In the following log, modification dates are listed using the European
convention in which the day comes before the month (ie. DD/MM/YYYY).
The most recent modifications are listed first.

             Dominyk Tiller reported that libtecla didn't compile on
             Mac OS X, due to libgcc.a being cited in the makefile
             without any path. This turned out to be because Gnu
             autoconf claims that the clang compiler is gcc, and clang
             has a -print-libgcc-file-name, but this only prints
             libgcc.a. I have modified the configure script to check
             whether the path returned by -print-libgcc-file-name
             actually exists and only use it if it does.

27/10/2014 Jon Szymaniak (documented here by
             Use $(AR) instead of plain ar, so that cross-compilation
             uses the correct ar program when cross-compiling.

             Add $(TARGETS) dependency to building the demo programs
             and the enhance program. When using parallel compilation
             this is needed to ensure that the library is compiled
             before the demos.

             Jonathan Niehof reported that libtecla wouldn't compile if
             there were spaces in LDFLAGS and pointed out that there should
             be quotes around $(LDFLAGS) in This also applied
             to a few other variables cited in the same way.

	     I had incorrectly assumed that system-V pseudo-terminal
	     allocation and system-V streams terminals always went
	     together.  However system-V pseudo terminal allocation is
	     now part of UNIX98, and this has been adopted into many BSD
	     style operating systems, without the use of system-V
	     streams. On such systems the lack of system-V streams IOCTL
             opcodes prevented system-V pseudo-terminal allocation being
             used. This was hidden under Linux until recently, because
             it had a stropts.h file, which made it appear as though
             Linux supported system-V streams terminals.

	     I have now created separate configuration tests and options
	     in the configure script for system-V terminal allocation
	     and system-V streams. On systems that only have the former,
             the latter won't be used.

	     When an initial input line was presented to gl_get_line()
	     for editing, the new input line was incorrectly appended to
	     the previous input line, instead of replacing it.

10/01/2004 Derek Jones (documented here by
             Derek discovered that the function that computes the
             width of the prompt, was not correctly skipping over 3 of
             the 6 possible prompt-formatting directives. Thus, when
             the %f,%p or %v prompt-formatting directives were used,
             the width of the prompt was incorrectly calculated. The
             fix was to copy the list of directives from
             gl_display_prompt(). I have also added a comment to
             gl_display_prompt(), to warn anybody who adds or removes
             formatting directives there, to also do the same to

31/10/2004 (problem reported by Godfrey van der Linden) 
             The gl_event_handler() function had the endif of a
             conditional compilation clause in the wrong place. This
             only upset the compiler on unusual systems that don't
             have select(). The problem was seen under Mac OS X, due
             to the configuration problem in 1.6.0 that caused the
             configure script to mistakenly report that select wasn't

31/10/2004 (info provided by Ivan Rayner)
             Ivan reported that under IRIX 6.5 it is necessary to add
             -D_XOPEN_SOURCE=500 to the compiler flags, when compiling
             the reentrant version of the library. Thus, whereas
             previously I hardwired the value of DEFINES_R in
   , I have now made this a variable in the
             configure script, which is augmented with the above
             addition, within an IRIX-specific switch clause.

             Also apparently configure leaves the RANLIB variable
             blank, instead of setting it to ":", so I have now
             explicitly set this to ":", within the new IRIX clause of
             the configure script.

31/10/2004 (info provided by Ivan Rayner)
             Under IRIX, the compiler warned that gl_read_unmasked()
             was returning an int, which was then being assigned to an
             enumeration type. This is techically fine, but it
             highlighted the fact that I had meant to declare
             gl_read_unmasked() to directly return the enumerated
             type. I have now done so.

             Users can now turn off interactive command-line editing
             by setting the TERM environment variable to the word "dumb".

18/07/2004 (problem noted by Michael MacFaden)
             Calling gl_terminal_size() on a system without support
             for SIGWINCH caused a divide-by-zero error in an unintended
             call to gl_erase_line(), because gl_update_size() was
             incorrectly being called to query the terminal size,
             instead of gl_query_size().

18/07/2004 Padraig Brady  (documented here by
             The suspend and termination signal-handlers installed by
             gl_tty_signals(), were being installed swapped.

03/06/2004 Mike Meaney  (documented here by
             Mike pointed out the fact that the curses setupterm()
             function is actually documented to exit the application
             if an error occurs while its optional errret argument is
             NULL. I hadn't noticed this, and because I didn't need
             the extra information returned in the errret argument, I
             was passing it a NULL. As suggested by Mike, I now pass
             this argument a pointer to a dummy errret variable.

23/05/2004 (problem noted by John Beck)
             Some of the prototypes of functions and types documented
             by the cpl_complete_word man page, weren't listed in the
             Synopsis section of this man page. They are now listed
           getline.c man/func/
             I have now added support for calling gl_normal_io() from
             any callback functions that the application installs by
             calling either gl_inactivity_timeout(), or gl_watch_fd().
             Previously, if one of these callback functions called
             gl_normal_io(), then after returning to gl_get_line(),
             gl_get_line() would incorrectly assume that the terminal
             was still in raw I/O mode. Now, gl_get_line() checks to
             see if gl_normal_io() was called by the callback, and
             if so, calls _gl_raw_io() to reinstate raw I/O mode.

             On Mac OS X the code that the configure script used to
             check for select() failed due to missing symbols in
             sys/select.h. Moving the inclusion of sys/select.h to
             after the inclusion of sys/time.h, sys/types.h and
             sys/unistd.h fixed this.

           getline.c man/func/
             If the line buffer returned by one call to gl_get_line()
             was passed as the start_line argument of the next call to
             gl_get_line(), then instead of the just-entered line
             being presented back to the user for further editing, the
             start_line argument was effectively ignored, because the
             line buffer whose pointer was being passed back, was
             being cleared before the start_line pointer was examined.
             This appears to have been a case of me incorrectly
             thinking that I had forgotten to initialize gl->line[]
             and gl->ntotal in the gl_reset_input_line() function, and
             then "fixing" this supposed omission. Removing this
             erroneous fix, restored things to how they were meant to
             be. To make it unlikely that I will make the same mistake
             again, I have renamed the function from
             gl_reset_input_line() to gl_reset_editor(), to stop it
             looking as though it is meant to reset the contents of
             the input line (that is what gl_truncate_buffer() is
             for), explicitly stated that it doesn't clear the input
             line, in the header comments of the function, and added a
             prominent warning comment in the body of the function.

             Also, since support for passing back the returned line
             pointer via the start_line argument of the next call to
             gl_get_line(), wasn't documented in the man page, but was
             meant to be supported, and definitely used to work, I
             have now amended the man page documentation of
             gl_get_line() to explicitly state that this feature is
             officially supported.

2?/04/2004 Released 1.6.0

22/04/2004  (Fixed a bug reported by John Beck)
             When an error, signal, or other abnormal event aborted
             gl_get_line(), the cleanup code that restored the
             terminal to a sane state, also overwrote the value of
             errno that was associated with the aborting event. An
             I/O error occurring in the cleanup code would have also
             overwritten the value to be returned by
             gl_return_status(), and thus remove any possibility of
             the caller finding out what really caused gl_get_line()
             to abort. I have now written a new internal function
             called, gl_record_status(), which records the completion
             status to be returned by gl_return_status(), and the
             value to assign to errno just before gl_get_line()
             returns. This is called wherever code detects conditions
             that require gl_get_line() to return early. The function
             ensures that once an abnormal completion status has been
             recorded for return, subsequent completions statuses
             aren't recorded. This ensures that the caller sees the
             original cause of the abnormal return, rather than any
             error that occurs during cleaning up from this before

             If an application's callback called gl_read_char() after
             calling gl_normal_io(), it would inappropriately
             redisplay the input line, when it called _gl_raw_io() to
             temporarily switch the terminal back into raw mode.

             To fix this, _gl_raw_io() now takes a new 'redisplay'
             argument, which specifies whether or not to queue a
             redisplay of the input line. I also created a new
             gl->postpone flag, which is set by gl_normal_io(), and
             cleared by _gl_raw_io() (when its redisplay argument is
             true). When this flag is set, gl_flush_output() ignores
             queued redisplays, as it generally should between calls
             to gl_normal_io() and gl_raw_io(). Thus its effect is to
             postpone redisplays while line editing is suspended.

           history.c man/misc/
             History searches can now include the globbing operators
             *, ?, []. When a search prefix is found to have at least
             one of these characters, then only history lines that
             completely match that pattern are returned.

11/04/2004  (issue raised by Mark Coiley)
           getline.c ioutil.c
             There appears to be a bug in Solaris's terminal I/O.
             When the terminal file descriptor is placed in
             non-blocking I/O mode, and the terminal is switched from
             canonical to raw mode, characters that were previously
             entered in canonical I/O mode don't become available to
             be read until the user types one character more. Select()
             incorrectly says that there are no characters available,
             and read() returns EAGAIN. This is only a problem for
             gl_get_line() when gl_get_line() is in non-blocking
             server I/O mode, so most users won't have experienced any
             problems with this.

             The only way that I have found to get read() to return
             the characters, without the user first having to type
             another character, is to turn off non-blocking I/O before
             calling read(). Select() still claims that there are no
             characters available to be read, but read happily returns
             them anyway. Fortunately, one can perform non-blocking
             terminal reads without setting the non-blocking I/O flag
             of the file descriptor, simply by setting the VTIME
             terminal attribute to zero (which I already was
             doing). Thus, when in non-blocking server I/O, I now turn
             off the non-blocking I/O flag, attempt to read one
             character and only if this fails, do I then call the
             select() based event handler to implement any configured
             non-zero timeout, before attempting the read again. Of
             course the non-blocking I/O flag is still needed for
             writing, so I only turn it off temporarily while reading.

25/03/2004  (bug reported by Gregory Harris)
             It appears that when in February, I patched
             to add abolute paths to the install-sh shell-script,
             I accidentally replaced install-sh with I
             corrected the name in the Makefile.

25/03/2004 Gregory Harris  (documented here by mcs)
             Greg added the configuration parameters needed to build
             the shared version of the libtecla library under FreeBSD.

           getline.c libtecla.h man/func/
             I wrote a public function called gl_read_char(). Unlike
             gl_query_char(), this function neither prompts the user
             for input, nor displays the character that was entered.
             In fact it doesn't write anything to the terminal, and
             takes pains not to disturb any incompletely entered
             input line, and can safely be called from application
             callback functions.

           getline.c libtecla.h man/func/
             I wrote a public function called gl_query_char(), which
             prompts the user and awaits a single-character reply,
             without the user having to hit return.

23/02/2004 (bug reported by Gregory Harris)
  configure getline.c enhance.c demo3.c
             The configure script now checks for the sys/select.h
             header file, and arranges for a C macro called
             HAVE_SYS_SELECT_H to be set if it exists. Thus the files
             that use select() now use this macro to conditionally
             include sys/select.h where available. Apparently this
             header is required under FreeBSD 5.1.

           getline.c libtecla.h man/func/
             I wrote two new public functions, gl_append_history() and
             gl_automatic_history(). Together these allow the
             application to take over the responsibility of adding
             lines to the history list from gl_get_line(). I then
             documented their functionality in the gl_get_line man
           Version 1.6.0
             I incremented the minor version number of the library, to
             comply with the requirement to do so when additions are
             made to the public interface. See for
             I added a new 1.6.0 group for the new minor version, and
             added the above pair of functions to it.

15/02/2004 (fixes a bug reported by Satya Sahoo)
             Calling gl_load_history() multiple times, eventually led
             to a segmentation fault. This was due to the head of the
             list of unused history string segments not getting
             reset when the history buffer was cleared. While
             debugging this problem I also noticed that the history
             resizing function was way too complicated to verify, so
             after fixing the above bug, I heavily simplified the
             history resizing function, trading off a small reduction
             in memory efficiency, for greatly improved clarity, and
             thus made it much more verifiable and maintainable.

14/02/2004 (fixes a bug reported by Tim Burress).
             If gl_change_terminal() was first used to tell
             gl_get_line to read input from a file, then called later
             to tell it to read subsequent input from a terminal, no
             prompt would be displayed for the first line of
             interactive input. The problem was that on reaching the
             end of the input file, gl_get_line() should have called
             gl_abandon_line(), to tell the next call to gl_get_line()
             to start inputting a new line from scratch. I have added
             this now.

14/02/2004 Krister Walfridsson (documented here by
             Krister noticed that I had failed to put $(srcdir)/ in front
             of some invokations of I have remedied this.
           config.guess config.sub
             I hadn't updated these for a long time, so apparently they
             didn't recognise the BSD system that Krister was using.
             I have now updated them to the versions that come with

             When parsing key-binding specifications, backslash escaped
             characters following ^ characters were not being expanded.
             Thus ^\\ got interpretted as a control-\ character followed
             by a \ character, rather than simply as a control-\

           cplfile.c cplmatch.c demo2.c demo3.c demo.c direader.c
           expand.c getline.c history.c homedir.c pathutil.c pcache.c
  configure INSTALL
             The configuration script now takes a
             "--without-file-system" argument. This is primarily for
             intended for embedded systems that either don't have
             filesystems, or where the file-system code in libtecla is
             unwanted bloat. It sets the WITHOUT_FILE_SYSTEM
             macro. This removes all code related to filesystem
             access, including the entire public file-expansion,
             file-completion and path-lookup facilities. Note that the
             general word completion facility is still included, but
             without the normally bundled file completion
             callback. Actually the callback is still there, but it
             reports no completions, regardless of what string you ask
             it to complete.

             This option is described in the INSTALL document.

           getline.c configure INSTALL
             The configuration script now takes a
             "--without-file-actions" argument. This allows an
             application author/installer to prevent users of
             gl_get_line() from accessing the filesystem from the
             builtin actions of gl_get_line(). It defines a macro
             called HIDE_FILE_SYSTEM. This causes the
             "expand-filename", "read-from-file", "read-init-files",
             and "list-glob" action functions to be completely
             removed. It also changes the default behavior of actions
             such as "complete-word" and "list-or-eof" to show no
             completions, instead of the normal default of showing
             filename completions.

             This option is described in the INSTALL document.

           getline.c man/func/
             In case an application's customized completion handler
             needs to write to the terminal for some unforseen reason,
             there needs to be a way for the it to cleanly suspend raw
             line editing, before writing to the terminal, and the
             caller then needs to be aware that it may need to
             resurrect the input line when the callback returns. I
             have now arranged that the completion callback functions
             can call the gl_normal_io() function for this purpose,
             and documented this in the gl_get_line() man page.

11/01/2004  (In response to a bug report by Satya Sahoo)
             The gl_configure_getline() function makes a malloc'd copy
             of the names of the configuration files that it is asked
             to read. Before the bug fix, if the application made one
             or more calls to this function, the memory allocated by
             the final call that it made before calling del_GetLine(),
             wasn't being freed. Note that memory allocated in all but
             the final call was being correctly freed, so the maximum
             extent of the memory leak was the length of the file
             name(s) passed in the final call to
             gl_configure_getline(), and an application that didn't
             call gl_configure_getline() didn't suffer any leak.

             Ellen tested the history fix that I reported below, and
             pointed out that it still had a problem. This turned out
             to be because getline.c was making some incorrect
             assumptions about the new behavior of history.c. This
             problem and the previous one both revolved around how
             search prefixes were stored and discarded, so I have now
             re-written this part of the code. Previously the search
             prefix was retained by looking for a line with that
             prefix, and keeping a pointer to that line. This saved
             memory, compared to storing a separate copy of the
             prefix, but it led to all kinds of hairy
             interdependencies, so I have now changed the code to keep
             a separate copy of search prefixes. To keep the memory
             requirements constant, the search prefix is stored in the
             history buffer, like normal history lines, but not
             referenced by the time-ordered history list. The prefix
             can now be kept around indefinitely, until a new search
             prefix is specified, regardless of changes to the
             archived lines in the history buffer. This is actually
             necessary to make the vi-mode re-search actions work
             correctly. In particular, I no longer discard the search
             prefix whenever a history search session ends. Also,
             rather than have getline.c keep its own record of when a
             history session is in progress, it now consults
             history.c, so that failed assumptions can't cause the
             kind of discrepancy that occurred before. For this to
             work, getline.c now explicitly tells history.c to cancel
             search sessions whenever it executes any non-history

14/12/2003 (bug reported by Ellen Oschmann)
             If one searched backwards for a prefix, then returned to
             the original line, changed that line, then started
             another backwards prefix search, getline incorrectly
             discarded the new search prefix in the process of
             throwing away its cached copy of the previous pre-search
             input line. In other words getline was belatedly
             cancelling a previous search, after a new search had
             already partially begun, and thus messed up the new
             search. The obvious fix was to arrange for the current
             search to be cancelled whenever the history pointer
             returns to its starting point, rather than waiting for
             the next search to begin from there.

             _glh_recall_line() was returning the last line in the
             history buffer instead of the line requested by the
             caller. This only affected the obscure "repeat-history"
             action-function, which probably isn't used by anybody.

09/12/2003 Version 1.5.0 released.

             When the home directory of the login user is requested,
             see if the HOME environment variable exists, and if so
             return its value, rather than looking up the user's home
             directory in the password file. This seems to be the
             convention adopted by other unix programs that perform
             tilde expansion, and it works around a strange problem,
             where a third-party libtecla program, statically compiled
             under an old version of RedHat, unexpectedly complained
             that getpwd() returned an error when the program was run
             under RedHat 9.

           getline.c libtecla.h man/func/
             It is now possible for an application to register
             external functions as action functions. These actions are
             initially bound to specified key-sequences, but if they
             are registered before the user's configuration file is
             loaded, they can also be re-bound by the user to
             different key-sequences. The function used to register a
             new action, is called gl_register_action().  Action
             functions are passed a readonly copy of the input line
             and the cursor position. They can display text to the
             terminal, or perform other operations on the application
             environment. Currently, they can't edit the input line or
             move the cursor. This will require the future addition of
             functions to queue the invokation of the built-in action

             I modified gl_update_buffer() to ensure that the cursor
             stays within the input line after external line
             modifications, and to queue a redisplay of the
             potentially modified input line.

  configure Makefile.stub INSTALL
             By specifying --without-man-pages or --with-man-pages=no
             as command-line arguments to the configure script, it is
             now possible to have the configure script skip the
             man-page preprocessing step, and arrange for the man-page
             installation targets in the Makefile to do nothing. This
             option is designed for people who embed libtecla within
             other packages. It is also used by Makefile.stub when
             the distclean target is specified.

             The previous workaround for recent versions of gcc
             placing /usr/local/include at the start of the system
             inlcude-file search path, broke something else.  The fix
             placed /usr/include before gcc's include area, which
             meant that gcc's modified version of stdarg.h was being
             ignored in deference to the version in /usr/include. I
             have changed the fix to have gcc report the search path,
             then have awk add options to CFLAGS to reorder this path,
             plaing /usr/local/include at the end.

             Also, under Solaris 9, including term.h without first
             including curses.h results in complaints about undefined
             symbols, such as bool. As a result the configure script's
             test for term.h was failing. I have now modified it to
             include curses.h in the test code that it uses to check
             for term.h. In the process I also improved the tests for
             curses.h and term.h to prevent an ncurses version of
             term.h from being used with the system-default version of

  direader.c homedir.c
             On some systems (eg. linux) the _POSIX_C_SOURCE
             feature-test macro is set by system headers, rather than
             being an option set by a project's Makefile at
             compilation time.  In software, such as tecla, where the
             definition of this macro is used as an indication of
             whether to use the non-reentrant or reentrant versions of
             system functions, this means that the reentrant functions
             are always used, regardless of whether this macro is set
             or not by the project Makefile. Thus, on such systems the
             reentrant and non-reentrant versions of the tecla library
             are essentially identical. This has a couple of
             drawbacks.  First, since thread-safe functions for
             traversing the password file don't exist, the supposedly
             non-reentrant version of the tecla library can't support
             ambiguous tab-completion of usernames in ~username/
             constructions. Secondly, on some systems the use of
             reentrant system functions dictates the use of a shared
             library that isn't needed for the non-reentrant
             functions, thus making it more difficult to distribute
             binary versions of the library.

             To remedy this situation I have modified the DEFINES_R
             variable in to arrange for the compiler to
             define a C macro called PREFER_REENTRANT when it is
             compiling the reentrant version of the tecla library.
             This macro is now used in the source code to determine
             when to require reentrant code. Whithin the source code,
             wherever a potentially non-reentrant interface is used,
             the existance of both this macro and a suitably valued
             _POSIX_C_SOURCE macro, are tested for to see if a
             reentrant alternative to the problem code should be used.

             I changed the way that redisplays are requested and
             performed.  Redisplays are now queued by calling
             gl_queue_redisplay(), and subsequently performed by
             gl_flush_output(), when the queue of already pending
             output has been completely dispatched. This was necessary
             to prevent event handlers from filling up the output
             queue with redisplays, and it also simplifies a number of
             things. In the process I removed the gl_queue_display()
             function. I also wrote a gl_line_erased() function, which
             is now called by all functions that erase the input
             line. I also split the gl_abandon_line() function into
             public and private callable parts, and used the private
             version internally to arrange to discard the input line
             after errors.

             The raw_mode flag was not being initialized by new_GetLine().
             It is now initialized to zero.

             I removed the zapline flag, since using the endline flag to
             communicate the desire to terminate the line, did the same

             gl_terminal_move_cursor() now does nothing when the input
             line isn't displayed.

             Fixed bug which was causing newlines not to be output
             at the end of each newly entered line. I was
             interpreting the gl->endline flag in conflicting ways in
             two places. To fix this I have created a gl->displayed
             flag. This flags whether an input line is currently

           getline.c libtecla.h man/func/
             I added a new function that programs can call to clear
             the terminal between calls to gl_get_line().

             Under linux when _POSIX_C_SOURCE is defined, getpwent()
             and associated functions become undefined, because
             _SVID_SOURCE and _BSD_SOURCE become undefined. Adding
             these feature macros back to CFLAGS resolves this.

           getline.c man/func/
             Following the lead of Edward Chien, I wrote a function
             called gl_bind_keyseq(), which binds a specified
             key-sequence to a given action, or unbinds the

           getline.c man/func/
             I implemented a simple function called
             cpl_recall_matches().  This recalls the return value of
             the last call to cpl_complete_word().

             The documented signal handling, fd event-handling,
             inactivity timeout handling, and server-mode non-blocking
             I/O features are now implemented for non-interactive
             input streams, such as pipes and files.

           getline.c libtecla.h man/func/ demo3.c
             I added a new return status enumerator to report
             when an end-of-file condition causes gl_get_line()
             to return NULL.

             I rewrote the history facility. The previous
             circular buffer implementation was a nightmare to change,
             and it couldn't efficiently support certain newly
             requested features. The new implementation stores history
             lines in linked lists of fixed sized string segments,
             taken from the buffer, with each line being reference
             counted and recorded in a hash table. If the user enters
             a line multiple times, only one copy of the line is now
             stored. Not only does this make better use of the
             available buffer space, but it also makes it easy to
             ensure that a line whose prefix matches the current
             search prefix, isn't returned more than once in sequence,
             since we can simply see if the latest search result has
             the same hash-table pointer as the previous one, rather
             than having to compare strings. Another plus is that due
             to the use of linked lists of nodes of fixed size line
             segments, there is no longer any need to continually
             shuffle the contents of the buffer in order to defragment
             it. As far as the user is concerned, the visible
             differences are as follows:

             1. If the user enters a given line multiple times in a
                row, each one will be recorded in the history list,
                and will thus be listed by gl_show_history(), and
                saved in the history file. Previously only one line
                was recorded when consecutive duplicates were entered.
                This was a kludge to prevent history recall from
                recalling the same line multiple times in a row. This
                only achieved the desired result when not recalling by

             2. Not only simple recall, but prefix-based history line
                recalls now don't return the same line multiple times
                in a row. As mentioned in (1) above, previously this
                only worked when performing a simple recall, without a
                search prefix.

             The one-line function, gl_buff_curpos_to_term_curpos()
             was only being used by gl_place_cursor(), so I inlined it
             in that function, and removed it.

             gl_suspend_process() was calling the application-level
             gl_normal_io() and gl_raw_io() functions, where it should
             have been calling the internal versions _gl_normal_io()
             and _gl_raw_io().
             Also gl_handle_signal() was masking and unmasking just
             the signals of the first element of the gl[] array
             argument. It now masks and unmasks all trappable signals.

             Now that the number of terminal characters used to
             display the current input line, is recorded, the relative
             line on which the last character of the input line
             resides can be determined without having to call
             gl_buff_curpos_to_term_curpos(). This is now used by
             gl_normal_io() via gl_start_newline(), so there is now no
             need for gl_buff_curpos_to_term_curpos() to be
             async-signal safe. I have thus removed the annoying
             gl->cwidth[] array, and gl_buff_curpos_to_term_curpos()
             now calls gl_width_of_char() directly again. There is
             also now no need for the gl_line_of_char_start() and
             gl_line_of_char_end() functions, so I have removed them.

             Unfortunately it turns out that the terminfo/termcap
             control sequence which is defined to delete everything
             from the current position to the end of the terminal, is
             only defined to work when at the start of a terminal
             line. In gnome terminals in RedHat 8.0, if it is used
             within a terminal line, it erases the whole terminal
             line, rather than just what follows the cursor. Thus to
             portably truncate the displayed input line it is
             necessary to first use the control sequence which deletes
             from the cursor position to the end of the line, then if
             there are more terminal lines, move to the start of the
             next line, and use the delete to end-of-terminal control
             sequence, then restore the cursor position. This requires
             that one know how many physical terminal lines are used
             by the current input line, so I now keep a record of the
             number of characters so far displayed to the terminal
             following the start of the prompt, and the new
             gl_truncate_display() function uses this information to
             truncate the displayed input line from the current cursor

             gl_start_newline() now moves to an empty line following
             the input line, rather than just to the next line. It
             also arranges for the input line to be redisplayed before
             editing resumes. A major user of this is gl_print_info(),
             which now need not be followed by an explicit call to
             gl_redisplay(), since the terminal input loop in
             gl_get_input_line() ensures that gl_redisplay() is called
             after any action function that asserts gl->redisplay.
             Also, all functions that erase the displayed input line
             can now call the gl_erase_line() function, which is
             designed to work correctly even when a terminal resize
             invalidates the horizontal cursor position.  Finally, the
             new gl_queue_display() function is now used by functions
             that need to arrange for the input line to be displayed
             from scratch after the displayed line has been erased or
             invalidated by other text being written to the terminal.
             All of these changes are aimed at reducing the number of
             places that directly modify gl->term_curpos and

22/12/2002 Markus Gyger   (logged here by mcs)
	     In places where echo and sed were being used to extract
	     the base names of files, Markus substituted the basename
	     command. He also replaced explicit cp and chmod commands
	     with invokations of the install-sh script.
             Use $target_os and $target_cpu, where appropriate,
	     instead of $target.
             The Solaris man function and library man pages should
	     be in sections 3lib and 3tecla respectively, only in
	     Solaris version 2.8 and above.
             Markus provided values for the man page configuration
             variables for HPUX.
             I had missed parameterizing man page section numbers in
	     the man page titles, Markus corrected this.
             Fixed incorrect section number in the link to the
             libtecla man page.
             When compiled to be reentrant, although one can't use the
	     non-reentrant getpwent() function to scan the password
	     file for username completions, one can at least see if
	     the prefix being completed is a valid username, and if
	     the username of the current user minimally matches the
	     prefix, and if so list them. I simplified Markus'
	     modification by adding a prefix argument to the
             _hd_scan_user_home_dirs() function, and redefining the
	     function description accordingly, such that now it
	     reports only those password file entries who's usernames
	     minimally match the specified prefix. Without this, it
	     would have been necessary to peak inside the private data
	     argument passed in by cf_complete_username().
             Markus also provided code which under Solaris uses the
             non-reentrant interfaces if the reentrant version of the
             library isn't linked with the threads library.

             Markus pointed out that LDFLAGS was being picked up by
             the configure script, but not then being interpolated
             into te Makefile. I have thus added the necessary
             assignment to and arranged for the value of
             LDFLAGS to be passed on to recursive make's. I also did
             the same for CPPFLAGS, which had also been omitted.

           man/* man/*/* configure
             It turns out that the assignment of man page sections to
             topics differs somewhat from system to system, so this is
             another thing that needs to be configured by the main
             configuration script, rather than being hardwired. All
             man pages have now been moved into suitably named
             topic-specific sub-directories of the top-level man
             directory, and instead of having a numeric suffix, now
             have the .in suffix, since they are now preprocessed by
             the configure script, in the same fashion as
             Whithin these *.in versions of the man pages, and within
   , the installation subdirectory (eg. man1) and
             the file-name suffix (eg. 1), are written using
             configuration macros, so that they get expanded to the
             appropriate tokens when the configure script is run. In
             principle, the man pages could also take advantage of
             other configuration macros, such as the one which expands
             to the library installation directory, to include full
             path names to installed files in the documentation, so in
             the future this feature could have more uses than just
             that of parameterizing man page sections.

           man3 man3/* html/index.html update_html
             Markus suggested splitting the gl_get_line(3) man page
             into user and developer sections, and also pointed out
             that the enhance man page should be in section 1, not
             section 3. I have thus created a top-level man
             directory in which to place the various sections, and
             moved the man3 directory into it. The enhance.3 man page
             is now in man/man1/enhance.1. I have extracted all
             user-oriented sections from the gl_get_line(3) man page
             and placed them in a new man7/tecla.7 man page.

             Terminal resizing was broken in normal mode, due to
             me forcing the terminal cursor position to zero in the
             wrong place in gl_check_caught_signal().

14/12/2002 Markus Gyger  (logged here by mcs)
             Under Solaris, recent versions of gcc search
             /usr/local/include for header files before the system
             directories. This caused a problem if ncurses was
             installed under Solaris, since the termcap.h include file
             in /usr/local/include ended up being used at compile
             time, whereas the system default version of the curses
             library was used at link time. Since the two libraries
             declare tputs() differently, this evoked a complaint from
             gcc. Markus came up with a way to force Gnu cpp to move
             /usr/local/include to the end of the system-include-file
             search path, where it belongs.

             I rewrote the man page which documents the new non-blocking
             server I/O mode.

             I wrote a new version of demo3.c, using signal handlers
             that call gl_handle_signal() and gl_abandon_line(), where
             previously in this demo, these functions were called from
             the application code.

             gl_normal_io(), gl_raw_io() and gl_handle_signal() and
             gl_abandon_line() are now signal safe, provided that
             signal handlers that call them are installed with sa_mask's
             that block all other signals who's handlers call them.
             This is the case if gl_tty_signals() is used to install
             signal handlers that call any of these functions.

             A major stumbling block that had to be overcome was that
             gl_displayed_char_width() calls isprint(), which can't
             safely be called from a signal handler (eg. under linux,
             the is*() functions all use thread-specific data
             facilities to support per-thread locales, and the
             thread-specific data facilities aren't signal safe). To
             work around this, all functions that modify the
             input-line buffer, now do so via accessor functions which
             also maintain a parallel array of character widths, for
             use by gl_buff_curpos_to_term_curpos() in place of
             gl_displayed_char_width(). Other minor problems were the
             need to avoid tputs(), who's signal safety isn't defined.

05/12/2002 Eric Norum        (logged here by
             Eric provided the configuration information needed
             to build shared libraries under Darwin (Max OS X).

05/12/2002 Richard Mlynarik  (logged here by
             AC_PROG_RANLIB gets the wrong version of ranlib when
             cross compiling, so has now been replaced by an
             invokation of AC_CHECK_TOOL. In addition, AC_CHECK_TOOL
             is also now used to find an appropriate version of LD.

05/12/2002 (based on patch by Pankaj Rathore)
           getline.c libtecla.h man3/gl_get_line.3
             The new gl_set_term_size() function provides a way
             to tell gl_get_line() about changes in the size of
             the terminal in cases where the values returned by
             ioctl(TIOCGWINSZ) isn't correct.

             Rather than calling sprintf() to see how much space would
             be needed to print a given number in octal, I wrote a
             gl_octal_width() function, for use by
             gl_displayed_char_width().  This makes the latter
             function async signal safe.

             Whenever the buffer is exhausted, and getting a new
             buffer node would require a call to malloc(), attempt
             to flush the buffer to the terminal. In blocking I/O
             mode this means that the buffer never grows. In
             non-blocking I/O mode, it just helps keep the buffer
             size down.

           freelist.h freelist.c
             The new _idle_FreeListNodes() function queries the
             number of nodes in the freelist which aren't currently
             in use.

             This now accepts all of the targets that the configured
             makefile does, and after configuring the latter makefile,
             it invokes it with the same options.

             I completed the man page for all of the new functions
             related to non-blocking I/O.

             I wrote a long section on reliable signal handling,
             explaining how gl_get_line() does this, how to make
             use of this in a program, and how to handle signals
             reliably when faced with other blocking functions.
             This basically documents what I have learnt about
             signal handling while working on this library.

           getline.c man3/gl_get_line.3
             In non-blocking server mode, the gl_replace_prompt()
             function can now be used between calls to gl_get_line()
             if the application wants to change the prompt of the
             line that is being edited.

             I documented the new gl_return_status() and
             gl_error_message() functions.

           getline.c man3/gl_get_line.3
             Added SIGPOLL and SIGXFSZ to the list of signals that
             are trapped by default. These are process termination
             signals, so the terminal needs to be restored to a
             usable state before they terminate the process.

           getline.c libtecla.h
             Completed the essential changes needed to support
             non-blocking server-I/O mode.

             The new gl_io_mode() function allows one to switch to
             and from non-blocking server-I/O mode.

             The new gl_raw_io() function is used in non-blocking
             server-I/O mode to switch the terminal into non-blocking
             raw I/O mode.

             The new gl_normal_io() function is used in non-blocking
             server-I/O mode to switch the restore the terminal to
             a normal, blocking state. This is used to suspend line
             input before suspending the process or writing messages
             to the terminal.

             The new gl_tty_signals() function installs specified
             signals handlers for all signals that suspend, terminate
             or resume processes, and also for signals that indicate
             that the terminal has been resized. This not only saves
             the application from having to keep its own ifdef'd list
             of such signals, of which there are many, but it also
             makes sure that these signal handlers are registered
             correctly. This includes using the sa_mask member of each
             sigaction structure to ensure that only one of these
             handlers runs at a time. This is essential to avoid the
             signal handlers all trying to simultaneously modify
             shared global data.

             The new gl_handle_signal() function is provided for
             responding (from application level) to signals caught by
             the application. It handles process suspension, process
             termination and terminal resize signals.

             The new gl_pending_io() function tells the application
             what direction of I/O gl_get_line() is currently waiting

             In non-blocking server I/O mode, the new
             gl_abandon_line() function can be called between calls to
             gl_get_line() to discard an input line and force the next
             call to gl_get_line() to start the input of a new line.

             Also, in non-blocking server-I/O gl_get_line() doesn't
             attempt to do anything but return when one of the signals
             that it is configured to catch is caught. This is
             necessary because when in this mode, the application is
             required to handle these signals when gl_get_line() is
             running, and the default configuration of most of these
             signals in gl_get_line() is to restore the terminal then
             call the application signal handlers. This would be a
             case of too many cooks spoiling the broth, so in this
             mode, gl_get_line() always defers to the application's
             signal handlers.
           getline.c libtecla.h
             I implemented a couple of new functions to support
             reliable signal handling, as now documented
             (see above) in the gl_get_line(3) man page.

             The new gl_catch_blocked() function tells gl_get_line()
             to unblock all configured signals around calls to
             long-running functions, not only those that aren't
             blocked when gl_get_line() is called. This allows
             the caller to implement reliable signal handling,
             since the unblocking is only done from within code
             protected by sigsetjmp(), which avoids race conditions.

             The new gl_list_signals() function fills a provided
             sigset_t with the set of signals that gl_get_line() is
             currently configured to catch. This allows callers to
             block said signals, such that they are only unblocked by
             gl_get_line() when it is waiting for I/O. When used in
             conjunction with the gl_catch_blocked() function, this
             removes the potential for race conditions.

             Also, when gl_get_line() installs its signal handler,
             it uses the sa_mask member of the sigaction structure
             to ensure that only one instance of this signal handler
             will ever be executing at a time.

25/11/2002 (bug reported by Pankaj Rathore)
             When any history recall action was invoked when the
             input line buffer was full, an error message would be
             displayed complaining about the length of the string
             in the line input buffer being inconsistent with the
             specified allocated size. This was because instead of
             sending the allocated size of the input line, I was
             sending the length excluding the element that is
             reserved for the '\0' terminator. Sending it the
             correct size corrected the problem.

             All public functions which take GetLine objects as
             arguments now block signals on entry and restore the
             signal mask on return. This was an attempt to make it
             safe to call getline functions from signal handlers, but
             the fact is that the functions that I really wanted this
             to apply to, potentially call malloc(), so this currently
             isn't the case.

           getline.c libtecla.h
             The new gl_return_status() function returns an enumerated
             return status which can be used to query what caused
             gl_get_line() to return.

           Most existing .c and .h files, plus errmsg.c errmsg.h
             Until now, many library functions would report error
             messages to stderr. This isn't appropriate for library
             functions, so in place of this behavior, error messages
             are now recorded in internal ErrMsg objects, and passed
             between modules via new module-specific error querying
             functions. In addition, errno is now set appropriately.
             Thus when gl_get_line() and related functions return an
             error, strerror() can be used to look up system errors,
             and gl_error_message() can be used to recover a higher level
             error message. Note that error messages that are
             responses to user actions continue to be reported to the
             terminal, as before.

           getline.c keytab.h keytab.c Makefile.rules
             I wrote a new version of _kt_lookup_binding() that didn't
             require the caller to have access to the innards of a
             KeyTab object. This then enabled me to move the definition
             of KeyTab objects into keytab.c and make the typedef in
             keytab.h opaque. Many nested includes were also moved from
             keytab.h into keytab.c.

           getline.c libtecla.h demo3.c
             I split the old gl_resize_terminal() function into
             two parts, gl_query_size() and gl_update_size(), with
             the latter calling the former to get the new terminal

             I fixed a long time bug in the terminal resizing code.
             When the cursor wasn't on the last terminal line of the
             input line, the resizing code would redisplay the
             the line one or more lines above where it should be
             restored. This was due to an error in the calculation of
             the number of lines above the cursor position.

           demo.c demo2.c demo3.c
             I used the new gl_display_text() function to display
             introductory text at the startup of each of the demo
             programs. The text is enclosed within a box of asterixes,
             drawn dynamically to fit within the confines of the
             available terminal width.

           libtecla.h getline.c ioutil.c ioutil.h Makefile.rules
  man3/gl_get_line.3 man3/gl_display_text.3
             Needing a way to display introductory text intelligently
             in the demo programs, I wrote and documented the
             gl_display_text() function. This justifies arbitrary
             length text within the bounds of the terminal width,
             with or without optional indentation, prefixes and

           demo3.c Makefile.rules
             I wrote a new demonstration program. This program acts
             exactly like the main demonstration program, except that
             it uses an external event loop instead of using the
             gl_get_line() internal event loop. This is thus an example
             of the new non-blocking server I/O facility.

           getline.c keytab.c keytab.h libtecla.h man3/gl_get_line.3
             I added the ability to register additional word
             completion actions via the new function
             gl_completion_action().  All action functions now take a
             new (void *data) argument, which is stored with the
             function in the symbol table of actions. The new
             gl_completion_action() function uses this feature to
             record dynamically allocated objects containing the
             specified completion function and callback data along
             with either the gl_complete_word() action function, or
             the gl_list_completions() action function.  These two
             actions continue to use the builtin completion functions
             when their data pointer is NULL.

           The following are changes merged from the non-blocking
           gl_get_line() development branch.

             I wrote a gl_start_newline() function, to replace all of
             the explicit calls to output \r\n to stdout.

             Informational messages are now written to the terminal
             using a new variadic function called gl_print_info().
             This starts a newline, writes string arguments until a
             special argument, GL_END_INFO, is seen, then starts
             another newline.

             Changed _output_ to _print_ in the following function
             names gl_output_control_sequence(), gl_output_char(),
             gl_output_string() and gl_output_raw_string().

             gl_print_raw_string() now has a length argument, so that
             strings that aren't terminated with '\0' can be printed.

             The display of the initial contents of a new line to be
             edited has been moved into a new function called

             The gl_get_input_line() function now takes the prompt
             string as an argument so that gl_replace_prompt() can be
             called from within this function instead of from

             Keyboard input is now buffered in a persistent buffer in
             the parent GetLine object. gl_read_character() checks
             this for unprocessed characters in preference to calling
             gl_read_terminal() to append characters to it.  A new
             function, gl_discard_chars(), removes processed
             characters from this buffer. This change is in
             preparation for a non-blocking version of gl_get_line(),
             where partially input key-sequences must be stored
             between calls to gl_get_line().

           getline.c getline.h history.c history.h cplmatch.c \
           cplmatch.h expand.c expand.h
             All terminal output from gl_get_line() is now routed
             through a GL_WRITE_FN() callback function called
             gl_write_fn. Internal functions in cplmatch.c,
             expand.c and history.c have been created which take
             such callbacks to write output. These are used both
             by functions in getline.c, to display file completions,
             expansions, history etc, and as the internals of existing
             public functions in these files that print to stdio
             streams. In the latter case an internal stdio
             GL_WRITE_FN() callback is substituted, so that the
             functions behave as before.

           getline.c chrqueue.c chrqueue.h
             The gl_write_fn() callback used by gl_get_line() now
             writes to a queue, implemented in chrqueue.c. This queue
             is implemented as a list of blocks of buffer segments,
             the number of which shrink and grow as
             needed. The contents of the queue are flushed to the
             terminal via another GL_WRITE_FN() callback passed to the
             queue object. Currently gl_get_line() passes an internal
             function assigned to gl->flush_fn, called
             gl_flush_terminal(), which writes the contents of the
             queue to the terminal, and knows how to handle both
             blocking and non-blocking I/O. The output queue is
             designed to be flushed to the terminal incrementally, and
             thereby also facilitates non-blocking I/O.

           getline.c getline.h
             gl_get_line() now reads all input via the GL_READ_FN()
             callback, assigned to gl->read_fn. Currently this is
             set to an internal function called gl_read_terminal(),
             which knows how to handle both blocking and
             non-blocking I/O.

           getline.c libtecla.h
             The new gl_set_nonblocking() function can be used to
             enable or disable non-blocking I/O. The default is still
             blocking I/O. In non-blocking mode, the terminal is told
             not to wait when either reading or writing would block.
             gl_get_line() then returns, with a return value of NULL,
             but with the terminal left in raw mode, so that the
             caller's event loop can detect key presses. The caller
             should call gl_return_status() to check whether the NULL
             return value was due to an error, lack of input, or
             inability to write to the terminal without waiting. If
             either reading or writing was said to have blocked, the
             user then should check for I/O readiness in the specified
             direction before calling gl_get_line() again to
             incrementally build up the input line.

           man3/gl_get_line.3 man3/gl_inactivity_timeout.3
             I documented the new gl_inactivity_timeout() function.

           libtecla.h getline.c
             I added a new gl_inactivity_timeout() function. On
             systems that have the select system call, this provides
             the option of registering a function that is then called
             whenever no I/O activity has been seen for more than a
             specified period of time. Like the gl_watch_fd()
             facility, timeout callbacks return a code which tells
             gl_get_line() how to proceed after the timeout has been
04/07/2002  (based on a bug report from Michael MacFaden)
             The internal event handler wasn't responding to write
             events on client file descriptors, due to a typo which
             resulted in read events being checked for twice, and
             writes not checked for at all.
             The amount of space to allocate for pathnames is supposed
             to come from PATH_MAX in limits.h, but I had neglected to
             include limits.h. This went unnoticed because on most
             systems the equivalent number is deduced by calling
             pathconf(). Apparently under NetBSD this function doesn't
             work correctly over NFS mounts.

30/05/2002 Version 1.4.1 released.

25/05/2002  (based on suggestions by Paul Smith)
             Apparently, under QNX pathconf("/",_PC_PATH_MAX) returns
             EINVAL. At Paul's suggestion I have modified the code to
             silently substitute the existing MAX_PATHLEN_FALLBACK
             value if pathconf() returns an error of any kind.
             Under QNX, sysconf(_SC_GETPW_R_SIZE_MAX) also apparently
             returns EINVAL, so as with pathconf() I modified the code
             to substitute a fallback default, rather than
             complaining and failing.
             Paul told me that the inclusion of sys/termios.h was
             causing compilation of enhance.c to fail under QNX. This
             line is a bug.  The correct thing to do is include
             termios.h without a sub-directory prefix, as I was
             already doing futher up in the file, so I have just
             removed the errant include line.

07/05/2002  (async development branch only)
             gl_read_character() now caches and reads unprocessed
             characters from a key-press lookahead buffer. Whenever
             gl_intepret_char() receives a new character which makes
             an initially promising key-sequence no longer match the
             prefix of any binding, it now simply discards the first
             character from the key-press buffer and resets the buffer
             pointer so that the next call to gl_read_character()
             returns the character that followed it, from the buffer.
             The part of gl_get_input_line() which preloads, displays
             and prepares to edit a new input line, has now been moved
             into a function called gl_present_line().

           getline.c configure
             Mac OS X doesn't have a term.h or termcap.h, but it does
             define prototypes for tputs() and setupterm(), so the
             default prototypes that I was including if no headers
             where available, upset it. I've removed these prototypes.
             I also now conditionally include whichever is found of
             curses.h and ncurses/curses.h for both termcap and
             terminfo (before I wasn't including curses.h when
             termcap was selected).

           Updated version number to 1.4.1, ready for a micro

             Added Mac OS X and Cygwin to the list of systems that
             can compile libtecla.

             Under Mac OS X, the tputs() callback function returns
             void, instead of the int return value used by other
             systems. This declaration is now used if both __MACH__
             and __APPLE__ are defined. Hopefully these are the
             correct system macros to check. Thanks for Stephan
             Fiedler for providing information on Mac OS X.

  configure getline.c
             Some systems don't have term.h, and others have it hidden
             in an ncurses sub-directory of the standard system include
             directory. If term.h can't be found, simply don't include
             it. If it is in an ncurses sub-directory, include
             ncurses/term.h instead of term.h.

  configure Makefile.rules
             Use ranlib on systems that need it (Mac OS X).  Also,
             make all components of the installation directories where
             needed, instead of assuming that they exist.

             When the tab completion binding was unbound from the tab
             key, hitting the tab key caused gl_get_line() to ring the
             bell instead of inserting a tab character. This is
             problematic when using the 'enhance' program with
             Jython, since tabs are important in Python. I have
             corrected this.

10/12/2001 Version 1.4.0 released.

             If the TIOCGWINSZ ioctl doesn't work, as is the case when
             running in an emacs shell, leave the size unchanged, rather
             than returning a fatal error.

             Now that the configure version of CFLAGS is included in
             the makefile, I noticed that the optimization flags -g
             and -O2 had been added. It turns out that if CFLAGS isn't
             already set, the autoconf AC_PROG_CC macro initializes it
             with these two optimization flags. Since this would break
             backwards compatibility in embedded distributions that
             already use the OPT= makefile argument, and because
             turning debugging on needlessly bloats the library, I now
             make sure that CFLAGS is set before calling this macro.

             Use argv[0] in error reports instead of using a
             hardcoded macro.

             The cut buffer wasn't being cleared after being
             used as a work buffer by gl_load_history().

             I removed my now redundant definition of SUN_TPUTS from
             CFLAGS. I also added "-I/usr/include" to CFLAGS under
             Solaris to prevent gcc from seeing conflicting versions
             of system header files in /usr/local/include.

06/12/2001 Markus Gyger (logged here by mcs)
           Lots of files.
             Lots of corrections to misspellings and typos in the
             Markus reverted a supposed fix that I added a day or two
             ago. I had incorrectly thought that in Solaris 8, Sun had
             finally brought their declaration of the callback
             function of tputs() into line with other systems, but it
             turned out that gcc was pulling in a GNU version of
             term.h from /usr/local/include, and this was what
             confused me.

             I added @CFLAGS@ to the CFLAGS assignment, so that
             if CFLAGS is set as an environment variable when
             configure is run, the corresponding make variable
             includes its values in the output makefile.

           getline.c libtecla.h man3/gl_get_line.3
             I added a function that programs can use to find out
             which signal caused gl_get_line() to return EINTR.

             When the newline action was triggered by a printable
             character, it failed to display that character. It now
             does. Also, extra control codes that I had added, to
             clear to the end of the display after the carriage return,
             but before displaying the prompt, were confusing expect
             scripts, so I have removed them. This step is now done
             instead in gl_redisplay() after displaying the full input

           getline.c man3/gl_get_line.3
             A user convinced me that continuing to invoke meta
             keybindings for meta characters that are printable is a
             bad idea, as is allowing users to ask to have setlocale()
             called behind the application's back. I have thus changed
             this. The setlocale configuration option has gone, and
             gl_get_line() is now completely 8-bit clean, by default.
             This means that if a meta character is printable, it is
             treated as a literal character, rather than a potential
             M-c binding.  Meta bindings can still be invoked via
             their Esc-c equivalents, and indeed most terminal
             emulators either output such escape pairs by default when
             the meta character is pressed, or can be configured to do
             so. I have documented how to configure xterm to do this,
             in the man page.

           getline.c man3/gl_get_line.3
             gl_get_line() by default now prints any 8-bit printable
             characters that don't match keybindings. Previously
             characters > 127 were only printed if preceded by the
             literal-next action.  Alternatively, by placing the
             command literal_if_printable in the tecla configuration
             file, all printable characters are treated as literal
             characters, even if they are bound to action functions.

             For international users of programs written by
             programmers that weren't aware of the need to call
             setlocale() to support alternate character sets, the
             configuration file can now also contain the single-word
             command "setlocale", which tells gl_get_line() to remedy

           demo.c demo2.c enhance man3/gl_get_line.3
             All demos and programs now call setlocale(LC_CTYPE,"").
             This makes them support character sets of different
             locales, where specified with the LC_CTYPE, LC_ALL, or
             LANG environment variables. I also added this to the demo
             in the man page, and documented its effect.

             When displaying unsigned characters with values over
             127 literally, previously it was assumed that they would
             all be displayable. Now isprint() is consulted, and if it
             says that a character isn't printable, the character code
             is displayed in octal like \307. In non-C locales, some
             characters with values > 127 are displayable, and
             isprint() tells gl_get_line() which are and which aren't.

           getline.c pathutil.c history.c enhance.c demo2.c
             All arguments of the ctype.h character class functions
             are now cast to (int)(unsigned char). Previously they
             were cast to (int), which doesn't correctly conform to
             the requirements of the C standard, and could cause
             problems for characters with values > 127 on systems
             with signed char's.

           man3/enhance.3 man3/libtecla.3
             I started writing a man page for the enhance program.

  Makefile.rules INSTALL
             It is now possible to specify whether the demos and other
             programs are to be built, by overriding the default
             values of the DEMOS, PROGRAMS and PROGRAMS_R variables.
             I have also documented the BINDIR variable and the
             install_bin makefile target.

           getline.c libtecla.h man3/gl_get_line.3
           man3/gl_ignore_signal.3 man3/gl_trap_signal.3
             Signal handling has now been modified to be customizable.
             Signals that are trapped by default can be removed from
             the list of trapped signals, and signals that aren't
             currently trapped, can be added to the list. Applications
             can also specify the signal and terminal environments in
             which an application's signal handler is invoked, and
             what gl_get_line() does after the signal handler returns.

           getline.c man3/gl_get_line.3
             Added half-bright, reverse-video and blinking text to the
             available prompt formatting options.
             Removed ^O from the default VT100 sgr0 capability
             string.  Apparently it can cause problems with some
             terminal emulators, and we don't need it, since it turns
             off the alternative character set mode, which we don't
             gl_tigetstr() and gl_tgetstr() didn't guard against the
             error returns of tigetstr() and tgetstr() respectively.
             They now do.

           getline.c libtecla.h man3/gl_get_line.3
             Although the default remains to display the prompt string
             literally, the new gl_prompt_style() function can be used
             to enable text attribute formatting directives in prompt
             strings, such as underlining, bold font, and highlighting

           enhance.c Makefile.rules configure
             I added a new program to the distribution that allows one
             to run most third party programs with the tecla library
             providing command-line editing.

           libtecla.h getline.c man3/gl_get_line.3 history.c history.h
             I added a max_lines argument to gl_show_history() and
             _glh_show_history(). This can optionally be used to
             set a limit on the number of history lines displayed.
           libtecla.h getline.c man3/gl_get_line.3
             I added a new function called gl_replace_prompt(). This
             can be used by gl_get_line() callback functions to
             request that a new prompt be use when they return.

           getline.c man3/gl_get_line.3
             I implemented, bound and documented the list-history
             action, used for listing historical lines of the current
             history group.
           getline.c man3/gl_get_line.3 man3/gl_echo_mode.3
             I wrote functions to specify and query whether subsequent
             lines will be visible as they are being typed.

           getline.c man3/gl_get_line.3
             For those cases where a terminal provides its own
             high-level terminal editing facilities, you can now
             specify an edit-mode argument of 'none'. This disables
             all tecla key bindings, and by using canonical terminal
             input mode instead of raw input mode, editing is left up
             to the terminal driver.

           libtecla.h getline.c history.c history.h
           man3/gl_get_line.3 man3/gl_history_info.3
             I added the new gl_state_of_history(),
             gl_range_of_history() and gl_size_of_history()
             functions for querying information about the
             history list.
             While testing the new gl_size_of_history()
             function, I noticed that when the history buffer
             wrapped, any location nodes of old lines between
             the most recent line and the end of the buffer
             weren't being removed. This could result in bogus
             entries appearing at the start of the history list.
             Now fixed.


           libtecla.h getline.c history.c history.h
           man3/gl_get_line.3 man3/gl_lookup_history.3
             I added a function called gl_lookup_history(), that
             the application can use to lookup lines in the history
           libtecla.h getline.c history.c history.h man3/gl_get_line.3
             gl_show_history() now takes a format string argument
             to control how the line is displayed, and with what
             information. It also now provides the option of either
             displaying all history lines or just those of the
             current history group.
           getline.c man3/gl_get_line.3
             gl_get_line() only archives lines in the history buffer
             if the newline action was invoked by a newline or
             carriage return character.


           history.c history.h getline.c libtecla.h
           man3/gl_get_line.3 man3/gl_resize_history.3
           man3/gl_limit_history.3 man3/gl_clear_history.3
	     I added a number of miscellaneous history configuration
	     functions. You can now resize or delete the history
	     buffer, limit the number of lines that are allowed in the
	     buffer, clear either all history or just the history of
	     the current history group, and temporarily enable and
	     disable the history mechanism.


             tputs_fp is now only declared if using termcap or
           getline.c man3/gl_get_line.3
             I added a public gl_terminal_size() function for
             updating and querying the current size of the terminal.
           update_version libtecla.h
             A user noted that on systems where the configure script
             couldn't be used, it was inconvenient to have the version
             number macros set by the configure script, so they are
             now specified in libtecla.h. To reduce the likelihood
             that the various files where the version number now
             appears might get out of sync, I have written the
             update_version script, which changes the version number
             in all of these files to a given value.


           getline.c history.c history.h man3/gl_get_line.3
             I added a max_lines argument to gl_save_history(), to
             allow people to optionally place a ceiling on the number
             of history lines saved. Specifying this as -1 sets the
             ceiling to infinity.


             Under digital unix, getline wouldn't compile with
             _POSIX_C_SOURCE set, due to type definitions needed by
             select being excluded by this flag. Defining the
             _OSF_SOURCE macro as well on this system, resolved this.


           getline.c libtecla.h history.c history.h man3/gl_get_line.3
             I implemented history streams. History streams
             effectively allow multiple history lists to be stored in
             a single history buffer. Lines in the buffer are tagged
             with the current stream identification number, and
             lookups only consider lines that are marked with the
             current stream identifier.
           getline.c libtecla.h history.c history.h man3/gl_get_line.3
             The new gl_show_history function displays the current
             history to a given stdio output stream.


             Previously new_GetLine() installed a persistent signal
             handler to be sure to catch the SIGWINCH (terminal size
             change) signal between calls to gl_get_line(). This had
             the drawback that if multiple GetLine objects were
             created, only the first GetLine object used after the
             signal was received, would see the signal and adapt to
             the new terminal size. Instead of this, a signal handler
             for sigwinch is only installed while gl_get_line() is
             running, and just after installing this handler,
             gl_get_line() checks for terminal size changes that
             might have occurred while the signal handler wasn't
             Dynamically allocated copies of capability strings looked
             up in the terminfo or termcap databases are now made, so
             that calls to setupterm() etc for one GetLine object
             don't get trashed when another GetLine object calls
             setupterm() etc. It is now safe to allocate and use
             multiple GetLine objects, albeit only within a single

           version.c Makefile.rules
             I added a function for querying the version number of
             the library.


           getline.c man3/gl_get_line.3
             I added the new gl_watch_fd() function, which allows
             applications to register callback functions to be invoked
             when activity is seen on arbitrary file descriptors while
             gl_get_line() is awaiting keyboard input from the user.

             If a request is received to delete a non-existent
             binding, which happens to be an ambiguous prefix of other
             bindings no complaint is now generated about it being


           getline.c history.c history.h man3/gl_get_line.3
             I added new public functions for saving and restoring the
             contents of the history list. The demo program now uses
             these functions to load and save history in ~/.demo_history.


             On trying the demo for the first time on a KDE konsole
             terminal, I discovered that the default M-O binding
             to repeat history was hiding the arrow keys, which are
             M-OA etc. I have removed this binding. The M-o (ie the
             lower case version of this), is still bound.


           getline.c man3/gl_get_line.3
             Automatic reading of ~/.teclarc is now postponed until
             the first call to gl_get_line(), to give the application
             the chance to specify alternative configuration sources
             with the new function gl_configure_getline(). The latter
             function allows configuration to be done with a string, a
             specified application-specific file, and/or a specified
             user-specific file. I also added a read-init-files action
             function, for re-reading the configuration files, if any.
             This is by default bound to ^X^R. This is all documented
             in gl_get_line.3.


           getline.c man3/gl_get_line.3
             It is now possible to bind actions to key-sequences
             that start with printable characters. Previously
             keysequences were required to start with meta or control
             characters. This is documented in gl_get_line.3.

           getline.c man3/gl_get_line.3
             A customized completion function can now arrange for
             gl_get_line() to return the current input line whenever a
             successful completion has been made. This is signalled by
             setting the last character of the optional continuation
             suffix to a newline character. This is documented in

05/07/2001 Bug reported by Mike MacFaden, fixed by mcs

             There was a bug in the configure script that only
             revealed itself on systems without termcap but not
             terminfo (eg. NetBSD). I traced the bug back to a lack of
             sufficient quoting of multi-line m4 macro arguments in
   , and have now fixed this and recreated the
             configure script.

05/07/2001 Bug reported and patched by Mike MacFaden (patch modified
           by mcs to match original intentions).

             getline.c wouldn't compile when termcap was selected as
             the terminal information database. setupterm() was being
             passed a non-existent variable, in place of the term[]
             argument of gl_control_strings(). Also if
             gl_change_terminal() is called with term==NULL, "ansi"
             is now substituted.

02/07/2001 Version 1.3.3 released.


           getline.c expand.c cplmatch.c
             Added checks to fprintf() statements that write to the
             Move the cursor to the end of the line before suspending,
             so that the cursor doesn't get left in the middle of the
             input line.
             On systems that don't support shared libraries, the
             distclean target of make deleted libtecla.h. This has
             now been fixed.
             gl_change_terminal() was being called by gl_change_editor(),
             with the unwanted side effect that raw terminal modes were
             stored as those to be restored later, if called by an
             action function. gl_change_terminal() was being called in
             this case to re-establish terminal-specific key bindings,
             so I have just split this part of the function out into
             a separate function for both gl_change_editor() and
             gl_change_terminal() to call.


             Signal handling has been improved. Many more signals are
             now trapped, and instead of using a simple flag set by a
             signal handler, race conditions are avoided by blocking
             signals during most of the gl_get_line() code, and
             unblocking them via calls to sigsetjmp(), just before
             attempting to read each new character from the user.
             The matching use of siglongjmp() in the signal
             handlers ensures that signals are reblocked correctly
             before they are handled. In most cases, signals cause
             gl_get_line() to restore the terminal modes and signal
             handlers of the calling application, then resend the
             signal to the application. In the case of SIGINT, SIGHUP,
             SIGPIPE, and SIGQUIT, if the process still exists after
             the signals are resent, gl_get_line() immediately returns
             with appropriate values assigned to errno. If SIGTSTP,
             SIGTTIN or SIGTTOU signals are received, the process is
             suspended. If any other signal is received, and the
             process continues to exist after the signal is resent to
             the calling application, line input is resumed after the
             terminal is put back into raw mode, the gl_get_line()
             signal handling is restored, and the input line redrawn.
             I added a SIGNAL HANDLING section to the gl_get_line()
             man page, describing the new signal handling features.

21/05/2001 Version 1.3.2 released.


             When vi-replace-char was used to replace the character at
             the end of the line, it left the cursor one character to
             its right instead of on top of it. Now rememdied.
             When undoing, to properly emulate vi, the cursor is now
             left at the leftmost of the saved and current cursor
           getline.c man3/gl_get_line.3
             Implemented find-parenthesis (%), delete-to-paren (M-d%),
             vi-change-to-paren (M-c%), copy-to-paren (M-y%).
           cplfile.c pcache.c
             In three places I was comparing the last argument of
             strncmp() to zero instead of the return value of


           getline.c man3/gl_get_line.3
             Implemented and documented the vi-repeat-change action,
             bound to the period key. This repeats the last action
             that modified the input line.


             I documented the new action functions and bindings
             provided by Tim Eliseo, plus the ring-bell action and
             the new "nobeep" configuration option.
             I modified gl_change_editor() to remove and reinstate the
             terminal settings as well as the default bindings, since
             these have editor-specific differences. I also modified
             it to not abort if a key-sequence can't be bound for some
             reason. This allows the new vi-mode and emacs-mode
             bindings to be used safely.
             When the line was re-displayed on receipt of a SIGWINCH
             signal, the result wasn't visible until the next
             character was typed, since a call to fflush() was needed.
             gl_redisplay_line() now calls gl_flush_output() to remedy


             Under Linux, calling fflush(gl->output_fd) hangs if
             terminal output has been suspended with ^S. With the
             tecla library taking responsability for reading the stop
             and start characters this was a problem, because once
             hung in fflush(), the keyboard input loop wasn't entered,
             so the user couldn't type the start character to resume
             output.  To remedy this, I now have the terminal process
             these characters, rather than the library.


             The literal-next action is now implemented as a single
             function which reads the next character itself.
             Previously it just set a flag which effected the
             interpretation of the next character read by the input
             Added a ring-bell action function. This is currently
             unbound to any key by default, but it is used internally,
             and can be used by users that want to disable any of the
             default key-bindings.

12/05/2001 Tim Eliseo    (logged here by mcs)

             Don't reset gl->number until after calling an action
             function. By looking at whether gl->number is <0 or
             not, action functions can then tell whether the count
             that they were passed was explicitly specified by the
             user, as opposed to being defaulted to 1.
             In vi, the position at which input mode is entered
             acts as a barrier to backward motion for the few
             backward moving actions that are enabled in input mode.
             Tim added this barrier to getline.
             In gl_get_line() after reading an input line, or
             having the read aborted by a signal, the sig_atomic_t
             gl_pending_signal was being compared to zero instead
             of -1 to see if no signals had been received.
             gl_get_line() will thus have been calling raise(-1),
             which luckily didn't seem to do anything. Tim also
             arranged for errno to be set to EINTR when a signal
             aborts gl_get_line().
             The test in gl_add_char_to_line() for detecting
             when overwriting a character with a wider character,
             had a < where it needed a >. Overwriting with a wider
             character thus overwrote trailing characters. Tim also
             removed a redundant copy of the character into the
             line buffer.
             gl_cursor_left() and gl->cursor_right() were executing
             a lot of redundant code, when the existing call to the
             recently added gl_place_cursor() function, does all that
             is necessary.
             Remove redundant code from backward_kill_line() by
             re-implimenting in terms of gl_place_cursor() and
             gl_forward_delete_char() now records characters in cut
             buffer when in vi command mode.
             In vi mode gl_backward_delete_char() now only deletes
             up to the point at which input mode was entered. Also
             gl_delete_chars() restores from the undo buffer when
             deleting in vi insert mode.
             Added action functions, vi-delete-goto-column,
             vi-change-to-bol, vi-change-line, emacs-mode, vi-mode,
             vi-forward-change-find, vi-backward-change-find,
             vi-forward-change-to, vi-backward-change-to,
             vi-change-goto-col, forward-delete-find, backward-delete-find,
             forward-delete-to, backward-delete-to,
             delete-refind, delete-invert-refind, forward-copy-find,
             backward-copy-find, forward-copy-to, backward-copy-to
             copy-goto-column, copy-rest-of-line, copy-to-bol, copy-line,
             history-re-search-forward, history-re-search-backward.

06/05/2001 Version 1.3.1 released.


             Old versions of GNU ld don't accept version scripts.
             Under Linux I thus added a test to try out ld with
             the --version-script argument to see if it works.
             If not, version scripts aren't used.
             My test for versions of Solaris earlier than 7
             failed when confronted by a three figure version
             number (2.5.1). Fixed.


             In vi mode, history-search-backward and
             history-search-forward weren't doing anything when
             invoked at the start of an empty line, whereas
             they should have acted like up-history and down-history.
             When shared libraries are being created, the build
             procedure now arranges for any alternate library
             links to be created as well, before linking the
             demos. Without this the demos always linked to the
             static libraries (which was perfectly ok, but wasn't a
             good example).
             On systems on which shared libraries were being created,
             if there were no alternate list of names, make would
             abort due to a Bourne shell 'for' statement that didn't
             have any arguments. Currently there are no systems who's
             shared library configurations would trigger this
             The demos now relink to take account of changes to the
             When determining whether the reentrant version of the
             library should be compiled by default, the configure
             script now attempts to compile a dummy program that
             includes all of the appropriate system headers and
             defines _POSIX_C_SOURCE. This should now be a robust test
             on systems which use C macros to alias these function
             names to other internal functions.
             Under Solaris 2.6 and earlier, the curses library is in
             /usr/ccs/lib. Gcc wasn't finding this. In addition to
             remedying this, I had to remove "-z text" from
             LINK_SHARED under Solaris to get it to successfully
             compile the shared library against the static curses
             Under Linux the -soname directive was being used
             incorrectly, citing the fully qualified name of the
             library instead of its major version alias. This will
             unfortunately mean that binaries linked with the 1.2.3
             and 1.2.4 versions of the shared library won't use
             later versions of the library unless relinked.


             In gl_get_input_line(), don't redundantly copy the
             start_line if start_line == gl->line.

30/04/2001 Version 1.3.0 released.


             I removed the --no-undefined directive from the Linux
             LINK_SHARED command. After recent patches to our RedHat
             7.0 systems ld started reporting some internal symbols of
             libc as being undefined.  Using nm on libc indicated that
             the offending symbols are indeed defined, albeit as
             "common" symbols, so there appears to be a bug in
             RedHat's ld. Removing this flag allows the tecla shared
             library to compile, and programs appear to function fine.
             The default key-sequence used to invoke the
             read-from-file action was incorrectly cited as ^Xi
             instead of ^X^F.


           getline.c man3/gl_get_line.3
             A new vi-style editing mode was added. This involved
             adding many new action functions, adding support for
             specifying editing modes in users' ~/.teclarc files,
             writing a higher level cursor motion function to support
             the different line-end bounds required in vi command
             mode, and a few small changes to support the fact that vi
             has two modes, input mode and command mode with different

             When vi editing mode is enabled, any binding that starts
             with an escape or a meta character, is interpreted as a
             command-mode binding, and switches the library to vi
             command mode if not already in that mode. Once in command
             mode the first character of all keysequences entered
             until input mode is re-enabled, are quietly coerced to
             meta characters before being looked up in the key-binding
             table. So, for example, in the key-binding table, the
             standard vi command-mode 'w' key, which moves the cursor
             one word to the right, is represented by M-w. This
             emulates vi's dual sets of bindings in a natural way
             without needing large changes to the library, or new
             binding syntaxes. Since cursor keys normally emit
             keysequences which start with escape, it also does
             something sensible when a cursor key is pressed during
             input mode (unlike true vi, which gets upset).

             I also added a ^Xg binding for the new list-glob action
             to both the emacs and vi key-binding tables. This lists
             the files that match the wild-card expression that
             precedes it on the command line.

             The function that reads in ~/.teclarc used to tell
             new_GetLine() to abort if it encountered anything that it
             didn't understand in this file. It now just reports an
             error and continues onto the next line.
             When passing LIBS=$(LIBS) to recursive invokations of
             make, quotes weren't included around the $(LIBS) part.
             This would cause problems if LIBS ever contained more
             than one word (with the supplied configure script this
             doesn't happen currently). I added these quotes.
           expand.c man3/ef_expand_file.3:
             I wrote a new public function called ef_list_expansions(),
             to list the matching filenames returned by

             I also fixed the example in the man page, which cited
             exp->file instead of exp->files, and changed the
             dangerous name 'exp' with 'expn'.
             Key-binding tables start with 100 elements, and are
             supposedly incremented in size by 100 elements whenever
             the a table runs out of space. The realloc arguments to
             do this were wrong. This would have caused problems if
             anybody added a lot of personal bindings in their
             ~/.teclarc file. I only noticed it because the number of
             key bindings needed by the new vi mode exceeded this
             ef_expand_file() is now reported as having been added in
             the upcoming 1.3.0 release.

25/03/2001 Markus Gyger  (logged here by mcs)

             Make symbolic links to alternative shared library names
             relative instead of absolute.
             The HP-UX file should be made in the
             compilation directory, to allow the source code directory
             to be on a readonly filesystem.
           cplmatch.c demo2.c history.c pcache.c
             To allow the library to be compiled with a C++ compiler,
             without generating warnings, a few casts were added where
             void* return values were being assigned directly to
             none void* pointer variables.


             Added comment header to explain the purpose of the file.
             Also added cpl_init_FileArgs to the list of exported
             symbols. This symbol is deprecated, and no longer
             documented, but for backwards compatibility, it should
             still be exported.
             I had forgotten to run autoconf before releasing version
             1.2.4, so I have just belatedly done so.  This enables
             Markus' changes to "" documented previously,
             (see 17/03/2001).

20/03/2001 John Levon   (logged here by mcs)

             A couple of the function prototypes in libtecla.h have
             (FILE *) argument declarations, which means that stdio.h
             needs to be included. The header file should be self
             contained, so libtecla.h now includes stdio.h.

18/03/2001 Version 1.2.4 released.

           README html/index.html
             Incremented minor version from 3 to 4.


             The fix for the end-of-line problem that I released a
             couple of weeks ago, only worked for the first line,
             because I was handling this case when the cursor position
             was equal to the last column, rather than when the cursor
             position modulo ncolumn was zero.
             The demos are now made by default, their rules now being
             int Makefile.rules instead of
             I documented how to compile the library in a different
             directory than the distribution directory.
             I also documented features designed to facilitate
             configuring and building the library as part of another

17/03/2001 Markus Gyger (logged here by mcs)

             Until now cursor motions were done one at a time. Markus
             has added code to make use the of the terminfo capability
             that moves the cursor by more than one position at a
             time. This greatly improves performance when editing near
             the start of long lines.
             To further improve performance, Markus switched from
             writing one character at a time to the terminal, using
             the write() system call, to using C buffered output
             streams. The output buffer is only flushed when
             Added support for compiling for different architectures
             in different directories. Simply create another directory
             and run the configure script located in the original
             Under Solaris, Linux and HP-UX, symbols that are to be
             exported by tecla shared libraries are explicitly specified
             via symbol map files. Only publicly documented functions
             are thus visible to applications.
             When linking shared libraries under Solaris SPARC,
             registers that are reserved for applications are marked
             as off limits to the library, using -xregs=no%appl when
             compiling with Sun cc, or -mno-app-regs when compiling
             with gcc. Also removed -z redlocsym for Solaris, which
             caused problems under some releases of ld.
           homedir.c  (after minor changes by mcs)
             Under ksh, ~+ expands to the current value of the ksh
             PWD environment variable, which contains the path of
             the current working directory, including any symbolic
             links that were traversed to get there. The special
             username "+" is now treated equally by tecla, except
             that it substitutes the return value of getcwd() if PWD
             either isn't set, or if it points at a different
             directory than that reported by getcwd().

08/03/2001 Version 1.2.3 released.


             On compiling the library under HP-UX for the first time
             I encountered and fixed a couple of bugs:

             1. On all systems except Solaris, the callback function
                required by tputs() takes an int argument for the
                character that is to be printed. Under Solaris it
                takes a char argument. The callback function was
                passing this argument, regardless of type, to write(),
                which wrote the first byte of the argument.  This was
                fine under Solaris and under little-endian systems,
                because the first byte contained the character to be
                written, but on big-endian systems, it always wrote
                the zero byte at the other end of the word. As a
                result, no control characters were being written to
                the terminal.
             2. While attempting to start a newline after the user hit
                enter, the library was outputting the control sequence
                for moving the cursor down, instead of the newline
                character. On many systems the control sequence for
                moving the cursor down happends to be a newline
                character, but under HP-UX it isn't. The result was
                that no new line was being started under HP-UX.


  Makefile.stub configure config.guess
           config.sub Makefile.rules install-sh PORTING README INSTALL
             Configuration and compilation of the library is now
             performed with the help of an autoconf configure
             script. In addition to relieving the user of the need to
             edit the Makefile, this also allows automatic compilation
             of the reentrant version of the library on platforms that
             can handle it, along with the creation of shared
             libraries where configured. On systems that aren't known
             to the configure script, just the static tecla library is
             compiled. This is currently the case on all systems
             except Linux, Solaris and HP-UX. In the hope that
             installers will provide specific conigurations for other
             systems, the script is heavily commented,
             and instructions on how to use are included in a new
             PORTING file.

24/02/2001 Version 1.2b released.


             It turns out that most terminals, but not all, on writing
             a character in the rightmost column, don't wrap the
             cursor onto the next line until the next character is
             output. This library wasn't aware of this and thus if one
             tried to reposition the cursor from the last column,
             gl_get_line() thought that it was moving relative to a
             point on the next line, and thus moved the cursor up a
             line. The fix was to write one extra character when in
             the last column to force the cursor onto the next line,
             then backup the cursor to the start of the new line.
             On terminal initialization, the dynamic LINES and COLUMNS
             environment variables were ignored unless
             terminfo/termcap didn't return sensible dimensions. In
             practice, when present they should override the static
             versions in the terminfo/termcap databases. This is the
             new behavior. In reality this probably won't have caused
             many problems, because a SIGWINCH signal which informs of
             terminal size changes is sent when the terminal is
             opened, so the dimensions established during
             initialization quickly get updated on most systems.

18/02/2001 Version 1.2a released.


             Three months ago I moved the point at which termios.h
             was included in getline.c. Unfortunately, I didn't notice
             that this moved it to after the test for TIOCGWINSZ being
             defined. This resulted in SIGWINCH signals not being
             trapped for, and thus terminal size changes went
             unnoticed. I have now moved the test to after the 
             inclusion of termios.h.

12/02/2001 Markus Gyger     (described here by mcs)

           man3/pca_lookup_file.3 man3/gl_get_line.3
           man3/ef_expand_file.3 man3/cpl_complete_word.3
             In the 1.2 release of the library, all functions in the
             library were given man pages. Most of these simply
             include one of the above 4 man pages, which describe the
             functions while describing the modules that they are in.
             Markus added all of these function names to the lists in
             the "NAME" headers of the respective man pages.
             Previously only the primary function of each module was
             named there.


             On entering a line that wrapped over two or more
             terminal, if the user pressed enter when the cursor
             wasn't on the last of the wrapped lines, the text of the
             wrapped lines that followed it got mixed up with the next
             line written by the application, or the next input
             line. Somehow this slipped through the cracks and wasn't
             noticed until now. Anyway, it is fixed now.

09/02/2001 Version 1.2 released.


           pcache.c libtecla.h
             With all filesystems local, demo2 was very fast to start
             up, but on a Sun system with one of the target
             directories being on a remote nfs mounted filesystem, the
             startup time was many seconds. This was due to the
             executable selection callback being applied to all files
             in the path at startup. To avoid this, all files are now
             included in the cache, and the application specified
             file-selection callback is only called on files as they
             are matched. Whether the callback rejected or accepted
             them is then cached so that the next time an already
             checked file is looked at, the callback doesn't have to
             be called. As a result, startup is now fast on all
             systems, and since usually there are only a few matching
             file completions at a time, the delay during completion
             is also usually small. The only exception is if the user
             tries to complete an empty string, at which point all
             files have to be checked. Having done this once, however,
             doing it again is fast.
             I added a man page documenting the new PathCache module.
             I have added man pages for all of the functions in each
             of the modules. These 1-line pages use the .so directive
             to redirect nroff to the man page of the parent module.
           man Makefile update_html
             I renamed man to man3 to make it easier to test man page
             rediction, and updated Makefile and update_html
             accordingly. I also instructed update_html to ignore
             1-line man pages when making html equivalents of the man
             In cpl_list_completions() the size_t return value of
             strlen() was being used as the length argument of a "%*s"
             printf directive. This ought to be an int, so the return
             value of strlen() is now cast to int. This would have
             caused problems on architectures where the size of a
             size_t is not equal to the size of an int.


             Under UNIX, certain terminal bindings are set using the
             stty command. This, for example, specifies which control
             key generates a user-interrupt (usually ^C or ^Y). What I
             hadn't realized was that ASCII NUL is used as the way to
             specify that one of these bindings is unset. I have now
             modified the code to skip unset bindings, leaving the
             corresponding action bound to the built-in default, or a
             user provided binding.


           pcache.c libtecla.h
             A new module was added which supports searching for files
             in any colon separated list of directories, such as the
             unix execution PATH environment variable. Files in these
             directories, after being individually okayed for
             inclusion via an application provided callback, are
             cached in a PathCache object. You can then look up the
             full pathname of a given filename, or you can use the
             provided completion callback to list possible completions
             in the path-list. The contents of relative directories,
             such as ".", obviously can't be cached, so these
             directories are read on the fly during lookups and
             completions. The obvious application of this facility is
             to provide Tab-completion of commands, and thus a
             callback to place executable files in the cache, is
             This new program demonstrates the new PathCache
             module. It reads and processes lines of input until the
             word 'exit' is entered, or C-d is pressed. The default
             tab-completion callback is replaced with one which at the
             start of a line, looks up completions of commands in the
             user's execution path, and when invoked in other parts of
             the line, reverts to normal filename completion. Whenever
             a new line is entered, it extracts the first word on the
             line, looks it up in the user's execution path to see if
             it corresponds to a known command file, and if so,
             displays the full pathname of the file, along with the
             remaining arguments.
             I added an optional pair of callback function/data
             members to the new cpl_file_completions() configuration
             structure. Where provided, this callback is asked
             on a file-by-file basis, which files should be included
             in the list of file completions. For example, a callback
             is provided for listing only completions of executable
             When listing completions, the length of the type suffix
             of each completion wasn't being taken into account
             correctly when computing the column widths. Thus the
             listing appeared ragged sometimes. This is now fixed.
             I added a function for prepending a string to a path,
             and another for testing whether a pathname referred to
             an executable file.


           libtecla.h cplmatch.c man/cpl_complete_word.3
             The use of a publically defined structure to configure
             the cpl_file_completions() callback was flawed, so a new
             approach has been designed, and the old method, albeit
             still supported, is no longer documented in the man
             pages. The definition of the CplFileArgs structure in
             libtecla.h is now accompanied by comments warning people
             not to modify it, since modifications could break
             applications linked to shared versions of the tecla
             library. The new method involves an opaque CplFileConf
             object, instances of which are returned by a provided
             constructor function, configured with provided accessor
             functions, and when no longer needed, deleted with a
             provided destructor function. This is documented in the
             cpl_complete_word man page. The cpl_file_completions()
             callback distinguishes what type of configuration
             structure it has been sent by virtue of a code placed at
             the beginning of the CplFileConf argument by its

04/01/2001 (Release of version 1.1j)

             I added upper-case bindings for the default meta-letter
             keysequences such as M-b. They thus continue to work
             when the user has caps-lock on.
             I re-implemented the "install" target in terms of new
             install_lib, install_inc and install_man targets. When
             distributing the library with other packages, these new
             targets allows for finer grained control of the
             installation process.


           getline.c man/gl_get_line.3
             I realized that the recall-history action that I
             implemented wasn't what Markus had asked me for. What he
             actually wanted was for down-history to continue going
             forwards through a previous history recall session if no
             history recall session had been started while entering
             the current line. I have thus removed the recall-history
             action and modified the down-history action function


             I modified gl_get_line() to allow the previously returned
             line to be passed in the start_line argument.
           getline.c man/gl_get_line.3
             I added a recall-history action function, bound to M^P.
             This recalls the last recalled history line, regardless
             of whether it was from the current or previous line.

13/12/2000 (Release of version 1.1i)

           getline.c history.h history.c man/gl_get_line.3
             I implemented the equivalent of the ksh Operate action. I
             have named the tecla equivalent "repeat-history". This
             causes the line that is to be edited to returned, and
             arranges for the next most recent history line to be
             preloaded on the next call to gl_get_line(). Repeated
             invocations of this action thus result in successive
             history lines being repeated - hence the
             name. Implementing the ksh Operate action was suggested
             by Markus Gyger. In ksh it is bound to ^O, but since ^O
             is traditionally bound by the default terminal settings,
             to stop-output, I have bound the tecla equivalent to M-o.

01/12/2000 (Release of version 1.1h)

           getline.c keytab.c keytab.h man/gl_get_line.3
             I added a digit-argument action, to allow repeat
             counts for actions to be entered. As in both tcsh
             and readline, this is bound by default to each of
             M-0, M-1 through to M-9, the number being appended
             to the current repeat count. Once one of these has been
             pressed, the subsequent digits of the repeat count can be
             typed with or without the meta key pressed. It is also
             possible to bind digit-argument to other keys, with or
             without a numeric final keystroke. See man page for

           getline.c man/gl_get_line.3
             Markus noted that my choice of M-< for the default
             binding of read-from-file, could be confusing, since
             readline binds this to beginning-of-history. I have
             thus rebound it to ^X^F (ie. like find-file in emacs).

           getline.c history.c history.h man/gl_get_line.3
             I have now implemented equivalents of the readline
             beginning-of-history and end-of-history actions.
             These are bound to M-< and M-> respectively.

           history.c history.h
             I Moved the definition of the GlHistory type, and
             its subordinate types from history.h to history.c.
             There is no good reason for any other module to
             have access to the innards of this structure.

27/11/2000 (Release of version 1.1g)

           getline.c man/gl_get_line.3
             I added a "read-from-file" action function and bound it
             by default to M-<. This causes gl_get_line() to
             temporarily return input from the file who's name
             precedes the cursor.

           getline.c keytab.c keytab.h man/gl_get_line.3
             I have reworked some of the keybinding code again.

             Now, within key binding strings, in addition to the
             previously existing notation, you can now use M-a to
             denote meta-a, and C-a to denote control-a. For example,
             a key binding which triggers when the user presses the
             meta key, the control key and the letter [
             simultaneously, can now be denoted by M-C-[, or M-^[ or
             \EC-[ or \E^[.

             I also updated the man page to use M- instead of \E in
             the list of default bindings, since this looks cleaner.

           getline.c man/gl_get_line.3
             I added a copy-region-as-kill action function and
             gave it a default binding to M-w.


             Markus Gyger sent me a copy of a previous version of
             the library, with const qualifiers added in appropriate
             places. I have done the same for the latest version.
             Among other things, this gets rid of the warnings
             that are generated if one tells the compiler to
             const qualify literal strings.

           getline.c getline.h glconf.c
             I have moved the contents of glconf.c and the declaration
             of the GetLine structure into getline.c. This is cleaner,
             since now only functions in getline.c can mess with the
             innards of GetLine objects. It also clears up some problems
             with system header inclusion order under Solaris, and also
             the possibility that this might result in inconsistent
             system macro definitions, which in turn could cause different
             declarations of the structure to be seen in different files.

             I wrote a wrapper function to go around strcmp(), such that
             when hash.c is compiled with a C++ compiler, the pointer
             to the wrapper function is a C++ function pointer.
             This makes it compatible with comparison function pointer
             recorded in the hash table.

           cplmatch.c getline.c libtecla.h
             Markus noted that the Sun C++ compiler wasn't able to
             match up the declaration of cpl_complete_word() in
             libtecla.h, where it is surrounded by a extern "C" {}
             wrapper, with the definition of this function in
             cplmatch.c. My suspicion is that the compiler looks not
             only at the function name, but also at the function
             arguments to see if two functions match, and that the
             match_fn() argument, being a fully blown function pointer
             declaration, got interpetted as that of a C function in
             one case, and a C++ function in the other, thus
             preventing a match.

             To fix this I now define a CplMatchFn typedef in libtecla.h,
             and use this to declare the match_fn callback.

20/11/2000 (Changes suggested by Markus Gyger to support C++ compilers):
             Renamed a variable called "explicit" to "xplicit", to
             avoid conflicts when compiling with C++ compilers.
             Added explicit casts when converting from (void *) to
             other pointer types. This isn't needed in C but it is
             in C++.
             tputs() has a strange declaration under Solaris. I was
             enabling this declaration when the SPARC feature-test
             macro was set. Markus changed the test to hinge on the
             __sun and __SVR4 macros.
           direader.c glconf.c stringrp.c
             I had omitted to include string.h in these two files.

           Markus also suggested some other changes, which are still
           under discussion. With the just above changes however, the
           library compiles without complaint using g++.

           getline.h getline.c keytab.c keytab.h glconf.c
             I added support for backslash escapes (include \e
             for the keyboard escape key) and literal binary
             characters to the characters allowed within key sequences
             of key bindings.

           getline.h getline.c keytab.c keytab.h glconf.c
             I introduced symbolic names for the arrow keys, and
             modified the library to use the cursor key sequences
             reported by terminfo/termcap in addition to the default
             ANSI ones. Anything bound to the symbolically named arrow
             keys also gets bound to the default and terminfo/termcap
             cursor key sequences. Note that under Solaris
             terminfo/termcap report the properties of hardware X
             terminals when TERM is xterm instead of the terminal
             emulator properties, and the cursor keys on these two
             systems generate different key sequences. This is an
             example of why extra default sequences are needed.

           getline.h getline.c keytab.c
             For some reason I was using \e to represent the escape
             character. This is supported by gcc, which thus doesn't
             emit a warning except with the -pedantic flag, but isn't
             part of standard C. I now use a macro to define escape
             as \033 in getline.h, and this is now used wherever the
             escape character is needed.

17/11/2000 (Release of version 1.1d)

           getline.c, man/gl_get_line(3), html/gl_get_line.html
             In tcsh ^D is bound to a function which does different
             things depending on where the cursor is within the input
             line. I have implemented its equivalent in the tecla
             library. When invoked at the end of the line this action
             function displays possible completions. When invoked on
             an empty line it causes gl_get_line() to return NULL,
             thus signalling end of input. When invoked within a line
             it invokes forward-delete-char, as before. The new action
             function is called del-char-or-list-or-eof.

           getline.c, man/gl_get_line(3), html/gl_get_line.html
             I found that the complete-word and expand-file actions
             had underscores in their names instead of hyphens. This
             made them different from all other action functions, so I
             have changed the underscores to hyphens.

             On SCO UnixWare while getpwuid_r() is available, the
             associated _SC_GETPW_R_SIZE_MAX macro used by sysconf()
             to find out how big to make the buffer to pass to this
             function to cater for any password entry, doesn't
             exist. I also hadn't catered for the case where sysconf()
             reports that this limit is indeterminate. I have thus
             change the code to substitute a default limit of 1024 if
             either the above macro isn't defined or if sysconf() says
             that the associated limit is indeterminate.
17/11/2000 (Release of version 1.1c)

           getline.c, getline.h, history.c, history.h
             I have modified the way that the history recall functions
             operate, to make them better emulate the behavior of
             tcsh. Previously the history search bindings always
             searched for the prefix that preceded the cursor, then
             left the cursor at the same point in the line, so that a
             following search would search using the same prefix. This
             isn't how tcsh operates. On finding a matching line, tcsh
             puts the cursor at the end of the line, but arranges for
             the followup search to continue with the same prefix,
             unless the user does any cursor motion or character
             insertion operations in between, in which case it changes
             the search prefix to the new set of characters that are
             before the cursor. There are other complications as well,
             which I have attempted to emulate. As far as I can
             tell, the tecla history recall facilities now fully
             emulate those of tcsh.

16/11/2000 (Release of version 1.1b)

             One can now quit from the demo by typing exit.

             The first entry of the table was getting deleted
             by _kt_clear_bindings() regardless of the source
             of the binding. This deleted the up-arrow binding.
             Symptoms noted by

             Depending on which system include files were include
             before the inclusion of getline.h, SIGWINCH and
             TIOCGWINSZ might or might not be defined. This resulted
             in different definitions of the GetLine object in
             different files, and thus some very strange bugs! I have
             now added #includes for the necessary system header files
             in getline.h itself. The symptom was that on creating a
             ~/.teclarc file, the demo program complained of a NULL
             argument to kt_set_keybinding() for the first line of the

15/11/2000 (Release of version 1.1a)

             I had neglected to check the return value of
             new_GetLine() in the demo program. Oops.

           getline.c libtecla.h:
             I wrote gl_change_terminal(). This allows one to change to
             a different terminal or I/O stream, by specifying the
             stdio streams to use for input and output, along with the
             type of terminal that they are connected to.

           getline.c libtecla.h:
             Renamed GetLine::isterm to GetLine::is_term. Standard
             C reserves names that start with "is" followed by
             alphanumeric characters, so this avoids potential
             clashes in the future.

           keytab.c keytab.h
             Each key-sequence can now have different binding
             functions from different sources, with the user provided
             binding having the highest precedence, followed by the
             default binding, followed by any terminal specific
             binding. This allows gl_change_terminal() to redefine the
             terminal-specific bindings each time that
             gl_change_terminal() is called, without overwriting the
             user specified or default bindings. In the future, it will
             also allow for reconfiguration of user specified
             bindings after the call to new_GetLine(). Ie. deleting a
             user specified binding should reinstate any default or
             terminal specific binding.

           man/cpl_complete_word.3 html/cpl_complete_word.html
           man/ef_expand_file.3    html/ef_expand_file.html
           man/gl_get_line.3       html/gl_get_line.html
             I added sections on thread safety to the man pages of the
             individual modules.

           man/gl_get_line.3       html/gl_get_line.html
             I documented the new gl_change_terminal() function.

           man/gl_get_line.3       html/gl_get_line.html
             In the description of the ~/.teclarc configuration file,
             I had omitted the 'bind' command word in the example
             entry. I have now remedied this.