Merge commit 'db1c88f6dab43484b6c33636600ac4596ff4c354'

[unleashed.git] / bin / ksh / NOTES

$OpenBSD: NOTES,v 1.16 2018/01/12 14:20:57 jca Exp $

General features of at&t ksh88 that are not (yet) in pdksh:
    - exported aliases and functions (not in ksh93).
    - set -t.
    - signals/traps not cleared during functions.
    - trap DEBUG, local ERR and EXIT traps in functions.
    - ERRNO parameter.
    - read/select aren't hooked in to the command line editor
    - the last command of a pipeline is not run in the parent shell

Known bugs (see also PROJECTS files):
    Variable parsing, Expansion:
        - some specials behave differently when unset (eg, IFS behaves like
          " \t\n") others lose their special meaning.  IFS/PATH taken care of,
          still need to sort out some others (eg, TMOUT).
    Parsing,Lexing:
        - line numbers in errors are wrong for nested constructs.  Need to
          keep track of the line a command started on (can use for LINENO
          parameter as well).
        - a $(..) expression nested inside double quotes inside another $(..)
          isn't parsed correctly (eg, $(echo "foo$(echo ")")") )
    Commands,Execution:
        - setting special parameters that have side effects when
          changed/restored (ie, HISTFILE, OPTIND, RANDOM) in front
          of a command (eg, HISTFILE=/foo/bar echo hi) affects the parent
          shell.  Note that setting other (not so special) parameters
          does not affect the parent shell.
        - `echo hi | exec cat -n' causes at&t to exit, `exec echo hi | cat -n'
          does not.  pdksh exits for neither.  Don't think POSIX requires
          an exit, but not sure.
        - `echo foo | read bar; echo $bar' prints foo in at&t ksh, nothing
          in pdksh (ie, the read is done in a separate process in pdksh).

Known problems not caused by ksh:
    - after stoping a job, emacs/vi is not re-entered.  Hitting return
      prints the prompt and everything is fine again.  Problem (often
      involving a pager like less) is related to order of process
      scheduling (shell runs before `stop'ed (sub) processes have had a chance
      to clean up the screen/terminal).

Known differences between pdksh & at&t ksh (that may change)
    - vi:
        - `^U': at&t: kills only what has been inserted, pdksh: kills to
          start of line
    - at&t ksh login shells say "Warning: you have running jobs" if you
      try to exit when there are running jobs.  An immediate second attempt
      to exit will kill the jobs and exit.  pdksh does not print a warning,
      nor does it kill running jobs when it exits (it does warn/kill for
      stopped jobs).
    - TMOUT: at&t prints warning, then waits another 60 seconds.  If on screwed
      up serial line, the output could cause more input, so pdksh just
      prints a message and exits.  (Also, in at&t ksh, setting TMOUT has no
      effect after the sequence "TMOUT=60; unset TMOUT", which could be
      useful - pdksh may do this in the future).
    - in pdksh, if the last command of a pipeline is a shell builtin, it is
      not executed in the parent shell, so "echo a b | read foo bar" does not
      set foo and bar in the parent shell (at&t ksh will).
      This may get fixed in the future, but it may take a while.
    - in pdksh, set +o lists the options that are currently set, in at&t ksh
      it is the same as set -o.
    - in pdksh emacs mode, ^T does what gnu emacs does, not what at&t ksh
      does.
    - in ksh93, `. name' calls a function (defined with function) with POSIX
      semantics (instead of ksh semantics).  in pdksh, . does not call
      functions.
    - test: "test -f foo bar blah" is the same as "test -f foo" (the extra
      arguments, of which there must be at least 2, are ignored) - pdksh
      generates an error message (unexpected operator/operand "bar") as it
      should.  Sometimes used to test file globs (e.g., if test -f *.o; ...).
    - if the command 'sleep 5 && /bin/echo blah' is run interactively and
      is the sleep is stopped (^Z), the echo is run immediately in pdksh.
      In at&t ksh, the whole thing is stopped.
    - LINENO:
        - in ksh88 variable is always 1 (can't be changed) in interac mode;
          in pdksh it changes.
        - Value of LINENO after it has been set by the script in one file
          is bizarre when used in another file.

Known differences between pdksh & at&t ksh (that are not likely to change)
    - at&t ksh seems to catch or ignore SIGALRM - pdksh dies upon receipt
      (unless it's traped of course)
    - typeset:
        - at&t ksh overloads -u/-l options: for integers, means unsigned/long,
          for strings means uppercase/lowercase; pdksh just has the
          upper/lower case (which can be useful for integers when base > 10).
          unsigned/long really should have their own options.
        - at&t ksh can't have justified integer variables
          (eg, typeset -iR5 j=10), pdksh can.
        - in pdksh, number arguments for -L/-R/-Z/-i must follow the option
          character, at&t allows it at the end of the option group (eg,
          at&t ksh likes "typeset -iu5 j", pdksh wants "typeset -i5 -u j"
          or "typeset -ui5 j").  Also, pdksh allows "typeset -i 5 j" (same
          as "typeset -i5 j"), at&t ksh does not allow this.
        - typeset -R: pdksh strips trailing space type characters (ie,
          uses isspace()), at&t ksh only skips blanks.
        - at&t ksh allows attributes of read-only variables to be changed,
          pdksh allows only the export attribute to be set.
    - (some) at&t ksh allows set -A of readonly variables, pdksh does not.
    - at&t ksh allows command assignments of readonly variables (eg, YY=2 cat),
      pdksh does not.
    - at&t ksh does not exit scripts when an implicit assignment to an integer
      variable fails due to an expression error: eg,
                echo 2+ > /tmp/x
                unset x; typeset -i x
                read x < /tmp/x
                echo still here
      prints an error and then prints "still here", similarly for
                unset x; typeset -i x
                set +A x 1 2+ 3
                echo still here
      and
                unset x y; typeset -i x y; set +A y 10 20 30
                set +A x 1 1+y[2+] 3
                echo still here
      pdksh exits a script in all the above cases. (note that both shells
      exit for:
                unset x; typeset -i x
                for x in 1 2+ 3; do echo x=$x; done
                echo still here
      ).
    - at&t ksh seems to allow function calls inside expressions
      (eg, typeset -i x='y(2)') but they do not seem to be regular functions
      nor math functions (eg, pow, exp) - anyone known anything about this?
    - `set -o nounset; unset foo; echo ${#foo}`: at&t ksh prints 0; pdksh
      generates error.  Same for ${#foo[*]} and ${#foo[@]}.
    - . file: at&t ksh parses the whole file before executing anything,
      pdksh executes as it parses.  This means aliases defined in the file
      will affect how pdksh parses the file, but won't affect how at&t ksh
      parses the file.  Also means pdksh will not parse statements occurring
      after a (executed) return statement.
    - a return in $ENV in at&t ksh will cause the shell to exit, while in
      pdksh it will stop executing the script (this is consistent with
      what a return in .profile does in both shells).
    - at&t ksh does file globbing for `echo "${foo:-"*"}"`, pdksh does not
      (POSIX would seem to indicate pdksh is right).
    - at&t ksh thinks ${a:##foo} is ok, pdksh doesn't.
    - at&t does tilde expansion on here-document delimiters, pdksh does
      not.  eg.
        $ cat << ~michael
        ~michael
        $
      works for pdksh, not for at&t ksh (POSIX seems to agree with pdksh).
    - in at&t ksh, tracked aliases have the export flag implicitly set
      and tracked aliases and normal aliases live in the same name space
      (eg, "alias" will list both tracked and normal aliases).
      in pdksh, -t does not imply -x (since -x doesn't do anything yet), and
      tracked/normal aliases live in separate name spaces.
      in at&t ksh, alias accepts + options (eg, +x, +t) - pdksh does not.
      in pdksh, alias has a -d option to allow examination/changing of
      cached ~ entries, also unalias has -d and -t options (unalias -d
      is useful if the ~ cache gets out of date - not sure how at&t deals
      with this problem (it does cache ~ entries)).
    - at&t ksh will stop a recursive function after about 60 calls; pdksh
      will not since the limit is arbitrary and can't be controlled
      by the user (hit ^C if you get in trouble).
    - the wait command (with and without arguments) in at&t ksh will wait for
      stopped jobs when job control is enabled.  pdksh doesn't.
    - at&t ksh automatically sets the bgnice option for interactive shells;
      pdksh does not.
    - in at&t ksh, "eval `false`; echo $?" prints 1, pdksh prints 0 (which
      is what POSIX says it should).  Same goes for "wait `false`; echo $?".
      (same goes for "set `false`; echo $?" if posix option is set - some
      scripts that use the old getopt depend on this, so be careful about
      setting the posix option).
    - in at&t ksh, print -uX and read -uX are interrperted as -u with no
      argument (defaults to 1 and 0 respectively) and -X (which may or
      may not be a valid flag).  In pdksh, -uX is interpreted as file
      descriptor X.
    - in at&t ksh, some signals (HUP, INT, QUIT) cause the read to exit, others
      (ie, everything else) do not.  When it does cause exiting, anything read
      to that point is used (usually an empty line) and read returns with 0
      status.  pdksh currently does similar things, but for TERM as well and
      the exit status is 128+<signal-number> - in future, pdksh's read will
      do this for all signals that are normally fatal as required by POSIX.
      (POSIX does not require the setting of variables to null so applications
      shouldn't rely on this).
    - in pdksh, ! substitution done before variable substitution; in at&t ksh
      it is done after substitution (and therefore may do ! substitutions on
      the result of variable substitutions).  POSIX doesn't say which is to be
      done.
    - pwd: in at&t ksh, it ignores arguments; in pdksh, it complains when given
      arguments.
    - the at&t ksh does not do command substition on PS1, pdksh does.
    - ksh93 allows ". foo" to run the function foo if there is no file
      called foo (go figure).
    - field splitting (IFS): ksh88/ksh93 strip leading non-white space IFS
      chars, pdksh (and POSIX, I think) leave them intact. e.g.
        $ IFS="$IFS:"; read x; echo "<$x>"
        ::
      prints "<>" in at&t ksh, "<::>" in pdksh.
    - command completion: at&t ksh will do completion on a blank line (matching
      all commands), pdksh does not (as this isn't very useful - use * if
      you really want the list).
    - co-processes: if ksh93, the write portion of the co-process output is
      closed when the most recently started co-process exits. pdksh closes
      it when all the co-processes using it have exited.
    - pdksh accepts empty command lists for while and for statements, while
      at&t ksh (and sh) don't.  Eg., pdksh likes
        while false ; do done
      but ksh88 doesn't like it.
    - pdksh bumps RANDOM in parent after a fork, at&t ksh bumps it in both
      parent and child:
        RANDOM=1
        echo child: `echo $RANDOM`
        echo parent: $RANDOM
      will produce "child: 16838 parent: 5758" in pdksh, while at&t ksh
      will produce "child: 5758 parent: 5758".

Oddities in ksh (pd & at&t):
    - array references inside (())/$(()) are strange:
          $(( x[2] )) does the expected, $(( $x[2] )) doesn't.
    - `typeset -R3 X='x '; echo "($X)"` produces (  x) - trailing
      spaces are stripped.
    - typeset -R turns off Z flag.
    - both shells have the following mis-feature:
        $ x='function xx {
                cat -n <<- EOF
                here we are in xx
                EOF
                }'
        $ (eval "$x"; (sleep 2; xx) & echo bye)
        [1] 1234
        bye
        $ xx: /tmp/sh1234.1: cannot open
    - bizarre special handling of alias/export/readonly/typeset arguments
        $ touch a=a; typeset a=[ab]; echo "$a"
        a=[ab]
        $ x=typeset; $x a=[ab]; echo "$a"
        a=a
        $
    - both ignore SIGTSTP,SIGTTIN,SIGTTOU in exec'd processes when talking
      and not monitoring (at&t ksh kind of does this).  Doesn't really make
      sense.
      (Note that ksh.att -ic 'set +m; check-sigs' shows TSTP et al aren't
       ignored, while ksh.att -ic 'set +m^J check-sigs' does... very strange)
    - when tracing (set -x), and a command's stderr is redirected, the trace
      output is also redirected. so "set -x; echo foo 2> /tmp/O > /dev/null"
      will create /tmp/foo with the lines "+ > /dev/null" and "+ echo foo".
    - undocumented at&t ksh88, documented in ksh93: FPATH is searched
      after PATH if no executable is found, even if typeset -uf wasn't used.

POSIX sh questions (references are to POSIX 1003.2-1992)
        - arithmetic expressions: how are empty expressions treated?
          (eg, echo $((  ))).  at&t ksh (and now pdksh) echo 0.
          Same question goes for `test "" -eq 0' - does this generate an error
          or, if not, what is the exit code?
        - if a signal is received during the execution of a built-in,
          does the builtin command exit or the whole shell?
        - is it legal to execute last command of pipeline in current
          execution environment (eg, can "echo foo | read bar" set
          bar?)
        - what action should be taken if there is an error doing a dup due
          to system limits (eg, not enough file destriptors): is this
          a "redirection error" (in which case a script will exit iff the
          error occured while executing a special built-in)?
          IMHO, shell should exit script.  Couldn't find a blanket statement
          like "if shell encounters an unexpected system error, it shall
          exit non-interactive scripts"...

POSIX sh bugs (references are to POSIX 1003.2-1992)
        - in vi insert mode, ^W deletes to beginning of line or to the first
          blank/punct character (para at line 9124, section 3).  This means
          "foo     ^W" will do nothing.  This is inconsistent with the vi
          spec, which says delete preceding word including and interceding
          blanks (para at line 5189, section 5).
        - parameter expansion, section 3.6.2, line 391: `in each case that a
          value of word is needed (..), word shall be subjected to tilde
          expansion, parameter expansion, ...'.  Various expansions should not
          be performed if parameter is in double quotes.
        - the getopts description says assigning OPTIND a value other than 1
          produces undefined results, while the rationale for getopts suggests
          saving/restoring the OPTIND value inside functions (since POSIX
          functions don't do the save/restore automatically).  Restoring
          OPTIND is kind of dumb since getopts may have been in the middle
          of parsing a group of flags (eg, -abc).
        - `...` definition (3.6.3) says nothing about backslash followed by
          a newline, which sh and at&t ksh strip out completely.  e.g.,
                $ show-args `echo 'X
                Y'`
                Number of args: 1
                        1: <XY>
                $
          POSIX would indicate the backslash-newline would be preserved.
        - does not say how "cat << ''" is to be treated (illegal, read 'til
          blank line, or read 'til eof).  at&t ksh reads til eof, bourne shell
          reads 'til blank line.  pdksh reads 'til blank line.