path: root/txr.1

.\"Copyright (C) 2011, Kaz Kylheku <kaz@kylheku.com>.
.\"All rights reserved.
.\"
.\"BSD License:
.\"
.\"Redistribution and use in source and binary forms, with or without
1\"modification, are permitted provided that the following conditions
.\"are met:
.\"
.\"  1. Redistributions of source code must retain the above copyright
.\"     notice, this list of conditions and the following disclaimer.
.\"  2. Redistributions in binary form must reproduce the above copyright
.\"     notice, this list of conditions and the following disclaimer in
.\"     the documentation and/or other materials provided with the
.\"     distribution.
.\"  3. The name of the author may not be used to endorse or promote
.\"     products derived from this software without specific prior
.\"     written permission.
.\"
.\"THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR
.\"IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED
.\"WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.

.TH "txr" 1 2011-10-30 "Utility Commands" "Txr Text Extractor" "Kaz Kylheku"
.SH NAME
txr \- text extractor (version 041)
.SH SYNOPSIS
.B txr [ options ] query-file { data-file }*
.sp
.SH DESCRIPTION
.B txr
is a query tool for extracting pieces of text buried in one or more text
file based on pattern matching.  A
.B txr
query specifies a pattern which matches (a prefix of) an entire file, or
multiple files. The pattern is matched against the material in the files, and
free variables occurring in the pattern are bound to the pieces of text
occurring in the corresponding positions. If the overall match is
successful, then
.B txr
can do one of two things: it can report the list of variables which were bound,
in the form of a set of variable assignments which can be evaluated by the
.B eval
command of the POSIX shell language, or generate a custom report according
to special directives in the query.

In addition to embedded variables which implicitly match text, the
.B txr
query language supports a number of directives, for matching text using regular
expressions, for continuing a match in another file, for searching through a
file for the place where an entire sub-query matches, for collecting lists, and
for combining sub-queries using logical conjunction, disjunction and negation.

When
.B txr
finds a match for a variable and binds it, if that variable occurs again
later in the query, the variable's text is substituted, forcing a match for
that exact text. Thus txr supports a rudimentary form of backreferencing
unification, if you will. For example, the query

  @FOO=@FOO

will match material from the start of the line until the first equal sign,
and bind it to the variable
.IR FOO.
Then, the material which follows the equal sign to the end of the line must
match the contents bound to FOO. Hence the line "abc=abc" will match, but
"abc=xyz" will fail to match.

Generally, the scope of a variable's binding
extends from its first successful match where the binding is established, to
the end of the query. Unsuccessful subqueries have no effect on the
bindings.  Even if a failed subquery is partially successful, all of its
bindings are thrown away.  Some directives treat the bindings emanating
from their subqueries in special ways.

.SH ARGUMENTS AND OPTIONS

Options which don't take an argument may be combined together.
The -v and -q options are mutually exclusive. Of these two, the one which
occurs in the rightmost position in the argument list dominates.
The -c and -f options are also mutually exclusive; if both are specified,
it is a fatal error.

.IP -Dvar=value
Bind the variable
.IR var
to the value
.IR value
prior to processing the query. The name is in scope over the entire
query, so that all occurrence of the variable are substituted and
match the equivalent text.  If the value contains commas, these
are interpreted as separators, which give rise to a list value.
For instance -Da,b,c creates a list of the strings "a", "b" and "c".
(See Collect Directive bellow). List variables provide a multiple
match. That is to say, if a list variable occurs in a query, a successful
match occurs if any of its values matches the text. If more than one
value matches the text, the first one is taken.

.IP -Dvar
Binds the variable
.IR var
to an empty string value prior to processing the query.

.IP -q
Quiet operation during matching. Certain error messages are not reported on the
standard error device (but the if the situations occur, they still fail the
query). This option does not suppress error generation during the parsing
of the query, only during its execution.

.IP -v
Verbose operation. Detailed logging is enabled.

.IP -b
Suppresses the printing of variable bindings for a successful query, and the
word .
IR false
for a failed query. The program still sets an appropriate
termination status.

.IP -l or --lisp-bindings
Print the variable bindings in Lisp syntax instead of shell syntax.

.IP -a num
Specifies the maximum number of array dimensions to use for variables
arising out of collect. The default is 1. Additional dimensions are
expressed using numeric suffixes in the generated variable names.
For instance, consider the three-dimensional list arising out of a triply
nested collect: ((("a" "b") ("c" "d")) (("e" "f") ("g" "h"))).
Suppose this is bound to a variable V.  With -a 1, this will be
reported as:

  V_0_0[0]="a"
  V_0_1[0]="b"
  V_1_0[0]="c"
  V_1_1[0]="d"
  V_0_0[1]="e"
  V_0_1[1]="f"
  V_1_0[1]="g"
  V_1_1[1]="h"

The leftmost bracketed index is the most major index. That is to say,
the dimension order is: NAME_m_m+1_..._n[1][2]...[m-1].

.IP -c query
Specifies the query in the form of a command line argument. If this option is
used, the query-file argument is omitted. The first non-option argument,
if there is one, now specifies the first input source rather than a query.
Unlike queries read from a file, (non-empty) queries specified as arguments
using -c do not have to properly end in a newline. Internally, txr
adds the missing newline before parsing the query. Thus -c "@a"
is a valid query which matches a line.

Example:

  # read two lines "1" and "2" from standard input,
  # binding them to variables a and b. Standard
  # input is specified as - and the data
  # comes from shell "here document" redirection.

  txr -c "@a
  @b" - <<!
  1
  2
  !

  Output:
  a=1
  b=2

The @# comment syntax can be used for better formatting:

  txr -c "@#
  @a
  @b"

.IP -f query-file
Specifies the file from which the query is to be read, instead of the
query-file argument. This is useful in #! scripts. (See Hash Bang Support
below).

.IP --help
Prints usage summary on standard output, and terminates successfully.

.IP --version
Prints program version standard output, and terminates successfully.

.IP --
Signifies the end of the option list. This option does not combine with others, so for instance -b- does not mean -b --, but is an error.

.IP -
This argument is not interpreted as an option, but treated as a filename
argument. After the first such argument, no more options are recognized. Even
if another argument looks like an option, it is treated as a name.
This special argument - means "read from standard input" instead of a file.
The query file, or any of the data files, may be specified using this option.
If two or more files are specified as -, the behavior is system-dependent.
It may be possible to indicate EOF from the interactive terminal, and
then specify more input which is interpreted as the second file, and so forth.

.PP
After the options, the remaining arguments are files. The first file argument
specifies the query, and is mandatory.  A file argument consisting of a single
- means to read the standard input instead of opening a file. A file argument
which begins with an exclamation symbol means that the rest of the argument is
a shell command which is to be run as a coprocess, and its output read like a
file.

.PP
.B txr
begins by reading the query. The entire query is scanned, internalized
and then begins executing, if it is free of syntax errors.  The reading of
data, on the other hand, is lazy.  A file isn't opened until the query demands
material from that file, and then the contents are read on demand, not all at
once.

If no files arguments are specified on the command line, it is up to the
query to open a file, pipe or standard input via the @(next) directive
prior to attempting to make a match. If a query attempts to match text,
but has run out of files to process, the match fails.

.SH STATUS AND ERROR REPORTING
.B txr
sends errors and verbose logs to the standard error device.  The following paragraphs apply when
.B txr
is run without enabling verbose mode. If verbose mode is enabled, then
.B txr
issues diagnostics on the standard error device even in situations which are
not erroneous.

If the command line arguments are incorrect, or the query has a malformed
syntax, or fails to match,
.B txr
issues an error diagnostic and terminates with a failed status.

If the query is accepted, but fails to execute, either due to a
semantic error or due to a mismatch against the data,
.B txr
terminates with a failed status, it also prints the word
.IR false
on standard output. (See NOTES ON FALSE below).  Printing of false
is suppressed if the query executed one or more @(output) directive
directed to standard output.

If the query is well-formed, and matches, then
.B txr
issues no diagnostics on standard error (except in the case of verbose
reporting enabled by -v).  If no variables were bound in the query, then
nothing is printed on standard output.  If the query has matched one or more
variables, then these variables are printed on standard output, in the form of
a shell script which, when evaluated, will cause shell variables to be
assigned.  Printing of these variables is suppressed if the query executed one
or more @(output) directive directed to standard output.

.SH BASIC QUERY SYNTAX AND SEMANTICS

.SS Comments

A query may contain comments which are delimited by the sequence @# and
extend to the end of the line. No whitespace can occur between the @ and #.
A comment which begins on a line swallows that entire line, as well as the
newline which terminates it. In essence, the entire comment disappears.
If the comment follows some material in a line, then it does not consume
the newline. Thus, the following two queries are equivalent:

 1.  @a@# comment: match whole line against variable @a
     @# this comment disappears entirely
     @b

 2.  @a
     @b

The comment after the @a does not consume the newline, but the
comment which follows does. Without this intuitive behavior,
line comment would give rise to empty lines that must match empty
lines in the data, leading to spurious mismatches.

.SS Hash Bang Support

If the first line of a query begins with the characters #!,
that entire line is deleted from the query. This allows
for txr queries to be turned into standalone executable programs in the POSIX
environment.

Shell example: create a simple executable program called "twoline.txr" and
run it. This assumes txr is installed in /usr/bin.

  $ cat > twoline.txr
  #!/usr/bin/txr
  @a
  @b
  [Ctrl-D]
  $ chmod a+x twoline.txr
  $ ./twoline.txr -
  1
  2
  [Ctrl-D]
  a=1
  b=2

A script written in this manner will not pass options to txr.  For
instance, if the above script is invoked like this

  ./twoline.txr -Da=42

the -D option isn't passed down to txr; -Da=42 is an ordinary
argument (which the script will try to open as an input file).
This behavior is useful if the script author wants not to
expose the txr options to the user of the script.

However, if the hash bang line can use the -f option:

  #!/usr/bin/txr -f

Now, the name of the script is passed as an argument to the -f option,
and txr will look for more options after that.

.SS Whitespace

Outside of directives, whitespace is significant in TXR queries, and represents
a pattern match for whitespace in the input.  An extent of text consisting of
an undivided mixture of tabs and spaces is a whitespace token.  

Whitespace tokens match a precisely identical piece of whitespace in the input,
with one exception: a whitespace token consisting of precisely one space has a
special meaning. It is equivalent to the regular expression @/[ ]+/: match
an extent of one or more spaces (but not tabs!)

Thus, the query line "a b" (one space) matches texts like "a b", "a   b", et
cetera (arbitrary number of tabs and spaces between a and b).  However "a  b"
(two spaces) matches only "a  b" (two spaces).

For matching a single space, the syntax @\ can be used (backslash-escaped
space).

It is more often necessary to match multiple spaces, than to exactly
match one space, so this rule simplifies many queries and adds inconvenience
to only few.

In output clauses, string and character literals and quasiliterals, a space
token denotes a space.

.SS Text

Query material which is not escaped by the special character @ is
literal text, which matches input character for character. Text which occurs at
the beginning of a line matches the beginning of a line.  Text which starts in
the middle of a line, other than following a variable, must match exactly at
the current position, where the previous match left off. Moreover, if the text
is the last element in the line, its match is anchored to the end of the line.

An empty query line matches an empty line in the input. Note that an
empty input stream does not contain any lines, and therefore is not matched
by an empty line. An empty line in the input is represented by a newline
character which is either the first character of the file, or follows
a previous newline-terminated line.  

Input streams which end without terminating their last line with a newline are
tolerated, and are treated as if they had the terminator.

Text which follows a variable has special semantics, discussed in the
section Variables below.

A query may not leave unmatched material in a line which is covered by the
query.  However, a query may leave unmatched lines.

In the following example, the query matches the text, even though
the text has an extra line.

 Query:         Four score and seven
                years ago our

 Text:          Four score and seven
                years ago our
                forefathers

In the following example, the query
.B fails
to match the text, because the text has extra material on one
line.

 Query:         I can carry nearly eighty gigs
                in my head

 Text:          I can carry nearly eighty gigs of data
                in my head

Needless to say, if the text has insufficient material relative
to the query, that is a failure also.

To match arbitrary material from the current position to the end
of a line, the "match any sequence of characters, including empty"
regular expression @/.*/ can be used. Example:

 Query:         I can carry nearly eighty gigs@/.*/

 Text:          I can carry nearly eighty gigs of data

In this example, the query matches, since the regular expression
matches the string "of data". (See Regular Expressions section below).

Another way to do this is:

 Query:         I can carry nearly eighty gigs@(skip)


.SS Special Characters in Text

Control characters may be embedded directly in a query (with the exception of
newline characters). An alternative to embedding is to use escape syntax.
The following escapes are supported:

.IP @\e<newline>
A backslash immediately followed by a newline introduces a physical line
break without breaking up the logical line. Material following this sequence
continues to be interpreted as a continuation of the previous line, so
that indentation can be introduced to show the continuation without appearing
in the data.
.IP @\e<space>
A backslash followed by a space encodes a space. This is useful in line
continuations when it is necessary for leading spaces to be preserved.
For instance the two line sequence

   abcd@\
     @\  efg

is equivalent to the line

  abcd  efg

The two spaces before the @\ in the second line are consumed. The
spaces after are preserved.

.IP @\ea
Alert character (ASCII 7, BEL).
.IP @\eb
Backspace (ASCII 8, BS).
.IP @\et
Horizontal tab (ASCII 9, HT).
.IP @\en
Line feed (ASCII 10, LF). Serves as abstract newline on POSIX systems.
.IP @\ev
Vertical tab (ASCII 11, VT).
.IP @\ef
Form feed (ASCII 12, FF). This character clears the screen on many
kinds of terminals, or ejects a page of text from a line printer.
.IP @\er
Carriage return (ASCII 13, CR).
.IP @\ee
Escape (ASCII 27, ESC)
.IP @\exHEX
A @\ex followed by a sequence of hex digits is interpreted as a hexadecimal
numeric character code. For instance @\ex41 is the ASCII character A.
.IP @\eOCTAL
A @\e followed by a sequence of octal digits (0 through 7) is interpreted
as an octal character code. For instance @\e010 is character 8, same as @\eb.
.PP

Note that if a newline is embedded into a query line with @\en, this
does not split the line into two; it's embedded into the line and
thus cannot match anything. However, @\en may be useful in the @(cat)
directive and in @(output).

.SS International Characters

.B txr
represents text internally using wide characters, which are used to represent
Unicode code points. The query language, as well as all data sources, are
assumed to be in the UTF-8 encoding.  In the query language, extended
characters can be used directly in comments, literal text, string literals,
quasiliterals and regular expressions.  Extended characters can also be
expressed indirectly using hexadecimal or octal escapes.
On some platforms, wide characters may be restricted to 16 bits, so that
.B txr
can only work with characters in the BMP (Basic Multilingual Plane)
subset of Unicode.

.B txr
does not use the localization features of the system library;
its handling of extended characters is not affected by environment variables
like LANG and L_CTYPE. The program reads and writes only the UTF-8 encoding.

If
.B txr
encounters an invalid bytes in the UTF-8 input, what happens depends on the
context in which this occurs. In a query, comments are read without regard
for encoding, so invalid encoding bytes are not detected. A comment is
simply a sequence of bytes terminated by a newline.  Invalid
encoding bytes in significant query text are diagnosed as syntax errors.
When the scanner is faced with input that isn't a valid multibyte character, it
issues an error message, skips one byte, and resumes scanning.

Invalid bytes in data are treated as follows: when an invalid byte is
encountered in the middle of a multibyte character, or if the input
ends in the middle of a multibyte character, the UTF-8 decoder returns
to the starting byte of the ill-formed multibyte character, and decodes just
that byte, by mapping it to the Unicode character range U+DC00 through U+DCFF.
The decoding resumes at the following character, expecting that byte to be the
start of another multibyte character.

.SS Regular Expression Directives

In place of a piece of text (see section Text above), a regular expression
directive may be used, which has the following syntax:

  @/RE/

where the RE part enclosed in slashes represents regular expression
syntax (described in the section Regular Expressions below).

Long regular expressions can be broken into multiple lines using a
backslash-newline sequence.  Whitespace before the sequence or after the
sequence is not significant, so the following two are equivalent:

  @/reg \e
    ular/

  @/regular/

There may not be whitespace between the backslash and newline.

Whereas literal text simply represents itself, regular expression denotes a
(potentially infinite) set of texts.  The regular expression directive
matches the longest piece of text (possibly empty) which belongs to the set
denoted by the regular expression. The match is anchored to the current
position; thus if the directive is the first element of a line, the match is
anchored to the start of a line. If the regular expression directive is the
last element of a line, it is anchored to the end of the line also: the regular
expression must match the text from the current position to the end of the
line.

Even if the regular expression matches the empty string, the match will fail if
the input is empty, or has run out of data. For instance suppose the third line
of the query is the regular expression @/.*/, but the input is a file which has
only two lines. This will fail: the data has line for the regular expression to
match. A line containing no characters is not the same thing as the absence of
a line, even though both abstractions imply an absence of characters.

Like text which follows a variable, a regular expression directive which
follows a variable has special semantics, discussed in the section Variables
below.

.SS Variables

Much of the query syntax consists of arbitrary text, which matches file data
character for character. Embedded within the query may be variables and
directives which are introduced by a @ character.  Two consecutive @@
characters encode a literal @.

A variable matching or substitution directive is written in one of several
ways:

  @NAME
  @{NAME}
  @*NAME
  @*{NAME}
  @{NAME /RE/}
  @{NAME NUMBER}

The forms with an * indicate a long match, see Longest Match below.
The last two forms with the embedded regexp /RE/ or number have special
semantics, see Positive Match below.

The name itself may consist of any combination of one or more letters, numbers,
and underscores, and must begin with a letter or underscore.  Case is
sensitive, so that @FOO is different from @foo, which is different from @Foo.
The braces around a name can be used when material which follows would
otherwise be interpreted as being part of the name. For instance @FOO_bar
introduces the name "FOO_bar", whereas @{FOO}_bar means the variable named
"FOO" followed by the text "_bar".   There may be whitespace between the @ and
the name, or opening brace. Whitespace is also allowed in the interior of the
braces. It is not significant.

If a variable has no prior binding, then it specifies a match. The
match is determined from some current position in the data: the
character which immediately follows all that has been matched previously.
If a variable occurs at the start of a line, it matches some text
at the start of the line. If it occurs at the end of a line, it matches
everything from the current position to the end of the line.

The extent of the matched text (the text bound to the variable) is determined
by looking at what follows the variable.  A variable may be followed by a piece
of text, a regular expression directive, another variable, or nothing (i.e.
occurs at the end of a line).

If the variable is followed by nothing, the
match extends from the current position in the data, to the end of the line.
Example:

  pattern:      "a b c @FOO"
  data:         "a b c defghijk"
  result:       FOO="defghijk"

If the variable is followed by text (all non-directive material extending to
the end of the line, or to the start of another directive), then the extent of
the match is determined by searching for the first occurrence of that text
within the line, starting at the current position. The variable matches
everything between the current position and the matching position (not
including the matching position). Any whitespace which follows the
variable (and is not enclosed inside braces that surround the variable
name) is part of the text. For example:

  pattern:      "a b @FOO e f"
  data:         "a b c d e f"
  result:       FOO="c d"

In the above example, the pattern text "a b " matches the
data "a b ". So when the @FOO variable is processed, the data being
matched is the remaining "c d e f". The text which follows @FOO
is " e f". This is found within the data "c d e f" at position 3
(counting from 0).  So positions 0-2 ("c d") constitute the matching
text which is bound to FOO.

If the variable is followed by a regular expression directive,
the extent is determined by finding the closest match for the
regular expression. (See Regular Expressions section below).

To match successfully, 

.SS Special Symbols

Just like in the programming language Lisp, the names nil and t cannot be used
as variables. They always represent themselves, and have many uses, internal to
the program as well as externally visible. The nil symbol stands for the empty
list object, an object which marks the end of a list, and boolean false. It is
synonymous with the syntax () which may be used interchangeably with nil in
most constructs.

Names whose names begin with the : character are keyword symbols.  These also
may not be used as variables either and stand for themselves. Keywords are
useful for labeling information and situations. 

.SS Consecutive Variables

If an unbound variable specified a fixed-width match or a regular expression,
then the issue of consecutive variables does not arise. Such a variable
consumes text regardless of any context which follows it.

However, what if an unbound variable with no modifier is followed by another
variable? The behavior depends on the nature of the other variable.

If the other variable also has no modifier, this is a semantic error which
will cause the query to fail.  A diagnostic message will be issued, unless
operating in quiet mode via -q.  The reason is that there is no way to bind two
consecutive variables to an extent of text; this is an ambiguous situation,
since there is no matching criterion for dividing the text between two
variables.  (In theory, a repetition of the same variable, like @FOO@FOO, could
find a solution by dividing the match extent in half, which would work only in
the case when it contains an even number of characters.  This behavior seems to
have dubious value).

An unbound variable may be followed by one which is bound. The bound
variable is replaced by the text which it denotes, and the logic proceeds
accordingly.  Variables are never bound to regular expressions, so
the regular expression match does not arise in this case.
The @* syntax for longest match is available. Example:

  pattern:      "@FOO:@BAR@FOO"
  data:         "xyz:defxyz"
  result:       FOO=xyz, BAR=def

Here, FOO is matched with "xyz", based on the delimiting around the
colon. The colon in the pattern then matches the colon in the data,
so that BAR is considered for matching against "defxyz".
BAR is followed by FOO, which is already bound to "xyz".
Thus "xyz" is located in the "defxyz" data following "def",
and so BAR is bound to "def".

If an unbound variable is followed by a variable which is bound to a list, or
nested list, then each character string in the list is tried in turn to produce
a match. The first match is taken.

An unbound variable may be followed by another unbound variable which specifies
a regular expression match. This is a special case called a "double variable
match".  What happens is that the text is searched using the regular
expression.  If the search fails, than neither variable is bound: it is a
matching failure.  If the search succeeds, than the first variable is bound to
the text which is skipped by the regular expression search.  The second
variable is bound to the text matched by the regular expression.

.SS Longest Match

The closest-match behavior for text and regular expressions can be
overridden to longest match behavior. A special syntax is provided
for this: an asterisk between the @ and the variable, e.g:

  pattern:      "a @*{FOO}cd"
  data:         "a b cdcdcdcd"
  result:       FOO="b cdcdcd"

  pattern:      "a @{FOO}cd"
  data:         "a b cdcdcd"
  result:       FOO="b "

In the former example, the match extends to the rightmost occurrence of "cd",
and so FOO receives "b cdcdcd".  In the latter example, the *
syntax isn't used, and so a leftmost match takes place. The extent
covers only the "b ", stopping at the first "cd" occurrence.

.SS Positive Match

The syntax variants

 @{NAME /RE/}
 @{NAME NUMBER}

specify a variable binding that is driven by a positive match derived
from a regular expression or character count, rather than from trailing
material (which may be regarded as a "negative" match, since the variable is
bound to material which is
.B skipped
in order to match the trailing material). In the /RE/ form, the match
extends over all characters from the current position which match
the regular expression RE. (see Regular Expressions section below).

In the NUMBER form, the match processes a field of text which
consists of the specified number of characters, which must be nonnegative
number.  If the data line doesn't have that many characters starting at the
current position, the match fails. A match for zero characters produces an
empty string.  The text which is actually matched by this construct
is all text within the specified field, but excluding leading and
trailing whitespace. If the field contains only spaces, then an empty
string is extracted.

A number is made up of digits, optionally preceded by a + or - sign.

This syntax is processed without consideration of what other
syntax follows.  A positive match may be directly followed by an unbound
variable.

.SS Regular Expressions

Regular expressions are a language for specifying sets of character strings.
Through the use of pattern matching elements, regular expression is
able to denote an infinite set of texts.
.B txr
contains an original implementation of regular expressions, which
supports the following syntax:
.IP .
(period) is a "wildcard" that matches any character.
.IP []
Character class: matches a single character, from the set specified by
special syntax written between the square brackets.
Supports basic regexp character class syntax; no POSIX
notation like [:digit:]. The class [a-zA-Z] means match an uppercase
or lowercase letter; the class [0-9a-f] means match a digit or
a lowercase letter; the class [^0-9] means match a non-digit, et cetera.
A ] or - can be used within a character class, but must be escaped
with a backslash. A ^ in the first position denotes a complemented
class, unless it is escaped by backslash. In any other position, it denotes
itself.  Two backslashes code for one backslash. So for instance
[\e[\e-] means match a [ or - character, [^^] means match any character other
than ^, and [\e^\e\e] means match either a ^ or a backslash. Regex operators
such as *, + and & appearing in a character class represent ordinary
characters. The characters -, ] and ^ occurring outside of a character class
are ordinary. Unescaped / characters can appear within a character class. The
empty character class [] matches no character at all, and its complement [^]
matches any character, and is treated as a synonym for the . (period) wildcard
operator.
.IP empty
An empty expression is a regular expression. It represents the set of strings
consisting of the empty string; i.e. it matches just the empty string. The
empty regex can appear alone as a full regular expression (for instance the
.B txr
syntax @// with nothing between the slashes)
and can also be passed as a subexpression to operators, though this
may require the use of parentheses to make the empty regex explicit.  For
example, the expression a| means: match either a, or nothing.  The forms
* and (*)  are syntax errors; though not useful, the correct way to match the
empty expression zero or more times is the syntax ()*.
.IP nomatch
The nomatch regular expression represents the
empty set: it matches no strings at all, not even the empty string.
There is no dedicated syntax to directly express nomatch in the regex language.
However, the empty character class [] is equivalent to nomatch, and may be
considered to be a notation for it. Other representations of nomatch are
possible: for instance, the regex ~.* which is the complement of the regex that
denotes the set of all possible strings, and thus denotes the empty set. A
nomatch has uses; for instance, it can be used to temporarily "comment out"
regular expressions. The regex ([]abc|xyz) is equivalent to (xyz), since the
[]abc branch cannot match anything. Using [] to "block" a subexpression allows
you to leave it in place, then enable it later by removing the "block".
.IP (R)
If R is a regular expression, then so is (R).
The contents of parentheses denote one regular expression unit, so that for
instance in (RE)*, the * operator applies to the entire parenthesized group.
The syntax () is valid and equivalent to the empty regular expression.
.IP R?
optionally match the preceding regular expression R.
.IP R*
match the expression R zero or more times. This
operator is sometimes called the "Kleene star", or "Kleene closure".
The Kleene closure favors the longest match. Roughly speaking, if there are two
or more ways in which R1*R2 can match, than that match occurs in which
R1* matches the longest possible text.
.IP R+
match the preceding expression R one or more times.
Like R*, this favors the longest possible match: R+ is equivalent to RR*.
.IP R1%R2
match R1 zero or more times, then match R2. If this match can occur in
more than one way, then it occurs such that R1 is matched the fewest
number of times, which is opposite from the behavior of R1*R2.
Repetitions of R1 terminate at the earliest
point in the text where a non-empty match for R2 occurs. Because
it favors shorter matches, % is termed a non-greedy operator.  If R2 is the
empty expression, or equivalent to it, then R1%R2 reduces to R1*.  So for
instance (R%) is equivalent to (R*), since the missing right operand is
interpreted as the empty regex. Note that whereas the expression
(R1*R2) is equivalent to (R1*)R2, the expression (R1%R2) is 
.B not
equivalent to (R1%)R2.
.IP ~R
match the opposite of the following expression R; i.e. match exactly
those texts that R does not match. This operator is called complement,
or logical not.
.IP R1R2
Two consecutive regular expressions denote catenation:
the left expression must match, and then the right.
.IP R1|R2
match either the expression R1 or R2. This operator is known by
a number of names: union, logical or, disjunction, branch, or alternative.
.IP R1&R2
match both the expression R1 and R2 simultaneously; i.e. the
matching text must be one of the texts which are in the intersection of the set
of texts matched by R1 and the set matched by R2. This operator is called
intersection, logical and, or conjunction.

.PP
Any of the special characters, including the delimiting /,  can be escaped with
a backslash to suppress its meaning and denote the character itself.

Furthermore, all of the same escapes are as described in the section Special
Characters in Text above---the difference is that in regular expressions, the @
character is not required, so for example a tab is coded as \et rather
than @\e\t.

Any escaped character which does not fall into the above escaping conventions,
or any unescaped character which is not a regular expression operator, denotes
one-position match of that character itself.

Precedence table, highest to lowest:
.TS
tab(!);
l l l.
operators!class!associativity
(R) []!primary!
R? R+ R* R%...!postfix!left-to-right
R1R2!catenation!left-to-right
~R ...%R!unary!right-to-left
R1&R2!intersection!left-to-right
R1|R2!union!left-to-right
.TE

The % operator is like a postfix operator with respect to its left
operand, but like a unary operator with respect to its right operand.
Thus a~b%c~d  is a(~(b%(c(~d)))), demonstrating right-to-left associativity,
where all of b% may be regarded as a unary operator being applied to c~d. 
Similarly, a?*+%b  means (((a?)*)+)%b, where the trailing %b behaves
like a postfix operator.

In
.B txr,
regular expression matches do not span multiple lines. The regex language has
no feature for multi-line matching. However, the @(freeform) directive
allows the remaining portion of the input to be treated as one string
in which line terminators appear as explicit characters. Regular expressions
may freely match through this sequence.

It's possible for a regular expression to match an empty string.
For instance, if the next input character is z, facing a
the regular expression /a?/, there is a zero-character match:
the regular expression's state machine can reach an acceptance
state without consuming any characters. Examples:

  pattern:      @A@/a?/@/.*/
  data:         zzzzz
  result:       A=""

  pattern:      @{A /a?/}@B
  data:         zzzzz
  result:       A="", B="zzzz"

  pattern:      @*A@/a?/
  data:         zzzzz
  result:       A="zzzzz"

In the first example, variable @A is followed by a regular expression
which can match an empty string. The expression faces the letter "z"
at position 0 in the data line. A zero-character match occurs there,
therefore the variable A takes on the empty string. The @/.*/ regular
expression then consumes the line.

Similarly, in the second example, the /a?/ regular expression faces
a "z", and thus yields an empty string which is bound to A. Variable
@B consumes the entire line.

The third example requests the longest match for the variable binding.
Thus, a search takes place for the rightmost position where the
regular expression matches. The regular expression matches anywhere,
including the empty string after the last character, which is
the rightmost place. Thus variable A fetches the entire line.

For additional information about the advanced regular expression
operators, NOTES ON EXOTIC REGULAR EXPRESSIONS below.

.SS Directives

The general syntax of a directive is:

  @EXPR

where expr is a parenthesized list of subexpressions. A subexpression
is an symbol, number, string literal, character literal, quasiliteral, regular
expression, or a parenthesized expression.  So, examples of syntactically valid
directives are:

  @(banana)

  @(a b c (d e f))

  @(  a (b (c d) (e  ) ))

  @("apple" #\eb #\espace 3)

  @(a /[a-z]*/ b)

  @(_ `@file.txt`)

A symbol is lexically the same thing as a variable and the same rules
apply. Tokens that look like numbers are treated as numbers.

Character literals are introduced by the #\ syntax, which is either
followed by a character name, the letter x followed by hex digits,
or a single character. Valid character names are: nul, alarm, backspace, tab,
linefeed, newline, vtab, page, return, esc, space. This convention
for character literals is similar to that of the Scheme language.

String literals are delimited by double respectively, and may not span multiple
lines. A double quote within a string literal is encoded using \e"
and a backslash is encoded as \e\e. Backslash escapes like \en and \et
are recognized, as are hexadecimal escapes like \exFF and octal
escapes like \e123.

Quasiliterals are similar to string literals, except that they may
contain variable references denoted by the usual @ syntax. The quasiliteral
represents a string formed by substituting the values of those variables
into the literal template. If a is bound to "apple" and b to "banana",
the quasiliteral `one@a and two @{b}s` represents the string
"one apple and two bananas". A backquote escaped by a backslash represents
itself, and two consecutive @ characters code for a literal @.
There is no \e@ escape.

Some directives are involved in structuring the overall syntax of the query.

There are syntactic constraints that depend on the directive. Some directives
are "vertical only". They must occur on a line by themselves. If they are
involved in additional syntax, it is line-oriented.  Others work horizontally.
They can occur anywhere in a line, and if they are involved in syntax, it hs
character-oriented.  Some work in both modes, with slightly different
semantics.

A summary of the available directives follows:

.IP @(eof)
Explicitly match the end of file. Fails if unmatched data remains in
the input stream.

.IP @(eol)
Explicitly match the end of line. Fails if the the current position is not the
end of a line. Also Fails if no data remains (there is no current line).

.IP @(next)
Continue matching in another file or other data source.

.IP @(block)
The remaining query is treated as an anonymous or named block.
Blocks may be referenced by @(accept) and @(fail) directives.
Blocks are discussed in the section BLOCKS below.

.IP @(skip)
Treat the remaining query as a subquery unit, and search the lines (or
characters) of the input file until that subquery matches somewhere.  A skip is
also an anonymous block.

.IP @(trailer)
Treat the remaining query or subquery as a match for a trailing context. That
is to say, if the remainder matches, the data position is not advanced.

.IP @(freeform)
Treat the remainder of the input as one big string, and apply the following
query line to that string. The newline characters (or custom separators) appear
explicitly in that string.

.IP @(some)
Multiple clauses are each applied to the same input. Succeeds if at least one
of the clauses matches the input. The bindings established by earlier
successful clauses are visible to the later clauses.

.IP @(all)
Multiple clauses are applied to the same input. Succeeds if and only if each
one of the clauses matches. The clauses are applied in sequence, and evaluation stops on the first failure.  The bindings established by earlier
successful clauses are visible to the later clauses.

.IP @(none)
Multiple clauses are applied to the same input. Succeeds if and only if none of
them match. The clauses are applied in sequence, and evaluation stops on the
first success. No bindings are ever produced by this construct.

.IP @(maybe)
Multiple clauses are applied to the same input. No failure occurs if none of
them match.  The bindings established by earlier successful clauses are visible
to the later clauses.

.IP @(cases)
Multiple clauses are applied to the same input. Evaluation stops on the
first successful clause.

.IP @(choose)
Multiple clauses are applied to the same input. The one whose effect persists
is the one which maximizes or minimizes the length of a particular variable.

.IP @(define\ NAME\ (\ ARGUMENTS\ ...))
Introduces a function. Functions are discussed in the FUNCTIONS section below.

.IP @(gather)
Searches text for matches for multiple clauses which may occur in arbitrary
order. For convenience, lines of the first clause are treated as separate
clauses.

.IP @(collect)
Search the data for multiple matches of a clause. Collect the
bindings in the clause into lists, which are output as array variables.
The @(collect) directive is line oriented. It works with a multi-line
pattern and scans line by line. A similar directive called @(coll)
works within one line.

A collect is an anonymous block.

.IP @(and)
Separator of clauses for @(some), @(all), @(none), @(maybe) and @(cases).
Equivalent to @(or). Choice is stylistic.

.IP @(or)
Separator of clauses for @(some), @(all), @(none), @(maybe) and @(cases).
Equivalent to @(and). Choice is stylistic.

.IP @(end)
Required terminator for @(some), @(all), @(none), @(maybe), @(cases),
@(collect), @(coll), @(output), and @(repeat).

.IP @(fail)
Terminate the processing of a block, as if it were a failed match.
Blocks are discussed in the section BLOCKS below.

.IP @(accept)
Terminate the processing of a block, as if it were a successful match.
What bindings emerge may depend on the kind of block: collect
has special semantics.  Blocks are discussed in the section BLOCKS below.

.IP @(try)
Indicates the start of a try block, which is related to exception
handling, discussed in the EXCEPTIONS section below.

.IP @(catch),\ @(finally)
Special clauses within @(try). See EXCEPTIONS below.

.IP @(defex),\ @(throw)
Define custom exception types; throw an exception.  See EXCEPTIONS below.

.IP @(flatten)
Normalizes a set of specified variables to one-dimensional lists. Those
variables which have scalar value are reduced to lists of that value.
Those which are lists of lists (to an arbitrary level of nesting) are converted
to flat lists of their leaf values.

.IP @(merge)
Binds a new variable which is the result of merging two or more
other variables. Merging has somewhat complicated semantics.

.IP @(cat)
Decimates a list (any number of dimensions) to a string, by catenating its
constituent strings, with an optional separator string between all of the
values.

.IP @(bind)
Binds one or more variables against another variable using a structural
pattern. A limited form of unification takes place which can cause a match to
fail.

.IP @(set)
Destructively assigns one or more existing variables using a structural
pattern, using syntax similar to bind. Assignment to unbound
variables triggers an error.

.IP @(output)
A directive which encloses an output clause in the query. An output section
does not match text, but produces text. The directives above are not
understood in an output clause.

.IP @(repeat)
A directive understood within an @(output) section, for repeating multi-line
text, with successive substitutions pulled from lists. The directive @(rep)
produces iteration over lists horizontally within one line.

.IP @(deffilter)
This directive is used for defining named filters, which are useful
for filtering variable substitutions in output blocks. Filters are useful
when data must be translated between different representations that
have different special characters or other syntax, requiring escaping
or similar treatment. Note that it is also possible to use a function
as a filter. See Function Filters below.

.IP @(filter)
The filder directive passes one or more variables through a given
filter or chain or filters, updating them with the filtered values.

.PP

.SS The Next Directive

The next directive comes in two forms, one of which is obsolescent
syntax.  The directive indicates that the remainder of the query
is to be applied to a new input source.

In the first form, it can occur by itself as the only element in a query line,
with, or without arguments:

  @(next)
  @(next SOURCE)
  @(next SOURCE :nothrow)
  @(next :args)
  @(next :env)
  @(next :list EXPR)
  @(next :string EXPR)

The lone @(next) without arguments switches to the next file in the
argument list which was passed to the
.B txr
utility. If SOURCE is given, it must be text-valued expression which denotes an
input source; it may be a string literal, quasiliteral or a variable.
For instance, if variable A contains the text "data", then

  @(next A)

means switch to the file called "data", and

  @(next `@A.txt`)

means to switch to the file "data.txt".

If the input source cannot be opened for whatever reason,
.B txr
throws an exception (see EXCEPTIONS below). An unhandled exception will
terminate the program.  Often, such a drastic measure is inconvenient;
if @(next) is invoked with the :nothrow keyword, then if the input
source cannot be opened, the situation is treated as a simple
match failure.

The variant @(next :args) means that the remaining command line arguments are to
be treated as a data source. For this purpose, each argument is considered to
be a line of text. If an argument is currently being processed as an input
source, that argument is included. Note that if the first entry in the argument
list does not name an input source, then the query should begin with
@(next :args) or some other form of next directive, to prevent an attempt to
open the input source named by that argument. If the very first directive of a query is any variant of the next directive, then
.B txr
avoids opening the first input source, but it does open the input source for
any other directive, even one which does not consume any data.

The variant @(next :env) means that the list of process enviornment variables
is treated as a source of data. It looks like a text file stream
consisting of lines of the form "name=value".  If this feature is not available
on a given platform, an exception is thrown.

The syntax @(next :list EXPR) treats the expression as a source of
text. The value of the expression is flattened to a list in a way similar
to the @(flatten) directive.  The resulting list is treated as if it were the
lines of a text file: each element of the list is a line. If the lines
happen contain embedded newline characters, they are a visible constituent
of the line, and do not act as line separators.

The syntax @(next :string EXPR) treats the expression as a source of
text. The value of the expression must be a string. Newlines in the string are
interpreted as line terminators. 

A string which is not terminated by
a newline is tolerated, so that:

  @(next :string "abc")
  @a

binds a to "abc".  Likewise, this is also the case with input files and other
streams whose last line is not terminated by a newline.

However, watch out for empty strings, which are analogous to a correctly formed
empty file which contains no lines:

  @(next :string "")
  @a

This will not bind a to ""; it is a matching failure.  The behavior of :list is
different. The query

  @(next :list "")
  @a

binds a to "".  The reason is that under :list the string "" is flattened to
the list ("") which is not an empty input stream, but a stream consisting of
one empty line.

Note that "remainder of the query" which is applied to the stream opened
by @(next) refers to the subquery in which the next directive appears, not
necessarily the entire query.  For example, the following query looks for the
line starting with "xyz" at the top of the file "foo.txt", within a some
directive.  After the @(end) which terminates the @(some), the "abc" is matched
in the previous file again.

  @(some)
  @(next "foo.txt")
  xyz@suffix
  @(end) abc

However, if the @(some) subquery successfully matched "xyz@suffix" within the
file foo.text,  there is now a binding for the suffix variable, which
is visible to the remainder of the entire query. The variable bindings
survive beyond the clause, but the data stream does not.

The @(next) directive supports the file name conventions as the command
line. The name - means standard input. Text which starts with a ! is
interpreted as a shell command whose output is read like a file.  These
interpretations are applied after variable substitution. If the file is
specified as @a, but the variable a expands to "!echo foo", then the output of
the "echo foo" command will be processed.

.SS The Skip Directive

The skip directive considers the remainder of the query as a search
pattern. The remainder is no longer required to strictly match at the
current line in the current file. Rather, the current file is searched,
starting with the current line, for the first line where the entire remainder
of the query will successfully match. If no such line is found, the skip
directive fails. If a matching position is found, the remainder of
the query is understood to be processed there.

Of course, the remainder of the query can itself contain skip directives.
Each such directive performs a recursive subsearch.

Skip comes in vertical and horizontal flavors. For instance, skip and match the
last line:

  @(skip)
  @last
  @(eof)

Skip and match the last character of the line:

  @(skip)@{last 1}@(eol)

The skip directive has two optional arguments. If the first argument is a
number, its value limits the range of lines scanned for a match. Judicious use
of this feature can improve the performance of queries.

Example: scan until "size: @SIZE" matches, which must happen within
the next 15 lines:

  @(skip 15)
  size: @SIZE

Without the range limitation skip will keep searching until it consumes
the entire input source. While sometimes this is what is intended,
often it is not. Sometimes a skip is nested within a collect, or
following another skip. For instance, consider:

  @(collect)
  begin @BEG_SYMBOL
  @(skip)
  end @BEG_SYMBOL
  @(end)

The collect iterates over the entire input. But, potentially, so does
the skip. Suppose that "begin x" is matched, but the data has no
matching "end x". The skip will search in vain all the way to the end of the
data, and then the collect will try another iteration back at the
beginning, just one line down from the original starting point.  If it is a
reasonable expectation that an "end x" occurs 15 lines of a "begin x", this can
be written instead:

  @(collect)
  begin @BEG_SYMBOL
  @(skip 15)
  end @BEG_SYMBOL
  @(end)

If the symbol nil is used in place of a number, it means to scan
an unlimited range of lines; thus, @(skip nil) is equivalent to @(skip).

If the symbol :greedy is used, it changes the semantics of the skip
to longest match semantics, like the regular expression * operator.
For instance, match the last three space-separated tokens of the line:

  @(skip :greedy) @a @b @c

Without :greedy, the variable @c will can match multiple tokens,
and end up with spaces in it, because nothign follows @c and
so it matches from any position which follows a space to the
end of the line.  Also note the space in front of @a. Without this
space, @a will get an empty string.

A line oriented example of greedy skip: match the last line without
using @eof:

  @(skip :greedy)
  @last_line

There may be a second numeric argument. This specifies a minimum
number of lines to skip before looking for a match. For instance,
skip 15 lines and then search indefinitely for "begin ...":

  @(skip nil 15)
  begin @BEG_SYMBOL

The two arguments may be used together. For instance, the following
matches if, and only if, the 15th line of input starts with "begin ":

  @(skip 1 15)
  begin @BEG_SYMBOL

Essentially, @(skip 1 <n>) means "hard skip by <n>" lines, then
match the query without scanning.  @(skip 1 0) is the same as @(skip 1), which
is a noop, because it means: "the remainder of the query must match starting on
the very next line", or, more briefly, "skip exactly zero lines", which is the
behavior if the skip directive is omitted altogether.

Here is one trick for grabbing the fourth line from the bottom of the input:

  @(skip)
  @fourth_from_bottom
  @(skip 1 3)
  @(eof)

Or using greedy skip:

  @(skip :greedy)
  @fourth_from_bottom
  @(skip 1 3)

Nongreedy skip with the @(eof) has a slight advantage because the greedy skip
will keep scanning even though it has found the correct match, then backtrack
to the last good match once it runs out of data. The regular skip with explicit
@(eof) will stop when the @(eof) matches.

.SS The Trailer Directive

The trailer directive introduces a trailing portion of a query or subquery
which matches input material normally, but in the event of a successful match,
does not advance the current position. This can be used, for instance, to
cause @(collect) to match partially overlapping regions.

Example:

  @(collect)
  @line
  @(trailer)
  @(skip)
  @line
  @(end)

This script collects each line which has a duplicate somewhere later
in the input. Without the @(trailer) directive, this does not work properly
for inputs like:

  111
  222
  111
  222

Without @(trailer), the first duplicate pair constitutes a match which
spans over the 222. After that pair is found, the matching continues
after the second 111.

With the @(trailer) directive in place, the collect body, on each
iteration, only consumes the lines matched prior to @(trailer).

.SS The Freeform Directive

The freeform directive provides a useful alternative to
.B txr's
line-oriented matching discipline. The freeform directive treats all remaining
input from the current input source as one big line. The query line which
immediately follows freeform is applied to that line.

The syntax variations are:

  @(freeform)
  ... query line ..

  @(freeform NUMBER)
  ... query line ..

  @(freeform STRING)
  ... query line ..

  @(freeform NUMBER STRING)
  ... query line ..

The string and numeric arguments, if both are present, may be given in either
order.

If a numeric argument is given, it limits the range of lines which are combined
together. For instance @(freeform 5) means to only consider the next five lines
to to be one big line. Without a numeric argument, freeform is "bottomless". It
can match the entire file, which creates the risk of allocating a large amount
of memory.

If a string argument is given, it specifies a custom line terminator. The
default terminator is "\en". The terminator does not have to be one character
long.

Freeform does not convert the entire remainder of the input into one big line
all at once, but does so in a dynamic, lazy fashion, which takes place as the
data is accessed. So at any time, only some prefix of the data exists as a flat
line in which newlines are replaced by the terminator string, and the remainder
of the data still remains as a list of lines.

After the subquery is applied to the virtual line, the unmatched remainder
of that line is broken up into multiple lines again, by looking for and
removing all occurences of the terminator string within the flattened portion.

Care must be taken if the terminator is other than the default "\en". All
occurences of the terminator string are treated as line terminators in
the flattened portion of the data, so extra line breaks may be introduced.
Likewise, in the yet unflattened portion, no breaking takes place, even if
the text contains occurences of the terminator string. The extent of data which
is flattened, and the amount of it which remains, depends entirely on the
query line underneath @(flatten).

In the following example, lines of data are flattened using $ as the line
terminator.

  Query:        @(freeform "$")
                @a$@b:
                @c
                @d

  Data:         1
                2:3
                4

  Output:       a="1"
                b="2"
                c="3"
                d="4"

The data is turned into the virtual line 1$2:3$4$.  The @a$@b: subquery matches
the 1$2: portion, binding a to 1, and b to 2.  The remaining portion 3$4$ is
then split into separate lines again according to the line terminator $:

                3
                4

Thus the remainder of the query

                @c
                @d

faces these lines, binding c to 3 and d to 4. Note that since the data
does not contain dollar signs, there is no ambiguity; the meaning may be
understood in terms of the entire data being flattened and split again.

In the following example, freeform is used to solve a tokenizing problem.  The
Unix password file has fields separated by colons. Some fields may be empty.
Using freeform, we can join the password file using ":" as a terminator.
By restricting freeform to one line, we can obtain each line of the password
file with a terminating ":", allowing for a simple tokenization, because
now the fields are colon-terminated rather than colon-separated.

Example:

  @(next "/etc/passwd")
  @(collect)
  @(freeform 1 ":")
  @(coll)@{token /[^:]*/}:@(end)
  @(end)

.SS The Some, All, None, Maybe, Cases and Choose directives

These directives, called the parallel directives, combine multiple subqueries,
which are applied at the same input position, rather than to consecutive input.

They come in vertical (line mode) and horizontal (character mode) flavors.

In horizontal mode, the current position is understood to be a character
position in the line being processed. The clauses advance this character
position by moving it to the right.  In vertical mode, the current position is
understood to be a line of text within the stream. A clause advances the
position by some whole number of lines.

The syntax of these parallel directives follows this example:

  @(some)
  subquery1
  .
  .
  .
  @(and)
  subquery2
  .
  .
  .
  @(and)
  subquery3
  .
  .
  .
  @(end)

And in horizontal mode:

  @(some)subquery1...@(and)subquery2...@(and)subquery3...@(end)

Long horizontal lines can be broken up with line continuations, allowing the
above example to be written like this, which is considered a single logical
line:

  @(some)@\
     subquery1...@\
  @(and)@\
     subquery2...@\
  @(and)@\
     subquery3...@\
  @(end)

The @(some), @(all), @(none), @(maybe), @(cases) or @(choose) must be followed
by at least one subquery clause, and be terminated by @(end). If there are two
or more subqueries, these additional clauses are indicated by @(and) or @(or),
which are interchangeable.  The separator and terminator directives also must
appear as the only element in a query line.

The choose directive requires keyword arguments. See below.

The syntax supports arbitrary nesting. For example:

  QUERY:            SYNTAX TREE:

  @(all)            all -+
  @  (skip)              +- skip -+
  @  (some)              |        +- some -+
  it                     |        |        +- TEXT
  @  (and)               |        |        +- and
  @    (none)            |        |        +- none -+
  was                    |        |        |        +- TEXT
  @    (end)             |        |        |        +- end
  @  (end)               |        |        +- end
  a dark                 |        +- TEXT
  @(end)                 *- end

nesting can be indicated using whitespace between @ and the
directive expression. Thus, the above is an @(all) query containing a @(skip)
clause which applies to a @(some) that is followed by the the text
line "a dark". The @(some) clause combines the text line "it",
and a @(none) clause which contains just one clause consisting of
the line "was".

The semantics of the parallel directives is:

.IP @(all)
Each of the clauses is matched at the current position. If any of the
clauses fails to match, the directive fails (and thus does not produce
any variable bindings). Clauses following the failed directive are not
evaluated. Bindings extracted by a successful clause are visible the clauses
which follow, and if the directive succeeds, all of the combined bindings
emerge.

.IP @(some)
Each of the clauses is matched at the current position. If any of the clauses
succeed, the directive succeeds, retaining the bindings accumulated by the
successully matching clauses.  Evaluation does not stop on the first successful
clause. Bindings extracted by a successful clause are visible to the clauses
which follow.

.IP @(none)
Each of the clauses is matched at the current position. The
directive succeeds only if all of the clauses fail. If
any clause succeeds, the directive fails, and subsequent clauses are not
evaluated. Thus, this directive never produces variable bindings, only matching
success or failure.

.IP @(maybe)
Each of the clauses is matched at the current position.  The directive always
succeeds, even if all of the clauses fail.  Whatever bindings are found in any
of the clauses are retained. Bindings extracted by any successful clause are
visible the clauses which follow.

.IP @(cases)
Each of the clauses is matched at the current position. The
The clauses are matched, in order, at the current position.
If any clause matches, the matching stops and the bindings
collected from that clause are retained. Any remaining clauses
after that one are not processed. If no clause matches, the
directive fails, and produces no bindings.

.IP @(choose { :longest <var> | :shortest <var> })
Each of the clauses is matched at the current position in order. In this
construct, bindings established an earlier clause are not visible to later
clauses.  Although any or all of the clauses can potentially match, the clause
which succeeds is the one which maximizes or minimizes the length of the
text bound to the specified variable. The other clauses have no effect.

For all of the parallel directives other than @(none) and @(choose), the query
advances the input position by the greatest number of lines that match in any
of the successfully matching subclauses that are evaluated.  The @(none)
directive does not advance the input position.

For instance if there are two subclauses, and one of them matches three lines,
but the other one matches five lines, then the overall clause is considered to
have made a five line match at its position. If more directives follow, they
begin matching five lines down from that position.

.SS The Gather Directive

Sometimes text is structured as items that can appear in an arbitrary order.
When multiple matches need to be extracted, there is a combinatorial explosion
of possible orders, making it impractical to write pattern matches for all
the possible orders.

The gather directive is for these situations. It specifies multiple clauses
which all have to match somewhere in the data, but in any order.

For further convenience, the lines of the first clause of the gather directive
are implicitly treated as separate clauses.

The syntax follows this pattern

  @(gather)
  one-line-query1
  one-line-query2
  .
  .
  .
  one-line-queryN
  @(and)
  multi
  line
  query1
  .
  .
  .
  @(and)
  multi
  line
  query2
  .
  .
  .
  @(end)

Of course the multi-line clauses are optional.   The gather directive takes
keyword parameters, see below.

How gather works is that the text is searched for matches for the single line
and multi-line queries. The clauses are applied in the order in which they appear.
Whenever one of the clauses matches, any bindings it produces are retained and
it is removed from further consideration. Multiple clauses can match at the
same text position.  The position advances by the longest match from among the
clauses which matched.  If no clauses match, the position advances by one line.
The search stops when all clauses are eliminated, and then the cumulative
bindings are produced.  If the data runs out, but unmatched clauses remain, the
directive fails.

Example: extract several environment variables, which do not appear in a particular
order:

  @(next :env)
  @(gather)
  USER=@USER
  HOME=@HOME
  SHELL=@SHELL
  @(end)

.SS Gather Keyword Parameters

The gather diretive accepts the keyword parameter :vars. The argument to vars is a list
of required and optional variables.  Optional variables are denoted by the
specification of a default value. Example:

  @(gather :vars (a b c (d "foo")))
  ...
  @(end)

Here, a, b, c and e are required variables, and d is optional.  Variable e is
required because its default value is the empty list (), same as the symbol
nil. 

The presence of vars changes the behavior in three ways.

Firstly, even if all the clauses in the gather match successfully and are
eliminated, the directive will fail if the required variables do not have
bindings. It doesn't matter whether the bindings are existing, or whether they
are established by the gather.

Secondly, if some of the clauses of the gather did not match, but all
of the required variables have bindings, then the directive succeeds.
Without the presence of :vars, it would fail in this situation.

Thirdly, if the the gather succeeds (all required variables have bindings),
then all of the optional variables which do not have bindings are given
bindings to their default values.

.SS The Collect Directive

The syntax of the collect directive is:

  @(collect)
  ... lines of subquery
  @(end)

or with an until or last clause:

  @(collect)
  ... lines of subquery: main clause
  @(until)
  ... lines of subquery: until clause
  @(end)

  @(collect)
  ... lines of subquery: main clause
  @(last)
  ... lines of subquery: last clause
  @(end)

The subquery is matched repeatedly, starting at the current line.
If it fails to match, it is tried starting at the subsequent line.
If it matches successfully, it is tried at the line following the
entire extent of matched data, if there is one. Thus, the collected regions do
not overlap.

Unless certain keywords are specified, or unless the collect is explicitly
failed with @(fail), it always succeeds, even if it collects nothing,
and even if the until/last clause never finds a match.

If no until/last clause is specified, and the collect is not limited
using parameters, the collect is unbounded. It consumes the entire data
file. If any query material follows such the collect clause, it will fail if it
tries to match anything in the current file; but of course, it is possible to
continue matching in another file by means of @(next).

If an until/last clause is specified, the collection stops when that clause
matches at the current position. 

If it is an until clause, no bindings are collected at that position, even if
the main clause matches at that position also. Moreover, the position is not
advanced.  The remainder of the query begins matching at that position.  

If it is a last clause matches, the behavior is different. Any bindings
captured by the main clause thrown away, just like with the until clause.
However, the bindings in the last clause itself survive, and the position
is advanced to skip over that material.

Example:

  Query:        @(collect)
                @a
                @(until)
                42
                @b
                @(end)
                @c

  Data:         1
                2
                3
                42
                5
                6

  Output:       a[0]="1"
                a[1]="2"
                a[2]="3"
                c="42"

The line 42 is not collected, even though it matches @a. Furthermore,
the until does not advance the position, so variable c takes 42.

If the @(until) is changed to @(last) the output will be different:

  Output:       a[0]="1"
                a[1]="2"
                a[2]="3"
                b=5
                c=6

The 42 is not collected into the a list, just like before. But now
the binding captured by @b emerges. Furthermore, the position advances
so variable now takes 6.

The binding variables within the clause of a collect are treated specially.
The multiple matches for each variable are collected into lists,
which then appear as array variables in the final output.

Example:

  Query:        @(collect)
                @a:@b:@c
                @(end)

  Data:         John:Doe:101
                Mary:Jane:202
                Bob:Coder:313

  Output:
                a[0]="John"
                a[1]="Mary"
                a[2]="Bob"
                b[0]="Doe"
                b[1]="Jane"
                b[2]="Coder"
                c[0]="101"
                c[1]="202"
                c[2]="313"

The query matches the data in three places, so each variable becomes
a list of three elements, reported as an array.

Variables with list bindings may be referenced in a query. They denote a
multiple match. The -D command line option can establish a one-dimensional
list binding.

Collect clauses may be nested.   Variable matches collated into lists in an
inner collect, are again collated into nested lists in the outer collect.
Thus an unbound variable wrapped in N nestings of @(collect) will
be an N-dimensional list. A one dimensional list is a list of strings;
a two dimensional list is a list of lists of strings, etc.

It is important to note that the variables which are bound within the main
clause of a collect---i.e. the variables which are subject to
collection---appear, within the collect, as normal one-value bindings. The
collation into lists happens outside of the collect. So for instance in the
query:

 @(collect)
 @x=@x
 @(end)

The left @x establishes a binding for some material preceding an equal sign.
The right @x refers to that binding. The value of @x is different in each
iteration, and these values are collected. What finally comes out of the
collect clause is list variable called x which holds each value that
was ever instantiated under that name within the collect clause.

Also note that the until clause has visibility over the bindings
established in the main clause. This is true even in the terminating
case when the until clause matches, and the bindings of the main clause
are discarded.

.SS Collect Keyword Parameters

By default, collect searches the rest of the input indefinitely,
or until the @(until) clause matches. It skips arbitrary amounts of
nonmatching material before the first match, and between matches.

Within the @(collect) syntax, it is possible to specify some useful keyword
parameters for additional control of the behavior. For instance

  @(collect :maxgap 5)

means that the collect will terminate if it does not find a match within five
lines of the starting position, or if more than five lines are skipped since
any successful match. A :maxgap of 0 means that the collected regions must be
adjacent. For instance:

  @(collect :maxgap 0)
  M @a
  @(end)

means: from here, collect consecutive lines of the form "M ...". This will not
search for the first such line, nor will it skip lines which do not match this
form.

Other keywords are :mingap, and :gap.  The :mingap keyword specifies a minimum
gap between matches, but has no effect on the distance to the first match. The
:gap keyword specifies  :mingap and :maxgap  at the same time, and can only be
used if these other two are not used. Thus:

 @(collect :gap 1)
 @a
 @(end)

means collect every other line starting with the current line. Several
other supported keywords are :times, :mintimes, :maxtimes and lines.
The shorthand :times N means the same thing as :mintimes N :maxtimes N.
These specify how many matches should be collected. If there are fewer
than mintimes matches, the collect fails. If maxtimes matches are collected,
collect stops collecting immediately. Example:

 @(collect :times 3)
 @a @b
 @(end)

This will collect a match for "@a @b" exactly three times. If three
matches are not found, it will fail.

The :lines parameter specifies the upper bound on how many lines
should be scanned by collect, measuring from the starting position.
The extent of the collect body is not counted. Example:

 @(collect :lines 2)
 foo: @a
 bar: @b
 baz: @c
 @(end)

The above collect will look for a match only twice: at the current position,
and one line down.

There is one more keyword, :vars, discussed in the following section.

.SS Specifying Variables in Collect

Normally, any variable for which a new binding occurs in a collect is
collected. A collect clause may be sloppy: it can neglect to collect some
variables on some iterations, or bind some variables which behave like
local temporaries, but end up collated into lists. Another issue is that
the collect clause might not match anything at all, and then none of
the variables are bound.

The :vars keyword allows the query writer to add discipline the collect body.

The argument to :vars is a list of variable specs. A variable spec is either a
symbol, or a (<symbol> <expression>) pair, where the expression specifies a
default value.

When a :vars list is specified, it means that only the given variables can
emerge from the successful collect. Any newly introduced bindings for other
variables do not propagate.

Furthermore, for any variable which is not specified with a default value, the
collect body, whenever it matches successfully, must bind that variable. If it
neglects to bind the variable, an exception of type query_error is thrown.

For any variable which does have a default value, if the collect body neglects
to bind that variable, the behavior is as if the collect did bind that variable
to that default value.

The default values are expressions, and so can be quasiliterals.

Lastly, if in the event that the collect does not match anything, the variables
specified in vars (whether or not they have a default value) are all bound to
empty lists. (These bindings are established after the processing of the
until/last clause, if present.)

Example:

  @(collect :vars (a b (c "foo")))
  @a @c
  @(end)

Here, if the body "@a @c" matches, an error will be thrown because one of the
mandatory variables is b, and the body neglects to produce a binding for b.

Example:

  @(collect :vars (a (c "foo")))
  @a @b
  @(end)

Here, if "@a @b" matches, only a will be collected, but not b, because b is not
in the variable list. Furthermore, because there is no binding for c in the
body, a binding is created with the value "foo", exactly as if c matched
such a piece of text.

In the following example, the assumption is that THIS NEVER MATCHES
is not found anywhere in the input but the line THIS DOES MATCH is
found and has a successor which is bound to a. Because the body did not
match, the :vars a and b should be bound to empty lists. But a is bound
by the last clause to some text, so this takes precedence. Only b is bound to a
an empty list.

  @(collect :vars (a b)
  THIS NEVER MATCHES
  @(last)
  THIS DOES MATCH
  @a
  @(end)

.SS The Coll Directive

The coll directive is a kind of miniature version of the collect directive.
Whereas the collect directive works with multi-line clauses on line-oriented
material, coll works within a single line. With coll, it is possible to
recognize repeating regularities within a line and collect lists.

Regular-expression based Positive Match variables work well with coll.

Example: collect a comma-separated list, terminated by a space.

  pattern:  @(coll)@{A /[^, ]+/}@(until) @(end)@B
  data:     foo,bar,xyzzy blorch
  result:   A[0]="foo"
            A[1]="bar"
            A[2]="xyzzy"
            B=blorch

Here, the variable A is bound to tokens which match the regular
expression /[^, ]+/: non-empty sequence of characters other than commas or
spaces.

Like its big cousin, the coll directive searches for matches.  If no match
occurs at the current character position, it tries at the next character
position. Whenever a match occurs, it continues at the character position which
follows the last character of the match, if such a position exists.

If not bounded by an until clause, it will exhaust the entire line.  If the
until clause matches, then the collection stops at that position,
and any bindings from that iteration are discarded.
Like collect, coll also supports a last clause, which propagates varaible
bindings and advances the position.

Coll clauses nest, and variables bound within a coll are available to within
the rest of the coll clause, including the until clause, and appear as single
values.  The final list aggregation is only visible after the coll clause.

The behavior of coll is troublesome, when delimited variables are used,
because in text file formats, the material which separates items is not
repeated after the last item. For instance, a comma-separated list usually
not appear as "a,b,c," but rather "a,b,c". There might not be any explicit
termination---the last item might be at the very end of the line.

So for instance, the following result is not satisfactory:

  pattern:      @(coll)@a @(end)
  data:         1 2 3 4 5
  result:       a[0]="1"
                a[1]="2"
                a[2]="3"
                a[3]="4"

What happened to the 5? After matching "4 ", coll continues to look for
matches. It tries "5", which does not match, because it is not followed by a
space. Then the line is consumed.  So in this sequence, a valid item is either
followed by a space, or by nothing. So it is tempting to try this:

  pattern:      @(coll)@a@/ ?/@(end)
  data:         1 2 3 4 5
  result:       a[0]=""
                a[1]=""
                a[2]=""
                a[3]=""
                a[4]=""
                a[5]=""
                a[6]=""
                a[7]=""
                a[8]=""

however, the problem is that the regular expression / ?/ (match either a space
or nothing), matches at any position.  So when it is used as a variable
delimiter, it matches at the current position, which binds the empty string to
the variable, the extent of the match being zero. In this situation, the coll
directive proceeds character by character. The solution is to use
positive matching: specify the regular expression which matches the item,
rather than a trying to match whatever follows.  The collect directive will
recognize all items which match the regular expression.

  pattern:      @(coll)@{a /[^ ]+/}@(end)
  data:         1 2 3 4 5
  result:       a[0]="1"
                a[1]="2"
                a[2]="3"
                a[3]="4"
                a[4]="5"

The until clause can specify a pattern which, when recognized, terminates
the collection. So for instance, suppose that the list of items may
or may not be terminated by a semicolon. We must exclude
the semicolon from being a valid character inside an item, and
add an until clause which recognizes a semicolon:

  pattern:      @(coll)@{a /[^ ;]+/}@(until);@(end);

  data:         1 2 3 4 5;
  result:       a[0]="1"
                a[1]="2"
                a[2]="3"
                a[3]="4"
                a[4]="5"

  data:         1 2 3 4 5;
  result:       a[0]="1"
                a[1]="2"
                a[2]="3"
                a[3]="4"
                a[4]="5"

Semicolon or not, the items are collected properly.

Note that the @(end) is followed by a semicolon. That's because
when the @(until) clause meets a match, the matching material
is not consumed.

Instead of regular expression hacks, this problem can be nicely
solved with cases:

  pattern:      @(coll)@(cases)@a @(or)@a@(end)@(end)
  data:         1 2 3 4 5
  result:       a[0]="1"
                a[1]="2"
                a[2]="3"
                a[3]="4"
                a[4]="5"

.SS Coll Keyword Parameters

The @(coll) directive takes most of the same parameters as @(collect).
See the section Collect Keyword Parameters above.
So for instance @(coll :gap 0) means that the collects must be
consecutive, and @(coll :maxtimes 2) means that at most two matches
will be collected.  The :lines keyword does not exist, but there is
an analogous :chars keyword.

.SS The Flatten Directive.

The flatten directive can be used to convert variables to one dimensional
lists. Variables which have a scalar value are converted to lists containing
that value. Variables which are multidimensional lists are flattened to
one-dimensional lists.

Example (without @(flatten))

  pattern:      @b
                @(collect)
                @(collect)
                @a
                @(end)
                @(end)

  data:         0
                1
                2
                3
                4
                5

  result:       b="0"
                a_0[0]="1"
                a_1[0]="2"
                a_2[0]="3"
                a_3[0]="4"
                a_4[0]="5"

Example (with flatten):

  pattern:      @b
                @(collect)
                @(collect)
                @a
                @(end)
                @(end)
                @(flatten a b)

  data:         0
                1
                2
                3
                4
                5

  result:       b[0]="0"
                a[0]="1"
                a[1]="2"
                a[2]="3"
                a[3]="4"
                a[4]="5"

.SS The Merge Directive

The merge directive provides a way of combining two or more variables
in a somewhat complicated but very useful way. 

To understand what merge does we first have to define a property called depth.
The depth of an atom such as a string is defined as 1. The depth of an empty
list is 0. The depth of a nonempty list is one plus the depth of its deepest
element. So for instance "foo" has depth 1, ("foo") has depth 2, and ("foo"
("bar")) has depth three.

We can now define the binary (two argument) merge operation as follows.  (merge
A B) first normalizes the values A and B such that they have normal depth.
1. A value which has depth zero is put into a one element list. 
2. If either value has a smaller depth than the other, it is wrapped
in a list as many times as needed to give it equal depth.
Finally, the values are appended together.

Merge takes more than two arguments. These are merged by a left reduction. The
leftmost two values are merged, and then this result is merged with the third
value, and so on.

Merge is useful for combining the results from collects at different
levels of nesting such that elements are at the appropriate depth.

.SS The Cat Directive

The @(cat) directive converts a list variable into a single
piece of text. The syntax is:

 @(cat VAR [ SEP ])

The SEP argument specifies a separating piece of text. If no separator
is specified, then a single space is used.

Example:

  pattern:      @(coll)@{a /[^ ]+/}@(end)
                @(cat a ":")
  data:         1 2 3 4 5
  result:       a="1:2:3:4:5"


.SS The Bind Directive

The syntax of the @(bind) directive is:

  @(bind pattern expression { keyword value }*)

The @(bind) directive is a kind of pattern match, which matches one or more
variables on in the left hand side pattern to the value of a variable on the
right hand side.  The right hand side variable must have a binding, or else the
directive fails. Any variables on the left hand side which are unbound receive
a matching piece of the right hand side value. Any variables on the left which
are already bound must match their corresponding value, or the bind fails. Any
variables which are already bound and which do match their corresponding value
remain unchanged (the match can be inexact).

The simplest bind is of one variable against itself, for instance bind A
against A:

  @(bind A A)

This will fail if A is not bound, (and complain loudly). If A is bound, it
succeeds, since A matches A.

The next simplest bind binds one variable to another:

  @(bind A B)

Here, if A is unbound, it takes on the same value as B. If A is bound, it has
to match B, or the bind fails. Matching means that either
.IP - 
A and B are the same text
.IP -
A is text, B is a list, and A occurs within B.
.IP -
vice versa: B is text, A is a list, and B occurs within A.
.IP -
A and B are lists and are either identical, or one is
found as substructure within the other.
.PP
The right hand side does not have to be a variable. It may be some other
object, like a string, quasiliteral, regexp, or list of strings, et cetera. For
instance

  @(bind A "ab\tc")

will bind the string "ab\tc" (the letter a, b, a tab character, and c)
to the variable A if A is unbound. If A is bound, this will fail unless
A already contains an identical string. However, the right hand side of
cannot be an unbound variable, nor a complex expression that contains unbound
variables.

The left hand side of a bind can be a nested list pattern containing variables.
The last item of a list at any nesting level can be preceded by a dot, which
means that the variable matches the rest of the list from that position.

Example: suppose that the list A contains ("now" "now" "brown" "cow"). Then the
directive @(bind (H N . C) A), assuming that H, N and C are unbound variables,
will bind H to "how", N to "now", and C to the remainder of the list ("brown"
"cow").

Example: suppose that the list A is nested to two dimensions and  contains
(("how" "now") ("brown" "cow")). Then @(bind ((H N) (B C)) A)
binds H to "how", N to "now", B to "brown" and C to "cow".

The dot notation may be used at any nesting level. it must be preceded and
followed by a symbol: the forms (.) (. X) and (X .) are invalid.

The number of items in a left pattern match must match the number of items in
the corresponding right side object. So the pattern () only matches
an empty list. The notation () and nil means exactly the same thing.

The symbols nil, t and keyword symbols may be used on either side.
They represent themselves.  For example @(bind :foo :bar) fails,
but @(bind :foo :foo) succeeds since the two sides denote the same
keyword symbol object.

.SS Keyword in The Bind Directive

The Bind directive accepts these keywords

.IP :lfilt
The argument to :lfilt is a filter specification. When the left side pattern
contains a binding which is therefore matched against its counterpart from the
right side expression, the left side is filtered through the filter specified
by :lfilt for the purposes of the comparison. For example:

  @(bind "a" "A" :lfilt :upcase)

produces a match, since the left side is the same as the right after
filtering through the :upcase filter.

.IP :rfilt
The argument to :rfilt is a filter specification. The specified filter is
applied to the right hand side material prior to matching it against
the left side. The filter is not applied if the left side is a variable
with no binding. It is only applied to determine a match. Binding takes
place the unmodified right hand side object.

Example, the following produces a match:

  @(bind "A" "a" :rfilt :upcase)

.IP :filter
This keyword is a shorthand to specify both filters to the same value.
So for instance :filter :upcase is equivalent to :lfilt :upcase :rfilt :upcase.

For a description of filters, see Output Filtering below.

Of course, compound filters like (:from_html :upcase) are supported with
all these keywords. The filters apply across arbitrary patterns and nested data.

Example:

  @(bind (a b c) ("A" "B" "C"))
  @(bind (a b c) (("z" "a") "b" "c") :rfilt :upcase)

Here, the first bind establishes the values for a, b and c, and the second bind
succeeds, because the value of a matches the second element of the list ("z"
"a") if it is upcased, and likewise b matches "b" and c matches "c" if these
are upcased.

.SS The Set Directive

The @(set) directive resembles bind, but is not a pattern match. It overwrites
the previous values of variables with new values from the right hand side.
Each variable that is assigned must have an existing binding.

Examples follow.

Store the value of A back into A, achieving nothing:

  @(set A A)

Exchange the values of A and B:

  @(set (A B) (B A))

Store a string into A:

  @(set A "text")

Store a list into A:

  @(set A ("line1" "line2"))

Destructuring assignment. D assumed to contain the list 

  @(bind D ("A" ("B1" "B2") "C1" "C2"))
  @(bind (A B C) (() () ()))
  @(set (A B . C) D)

A ends up with "A", B ends up with ("B1" "B2") and C gets ("C1" and "C2").


.SH BLOCKS

.SS Introduction

Blocks are sections of a query which are denoted by a name. Blocks denoted by
the name nil are understood as anonymous.

The @(block NAME) directive introduces a named block, except when the name is
the word nil.  The @(block) directive introduces an unnamed block, equivalent
to @(block nil).

The @(skip) and @(collect) directives introduce implicit anonymous blocks,
as do function bodies.

.SS Block Scope

The names of blocks are in a distinct namespace from the variable binding
space. So @(block foo) has no interaction with the variable @foo.

A block extends from the @(block ...) directive which introduces it,
to the end of the subquery in which that directive is contained. For instance:

  @(some)
  abc
  @(block foo)
  xyz
  @(end)

Here, the block foo occurs in a @(some) clause, and so it extends to the @(end)
which terminates that clause.  After that @(end), the name foo is not
associated with a block (is not "in scope"). A block which is not contained in
any subquery extends to the end of the overall query.  Blocks are never
terminated by @(end).

The implicit anonymous block introduced by @(skip) has the same scope
as the @(skip): they extends over all of the material which follows the skip,
to the end of the containing subquery.

.SS Block Nesting

Blocks may nest, and nested blocks may have the same names as blocks in
which they are nested. For instance:

  @(block)
  @(block)
  ...

is a nesting of two anonymous blocks, and

  @(block foo)
  @(block foo)

is a nesting of two named blocks which happen to have the same name.
When a nested block has the same name as an outer block, it creates
a block scope in which the outer block is "shadowed"; that is to say,
directives which refer to that block name within the nested block refer to the
inner block, and not to the outer one.

A more complicated example of nesting is:

  @(skip)
  abc
  @(block)
  @(some)
  @(block foo)
  @(end)

Here, the @(skip) introduces an anonymous block. The explicit anonymous
@(block) is nested within skip's anonymous block and shadows it.
The foo block is nested within both of these.

.SS Block Semantics

A block normally does nothing. The query material in the block is evaluated
normally. However, a block serves as a termination point for @(fail) and
@(accept) directives which are in scope of that block and refer to it.

The precise meaning of these directives is:

.IP @(fail\ NAME)

Immediately terminate the enclosing query block called NAME, as if that block failed to match anything. If more than one block by that name encloses
the directive, the inner-most block is terminated. No bindings
emerge from a failed block.

.IP @(fail)

Immediately terminate the innermost enclosing anonymous block, as if
that block failed to match.

If the implicit block introduced by @(skip) is terminated in this manner,
this has the effect of causing the skip itself to fail. I.e. the behavior
is as if skip search did not find a match for the trailing material,
except that it takes place prematurely (before the end of the available
data source is reached).

If the implicit block associated with a @(collect)  is terminated this way,
then the entire collect fails. This is a special behavior, because a
collect normally does not fail, even if it matches and collects nothing!

To prematurely terminate a collect by means of its anonymous block, without
failing it, use @(accept).

.IP @(accept\ NAME)

Immediately terminate the enclosing query block called NAME, as if that block
successfully matched. If more than one block by that name encloses the
directive, the inner-most block is terminated.  Any bindings established within
that block until this point emerge from that block.

.IP @(accept)

Immediately terminate the innermost enclosing anonymous block, as if
that block successfully matched. Any bindings established within
that block until this point emerge from that block.

If the implicit block introduced by @(skip) is terminated in this manner,
this has the effect of causing the skip itself to succeed, as if
all of the trailing material successfully matched.

If the implicit block associated with a @(collect)  is terminated this way,
then the collection stops. All bindings collected in the current iteration of
the collect are discarded. Bindings collected in previous iterations are
retained, and collated into lists in accordance with the semantics of collect.

Example: alternative way to @(until) termination:

  @(collect)
  @  (maybe)
  ---
  @  (accept)
  @  (end)
  @LINE
  @(end)

This query will collect entire lines into a list called LINE. However,
if the line --- is matched (by the embedded @(maybe)), the collection
is terminated. Only the lines up to, and not including the --- line,
are collected. The effect is identical to:

  @(collect)
  @LINE
  @(until)
  ---
  @(end)

The difference (not relevant in these examples) is that the until clause has
visibility into the bindings set up by the main clause.

However, the following example has a different meaning:

  @(collect)
  @LINE
  @  (maybe)
  ---
  @  (accept)
  @  (end)
  @(end)

Now, lines are collected until the end of the data source, or until a line is
found which is followed by a --- line. If such a line is found,
the collection stops, and that line is not included in the collection!
The @(accept) terminates the process of the collect body, and so the
action of collecting the last @LINE binding into the list is not performed.

.SS Data Extent of Terminated Blocks

A query block may have matched some material prior to being terminated by
accept. In that case, it is deemed to have only matched that material,
and not any material which follows. This may matter, depending on the context
in which the block occurs.

Example:

  Query:        @(some)
                @(block foo)
                @first
                @(accept foo)
                @ignored
                @(end)
                @second

  Data:         1
                2
                3

  Output:       first="1"
                second="2"

At the point where the accept occurs, the foo block has matched the first line,
bound the text "1" to the variable @first. The block is then terminated.
Not only does the @first binding emerge from this terminated block, but
what also emerges is that the block advanced the data past the first line to
the second line. So next, the @(some) directive ends, and propagates the
bindings and position. Thus the @second which follows then matches the second
line and takes the text "2".

In the following query, the foo block occurs inside a maybe clause.
Inside the foo block there is a @(some) clause. Its first subclause
matches variable @first and then terminates block foo. Since block foo is
outside of the @(some) directive, this has the effect of terminating the
@(some) clause:

  Query:        @(maybe)
                @(block foo)
                @  (some)
                @first
                @  (accept foo)
                @  (or)
                @one
                @two
                @three
                @four
                @  (end)
                @(end)
                @second

  Data:         1
                2
                3
                4
                5

  Output:       first="1"
                second="2"

The second clause of the @(some) directive, namely:

  @one
  @two
  @three
  @four

is never processed. The reason is that subclauses are processed in top
to bottom order, but the processing was aborted within the
first clause the @(accept foo). The @(some) construct never had the
opportunity to match four lines.

If the @(accept foo) line is removed from the above query, the output
is different:

  Query:        @(maybe)
                @(block foo)
                @  (some)
                @first
                @#          <--  @(accept foo) removed from here!!!
                @  (or)
                @one
                @two
                @three
                @four
                @  (end)
                @(end)
                @second

  Data:         1
                2
                3
                4
                5

  Output:       first="1"
                one="1"
                two="2"
                three="3"
                four="4"
                second="5"

Now, all clauses of the @(some) directive have the opportunity to match.
The second clause grabs four lines, which is the longest match.
And so, the next line of input available for matching is 5, which goes
to the @second variable.

.SH FUNCTIONS

.SS Introduction

.B txr
functions allow a query to be structured to avoid repetition.
On a theoretical note, because
.B txr
functions support recursion, functions enable txr to match some
kinds of patterns which exhibit self-embedding, or nesting,
and thus cannot be matched by a regular language.

Functions in
.B txr
are not exactly like functions in mathematics or functional languages, and are
not like procedures in imperative programming languages. They are not exactly
like macros either. What it means for a
.B txr
function to take arguments and produce a result is different from
the conventional notion of a function.

A
.B txr
function may have one or more parameters. When such a function is invoked, an
argument must be specified for each parameter.  However, a special behavior is
at play here. Namely, some or all of the argument expressions may be unbound
variables.  In that case, the corresponding parameters behave like unbound
variables also.  Thus
.B txr
function calls can transmit the "unbound" state from argument to parameter.

It should be mentioned that functions have access to all bindings that are
visible in the caller; functions may refer to variables which are not
mentioned in their parameter list.

With regard to returning,
.B txr
functions are also unconventional. If the function fails, then the function
call is considered to have failed. The function call behaves like a kind of
match; if the function fails, then the call is like a failed match.

When a function call succeeds, then the bindings emanating from that function
are processed specially. Firstly, any bindings for variables which do not
correspond to one of the function's parameters are thrown away. Functions may
internally bind arbitrary variables in order to get their job done, but only
those variables which are named in the function argument list may propagate out
of the function call.  Thus, a function with no arguments can only indicate
matching success or failure, but not produce any bindings. Secondly,
variables do not propagate out of the function directly, but undergo
a renaming. For each parameter which went into the function as an unbound
variable (because its corresponding argument was an unbound variable),
if that parameter now has a value, that value is bound onto the corresponding
argument.

Example:

  @(define collect_words (list))
  @(coll)@{list /[^ \t]/}@(end)
  @(end)

The above function "collect_words" contains a query which collects words from a
line (sequences of characters other than space or tab), into the list variable
called "list".  This variable is named in the parameter list of the function,
therefore, its value, if it has one, is permitted to escape from the function
call.

Suppose the input data is:

  Fine summer day

and the function is called like this:

  @(collect_words wordlist)

The result is:

  wordlist[0]=Fine
  wordlist[1]=summer
  wordlist[1]=day

How it works is that in the function call @(collect_words wordlist),
"wordlist" is an unbound variable. The parameter corresponding to that
unbound variable is the parameter "list". Therefore, that parameter
is unbound over the body of the function.  The function body collects the
words of "Fine summer day" into the variable "list", and then
yields the that binding.   Then the function call completes by
noticing that the function parameter "list" now has a binding, and
that the corresponding argument "wordlist" has no binding. The binding
is thus transferred to the "wordlist" variable.  After that, the
bindings produced by the function are thrown away. The only enduring
effects are:
.IP -
the function matched and consumed some input; and
.IP -
the function succeeded; and
.IP -
the wordlist variable now has a binding.
.PP
Another way to understand the parameter behavior is that function
parameters behave like proxies which represent their arguments.  If an argument
is an established value, such as a character string or bound variable, the
parameter is a proxy for that value and behaves just like that value. If an
argument is an unbound variable, the function parameter acts as a proxy
representing that unbound variable. The effect of binding the proxy is
that the variable becomes bound, an effect which is settled when the
function goes out of scope.

Within the function, both the original variable and the proxy are
visible simultaneously, and are independent.  What if a function binds both of
them? Suppose a function has a parameter called P, which is called
with an argument A, and then in the function @A and @P are bound.  This is
permitted, and they can even be bound to different values.  However, when the
function terminates, the local binding of A simply disappears (because,
remember, the symbol A is not a member of the list of parameters).
Only the value bound to P emerges, and is bound to A, which still appears
unbound at that point.

.SS Definition Syntax

Function definition syntax comes in two flavors: vertical and horizontal.
Horizontal definitions actually come in two forms the distinction
between which is hardly noticeable, and the need for which is
made clear below.

A function definition begins with a @(define ...) directive. For vertical
functions, this is the only element in a line.

The define symbol must be followed by a symbol, which is the name of the
function being defined. After the symbol, there is a parenthesized optional
argument list. If there is no such list, or if the list is specified as () or
the symbol "nil" then the function has no parameters. Examples of valid define
syntax are:

  @(define foo)
  @(define bar ())
  @(define match (a b c))

If the define directive is followed by material to the right, then
it it defines a horizontal function:

  @(define match_x)x@(end)

If the define material is followed by material below, then it
is a vertical function:

  @(define match_x)
  x
  @(end)

The difference between the two is that a horizontal function matches
characters within a line, whereas a vertical function matches lines
within a stream. The former match_x matches the character x, advancing
to the next character position.  The latter match_x matches a line consisting
of the character x, advancing to the next line.

Material between @(define) and @(end) is the function body.  The define
directive may be followed directly by the @(end) directive, in which case the
function has an empty body.

Functions may be nested within function bodies. Such local functions have
dynamic scope. They are visible in the function body in which they are defined,
and in any functions invoked from that body.

The body of a function is an anonymous block. (See BLOCKS above).

.SS Two Forms of The Horizontal Function

If a horizontal function is defined as the only element of a line,
it may not be followed by additional material. The following
construct is erroneous:

  @(define horiz (x))@foo:@bar@(end)lalala

This kind of definition is actually considered to be in the vertical context,
and like other directives that have special effects and that do not match
anything, it does not consume a line of input. If the above syntax were
allowed, it would mean that the line would not only define a function but also
match "lalala".  This would, in turn, would mean that the @(define)...@(end) is
actually in horizontal mode, and so it matches a span of zero characters within
a line (which means that is would require a line of input to match: a nasty
thing for a non-matching directive to do!) 

A horizontal function can be defined in an actual horizontal context. This
occurs if its is in a line where it is preceded by other material.
For instance:

  X@(define fun)...@(end)Y

This is a query line which must match the text XY. It also defines the function
fun. The main use of this form is for nested horizontal functions:

  @(define fun)@(define local_fun)...@(end)@(end)

.SS Vertical-Horizontal Overloading

A function of the same name may be defined as both vertical and horizontal.
Both functions are available at the same time. Which one is used by
a call is resolved by context. See the section Vertical Versus Horizontal Calls below.

.SS Call Syntax

A function is invoked by compound directive whose first symbol is the name of
that function. Additional elements in the directive are the arguments.
Arguments may be symbols, or other objects like string and character
literals, quasiliterals ore regular expressions.

Example:

  Query:         @(define pair (a b))
                 @a @b
                 @(end)
                 @(pair first second)
                 @(pair "ice" cream)
 
  Data:          one two
                 ice milk
 
  Output:        first="one"
                 second="two"
                 cream="milk"

The first call to the function takes the line "one two". The parameter "a"
takes "one" and parameter b takes "two". These are rebound to the arguments
first and second. The second call to the function binds the a parameter
to the word "ice", and the b is unbound, because the
corresponding argument "cream" is unbound. Thus inside the function, @a
is forced to match "ice". Then a space is matched and @b collects the text
"milk". When the function returns, the unbound "cream" variable gets this value.

If a symbol occurs multiple times in the argument list, it constrains
both parameters to bind to the same value. That is to say, all parameters
which, in the body of the function, bind a value, and which are all derived
from the same argument symbol must bind to the same value. This is settled when
the function terminates, not while it is matching. Example:

  Query:         @(define pair (a b))
                 @a @b
                 @(end)
                 @(pair same same)
 
  Data:          one two

  Output:        [query fails, prints "false"]


.SS Vertical Versus Horizontal Calls

A function call which is the only element of the query line in
which it occurs is ambiguous. It can go either to a vertical
function or to the horizontal one. If both are defined, then
it goes to the vertical one.

Example:

  Query:          @(define which (x))@(bind x "horizontal")@(end)
                  @(define which (x))
                  @(bind x "vertical")
                  @(end)
                  @(which fun)

  Output:         fun="vertical"

Not only does this call go to the vertical function, but
it is in a vertical context.

If only a horizontal function is defined, then that is the one which is called,
even if the call is the only element in the line. This takes place in a
horizontal character-matching context, which requires a line of input which can
be traversed:

Example:

  Query:          @(define which (x))@(bind x "horizontal")@(end)
                  @(which fun)

  Data:           ABC

  Output:         false

The query failed. Why? Because since @(which fun) is in horizontal mode,
it matches characters in a line. Since the function body consists
of @(bind ...) which doesn't match any characters, the function
call requires an empty line to match. The line ABC is not empty,
and so there is a matching failure. The following
example corrects this:

Example:

  Query:          @(define which (x))@(bind x "horizontal")@(end)
                  @(which fun)

  Data:           <empty line>

  Output:         fun="horizontal"

A call made in a clearly horizontal context will ignore
the vertical definition, if any, and require the horizontal function.
In the next example, the call is followed by trailing material,
placing it in a horizontal context. Leading material will
do the same thing:

Example:

  Query:          @(define which (x))@(bind x "horizontal")@(end)
                  @(define which (x))
                  @(bind x "vertical")
                  @(end)
                  @(which fun)B

  Data:           B

  Output:         fun="horizontal"

.SS Nested Functions

Function definitions may appear in a function. Such definitions
are visible in all functions which are invoked from the body
(and not necessarily enclosed in the body). In other words, the
scope is dynamic, not lexical.  Inner definitions shadow outer
definitions. This means that a caller can redirect the function
calls that take place in a callee, by defining local functions
which capture the references.

Example:

  Query:        @(define which)
                @  (fun)
                @(end)
                @(define fun)
                @  (output)
                toplevel fun!
                @  (end)
                @(end)
                @(define callee)
                @  (define fun)
                @    (output)
                local fun!
                @    (end)
                @  (end)
                @  (which)
                @(end)
                @(callee)
                @(which)

   Output:      local fun!
                toplevel fun!

Here, the function "which" is defined which calls "fun".
A toplevel definition of "fun" is introduced which
outputs "toplevel fun!".  The function "callee" provides its own local
definition of "fun" which outputs "local fun!" before calling "which".  When
callee is invoked, it calls @(which), whose @(fun) call is routed to callee's
local definition.  When @(which) is called directly from the top level, its
@(fun) call goes to the toplevel definition.

.SH OUTPUT

.SS Introduction

A
.B txr
query may perform custom output. Output is performed by @(output) clauses,
which may be embedded anywhere in the query, or placed at the end.  Output
occurs as a side effect of producing a part of a query which contains an
@(output) directive, and is executed even if that part of the query ultimately
fails to find a match. Thus output can be useful for debugging.
An output clause specifies that its output goes to a file, pipe, or (by
default) standard output. If any output clause is executed whose destination is
standard output,
.B txr
makes a note of this, and later, just prior to termination, suppresses the
usual printing of the variable bindings or the word false.

.SS The Output Directive

The syntax of the @(output) directive is:

  @(output [ DESTINATION ] { bool-keyword | keyword value }* )
  .
  . one or more output directives or lines
  .
  @(end)

The optional destination is a filename, the special name, - which
redirects to standard output, or a shell command preceded by the ! symbol.
In the first form, the destination may be specified as a variable
which holds text, a string literal or a quasiliteral

The keyword list consists of a mixture of boolean keywords which
do not have an argument, or keywords with arguments.

The following boolean keywords are supported:

.IP :nothrow

The output directive throws an exception if the output destination
cannot be opened, unless the :nothrow keyword is present, in which
case the situation is treated as a match failure.  

Note that since command pipes are processes that report errors
asynchronously, a failing command will not throw an immediate exception that
can be suppressed with :nothrow.   This is for synchronous errors, like
trying to open a destination file, but not having permissions, etc.

.IP :append

This keyword is meaningful for files, specifying append mode: the output is to
be added to the end of the file rather than overwriting the file.

The following value keywords are supported:

.IP :filter

The argument can be a symbol, which specifies a filter to be applied to
the variable substitutions occuring within the output clause.
The argument can also be a list of filter symbols, which specifies
that multiple filters are to be applied, in left to right order.

See the later sections Output Filtering below, and The Deffilter Directive.

.IP :into

The argument of :into is a symbol which denotes a variable.
The output will go into that variable.  If the variable is unbound,
it will be created. Otherwise, its contents are overwritten
unless the :append keyword is used. If :append is used, then
the new content will be appened to the previous content of
the variable, after flattening the content to a list,
as if by the @(flatten) directive.

.SS Output Text

Text in an output clause is not matched against anything, but is output
verbatim to the destination file, device or command pipe.

.SS Output Variables

Variables occurring in an output clause do not match anything, but instead their
contents are output. A variable being output must be a simple string, not a
list. Lists may be output within @(repeat) or @(rep) clauses. A list variable
must be wrapped in as many nestings of these clauses as it has dimensions.  For
instance, a two-dimensional list may be mentioned in output if it is inside a
@(rep) or @(repeat) clause which is itself wrapped inside another @(rep) or
@(repeat) clause.

In an output clause, the @{NAME NUMBER} variable syntax generates fixed-width
field, which contains the variable's text.  The absolute value of the
number specifies the field width. For instance -20 and 20 both specify a field
width of twenty.  If the text is longer than the field, then it overflows the
field. If the text is shorter than the field, then it is left-adjusted within
that field, if the width is specified as a positive number, and right-adjusted
if the width is specified as negative.

An output variable may specify a filter which overrides any filter established
for the output clause. The syntax for this is @(NAME :filter <filterspec>}.
The filter specification syntax is the same as in the output clause.
See Output Filtering below.

.SS The Repeat Directive

The repeat directive is generates repeated text from a ``boilerplate'',
by taking successive elements from lists. The syntax of repeat is
like this:

  @(repeat)
  .
  .
  main clause material, required
  .
  .
  special clauses, optional
  .
  .
  @(end)

Repeat has four types of special clauses, any of which may be
specified with empty contents, or omitted entirely. They are explained
below.

All of the material in the main clause and optional clauses
is examined for the presence of variables.  If none of the variables
hold lists which contain at least one item, then no output is performed,
(unless the repeat specifies an @(empty) clause, see below).
Otherwise, among those variables which contain non-empty lists, repeat finds
the length of the longest list. This length of this list determines the number
of repetitions, R.

If the repeat contains only a main clause, then the lines of this clause is
output R times. Over the first repetition, all of the variables which, outside
of the repeat, contain lists are locally rebound to just their first item. Over
the second repetition, all of the list variables are bound to their second
item, and so forth. Any variables which hold shorter lists than the longest
list eventually end up with empty values over some repetitions.

Example: if the list A holds "1", "2" and "3"; the list B holds "A", "B";
and the variable C holds "X", then

  @(repeat)
  >> @C
  >> @A @B
  @(end)

will produce three repetitions (since there are two lists, the longest
of which has three items). The output is:

  >> X
  >> 1 A
  >> X
  >> 2 B
  >> X
  >> 3

The last line has a trailing space, since it is produced by "@A @B",
where @B has an empty value. Since C is not a list variable, it
produces the same value in each repetition.

The special clauses are:

.IP @(single)
If the repeat produces exactly one repetition, then the contents of this clause
are processed for that one and only repetition, instead of the main clause
or any other clause which would otherwise be processed.

.IP @(first)
The body of this clause specifies an alternative body to be used for the first
repetition, instead of the material from the main clause.

.IP @(last)
The body of this clause is used instead of the main clause for the last
repetition.

.IP @(empty)
If the repeat produces no repetitions, then the body of this clause is output.
If this clause is absent or empty, the repeat produces no output.

.PP
The precedence among the clauses which take an iteration is:
single > first > last > main.   That is if two or more of these clauses
can apply to a repetition, then the leftmost one in this precedence list
applies. For instance, if there is just a single repetition, then any of these
special clause types can apply to that repetition, since it is the only
repetition, as well as the first and last one. In this situation, if
there is a single clause present, then the repetition is processed
using that clause. Otherwise, if there is a first clause present, that
clause is used. Failing that, a last clause applies. Only if none of these
clauses are present will the repetition be processed using the main clause.

.SS Nested Repeats

If a repeat clause encloses variables which holds multidimensional lists,
those lists require additional nesting levels of repeat (or rep).
It is an error to attempt to output a list variable which has not been
decimated into primary elements via a repeat construct.

Suppose that a variable X is two-dimensional (contains a list of lists).  X
must be twice nested in a repeat. The outer repeat will walk over the lists
contained in X. The inner repeat will walk over the elements of each of these
lists.

A nested repeat may be embedded in any of the clauses of a repeat,
not only the main clause.

.SS The Rep Directive

The @(rep) directive is similar to @(repeat), but whereas @(repeat) is line
oriented, @(rep) generates material within a line. It has all the same clauses,
but everything is specified within one line:

  @(rep)... main material ... .... special clauses ...@(end)

More than one @(rep) can occur within a line, mixed with other material.
A @(rep) can be nested within a @(repeat) or within another @(rep).

.SS Repeat and Rep Examples

Example 1: show the list L in parentheses, with spaces between
the elements, or the symbol NIL if the list is empty:

  @(output)
  @(rep)@L @(single)(@L)@(first)(@L @(last)@L)@(empty)NIL@(end)
  @(end)

Here, the @(empty) clause specifies NIL. So if there are no repetitions,
the text NIL is produced. If there is a single item in the list L,
then  @(single)(@L) produces that item between parentheses.  Otherwise
if there are two or more items, the first item is produced with
a leading parenthesis followed by a space by @(first)(@L , and
the last item is produced with a closing parenthesis: @(last)@L).
All items in between are emitted with a trailing space by
the main clause: @(rep)@L .

Example 2: show the list L like Example 1 above, but the empty list is ().

  @(output)
  (@(rep)@L @(last)@L@(end))
  @(end)

This is simpler. The parentheses are part of the text which
surrounds the @(rep) construct, produced unconditionally.
If the list L is empty, then @(rep) produces no output, resulting in ().
If the list L has one or more items, then they are produced with
spaces each one, except the last which has no space.
If the list has exactly one item, then the @(last) applies to it
instead of the main clause: it is produced with no trailing space.

.SS Output Filtering

Often it is necessary to transform the output to preserve its meaning
under the convention of a given data format. For instance, if a piece of
text contains the characters < or >, then if that text is being
substituted into HTML, these should be replaced by &lt; and &gt;.
This is what filtering is for.  Filtering is applied to the contents of output
variables, not to any template text.
.B txr
implements named filters.  Built-in filters are named by keywords,
given below. User-defined filters are possible, however.  See notes on the
deffilter directive below.

Built-in filters:

.IP :to_html
Filter text to HTML, representing special characters using HTML
ampersand sequences. For instance '>' is replaced by '&gt;'.

.IP :from_html
Filter text with HTML codes into text in which the codes are replaced by the
corresponding characters. For instance '&gt;' is replaced by '>'.

.IP :upcase
Convert the 26 lower case letters of the English alphabet to upper case.

.IP :downcase
Convert the 26 upper case letters of the English alphabet to lower case.

Example: to escape HTML characters in all variable substitutions occuring in an
output clause, specify :filter :to_html in the directive:

  @(output :filter :to_html)
  ...
  @(end)

To filter an individual variable, add the syntax to the variable spec:

  @(output)
  @{x :filter :to_html}
  @(end)

Multiple filters can be applied at the same time. For instance:

  @(output)
  @{x :filter (:upcase :to_html)}
  @(end)

This will fold the contents of x to upper case, and then encode any special
characters into HTML. Beware of combinations that do not make sense.
For instance, suppose the original text is HTML, containing codes
like '&quot;'. The compound filter (:upcase :from_html) will not work
because '&quot;' will turn to '&QUOT;' which no longer be recognized
by the :from_html filter, because the entity names in HTML codes
are case-sensitive.

Instead of a filter name, the syntax (fun NAME) can be used. This
denotes that the function called NAME is to be used as a filter.
This is discussed in the next section Function Filters below.

.SS Function Filters

A function can be used as a filter. For this to be possible, the function must
conform to certain rules:

.IP 1.
The function must take two special arguments, which may be followed
by additional arguments.

.IP 2.
When the function is called, the first argument will be bound to a string,
and the second argument will be unbound. The function must produce a 
value by binding it to the second argument. If the filter is to be used
as the final filter in a chain, it must produce a string.

For instance, the following is a valid filter function:

  @(define foo_to_bar (in out)
  @  (next :string in)
  @  (cases)
  foo
  @    (bind out "bar")
  @  (or)
  @    (bind out in)
  @  (end)
  @(end)

This function binds the out parameter to "bar" if the in parameter
is "foo", otherwise it binds the out parameter to a copy of the in parameter.
This is a simple filter.

To use the filter, use the syntax (:fun foo_to_bar) in place of a filter name. 
For instance in the bind directive:

  @(bind "foo" "bar" :lfilt (:fun foo_to_bar))

The above should succeed since the left side is filtered from "foo"
to "bar", so that there is a match.

Of course, function filters can be used in a chain:

  @(output :filter (:downcase (:fun foo_to_bar) :upcase))
  ...
  @(end)

Here is a split function which takes an extra argument.

  @(define split (in out sep))
  @  (next :list in)
  @  (coll)@(maybe)@token@sep@(or)@token@(end)@(end)
  @  (bind out token)
  @(end)

Furthermore, note that it produces a list rather than a string.
This function separates the argument in into tokens according to the
separator text sep.

Here is another function, join, which catenates a list:

  @(define join (in out sep))
  @  (output :into out)
  @  (rep)@in@sep@(last)@in@(end)
  @  (end)
  @(end)

Now here is these two being used in a chain:

  @(bind text "how,are,you")
  @(output :filter (:fun split ",") (:fun join "-"))
  @text
  @(end)

Output:

  how-are-you

When the filter invokes a function, it generates the first two arguments
internally to pass in the input value and capture the output. The remaining
arguments from the (:fun ...) construct are also passed to the function.
Thus the "," and "-" are passed as the sep argument to split and join.

Note that split puts out a list, which join accepts. So the overall filter
chain operates on a string: a string goes into split, and a string comes out of
join.

.SS The Deffilter Directive

The deffilter directive allows a query to define a custom filter, which
can then be used in @(output) clauses to transform substituted data.

This directive's syntax is illustrated in this example:

  Query:        @(deffilter rot13 
                   ("a" "n")
                   ("b" "o")
                   ("c" "p")
                   ("d" "q")
                   ("e" "r")
                   ("f" "s")
                   ("g" "t")
                   ("h" "u")
                   ("i" "v")
                   ("j" "w")
                   ("k" "x")
                   ("l" "y")
                   ("m" "z")
                   ("n" "a")
                   ("o" "b")
                   ("p" "c")
                   ("q" "d")
                   ("r" "e")
                   ("s" "f")
                   ("t" "g")
                   ("u" "h")
                   ("v" "i")
                   ("w" "j")
                   ("x" "k")
                   ("y" "l")
                   ("z" "m"))
                @(collect)
                @line
                @(end)
                @(output :filter rot13)
                @(repeat)
                @line
                @(end)
                @(end)

  Input:        hey there!

  Output:       url gurer!


The deffilter symbol must be followed by the name of the filter to be defined,
followed by tuples of strings. Each tuple specifies one or more texts
which are mapped to a replacement text. For instance, the following specifies
a telephone keypad mapping from upper case letters to digits.
 
  @(deffilter alpha_to_phone ("E" "0") 
                             ("J" "N" "Q" "1")
                             ("R" "W" "X" "2")
                             ("D" "S" "Y" "3")
                             ("F" "T" "4")
                             ("A" "M" "5")
                             ("C" "I" "V" "6")
                             ("B" "K" "U" "7")
                             ("L" "O" "P" "8")
                             ("G" "H" "Z" "9"))

Filtering works using a longest match algorithm. The input is scanned from left
to right, and the longest piece of text is identified at every character
position which matches a string on the left hand side, and that text is
replaced with its associated replacement text. The scanning then continues
at the first character after the matched text.

If none of the strings matches at a given character position, then that
character is passed through the filter untranslated, and the scan continues at
the next character in the input.

Filtering is not in-place but rather instantiates a new text, and so
replacement text is not re-scanned for more replacements.

If a filter definition accidentally contains two or more repetitions of the
same left hand string with different right hand translations, the later ones
take precedence. No warning is issued.


.SS The Filter Directive

The syntax of the filter directive is:

  @(filter FILTER { VAR }+ }

A filter is specified, followed by one or more variables whose values
are filtered and stored back into each variable.

Example: convert a, b, and c to upper case and HTML encode:

  @(filter (:upcase :to_html) a b c)


.SH EXCEPTIONS

.SS Introduction

The exceptions mechanism in
.B txr
is another disciplined form of non-local transfer, in addition to the blocks
mechanism (see BLOCKS above).  Like blocks, exceptions provide a construct
which serves as the target for a dynamic exit.  Both blocks and exceptions
can be used to bail out of deep nesting when some condition occurs.
However, exceptions provide more complexity. Exceptions are useful for
error handling, and txr in fact maps certain error situations to exception
control transfers. However, exceptions are not inherently an error-handling
mechanism; they are a structured dynamic control transfer mechanism, one
of whose applications is error handling.

An exception control transfer (simply called an exception) is always identified
by a symbol, which is its type. Types are organized in a subtype-supertype
hierarchy.  For instance, the file_error exception type is a subtype of the
error type. This means that a file error is a kind of error. An exception
handling block which catches exceptions of type error will catch exceptions of
type file_error, but a block which catches file_error will not catch all
exceptions of type error. A query_error is a kind of error, but not a kind of
file_error. The symbol t is the supertype of every type: every exception type
is considered to be a kind of t.  (Mnemonic: t stands for type, as in any
type).

Exceptions are handled using @(catch) clauses within a @(try) directive.

In addition to being useful for exception handling, the @(try) directive
also provides unwind protection by means of a @(finally) clause,
which specifies query material to be executed unconditionally when
the try clause terminates, no matter how it terminates.

.SS The Try Directive

The general syntax of the try directive is

  @(try)
  ... main clause, required ...
  ... optional catch clauses ...
  ... optional finally clause
  @(end)

A catch clause looks like:

  @(catch TYPE)
  .
  .
  .

and also the this form, equivalent to @(catch (t)):

  @(catch)
  .
  .
  .

which catches all exceptions.

A finally clause looks like:

  @(finally)
  ...
  .
  .

The main clause may not be empty, but the catch and finally may be.

A try clause is surrounded by an implicit anonymous block (see BLOCKS section
above). So for instance, the following is a no-op (an operation with no effect,
other than successful execution):

  @(try)
  @(accept)
  @(end)

The @(accept) causes a successful termination of the implicit anonymous block.
Execution resumes with query lines or directives which follow, if any.

Try clauses and blocks interact. For instance, a block accept from within
a try clause invokes a finally.

  Query:        @(block foo)
                @  (try)
                @    (accept foo)
                @  (finally)
                @     (output)
                bye!
                @     (end)
                @  (end)

  Output:       bye!

How this works: the try block's main clause is @(accept foo). This causes
the enclosing block named foo to terminate, as a successful match.
Since the try is nested within this block, it too must terminate
in order for the block to terminate. But the try has a finally clause,
which executes unconditionally, no matter how the try block
terminates. The finally clause performs some output, which is seen.

.SS The Finally Clause

A try directive can terminate in one of three ways. The main clause
may match successfully, and possibly yield some new variable bindings.
The main clause may fail to match. Or the main clause may be terminated
by a non-local control transfer, like an exception being thrown or a block
return (like the block foo example in the previous section).

No matter how the try clause terminates, the finally clause is processed.

Now, the finally clause is itself a query which binds variables, which leads to
the question: what happens to such variables? What if the finally block fails
as a query? Another question is: what if a finally clause itself initiates a
control transfer?  Answers follow.

Firstly, a finally clause will contribute variable bindings only if the main
clause terminates normally (either as a successful or failed match).
If the main clause successfully matches, then the finally block continues
matching at the next position in the data, and contributes bindings.
If the main clause fails, then the finally block matches at the
same position.

The overall try directive succeeds as a match if either the main clause
or the finally clause succeed. If both fail, then the try directive is
a failed match. The subquery in which it is located fails, et cetera.

Example:

  Query:        @(try)
                @a
                @(finally)
                @b
                @(end)
                @c

  Data:         1
                2
                3

  Output:       a=1
                b=2
                c=3

In this example, the main clause of the try captures line "1" of the data as
variable a, then the finally clause captures "2" as b, and then the
query continues with the @c variable after try block, and captures "3".


Example:

  Query:        @(try)
                hello @a
                @(finally)
                @b
                @(end)
                @c

  Data:         1
                2

  Output:       b=1
                c=2

In this example, the main clause of the try fails to match, because
the input is not prefixed with "hello ". However, the finally clause
matches, binding b to "1". This means that the try block is a successful
match, and so processing continues with @c which captures "2".

When finally clauses are processed during a non-local return,
they have no externally visible effect if they do not bind variables.
However, their execution makes itself known if they perform side effects,
such as output.

A finally clause guards only the main clause and the catch clauses. It does not
guard itself.   Once the finally clause is executing, the try block is no
longer guarded.  This means if a nonlocal transfer, such as a block accept
or exception, is initiated within the finally clause, it will not re-execute
the finally clause. The finally clause is simply abandoned.

The disestablishment of blocks and try clauses is properly interleaved
with the execution of finally clauses. This means that all surrounding
exit points are visible in a finally clause, even if the finally clause
is being invoked as part of a transfer to a distant exit point.
The finally clause can make a control transfer to an exit point which
is more near than the original one, thereby "hijacking" the control
transfer. Also, the anonymous block established by the try directive
is visible in the finally clause.

Example:

  @(try)
  @  (try)
  @    (next "nonexistent-file")
  @  (finally)
  @    (accept)
  @  (end)
  @(catch file_error)
  @  (output)
  file error caught
  @  (end)
  @(end)

In this example, the @(next) directive throws an exception of type file_error,
because the given file does not exist. The exit point for this exception is the
@(catch file_error) clause in the outer-most try block. The inner block is
not eligible because it contains no catch clauses at all. However, the inner
try block has a finally clause, and so during the processing of this
exception which is headed for the @(catch file_error), the finally
clause performs an anonymous accept. The exit point for the accept
is the anonymous block surrounding the inner try.  So the original
transfer to the catch clause is forgotten. The inner try terminates
successfully, and since it constitutes the main clause of the outer try,
that also terminates successfully. The "file error caught" message is
never printed.

.SS Catch Clauses

Catch clauses establish a try block as a potential exit point for
an exception-induced control transfer (called a ``throw'').

A catch clause specifies an optional list of symbols which represent
the exception types which it catches. The catch clause will catch
exceptions which are a subtype of any one of those exception types.

If a try block has more than one catch clause which can match a given
exception, the first one will be invoked.

The exception protection of a try block does not extend over the
catch clauses.  Once a catch clause is being executed, if it throws
an exception, that exception will not re-enter any catch within the
same try block, even if it matches one.

Catches are processed prior to finally.

When a catch is invoked, it is of course understood that the main clause did
not terminate normally, and so the main clause could not have produced any
bindings.

So the success or failure of the try block depends on the behavior of the catch
clause or the finally, if there is one. If either of them succeed, then the try block is considered a successful match.

Example:

  Query:        @(try)
                @  (next "nonexistent-file")
                @  x
                @  (catch file_error)
                @a
                @(finally)
                @b
                @(end)
                @c

  Data:         1
                2
                3

  Output:       a=1
                b=2
                c=3

Here, the try block's main clause is terminated abruptly by a file_error
exception from the @(next) directive.   This is handled by the
catch clause, which binds variable a to the input line "1".
Then the finally clause executes, binding b to "2". The try block
then terminates successfully, and so @c takes "3".

.SS Catch Clauses with Parameters

A catch may have parameters following the type name, like this:

  @(catch pair (a b))

To write a catch-all with parameters, explicitly write the
master supertype t:

  @(catch t (arg ...))

Parameters are useful in conjunction with throw. The built-in
error exceptions generate one argument, which is a string containing
the error message. Using throw, arbitrary parameters can be passed
from the throw site to the catches.

.SS The Throw Directive

The throw directive generates an exception. A type must be specified,
followed by optional arguments. For example,

  @(throw pair "a" `@file.txt`)

throws an exception of type pair, with two arguments, being "a"
and the expansion of the quasiliteral `@file.txt`.

The selection of the target catch is performed purely using the type
name; the parameters are not involved in the selection.

Binding takes place between the arguments given in throw, and the
target catch.

If any catch parameter, for which a throw argument is given, is a bound
variable, it has to be identical to the argument, otherwise the catch fails.
(Control still passes to the catch, but the catch is a failed match).

  Query:        @(bind a "apple")
                @(try)
                @(throw e "banana")
                @(catch e (a))
                @(end)

  Output:       false

If any argument is an unbound variable, the corresponding parameter
in the catch is left alone: if it is an unbound variable, it remains
unbound, and if it is bound, it stays as is.

  Query:        @(try)
                @(trow e "honda" unbound)
                @(catch e (car1 car2))
                @car1 @car2
                @(end)

  Data:         honda toyota

  Output:       car1="honda"
                car2="toyota"

If a catch has fewer parameters than there are throw arguments,
the excess arguments are ignored.

  Query:         @(try)
                 @(throw e "banana" "apple" "pear")
                 @(catch e (fruit))
                 @(end)

  Output:        fruit="banana"

If a catch has more parameters than there are throw arguments, the excess
parameters are left alone. They may be bound or unbound variables.

  Query:        @(try)
                @(trow e "honda")
                @(catch e (car1 car2))
                @car1 @car2
                @(end)

  Data:         honda toyota

  Output:       car1="honda"
                car2="toyota"

A throw argument passing a value to a catch parameter which is unbound causes
that parameter to be bound to that value.

Throw arguments are evaluated in the context of the throw, and the bindings
which are available there. Consideration of what parameters are bound
is done in the context of the catch.

  Query:        @(bind c "c")
                @(try)
                @(forget c)
                @(bind (a c) ("a" "lc"))
                @(throw e a c)
                @(catch e (b a))
                @(end)

   Output:      c="c"
                b="a"
                a="lc"

In the above example, c has a toplevel binding to the string "c",
but is then unbound within the try construct, and rebound to the value "c".
Since the try construct is terminated by a throw, these modifications of the
binding environment are discarded. Hence, at the end of the query, variable
c ends up bound to the original value "c".  The throw still takes place
within the scope of the bindings set up by the try clause, so the values of
a and c that are thrown are "a" and "lc".  However, at the catch site, variable
a does not have a binding.  At that point, the binding to "a" established in
the try has disappeared already. Being unbound, the catch parameter a can take
whatever value the corresponding throw argument provides, so it ends up with
"lc".

.SS The Defex Directive

The defex directive allows the query writer to invent custom exception types,
which are arranged in a type hierarchy (meaning that some exception types are
considered subtypes of other types).

Subtyping means that if an exception type B is a subtype of A, then every
exception of type B is also considered to be of type A. So a catch for type A
will also catch exceptions of type B.  Every type is a supertype of itself: an
A is a kind of A. This of course implies that ever type is a subtype of itself
also.  Furthermore, every type is a subtype of the type t, which has no
supertype other than itself. Type nil is is a subtype of every type, including
itself.  The subtyping relationship is transitive also. If A is a subtype
of B, and B is a subtype of C, then A is a subtype of C.

Defex may be invoked with no arguments, in which case it does nothing:

  @(defex)

It may be invoked with one argument, which must be a symbol. This introduces a
new exception type. Strictly speaking, such an introduction is not necessary;
any symbol may be used as an exception type without being introduced by
@(defex):

  @(defex a)

Therefore, this also does nothing, other than document the intent to use
a as an exception.

If two or more argument symbols are given, the symbols are all introduced as
types, engaged in a subtype-supertype relationship from left to right.
That is to say, the first (leftmost) symbol is a subtype of the next one,
which is a subtype of the next one and so on. The last symbol, if it
had not been already defined as a subtype of some type, becomes a
direct subtype of the master supertype t. Example:

  @(defex d e)
  @(defex a b c d)

The fist directive defines d as a subtype of e, and e as a subtype of t.
The second defines a as a subtype of b, b as a subtype of c, and
c as a subtype of d, which is already defined as a subtype of e.
Thus a is now a subtype of e. It should be obvious that the above
could be condensed to:

  @(defex a b c d e)

Example:

  Query:        @(defex gorilla ape primate)
                @(defex monkey primate)
                @(defex human primate)
                @(collect)
                @(try)
                @(skip)
                @(cases)
                gorilla @name
                @(throw gorilla name)
                @(or)
                monkey @name
                @(throw monkey name)
                @(or)
                human @name
                @(throw human name)
                @(end)@#cases
                @(catch primate (name))
                @kind @name
                @(output)
                we have a primate @name of kind @kind
                @(end)@#output
                @(end)@#try
                @(end)@#collect


  Input:        gorilla joe
                human bob
                monkey alice

  Output:       we have a primate joe of kind gorilla
                we have a primate bob of kind human
                we have a primate alice of kind monkey

Exception types have a pervasive scope. Once a type relationship is introduced,
it is visible everywhere. Moreover, the defex directive is destructive,
meaning that the supertype of a type can be redefined. This is necessary so
that something like the following works right.

  @(defex gorilla ape)
  @(defex ape primate)

These directives are evaluated in sequence. So after the first one, the ape
type has the type t as its immediate supertype.  But in the second directive,
ape appears again, and is assigned the primate supertype, while retaining
gorilla as a subtype.  This situation could instead be diagnosed as an
error, forcing the programmer to reorder the statements, but instead
txr obliges. However, there are limitations.  It is an error to define a
subtype-supertype relationship between two types if they are already connected
by such a relationship, directly or transitively. So the following
definitions are in error:

  @(defex a b)
  @(defex b c)
  @(defex a c)@# error: a is already a subtype of c, through b

  @(defex x y)
  @(defex y x)@# error: circularity; y is already a supertype of x.

.SH NOTES ON EXOTIC REGULAR EXPRESSIONS

Users familiar with regular expressions may not be familiar with the complement
and intersection operators, which are often absent from text processing tools
that support regular expressions.  The following remarks are offered in hope
that they are of some use.

.SS Equivalence to Sets

Regexp intersection is not essential; it may be obtained from complement and
union as follows, since De Morgan's law applies to regular expression algebra:
(R1)&(R2) = ~(~(R1)|~(R2)). (The complement of the union of the complements of
R1 and R2 constitutes the intersection.) This law works because the regular
expression operators denote set operations in a straightforward way. A regular
expression denotes a set of strings (a potentially infinite one) in a condensed
way. The union of two regular expressions R1 | R2 denotes the union of the set
of texts denoted by R1 and that denoted by R2. Similarly R1 & R2 denotes a set
intersection, and ~R denotes a set complement. Thus algebraic laws
that apply to set operations apply to regular expressions. It's useful to keep
in mind this relationship between regular expressions and sets in understanding
intersection and complement.

Given a finite set of strings, like the set { "abc", "def" }, which corresponds
to the regular expression (abc|def), the complement is the set which contains
an infinite number of strings: it consists of all possible strings except "abc"
and "def". It includes the empty string, all strings of length 1, all strings
of length 2, all strings of length 3 other than "abc" and "def", all strings of
length 4, etc.  This means that a "harmless looking" expression like ~(abc|def)
can actually match arbitrarily long inputs.

.SS Set Difference

How about matching only three-character-long strings other than "abc" or "def"?
To express this, regex intersection can be used: these strings are the
intersection of the set of all three-character strings, and the set of all
strings which are not "abc" or "def". The straightforward set-based reasoning
leads us to this: ...&~(abc|def).   This A&~B idiom is also called set
difference, sometimes notated with a minus sign: A-B (which is not
supported in 
.B txr
regular expression syntax).  Elements which are in the set A, but not B, are
those elements which are in the intersection of A with the complement of B.
This is similar to the arithmetic rule A - B = A + -B: subtraction is
equivalent to addition of the additive inverse. Set difference is a useful
tool: it enables us to write a positive match which captures a more general set
than what is intended (but one whose regular expression is far simpler
than a positive match for the exact set we want), then we can
intersect this over-generalized set with the complemented set of
another regular expression which matches the particulars that we wish excluded.

.SS Expressivity versus Power

It turns out that regular expressions which do not make use of the
complement or intersection operators are just as powerful as expressions
that do. That is to say, with or without these operators, regular expressions
can match the same sets of strings (all regular languages).  This means that
for a given regular expression which uses intersection and complement, it is
possible to find a regular expression which doesn't use these operators, yet
matches the same set of strings.  But, though they exist, such equivalent
regular expressions are often much more complicated, which makes them difficult
to design.  Such expressions do not necessarily
. B express
what it is they match; they merely capture the equivalent set. They
perform a job, without making it obvious what it is they do.  The use of
complement and intersection leads to natural ways of expressing many kinds of
matching sets, which not only demonstrate the power to carry out an operation,
but also easily express the concept.

.SS Example: Matching C Language Comments

For instance, using complement, we can write a straightforward regular
expression which matches C language comments. A C language
comment is the digraph /*, followed by any string which does not contain the
closing sequence */, followed by that closing sequence.
Examples of valid comments are /**/, /* abc */ or /***/. But C
comments do not nest (cannot contain comments), so that 
/* /* nested */ */ actually consists of the comment /* /* nested */,
which is followed by the trailing junk */.
Our simple characterization of interior part of a C comment as a string
which does not contain the terminating digraph makes use of the
complement, and can be expressed using the complemented regular expression like
this: (~.*[*][/].*).  That is to say, strings which contain */ are matched by
the expression .*[*][/].*: zero or more arbitrary characters, followed by
*/, followed by zero or more arbitrary characters. Therefore, the complement of this expression matches all other strings: those which do not contain */.
These strings up the inside of a C comment between the /* and */.

The equivalent simple regex is quite a bit more complicated.
Without complement, we must somehow write a positive match for all strings such
that we avoid matching */. Obviously, sequences of characters other than *
are included: [^*]*.   Occurrences of * are also allowed, but only if followed
by something other than a slash, so let's include this via union:

  ([^*]|[*][^/])*. 

Alas, already, we have a bug in this expression. The
subexpression [*][^/] can match "**", since a * is not a /.  If the next
character in the input is /, we missed a comment close.  To fix the problem we
revise to this: 

  ([^*]|[*][^*/])*

(The interior of a C language comment is a any
mixture of zero or more non-asterisks, or digraphs consisting of an asterisk
followed by something other than a slash or another asterisk).  Oops, now we
have a problem again. What if two asterisks occur in a comment?  They are not
matched by [^*], and they are not matched by [*][^*/]. Actually, our regex must
not simply match asterisk-non-asterisk digraphs, but rather sequences of one or
more asterisks followed by a non-asterisk: 

  ([^*]|[*]*[^*/])*

This is still not right, because, for instance, it fails to match the interior
of a comment which is terminated by asterisks, including the simple test cases
where the comment interior is nothing but asterisks.  We have no provision in
our expression for this case; the expression requires all runs of asterisks to
be followed by something which is not a slash or asterisk. The way to fix this
is to add on a subexpression which optionally matches a run of zero or more
interior asterisks before the comment close: 

  ([^*]|[*]*[^*/])*[*]*

Thus our the semi-final regular expression is

  [/][*]([^*]|[*]*[^*/])*[*]*[*][/]

(A C comment is an interior string enclosed in /* */, where this interior part
consists of a mixture of non-asterisk characters, as well as runs of asterisk
characters which are terminated by a character other than a slash, except for
possibly one rightmost run of asterisks which extends to the end of the
interior, touching the comment close. Phew!) One final simplification is
possible: the tail part [*]*[*][/]  can be reduced to [*]+[/] such that the
final run of asterisks is regarded as part of an extended comment terminator
which consists of one or more asterisks followed by a slash.   The regular
expression works, but it's cryptic; to someone who has not developed it, it
isn't obvious what it is intended to match.  Working out complemented matching
without complement support from the language is not impossible, but it may be
difficult and error-prone, possibly requiring multiple iterations of
trial-and-error development involving numerous test cases, resulting in an
expression that doesn't have a straightforward relationship to the original
idea.

.SS The Non-Greedy Operator

The non-greedy operator % is actually defined in terms of a set difference,
which is in turn based on intersection and complement. The uninteresting case
(R%) where the right operand is empty reduces to (R*): if there is no trailing
context, the non-greedy operator matches R as far as possible, possibly to the
end of the input, exactly like the greedy Kleene.  The interesting case (R%T)
is defined as a "syntactic sugar" for the equivalent expression
((R*)&(~.*(T&.+).*))T   which means: match the longest string which is matched
by R*, but which does not contain a non-empty match for T; then, match T.  This
is a useful and expressive notation. With it, we can write the regular
expression for matching C language comments simply like this: [/][*].%[*][/]
(match the opening sequence /*, then match a sequence of zero or more
characters non-greedily, and then the closing sequence */.  With the non-greedy
operator, we don't have to think about the interior of the comment as set of
strings which excludes */.  Though the non-greedy operator appears expressive,
its apparent simplicity may be deceptive.  It looks as if it works "magically"
by itself; "somehow" this .% "knows" only to consume enough characters so that
it doesn't swallow an occurrence of the trailing context. Care must be taken
that the trailing context passed to the operator really is the correct text
that should be excluded by the non-greedy match. For instance, take the
expression .%abc.  If you intend the trailing context to be merely a, you must
be careful to write (.%a)bc. Otherwise the trailing context is abc, and this
means that the .% match will consume the longest string that does not contain
"abc", when in fact what was intended was to consume the longest string that
does not contain a. The change in behavior of the % operator upon modifying the
trailing context is not as intuitive as that of the * operator, because the
trailing context is deeply involved in its logic.  For single-character
trailing contexts, it may be a good idea to use a complemented character class
instead. That is to say, rather than (.%a)bc, consider [^a]*bc. The set of
strings which don't contain the character a is adequately expressed by [^a]*.

.SH NOTES ON FALSE

The reason for printing the word
.IR false
on standard output when
a query doesn't match, in addition to returning a failed termination
status, is that the output of
.B txr
may be collected by a shell script, by the application of eval to command
substitution syntax. Printing
.IR false
will cause eval to evaluate the
.IR false
command, and thus failed status will propagate from the eval
itself.   The eval command conceals the termination status of a
program run via command substitution.  That is to say, if a program
fails, without producing output, its output is substituted into the eval
command which then succeeds, masking the failure of the program. For example:

  eval "$(false)"

appears successful: the false utility indicates a failed status, but
produces no output. Eval evaluates an empty script and reports success;
the failed status of the false program is forgotten.
Note the difference between the above and this:

  eval "$(echo false)"

This command has a failed status. The echo prints the word false and succeeds;
this false word is then evaluated as a script, and thus interpreted as the
false command which fails. This failure
.B is
propagated as the result of the eval
command.