C.1.14 Limitations of Shell Builtins

No, no, we are serious: some shells do have limitations! :)

You should always keep in mind that any builtin or command may support options, and therefore differ in behavior with arguments starting with a dash. For instance, even the innocent ‘echo "$word"’ can give unexpected results when word starts with a dash. It is often possible to avoid this problem using ‘echo "x$word"’, taking the ‘x’ into account later in the pipe. Many of these limitations can be worked around using M4sh (see Programming in M4sh).

.

Use . only with regular files (use ‘test -f’). Bash 2.03, for instance, chokes on ‘. /dev/null’. Remember that . uses PATH if its argument contains no slashes. Also, some shells, including bash 3.2, implicitly append the current directory to this PATH search, even though Posix forbids it. So if you want to use . on a file foo in the current directory, you must use ‘. ./foo’.

Not all shells gracefully handle syntax errors within a sourced file. On one extreme, some non-interactive shells abort the entire script. On the other, zsh 4.3.10 has a bug where it fails to react to the syntax error.

$ echo 'fi' > syntax
$ bash -c '. ./syntax; echo $?'
./syntax: line 1: syntax error near unexpected token `fi'
./syntax: line 1: `fi'
1
$ ash -c '. ./syntax; echo $?'
./syntax: 1: Syntax error: "fi" unexpected
$ zsh -c '. ./syntax; echo $?'
./syntax:1: parse error near `fi'
0

!

The Unix version 7 shell did not support negating the exit status of commands with !, and this feature is still absent from some shells (e.g., Solaris /bin/sh). Other shells, such as FreeBSD /bin/sh or ash, have bugs when using !:

$ sh -c '! : | :'; echo $?
1
$ ash -c '! : | :'; echo $?
0
$ sh -c '! { :; }'; echo $?
1
$ ash -c '! { :; }'; echo $?
{: not found
Syntax error: "}" unexpected
2

Shell code like this:

if ! cmp file1 file2 >/dev/null 2>&1; then
  echo files differ or trouble
fi

is therefore not portable in practice. Typically it is easy to rewrite such code, e.g.:

cmp file1 file2 >/dev/null 2>&1 ||
  echo files differ or trouble

More generally, one can always rewrite ‘! command’ as:

if command; then (exit 1); else :; fi

{...}

Bash 3.2 (and earlier versions) sometimes does not properly set ‘$?’ when failing to write redirected output of a compound command. This problem is most commonly observed with ‘{…}’; it does not occur with ‘(…)’. For example:

$ bash -c '{ echo foo; } >/bad; echo $?'
bash: line 1: /bad: Permission denied
0
$ bash -c 'while :; do echo; done >/bad; echo $?'
bash: line 1: /bad: Permission denied
0

To work around the bug, prepend ‘:;’:

$ bash -c ':;{ echo foo; } >/bad; echo $?'
bash: line 1: /bad: Permission denied
1

Posix requires a syntax error if a brace list has no contents. However, not all shells obey this rule; and on shells where empty lists are permitted, the effect on ‘$?’ is inconsistent. To avoid problems, ensure that a brace list is never empty.

$ bash -c 'false; { }; echo $?' || echo $?
bash: line 1: syntax error near unexpected token `}'
bash: line 1: `false; { }; echo $?'
2
$ zsh -c 'false; { }; echo $?' || echo $?
1
$ pdksh -c 'false; { }; echo $?' || echo $?
0

break

The use of ‘break 2’ etc. is safe.

case

You don’t need to quote the argument; no splitting is performed.

You don’t need the final ‘;;’, but you should use it.

Posix requires support for case patterns with opening parentheses like this:

case $file_name in
  (*.c) echo "C source code";;
esac

but the ( in this example is not portable to many Bourne shell implementations, which is a pity for those of us using tools that rely on balanced parentheses. For instance, with Solaris /bin/sh:

$ case foo in (foo) echo foo;; esac
error→syntax error: `(' unexpected

The leading ‘(’ can be omitted safely. Unfortunately, there are contexts where unbalanced parentheses cause other problems, such as when using a syntax-highlighting editor that searches for the balancing counterpart, or more importantly, when using a case statement as an underquoted argument to an Autoconf macro. See Balancing Parentheses, for tradeoffs involved in various styles of dealing with unbalanced ‘)’.

Zsh handles pattern fragments derived from parameter expansions or command substitutions as though quoted:

$ pat=\?; case aa in ?$pat) echo match;; esac
$ pat=\?; case a? in ?$pat) echo match;; esac
match

Because of a bug in its fnmatch, Bash fails to properly handle backslashes in character classes:

bash-2.02$ case /tmp in [/\\]*) echo OK;; esac
bash-2.02$

This is extremely unfortunate, since you are likely to use this code to handle Posix or MS-DOS absolute file names. To work around this bug, always put the backslash first:

bash-2.02$ case '\TMP' in [\\/]*) echo OK;; esac
OK
bash-2.02$ case /tmp in [\\/]*) echo OK;; esac
OK

Many Bourne shells cannot handle closing brackets in character classes correctly.

Some shells also have problems with backslash escaping in case you do not want to match the backslash: both a backslash and the escaped character match this pattern. To work around this, specify the character class in a variable, so that quote removal does not apply afterwards, and the special characters don’t have to be backslash-escaped:

$ case '\' in [\<]) echo OK;; esac
OK
$ scanset='[<]'; case '\' in $scanset) echo OK;; esac
$

Even with this, Solaris ksh matches a backslash if the set contains any of the characters ‘|’, ‘&’, ‘(’, or ‘)’.

Conversely, Tru64 ksh (circa 2003) erroneously always matches a closing parenthesis if not specified in a character class:

$ case foo in *\)*) echo fail ;; esac
fail
$ case foo in *')'*) echo fail ;; esac
fail

Some shells, such as Ash 0.3.8, are confused by an empty case/esac:

ash-0.3.8 $ case foo in esac;
error→Syntax error: ";" unexpected (expecting ")")

Posix requires case to give an exit status of 0 if no cases match. However, /bin/sh in Solaris 10 does not obey this rule. Meanwhile, it is unclear whether a case that matches, but contains no statements, must also change the exit status to 0. The M4sh macro AS_CASE works around these inconsistencies.

$ bash -c 'case `false` in ?) ;; esac; echo $?'
0
$ /bin/sh -c 'case `false` in ?) ;; esac; echo $?'
255

cd

Posix 1003.1-2001 requires that cd must support the -L (“logical”) and -P (“physical”) options, with -L being the default. However, traditional shells do not support these options, and their cd command has the -P behavior.

Portable scripts should assume neither option is supported, and should assume neither behavior is the default. This can be a bit tricky, since the Posix default behavior means that, for example, ‘ls ..’ and ‘cd ..’ may refer to different directories if the current logical directory is a symbolic link. It is safe to use cd dir if dir contains no .. components. Also, Autoconf-generated scripts check for this problem when computing variables like ac_top_srcdir (see Configuration Actions), so it is safe to cd to these variables.

Posix states that behavior is undefined if cd is given an explicit empty argument. Some shells do nothing, some change to the first entry in CDPATH, some change to HOME, and some exit the shell rather than returning an error. Unfortunately, this means that if ‘$var’ is empty, then ‘cd "$var"’ is less predictable than ‘cd $var’ (at least the latter is well-behaved in all shells at changing to HOME, although this is probably not what you wanted in a script). You should check that a directory name was supplied before trying to change locations.

See Special Shell Variables, for portability problems involving cd and the CDPATH environment variable. Also please see the discussion of the pwd command.

echo

The simple echo is probably the most surprising source of portability troubles. It is not possible to use ‘echo’ portably unless both options and escape sequences are omitted. Don’t expect any option.

Do not use backslashes in the arguments, as there is no consensus on their handling. For ‘echo '\n' | wc -l’, the sh of Solaris outputs 2, but Bash and Zsh (in sh emulation mode) output 1. The problem is truly echo: all the shells understand ‘'\n'’ as the string composed of a backslash and an ‘n’. Within a command substitution, ‘echo 'string\c'’ will mess up the internal state of ksh88 on AIX 6.1 so that it will print the first character ‘s’ only, followed by a newline, and then entirely drop the output of the next echo in a command substitution.

Because of these problems, do not pass a string containing arbitrary characters to echo. For example, ‘echo "$foo"’ is safe only if you know that foo’s value cannot contain backslashes and cannot start with ‘-’.

Normally, printf is safer and easier to use than echo and echo -n. Thus, you should use printf "%s\n" instead of echo, and similarly use printf %s instead of echo -n.

Older scripts, written before printf was portable, sometimes used a here-document as a safer alternative to echo, like this:

cat <<EOF
$foo
EOF

eval

The eval command is useful in limited circumstances, e.g., using commands like ‘eval table_$key=\$value’ and ‘eval value=table_$key’ to simulate a hash table when the key is known to be alphanumeric.

You should also be wary of common bugs in eval implementations. In some shell implementations (e.g., older ash, OpenBSD 3.8 sh, pdksh v5.2.14 99/07/13.2, and zsh 4.2.5), the arguments of ‘eval’ are evaluated in a context where ‘$?’ is 0, so they exhibit behavior like this:

$ false; eval 'echo $?'
0

The correct behavior here is to output a nonzero value, but portable scripts should not rely on this.

You should not rely on LINENO within eval. See Special Shell Variables.

Note that, even though these bugs are easily avoided, eval is tricky to use on arbitrary arguments. It is obviously unwise to use ‘eval $cmd’ if the string value of ‘cmd’ was derived from an untrustworthy source. But even if the string value is valid, ‘eval $cmd’ might not work as intended, since it causes field splitting and file name expansion to occur twice, once for the eval and once for the command itself. It is therefore safer to use ‘eval "$cmd"’. For example, if cmd has the value ‘cat test?.c’, ‘eval $cmd’ might expand to the equivalent of ‘cat test;.c’ if there happens to be a file named test;.c in the current directory; and this in turn mistakenly attempts to invoke cat on the file test and then execute the command .c. To avoid this problem, use ‘eval "$cmd"’ rather than ‘eval $cmd’.

However, suppose that you want to output the text of the evaluated command just before executing it. Assuming the previous example, ‘echo "Executing: $cmd"’ outputs ‘Executing: cat test?.c’, but this output doesn’t show the user that ‘test;.c’ is the actual name of the copied file. Conversely, ‘eval "echo Executing: $cmd"’ works on this example, but it fails with ‘cmd='cat foo >bar'’, since it mistakenly replaces the contents of bar by the string ‘cat foo’. No simple, general, and portable solution to this problem is known.

exec

Posix describes several categories of shell built-ins. Special built-ins (such as exit) must impact the environment of the current shell, and need not be available through exec. All other built-ins are regular, and must not propagate variable assignments to the environment of the current shell. However, the group of regular built-ins is further distinguished by commands that do not require a PATH search (such as cd), in contrast to built-ins that are offered as a more efficient version of something that must still be found in a PATH search (such as echo). Posix is not clear on whether exec must work with the list of 17 utilities that are invoked without a PATH search, and many platforms lack an executable for some of those built-ins:

$ sh -c 'exec cd /tmp'
sh: line 0: exec: cd: not found

All other built-ins that provide utilities specified by Posix must have a counterpart executable that exists on PATH, although Posix allows exec to use the built-in instead of the executable. For example, contrast bash 3.2 and pdksh 5.2.14:

$ bash -c 'pwd --version' | head -n1
bash: line 0: pwd: --: invalid option
pwd: usage: pwd [-LP]
$ bash -c 'exec pwd --version' | head -n1
pwd (GNU coreutils) 6.10
$ pdksh -c 'exec pwd --version' | head -n1
pdksh: pwd: --: unknown option

When it is desired to avoid a regular shell built-in, the workaround is to use some other forwarding command, such as env or nice, that will ensure a path search:

$ pdksh -c 'exec true --version' | head -n1
$ pdksh -c 'nice true --version' | head -n1
true (GNU coreutils) 6.10
$ pdksh -c 'env true --version' | head -n1
true (GNU coreutils) 6.10

exit

The default value of exit is supposed to be $?; unfortunately, some shells, such as the DJGPP port of Bash 2.04, just perform ‘exit 0’.

bash-2.04$ foo=`exit 1` || echo fail
fail
bash-2.04$ foo=`(exit 1)` || echo fail
fail
bash-2.04$ foo=`(exit 1); exit` || echo fail
bash-2.04$

Using ‘exit $?’ restores the expected behavior.

Some shell scripts, such as those generated by autoconf, use a trap to clean up before exiting. If the last shell command exited with nonzero status, the trap also exits with nonzero status so that the invoker can tell that an error occurred.

Unfortunately, in some shells, such as Solaris /bin/sh, an exit trap ignores the exit command’s argument. In these shells, a trap cannot determine whether it was invoked by plain exit or by exit 1. Instead of calling exit directly, use the AC_MSG_ERROR macro that has a workaround for this problem.

export

The builtin export dubs a shell variable environment variable. Each update of exported variables corresponds to an update of the environment variables. Conversely, each environment variable received by the shell when it is launched should be imported as a shell variable marked as exported.

Alas, many shells, such as Solaris /bin/sh, IRIX 6.3, IRIX 5.2, AIX 4.1.5, and Digital Unix 4.0, forget to export the environment variables they receive. As a result, two variables coexist: the environment variable and the shell variable. The following code demonstrates this failure:

#!/bin/sh
echo $FOO
FOO=bar
echo $FOO
exec /bin/sh $0

when run with ‘FOO=foo’ in the environment, these shells print alternately ‘foo’ and ‘bar’, although they should print only ‘foo’ and then a sequence of ‘bar’s.

Therefore you should export again each environment variable that you update; the export can occur before or after the assignment.

Posix is not clear on whether the export of an undefined variable causes the variable to be defined with the value of an empty string, or merely marks any future definition of a variable by that name for export. Various shells behave differently in this regard:

$ sh -c 'export foo; env | grep foo'
$ ash -c 'export foo; env | grep foo'
foo=

Posix requires export to honor assignments made as arguments, but older shells do not support this, including /bin/sh in Solaris 10. Portable scripts should separate assignments and exports into different statements.

$ bash -c 'export foo=bar; echo $foo'
bar
$ /bin/sh -c 'export foo=bar; echo $foo'
/bin/sh: foo=bar: is not an identifier
$ /bin/sh -c 'export foo; foo=bar; echo $foo'
bar

false

Don’t expect false to exit with status 1: in native Solaris /bin/false exits with status 255.

for

To loop over positional arguments, use:

for arg
do
  echo "$arg"
done

You may not leave the do on the same line as for, since some shells improperly grok:

for arg; do
  echo "$arg"
done

If you want to explicitly refer to the positional arguments, given the ‘$@’ bug (see Shell Substitutions), use:

for arg in ${1+"$@"}; do
  echo "$arg"
done

But keep in mind that Zsh, even in Bourne shell emulation mode, performs word splitting on ‘${1+"$@"}’; see Shell Substitutions, item ‘$@’, for more.

In Solaris /bin/sh, when the list of arguments of a for loop starts with unquoted tokens looking like variable assignments, the loop is not executed on those tokens:

$ /bin/sh -c 'for v in a=b c=d x e=f; do echo $v; done'
x
e=f

Thankfully, quoting the assignment-like tokens, or starting the list with other tokens (including unquoted variable expansion that results in an assignment-like result), avoids the problem, so it is easy to work around:

$ /bin/sh -c 'for v in "a=b"; do echo $v; done'
a=b
$ /bin/sh -c 'x=a=b; for v in $x c=d; do echo $v; done'
a=b
c=d

if

Using ‘!’ is not portable. Instead of:

if ! cmp -s file file.new; then
  mv file.new file
fi

use:

if cmp -s file file.new; then :; else
  mv file.new file
fi

Or, especially if the else branch is short, you can use ||. In M4sh, the AS_IF macro provides an easy way to write these kinds of conditionals:

AS_IF([cmp -s file file.new], [], [mv file.new file])

This is especially useful in other M4 macros, where the then and else branches might be macro arguments.

Some very old shells did not reset the exit status from an if with no else:

$ if (exit 42); then true; fi; echo $?
42

whereas a proper shell should have printed ‘0’. But this is no longer a portability problem; any shell that supports functions gets it correct. However, it explains why some makefiles have lengthy constructs:

if test -f "$file"; then
  install "$file" "$dest"
else
  :
fi

printf

A format string starting with a ‘-’ can cause problems. Bash interprets it as an option and gives an error. And ‘--’ to mark the end of options is not good in the NetBSD Almquist shell (e.g., 0.4.6) which takes that literally as the format string. Putting the ‘-’ in a ‘%c’ or ‘%s’ is probably easiest:

printf %s -foo

Bash 2.03 mishandles an escape sequence that happens to evaluate to ‘%’:

$ printf '\045'
bash: printf: `%': missing format character

Large outputs may cause trouble. On Solaris 2.5.1 through 10, for example, /usr/bin/printf is buggy, so when using /bin/sh the command ‘printf %010000x 123’ normally dumps core.

Since printf is not always a shell builtin, there is a potential speed penalty for using printf '%s\n' as a replacement for an echo that does not interpret ‘\’ or leading ‘-’. With Solaris ksh, it is possible to use print -r -- for this role instead.

See Limitations of Shell Builtins, for a discussion of portable alternatives to both printf and echo.

pwd

With modern shells, plain pwd outputs a “logical” directory name, some of whose components may be symbolic links. These directory names are in contrast to “physical” directory names, whose components are all directories.

Posix 1003.1-2001 requires that pwd must support the -L (“logical”) and -P (“physical”) options, with -L being the default. However, traditional shells do not support these options, and their pwd command has the -P behavior.

Portable scripts should assume neither option is supported, and should assume neither behavior is the default. Also, on many hosts ‘/bin/pwd’ is equivalent to ‘pwd -P’, but Posix does not require this behavior and portable scripts should not rely on it.

Typically it’s best to use plain pwd. On modern hosts this outputs logical directory names, which have the following advantages:

• Logical names are what the user specified.
• Physical names may not be portable from one installation host to another due to network file system gymnastics.
• On modern hosts ‘pwd -P’ may fail due to lack of permissions to some parent directory, but plain pwd cannot fail for this reason.

Also please see the discussion of the cd command.

read

No options are portable, not even support -r (Solaris /bin/sh for example). Tru64/OSF 5.1 sh treats read as a special built-in, so it may exit if input is redirected from a non-existent or unreadable file.

set

With the FreeBSD 6.0 shell, the set command (without any options) does not sort its output.

The set builtin faces the usual problem with arguments starting with a dash. Modern shells such as Bash or Zsh understand -- to specify the end of the options (any argument after -- is a parameter, even ‘-x’ for instance), but many traditional shells (e.g., Solaris 10 /bin/sh) simply stop option processing as soon as a non-option argument is found. Therefore, use ‘dummy’ or simply ‘x’ to end the option processing, and use shift to pop it out:

set x $my_list; shift

Avoid ‘set -’, e.g., ‘set - $my_list’. Posix no longer requires support for this command, and in traditional shells ‘set - $my_list’ resets the -v and -x options, which makes scripts harder to debug.

Some nonstandard shells do not recognize more than one option (e.g., ‘set -e -x’ assigns ‘-x’ to the command line). It is better to combine them:

set -ex

The option -e has historically been underspecified, with enough ambiguities to cause numerous differences across various shell implementations; see for example this overview, or this link, documenting a change to Posix 2008 to match ksh88 behavior. Note that mixing set -e and shell functions is asking for surprises:

set -e
doit()
{
  rm file
  echo one
}
doit || echo two

According to the recommendation, ‘one’ should always be output regardless of whether the rm failed, because it occurs within the body of the shell function ‘doit’ invoked on the left side of ‘||’, where the effects of ‘set -e’ are not enforced. Likewise, ‘two’ should never be printed, since the failure of rm does not abort the function, such that the status of ‘doit’ is 0.

The BSD shell has had several problems with the -e option. Older versions of the BSD shell (circa 1990) mishandled ‘&&’, ‘||’, ‘if’, and ‘case’ when -e was in effect, causing the shell to exit unexpectedly in some cases. This was particularly a problem with makefiles, and led to circumlocutions like ‘sh -c 'test -f file || touch file'’, where the seemingly-unnecessary ‘sh -c '…'’ wrapper works around the bug (see Failure in Make Rules).

Even relatively-recent versions of the BSD shell (e.g., OpenBSD 3.4) wrongly exit with -e if the last command within a compound statement fails and is guarded by an ‘&&’ only. For example:

#! /bin/sh
set -e
foo=''
test -n "$foo" && exit 1
echo one
if :; then
  test -n "$foo" && exit 1
  echo two
  test -n "$foo" && exit 1
fi
echo three

does not print ‘three’. One workaround is to change the last instance of ‘test -n "$foo" && exit 1’ to be ‘if test -n "$foo"; then exit 1; fi’ instead. Another possibility is to warn BSD users not to use ‘sh -e’.

When ‘set -e’ is in effect, a failed command substitution in Solaris /bin/sh cannot be ignored, even with ‘||’.

$ /bin/sh -c 'set -e; foo=`false` || echo foo; echo bar'
$ bash -c 'set -e; foo=`false` || echo foo; echo bar'
foo
bar

Moreover, a command substitution, successful or not, causes this shell to exit from a failing outer command even in presence of an ‘&&’ list:

$ bash -c 'set -e; false `true` && echo notreached; echo ok'
ok
$ sh -c 'set -e; false `true` && echo notreached; echo ok'
$

Portable scripts should not use ‘set -e’ if trap is used to install an exit handler. This is because Tru64/OSF 5.1 sh sometimes enters the trap handler with the exit status of the command prior to the one that triggered the errexit handler:

$ sh -ec 'trap '\''echo $?'\'' 0; false'
0
$ sh -c 'set -e; trap '\''echo $?'\'' 0; false'
1

Thus, when writing a script in M4sh, rather than trying to rely on ‘set -e’, it is better to append ‘|| AS_EXIT’ to any statement where it is desirable to abort on failure.

Job control is not provided by all shells, so the use of ‘set -m’ or ‘set -b’ must be done with care. When using zsh in native mode, asynchronous notification (‘set -b’) is enabled by default, and using ‘emulate sh’ to switch to Posix mode does not clear this setting (although asynchronous notification has no impact unless job monitoring is also enabled). Also, zsh 4.3.10 and earlier have a bug where job control can be manipulated in interactive shells, but not in subshells or scripts. Furthermore, some shells, like pdksh, fail to treat subshells as interactive, even though the parent shell was.

$ echo $ZSH_VERSION
4.3.10
$ set -m; echo $?
0
$ zsh -c 'set -m; echo $?'
set: can't change option: -m
$ (set -m); echo $?
set: can't change option: -m
1
$ pdksh -ci 'echo $-; (echo $-)'
cim
c

Use of set -n (typically via sh -n script) to validate a script is not foolproof. Modern ksh93 tries to be helpful by informing you about better syntax, but switching the script to use the suggested syntax in order to silence the warnings would render the script no longer portable to older shells:

$ ksh -nc '``'
ksh: warning: line 1: `...` obsolete, use $(...)
0

Furthermore, on ancient hosts, such as SunOS 4, sh -n could go into an infinite loop; even with that bug fixed, Solaris 8 /bin/sh takes extremely long to parse large scripts. Autoconf itself uses sh -n within its testsuite to check that correct scripts were generated, but only after first probing for other shell features (such as test -n "${BASH_VERSION+set}") that indicate a reasonably fast and working implementation.

shift

Not only is shifting a bad idea when there is nothing left to shift, but in addition it is not portable: the shell of MIPS RISC/OS 4.52 refuses to do it.

Don’t use ‘shift 2’ etc.; while it in the SVR1 shell (1983), it is also absent in many pre-Posix shells.

source

This command is not portable, as Posix does not require it; use . instead.

test

The test program is the way to perform many file and string tests. It is often invoked by the alternate name ‘[’, but using that name in Autoconf code is asking for trouble since it is an M4 quote character.

The -a, -o, ‘(’, and ‘)’ operands are not present in all implementations, and have been marked obsolete by Posix 2008. This is because there are inherent ambiguities in using them. For example, ‘test "$1" -a "$2"’ looks like a binary operator to check whether two strings are both non-empty, but if ‘$1’ is the literal ‘!’, then some implementations of test treat it as a negation of the unary operator -a.

Thus, portable uses of test should never have more than four arguments, and scripts should use shell constructs like ‘&&’ and ‘||’ instead. If you combine ‘&&’ and ‘||’ in the same statement, keep in mind that they have equal precedence, so it is often better to parenthesize even when this is redundant. For example:

# Not portable:
test "X$a" = "X$b" -a \
  '(' "X$c" != "X$d" -o "X$e" = "X$f" ')'

# Portable:
test "X$a" = "X$b" &&
  { test "X$c" != "X$d" || test "X$e" = "X$f"; }

test does not process options like most other commands do; for example, it does not recognize the -- argument as marking the end of options.

It is safe to use ‘!’ as a test operator. For example, ‘if test ! -d foo; …’ is portable even though ‘if ! test -d foo; …’ is not.

test (files)

To enable configure scripts to support cross-compilation, they shouldn’t do anything that tests features of the build system instead of the host system. But occasionally you may find it necessary to check whether some arbitrary file exists. To do so, use ‘test -f’, ‘test -r’, or ‘test -x’. Do not use ‘test -e’, because Solaris 10 /bin/sh lacks it. To test for symbolic links on systems that have them, use ‘test -h’ rather than ‘test -L’; either form conforms to Posix 1003.1-2001, but older shells like Solaris 8 /bin/sh support only -h.

For historical reasons, Posix reluctantly allows implementations of ‘test -x’ that will succeed for the root user, even if no execute permissions are present. Furthermore, shells do not all agree on whether Access Control Lists should affect ‘test -r’, ‘test -w’, and ‘test -x’; some shells base test results strictly on the current user id compared to file owner and mode, as if by stat(2); while other shells base test results on whether the current user has the given right, even if that right is only granted by an ACL, as if by faccessat(2). Furthermore, there is a classic time of check to time of use race between any use of test followed by operating on the just-checked file. Therefore, it is a good idea to write scripts that actually attempt an operation, and are prepared for the resulting failure if permission is denied, rather than trying to avoid an operation based solely on whether test guessed that it might not be permitted.

test (strings)

Posix says that ‘test "string"’ succeeds if string is not null, but this usage is not portable to traditional platforms like Solaris 10 /bin/sh, which mishandle strings like ‘!’ and ‘-n’.

Posix also says that ‘test ! "string"’, ‘test -n "string"’ and ‘test -z "string"’ work with any string, but many shells (such as Solaris, AIX 3.2, UNICOS 10.0.0.6, Digital Unix 4, etc.) get confused if string looks like an operator:

$ test -n =
test: argument expected
$ test ! -n
test: argument expected
$ test -z ")"; echo $?
0

Similarly, Posix says that both ‘test "string1" = "string2"’ and ‘test "string1" != "string2"’ work for any pairs of strings, but in practice this is not true for troublesome strings that look like operators or parentheses, or that begin with ‘-’.

It is best to protect such strings with a leading ‘X’, e.g., ‘test "Xstring" != X’ rather than ‘test -n "string"’ or ‘test ! "string"’.

It is common to find variations of the following idiom:

test -n "`echo $ac_feature | sed 's/[-a-zA-Z0-9_]//g'`" &&
  action

to take an action when a token matches a given pattern. Such constructs should be avoided by using:

case $ac_feature in
  *[!-a-zA-Z0-9_]*) action;;
esac

If the pattern is a complicated regular expression that cannot be expressed as a shell pattern, use something like this instead:

expr "X$ac_feature" : 'X.*[^-a-zA-Z0-9_]' >/dev/null &&
  action

‘expr "Xfoo" : "Xbar"’ is more robust than ‘echo "Xfoo" | grep "^Xbar"’, because it avoids problems when ‘foo’ contains backslashes.

trap

It is safe to trap at least the signals 1, 2, 13, and 15. You can also trap 0, i.e., have the trap run when the script ends (either via an explicit exit, or the end of the script). The trap for 0 should be installed outside of a shell function, or AIX 5.3 /bin/sh will invoke the trap at the end of this function.

Posix says that ‘trap - 1 2 13 15’ resets the traps for the specified signals to their default values, but many common shells (e.g., Solaris /bin/sh) misinterpret this and attempt to execute a “command” named - when the specified conditions arise. Posix 2008 also added a requirement to support ‘trap 1 2 13 15’ to reset traps, as this is supported by a larger set of shells, but there are still shells like dash that mistakenly try to execute 1 instead of resetting the traps. Therefore, there is no portable workaround, except for ‘trap - 0’, for which ‘trap '' 0’ is a portable substitute.

Although Posix is not absolutely clear on this point, it is widely admitted that when entering the trap ‘$?’ should be set to the exit status of the last command run before the trap. The ambiguity can be summarized as: “when the trap is launched by an exit, what is the last command run: that before exit, or exit itself?”

Bash considers exit to be the last command, while Zsh and Solaris /bin/sh consider that when the trap is run it is still in the exit, hence it is the previous exit status that the trap receives:

$ cat trap.sh
trap 'echo $?' 0
(exit 42); exit 0
$ zsh trap.sh
42
$ bash trap.sh
0

The portable solution is then simple: when you want to ‘exit 42’, run ‘(exit 42); exit 42’, the first exit being used to set the exit status to 42 for Zsh, and the second to trigger the trap and pass 42 as exit status for Bash. In M4sh, this is covered by using AS_EXIT.

The shell in FreeBSD 4.0 has the following bug: ‘$?’ is reset to 0 by empty lines if the code is inside trap.

$ trap 'false

echo $?' 0
$ exit
0

Fortunately, this bug only affects trap.

Several shells fail to execute an exit trap that is defined inside a subshell, when the last command of that subshell is not a builtin. A workaround is to use ‘exit $?’ as the shell builtin.

$ bash -c '(trap "echo hi" 0; /bin/true)'
hi
$ /bin/sh -c '(trap "echo hi" 0; /bin/true)'
$ /bin/sh -c '(trap "echo hi" 0; /bin/true; exit $?)'
hi

Likewise, older implementations of bash failed to preserve ‘$?’ across an exit trap consisting of a single cleanup command.

$ bash -c 'trap "/bin/true" 0; exit 2'; echo $?
2
$ bash-2.05b -c 'trap "/bin/true" 0; exit 2'; echo $?
0
$ bash-2.05b -c 'trap ":; /bin/true" 0; exit 2'; echo $?
2

Be aware that a trap can be called from any number of places in your script, and therefore the trap handler should not make assumptions about shell state. For some examples, if your script temporarily modifies IFS, then the trap should include an initialization back to its typical value of space-tab-newline (autoconf does this for generated configure files). Likewise, if your script changes the current working directory at some point after the trap is installed, then your trap cannot assume which directory it is in, and should begin by changing directories to an absolute path if that is important to the cleanup efforts (autotest does this for generated testsuite files).

true

Don’t worry: as far as we know true is portable. Nevertheless, it’s not always a builtin (e.g., Bash 1.x), and the portable shell community tends to prefer using :. This has a funny side effect: when asked whether false is more portable than true Alexandre Oliva answered:

In a sense, yes, because if it doesn’t exist, the shell will produce an exit status of failure, which is correct for false, but not for true.

Remember that even though ‘:’ ignores its arguments, it still takes time to compute those arguments. It is a good idea to use double quotes around any arguments to ‘:’ to avoid time spent in field splitting and file name expansion.

unset

In some nonconforming shells (e.g., Solaris 10 /bin/ksh and /usr/xpg4/bin/sh, NetBSD 5.99.43 sh, or Bash 2.05a), unset FOO fails when FOO is not set. This can interfere with set -e operation. You can use

FOO=; unset FOO

if you are not sure that FOO is set.

A few ancient shells lack unset entirely. For some variables such as PS1, you can use a neutralizing value instead:

PS1='$ '

Usually, shells that do not support unset need less effort to make the environment sane, so for example is not a problem if you cannot unset CDPATH on those shells. However, Bash 2.01 mishandles unset MAIL and unset MAILPATH in some cases and dumps core. So, you should do something like

( (unset MAIL) || exit 1) >/dev/null 2>&1 && unset MAIL || :

See Special Shell Variables, for some neutralizing values. Also, see Limitations of Builtins, for the case of environment variables.

wait

The exit status of wait is not always reliable.