  */
  i += fts5GetVarint32(&pData[i], nLevel);
  i += fts5GetVarint32(&pData[i], nSegment);
  nByte = (
      sizeof(Fts5Structure) +                    /* Main structure */
      sizeof(Fts5StructureLevel) * (nLevel)      /* aLevel[] array */
  );
  pRet = (Fts5Structure*)sqlite3Fts5MallocZero(&rc, nByte);

  if( pRet ){
    pRet->nLevel = nLevel;
    pRet->nSegment = nSegment;
    i += sqlite3Fts5GetVarint(&pData[i], &pRet->nWriteCounter);

    for(iLvl=0; rc==SQLITE_OK && iLvl<nLevel; iLvl++){
      Fts5StructureLevel *pLvl = &pRet->aLevel[iLvl];
      int nTotal;
      int iSeg;

      i += fts5GetVarint32(&pData[i], pLvl->nMerge);
      i += fts5GetVarint32(&pData[i], nTotal);
      assert( nTotal>=pLvl->nMerge );
      pLvl->aSeg = (Fts5StructureSegment*)sqlite3Fts5MallocZero(&rc, 
          nTotal * sizeof(Fts5StructureSegment)
      );

      if( rc==SQLITE_OK ){
        pLvl->nSeg = nTotal;
        for(iSeg=0; iSeg<nTotal; iSeg++){
          i += fts5GetVarint32(&pData[i], pLvl->aSeg[iSeg].iSegid);
          i += fts5GetVarint32(&pData[i], pLvl->aSeg[iSeg].nHeight);
          i += fts5GetVarint32(&pData[i], pLvl->aSeg[iSeg].pgnoFirst);
          i += fts5GetVarint32(&pData[i], pLvl->aSeg[iSeg].pgnoLast);
        }
      }else{
        fts5StructureRelease(pRet);
        pRet = 0;
      }
    }
  }

  *ppOut = pRet;
  return rc;
}

/*
**
*/
static void fts5StructureAddLevel(int *pRc, Fts5Structure **ppStruct){
  if( *pRc==SQLITE_OK ){
    Fts5Structure *pStruct = *ppStruct;
    int nLevel = pStruct->nLevel;
    int nByte = (
        sizeof(Fts5Structure) +                  /* Main structure */
        sizeof(Fts5StructureLevel) * (nLevel+1)  /* aLevel[] array */
    );

    pStruct = sqlite3_realloc(pStruct, nByte);
    if( pStruct ){
      memset(&pStruct->aLevel[nLevel], 0, sizeof(Fts5StructureLevel));
      pStruct->nLevel++;
      *ppStruct = pStruct;
    }else{
      *pRc = SQLITE_NOMEM;
    }
  }
}

/*
** Extend level iLvl so that there is room for at least nExtra more
** segments.
*/
static void fts5StructureExtendLevel(
  int *pRc, 
  Fts5Structure *pStruct, 
  int iLvl, 
  int nExtra, 
  int bInsert
){
  if( *pRc==SQLITE_OK ){
    Fts5StructureLevel *pLvl = &pStruct->aLevel[iLvl];
    Fts5StructureSegment *aNew;
    int nByte;

    nByte = (pLvl->nSeg + nExtra) * sizeof(Fts5StructureSegment);
    aNew = sqlite3_realloc(pLvl->aSeg, nByte);
    if( aNew ){
      if( bInsert==0 ){
        memset(&aNew[pLvl->nSeg], 0, sizeof(Fts5StructureSegment) * nExtra);
      }else{
        int nMove = pLvl->nSeg * sizeof(Fts5StructureSegment);
        memmove(&aNew[nExtra], aNew, nMove);
        memset(aNew, 0, sizeof(Fts5StructureSegment) * nExtra);
      }
      pLvl->aSeg = aNew;
    }else{
      *pRc = SQLITE_NOMEM;
    }
  }
}

/*
** Read, deserialize and return the structure record.
**
** The Fts5Structure.aLevel[] and each Fts5StructureLevel.aSeg[] array
** are over-allocated as described for function fts5StructureDecode() 
** above.
**
** If an error occurs, NULL is returned and an error code left in the
** Fts5Index handle. If an error has already occurred when this function
** is called, it is a no-op.
*/
static Fts5Structure *fts5StructureRead(Fts5Index *p){
  Fts5Config *pConfig = p->pConfig;
  Fts5Structure *pRet = 0;        /* Object to return */
  Fts5Data *pData;                /* %_data entry containing structure record */
  int iCookie;                    /* Configuration cookie */

  pData = fts5DataRead(p, FTS5_STRUCTURE_ROWID);
  if( !pData ) return 0;
  p->rc = fts5StructureDecode(pData->p, pData->n, &iCookie, &pRet);

  if( p->rc==SQLITE_OK && pConfig->iCookie!=iCookie ){
    p->rc = sqlite3Fts5ConfigLoad(pConfig, iCookie);
  }

  fts5DataRelease(pData);
  if( p->rc!=SQLITE_OK ){
    fts5StructureRelease(pRet);
    pRet = 0;
  }
  return pRet;
}

/*
** Return the total number of segments in index structure pStruct. This
** function is only ever used as part of assert() conditions.
*/
#ifdef SQLITE_DEBUG
static int fts5StructureCountSegments(Fts5Structure *pStruct){
  int nSegment = 0;               /* Total number of segments */
  if( pStruct ){
    int iLvl;                     /* Used to iterate through levels */
    for(iLvl=0; iLvl<pStruct->nLevel; iLvl++){
      nSegment += pStruct->aLevel[iLvl].nSeg;
    }
  }

  return nSegment;
}
#endif

/*
** Serialize and store the "structure" record.
**
** If an error occurs, leave an error code in the Fts5Index object. If an
** error has already occurred, this function is a no-op.
*/
static void fts5StructureWrite(Fts5Index *p, Fts5Structure *pStruct){
  if( p->rc==SQLITE_OK ){
    Fts5Buffer buf;               /* Buffer to serialize record into */
    int iLvl;                     /* Used to iterate through levels */
    int iCookie;                  /* Cookie value to store */

    assert( pStruct->nSegment==fts5StructureCountSegments(pStruct) );
    memset(&buf, 0, sizeof(Fts5Buffer));

    /* Append the current configuration cookie */
    iCookie = p->pConfig->iCookie;
    if( iCookie<0 ) iCookie = 0;
    fts5BufferAppend32(&p->rc, &buf, iCookie);

    fts5BufferAppendVarint(&p->rc, &buf, pStruct->nLevel);
    fts5BufferAppendVarint(&p->rc, &buf, pStruct->nSegment);
    fts5BufferAppendVarint(&p->rc, &buf, (i64)pStruct->nWriteCounter);

    for(iLvl=0; iLvl<pStruct->nLevel; iLvl++){
      int iSeg;                     /* Used to iterate through segments */
      Fts5StructureLevel *pLvl = &pStruct->aLevel[iLvl];
      fts5BufferAppendVarint(&p->rc, &buf, pLvl->nMerge);
      fts5BufferAppendVarint(&p->rc, &buf, pLvl->nSeg);
      assert( pLvl->nMerge<=pLvl->nSeg );

      for(iSeg=0; iSeg<pLvl->nSeg; iSeg++){
        fts5BufferAppendVarint(&p->rc, &buf, pLvl->aSeg[iSeg].iSegid);
        fts5BufferAppendVarint(&p->rc, &buf, pLvl->aSeg[iSeg].nHeight);
        fts5BufferAppendVarint(&p->rc, &buf, pLvl->aSeg[iSeg].pgnoFirst);
        fts5BufferAppendVarint(&p->rc, &buf, pLvl->aSeg[iSeg].pgnoLast);
      }
    }

    fts5DataWrite(p, FTS5_STRUCTURE_ROWID, buf.p, buf.n);
    fts5BufferFree(&buf);
  }
}

#if 0
static void fts5DebugStructure(int*,Fts5Buffer*,Fts5Structure*);
static void fts5PrintStructure(const char *zCaption, Fts5Structure *pStruct){
  int rc = SQLITE_OK;
  Fts5Buffer buf;
  memset(&buf, 0, sizeof(buf));
  fts5DebugStructure(&rc, &buf, pStruct);
  fprintf(stdout, "%s: %s\n", zCaption, buf.p);
  fflush(stdout);
  fts5BufferFree(&buf);
}
#else
# define fts5PrintStructure(x,y)
#endif

static int fts5SegmentSize(Fts5StructureSegment *pSeg){
  return 1 + pSeg->pgnoLast - pSeg->pgnoFirst;
}

/*
** Return a copy of index structure pStruct. Except, promote as many 
** segments as possible to level iPromote. If an OOM occurs, NULL is 
** returned.
*/
static void fts5StructurePromoteTo(
  Fts5Index *p,
  int iPromote,
  int szPromote,
  Fts5Structure *pStruct
){
  int il, is;
  Fts5StructureLevel *pOut = &pStruct->aLevel[iPromote];

  if( pOut->nMerge==0 ){
    for(il=iPromote+1; il<pStruct->nLevel; il++){
      Fts5StructureLevel *pLvl = &pStruct->aLevel[il];
      if( pLvl->nMerge ) return;
      for(is=pLvl->nSeg-1; is>=0; is--){
        int sz = fts5SegmentSize(&pLvl->aSeg[is]);
        if( sz>szPromote ) return;
        fts5StructureExtendLevel(&p->rc, pStruct, iPromote, 1, 1);
        if( p->rc ) return;
        memcpy(pOut->aSeg, &pLvl->aSeg[is], sizeof(Fts5StructureSegment));
        pOut->nSeg++;
        pLvl->nSeg--;
      }
    }
  }
}

/*
** A new segment has just been written to level iLvl of index structure
** pStruct. This function determines if any segments should be promoted
** as a result. Segments are promoted in two scenarios:
**
**   a) If the segment just written is smaller than one or more segments
**      within the previous populated level, it is promoted to the previous
**      populated level.
**
**   b) If the segment just written is larger than the newest segment on
**      the next populated level, then that segment, and any other adjacent
**      segments that are also smaller than the one just written, are 
**      promoted. 
**
** If one or more segments are promoted, the structure object is updated
** to reflect this.
*/
static void fts5StructurePromote(
  Fts5Index *p,                   /* FTS5 backend object */
  int iLvl,                       /* Index level just updated */
  Fts5Structure *pStruct          /* Index structure */
){
  if( p->rc==SQLITE_OK ){
    int iTst;
    int iPromote = -1;
    int szPromote;                /* Promote anything this size or smaller */
    Fts5StructureSegment *pSeg;   /* Segment just written */
    int szSeg;                    /* Size of segment just written */


    pSeg = &pStruct->aLevel[iLvl].aSeg[pStruct->aLevel[iLvl].nSeg-1];
    szSeg = (1 + pSeg->pgnoLast - pSeg->pgnoFirst);

    /* Check for condition (a) */
    for(iTst=iLvl-1; iTst>=0 && pStruct->aLevel[iTst].nSeg==0; iTst--);
    if( iTst>=0 ){
      int i;
      int szMax = 0;
      Fts5StructureLevel *pTst = &pStruct->aLevel[iTst];
      assert( pTst->nMerge==0 );
      for(i=0; i<pTst->nSeg; i++){
        int sz = pTst->aSeg[i].pgnoLast - pTst->aSeg[i].pgnoFirst + 1;
        if( sz>szMax ) szMax = sz;
      }
      if( szMax>=szSeg ){
        /* Condition (a) is true. Promote the newest segment on level 
        ** iLvl to level iTst.  */
        iPromote = iTst;
        szPromote = szMax;
      }
    }

    /* If condition (a) is not met, assume (b) is true. StructurePromoteTo()
    ** is a no-op if it is not.  */
    if( iPromote<0 ){
      iPromote = iLvl;
      szPromote = szSeg;
    }
    fts5StructurePromoteTo(p, iPromote, szPromote, pStruct);
  }
}


/*
** If the pIter->iOff offset currently points to an entry indicating one
** or more term-less nodes, advance past it and set pIter->nEmpty to
** the number of empty child nodes.
*/
static void fts5NodeIterGobbleNEmpty(Fts5NodeIter *pIter){
  if( pIter->iOff<pIter->nData && 0==(pIter->aData[pIter->iOff] & 0xfe) ){
    pIter->bDlidx = pIter->aData[pIter->iOff] & 0x01;
    pIter->iOff++;
    pIter->iOff += fts5GetVarint32(&pIter->aData[pIter->iOff], pIter->nEmpty);
  }else{
    pIter->nEmpty = 0;
    pIter->bDlidx = 0;
  }
}

/*
** Advance to the next entry within the node.
*/
static void fts5NodeIterNext(int *pRc, Fts5NodeIter *pIter){
  if( pIter->iOff>=pIter->nData ){
    pIter->aData = 0;
    pIter->iChild += pIter->nEmpty;
  }else{
    int nPre, nNew;
    pIter->iOff += fts5GetVarint32(&pIter->aData[pIter->iOff], nPre);
    pIter->iOff += fts5GetVarint32(&pIter->aData[pIter->iOff], nNew);
    pIter->term.n = nPre-2;
    fts5BufferAppendBlob(pRc, &pIter->term, nNew, pIter->aData+pIter->iOff);
    pIter->iOff += nNew;
    pIter->iChild += (1 + pIter->nEmpty);
    fts5NodeIterGobbleNEmpty(pIter);
    if( *pRc ) pIter->aData = 0;
  }
}


/*
** Initialize the iterator object pIter to iterate through the internal
** segment node in pData.
*/
static void fts5NodeIterInit(const u8 *aData, int nData, Fts5NodeIter *pIter){
  memset(pIter, 0, sizeof(*pIter));
  pIter->aData = aData;
  pIter->nData = nData;
  pIter->iOff = fts5GetVarint32(aData, pIter->iChild);
  fts5NodeIterGobbleNEmpty(pIter);
}

/*
** Free any memory allocated by the iterator object.
*/
static void fts5NodeIterFree(Fts5NodeIter *pIter){
  fts5BufferFree(&pIter->term);
}

/*
** Advance the iterator passed as the only argument. If the end of the 
** doclist-index page is reached, return non-zero.
*/
static int fts5DlidxLvlNext(Fts5DlidxLvl *pLvl){
  Fts5Data *pData = pLvl->pData;

  if( pLvl->iOff==0 ){
    assert( pLvl->bEof==0 );
    pLvl->iOff = 1;
    pLvl->iOff += fts5GetVarint32(&pData->p[1], pLvl->iLeafPgno);
    pLvl->iOff += fts5GetVarint(&pData->p[pLvl->iOff], (u64*)&pLvl->iRowid);
    pLvl->iFirstOff = pLvl->iOff;
  }else{
    int iOff;
    for(iOff=pLvl->iOff; iOff<pData->n; iOff++){
      if( pData->p[iOff] ) break; 
    }

    if( iOff<pData->n ){
      i64 iVal;
      pLvl->iLeafPgno += (iOff - pLvl->iOff) + 1;
      iOff += fts5GetVarint(&pData->p[iOff], (u64*)&iVal);
      pLvl->iRowid += iVal;
      pLvl->iOff = iOff;
    }else{
      pLvl->bEof = 1;
    }
  }

  return pLvl->bEof;
}

/*
** Advance the iterator passed as the only argument.
*/
static int fts5DlidxIterNextR(Fts5Index *p, Fts5DlidxIter *pIter, int iLvl){
  Fts5DlidxLvl *pLvl = &pIter->aLvl[iLvl];

  assert( iLvl<pIter->nLvl );
  if( fts5DlidxLvlNext(pLvl) ){
    if( (iLvl+1) < pIter->nLvl ){
      fts5DlidxIterNextR(p, pIter, iLvl+1);
      if( pLvl[1].bEof==0 ){
        fts5DataRelease(pLvl->pData);
        memset(pLvl, 0, sizeof(Fts5DlidxLvl));
        pLvl->pData = fts5DataRead(p, 
            FTS5_DLIDX_ROWID(pIter->iSegid, iLvl, pLvl[1].iLeafPgno)
        );
        if( pLvl->pData ) fts5DlidxLvlNext(pLvl);
      }
    }
  }

  return pIter->aLvl[0].bEof;
}
static int fts5DlidxIterNext(Fts5Index *p, Fts5DlidxIter *pIter){
  return fts5DlidxIterNextR(p, pIter, 0);
}

/*
** The iterator passed as the first argument has the following fields set
** as follows. This function sets up the rest of the iterator so that it
** points to the first rowid in the doclist-index.
**
**   pData:
**     pointer to doclist-index record, 
**
** When this function is called pIter->iLeafPgno is the page number the
** doclist is associated with (the one featuring the term).
*/
static int fts5DlidxIterFirst(Fts5DlidxIter *pIter){
  int i;
  for(i=0; i<pIter->nLvl; i++){
    fts5DlidxLvlNext(&pIter->aLvl[i]);
  }
  return pIter->aLvl[0].bEof;
}


static int fts5DlidxIterEof(Fts5Index *p, Fts5DlidxIter *pIter){
  return p->rc!=SQLITE_OK || pIter->aLvl[0].bEof;
}

static void fts5DlidxIterLast(Fts5Index *p, Fts5DlidxIter *pIter){
  int i;

  /* Advance each level to the last entry on the last page */
  for(i=pIter->nLvl-1; p->rc==SQLITE_OK && i>=0; i--){
    Fts5DlidxLvl *pLvl = &pIter->aLvl[i];
    while( fts5DlidxLvlNext(pLvl)==0 );
    pLvl->bEof = 0;

    if( i>0 ){
      Fts5DlidxLvl *pChild = &pLvl[-1];
      fts5DataRelease(pChild->pData);
      memset(pChild, 0, sizeof(Fts5DlidxLvl));
      pChild->pData = fts5DataRead(p, 
          FTS5_DLIDX_ROWID(pIter->iSegid, i-1, pLvl->iLeafPgno)
      );
    }
  }
}

/*
** Move the iterator passed as the only argument to the previous entry.
*/
static int fts5DlidxLvlPrev(Fts5DlidxLvl *pLvl){
  int iOff = pLvl->iOff;

  assert( pLvl->bEof==0 );
  if( iOff<=pLvl->iFirstOff ){
    pLvl->bEof = 1;
  }else{
    u8 *a = pLvl->pData->p;
    i64 iVal;
    int iLimit;
    int ii;
    int nZero = 0;

    /* Currently iOff points to the first byte of a varint. This block 
    ** decrements iOff until it points to the first byte of the previous 
    ** varint. Taking care not to read any memory locations that occur
    ** before the buffer in memory.  */
    iLimit = (iOff>9 ? iOff-9 : 0);
    for(iOff--; iOff>iLimit; iOff--){
      if( (a[iOff-1] & 0x80)==0 ) break;
    }

    fts5GetVarint(&a[iOff], (u64*)&iVal);
    pLvl->iRowid -= iVal;
    pLvl->iLeafPgno--;

    /* Skip backwards past any 0x00 varints. */
    for(ii=iOff-1; ii>=pLvl->iFirstOff && a[ii]==0x00; ii--){
      nZero++;
    }
    if( ii>=pLvl->iFirstOff && (a[ii] & 0x80) ){
      /* The byte immediately before the last 0x00 byte has the 0x80 bit
      ** set. So the last 0x00 is only a varint 0 if there are 8 more 0x80
      ** bytes before a[ii]. */
      int bZero = 0;              /* True if last 0x00 counts */
      if( (ii-8)>=pLvl->iFirstOff ){
        int j;
        for(j=1; j<=8 && (a[ii-j] & 0x80); j++);
        bZero = (j>8);
      }
      if( bZero==0 ) nZero--;
    }
    pLvl->iLeafPgno -= nZero;
    pLvl->iOff = iOff - nZero;
  }

  return pLvl->bEof;
}

static int fts5DlidxIterPrevR(Fts5Index *p, Fts5DlidxIter *pIter, int iLvl){
  Fts5DlidxLvl *pLvl = &pIter->aLvl[iLvl];

  assert( iLvl<pIter->nLvl );
  if( fts5DlidxLvlPrev(pLvl) ){
    if( (iLvl+1) < pIter->nLvl ){
      fts5DlidxIterPrevR(p, pIter, iLvl+1);
      if( pLvl[1].bEof==0 ){
        fts5DataRelease(pLvl->pData);
        memset(pLvl, 0, sizeof(Fts5DlidxLvl));
        pLvl->pData = fts5DataRead(p, 
            FTS5_DLIDX_ROWID(pIter->iSegid, iLvl, pLvl[1].iLeafPgno)
        );
        if( pLvl->pData ){
          while( fts5DlidxLvlNext(pLvl)==0 );
          pLvl->bEof = 0;
        }
      }
    }
  }

  return pIter->aLvl[0].bEof;
}
static int fts5DlidxIterPrev(Fts5Index *p, Fts5DlidxIter *pIter){
  return fts5DlidxIterPrevR(p, pIter, 0);
}

/*
** Free a doclist-index iterator object allocated by fts5DlidxIterInit().
*/
static void fts5DlidxIterFree(Fts5DlidxIter *pIter){
  if( pIter ){
    int i;
    for(i=0; i<pIter->nLvl; i++){
      fts5DataRelease(pIter->aLvl[i].pData);
    }
    sqlite3_free(pIter);
  }
}

static Fts5DlidxIter *fts5DlidxIterInit(
  Fts5Index *p,                   /* Fts5 Backend to iterate within */
  int bRev,                       /* True for ORDER BY ASC */
  int iSegid,                     /* Segment id */
  int iLeafPg                     /* Leaf page number to load dlidx for */
){
  Fts5DlidxIter *pIter = 0;
  int i;
  int bDone = 0;

  for(i=0; p->rc==SQLITE_OK && bDone==0; i++){
    int nByte = sizeof(Fts5DlidxIter) + i * sizeof(Fts5DlidxLvl);
    Fts5DlidxIter *pNew;

    pNew = (Fts5DlidxIter*)sqlite3_realloc(pIter, nByte);
    if( pNew==0 ){
      p->rc = SQLITE_NOMEM;
    }else{
      i64 iRowid = FTS5_DLIDX_ROWID(iSegid, i, iLeafPg);
      Fts5DlidxLvl *pLvl = &pNew->aLvl[i];
      pIter = pNew;
      memset(pLvl, 0, sizeof(Fts5DlidxLvl));
      pLvl->pData = fts5DataRead(p, iRowid);
      if( pLvl->pData && (pLvl->pData->p[0] & 0x0001)==0 ){
        bDone = 1;
      }
      pIter->nLvl = i+1;
    }
  }

  if( p->rc==SQLITE_OK ){
    pIter->iSegid = iSegid;
    if( bRev==0 ){
      fts5DlidxIterFirst(pIter);
    }else{
      fts5DlidxIterLast(p, pIter);
    }
  }

  if( p->rc!=SQLITE_OK ){
    fts5DlidxIterFree(pIter);
    pIter = 0;
  }

  return pIter;
}

static i64 fts5DlidxIterRowid(Fts5DlidxIter *pIter){
  return pIter->aLvl[0].iRowid;
}
static int fts5DlidxIterPgno(Fts5DlidxIter *pIter){
  return pIter->aLvl[0].iLeafPgno;
}

static void fts5LeafHeader(Fts5Data *pLeaf, int *piRowid, int *piTerm){
  *piRowid = (int)fts5GetU16(&pLeaf->p[0]);
  *piTerm = (int)fts5GetU16(&pLeaf->p[2]);
}

/*
** Load the next leaf page into the segment iterator.
*/
static void fts5SegIterNextPage(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5SegIter *pIter              /* Iterator to advance to next page */
){
  Fts5StructureSegment *pSeg = pIter->pSeg;
  fts5DataRelease(pIter->pLeaf);
  pIter->iLeafPgno++;
  if( pIter->pNextLeaf ){
    assert( pIter->iLeafPgno<=pSeg->pgnoLast );
    pIter->pLeaf = pIter->pNextLeaf;
    pIter->pNextLeaf = 0;
  }else if( pIter->iLeafPgno<=pSeg->pgnoLast ){
    pIter->pLeaf = fts5DataRead(p, 
        FTS5_SEGMENT_ROWID(pSeg->iSegid, 0, pIter->iLeafPgno)
    );
  }else{
    pIter->pLeaf = 0;
  }
}

/*
** Argument p points to a buffer containing a varint to be interpreted as a
** position list size field. Read the varint and return the number of bytes
** read. Before returning, set *pnSz to the number of bytes in the position
** list, and *pbDel to true if the delete flag is set, or false otherwise.
*/
static int fts5GetPoslistSize(const u8 *p, int *pnSz, int *pbDel){
  int nSz;
  int n = fts5GetVarint32(p, nSz);
  assert_nc( nSz>=0 );
  *pnSz = nSz/2;
  *pbDel = nSz & 0x0001;
  return n;
}

/*
** Fts5SegIter.iLeafOffset currently points to the first byte of a
** position-list size field. Read the value of the field and store it
** in the following variables:
**
**   Fts5SegIter.nPos
**   Fts5SegIter.bDel
**
** Leave Fts5SegIter.iLeafOffset pointing to the first byte of the 
** position list content (if any).
*/
static void fts5SegIterLoadNPos(Fts5Index *p, Fts5SegIter *pIter){
  if( p->rc==SQLITE_OK ){
    int iOff = pIter->iLeafOffset;  /* Offset to read at */
    if( iOff>=pIter->pLeaf->n ){
      p->rc = FTS5_CORRUPT;
    }else{
      const u8 *a = &pIter->pLeaf->p[iOff];
      pIter->iLeafOffset += fts5GetPoslistSize(a, &pIter->nPos, &pIter->bDel);
    }
  }
}

/*
** Fts5SegIter.iLeafOffset currently points to the first byte of the 
** "nSuffix" field of a term. Function parameter nKeep contains the value
** of the "nPrefix" field (if there was one - it is passed 0 if this is
** the first term in the segment).
**
** This function populates:
**
**   Fts5SegIter.term
**   Fts5SegIter.rowid
**
** accordingly and leaves (Fts5SegIter.iLeafOffset) set to the content of
** the first position list. The position list belonging to document 
** (Fts5SegIter.iRowid).
*/
static void fts5SegIterLoadTerm(Fts5Index *p, Fts5SegIter *pIter, int nKeep){
  u8 *a = pIter->pLeaf->p;        /* Buffer to read data from */
  int iOff = pIter->iLeafOffset;  /* Offset to read at */
  int nNew;                       /* Bytes of new data */

  iOff += fts5GetVarint32(&a[iOff], nNew);
  pIter->term.n = nKeep;
  fts5BufferAppendBlob(&p->rc, &pIter->term, nNew, &a[iOff]);
  iOff += nNew;
  pIter->iTermLeafOffset = iOff;
  pIter->iTermLeafPgno = pIter->iLeafPgno;
  if( iOff>=pIter->pLeaf->n ){
    fts5SegIterNextPage(p, pIter);
    if( pIter->pLeaf==0 ){
      if( p->rc==SQLITE_OK ) p->rc = FTS5_CORRUPT;
      return;
    }
    iOff = 4;
    a = pIter->pLeaf->p;
  }
  iOff += sqlite3Fts5GetVarint(&a[iOff], (u64*)&pIter->iRowid);
  pIter->iLeafOffset = iOff;
}

/*
** Initialize the iterator object pIter to iterate through the entries in
** segment pSeg. The iterator is left pointing to the first entry when 
** this function returns.
**
** If an error occurs, Fts5Index.rc is set to an appropriate error code. If 
** an error has already occurred when this function is called, it is a no-op.
*/
static void fts5SegIterInit(
  Fts5Index *p,                   /* FTS index object */
  Fts5StructureSegment *pSeg,     /* Description of segment */
  Fts5SegIter *pIter              /* Object to populate */
){
  if( pSeg->pgnoFirst==0 ){
    /* This happens if the segment is being used as an input to an incremental
    ** merge and all data has already been "trimmed". See function
    ** fts5TrimSegments() for details. In this case leave the iterator empty.
    ** The caller will see the (pIter->pLeaf==0) and assume the iterator is
    ** at EOF already. */
    assert( pIter->pLeaf==0 );
    return;
  }

  if( p->rc==SQLITE_OK ){
    memset(pIter, 0, sizeof(*pIter));
    pIter->pSeg = pSeg;
    pIter->iLeafPgno = pSeg->pgnoFirst-1;
    fts5SegIterNextPage(p, pIter);
  }

  if( p->rc==SQLITE_OK ){
    u8 *a = pIter->pLeaf->p;
    pIter->iLeafOffset = fts5GetU16(&a[2]);
    fts5SegIterLoadTerm(p, pIter, 0);
    fts5SegIterLoadNPos(p, pIter);
  }
}

/*
** This function is only ever called on iterators created by calls to
** Fts5IndexQuery() with the FTS5INDEX_QUERY_DESC flag set.
**
** The iterator is in an unusual state when this function is called: the
** Fts5SegIter.iLeafOffset variable is set to the offset of the start of
** the position-list size field for the first relevant rowid on the page.
** Fts5SegIter.rowid is set, but nPos and bDel are not.
**
** This function advances the iterator so that it points to the last 
** relevant rowid on the page and, if necessary, initializes the 
** aRowidOffset[] and iRowidOffset variables. At this point the iterator
** is in its regular state - Fts5SegIter.iLeafOffset points to the first
** byte of the position list content associated with said rowid.
*/
static void fts5SegIterReverseInitPage(Fts5Index *p, Fts5SegIter *pIter){
  int n = pIter->pLeaf->n;
  int i = pIter->iLeafOffset;
  u8 *a = pIter->pLeaf->p;
  int iRowidOffset = 0;

  while( 1 ){
    i64 iDelta = 0;
    int nPos;
    int bDummy;

    i += fts5GetPoslistSize(&a[i], &nPos, &bDummy);
    i += nPos;
    if( i>=n ) break;
    i += fts5GetVarint(&a[i], (u64*)&iDelta);
    if( iDelta==0 ) break;
    pIter->iRowid += iDelta;

    if( iRowidOffset>=pIter->nRowidOffset ){
      int nNew = pIter->nRowidOffset + 8;
      int *aNew = (int*)sqlite3_realloc(pIter->aRowidOffset, nNew*sizeof(int));
      if( aNew==0 ){
        p->rc = SQLITE_NOMEM;
        break;
      }
      pIter->aRowidOffset = aNew;
      pIter->nRowidOffset = nNew;
    }

    pIter->aRowidOffset[iRowidOffset++] = pIter->iLeafOffset;
    pIter->iLeafOffset = i;
  }
  pIter->iRowidOffset = iRowidOffset;
  fts5SegIterLoadNPos(p, pIter);
}

/*
**
*/
static void fts5SegIterReverseNewPage(Fts5Index *p, Fts5SegIter *pIter){
  assert( pIter->flags & FTS5_SEGITER_REVERSE );
  assert( pIter->flags & FTS5_SEGITER_ONETERM );

  fts5DataRelease(pIter->pLeaf);
  pIter->pLeaf = 0;
  while( p->rc==SQLITE_OK && pIter->iLeafPgno>pIter->iTermLeafPgno ){
    Fts5Data *pNew;
    pIter->iLeafPgno--;
    pNew = fts5DataRead(p, FTS5_SEGMENT_ROWID(
          pIter->pSeg->iSegid, 0, pIter->iLeafPgno
    ));
    if( pNew ){
      if( pIter->iLeafPgno==pIter->iTermLeafPgno ){
        if( pIter->iTermLeafOffset<pNew->n ){
          pIter->pLeaf = pNew;
          pIter->iLeafOffset = pIter->iTermLeafOffset;
        }
      }else{
        int iRowidOff, dummy;
        fts5LeafHeader(pNew, &iRowidOff, &dummy);
        if( iRowidOff ){
          pIter->pLeaf = pNew;
          pIter->iLeafOffset = iRowidOff;
        }
      }

      if( pIter->pLeaf ){
        u8 *a = &pIter->pLeaf->p[pIter->iLeafOffset];
        pIter->iLeafOffset += fts5GetVarint(a, (u64*)&pIter->iRowid);
        break;
      }else{
        fts5DataRelease(pNew);
      }
    }
  }

  if( pIter->pLeaf ){
    fts5SegIterReverseInitPage(p, pIter);
  }
}

/*
** Return true if the iterator passed as the second argument currently
** points to a delete marker. A delete marker is an entry with a 0 byte
** position-list.
*/
static int fts5MultiIterIsEmpty(Fts5Index *p, Fts5MultiSegIter *pIter){
  Fts5SegIter *pSeg = &pIter->aSeg[pIter->aFirst[1].iFirst];
  return (p->rc==SQLITE_OK && pSeg->pLeaf && pSeg->nPos==0);
}

/*
** Advance iterator pIter to the next entry. 
**
** If an error occurs, Fts5Index.rc is set to an appropriate error code. It 
** is not considered an error if the iterator reaches EOF. If an error has 
** already occurred when this function is called, it is a no-op.
*/
static void fts5SegIterNext(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5SegIter *pIter,             /* Iterator to advance */
  int *pbNewTerm                  /* OUT: Set for new term */
){
  assert( pbNewTerm==0 || *pbNewTerm==0 );
  if( p->rc==SQLITE_OK ){
    if( pIter->flags & FTS5_SEGITER_REVERSE ){
      assert( pIter->pNextLeaf==0 );
      if( pIter->iRowidOffset>0 ){
        u8 *a = pIter->pLeaf->p;
        int iOff;
        int nPos;
        int bDummy;
        i64 iDelta;

        pIter->iRowidOffset--;
        pIter->iLeafOffset = iOff = pIter->aRowidOffset[pIter->iRowidOffset];
        iOff += fts5GetPoslistSize(&a[iOff], &nPos, &bDummy);
        iOff += nPos;
        fts5GetVarint(&a[iOff], (u64*)&iDelta);
        pIter->iRowid -= iDelta;
        fts5SegIterLoadNPos(p, pIter);
      }else{
        fts5SegIterReverseNewPage(p, pIter);
      }
    }else{
      Fts5Data *pLeaf = pIter->pLeaf;
      int iOff;
      int bNewTerm = 0;
      int nKeep = 0;

      /* Search for the end of the position list within the current page. */
      u8 *a = pLeaf->p;
      int n = pLeaf->n;

      iOff = pIter->iLeafOffset + pIter->nPos;

      if( iOff<n ){
        /* The next entry is on the current page */
        u64 iDelta;
        iOff += sqlite3Fts5GetVarint(&a[iOff], &iDelta);
        pIter->iLeafOffset = iOff;
        if( iDelta==0 ){
          bNewTerm = 1;
          if( iOff>=n ){
            fts5SegIterNextPage(p, pIter);
            pIter->iLeafOffset = 4;
          }else if( iOff!=fts5GetU16(&a[2]) ){
            pIter->iLeafOffset += fts5GetVarint32(&a[iOff], nKeep);
          }
        }else{
          pIter->iRowid += iDelta;
        }
      }else if( pIter->pSeg==0 ){
        const u8 *pList = 0;
        const char *zTerm;
        int nList;
        if( 0==(pIter->flags & FTS5_SEGITER_ONETERM) ){
          sqlite3Fts5HashScanNext(p->pHash);
          sqlite3Fts5HashScanEntry(p->pHash, &zTerm, &pList, &nList);
        }
        if( pList==0 ){
          fts5DataRelease(pIter->pLeaf);
          pIter->pLeaf = 0;
        }else{
          pIter->pLeaf->p = (u8*)pList;
          pIter->pLeaf->n = nList;
          sqlite3Fts5BufferSet(&p->rc, &pIter->term, strlen(zTerm), (u8*)zTerm);
          pIter->iLeafOffset = fts5GetVarint(pList, (u64*)&pIter->iRowid);
        }
      }else{
        iOff = 0;
        /* Next entry is not on the current page */
        while( iOff==0 ){
          fts5SegIterNextPage(p, pIter);
          pLeaf = pIter->pLeaf;
          if( pLeaf==0 ) break;
          if( (iOff = fts5GetU16(&pLeaf->p[0])) && iOff<pLeaf->n ){
            iOff += sqlite3Fts5GetVarint(&pLeaf->p[iOff], (u64*)&pIter->iRowid);
            pIter->iLeafOffset = iOff;
          }
          else if( (iOff = fts5GetU16(&pLeaf->p[2])) ){
            pIter->iLeafOffset = iOff;
            bNewTerm = 1;
          }
          if( iOff>=pLeaf->n ){
            p->rc = FTS5_CORRUPT;
            return;
          }
        }
      }

      /* Check if the iterator is now at EOF. If so, return early. */
      if( pIter->pLeaf ){
        if( bNewTerm ){
          if( pIter->flags & FTS5_SEGITER_ONETERM ){
            fts5DataRelease(pIter->pLeaf);
            pIter->pLeaf = 0;
          }else{
            fts5SegIterLoadTerm(p, pIter, nKeep);
            fts5SegIterLoadNPos(p, pIter);
            if( pbNewTerm ) *pbNewTerm = 1;
          }
        }else{
          fts5SegIterLoadNPos(p, pIter);
        }
      }
    }
  }
}

#define SWAPVAL(T, a, b) { T tmp; tmp=a; a=b; b=tmp; }

/*
** Iterator pIter currently points to the first rowid in a doclist. This
** function sets the iterator up so that iterates in reverse order through
** the doclist.
*/
static void fts5SegIterReverse(Fts5Index *p, Fts5SegIter *pIter){
  Fts5DlidxIter *pDlidx = pIter->pDlidx;
  Fts5Data *pLast = 0;
  int pgnoLast = 0;

  if( pDlidx ){
    int iSegid = pIter->pSeg->iSegid;
    pgnoLast = fts5DlidxIterPgno(pDlidx);
    pLast = fts5DataRead(p, FTS5_SEGMENT_ROWID(iSegid, 0, pgnoLast));
  }else{
    int iOff;                               /* Byte offset within pLeaf */
    Fts5Data *pLeaf = pIter->pLeaf;         /* Current leaf data */

    /* Currently, Fts5SegIter.iLeafOffset (and iOff) points to the first 
    ** byte of position-list content for the current rowid. Back it up
    ** so that it points to the start of the position-list size field. */
    pIter->iLeafOffset -= sqlite3Fts5GetVarintLen(pIter->nPos*2+pIter->bDel);
    iOff = pIter->iLeafOffset;
    assert( iOff>=4 );

    /* Search for a new term within the current leaf. If one can be found,
    ** then this page contains the largest rowid for the current term. */
    while( iOff<pLeaf->n ){
      int nPos;
      i64 iDelta;
      int bDummy;

      /* Read the position-list size field */
      iOff += fts5GetPoslistSize(&pLeaf->p[iOff], &nPos, &bDummy);
      iOff += nPos;
      if( iOff>=pLeaf->n ) break;

      /* Rowid delta. Or, if 0x00, the end of doclist marker. */
      nPos = fts5GetVarint(&pLeaf->p[iOff], (u64*)&iDelta);
      if( iDelta==0 ) break;
      iOff += nPos;
    }

    /* If this condition is true then the largest rowid for the current
    ** term may not be stored on the current page. So search forward to
    ** see where said rowid really is.  */
    if( iOff>=pLeaf->n ){
      int pgno;
      Fts5StructureSegment *pSeg = pIter->pSeg;

      /* The last rowid in the doclist may not be on the current page. Search
      ** forward to find the page containing the last rowid.  */
      for(pgno=pIter->iLeafPgno+1; !p->rc && pgno<=pSeg->pgnoLast; pgno++){
        i64 iAbs = FTS5_SEGMENT_ROWID(pSeg->iSegid, 0, pgno);
        Fts5Data *pNew = fts5DataRead(p, iAbs);
        if( pNew ){
          int iRowid, iTerm;
          fts5LeafHeader(pNew, &iRowid, &iTerm);
          if( iRowid ){
            SWAPVAL(Fts5Data*, pNew, pLast);
            pgnoLast = pgno;
          }
          fts5DataRelease(pNew);
          if( iTerm ) break;
        }
      }
    }
  }

  /* If pLast is NULL at this point, then the last rowid for this doclist
  ** lies on the page currently indicated by the iterator. In this case 
  ** pIter->iLeafOffset is already set to point to the position-list size
  ** field associated with the first relevant rowid on the page.
  **
  ** Or, if pLast is non-NULL, then it is the page that contains the last
  ** rowid. In this case configure the iterator so that it points to the
  ** first rowid on this page.
  */
  if( pLast ){
    int dummy;
    int iOff;
    fts5DataRelease(pIter->pLeaf);
    pIter->pLeaf = pLast;
    pIter->iLeafPgno = pgnoLast;
    fts5LeafHeader(pLast, &iOff, &dummy);
    iOff += fts5GetVarint(&pLast->p[iOff], (u64*)&pIter->iRowid);
    pIter->iLeafOffset = iOff;
  }

  fts5SegIterReverseInitPage(p, pIter);
}

/*
** Iterator pIter currently points to the first rowid of a doclist.
** There is a doclist-index associated with the final term on the current 
** page. If the current term is the last term on the page, load the 
** doclist-index from disk and initialize an iterator at (pIter->pDlidx).
*/
static void fts5SegIterLoadDlidx(Fts5Index *p, Fts5SegIter *pIter){
  int iSeg = pIter->pSeg->iSegid;
  int bRev = (pIter->flags & FTS5_SEGITER_REVERSE);
  Fts5Data *pLeaf = pIter->pLeaf; /* Current leaf data */

  assert( pIter->flags & FTS5_SEGITER_ONETERM );
  assert( pIter->pDlidx==0 );

  /* Check if the current doclist ends on this page. If it does, return
  ** early without loading the doclist-index (as it belongs to a different
  ** term. */
  if( pIter->iTermLeafPgno==pIter->iLeafPgno ){
    int iOff = pIter->iLeafOffset + pIter->nPos;
    while( iOff<pLeaf->n ){
      int bDummy;
      int nPos;
      i64 iDelta;

      /* iOff is currently the offset of the start of position list data */
      iOff += fts5GetVarint(&pLeaf->p[iOff], (u64*)&iDelta);
      if( iDelta==0 ) return;
      assert_nc( iOff<pLeaf->n );
      iOff += fts5GetPoslistSize(&pLeaf->p[iOff], &nPos, &bDummy);
      iOff += nPos;
    }
  }

  pIter->pDlidx = fts5DlidxIterInit(p, bRev, iSeg, pIter->iTermLeafPgno);
}

/*
** Initialize the object pIter to point to term pTerm/nTerm within segment
** pSeg. If there is no such term in the index, the iterator is set to EOF.
**
** If an error occurs, Fts5Index.rc is set to an appropriate error code. If 
** an error has already occurred when this function is called, it is a no-op.
*/
static void fts5SegIterSeekInit(
  Fts5Index *p,                   /* FTS5 backend */
  const u8 *pTerm, int nTerm,     /* Term to seek to */
  int flags,                      /* Mask of FTS5INDEX_XXX flags */
  Fts5StructureSegment *pSeg,     /* Description of segment */
  Fts5SegIter *pIter              /* Object to populate */
){
  int iPg = 1;
  int h;
  int bGe = (flags & FTS5INDEX_QUERY_SCAN);
  int bDlidx = 0;                 /* True if there is a doclist-index */
  Fts5Data *pLeaf;

  assert( bGe==0 || (flags & FTS5INDEX_QUERY_DESC)==0 );
  assert( pTerm && nTerm );
  memset(pIter, 0, sizeof(*pIter));
  pIter->pSeg = pSeg;

  /* This block sets stack variable iPg to the leaf page number that may
  ** contain term (pTerm/nTerm), if it is present in the segment. */
  for(h=pSeg->nHeight-1; h>0; h--){
    Fts5NodeIter node;              /* For iterating through internal nodes */
    i64 iRowid = FTS5_SEGMENT_ROWID(pSeg->iSegid, h, iPg);
    Fts5Data *pNode = fts5DataRead(p, iRowid);
    if( pNode==0 ) break;

    fts5NodeIterInit(pNode->p, pNode->n, &node);
    assert( node.term.n==0 );

    iPg = node.iChild;
    bDlidx = node.bDlidx;
    for(fts5NodeIterNext(&p->rc, &node);
        node.aData && fts5BufferCompareBlob(&node.term, pTerm, nTerm)<=0;
        fts5NodeIterNext(&p->rc, &node)
    ){
      iPg = node.iChild;
      bDlidx = node.bDlidx;
    }
    fts5NodeIterFree(&node);
    fts5DataRelease(pNode);
  }

  if( iPg<pSeg->pgnoFirst ){
    iPg = pSeg->pgnoFirst;
    bDlidx = 0;
  }

  pIter->iLeafPgno = iPg - 1;
  fts5SegIterNextPage(p, pIter);

  if( (pLeaf = pIter->pLeaf) ){
    int res;
    pIter->iLeafOffset = fts5GetU16(&pLeaf->p[2]);
    if( pIter->iLeafOffset<4 || pIter->iLeafOffset>=pLeaf->n ){
      p->rc = FTS5_CORRUPT;
    }else{
      fts5SegIterLoadTerm(p, pIter, 0);
      fts5SegIterLoadNPos(p, pIter);
      do {
        res = fts5BufferCompareBlob(&pIter->term, pTerm, nTerm);
        if( res>=0 ) break;
        fts5SegIterNext(p, pIter, 0);
      }while( pIter->pLeaf && p->rc==SQLITE_OK );

      if( bGe==0 && res ){
        /* Set iterator to point to EOF */
        fts5DataRelease(pIter->pLeaf);
        pIter->pLeaf = 0;
      }
    }
  }

  if( p->rc==SQLITE_OK && bGe==0 ){
    pIter->flags |= FTS5_SEGITER_ONETERM;
    if( pIter->pLeaf ){
      if( flags & FTS5INDEX_QUERY_DESC ){
        pIter->flags |= FTS5_SEGITER_REVERSE;
      }
      if( bDlidx ){
        fts5SegIterLoadDlidx(p, pIter);
      }
      if( flags & FTS5INDEX_QUERY_DESC ){
        fts5SegIterReverse(p, pIter);
      }
    }
  }
}

/*
** Initialize the object pIter to point to term pTerm/nTerm within the
** in-memory hash table. If there is no such term in the hash-table, the 
** iterator is set to EOF.
**
** If an error occurs, Fts5Index.rc is set to an appropriate error code. If 
** an error has already occurred when this function is called, it is a no-op.
*/
static void fts5SegIterHashInit(
  Fts5Index *p,                   /* FTS5 backend */
  const u8 *pTerm, int nTerm,     /* Term to seek to */
  int flags,                      /* Mask of FTS5INDEX_XXX flags */
  Fts5SegIter *pIter              /* Object to populate */
){
  const u8 *pList = 0;
  int nList = 0;
  const u8 *z = 0;
  int n = 0;

  assert( p->pHash );
  assert( p->rc==SQLITE_OK );

  if( pTerm==0 || (flags & FTS5INDEX_QUERY_SCAN) ){
    p->rc = sqlite3Fts5HashScanInit(p->pHash, (const char*)pTerm, nTerm);
    sqlite3Fts5HashScanEntry(p->pHash, (const char**)&z, &pList, &nList);
    n = (z ? strlen((const char*)z) : 0);
  }else{
    pIter->flags |= FTS5_SEGITER_ONETERM;
    sqlite3Fts5HashQuery(p->pHash, (const char*)pTerm, nTerm, &pList, &nList);
    z = pTerm;
    n = nTerm;
  }

  if( pList ){
    Fts5Data *pLeaf;
    sqlite3Fts5BufferSet(&p->rc, &pIter->term, n, z);
    pLeaf = fts5IdxMalloc(p, sizeof(Fts5Data));
    if( pLeaf==0 ) return;
    pLeaf->p = (u8*)pList;
    pLeaf->n = nList;
    pIter->pLeaf = pLeaf;
    pIter->iLeafOffset = fts5GetVarint(pLeaf->p, (u64*)&pIter->iRowid);

    if( flags & FTS5INDEX_QUERY_DESC ){
      pIter->flags |= FTS5_SEGITER_REVERSE;
      fts5SegIterReverseInitPage(p, pIter);
    }else{
      fts5SegIterLoadNPos(p, pIter);
    }
  }
}

/*
** Zero the iterator passed as the only argument.
*/
static void fts5SegIterClear(Fts5SegIter *pIter){
  fts5BufferFree(&pIter->term);
  fts5DataRelease(pIter->pLeaf);
  fts5DataRelease(pIter->pNextLeaf);
  fts5DlidxIterFree(pIter->pDlidx);
  sqlite3_free(pIter->aRowidOffset);
  memset(pIter, 0, sizeof(Fts5SegIter));
}

#ifdef SQLITE_DEBUG

/*
** This function is used as part of the big assert() procedure implemented by
** fts5AssertMultiIterSetup(). It ensures that the result currently stored
** in *pRes is the correct result of comparing the current positions of the
** two iterators.
*/
static void fts5AssertComparisonResult(
  Fts5MultiSegIter *pIter, 
  Fts5SegIter *p1,
  Fts5SegIter *p2,
  Fts5CResult *pRes
){
  int i1 = p1 - pIter->aSeg;
  int i2 = p2 - pIter->aSeg;

  if( p1->pLeaf || p2->pLeaf ){
    if( p1->pLeaf==0 ){
      assert( pRes->iFirst==i2 );
    }else if( p2->pLeaf==0 ){
      assert( pRes->iFirst==i1 );
    }else{
      int nMin = MIN(p1->term.n, p2->term.n);
      int res = memcmp(p1->term.p, p2->term.p, nMin);
      if( res==0 ) res = p1->term.n - p2->term.n;

      if( res==0 ){
        assert( pRes->bTermEq==1 );
        assert( p1->iRowid!=p2->iRowid );
        res = ((p1->iRowid > p2->iRowid)==pIter->bRev) ? -1 : 1;
      }else{
        assert( pRes->bTermEq==0 );
      }

      if( res<0 ){
        assert( pRes->iFirst==i1 );
      }else{
        assert( pRes->iFirst==i2 );
      }
    }
  }
}

/*
** This function is a no-op unless SQLITE_DEBUG is defined when this module
** is compiled. In that case, this function is essentially an assert() 
** statement used to verify that the contents of the pIter->aFirst[] array
** are correct.
*/
static void fts5AssertMultiIterSetup(Fts5Index *p, Fts5MultiSegIter *pIter){
  if( p->rc==SQLITE_OK ){
    int i;
    for(i=0; i<pIter->nSeg; i+=2){
      Fts5SegIter *p1 = &pIter->aSeg[i];
      Fts5SegIter *p2 = &pIter->aSeg[i+1];
      Fts5CResult *pRes = &pIter->aFirst[(pIter->nSeg + i) / 2];
      fts5AssertComparisonResult(pIter, p1, p2, pRes);
    }

    for(i=1; i<(pIter->nSeg / 2); i+=2){
      Fts5CResult *pRes = &pIter->aFirst[i];
      Fts5SegIter *p1 = &pIter->aSeg[ pIter->aFirst[i*2].iFirst ];
      Fts5SegIter *p2 = &pIter->aSeg[ pIter->aFirst[i*2+1].iFirst ];

      fts5AssertComparisonResult(pIter, p1, p2, pRes);
    }
  }
}
#else
# define fts5AssertMultiIterSetup(x,y)
#endif

/*
** Do the comparison necessary to populate pIter->aFirst[iOut].
**
** If the returned value is non-zero, then it is the index of an entry
** in the pIter->aSeg[] array that is (a) not at EOF, and (b) pointing
** to a key that is a duplicate of another, higher priority, 
** segment-iterator in the pSeg->aSeg[] array.
*/
static int fts5MultiIterDoCompare(Fts5MultiSegIter *pIter, int iOut){
  int i1;                         /* Index of left-hand Fts5SegIter */
  int i2;                         /* Index of right-hand Fts5SegIter */
  int iRes;
  Fts5SegIter *p1;                /* Left-hand Fts5SegIter */
  Fts5SegIter *p2;                /* Right-hand Fts5SegIter */
  Fts5CResult *pRes = &pIter->aFirst[iOut];

  assert( iOut<pIter->nSeg && iOut>0 );
  assert( pIter->bRev==0 || pIter->bRev==1 );

  if( iOut>=(pIter->nSeg/2) ){
    i1 = (iOut - pIter->nSeg/2) * 2;
    i2 = i1 + 1;
  }else{
    i1 = pIter->aFirst[iOut*2].iFirst;
    i2 = pIter->aFirst[iOut*2+1].iFirst;
  }
  p1 = &pIter->aSeg[i1];
  p2 = &pIter->aSeg[i2];

  pRes->bTermEq = 0;
  if( p1->pLeaf==0 ){           /* If p1 is at EOF */
    iRes = i2;
  }else if( p2->pLeaf==0 ){     /* If p2 is at EOF */
    iRes = i1;
  }else{
    int res = fts5BufferCompare(&p1->term, &p2->term);
    if( res==0 ){
      assert( i2>i1 );
      assert( i2!=0 );
      pRes->bTermEq = 1;
      if( p1->iRowid==p2->iRowid ){
        p1->bDel = p2->bDel;
        return i2;
      }
      res = ((p1->iRowid > p2->iRowid)==pIter->bRev) ? -1 : +1;
    }
    assert( res!=0 );
    if( res<0 ){
      iRes = i1;
    }else{
      iRes = i2;
    }
  }

  pRes->iFirst = iRes;
  return 0;
}

/*
** Move the seg-iter so that it points to the first rowid on page iLeafPgno.
** It is an error if leaf iLeafPgno does not exist or contains no rowids.
*/
static void fts5SegIterGotoPage(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5SegIter *pIter,             /* Iterator to advance */
  int iLeafPgno
){
  assert( iLeafPgno>pIter->iLeafPgno );

  if( iLeafPgno>pIter->pSeg->pgnoLast ){
    p->rc = FTS5_CORRUPT;
  }else{
    fts5DataRelease(pIter->pNextLeaf);
    pIter->pNextLeaf = 0;
    pIter->iLeafPgno = iLeafPgno-1;
    fts5SegIterNextPage(p, pIter);
    assert( p->rc!=SQLITE_OK || pIter->iLeafPgno==iLeafPgno );

    if( p->rc==SQLITE_OK ){
      int iOff;
      u8 *a = pIter->pLeaf->p;
      int n = pIter->pLeaf->n;

      iOff = fts5GetU16(&a[0]);
      if( iOff<4 || iOff>=n ){
        p->rc = FTS5_CORRUPT;
      }else{
        iOff += fts5GetVarint(&a[iOff], (u64*)&pIter->iRowid);
        pIter->iLeafOffset = iOff;
        fts5SegIterLoadNPos(p, pIter);
      }
    }
  }
}

/*
** Advance the iterator passed as the second argument until it is at or 
** past rowid iFrom. Regardless of the value of iFrom, the iterator is
** always advanced at least once.
*/
static void fts5SegIterNextFrom(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5SegIter *pIter,             /* Iterator to advance */
  i64 iMatch                      /* Advance iterator at least this far */
){
  int bRev = (pIter->flags & FTS5_SEGITER_REVERSE);
  Fts5DlidxIter *pDlidx = pIter->pDlidx;
  int iLeafPgno = pIter->iLeafPgno;
  int bMove = 1;

  assert( pIter->flags & FTS5_SEGITER_ONETERM );
  assert( pIter->pDlidx );
  assert( pIter->pLeaf );

  if( bRev==0 ){
    while( !fts5DlidxIterEof(p, pDlidx) && iMatch>fts5DlidxIterRowid(pDlidx) ){
      iLeafPgno = fts5DlidxIterPgno(pDlidx);
      fts5DlidxIterNext(p, pDlidx);
    }
    assert_nc( iLeafPgno>=pIter->iLeafPgno || p->rc );
    if( iLeafPgno>pIter->iLeafPgno ){
      fts5SegIterGotoPage(p, pIter, iLeafPgno);
      bMove = 0;
    }
  }else{
    assert( pIter->pNextLeaf==0 );
    assert( iMatch<pIter->iRowid );
    while( !fts5DlidxIterEof(p, pDlidx) && iMatch<fts5DlidxIterRowid(pDlidx) ){
      fts5DlidxIterPrev(p, pDlidx);
    }
    iLeafPgno = fts5DlidxIterPgno(pDlidx);

    assert( fts5DlidxIterEof(p, pDlidx) || iLeafPgno<=pIter->iLeafPgno );

    if( iLeafPgno<pIter->iLeafPgno ){
      pIter->iLeafPgno = iLeafPgno+1;
      fts5SegIterReverseNewPage(p, pIter);
      bMove = 0;
    }
  }

  while( p->rc==SQLITE_OK ){
    if( bMove ) fts5SegIterNext(p, pIter, 0);
    if( pIter->pLeaf==0 ) break;
    if( bRev==0 && pIter->iRowid>=iMatch ) break;
    if( bRev!=0 && pIter->iRowid<=iMatch ) break;
    bMove = 1;
  }
}


/*
** Free the iterator object passed as the second argument.
*/
static void fts5MultiIterFree(Fts5Index *p, Fts5MultiSegIter *pIter){
  if( pIter ){
    int i;
    for(i=0; i<pIter->nSeg; i++){
      fts5SegIterClear(&pIter->aSeg[i]);
    }
    sqlite3_free(pIter);
  }
}

static void fts5MultiIterAdvanced(
  Fts5Index *p,                   /* FTS5 backend to iterate within */
  Fts5MultiSegIter *pIter,        /* Iterator to update aFirst[] array for */
  int iChanged,                   /* Index of sub-iterator just advanced */
  int iMinset                     /* Minimum entry in aFirst[] to set */
){
  int i;
  for(i=(pIter->nSeg+iChanged)/2; i>=iMinset && p->rc==SQLITE_OK; i=i/2){
    int iEq;
    if( (iEq = fts5MultiIterDoCompare(pIter, i)) ){
      fts5SegIterNext(p, &pIter->aSeg[iEq], 0);
      i = pIter->nSeg + iEq;
    }
  }
}

static int fts5MultiIterAdvanceRowid(
  Fts5Index *p,                   /* FTS5 backend to iterate within */
  Fts5MultiSegIter *pIter,        /* Iterator to update aFirst[] array for */
  int iChanged                    /* Index of sub-iterator just advanced */
){
  int i;
  Fts5SegIter *pNew = &pIter->aSeg[iChanged];
  Fts5SegIter *pOther = &pIter->aSeg[iChanged ^ 0x0001];

  for(i=(pIter->nSeg+iChanged)/2; p->rc==SQLITE_OK; i=i/2){
    Fts5CResult *pRes = &pIter->aFirst[i];

    assert( pNew->pLeaf );
    assert( pRes->bTermEq==0 || pOther->pLeaf );
    
    if( pRes->bTermEq ){
      if( pNew->iRowid==pOther->iRowid ){
        return 1;
      }else if( (pOther->iRowid>pNew->iRowid)==pIter->bRev ){
        pNew = pOther;
      }
    }
    pRes->iFirst = (pNew - pIter->aSeg);
    if( i==1 ) break;

    pOther = &pIter->aSeg[ pIter->aFirst[i ^ 0x0001].iFirst ];
  }

  return 0;
}

/*
** Move the iterator to the next entry. 
**
** If an error occurs, an error code is left in Fts5Index.rc. It is not 
** considered an error if the iterator reaches EOF, or if it is already at 
** EOF when this function is called.
*/
static void fts5MultiIterNext(
  Fts5Index *p, 
  Fts5MultiSegIter *pIter,
  int bFrom,                      /* True if argument iFrom is valid */
  i64 iFrom                       /* Advance at least as far as this */
){
  if( p->rc==SQLITE_OK ){
    int bUseFrom = bFrom;
    do {
      int iFirst = pIter->aFirst[1].iFirst;
      int bNewTerm = 0;
      Fts5SegIter *pSeg = &pIter->aSeg[iFirst];
      assert( p->rc==SQLITE_OK );
      if( bUseFrom && pSeg->pDlidx ){
        fts5SegIterNextFrom(p, pSeg, iFrom);
      }else{
        fts5SegIterNext(p, pSeg, &bNewTerm);
      }

      if( pSeg->pLeaf==0 || bNewTerm 
       || fts5MultiIterAdvanceRowid(p, pIter, iFirst)
      ){
        fts5MultiIterAdvanced(p, pIter, iFirst, 1);
      }
      fts5AssertMultiIterSetup(p, pIter);

      bUseFrom = 0;
    }while( pIter->bSkipEmpty && fts5MultiIterIsEmpty(p, pIter) );
  }
}

static Fts5MultiSegIter *fts5MultiIterAlloc(
  Fts5Index *p,                   /* FTS5 backend to iterate within */
  int nSeg
){
  Fts5MultiSegIter *pNew;
  int nSlot;                      /* Power of two >= nSeg */

  for(nSlot=2; nSlot<nSeg; nSlot=nSlot*2);
  pNew = fts5IdxMalloc(p, 
      sizeof(Fts5MultiSegIter) +          /* pNew */
      sizeof(Fts5SegIter) * nSlot +       /* pNew->aSeg[] */
      sizeof(Fts5CResult) * nSlot         /* pNew->aFirst[] */
  );
  if( pNew ){
    pNew->nSeg = nSlot;
    pNew->aSeg = (Fts5SegIter*)&pNew[1];
    pNew->aFirst = (Fts5CResult*)&pNew->aSeg[nSlot];
  }
  return pNew;
}

/*
** Allocate a new Fts5MultiSegIter object.
**
** The new object will be used to iterate through data in structure pStruct.
** If iLevel is -ve, then all data in all segments is merged. Or, if iLevel
** is zero or greater, data from the first nSegment segments on level iLevel
** is merged.
**
** The iterator initially points to the first term/rowid entry in the 
** iterated data.
*/
static void fts5MultiIterNew(
  Fts5Index *p,                   /* FTS5 backend to iterate within */
  Fts5Structure *pStruct,         /* Structure of specific index */
  int bSkipEmpty,                 /* True to ignore delete-keys */
  int flags,                      /* FTS5INDEX_QUERY_XXX flags */
  const u8 *pTerm, int nTerm,     /* Term to seek to (or NULL/0) */
  int iLevel,                     /* Level to iterate (-1 for all) */
  int nSegment,                   /* Number of segments to merge (iLevel>=0) */
  Fts5MultiSegIter **ppOut        /* New object */
){
  int nSeg = 0;                   /* Number of segment-iters in use */
  int iIter = 0;                  /* */
  int iSeg;                       /* Used to iterate through segments */
  Fts5StructureLevel *pLvl;
  Fts5MultiSegIter *pNew;

  assert( (pTerm==0 && nTerm==0) || iLevel<0 );

  /* Allocate space for the new multi-seg-iterator. */
  if( p->rc==SQLITE_OK ){
    if( iLevel<0 ){
      assert( pStruct->nSegment==fts5StructureCountSegments(pStruct) );
      nSeg = pStruct->nSegment;
      nSeg += (p->pHash ? 1 : 0);
    }else{
      nSeg = MIN(pStruct->aLevel[iLevel].nSeg, nSegment);
    }
  }
  *ppOut = pNew = fts5MultiIterAlloc(p, nSeg);
  if( pNew==0 ) return;
  pNew->bRev = (0!=(flags & FTS5INDEX_QUERY_DESC));
  pNew->bSkipEmpty = bSkipEmpty;

  /* Initialize each of the component segment iterators. */
  if( iLevel<0 ){
    Fts5StructureLevel *pEnd = &pStruct->aLevel[pStruct->nLevel];
    if( p->pHash ){
      /* Add a segment iterator for the current contents of the hash table. */
      Fts5SegIter *pIter = &pNew->aSeg[iIter++];
      fts5SegIterHashInit(p, pTerm, nTerm, flags, pIter);
    }
    for(pLvl=&pStruct->aLevel[0]; pLvl<pEnd; pLvl++){
      for(iSeg=pLvl->nSeg-1; iSeg>=0; iSeg--){
        Fts5StructureSegment *pSeg = &pLvl->aSeg[iSeg];
        Fts5SegIter *pIter = &pNew->aSeg[iIter++];
        if( pTerm==0 ){
          fts5SegIterInit(p, pSeg, pIter);
        }else{
          fts5SegIterSeekInit(p, pTerm, nTerm, flags, pSeg, pIter);
        }
      }
    }
  }else{
    pLvl = &pStruct->aLevel[iLevel];
    for(iSeg=nSeg-1; iSeg>=0; iSeg--){
      fts5SegIterInit(p, &pLvl->aSeg[iSeg], &pNew->aSeg[iIter++]);
    }
  }
  assert( iIter==nSeg );

  /* If the above was successful, each component iterators now points 
  ** to the first entry in its segment. In this case initialize the 
  ** aFirst[] array. Or, if an error has occurred, free the iterator
  ** object and set the output variable to NULL.  */
  if( p->rc==SQLITE_OK ){
    for(iIter=pNew->nSeg-1; iIter>0; iIter--){
      int iEq;
      if( (iEq = fts5MultiIterDoCompare(pNew, iIter)) ){
        fts5SegIterNext(p, &pNew->aSeg[iEq], 0);
        fts5MultiIterAdvanced(p, pNew, iEq, iIter);
      }
    }
    fts5AssertMultiIterSetup(p, pNew);

    if( pNew->bSkipEmpty && fts5MultiIterIsEmpty(p, pNew) ){
      fts5MultiIterNext(p, pNew, 0, 0);
    }
  }else{
    fts5MultiIterFree(p, pNew);
    *ppOut = 0;
  }
}

/*
** Create an Fts5MultiSegIter that iterates through the doclist provided
** as the second argument.
*/
static void fts5MultiIterNew2(
  Fts5Index *p,                   /* FTS5 backend to iterate within */
  Fts5Data *pData,                /* Doclist to iterate through */
  int bDesc,                      /* True for descending rowid order */
  Fts5MultiSegIter **ppOut        /* New object */
){
  Fts5MultiSegIter *pNew;
  pNew = fts5MultiIterAlloc(p, 2);
  if( pNew ){
    Fts5SegIter *pIter = &pNew->aSeg[1];

    pIter->flags = FTS5_SEGITER_ONETERM;
    if( pData->n>0 ){
      pIter->pLeaf = pData;
      pIter->iLeafOffset = fts5GetVarint(pData->p, (u64*)&pIter->iRowid);
      pNew->aFirst[1].iFirst = 1;
      if( bDesc ){
        pNew->bRev = 1;
        pIter->flags |= FTS5_SEGITER_REVERSE;
        fts5SegIterReverseInitPage(p, pIter);
      }else{
        fts5SegIterLoadNPos(p, pIter);
      }
      pData = 0;
    }

    *ppOut = pNew;
  }

  fts5DataRelease(pData);
}

/*
** Return true if the iterator is at EOF or if an error has occurred. 
** False otherwise.
*/
static int fts5MultiIterEof(Fts5Index *p, Fts5MultiSegIter *pIter){
  return (p->rc || pIter->aSeg[ pIter->aFirst[1].iFirst ].pLeaf==0);
}

/*
** Return the rowid of the entry that the iterator currently points
** to. If the iterator points to EOF when this function is called the
** results are undefined.
*/
static i64 fts5MultiIterRowid(Fts5MultiSegIter *pIter){
  assert( pIter->aSeg[ pIter->aFirst[1].iFirst ].pLeaf );
  return pIter->aSeg[ pIter->aFirst[1].iFirst ].iRowid;
}

/*
** Move the iterator to the next entry at or following iMatch.
*/
static void fts5MultiIterNextFrom(
  Fts5Index *p, 
  Fts5MultiSegIter *pIter, 
  i64 iMatch
){
  while( 1 ){
    i64 iRowid;
    fts5MultiIterNext(p, pIter, 1, iMatch);
    if( fts5MultiIterEof(p, pIter) ) break;
    iRowid = fts5MultiIterRowid(pIter);
    if( pIter->bRev==0 && iRowid>=iMatch ) break;
    if( pIter->bRev!=0 && iRowid<=iMatch ) break;
  }
}

/*
** Return a pointer to a buffer containing the term associated with the 
** entry that the iterator currently points to.
*/
static const u8 *fts5MultiIterTerm(Fts5MultiSegIter *pIter, int *pn){
  Fts5SegIter *p = &pIter->aSeg[ pIter->aFirst[1].iFirst ];
  *pn = p->term.n;
  return p->term.p;
}

static void fts5ChunkIterate(
  Fts5Index *p,                   /* Index object */
  Fts5SegIter *pSeg,              /* Poslist of this iterator */
  void *pCtx,                     /* Context pointer for xChunk callback */
  void (*xChunk)(Fts5Index*, void*, const u8*, int)
){
  int nRem = pSeg->nPos;          /* Number of bytes still to come */
  Fts5Data *pData = 0;
  u8 *pChunk = &pSeg->pLeaf->p[pSeg->iLeafOffset];
  int nChunk = MIN(nRem, pSeg->pLeaf->n - pSeg->iLeafOffset);
  int pgno = pSeg->iLeafPgno;
  int pgnoSave = 0;

  if( (pSeg->flags & FTS5_SEGITER_REVERSE)==0 ){
    pgnoSave = pgno+1;
  }

  while( 1 ){
    xChunk(p, pCtx, pChunk, nChunk);
    nRem -= nChunk;
    fts5DataRelease(pData);
    if( nRem<=0 ){
      break;
    }else{
      pgno++;
      pData = fts5DataRead(p, FTS5_SEGMENT_ROWID(pSeg->pSeg->iSegid, 0, pgno));
      if( pData==0 ) break;
      pChunk = &pData->p[4];
      nChunk = MIN(nRem, pData->n - 4);
      if( pgno==pgnoSave ){
        assert( pSeg->pNextLeaf==0 );
        pSeg->pNextLeaf = pData;
        pData = 0;
      }
    }
  }
}



/*
** Allocate a new segment-id for the structure pStruct. The new segment
** id must be between 1 and 65335 inclusive, and must not be used by 
** any currently existing segment. If a free segment id cannot be found,
** SQLITE_FULL is returned.
**
** If an error has already occurred, this function is a no-op. 0 is 
** returned in this case.
*/
static int fts5AllocateSegid(Fts5Index *p, Fts5Structure *pStruct){
  u32 iSegid = 0;

  if( p->rc==SQLITE_OK ){
    if( pStruct->nSegment>=FTS5_MAX_SEGMENT ){
      p->rc = SQLITE_FULL;
    }else{
      while( iSegid==0 ){
        int iLvl, iSeg;
        sqlite3_randomness(sizeof(u32), (void*)&iSegid);
        iSegid = (iSegid % ((1 << FTS5_DATA_ID_B) - 2)) + 1;
        assert( iSegid>0 && iSegid<=65535 );
        for(iLvl=0; iLvl<pStruct->nLevel; iLvl++){
          for(iSeg=0; iSeg<pStruct->aLevel[iLvl].nSeg; iSeg++){
            if( iSegid==pStruct->aLevel[iLvl].aSeg[iSeg].iSegid ){
              iSegid = 0;
            }
          }
        }
      }
    }
  }

  return (int)iSegid;
}

/*
** Discard all data currently cached in the hash-tables.
*/
static void fts5IndexDiscardData(Fts5Index *p){
  assert( p->pHash || p->nPendingData==0 );
  if( p->pHash ){
    sqlite3Fts5HashClear(p->pHash);
    p->nPendingData = 0;
  }
}

/*
** Return the size of the prefix, in bytes, that buffer (nNew/pNew) shares
** with buffer (nOld/pOld).
*/
static int fts5PrefixCompress(
  int nOld, const u8 *pOld,
  int nNew, const u8 *pNew
){
  int i;
  assert( fts5BlobCompare(pOld, nOld, pNew, nNew)<0 );
  for(i=0; i<nOld; i++){
    if( pOld[i]!=pNew[i] ) break;
  }
  return i;
}

static void fts5WriteDlidxClear(
  Fts5Index *p, 
  Fts5SegWriter *pWriter,
  int bFlush                      /* If true, write dlidx to disk */
){
  int i;
  assert( bFlush==0 || (pWriter->nDlidx>0 && pWriter->aDlidx[0].buf.n>0) );
  for(i=0; i<pWriter->nDlidx; i++){
    Fts5DlidxWriter *pDlidx = &pWriter->aDlidx[i];
    if( pDlidx->buf.n==0 ) break;
    if( bFlush ){
      assert( pDlidx->pgno!=0 );
      fts5DataWrite(p, 
          FTS5_DLIDX_ROWID(pWriter->iSegid, i, pDlidx->pgno),
          pDlidx->buf.p, pDlidx->buf.n
      );
    }
    sqlite3Fts5BufferZero(&pDlidx->buf);
    pDlidx->bPrevValid = 0;
  }
}

/*
** Grow the pWriter->aDlidx[] array to at least nLvl elements in size.
** Any new array elements are zeroed before returning.
*/
static int fts5WriteDlidxGrow(
  Fts5Index *p,
  Fts5SegWriter *pWriter,
  int nLvl
){
  if( p->rc==SQLITE_OK && nLvl>=pWriter->nDlidx ){
    Fts5DlidxWriter *aDlidx = (Fts5DlidxWriter*)sqlite3_realloc(
        pWriter->aDlidx, sizeof(Fts5DlidxWriter) * nLvl
    );
    if( aDlidx==0 ){
      p->rc = SQLITE_NOMEM;
    }else{
      int nByte = sizeof(Fts5DlidxWriter) * (nLvl - pWriter->nDlidx);
      memset(&aDlidx[pWriter->nDlidx], 0, nByte);
      pWriter->aDlidx = aDlidx;
      pWriter->nDlidx = nLvl;
    }
  }
  return p->rc;
}

/*
** If an "nEmpty" record must be written to the b-tree before the next
** term, write it now. 
*/
static void fts5WriteBtreeNEmpty(Fts5Index *p, Fts5SegWriter *pWriter){
  if( pWriter->nEmpty ){
    int bFlag = 0;
    Fts5PageWriter *pPg;
    pPg = &pWriter->aWriter[1];

    /* If there were FTS5_MIN_DLIDX_SIZE or more empty leaf pages written
    ** to the database, also write the doclist-index to disk.  */
    if( pWriter->aDlidx[0].buf.n>0 && pWriter->nEmpty>=FTS5_MIN_DLIDX_SIZE ){
      bFlag = 1;
    }
    fts5WriteDlidxClear(p, pWriter, bFlag);
    fts5BufferAppendVarint(&p->rc, &pPg->buf, bFlag);
    fts5BufferAppendVarint(&p->rc, &pPg->buf, pWriter->nEmpty);
    pWriter->nEmpty = 0;
  }else{
    fts5WriteDlidxClear(p, pWriter, 0);
  }

  assert( pWriter->nDlidx==0 || pWriter->aDlidx[0].buf.n==0 );
  assert( pWriter->nDlidx==0 || pWriter->aDlidx[0].bPrevValid==0 );
}

static void fts5WriteBtreeGrow(Fts5Index *p, Fts5SegWriter *pWriter){
  if( p->rc==SQLITE_OK ){
    Fts5PageWriter *aNew;
    Fts5PageWriter *pNew;
    int nNew = sizeof(Fts5PageWriter) * (pWriter->nWriter+1);

    aNew = (Fts5PageWriter*)sqlite3_realloc(pWriter->aWriter, nNew);
    if( aNew==0 ){
      p->rc = SQLITE_NOMEM;
      return;
    }

    pNew = &aNew[pWriter->nWriter];
    memset(pNew, 0, sizeof(Fts5PageWriter));
    pNew->pgno = 1;
    fts5BufferAppendVarint(&p->rc, &pNew->buf, 1);

    pWriter->nWriter++;
    pWriter->aWriter = aNew;
  }
}

/*
** This is called once for each leaf page except the first that contains
** at least one term. Argument (nTerm/pTerm) is the split-key - a term that
** is larger than all terms written to earlier leaves, and equal to or
** smaller than the first term on the new leaf.
**
** If an error occurs, an error code is left in Fts5Index.rc. If an error
** has already occurred when this function is called, it is a no-op.
*/
static void fts5WriteBtreeTerm(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5SegWriter *pWriter,         /* Writer object */
  int nTerm, const u8 *pTerm      /* First term on new page */
){
  int iHeight;
  for(iHeight=1; 1; iHeight++){
    Fts5PageWriter *pPage;

    if( iHeight>=pWriter->nWriter ){
      fts5WriteBtreeGrow(p, pWriter);
      if( p->rc ) return;
    }
    pPage = &pWriter->aWriter[iHeight];

    fts5WriteBtreeNEmpty(p, pWriter);

    if( pPage->buf.n>=p->pConfig->pgsz ){
      /* pPage will be written to disk. The term will be written into the
      ** parent of pPage.  */
      i64 iRowid = FTS5_SEGMENT_ROWID(pWriter->iSegid, iHeight, pPage->pgno);
      fts5DataWrite(p, iRowid, pPage->buf.p, pPage->buf.n);
      fts5BufferZero(&pPage->buf);
      fts5BufferZero(&pPage->term);
      fts5BufferAppendVarint(&p->rc, &pPage->buf, pPage[-1].pgno);
      pPage->pgno++;
    }else{
      int nPre = fts5PrefixCompress(pPage->term.n, pPage->term.p, nTerm, pTerm);
      fts5BufferAppendVarint(&p->rc, &pPage->buf, nPre+2);
      fts5BufferAppendVarint(&p->rc, &pPage->buf, nTerm-nPre);
      fts5BufferAppendBlob(&p->rc, &pPage->buf, nTerm-nPre, pTerm+nPre);
      fts5BufferSet(&p->rc, &pPage->term, nTerm, pTerm);
      break;
    }
  }
}

/*
** This function is called when flushing a leaf page that contains no
** terms at all to disk.
*/
static void fts5WriteBtreeNoTerm(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5SegWriter *pWriter          /* Writer object */
){
  /* If there were no rowids on the leaf page either and the doclist-index
  ** has already been started, append an 0x00 byte to it.  */
  if( pWriter->bFirstRowidInPage && pWriter->aDlidx[0].buf.n>0 ){
    Fts5DlidxWriter *pDlidx = &pWriter->aDlidx[0];
    assert( pDlidx->bPrevValid );
    sqlite3Fts5BufferAppendVarint(&p->rc, &pDlidx->buf, 0);
  }

  /* Increment the "number of sequential leaves without a term" counter. */
  pWriter->nEmpty++;
}

static i64 fts5DlidxExtractFirstRowid(Fts5Buffer *pBuf){
  i64 iRowid;
  int iOff;

  iOff = 1 + fts5GetVarint(&pBuf->p[1], (u64*)&iRowid);
  fts5GetVarint(&pBuf->p[iOff], (u64*)&iRowid);
  return iRowid;
}

/*
** Rowid iRowid has just been appended to the current leaf page. It is the
** first on the page. This function appends an appropriate entry to the current
** doclist-index.
*/
static void fts5WriteDlidxAppend(
  Fts5Index *p, 
  Fts5SegWriter *pWriter, 
  i64 iRowid
){
  int i;
  int bDone = 0;

  for(i=0; p->rc==SQLITE_OK && bDone==0; i++){
    i64 iVal;
    Fts5DlidxWriter *pDlidx = &pWriter->aDlidx[i];

    if( pDlidx->buf.n>=p->pConfig->pgsz ){
      /* The current doclist-index page is full. Write it to disk and push
      ** a copy of iRowid (which will become the first rowid on the next
      ** doclist-index leaf page) up into the next level of the b-tree 
      ** hierarchy. If the node being flushed is currently the root node,
      ** also push its first rowid upwards. */
      pDlidx->buf.p[0] = 0x01;    /* Not the root node */
      fts5DataWrite(p, 
          FTS5_DLIDX_ROWID(pWriter->iSegid, i, pDlidx->pgno),
          pDlidx->buf.p, pDlidx->buf.n
      );
      fts5WriteDlidxGrow(p, pWriter, i+2);
      pDlidx = &pWriter->aDlidx[i];
      if( p->rc==SQLITE_OK && pDlidx[1].buf.n==0 ){
        i64 iFirst = fts5DlidxExtractFirstRowid(&pDlidx->buf);

        /* This was the root node. Push its first rowid up to the new root. */
        pDlidx[1].pgno = pDlidx->pgno;
        sqlite3Fts5BufferAppendVarint(&p->rc, &pDlidx[1].buf, 0);
        sqlite3Fts5BufferAppendVarint(&p->rc, &pDlidx[1].buf, pDlidx->pgno);
        sqlite3Fts5BufferAppendVarint(&p->rc, &pDlidx[1].buf, iFirst);
        pDlidx[1].bPrevValid = 1;
        pDlidx[1].iPrev = iFirst;
      }

      sqlite3Fts5BufferZero(&pDlidx->buf);
      pDlidx->bPrevValid = 0;
      pDlidx->pgno++;
    }else{
      bDone = 1;
    }

    if( pDlidx->bPrevValid ){
      iVal = iRowid - pDlidx->iPrev;
    }else{
      i64 iPgno = (i==0 ? pWriter->aWriter[0].pgno : pDlidx[-1].pgno);
      assert( pDlidx->buf.n==0 );
      sqlite3Fts5BufferAppendVarint(&p->rc, &pDlidx->buf, !bDone);
      sqlite3Fts5BufferAppendVarint(&p->rc, &pDlidx->buf, iPgno);
      iVal = iRowid;
    }

    sqlite3Fts5BufferAppendVarint(&p->rc, &pDlidx->buf, iVal);
    pDlidx->bPrevValid = 1;
    pDlidx->iPrev = iRowid;
  }
}

static void fts5WriteFlushLeaf(Fts5Index *p, Fts5SegWriter *pWriter){
  static const u8 zero[] = { 0x00, 0x00, 0x00, 0x00 };
  Fts5PageWriter *pPage = &pWriter->aWriter[0];
  i64 iRowid;

  if( pWriter->bFirstTermInPage ){
    /* No term was written to this page. */
    assert( 0==fts5GetU16(&pPage->buf.p[2]) );
    fts5WriteBtreeNoTerm(p, pWriter);
  }

  /* Write the current page to the db. */
  iRowid = FTS5_SEGMENT_ROWID(pWriter->iSegid, 0, pPage->pgno);
  fts5DataWrite(p, iRowid, pPage->buf.p, pPage->buf.n);

  /* Initialize the next page. */
  fts5BufferZero(&pPage->buf);
  fts5BufferAppendBlob(&p->rc, &pPage->buf, 4, zero);
  pPage->pgno++;

  /* Increase the leaves written counter */
  pWriter->nLeafWritten++;

  /* The new leaf holds no terms or rowids */
  pWriter->bFirstTermInPage = 1;
  pWriter->bFirstRowidInPage = 1;
}

/*
** Append term pTerm/nTerm to the segment being written by the writer passed
** as the second argument.
**
** If an error occurs, set the Fts5Index.rc error code. If an error has 
** already occurred, this function is a no-op.
*/
static void fts5WriteAppendTerm(
  Fts5Index *p, 
  Fts5SegWriter *pWriter,
  int nTerm, const u8 *pTerm 
){
  int nPrefix;                    /* Bytes of prefix compression for term */
  Fts5PageWriter *pPage = &pWriter->aWriter[0];

  assert( pPage->buf.n==0 || pPage->buf.n>4 );
  if( pPage->buf.n==0 ){
    /* Zero the first term and first docid fields */
    static const u8 zero[] = { 0x00, 0x00, 0x00, 0x00 };
    fts5BufferAppendBlob(&p->rc, &pPage->buf, 4, zero);
    assert( pWriter->bFirstTermInPage );
  }
  if( p->rc ) return;
  
  if( pWriter->bFirstTermInPage ){
    /* Update the "first term" field of the page header. */
    assert( pPage->buf.p[2]==0 && pPage->buf.p[3]==0 );
    fts5PutU16(&pPage->buf.p[2], pPage->buf.n);
    nPrefix = 0;
    if( pPage->pgno!=1 ){
      /* This is the first term on a leaf that is not the leftmost leaf in
      ** the segment b-tree. In this case it is necessary to add a term to
      ** the b-tree hierarchy that is (a) larger than the largest term 
      ** already written to the segment and (b) smaller than or equal to
      ** this term. In other words, a prefix of (pTerm/nTerm) that is one
      ** byte longer than the longest prefix (pTerm/nTerm) shares with the
      ** previous term. 
      **
      ** Usually, the previous term is available in pPage->term. The exception
      ** is if this is the first term written in an incremental-merge step.
      ** In this case the previous term is not available, so just write a
      ** copy of (pTerm/nTerm) into the parent node. This is slightly
      ** inefficient, but still correct.  */
      int n = nTerm;
      if( pPage->term.n ){
        n = 1 + fts5PrefixCompress(pPage->term.n, pPage->term.p, nTerm, pTerm);
      }
      fts5WriteBtreeTerm(p, pWriter, n, pTerm);
      pPage = &pWriter->aWriter[0];
    }
  }else{
    nPrefix = fts5PrefixCompress(pPage->term.n, pPage->term.p, nTerm, pTerm);
    fts5BufferAppendVarint(&p->rc, &pPage->buf, nPrefix);
  }

  /* Append the number of bytes of new data, then the term data itself
  ** to the page. */
  fts5BufferAppendVarint(&p->rc, &pPage->buf, nTerm - nPrefix);
  fts5BufferAppendBlob(&p->rc, &pPage->buf, nTerm - nPrefix, &pTerm[nPrefix]);

  /* Update the Fts5PageWriter.term field. */
  fts5BufferSet(&p->rc, &pPage->term, nTerm, pTerm);
  pWriter->bFirstTermInPage = 0;

  pWriter->bFirstRowidInPage = 0;
  pWriter->bFirstRowidInDoclist = 1;

  assert( p->rc || (pWriter->nDlidx>0 && pWriter->aDlidx[0].buf.n==0) );
  pWriter->aDlidx[0].pgno = pPage->pgno;

  /* If the current leaf page is full, flush it to disk. */
  if( pPage->buf.n>=p->pConfig->pgsz ){
    fts5WriteFlushLeaf(p, pWriter);
  }
}

/*
** Append a docid and position-list size field to the writers output. 
*/
static void fts5WriteAppendRowid(
  Fts5Index *p, 
  Fts5SegWriter *pWriter,
  i64 iRowid,
  int nPos
){
  if( p->rc==SQLITE_OK ){
    Fts5PageWriter *pPage = &pWriter->aWriter[0];

    /* If this is to be the first docid written to the page, set the 
    ** docid-pointer in the page-header. Also append a value to the dlidx
    ** buffer, in case a doclist-index is required.  */
    if( pWriter->bFirstRowidInPage ){
      fts5PutU16(pPage->buf.p, pPage->buf.n);
      fts5WriteDlidxAppend(p, pWriter, iRowid);
    }

    /* Write the docid. */
    if( pWriter->bFirstRowidInDoclist || pWriter->bFirstRowidInPage ){
      fts5BufferAppendVarint(&p->rc, &pPage->buf, iRowid);
    }else{
      assert( p->rc || iRowid>pWriter->iPrevRowid );
      fts5BufferAppendVarint(&p->rc, &pPage->buf, iRowid - pWriter->iPrevRowid);
    }
    pWriter->iPrevRowid = iRowid;
    pWriter->bFirstRowidInDoclist = 0;
    pWriter->bFirstRowidInPage = 0;

    fts5BufferAppendVarint(&p->rc, &pPage->buf, nPos);

    if( pPage->buf.n>=p->pConfig->pgsz ){
      fts5WriteFlushLeaf(p, pWriter);
    }
  }
}

static void fts5WriteAppendPoslistData(
  Fts5Index *p, 
  Fts5SegWriter *pWriter, 
  const u8 *aData, 
  int nData
){
  Fts5PageWriter *pPage = &pWriter->aWriter[0];
  const u8 *a = aData;
  int n = nData;
  
  assert( p->pConfig->pgsz>0 );
  while( p->rc==SQLITE_OK && (pPage->buf.n + n)>=p->pConfig->pgsz ){
    int nReq = p->pConfig->pgsz - pPage->buf.n;
    int nCopy = 0;
    while( nCopy<nReq ){
      i64 dummy;
      nCopy += fts5GetVarint(&a[nCopy], (u64*)&dummy);
    }
    fts5BufferAppendBlob(&p->rc, &pPage->buf, nCopy, a);
    a += nCopy;
    n -= nCopy;
    fts5WriteFlushLeaf(p, pWriter);
  }
  if( n>0 ){
    fts5BufferAppendBlob(&p->rc, &pPage->buf, n, a);
  }
}

static void fts5WriteAppendZerobyte(Fts5Index *p, Fts5SegWriter *pWriter){
  fts5BufferAppendVarint(&p->rc, &pWriter->aWriter[0].buf, 0);
}

/*
** Flush any data cached by the writer object to the database. Free any
** allocations associated with the writer.
*/
static void fts5WriteFinish(
  Fts5Index *p, 
  Fts5SegWriter *pWriter,         /* Writer object */
  int *pnHeight,                  /* OUT: Height of the b-tree */
  int *pnLeaf                     /* OUT: Number of leaf pages in b-tree */
){
  int i;
  if( p->rc==SQLITE_OK ){
    Fts5PageWriter *pLeaf = &pWriter->aWriter[0];
    if( pLeaf->pgno==1 && pLeaf->buf.n==0 ){
      *pnLeaf = 0;
      *pnHeight = 0;
    }else{
      if( pLeaf->buf.n>4 ){
        fts5WriteFlushLeaf(p, pWriter);
      }
      *pnLeaf = pLeaf->pgno-1;
      if( pWriter->nWriter==1 && pWriter->nEmpty>=FTS5_MIN_DLIDX_SIZE ){
        fts5WriteBtreeGrow(p, pWriter);
      }
      if( pWriter->nWriter>1 ){
        fts5WriteBtreeNEmpty(p, pWriter);
      }
      *pnHeight = pWriter->nWriter;

      for(i=1; i<pWriter->nWriter; i++){
        Fts5PageWriter *pPg = &pWriter->aWriter[i];
        fts5DataWrite(p, 
            FTS5_SEGMENT_ROWID(pWriter->iSegid, i, pPg->pgno), 
            pPg->buf.p, pPg->buf.n
        );
      }
    }
  }
  for(i=0; i<pWriter->nWriter; i++){
    Fts5PageWriter *pPg = &pWriter->aWriter[i];
    fts5BufferFree(&pPg->term);
    fts5BufferFree(&pPg->buf);
  }
  sqlite3_free(pWriter->aWriter);

  for(i=0; i<pWriter->nDlidx; i++){
    sqlite3Fts5BufferFree(&pWriter->aDlidx[i].buf);
  }
  sqlite3_free(pWriter->aDlidx);
}

static void fts5WriteInit(
  Fts5Index *p, 
  Fts5SegWriter *pWriter, 
  int iSegid
){
  memset(pWriter, 0, sizeof(Fts5SegWriter));
  pWriter->iSegid = iSegid;

  pWriter->aWriter = (Fts5PageWriter*)fts5IdxMalloc(p, sizeof(Fts5PageWriter));
  if( fts5WriteDlidxGrow(p, pWriter, 1) ) return;
  pWriter->nWriter = 1;
  pWriter->nDlidx = 1;
  pWriter->aWriter[0].pgno = 1;
  pWriter->bFirstTermInPage = 1;
}

static void fts5WriteInitForAppend(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5SegWriter *pWriter,         /* Writer to initialize */
  Fts5StructureSegment *pSeg      /* Segment object to append to */
){
  int nByte = pSeg->nHeight * sizeof(Fts5PageWriter);
  memset(pWriter, 0, sizeof(Fts5SegWriter));
  pWriter->iSegid = pSeg->iSegid;
  pWriter->aWriter = (Fts5PageWriter*)fts5IdxMalloc(p, nByte);
  pWriter->aDlidx = (Fts5DlidxWriter*)fts5IdxMalloc(p, sizeof(Fts5DlidxWriter));

  if( p->rc==SQLITE_OK ){
    int pgno = 1;
    int i;
    pWriter->nDlidx = 1;
    pWriter->nWriter = pSeg->nHeight;
    pWriter->aWriter[0].pgno = pSeg->pgnoLast+1;
    for(i=pSeg->nHeight-1; i>0; i--){
      i64 iRowid = FTS5_SEGMENT_ROWID(pWriter->iSegid, i, pgno);
      Fts5PageWriter *pPg = &pWriter->aWriter[i];
      pPg->pgno = pgno;
      fts5DataBuffer(p, &pPg->buf, iRowid);
      if( p->rc==SQLITE_OK ){
        Fts5NodeIter ss;
        fts5NodeIterInit(pPg->buf.p, pPg->buf.n, &ss);
        while( ss.aData ) fts5NodeIterNext(&p->rc, &ss);
        fts5BufferSet(&p->rc, &pPg->term, ss.term.n, ss.term.p);
        pgno = ss.iChild;
        fts5NodeIterFree(&ss);
      }
    }
    assert( p->rc!=SQLITE_OK || (pgno+pWriter->nEmpty)==pSeg->pgnoLast );
    pWriter->bFirstTermInPage = 1;
    assert( pWriter->aWriter[0].term.n==0 );
  }
}

/*
** Iterator pIter was used to iterate through the input segments of on an
** incremental merge operation. This function is called if the incremental
** merge step has finished but the input has not been completely exhausted.
*/
static void fts5TrimSegments(Fts5Index *p, Fts5MultiSegIter *pIter){
  int i;
  Fts5Buffer buf;
  memset(&buf, 0, sizeof(Fts5Buffer));
  for(i=0; i<pIter->nSeg; i++){
    Fts5SegIter *pSeg = &pIter->aSeg[i];
    if( pSeg->pSeg==0 ){
      /* no-op */
    }else if( pSeg->pLeaf==0 ){
      /* All keys from this input segment have been transfered to the output.
      ** Set both the first and last page-numbers to 0 to indicate that the
      ** segment is now empty. */
      pSeg->pSeg->pgnoLast = 0;
      pSeg->pSeg->pgnoFirst = 0;
    }else{
      int iOff = pSeg->iTermLeafOffset;     /* Offset on new first leaf page */
      i64 iLeafRowid;
      Fts5Data *pData;
      int iId = pSeg->pSeg->iSegid;
      u8 aHdr[4] = {0x00, 0x00, 0x00, 0x04};

      iLeafRowid = FTS5_SEGMENT_ROWID(iId, 0, pSeg->iTermLeafPgno);
      pData = fts5DataRead(p, iLeafRowid);
      if( pData ){
        fts5BufferZero(&buf);
        fts5BufferAppendBlob(&p->rc, &buf, sizeof(aHdr), aHdr);
        fts5BufferAppendVarint(&p->rc, &buf, pSeg->term.n);
        fts5BufferAppendBlob(&p->rc, &buf, pSeg->term.n, pSeg->term.p);
        fts5BufferAppendBlob(&p->rc, &buf, pData->n - iOff, &pData->p[iOff]);
        fts5DataRelease(pData);
        pSeg->pSeg->pgnoFirst = pSeg->iTermLeafPgno;
        fts5DataDelete(p, FTS5_SEGMENT_ROWID(iId, 0, 1),iLeafRowid);
        fts5DataWrite(p, iLeafRowid, buf.p, buf.n);
      }
    }
  }
  fts5BufferFree(&buf);
}

static void fts5MergeChunkCallback(
  Fts5Index *p, 
  void *pCtx, 
  const u8 *pChunk, int nChunk
){
  Fts5SegWriter *pWriter = (Fts5SegWriter*)pCtx;
  fts5WriteAppendPoslistData(p, pWriter, pChunk, nChunk);
}

/*
**
*/
static void fts5IndexMergeLevel(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5Structure **ppStruct,       /* IN/OUT: Stucture of index */
  int iLvl,                       /* Level to read input from */
  int *pnRem                      /* Write up to this many output leaves */
){
  Fts5Structure *pStruct = *ppStruct;
  Fts5StructureLevel *pLvl = &pStruct->aLevel[iLvl];
  Fts5StructureLevel *pLvlOut;
  Fts5MultiSegIter *pIter = 0;    /* Iterator to read input data */
  int nRem = pnRem ? *pnRem : 0;  /* Output leaf pages left to write */
  int nInput;                     /* Number of input segments */
  Fts5SegWriter writer;           /* Writer object */
  Fts5StructureSegment *pSeg;     /* Output segment */
  Fts5Buffer term;
  int bRequireDoclistTerm = 0;    /* Doclist terminator (0x00) required */
  int bOldest;                    /* True if the output segment is the oldest */

  assert( iLvl<pStruct->nLevel );
  assert( pLvl->nMerge<=pLvl->nSeg );

  memset(&writer, 0, sizeof(Fts5SegWriter));
  memset(&term, 0, sizeof(Fts5Buffer));
  if( pLvl->nMerge ){
    pLvlOut = &pStruct->aLevel[iLvl+1];
    assert( pLvlOut->nSeg>0 );
    nInput = pLvl->nMerge;
    fts5WriteInitForAppend(p, &writer, &pLvlOut->aSeg[pLvlOut->nSeg-1]);
    pSeg = &pLvlOut->aSeg[pLvlOut->nSeg-1];
  }else{
    int iSegid = fts5AllocateSegid(p, pStruct);

    /* Extend the Fts5Structure object as required to ensure the output
    ** segment exists. */
    if( iLvl==pStruct->nLevel-1 ){
      fts5StructureAddLevel(&p->rc, ppStruct);
      pStruct = *ppStruct;
    }
    fts5StructureExtendLevel(&p->rc, pStruct, iLvl+1, 1, 0);
    if( p->rc ) return;
    pLvl = &pStruct->aLevel[iLvl];
    pLvlOut = &pStruct->aLevel[iLvl+1];

    fts5WriteInit(p, &writer, iSegid);

    /* Add the new segment to the output level */
    pSeg = &pLvlOut->aSeg[pLvlOut->nSeg];
    pLvlOut->nSeg++;
    pSeg->pgnoFirst = 1;
    pSeg->iSegid = iSegid;
    pStruct->nSegment++;

    /* Read input from all segments in the input level */
    nInput = pLvl->nSeg;
  }
  bOldest = (pLvlOut->nSeg==1 && pStruct->nLevel==iLvl+2);

  assert( iLvl>=0 );
  for(fts5MultiIterNew(p, pStruct, 0, 0, 0, 0, iLvl, nInput, &pIter);
      fts5MultiIterEof(p, pIter)==0;
      fts5MultiIterNext(p, pIter, 0, 0)
  ){
    Fts5SegIter *pSeg = &pIter->aSeg[ pIter->aFirst[1].iFirst ];
    int nPos;                     /* position-list size field value */
    int nTerm;
    const u8 *pTerm;

    /* Check for key annihilation. */
    if( pSeg->nPos==0 && (bOldest || pSeg->bDel==0) ) continue;

    pTerm = fts5MultiIterTerm(pIter, &nTerm);
    if( nTerm!=term.n || memcmp(pTerm, term.p, nTerm) ){
      if( pnRem && writer.nLeafWritten>nRem ){
        break;
      }

      /* This is a new term. Append a term to the output segment. */
      if( bRequireDoclistTerm ){
        fts5WriteAppendZerobyte(p, &writer);
      }
      fts5WriteAppendTerm(p, &writer, nTerm, pTerm);
      fts5BufferSet(&p->rc, &term, nTerm, pTerm);
      bRequireDoclistTerm = 1;
    }

    /* Append the rowid to the output */
    /* WRITEPOSLISTSIZE */
    nPos = pSeg->nPos*2 + pSeg->bDel;
    fts5WriteAppendRowid(p, &writer, fts5MultiIterRowid(pIter), nPos);

    /* Append the position-list data to the output */
    fts5ChunkIterate(p, pSeg, (void*)&writer, fts5MergeChunkCallback);
  }

  /* Flush the last leaf page to disk. Set the output segment b-tree height
  ** and last leaf page number at the same time.  */
  fts5WriteFinish(p, &writer, &pSeg->nHeight, &pSeg->pgnoLast);

  if( fts5MultiIterEof(p, pIter) ){
    int i;

    /* Remove the redundant segments from the %_data table */
    for(i=0; i<nInput; i++){
      fts5DataRemoveSegment(p, pLvl->aSeg[i].iSegid);
    }

    /* Remove the redundant segments from the input level */
    if( pLvl->nSeg!=nInput ){
      int nMove = (pLvl->nSeg - nInput) * sizeof(Fts5StructureSegment);
      memmove(pLvl->aSeg, &pLvl->aSeg[nInput], nMove);
    }
    pStruct->nSegment -= nInput;
    pLvl->nSeg -= nInput;
    pLvl->nMerge = 0;
    if( pSeg->pgnoLast==0 ){
      pLvlOut->nSeg--;
      pStruct->nSegment--;
    }
  }else{
    assert( pSeg->nHeight>0 && pSeg->pgnoLast>0 );
    fts5TrimSegments(p, pIter);
    pLvl->nMerge = nInput;
  }

  fts5MultiIterFree(p, pIter);
  fts5BufferFree(&term);
  if( pnRem ) *pnRem -= writer.nLeafWritten;
}

/*
** Do up to nPg pages of automerge work on the index.
*/
static void fts5IndexMerge(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5Structure **ppStruct,       /* IN/OUT: Current structure of index */
  int nPg                         /* Pages of work to do */
){
  int nRem = nPg;
  Fts5Structure *pStruct = *ppStruct;
  while( nRem>0 && p->rc==SQLITE_OK ){
    int iLvl;                   /* To iterate through levels */
    int iBestLvl = 0;           /* Level offering the most input segments */
    int nBest = 0;              /* Number of input segments on best level */

    /* Set iBestLvl to the level to read input segments from. */
    assert( pStruct->nLevel>0 );
    for(iLvl=0; iLvl<pStruct->nLevel; iLvl++){
      Fts5StructureLevel *pLvl = &pStruct->aLevel[iLvl];
      if( pLvl->nMerge ){
        if( pLvl->nMerge>nBest ){
          iBestLvl = iLvl;
          nBest = pLvl->nMerge;
        }
        break;
      }
      if( pLvl->nSeg>nBest ){
        nBest = pLvl->nSeg;
        iBestLvl = iLvl;
      }
    }

    /* If nBest is still 0, then the index must be empty. */
#ifdef SQLITE_DEBUG
    for(iLvl=0; nBest==0 && iLvl<pStruct->nLevel; iLvl++){
      assert( pStruct->aLevel[iLvl].nSeg==0 );
    }
#endif

    if( nBest<p->pConfig->nAutomerge 
        && pStruct->aLevel[iBestLvl].nMerge==0 
      ){
      break;
    }
    fts5IndexMergeLevel(p, &pStruct, iBestLvl, &nRem);
    if( p->rc==SQLITE_OK && pStruct->aLevel[iBestLvl].nMerge==0 ){
      fts5StructurePromote(p, iBestLvl+1, pStruct);
    }
  }
  *ppStruct = pStruct;
}

/*
** A total of nLeaf leaf pages of data has just been flushed to a level-0
** segment. This function updates the write-counter accordingly and, if
** necessary, performs incremental merge work.
**
** If an error occurs, set the Fts5Index.rc error code. If an error has 
** already occurred, this function is a no-op.
*/
static void fts5IndexAutomerge(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5Structure **ppStruct,       /* IN/OUT: Current structure of index */
  int nLeaf                       /* Number of output leaves just written */
){
  if( p->rc==SQLITE_OK && p->pConfig->nAutomerge>0 ){
    Fts5Structure *pStruct = *ppStruct;
    u64 nWrite;                   /* Initial value of write-counter */
    int nWork;                    /* Number of work-quanta to perform */
    int nRem;                     /* Number of leaf pages left to write */

    /* Update the write-counter. While doing so, set nWork. */
    nWrite = pStruct->nWriteCounter;
    nWork = (int)(((nWrite + nLeaf) / p->nWorkUnit) - (nWrite / p->nWorkUnit));
    pStruct->nWriteCounter += nLeaf;
    nRem = (int)(p->nWorkUnit * nWork * pStruct->nLevel);

    fts5IndexMerge(p, ppStruct, nRem);
  }
}

static void fts5IndexCrisismerge(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5Structure **ppStruct        /* IN/OUT: Current structure of index */
){
  const int nCrisis = p->pConfig->nCrisisMerge;
  Fts5Structure *pStruct = *ppStruct;
  int iLvl = 0;

  assert( p->rc!=SQLITE_OK || pStruct->nLevel>0 );
  while( p->rc==SQLITE_OK && pStruct->aLevel[iLvl].nSeg>=nCrisis ){
    fts5IndexMergeLevel(p, &pStruct, iLvl, 0);
    fts5StructurePromote(p, iLvl+1, pStruct);
    iLvl++;
  }
  *ppStruct = pStruct;
}

static int fts5IndexReturn(Fts5Index *p){
  int rc = p->rc;
  p->rc = SQLITE_OK;
  return rc;
}

typedef struct Fts5FlushCtx Fts5FlushCtx;
struct Fts5FlushCtx {
  Fts5Index *pIdx;
  Fts5SegWriter writer; 
};

/*
** Buffer aBuf[] contains a list of varints, all small enough to fit
** in a 32-bit integer. Return the size of the largest prefix of this 
** list nMax bytes or less in size.
*/
static int fts5PoslistPrefix(const u8 *aBuf, int nMax){
  int ret;
  u32 dummy;
  ret = fts5GetVarint32(aBuf, dummy);
  while( 1 ){
    int i = fts5GetVarint32(&aBuf[ret], dummy);
    if( (ret + i) > nMax ) break;
    ret += i;
  }
  return ret;
}

#define fts5BufferSafeAppendBlob(pBuf, pBlob, nBlob) { \
  assert( pBuf->nSpace>=(pBuf->n+nBlob) );             \
  memcpy(&pBuf->p[pBuf->n], pBlob, nBlob);             \
  pBuf->n += nBlob;                                    \
}

/*
** Flush the contents of in-memory hash table iHash to a new level-0 
** segment on disk. Also update the corresponding structure record.
**
** If an error occurs, set the Fts5Index.rc error code. If an error has 
** already occurred, this function is a no-op.
*/
static void fts5FlushOneHash(Fts5Index *p){
  Fts5Hash *pHash = p->pHash;
  Fts5Structure *pStruct;
  int iSegid;
  int pgnoLast = 0;                 /* Last leaf page number in segment */

  /* Obtain a reference to the index structure and allocate a new segment-id
  ** for the new level-0 segment.  */
  pStruct = fts5StructureRead(p);
  iSegid = fts5AllocateSegid(p, pStruct);

  if( iSegid ){
    const int pgsz = p->pConfig->pgsz;

    Fts5StructureSegment *pSeg;   /* New segment within pStruct */
    int nHeight;                  /* Height of new segment b-tree */
    Fts5Buffer *pBuf;             /* Buffer in which to assemble leaf page */
    const u8 *zPrev = 0;

    Fts5SegWriter writer;
    fts5WriteInit(p, &writer, iSegid);

    /* Pre-allocate the buffer used to assemble leaf pages to the target
    ** page size.  */
    assert( pgsz>0 );
    pBuf = &writer.aWriter[0].buf;
    fts5BufferGrow(&p->rc, pBuf, pgsz + 20);

    /* Begin scanning through hash table entries. This loop runs once for each
    ** term/doclist currently stored within the hash table. */
    if( p->rc==SQLITE_OK ){
      memset(pBuf->p, 0, 4);
      pBuf->n = 4;
      p->rc = sqlite3Fts5HashScanInit(pHash, 0, 0);
    }
    while( p->rc==SQLITE_OK && 0==sqlite3Fts5HashScanEof(pHash) ){
      const char *zTerm;          /* Buffer containing term */
      int nTerm;                  /* Size of zTerm in bytes */
      const u8 *pDoclist;         /* Pointer to doclist for this term */
      int nDoclist;               /* Size of doclist in bytes */
      int nSuffix;                /* Size of term suffix */

      sqlite3Fts5HashScanEntry(pHash, &zTerm, &pDoclist, &nDoclist);
      nTerm = strlen(zTerm);

      /* Decide if the term will fit on the current leaf. If it will not, 
      ** flush the leaf to disk here.  */
      if( (pBuf->n + nTerm + 2) > pgsz ){
        fts5WriteFlushLeaf(p, &writer);
        pBuf = &writer.aWriter[0].buf;
        if( (nTerm + 32) > pBuf->nSpace ){
          fts5BufferGrow(&p->rc, pBuf, nTerm + 32 - pBuf->n);
          if( p->rc ) break;
        }
      }

      /* Write the term to the leaf. And if it is the first on the leaf, and
      ** the leaf is not page number 1, push it up into the b-tree hierarchy 
      ** as well.  */
      if( writer.bFirstTermInPage==0 ){
        int nPre = fts5PrefixCompress(nTerm, zPrev, nTerm, (const u8*)zTerm);
        pBuf->n += sqlite3Fts5PutVarint(&pBuf->p[pBuf->n], nPre);
        nSuffix = nTerm - nPre;
      }else{
        fts5PutU16(&pBuf->p[2], pBuf->n);
        writer.bFirstTermInPage = 0;
        if( writer.aWriter[0].pgno!=1 ){
          int nPre = fts5PrefixCompress(nTerm, zPrev, nTerm, (const u8*)zTerm);
          fts5WriteBtreeTerm(p, &writer, nPre+1, (const u8*)zTerm);
          pBuf = &writer.aWriter[0].buf;
          assert( nPre<nTerm );
        }
        nSuffix = nTerm;
      }
      pBuf->n += sqlite3Fts5PutVarint(&pBuf->p[pBuf->n], nSuffix);
      fts5BufferSafeAppendBlob(pBuf, (const u8*)&zTerm[nTerm-nSuffix], nSuffix);

      /* We just wrote a term into page writer.aWriter[0].pgno. If a 
      ** doclist-index is to be generated for this doclist, it will be
      ** associated with this page. */
      assert( writer.nDlidx>0 && writer.aDlidx[0].buf.n==0 );
      writer.aDlidx[0].pgno = writer.aWriter[0].pgno;

      if( pgsz>=(pBuf->n + nDoclist + 1) ){
        /* The entire doclist will fit on the current leaf. */
        fts5BufferSafeAppendBlob(pBuf, pDoclist, nDoclist);
      }else{
        i64 iRowid = 0;
        i64 iDelta = 0;
        int iOff = 0;

        writer.bFirstRowidInPage = 0;

        /* The entire doclist will not fit on this leaf. The following 
        ** loop iterates through the poslists that make up the current 
        ** doclist.  */
        while( p->rc==SQLITE_OK && iOff<nDoclist ){
          int nPos;
          int nCopy;
          int bDummy;
          iOff += fts5GetVarint(&pDoclist[iOff], (u64*)&iDelta);
          nCopy = fts5GetPoslistSize(&pDoclist[iOff], &nPos, &bDummy);
          nCopy += nPos;
          iRowid += iDelta;
          
          if( writer.bFirstRowidInPage ){
            fts5PutU16(&pBuf->p[0], pBuf->n);   /* first docid on page */
            pBuf->n += sqlite3Fts5PutVarint(&pBuf->p[pBuf->n], iRowid);
            writer.bFirstRowidInPage = 0;
            fts5WriteDlidxAppend(p, &writer, iRowid);
          }else{
            pBuf->n += sqlite3Fts5PutVarint(&pBuf->p[pBuf->n], iDelta);
          }
          assert( pBuf->n<=pBuf->nSpace );

          if( (pBuf->n + nCopy) <= pgsz ){
            /* The entire poslist will fit on the current leaf. So copy
            ** it in one go. */
            fts5BufferSafeAppendBlob(pBuf, &pDoclist[iOff], nCopy);
          }else{
            /* The entire poslist will not fit on this leaf. So it needs
            ** to be broken into sections. The only qualification being
            ** that each varint must be stored contiguously.  */
            const u8 *pPoslist = &pDoclist[iOff];
            int iPos = 0;
            while( p->rc==SQLITE_OK ){
              int nSpace = pgsz - pBuf->n;
              int n = 0;
              if( (nCopy - iPos)<=nSpace ){
                n = nCopy - iPos;
              }else{
                n = fts5PoslistPrefix(&pPoslist[iPos], nSpace);
              }
              assert( n>0 );
              fts5BufferSafeAppendBlob(pBuf, &pPoslist[iPos], n);
              iPos += n;
              if( pBuf->n>=pgsz ){
                fts5WriteFlushLeaf(p, &writer);
                pBuf = &writer.aWriter[0].buf;
              }
              if( iPos>=nCopy ) break;
            }
          }
          iOff += nCopy;
        }
      }

      pBuf->p[pBuf->n++] = '\0';
      assert( pBuf->n<=pBuf->nSpace );
      zPrev = (const u8*)zTerm;
      sqlite3Fts5HashScanNext(pHash);
    }
    sqlite3Fts5HashClear(pHash);
    fts5WriteFinish(p, &writer, &nHeight, &pgnoLast);

    /* Update the Fts5Structure. It is written back to the database by the
    ** fts5StructureRelease() call below.  */
    if( pStruct->nLevel==0 ){
      fts5StructureAddLevel(&p->rc, &pStruct);
    }
    fts5StructureExtendLevel(&p->rc, pStruct, 0, 1, 0);
    if( p->rc==SQLITE_OK ){
      pSeg = &pStruct->aLevel[0].aSeg[ pStruct->aLevel[0].nSeg++ ];
      pSeg->iSegid = iSegid;
      pSeg->nHeight = nHeight;
      pSeg->pgnoFirst = 1;
      pSeg->pgnoLast = pgnoLast;
      pStruct->nSegment++;
    }
    fts5StructurePromote(p, 0, pStruct);
  }


  fts5IndexAutomerge(p, &pStruct, pgnoLast);
  fts5IndexCrisismerge(p, &pStruct);
  fts5StructureWrite(p, pStruct);
  fts5StructureRelease(pStruct);
}

/*
** Flush any data stored in the in-memory hash tables to the database.
*/
static void fts5IndexFlush(Fts5Index *p){
  /* Unless it is empty, flush the hash table to disk */
  if( p->nPendingData ){
    assert( p->pHash );
    p->nPendingData = 0;
    fts5FlushOneHash(p);
  }
}


int sqlite3Fts5IndexOptimize(Fts5Index *p){
  Fts5Structure *pStruct;
  Fts5Structure *pNew = 0;
  int nSeg = 0;

  assert( p->rc==SQLITE_OK );
  fts5IndexFlush(p);
  pStruct = fts5StructureRead(p);

  if( pStruct ){
    assert( pStruct->nSegment==fts5StructureCountSegments(pStruct) );
    nSeg = pStruct->nSegment;
    if( nSeg>1 ){
      int nByte = sizeof(Fts5Structure);
      nByte += (pStruct->nLevel+1) * sizeof(Fts5StructureLevel);
      pNew = (Fts5Structure*)sqlite3Fts5MallocZero(&p->rc, nByte);
    }
  }
  if( pNew ){
    Fts5StructureLevel *pLvl;
    int nByte = nSeg * sizeof(Fts5StructureSegment);
    pNew->nLevel = pStruct->nLevel+1;
    pNew->nWriteCounter = pStruct->nWriteCounter;
    pLvl = &pNew->aLevel[pStruct->nLevel];
    pLvl->aSeg = (Fts5StructureSegment*)sqlite3Fts5MallocZero(&p->rc, nByte);
    if( pLvl->aSeg ){
      int iLvl, iSeg;
      int iSegOut = 0;
      for(iLvl=0; iLvl<pStruct->nLevel; iLvl++){
        for(iSeg=0; iSeg<pStruct->aLevel[iLvl].nSeg; iSeg++){
          pLvl->aSeg[iSegOut] = pStruct->aLevel[iLvl].aSeg[iSeg];
          iSegOut++;
        }
      }
      pNew->nSegment = pLvl->nSeg = nSeg;
    }else{
      sqlite3_free(pNew);
      pNew = 0;
    }
  }

  if( pNew ){
    int iLvl = pNew->nLevel-1;
    while( p->rc==SQLITE_OK && pNew->aLevel[iLvl].nSeg>0 ){
      int nRem = FTS5_OPT_WORK_UNIT;
      fts5IndexMergeLevel(p, &pNew, iLvl, &nRem);
    }

    fts5StructureWrite(p, pNew);
    fts5StructureRelease(pNew);
  }

  fts5StructureRelease(pStruct);
  return fts5IndexReturn(p); 
}

int sqlite3Fts5IndexMerge(Fts5Index *p, int nMerge){
  Fts5Structure *pStruct;

  pStruct = fts5StructureRead(p);
  if( pStruct && pStruct->nLevel ){
    fts5IndexMerge(p, &pStruct, nMerge);
    fts5StructureWrite(p, pStruct);
  }
  fts5StructureRelease(pStruct);

  return fts5IndexReturn(p);
}

static void fts5PoslistCallback(
  Fts5Index *p, 
  void *pCtx, 
  const u8 *pChunk, int nChunk
){
  fts5BufferAppendBlob(&p->rc, (Fts5Buffer*)pCtx, nChunk, pChunk);
}

/*
** Iterator pIter currently points to a valid entry (not EOF). This
** function appends the position list data for the current entry to
** buffer pBuf. It does not make a copy of the position-list size
** field.
*/
static void fts5SegiterPoslist(
  Fts5Index *p,
  Fts5SegIter *pSeg,
  Fts5Buffer *pBuf
){
  fts5ChunkIterate(p, pSeg, (void*)pBuf, fts5PoslistCallback);
}

/*
** Iterator pMulti currently points to a valid entry (not EOF). This
** function appends a copy of the position-list of the entry pMulti 
** currently points to to buffer pBuf.
**
** If an error occurs, an error code is left in p->rc. It is assumed
** no error has already occurred when this function is called.
*/
static void fts5MultiIterPoslist(
  Fts5Index *p,
  Fts5MultiSegIter *pMulti,
  int bSz,                        /* Append a size field before the data */
  Fts5Buffer *pBuf
){
  if( p->rc==SQLITE_OK ){
    Fts5SegIter *pSeg = &pMulti->aSeg[ pMulti->aFirst[1].iFirst ];
    assert( fts5MultiIterEof(p, pMulti)==0 );

    if( bSz ){
      /* WRITEPOSLISTSIZE */
      fts5BufferAppendVarint(&p->rc, pBuf, pSeg->nPos*2);
    }
    fts5SegiterPoslist(p, pSeg, pBuf);
  }
}

static void fts5DoclistIterNext(Fts5DoclistIter *pIter){
  if( pIter->i<pIter->n ){
    int bDummy;
    if( pIter->i ){
      i64 iDelta;
      pIter->i += fts5GetVarint(&pIter->a[pIter->i], (u64*)&iDelta);
      pIter->iRowid += iDelta;
    }else{
      pIter->i += fts5GetVarint(&pIter->a[pIter->i], (u64*)&pIter->iRowid);
    }
    pIter->i += fts5GetPoslistSize(
        &pIter->a[pIter->i], &pIter->nPoslist, &bDummy
    );
    pIter->aPoslist = &pIter->a[pIter->i];
    pIter->i += pIter->nPoslist;
  }else{
    pIter->aPoslist = 0;
  }
}

static void fts5DoclistIterInit(
  Fts5Buffer *pBuf, 
  Fts5DoclistIter *pIter
){
  memset(pIter, 0, sizeof(*pIter));
  pIter->a = pBuf->p;
  pIter->n = pBuf->n;
  fts5DoclistIterNext(pIter);
}

/*
** Append a doclist to buffer pBuf.
*/
static void fts5MergeAppendDocid(
  int *pRc,                       /* IN/OUT: Error code */
  Fts5Buffer *pBuf,               /* Buffer to write to */
  i64 *piLastRowid,               /* IN/OUT: Previous rowid written (if any) */
  i64 iRowid                      /* Rowid to append */
){
  if( pBuf->n==0 ){
    fts5BufferAppendVarint(pRc, pBuf, iRowid);
  }else{
    fts5BufferAppendVarint(pRc, pBuf, iRowid - *piLastRowid);
  }
  *piLastRowid = iRowid;
}

/*
** Buffers p1 and p2 contain doclists. This function merges the content
** of the two doclists together and sets buffer p1 to the result before
** returning.
**
** If an error occurs, an error code is left in p->rc. If an error has
** already occurred, this function is a no-op.
*/
static void fts5MergePrefixLists(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5Buffer *p1,                 /* First list to merge */
  Fts5Buffer *p2                  /* Second list to merge */
){
  if( p2->n ){
    i64 iLastRowid = 0;
    Fts5DoclistIter i1;
    Fts5DoclistIter i2;
    Fts5Buffer out;
    Fts5Buffer tmp;
    memset(&out, 0, sizeof(out));
    memset(&tmp, 0, sizeof(tmp));

    fts5DoclistIterInit(p1, &i1);
    fts5DoclistIterInit(p2, &i2);
    while( p->rc==SQLITE_OK && (i1.aPoslist!=0 || i2.aPoslist!=0) ){
      if( i2.aPoslist==0 || (i1.aPoslist && i1.iRowid<i2.iRowid) ){
        /* Copy entry from i1 */
        fts5MergeAppendDocid(&p->rc, &out, &iLastRowid, i1.iRowid);
        /* WRITEPOSLISTSIZE */
        fts5BufferAppendVarint(&p->rc, &out, i1.nPoslist * 2);
        fts5BufferAppendBlob(&p->rc, &out, i1.nPoslist, i1.aPoslist);
        fts5DoclistIterNext(&i1);
      }
      else if( i1.aPoslist==0 || i2.iRowid!=i1.iRowid ){
        /* Copy entry from i2 */
        fts5MergeAppendDocid(&p->rc, &out, &iLastRowid, i2.iRowid);
        /* WRITEPOSLISTSIZE */
        fts5BufferAppendVarint(&p->rc, &out, i2.nPoslist * 2);
        fts5BufferAppendBlob(&p->rc, &out, i2.nPoslist, i2.aPoslist);
        fts5DoclistIterNext(&i2);
      }
      else{
        Fts5PoslistReader r1;
        Fts5PoslistReader r2;
        Fts5PoslistWriter writer;

        memset(&writer, 0, sizeof(writer));

        /* Merge the two position lists. */ 
        fts5MergeAppendDocid(&p->rc, &out, &iLastRowid, i2.iRowid);
        fts5BufferZero(&tmp);
        sqlite3Fts5PoslistReaderInit(-1, i1.aPoslist, i1.nPoslist, &r1);
        sqlite3Fts5PoslistReaderInit(-1, i2.aPoslist, i2.nPoslist, &r2);
        while( p->rc==SQLITE_OK && (r1.bEof==0 || r2.bEof==0) ){
          i64 iNew;
          if( r2.bEof || (r1.bEof==0 && r1.iPos<r2.iPos) ){
            iNew = r1.iPos;
            sqlite3Fts5PoslistReaderNext(&r1);
          }else{
            iNew = r2.iPos;
            sqlite3Fts5PoslistReaderNext(&r2);
            if( r1.iPos==r2.iPos ) sqlite3Fts5PoslistReaderNext(&r1);
          }
          p->rc = sqlite3Fts5PoslistWriterAppend(&tmp, &writer, iNew);
        }

        /* WRITEPOSLISTSIZE */
        fts5BufferAppendVarint(&p->rc, &out, tmp.n * 2);
        fts5BufferAppendBlob(&p->rc, &out, tmp.n, tmp.p);
        fts5DoclistIterNext(&i1);
        fts5DoclistIterNext(&i2);
      }
    }

    fts5BufferSet(&p->rc, p1, out.n, out.p);
    fts5BufferFree(&tmp);
    fts5BufferFree(&out);
  }
}

static void fts5BufferSwap(Fts5Buffer *p1, Fts5Buffer *p2){
  Fts5Buffer tmp = *p1;
  *p1 = *p2;
  *p2 = tmp;
}

static void fts5SetupPrefixIter(
  Fts5Index *p,                   /* Index to read from */
  int bDesc,                      /* True for "ORDER BY rowid DESC" */
  const u8 *pToken,               /* Buffer containing prefix to match */
  int nToken,                     /* Size of buffer pToken in bytes */
  Fts5IndexIter *pIter            /* Populate this object */
){
  Fts5Structure *pStruct;
  Fts5Buffer *aBuf;
  const int nBuf = 32;

  aBuf = (Fts5Buffer*)fts5IdxMalloc(p, sizeof(Fts5Buffer)*nBuf);
  pStruct = fts5StructureRead(p);

  if( aBuf && pStruct ){
    const int flags = FTS5INDEX_QUERY_SCAN;
    int i;
    i64 iLastRowid = 0;
    Fts5MultiSegIter *p1 = 0;     /* Iterator used to gather data from index */
    Fts5Data *pData;
    Fts5Buffer doclist;

    memset(&doclist, 0, sizeof(doclist));
    for(fts5MultiIterNew(p, pStruct, 1, flags, pToken, nToken, -1, 0, &p1);
        fts5MultiIterEof(p, p1)==0;
        fts5MultiIterNext(p, p1, 0, 0)
    ){
      i64 iRowid = fts5MultiIterRowid(p1);
      int nTerm;
      const u8 *pTerm = fts5MultiIterTerm(p1, &nTerm);
      assert( memcmp(pToken, pTerm, MIN(nToken, nTerm))<=0 );
      if( nTerm<nToken || memcmp(pToken, pTerm, nToken) ) break;

      if( doclist.n>0 && iRowid<=iLastRowid ){
        for(i=0; p->rc==SQLITE_OK && doclist.n; i++){
          assert( i<nBuf );
          if( aBuf[i].n==0 ){
            fts5BufferSwap(&doclist, &aBuf[i]);
            fts5BufferZero(&doclist);
          }else{
            fts5MergePrefixLists(p, &doclist, &aBuf[i]);
            fts5BufferZero(&aBuf[i]);
          }
        }
      }

      fts5MergeAppendDocid(&p->rc, &doclist, &iLastRowid, iRowid);
      fts5MultiIterPoslist(p, p1, 1, &doclist);
    }

    for(i=0; i<nBuf; i++){
      fts5MergePrefixLists(p, &doclist, &aBuf[i]);
      fts5BufferFree(&aBuf[i]);
    }
    fts5MultiIterFree(p, p1);

    pData = fts5IdxMalloc(p, sizeof(Fts5Data) + doclist.n);
    if( pData ){
      pData->p = (u8*)&pData[1];
      pData->n = doclist.n;
      memcpy(pData->p, doclist.p, doclist.n);
      fts5MultiIterNew2(p, pData, bDesc, &pIter->pMulti);
    }
    fts5BufferFree(&doclist);
  }

  fts5StructureRelease(pStruct);
  sqlite3_free(aBuf);
}


/*
** Indicate that all subsequent calls to sqlite3Fts5IndexWrite() pertain
** to the document with rowid iRowid.
*/
int sqlite3Fts5IndexBeginWrite(Fts5Index *p, i64 iRowid){
  assert( p->rc==SQLITE_OK );

  /* Allocate the hash table if it has not already been allocated */
  if( p->pHash==0 ){
    p->rc = sqlite3Fts5HashNew(&p->pHash, &p->nPendingData);
  }

  /* Flush the hash table to disk if required */
  if( iRowid<=p->iWriteRowid || (p->nPendingData > p->nMaxPendingData) ){
    fts5IndexFlush(p);
  }
  p->iWriteRowid = iRowid;
  return fts5IndexReturn(p);
}

/*
** Commit data to disk.
*/
int sqlite3Fts5IndexSync(Fts5Index *p, int bCommit){
  assert( p->rc==SQLITE_OK );
  fts5IndexFlush(p);
  if( bCommit ) fts5CloseReader(p);
  return fts5IndexReturn(p);
}

/*
** Discard any data stored in the in-memory hash tables. Do not write it
** to the database. Additionally, assume that the contents of the %_data
** table may have changed on disk. So any in-memory caches of %_data 
** records must be invalidated.
*/
int sqlite3Fts5IndexRollback(Fts5Index *p){
  fts5CloseReader(p);
  fts5IndexDiscardData(p);
  assert( p->rc==SQLITE_OK );
  return SQLITE_OK;
}

/*
** The %_data table is completely empty when this function is called. This
** function populates it with the initial structure objects for each index,
** and the initial version of the "averages" record (a zero-byte blob).
*/
int sqlite3Fts5IndexReinit(Fts5Index *p){
  Fts5Structure s;

  assert( p->rc==SQLITE_OK );
  p->rc = sqlite3Fts5IndexSetAverages(p, (const u8*)"", 0);

  memset(&s, 0, sizeof(Fts5Structure));
  fts5StructureWrite(p, &s);

  return fts5IndexReturn(p);
}

/*
** Open a new Fts5Index handle. If the bCreate argument is true, create
** and initialize the underlying %_data table.
**
** If successful, set *pp to point to the new object and return SQLITE_OK.
** Otherwise, set *pp to NULL and return an SQLite error code.
*/
int sqlite3Fts5IndexOpen(
  Fts5Config *pConfig, 
  int bCreate, 
  Fts5Index **pp,
  char **pzErr
){
  int rc = SQLITE_OK;
  Fts5Index *p;                   /* New object */

  *pp = p = (Fts5Index*)sqlite3Fts5MallocZero(&rc, sizeof(Fts5Index));
  if( rc==SQLITE_OK ){
    p->pConfig = pConfig;
    p->nWorkUnit = FTS5_WORK_UNIT;
    p->nMaxPendingData = 1024*1024;
    p->zDataTbl = sqlite3Fts5Mprintf(&rc, "%s_data", pConfig->zName);
    if( p->zDataTbl && bCreate ){
      rc = sqlite3Fts5CreateTable(
          pConfig, "data", "id INTEGER PRIMARY KEY, block BLOB", 0, pzErr
      );
      if( rc==SQLITE_OK ){
        rc = sqlite3Fts5IndexReinit(p);
      }
    }
  }

  assert( rc!=SQLITE_OK || p->rc==SQLITE_OK );
  if( rc ){
    sqlite3Fts5IndexClose(p);
    *pp = 0;
  }
  return rc;
}

/*
** Close a handle opened by an earlier call to sqlite3Fts5IndexOpen().
*/
int sqlite3Fts5IndexClose(Fts5Index *p){
  int rc = SQLITE_OK;
  if( p ){
    assert( p->pReader==0 );
    sqlite3_finalize(p->pWriter);
    sqlite3_finalize(p->pDeleter);
    sqlite3Fts5HashFree(p->pHash);
    sqlite3Fts5BufferFree(&p->scratch);
    sqlite3_free(p->zDataTbl);
    sqlite3_free(p);
  }
  return rc;
}

/*
** Argument p points to a buffer containing utf-8 text that is n bytes in 
** size. Return the number of bytes in the nChar character prefix of the
** buffer, or 0 if there are less than nChar characters in total.
*/
static int fts5IndexCharlenToBytelen(const char *p, int nByte, int nChar){
  int n = 0;
  int i;
  for(i=0; i<nChar; i++){
    if( n>=nByte ) return 0;      /* Input contains fewer than nChar chars */
    if( (unsigned char)p[n++]>=0xc0 ){
      while( (p[n] & 0xc0)==0x80 ) n++;
    }
  }
  return n;
}

/*
** pIn is a UTF-8 encoded string, nIn bytes in size. Return the number of
** unicode characters in the string.
*/
static int fts5IndexCharlen(const char *pIn, int nIn){
  int nChar = 0;            
  int i = 0;
  while( i<nIn ){
    if( (unsigned char)pIn[i++]>=0xc0 ){
      while( i<nIn && (pIn[i] & 0xc0)==0x80 ) i++;
    }
    nChar++;
  }
  return nChar;
}

/*
** Insert or remove data to or from the index. Each time a document is 
** added to or removed from the index, this function is called one or more
** times.
**
** For an insert, it must be called once for each token in the new document.
** If the operation is a delete, it must be called (at least) once for each
** unique token in the document with an iCol value less than zero. The iPos
** argument is ignored for a delete.
*/
int sqlite3Fts5IndexWrite(
  Fts5Index *p,                   /* Index to write to */
  int iCol,                       /* Column token appears in (-ve -> delete) */
  int iPos,                       /* Position of token within column */
  const char *pToken, int nToken  /* Token to add or remove to or from index */
){
  int i;                          /* Used to iterate through indexes */
  int rc = SQLITE_OK;             /* Return code */
  Fts5Config *pConfig = p->pConfig;

  assert( p->rc==SQLITE_OK );

  /* Add the entry to the main terms index. */
  rc = sqlite3Fts5HashWrite(
      p->pHash, p->iWriteRowid, iCol, iPos, FTS5_MAIN_PREFIX, pToken, nToken
  );

  for(i=0; i<pConfig->nPrefix && rc==SQLITE_OK; i++){
    int nByte = fts5IndexCharlenToBytelen(pToken, nToken, pConfig->aPrefix[i]);
    if( nByte ){
      rc = sqlite3Fts5HashWrite(p->pHash, 
          p->iWriteRowid, iCol, iPos, FTS5_MAIN_PREFIX+i+1, pToken, nByte
      );
    }
  }

  return rc;
}

/*
** Open a new iterator to iterate though all docids that match the 
** specified token or token prefix.
*/
int sqlite3Fts5IndexQuery(
  Fts5Index *p,                   /* FTS index to query */
  const char *pToken, int nToken, /* Token (or prefix) to query for */
  int flags,                      /* Mask of FTS5INDEX_QUERY_X flags */
  Fts5IndexIter **ppIter          /* OUT: New iterator object */
){
  Fts5Config *pConfig = p->pConfig;
  Fts5IndexIter *pRet;
  int iIdx = 0;
  Fts5Buffer buf = {0, 0, 0};

  /* If the QUERY_SCAN flag is set, all other flags must be clear. */
  assert( (flags & FTS5INDEX_QUERY_SCAN)==0
       || (flags & FTS5INDEX_QUERY_SCAN)==FTS5INDEX_QUERY_SCAN
  );

  if( sqlite3Fts5BufferGrow(&p->rc, &buf, nToken+1)==0 ){
    memcpy(&buf.p[1], pToken, nToken);
  }

#ifdef SQLITE_DEBUG
  if( flags & FTS5INDEX_QUERY_TEST_NOIDX ){
    assert( flags & FTS5INDEX_QUERY_PREFIX );
    iIdx = 1+pConfig->nPrefix;
  }else
#endif
  if( flags & FTS5INDEX_QUERY_PREFIX ){
    int nChar = fts5IndexCharlen(pToken, nToken);
    for(iIdx=1; iIdx<=pConfig->nPrefix; iIdx++){
      if( pConfig->aPrefix[iIdx-1]==nChar ) break;
    }
  }

  pRet = (Fts5IndexIter*)sqlite3Fts5MallocZero(&p->rc, sizeof(Fts5IndexIter));
  if( pRet ){
    pRet->pIndex = p;
    if( iIdx<=pConfig->nPrefix ){
      buf.p[0] = FTS5_MAIN_PREFIX + iIdx;
      pRet->pStruct = fts5StructureRead(p);
      if( pRet->pStruct ){
        fts5MultiIterNew(
            p, pRet->pStruct, 1, flags, buf.p, nToken+1, -1, 0, &pRet->pMulti
        );
      }
    }else{
      int bDesc = (flags & FTS5INDEX_QUERY_DESC)!=0;
      buf.p[0] = FTS5_MAIN_PREFIX;
      fts5SetupPrefixIter(p, bDesc, buf.p, nToken+1, pRet);
    }
  }

  if( p->rc ){
    sqlite3Fts5IterClose(pRet);
    pRet = 0;
  }
  *ppIter = pRet;
  sqlite3Fts5BufferFree(&buf);
  return fts5IndexReturn(p);
}

/*
** Return true if the iterator passed as the only argument is at EOF.
*/
int sqlite3Fts5IterEof(Fts5IndexIter *pIter){
  assert( pIter->pIndex->rc==SQLITE_OK );
  return fts5MultiIterEof(pIter->pIndex, pIter->pMulti);
}

/*
** Move to the next matching rowid. 
*/
int sqlite3Fts5IterNext(Fts5IndexIter *pIter){
  assert( pIter->pIndex->rc==SQLITE_OK );
  fts5MultiIterNext(pIter->pIndex, pIter->pMulti, 0, 0);
  return fts5IndexReturn(pIter->pIndex);
}

/*
** Move to the next matching term/rowid. Used by the fts5vocab module.
*/
int sqlite3Fts5IterNextScan(Fts5IndexIter *pIter){
  Fts5Index *p = pIter->pIndex;
  Fts5MultiSegIter *pMulti = pIter->pMulti;

  assert( pIter->pIndex->rc==SQLITE_OK );
  assert( pMulti );

  fts5MultiIterNext(p, pMulti, 0, 0);
  if( p->rc==SQLITE_OK ){
    Fts5SegIter *pSeg = &pMulti->aSeg[ pMulti->aFirst[1].iFirst ];
    if( pSeg->pLeaf && pSeg->term.p[0]!=FTS5_MAIN_PREFIX ){
      fts5DataRelease(pSeg->pLeaf);
      pSeg->pLeaf = 0;
    }
  }

  return fts5IndexReturn(pIter->pIndex);
}

/*
** Move to the next matching rowid that occurs at or after iMatch. The
** definition of "at or after" depends on whether this iterator iterates
** in ascending or descending rowid order.
*/
int sqlite3Fts5IterNextFrom(Fts5IndexIter *pIter, i64 iMatch){
  fts5MultiIterNextFrom(pIter->pIndex, pIter->pMulti, iMatch);
  return fts5IndexReturn(pIter->pIndex);
}

/*
** Return the current rowid.
*/
i64 sqlite3Fts5IterRowid(Fts5IndexIter *pIter){
  return fts5MultiIterRowid(pIter->pMulti);
}

/*
** Return the current term.
*/
const char *sqlite3Fts5IterTerm(Fts5IndexIter *pIter, int *pn){
  int n;
  const char *z = (const char*)fts5MultiIterTerm(pIter->pMulti, &n);
  *pn = n-1;
  return &z[1];
}


/*
** Return a pointer to a buffer containing a copy of the position list for
** the current entry. Output variable *pn is set to the size of the buffer 
** in bytes before returning.
**
** The returned position list does not include the "number of bytes" varint
** field that starts the position list on disk.
*/
int sqlite3Fts5IterPoslist(
  Fts5IndexIter *pIter, 
  const u8 **pp,                  /* OUT: Pointer to position-list data */
  int *pn,                        /* OUT: Size of position-list in bytes */
  i64 *piRowid                    /* OUT: Current rowid */
){
  Fts5MultiSegIter *pMulti = pIter->pMulti;
  Fts5SegIter *pSeg = &pMulti->aSeg[ pMulti->aFirst[1].iFirst ];
  assert( pIter->pIndex->rc==SQLITE_OK );
  *piRowid = pSeg->iRowid;
  *pn = pSeg->nPos;
  if( pSeg->iLeafOffset+pSeg->nPos <= pSeg->pLeaf->n ){
    *pp = &pSeg->pLeaf->p[pSeg->iLeafOffset];
  }else{
    fts5BufferZero(&pIter->poslist);
    fts5SegiterPoslist(pIter->pIndex, pSeg, &pIter->poslist);
    *pp = pIter->poslist.p;
  }
  return fts5IndexReturn(pIter->pIndex);
}

/*
** This function is similar to sqlite3Fts5IterPoslist(), except that it
** copies the position list into the buffer supplied as the second 
** argument.
*/
int sqlite3Fts5IterPoslistBuffer(Fts5IndexIter *pIter, Fts5Buffer *pBuf){
  Fts5Index *p = pIter->pIndex;
  Fts5MultiSegIter *pMulti = pIter->pMulti;

  assert( p->rc==SQLITE_OK );
  fts5BufferZero(pBuf);
  fts5MultiIterPoslist(p, pMulti, 0, pBuf);
  return fts5IndexReturn(p);
}

/*
** Close an iterator opened by an earlier call to sqlite3Fts5IndexQuery().
*/
void sqlite3Fts5IterClose(Fts5IndexIter *pIter){
  if( pIter ){
    fts5MultiIterFree(pIter->pIndex, pIter->pMulti);
    fts5StructureRelease(pIter->pStruct);
    fts5BufferFree(&pIter->poslist);
    fts5CloseReader(pIter->pIndex);
    sqlite3_free(pIter);
  }
}

/*
** Read the "averages" record into the buffer supplied as the second 
** argument. Return SQLITE_OK if successful, or an SQLite error code
** if an error occurs.
*/
int sqlite3Fts5IndexGetAverages(Fts5Index *p, Fts5Buffer *pBuf){
  assert( p->rc==SQLITE_OK );
  fts5DataReadOrBuffer(p, pBuf, FTS5_AVERAGES_ROWID);
  return fts5IndexReturn(p);
}

/*
** Replace the current "averages" record with the contents of the buffer 
** supplied as the second argument.
*/
int sqlite3Fts5IndexSetAverages(Fts5Index *p, const u8 *pData, int nData){
  assert( p->rc==SQLITE_OK );
  fts5DataWrite(p, FTS5_AVERAGES_ROWID, pData, nData);
  return fts5IndexReturn(p);
}

/*
** Return the total number of blocks this module has read from the %_data
** table since it was created.
*/
int sqlite3Fts5IndexReads(Fts5Index *p){
  return p->nRead;
}

/*
** Set the 32-bit cookie value stored at the start of all structure 
** records to the value passed as the second argument.
**
** Return SQLITE_OK if successful, or an SQLite error code if an error
** occurs.
*/
int sqlite3Fts5IndexSetCookie(Fts5Index *p, int iNew){
  int rc;                              /* Return code */
  Fts5Config *pConfig = p->pConfig;    /* Configuration object */
  u8 aCookie[4];                       /* Binary representation of iNew */
  sqlite3_blob *pBlob = 0;

  assert( p->rc==SQLITE_OK );
  sqlite3Fts5Put32(aCookie, iNew);

  rc = sqlite3_blob_open(pConfig->db, pConfig->zDb, p->zDataTbl, 
      "block", FTS5_STRUCTURE_ROWID, 1, &pBlob
  );
  if( rc==SQLITE_OK ){
    sqlite3_blob_write(pBlob, aCookie, 4, 0);
    rc = sqlite3_blob_close(pBlob);
  }

  return rc;
}

int sqlite3Fts5IndexLoadConfig(Fts5Index *p){
  Fts5Structure *pStruct;
  pStruct = fts5StructureRead(p);
  fts5StructureRelease(pStruct);
  return fts5IndexReturn(p);
}


/*************************************************************************
**************************************************************************
** Below this point is the implementation of the integrity-check 
** functionality.
*/

/*
** Return a simple checksum value based on the arguments.
*/
static u64 fts5IndexEntryCksum(
  i64 iRowid, 
  int iCol, 
  int iPos, 
  int iIdx,
  const char *pTerm,
  int nTerm
){
  int i;
  u64 ret = iRowid;
  ret += (ret<<3) + iCol;
  ret += (ret<<3) + iPos;
  if( iIdx>=0 ) ret += (ret<<3) + (FTS5_MAIN_PREFIX + iIdx);
  for(i=0; i<nTerm; i++) ret += (ret<<3) + pTerm[i];
  return ret;
}

static void fts5BtreeIterInit(
  Fts5Index *p, 
  Fts5StructureSegment *pSeg, 
  Fts5BtreeIter *pIter
){
  int nByte;
  int i;
  nByte = sizeof(pIter->aLvl[0]) * (pSeg->nHeight-1);
  memset(pIter, 0, sizeof(*pIter));
  if( nByte ){
    pIter->aLvl = (Fts5BtreeIterLevel*)fts5IdxMalloc(p, nByte);
  }
  if( p->rc==SQLITE_OK ){
    pIter->nLvl = pSeg->nHeight-1;
    pIter->p = p;
    pIter->pSeg = pSeg;
  }
  for(i=0; p->rc==SQLITE_OK && i<pIter->nLvl; i++){
    i64 iRowid = FTS5_SEGMENT_ROWID(pSeg->iSegid, i+1, 1);
    Fts5Data *pData;
    pIter->aLvl[i].pData = pData = fts5DataRead(p, iRowid);
    if( pData ){
      fts5NodeIterInit(pData->p, pData->n, &pIter->aLvl[i].s);
    }
  }

  if( pIter->nLvl==0 || p->rc ){
    pIter->bEof = 1;
    pIter->iLeaf = pSeg->pgnoLast;
  }else{
    pIter->nEmpty = pIter->aLvl[0].s.nEmpty;
    pIter->iLeaf = pIter->aLvl[0].s.iChild;
    pIter->bDlidx = pIter->aLvl[0].s.bDlidx;
  }
}

static void fts5BtreeIterNext(Fts5BtreeIter *pIter){
  Fts5Index *p = pIter->p;
  int i;

  assert( pIter->bEof==0 && pIter->aLvl[0].s.aData );
  for(i=0; i<pIter->nLvl && p->rc==SQLITE_OK; i++){
    Fts5BtreeIterLevel *pLvl = &pIter->aLvl[i];
    fts5NodeIterNext(&p->rc, &pLvl->s);
    if( pLvl->s.aData ){
      fts5BufferSet(&p->rc, &pIter->term, pLvl->s.term.n, pLvl->s.term.p);
      break;
    }else{
      fts5NodeIterFree(&pLvl->s);
      fts5DataRelease(pLvl->pData);
      pLvl->pData = 0;
    }
  }
  if( i==pIter->nLvl || p->rc ){
    pIter->bEof = 1;
  }else{
    int iSegid = pIter->pSeg->iSegid;
    for(i--; i>=0; i--){
      Fts5BtreeIterLevel *pLvl = &pIter->aLvl[i];
      i64 iRowid = FTS5_SEGMENT_ROWID(iSegid, i+1, pLvl[1].s.iChild);
      pLvl->pData = fts5DataRead(p, iRowid);
      if( pLvl->pData ){
        fts5NodeIterInit(pLvl->pData->p, pLvl->pData->n, &pLvl->s);
      }
    }
  }

  pIter->nEmpty = pIter->aLvl[0].s.nEmpty;
  pIter->bDlidx = pIter->aLvl[0].s.bDlidx;
  pIter->iLeaf = pIter->aLvl[0].s.iChild;
}

static void fts5BtreeIterFree(Fts5BtreeIter *pIter){
  int i;
  for(i=0; i<pIter->nLvl; i++){
    Fts5BtreeIterLevel *pLvl = &pIter->aLvl[i];
    fts5NodeIterFree(&pLvl->s);
    if( pLvl->pData ){
      fts5DataRelease(pLvl->pData);
      pLvl->pData = 0;
    }
  }
  sqlite3_free(pIter->aLvl);
  fts5BufferFree(&pIter->term);
}

#ifdef SQLITE_DEBUG
/*
** This function is purely an internal test. It does not contribute to 
** FTS functionality, or even the integrity-check, in any way.
**
** Instead, it tests that the same set of pgno/rowid combinations are 
** visited regardless of whether the doclist-index identified by parameters
** iSegid/iLeaf is iterated in forwards or reverse order.
*/
static void fts5TestDlidxReverse(
  Fts5Index *p, 
  int iSegid,                     /* Segment id to load from */
  int iLeaf                       /* Load doclist-index for this leaf */
){
  Fts5DlidxIter *pDlidx = 0;
  u64 cksum1 = 13;
  u64 cksum2 = 13;

  for(pDlidx=fts5DlidxIterInit(p, 0, iSegid, iLeaf);
      fts5DlidxIterEof(p, pDlidx)==0;
      fts5DlidxIterNext(p, pDlidx)
  ){
    i64 iRowid = fts5DlidxIterRowid(pDlidx);
    int pgno = fts5DlidxIterPgno(pDlidx);
    assert( pgno>iLeaf );
    cksum1 += iRowid + ((i64)pgno<<32);
  }
  fts5DlidxIterFree(pDlidx);
  pDlidx = 0;

  for(pDlidx=fts5DlidxIterInit(p, 1, iSegid, iLeaf);
      fts5DlidxIterEof(p, pDlidx)==0;
      fts5DlidxIterPrev(p, pDlidx)
  ){
    i64 iRowid = fts5DlidxIterRowid(pDlidx);
    int pgno = fts5DlidxIterPgno(pDlidx);
    assert( fts5DlidxIterPgno(pDlidx)>iLeaf );
    cksum2 += iRowid + ((i64)pgno<<32);
  }
  fts5DlidxIterFree(pDlidx);
  pDlidx = 0;

  if( p->rc==SQLITE_OK && cksum1!=cksum2 ) p->rc = FTS5_CORRUPT;
}

static int fts5QueryCksum(
  Fts5Index *p,                   /* Fts5 index object */
  int iIdx,
  const char *z,                  /* Index key to query for */
  int n,                          /* Size of index key in bytes */
  int flags,                      /* Flags for Fts5IndexQuery */
  u64 *pCksum                     /* IN/OUT: Checksum value */
){
  u64 cksum = *pCksum;
  Fts5IndexIter *pIdxIter = 0;
  int rc = sqlite3Fts5IndexQuery(p, z, n, flags, &pIdxIter);

  while( rc==SQLITE_OK && 0==sqlite3Fts5IterEof(pIdxIter) ){
    i64 dummy;
    const u8 *pPos;
    int nPos;
    i64 rowid = sqlite3Fts5IterRowid(pIdxIter);
    rc = sqlite3Fts5IterPoslist(pIdxIter, &pPos, &nPos, &dummy);
    if( rc==SQLITE_OK ){
      Fts5PoslistReader sReader;
      for(sqlite3Fts5PoslistReaderInit(-1, pPos, nPos, &sReader);
          sReader.bEof==0;
          sqlite3Fts5PoslistReaderNext(&sReader)
      ){
        int iCol = FTS5_POS2COLUMN(sReader.iPos);
        int iOff = FTS5_POS2OFFSET(sReader.iPos);
        cksum ^= fts5IndexEntryCksum(rowid, iCol, iOff, iIdx, z, n);
      }
      rc = sqlite3Fts5IterNext(pIdxIter);
    }
  }
  sqlite3Fts5IterClose(pIdxIter);

  *pCksum = cksum;
  return rc;
}


/*
** This function is also purely an internal test. It does not contribute to 
** FTS functionality, or even the integrity-check, in any way.
*/
static void fts5TestTerm(
  Fts5Index *p, 
  Fts5Buffer *pPrev,              /* Previous term */
  const char *z, int n,           /* Possibly new term to test */
  u64 expected,
  u64 *pCksum
){
  int rc = p->rc;
  if( pPrev->n==0 ){
    fts5BufferSet(&rc, pPrev, n, (const u8*)z);
  }else
  if( rc==SQLITE_OK && (pPrev->n!=n || memcmp(pPrev->p, z, n)) ){
    u64 cksum3 = *pCksum;
    const char *zTerm = (const char*)&pPrev->p[1];  /* term sans prefix-byte */
    int nTerm = pPrev->n-1;            /* Size of zTerm in bytes */
    int iIdx = (pPrev->p[0] - FTS5_MAIN_PREFIX);
    int flags = (iIdx==0 ? 0 : FTS5INDEX_QUERY_PREFIX);
    int rc;
    u64 ck1 = 0;
    u64 ck2 = 0;

    /* Check that the results returned for ASC and DESC queries are
    ** the same. If not, call this corruption.  */
    rc = fts5QueryCksum(p, iIdx, zTerm, nTerm, flags, &ck1);
    if( rc==SQLITE_OK ){
      int f = flags|FTS5INDEX_QUERY_DESC;
      rc = fts5QueryCksum(p, iIdx, zTerm, nTerm, f, &ck2);
    }
    if( rc==SQLITE_OK && ck1!=ck2 ) rc = FTS5_CORRUPT;

    /* If this is a prefix query, check that the results returned if the
    ** the index is disabled are the same. In both ASC and DESC order. */
    if( iIdx>0 && rc==SQLITE_OK ){
      int f = flags|FTS5INDEX_QUERY_TEST_NOIDX;
      ck2 = 0;
      rc = fts5QueryCksum(p, iIdx, zTerm, nTerm, f, &ck2);
      if( rc==SQLITE_OK && ck1!=ck2 ) rc = FTS5_CORRUPT;
    }
    if( iIdx>0 && rc==SQLITE_OK ){
      int f = flags|FTS5INDEX_QUERY_TEST_NOIDX|FTS5INDEX_QUERY_DESC;
      ck2 = 0;
      rc = fts5QueryCksum(p, iIdx, zTerm, nTerm, f, &ck2);
      if( rc==SQLITE_OK && ck1!=ck2 ) rc = FTS5_CORRUPT;
    }

    cksum3 ^= ck1;
    fts5BufferSet(&rc, pPrev, n, (const u8*)z);

    if( rc==SQLITE_OK && cksum3!=expected ){
      rc = FTS5_CORRUPT;
    }
    *pCksum = cksum3;
  }
  p->rc = rc;
}
 
#else
# define fts5TestDlidxReverse(x,y,z)
# define fts5TestTerm(u,v,w,x,y,z)
#endif

static void fts5IndexIntegrityCheckSegment(
  Fts5Index *p,                   /* FTS5 backend object */
  Fts5StructureSegment *pSeg      /* Segment to check internal consistency */
){
  Fts5BtreeIter iter;             /* Used to iterate through b-tree hierarchy */

  if( pSeg->pgnoFirst==0 ) return;

  /* Iterate through the b-tree hierarchy.  */
  for(fts5BtreeIterInit(p, pSeg, &iter);
      p->rc==SQLITE_OK && iter.bEof==0;
      fts5BtreeIterNext(&iter)
  ){
    i64 iRow;                     /* Rowid for this leaf */
    Fts5Data *pLeaf;              /* Data for this leaf */
    int iOff;                     /* Offset of first term on leaf */
    int i;                        /* Used to iterate through empty leaves */

    /* If the leaf in question has already been trimmed from the segment, 
    ** ignore this b-tree entry. Otherwise, load it into memory. */
    if( iter.iLeaf<pSeg->pgnoFirst ) continue;
    iRow = FTS5_SEGMENT_ROWID(pSeg->iSegid, 0, iter.iLeaf);
    pLeaf = fts5DataRead(p, iRow);
    if( pLeaf==0 ) break;

    /* Check that the leaf contains at least one term, and that it is equal
    ** to or larger than the split-key in iter.term.  Also check that if there
    ** is also a rowid pointer within the leaf page header, it points to a
    ** location before the term.  */
    iOff = fts5GetU16(&pLeaf->p[2]);
    if( iOff==0 ){
      p->rc = FTS5_CORRUPT;
    }else{
      int iRowidOff;
      int nTerm;                  /* Size of term on leaf in bytes */
      int res;                    /* Comparison of term and split-key */

      iRowidOff = fts5GetU16(&pLeaf->p[0]);
      if( iRowidOff>=iOff ){
        p->rc = FTS5_CORRUPT;
      }else{
        iOff += fts5GetVarint32(&pLeaf->p[iOff], nTerm);
        res = memcmp(&pLeaf->p[iOff], iter.term.p, MIN(nTerm, iter.term.n));
        if( res==0 ) res = nTerm - iter.term.n;
        if( res<0 ) p->rc = FTS5_CORRUPT;
      }
    }
    fts5DataRelease(pLeaf);
    if( p->rc ) break;


    /* Now check that the iter.nEmpty leaves following the current leaf
    ** (a) exist and (b) contain no terms. */
    for(i=1; p->rc==SQLITE_OK && i<=iter.nEmpty; i++){
      pLeaf = fts5DataRead(p, iRow+i);
      if( pLeaf && 0!=fts5GetU16(&pLeaf->p[2]) ){
        p->rc = FTS5_CORRUPT;
      }
      fts5DataRelease(pLeaf);
    }

    /* If there is a doclist-index, check that it looks right. */
    if( iter.bDlidx ){
      Fts5DlidxIter *pDlidx = 0;  /* For iterating through doclist index */
      int iPrevLeaf = iter.iLeaf;
      int iSegid = pSeg->iSegid;
      int iPg;
      i64 iKey;

      for(pDlidx=fts5DlidxIterInit(p, 0, iSegid, iter.iLeaf);
          fts5DlidxIterEof(p, pDlidx)==0;
          fts5DlidxIterNext(p, pDlidx)
      ){

        /* Check any rowid-less pages that occur before the current leaf. */
        for(iPg=iPrevLeaf+1; iPg<fts5DlidxIterPgno(pDlidx); iPg++){
          iKey = FTS5_SEGMENT_ROWID(iSegid, 0, iPg);
          pLeaf = fts5DataRead(p, iKey);
          if( pLeaf ){
            if( fts5GetU16(&pLeaf->p[0])!=0 ) p->rc = FTS5_CORRUPT;
            fts5DataRelease(pLeaf);
          }
        }
        iPrevLeaf = fts5DlidxIterPgno(pDlidx);

        /* Check that the leaf page indicated by the iterator really does
        ** contain the rowid suggested by the same. */
        iKey = FTS5_SEGMENT_ROWID(iSegid, 0, iPrevLeaf);
        pLeaf = fts5DataRead(p, iKey);
        if( pLeaf ){
          i64 iRowid;
          int iRowidOff = fts5GetU16(&pLeaf->p[0]);
          if( iRowidOff>=pLeaf->n ){
            p->rc = FTS5_CORRUPT;
          }else{
            fts5GetVarint(&pLeaf->p[iRowidOff], (u64*)&iRowid);
            if( iRowid!=fts5DlidxIterRowid(pDlidx) ) p->rc = FTS5_CORRUPT;
          }
          fts5DataRelease(pLeaf);
        }
      }

      for(iPg=iPrevLeaf+1; iPg<=(iter.iLeaf + iter.nEmpty); iPg++){
        iKey = FTS5_SEGMENT_ROWID(iSegid, 0, iPg);
        pLeaf = fts5DataRead(p, iKey);
        if( pLeaf ){
          if( fts5GetU16(&pLeaf->p[0])!=0 ) p->rc = FTS5_CORRUPT;
          fts5DataRelease(pLeaf);
        }
      }

      fts5DlidxIterFree(pDlidx);
      fts5TestDlidxReverse(p, iSegid, iter.iLeaf);
    }
  }

  /* Page iter.iLeaf must now be the rightmost leaf-page in the segment */
  if( p->rc==SQLITE_OK && iter.iLeaf!=pSeg->pgnoLast ){
    p->rc = FTS5_CORRUPT;
  }

  fts5BtreeIterFree(&iter);
}


/*
** Run internal checks to ensure that the FTS index (a) is internally 
** consistent and (b) contains entries for which the XOR of the checksums
** as calculated by fts5IndexEntryCksum() is cksum.
**
** Return SQLITE_CORRUPT if any of the internal checks fail, or if the
** checksum does not match. Return SQLITE_OK if all checks pass without
** error, or some other SQLite error code if another error (e.g. OOM)
** occurs.
*/
int sqlite3Fts5IndexIntegrityCheck(Fts5Index *p, u64 cksum){
  u64 cksum2 = 0;                 /* Checksum based on contents of indexes */
  Fts5Buffer poslist = {0,0,0};   /* Buffer used to hold a poslist */
  Fts5MultiSegIter *pIter;        /* Used to iterate through entire index */
  Fts5Structure *pStruct;         /* Index structure */

  /* Used by extra internal tests only run if NDEBUG is not defined */
  u64 cksum3 = 0;                 /* Checksum based on contents of indexes */
  Fts5Buffer term = {0,0,0};      /* Buffer used to hold most recent term */
  
  /* Load the FTS index structure */
  pStruct = fts5StructureRead(p);

  /* Check that the internal nodes of each segment match the leaves */
  if( pStruct ){
    int iLvl, iSeg;
    for(iLvl=0; iLvl<pStruct->nLevel; iLvl++){
      for(iSeg=0; iSeg<pStruct->aLevel[iLvl].nSeg; iSeg++){
        Fts5StructureSegment *pSeg = &pStruct->aLevel[iLvl].aSeg[iSeg];
        fts5IndexIntegrityCheckSegment(p, pSeg);
      }
    }
  }

  /* The cksum argument passed to this function is a checksum calculated
  ** based on all expected entries in the FTS index (including prefix index
  ** entries). This block checks that a checksum calculated based on the
  ** actual contents of FTS index is identical.
  **
  ** Two versions of the same checksum are calculated. The first (stack
  ** variable cksum2) based on entries extracted from the full-text index
  ** while doing a linear scan of each individual index in turn. 
  **
  ** As each term visited by the linear scans, a separate query for the
  ** same term is performed. cksum3 is calculated based on the entries
  ** extracted by these queries.
  */
  for(fts5MultiIterNew(p, pStruct, 0, 0, 0, 0, -1, 0, &pIter);
      fts5MultiIterEof(p, pIter)==0;
      fts5MultiIterNext(p, pIter, 0, 0)
  ){
    int n;                      /* Size of term in bytes */
    i64 iPos = 0;               /* Position read from poslist */
    int iOff = 0;               /* Offset within poslist */
    i64 iRowid = fts5MultiIterRowid(pIter);
    char *z = (char*)fts5MultiIterTerm(pIter, &n);

    poslist.n = 0;
    fts5MultiIterPoslist(p, pIter, 0, &poslist);
    while( 0==sqlite3Fts5PoslistNext64(poslist.p, poslist.n, &iOff, &iPos) ){
      int iCol = FTS5_POS2COLUMN(iPos);
      int iTokOff = FTS5_POS2OFFSET(iPos);
      cksum2 ^= fts5IndexEntryCksum(iRowid, iCol, iTokOff, -1, z, n);
    }

    /* If this is a new term, query for it. Update cksum3 with the results. */
    fts5TestTerm(p, &term, z, n, cksum2, &cksum3);
  }
  fts5TestTerm(p, &term, 0, 0, cksum2, &cksum3);

  fts5MultiIterFree(p, pIter);
  if( p->rc==SQLITE_OK && cksum!=cksum2 ) p->rc = FTS5_CORRUPT;

  fts5StructureRelease(pStruct);
  fts5BufferFree(&term);
  fts5BufferFree(&poslist);
  return fts5IndexReturn(p);
}


/*
** Calculate and return a checksum that is the XOR of the index entry
** checksum of all entries that would be generated by the token specified
** by the final 5 arguments.
*/
u64 sqlite3Fts5IndexCksum(
  Fts5Config *pConfig,            /* Configuration object */
  i64 iRowid,                     /* Document term appears in */
  int iCol,                       /* Column term appears in */
  int iPos,                       /* Position term appears in */
  const char *pTerm, int nTerm    /* Term at iPos */
){
  u64 ret = 0;                    /* Return value */
  int iIdx;                       /* For iterating through indexes */

  ret = fts5IndexEntryCksum(iRowid, iCol, iPos, 0, pTerm, nTerm);

  for(iIdx=0; iIdx<pConfig->nPrefix; iIdx++){
    int nByte = fts5IndexCharlenToBytelen(pTerm, nTerm, pConfig->aPrefix[iIdx]);
    if( nByte ){
      ret ^= fts5IndexEntryCksum(iRowid, iCol, iPos, iIdx+1, pTerm, nByte);
    }
  }

  return ret;
}

/*************************************************************************
**************************************************************************
** Below this point is the implementation of the fts5_decode() scalar
** function only.
*/

/*
** Decode a segment-data rowid from the %_data table. This function is
** the opposite of macro FTS5_SEGMENT_ROWID().
*/
static void fts5DecodeRowid(
  i64 iRowid,                     /* Rowid from %_data table */
  int *piSegid,                   /* OUT: Segment id */
  int *pbDlidx,                   /* OUT: Dlidx flag */
  int *piHeight,                  /* OUT: Height */
  int *piPgno                     /* OUT: Page number */
){
  *piPgno = (int)(iRowid & (((i64)1 << FTS5_DATA_PAGE_B) - 1));
  iRowid >>= FTS5_DATA_PAGE_B;

  *piHeight = (int)(iRowid & (((i64)1 << FTS5_DATA_HEIGHT_B) - 1));
  iRowid >>= FTS5_DATA_HEIGHT_B;

  *pbDlidx = (int)(iRowid & 0x0001);
  iRowid >>= FTS5_DATA_DLI_B;

  *piSegid = (int)(iRowid & (((i64)1 << FTS5_DATA_ID_B) - 1));
}

static void fts5DebugRowid(int *pRc, Fts5Buffer *pBuf, i64 iKey){
  int iSegid, iHeight, iPgno, bDlidx;       /* Rowid compenents */
  fts5DecodeRowid(iKey, &iSegid, &bDlidx, &iHeight, &iPgno);

  if( iSegid==0 ){
    if( iKey==FTS5_AVERAGES_ROWID ){
      sqlite3Fts5BufferAppendPrintf(pRc, pBuf, "(averages) ");
    }else{
      sqlite3Fts5BufferAppendPrintf(pRc, pBuf, "(structure)");
    }
  }
  else{
    sqlite3Fts5BufferAppendPrintf(pRc, pBuf, "(%ssegid=%d h=%d pgno=%d)",
        bDlidx ? "dlidx " : "", iSegid, iHeight, iPgno
    );
  }
}

static void fts5DebugStructure(
  int *pRc,                       /* IN/OUT: error code */
  Fts5Buffer *pBuf,
  Fts5Structure *p
){
  int iLvl, iSeg;                 /* Iterate through levels, segments */

  for(iLvl=0; iLvl<p->nLevel; iLvl++){
    Fts5StructureLevel *pLvl = &p->aLevel[iLvl];
    sqlite3Fts5BufferAppendPrintf(pRc, pBuf, 
        " {lvl=%d nMerge=%d", iLvl, pLvl->nMerge
    );
    for(iSeg=0; iSeg<pLvl->nSeg; iSeg++){
      Fts5StructureSegment *pSeg = &pLvl->aSeg[iSeg];
      sqlite3Fts5BufferAppendPrintf(pRc, pBuf, 
          " {id=%d h=%d leaves=%d..%d}", pSeg->iSegid, pSeg->nHeight, 
          pSeg->pgnoFirst, pSeg->pgnoLast
      );
    }
    sqlite3Fts5BufferAppendPrintf(pRc, pBuf, "}");
  }
}

/*
** This is part of the fts5_decode() debugging aid.
**
** Arguments pBlob/nBlob contain a serialized Fts5Structure object. This
** function appends a human-readable representation of the same object
** to the buffer passed as the second argument. 
*/
static void fts5DecodeStructure(
  int *pRc,                       /* IN/OUT: error code */
  Fts5Buffer *pBuf,
  const u8 *pBlob, int nBlob
){
  int rc;                         /* Return code */
  Fts5Structure *p = 0;           /* Decoded structure object */

  rc = fts5StructureDecode(pBlob, nBlob, 0, &p);
  if( rc!=SQLITE_OK ){
    *pRc = rc;
    return;
  }

  fts5DebugStructure(pRc, pBuf, p);
  fts5StructureRelease(p);
}

/*
** Buffer (a/n) is assumed to contain a list of serialized varints. Read
** each varint and append its string representation to buffer pBuf. Return
** after either the input buffer is exhausted or a 0 value is read.
**
** The return value is the number of bytes read from the input buffer.
*/
static int fts5DecodePoslist(int *pRc, Fts5Buffer *pBuf, const u8 *a, int n){
  int iOff = 0;
  while( iOff<n ){
    int iVal;
    iOff += fts5GetVarint32(&a[iOff], iVal);
    sqlite3Fts5BufferAppendPrintf(pRc, pBuf, " %d", iVal);
  }
  return iOff;
}

/*
** The start of buffer (a/n) contains the start of a doclist. The doclist
** may or may not finish within the buffer. This function appends a text
** representation of the part of the doclist that is present to buffer
** pBuf. 
**
** The return value is the number of bytes read from the input buffer.
*/
static int fts5DecodeDoclist(int *pRc, Fts5Buffer *pBuf, const u8 *a, int n){
  i64 iDocid;
  int iOff = 0;

  iOff = sqlite3Fts5GetVarint(&a[iOff], (u64*)&iDocid);
  sqlite3Fts5BufferAppendPrintf(pRc, pBuf, " rowid=%lld", iDocid);
  while( iOff<n ){
    int nPos;
    int bDummy;
    iOff += fts5GetPoslistSize(&a[iOff], &nPos, &bDummy);
    iOff += fts5DecodePoslist(pRc, pBuf, &a[iOff], MIN(n-iOff, nPos));
    if( iOff<n ){
      i64 iDelta;
      iOff += sqlite3Fts5GetVarint(&a[iOff], (u64*)&iDelta);
      if( iDelta==0 ) return iOff;
      iDocid += iDelta;
      sqlite3Fts5BufferAppendPrintf(pRc, pBuf, " rowid=%lld", iDocid);
    }
  }

  return iOff;
}

/*
** The implementation of user-defined scalar function fts5_decode().
*/
static void fts5DecodeFunction(
  sqlite3_context *pCtx,          /* Function call context */
  int nArg,                       /* Number of args (always 2) */
  sqlite3_value **apVal           /* Function arguments */
){
  i64 iRowid;                     /* Rowid for record being decoded */
  int iSegid,iHeight,iPgno,bDlidx;/* Rowid components */
  const u8 *aBlob; int n;         /* Record to decode */
  u8 *a = 0;
  Fts5Buffer s;                   /* Build up text to return here */
  int rc = SQLITE_OK;             /* Return code */
  int nSpace = 0;

  assert( nArg==2 );
  memset(&s, 0, sizeof(Fts5Buffer));
  iRowid = sqlite3_value_int64(apVal[0]);
  n = sqlite3_value_bytes(apVal[1]);
  aBlob = sqlite3_value_blob(apVal[1]);

  nSpace = n + FTS5_DATA_ZERO_PADDING;
  a = (u8*)sqlite3Fts5MallocZero(&rc, nSpace);
  if( a==0 ) goto decode_out;
  memcpy(a, aBlob, n);
  fts5DecodeRowid(iRowid, &iSegid, &bDlidx, &iHeight, &iPgno);

  fts5DebugRowid(&rc, &s, iRowid);
  if( bDlidx ){
    Fts5Data dlidx;
    Fts5DlidxLvl lvl;

    dlidx.p = a;
    dlidx.n = n;

    memset(&lvl, 0, sizeof(Fts5DlidxLvl));
    lvl.pData = &dlidx;
    lvl.iLeafPgno = iPgno;

    for(fts5DlidxLvlNext(&lvl); lvl.bEof==0; fts5DlidxLvlNext(&lvl)){
      sqlite3Fts5BufferAppendPrintf(&rc, &s, 
          " %d(%lld)", lvl.iLeafPgno, lvl.iRowid
      );
    }
  }else if( iSegid==0 ){
    if( iRowid==FTS5_AVERAGES_ROWID ){
      /* todo */
    }else{
      fts5DecodeStructure(&rc, &s, a, n);
    }
  }else{

    Fts5Buffer term;
    memset(&term, 0, sizeof(Fts5Buffer));

    if( iHeight==0 ){
      int iTermOff = 0;
      int iRowidOff = 0;
      int iOff;
      int nKeep = 0;

      if( n>=4 ){
        iRowidOff = fts5GetU16(&a[0]);
        iTermOff = fts5GetU16(&a[2]);
      }else{
        sqlite3Fts5BufferSet(&rc, &s, 8, (const u8*)"corrupt");
        goto decode_out;
      }

      if( iRowidOff ){
        iOff = iRowidOff;
      }else if( iTermOff ){
        iOff = iTermOff;
      }else{
        iOff = n;
      }
      fts5DecodePoslist(&rc, &s, &a[4], iOff-4);

      assert( iRowidOff==0 || iOff==iRowidOff );
      if( iRowidOff ){
        iOff += fts5DecodeDoclist(&rc, &s, &a[iOff], n-iOff);
      }

      assert( iTermOff==0 || iOff==iTermOff );
      while( iOff<n ){
        int nByte;
        iOff += fts5GetVarint32(&a[iOff], nByte);
        term.n= nKeep;
        fts5BufferAppendBlob(&rc, &term, nByte, &a[iOff]);
        iOff += nByte;

        sqlite3Fts5BufferAppendPrintf(
            &rc, &s, " term=%.*s", term.n, (const char*)term.p
        );
        iOff += fts5DecodeDoclist(&rc, &s, &a[iOff], n-iOff);
        if( iOff<n ){
          iOff += fts5GetVarint32(&a[iOff], nKeep);
        }
      }
      fts5BufferFree(&term);
    }else{
      Fts5NodeIter ss;
      for(fts5NodeIterInit(a, n, &ss); ss.aData; fts5NodeIterNext(&rc, &ss)){
        if( ss.term.n==0 ){
          sqlite3Fts5BufferAppendPrintf(&rc, &s, " left=%d", ss.iChild);
        }else{
          sqlite3Fts5BufferAppendPrintf(&rc,&s, " \"%.*s\"", 
              ss.term.n, ss.term.p
          );
        }
        if( ss.nEmpty ){
          sqlite3Fts5BufferAppendPrintf(&rc, &s, " empty=%d%s", ss.nEmpty,
              ss.bDlidx ? "*" : ""
          );
        }
      }
      fts5NodeIterFree(&ss);
    }
  }
  
 decode_out:
  sqlite3_free(a);
  if( rc==SQLITE_OK ){
    sqlite3_result_text(pCtx, (const char*)s.p, s.n, SQLITE_TRANSIENT);
  }else{
    sqlite3_result_error_code(pCtx, rc);
  }
  fts5BufferFree(&s);
}

/*
** The implementation of user-defined scalar function fts5_rowid().
*/
static void fts5RowidFunction(
  sqlite3_context *pCtx,          /* Function call context */
  int nArg,                       /* Number of args (always 2) */
  sqlite3_value **apVal           /* Function arguments */
){
  const char *zArg;
  if( nArg==0 ){
    sqlite3_result_error(pCtx, "should be: fts5_rowid(subject, ....)", -1);
  }else{
    zArg = (const char*)sqlite3_value_text(apVal[0]);
    if( 0==sqlite3_stricmp(zArg, "segment") ){
      i64 iRowid;
      int segid, height, pgno;
      if( nArg!=4 ){
        sqlite3_result_error(pCtx, 
            "should be: fts5_rowid('segment', segid, height, pgno))", -1
        );
      }else{
        segid = sqlite3_value_int(apVal[1]);
        height = sqlite3_value_int(apVal[2]);
        pgno = sqlite3_value_int(apVal[3]);
        iRowid = FTS5_SEGMENT_ROWID(segid, height, pgno);
        sqlite3_result_int64(pCtx, iRowid);
      }
    }else {
      sqlite3_result_error(pCtx, 
        "first arg to fts5_rowid() must be 'segment' "
        "or 'start-of-index'"
        , -1
      );
    }
  }
}

/*
** This is called as part of registering the FTS5 module with database
** connection db. It registers several user-defined scalar functions useful
** with FTS5.
**
** If successful, SQLITE_OK is returned. If an error occurs, some other
** SQLite error code is returned instead.
*/
int sqlite3Fts5IndexInit(sqlite3 *db){
  int rc = sqlite3_create_function(
      db, "fts5_decode", 2, SQLITE_UTF8, 0, fts5DecodeFunction, 0, 0
  );
  if( rc==SQLITE_OK ){
    rc = sqlite3_create_function(
        db, "fts5_rowid", -1, SQLITE_UTF8, 0, fts5RowidFunction, 0, 0
    );
  }
  return rc;
}

#endif /* SQLITE_ENABLE_FTS5 */