Word Splitting
1. Introduction
The shell's parser performs several operations on your commands before finally executing them. Understanding how your original command will be transformed by the shell is of paramount importance in writing robust scripts. From the bash man page:
The order of expansions is: brace expansion, tilde expansion, parameter, variable and arithmetic expansion and command substitution (done in a left-to-right fashion), word splitting, and pathname expansion.
For additional information on word splitting and argument handling in Bash, consider reading Arguments.
2. What is Word Splitting?
This page will focus on word splitting, of course. Before we get into the technical details, let's write a little helper script that will show us the arguments as passed by the shell:
#!/bin/sh
printf "%d args:" $#
printf " <%s>" "$@"
echo
If you create a file named args with the above contents, make it executable with chmod +x args , and put it in one of the directories listed in echo "$PATH" , then you could run the following command with the following output:
griffon:~$ args hello world "how are you?" 3 args: <hello> <world> <how are you?>
The ultimate result of most shell commands is to execute some program with a specific set of arguments (as well as setting up environment variables, opening file descriptors, etc.). Word splitting is part of the process that determines what each of those arguments will be -- after word splitting and pathname expansion, every resulting word becomes an argument to the program that the shell executes. Our helper program above receives the argument list as constructed by the shell, and shows it to us.
Word splitting is performed on the results of almost all unquoted expansions. The result of the expansion is broken into separate words based on the characters of the IFS variable. If IFS is not set, then it will be performed as if IFS contained a space, a tab, and a newline. For example:
griffon:~$ var="This is a variable" griffon:~$ args $var 4 args: <This> <is> <a> <variable>
An example using IFS:
griffon:~$ log=/var/log/qmail/current IFS=/ griffon:~$ args $log 5 args: <> <var> <log> <qmail> <current> griffon:~$ unset IFS
An example with CommandSubstitution:
griffon:/music/Yello$ ls -l total 2864 -rw-r--r-- 1 greg greg 2919154 2001-05-23 00:48 Yello - Oh Yeah.mp3 griffon:/music/Yello$ args $(ls) 4 args: <Yello> <-> <Oh> <Yeah.mp3>
3. Controlling Word Splitting
As you can see above, we usually do not want to let word splitting occur when filenames are involved. (See BashPitfalls for a discussion of this particular issue.)
Double quoting an expansion suppresses word splitting, except in the special cases of "$@" and "${array[@]}":
griffon:~$ var="This is a variable"; args "$var"
1 args: <This is a variable>
griffon:~$ array=(testing, testing, "1 2 3"); args "${array[@]}"
3 args: <testing,> <testing,> <1 2 3>"$@" causes each positional parameter to be expanded to a separate word; its array equivalent likewise causes each element of the array to be expanded to a separate word.
There are very complicated rules involving whitespace characters in IFS. Quoting the man page again:
If IFS is unset, or its value is exactly <space><tab><newline>, the default, then any sequence of IFS characters serves to delimit words. If IFS has a value other than the default, then sequences of the whitespace characters space and tab are ignored at the beginning and end of the word, as long as the whitespace character is in the value of IFS (an IFS whitespace character). Any character in IFS that is not IFS whitespace, along with any adjacent IFS whitespace characters, delimits a field. A sequence of IFS whitespace characters is also treated as a delimiter. If the value of IFS is null, no word splitting occurs.
We won't explore those rules in depth here, except to note the part about sequences of non-whitespace characters. If IFS contains non-whitespace characters, then empty words can be generated:
griffon:~$ getent passwd sshd sshd:x:100:65534::/var/run/sshd:/usr/sbin/nologin griffon:~$ IFS=:; args $(getent passwd sshd) 7 args: <sshd> <x> <100> <65534> <> </var/run/sshd> </usr/sbin/nologin> griffon:~$ unset IFS
There was another empty word generated in one of our previous examples, where IFS was set to /. The observant reader will have noticed, therefore, that non-whitespace IFS characters are not ignored at the beginning and end of expansions, the way whitespace IFS characters are.
Whitespace IFS characters get consolidated. Multiple spaces in a row, for example, have the same effect as a single space, when IFS contains a space (or is not set at all). Newlines also count as whitespace for this purpose, which has important ramifications when attempting to load an array with lines of input.
Finally, we note that pathname expansion happens after word splitting, and can produce some very shocking results.
griffon:~$ getent passwd qmaild qmaild:*:994:998::/var/qmail:/sbin/nologin griffon:~$ IFS=:; args $(getent passwd qmaild) 737 args: <qmaild> <00INDEX.lsof> <03> <037_ftpd.patch> ... griffon:~$ unset IFS
The * word, produced by the shell's word splitting, was then expanded as a glob, resulting in several hundred new and exciting words. This can be disastrous if it happens unexpectedly. As with most of the dangerous features of the shell, it is retained because "it's always worked that way". In fact, it can be used for good, if you're very careful:
griffon:/music/Yello$ files='*.mp3 *.ogg' griffon:/music/Yello$ args $files 2 args: <Yello - Oh Yeah.mp3> <*.ogg>
Pathname expansion can be disabled with set -f; though this can lead to surprising and confusing code.
4. Notes
Word splitting is not performed on expansions inside Bash keywords such as [[ ... ]] and case.
- Word splitting is not performed on expansions in assignments. Thus, one does not need to quote anything in a command like these:
foo=$bar
bar=$(a command)
logfile=$logdir/foo-$(date +%Y%m%d)
PATH=/usr/local/bin:$PATH ./myscript
In either case, quoting anyway will not break anything. So if in doubt, quote!
When using the read command, word splitting is performed on the input, but only when multiple variable names are given, or when read -a is used (to populate an array). Quoting is irrelevant here, though this behavior can be disabled by removing whitespace from IFS.