Bash arrays

Way too many people don’t understand Bash arrays. Many of them argue that if you need arrays, you shouldn’t be using Bash. If we reject the notion that one should never use Bash for scripting, then thinking you don’t need Bash arrays is what I like to call “wrong”. I don’t even mean real scripting; even these little stubs in /usr/bin:

#!/bin/sh
java -jar /…/something.jar $* # WRONG!

Command line arguments are exposed as an array, that little $* is accessing it, and is doing the wrong thing (for the lazy, the correct thing is -- "$@"). Arrays in Bash offer a safe way preserve field separation.

One of the main sources of bugs (and security holes) in shell scripts is field separation. That’s what arrays are about.

What? Field separation?

Field separation is just splitting a larger unit into a list of “fields”. The most common case is when Bash splits a “simple command” (in the Bash manual’s terminology) into a list of arguments. Understanding how this works is an important prerequisite to understanding arrays, and even why they are important.

Dealing with lists is something that is very common in Bash scripts; from dealing with lists of arguments, to lists of files; they pop up a lot, and each time, you need to think about how the list is separated. In the case of $PATH, the list is separated by colons. In the case of $CFLAGS, the list is separated by whitespace. In the case of actual arrays, it’s easy, there’s no special character to worry about, just quote it, and you’re good to go.

Bash word splitting

When Bash reads a “simple command”, it splits the whole thing into a list of “words”. “The first word specifies the command to be executed, and is passed as argument zero. The remaining words are passed as arguments to the invoked command.” (to quote bash(1))

It is often hard for those unfamiliar with Bash to understand when something is multiple words, and when it is a single word that just contains a space or newline. To help gain an intuitive understanding, I recommend using the following command to print a bullet list of words, to see how Bash splits them up:

printf ' -> %s\n' words…
 -> word one
 -> multiline
word
 -> third word

In a simple command, in absence of quoting, Bash separates the “raw” input into words by splitting on spaces and tabs. In other places, such as when expanding a variable, it uses the same process, but splits on the characters in the $IFS variable (which has the default value of space/tab/newline). This process is, creatively enough, called “word splitting”.

In most discussions of Bash arrays, one of the frequent criticisms is all the footnotes and “gotchas” about when to quote things. That’s because they usually don’t set the context of word splitting. Double quotes (") inhibit Bash from doing word splitting. That’s it, that’s all they do. Arrays are already split into words; without wrapping them in double quotes Bash re-word splits them, which is almost never what you want; otherwise, you wouldn’t be working with an array.

Normal array syntax

Now, for accessing the array. The most important things to understanding arrays is to quote them, and understanding the difference between @ and *.

It’s really that simple—that covers most usages of arrays, and most of the mistakes made with them.

To help you understand the difference between @ and *, here is a sample of each:

#!/bin/bash
array=(foo bar baz)
for item in "${array[@]}"; do
        echo " - <${item}>"
done

#!/bin/bash
array=(foo bar baz)
for item in "${array[*]}"; do
        echo " - <${item}>"
done

 - <foo>
 - <bar>
 - <baz>

 - <foo bar baz>

In most cases, @ is what you want, but * comes up often enough too.

To get individual entries, the syntax is ${array[n]}, where n starts at 0.

To get a subset of the array, there are a few options:

Notice that "${array[@]}" is equivalent to "${array[@]:0}".

Argument array syntax

Accessing the arguments is mostly that simple, but that array doesn’t actually have a variable name. It’s special. Instead, it is exposed through a series of special variables (normal variables can only start with letters and underscore), that mostly match up with the normal array syntax.

Setting the arguments array, on the other hand, is pretty different. That’s fine, because setting the arguments array is less useful anyway.

Did you notice what was inconsistent? The variables $*, $@, and $# behave like the n=0 entry doesn’t exist.

These make sense because argument 0 is the name of the script—we almost never want that when parsing arguments. You’d spend more code getting the values that it currently gives you.

Now, for an explanation of setting the arguments array. You cannot set argument n=0. The set command is used to manipulate the arguments passed to Bash after the fact—similarly, you could use set -x to make Bash behave like you ran it as bash -x; like most GNU programs, the -- tells it to not parse any of the options as flags. The shift command shifts each entry n spots to the left, using n=1 if no value is specified; and leaving argument 0 alone.

But you mentioned “gotchas” about quoting!

But I explained that quoting simply inhibits word splitting, which you pretty much never want when working with arrays. If, for some odd reason, you do what word splitting, then that’s when you don’t quote. Simple, easy to understand.

I think possibly the only case where you do want word splitting with an array is when you didn’t want an array, but it’s what you get (arguments are, by necessity, an array). For example:

# Usage: path_ls PATH1 PATH2…
# Description:
#   Takes any number of PATH-style values; that is,
#   colon-separated lists of directories, and prints a
#   newline-separated list of executables found in them.
# Bugs:
#   Does not correctly handle programs with a newline in the name,
#   as the output is newline-separated.
path_ls() {
    local IFS dirs
    IFS=:
    dirs=($@) # The odd-ball time that it needs to be unquoted
    find -L "${dirs[@]}" -maxdepth 1 -type f -executable \
        -printf '%f\n' 2>/dev/null | sort -u
}

Logically, there shouldn’t be multiple arguments, just a single $PATH value; but, we can’t enforce that, as the array can have any size. So, we do the robust thing, and just act on the entire array, not really caring about the fact that it is an array. Alas, there is still a field-separation bug in the program, with the output.

I still don’t think I need arrays in my scripts

Consider the common code:

ARGS=' -f -q'
…
command $ARGS  # unquoted variables are a bad code-smell anyway

Here, $ARGS is field-separated by $IFS, which we are assuming has the default value. This is fine, as long as $ARGS is known to never need an embedded space; which you do as long as it isn’t based on anything outside of the program. But wait until you want to do this:

ARGS=' -f -q'
…
if [[ -f "$filename" ]]; then
    ARGS+=" -F $filename"
fi
…
command $ARGS

Now you’re hosed if $filename contains a space! More than just breaking, it could have unwanted side effects, such as when someone figures out how to make filename='foo --dangerous-flag'.

Compare that with the array version:

ARGS=(-f -q)
…
if [[ -f "$filename" ]]; then
    ARGS+=(-F "$filename")
fi
…
command "${ARGS[@]}"

What about portability?

Except for the little stubs that call another program with "$@" at the end, trying to write for multiple shells (including the ambiguous /bin/sh) is not a task for mere mortals. If you do try that, your best bet is probably sticking to POSIX. Arrays are not POSIX; except for the arguments array, which is; though getting subset arrays from $@ and $* is not (tip: use set -- to re-purpose the arguments array).

Writing for various versions of Bash, though, is pretty do-able. Everything here works all the way back in bash-2.0 (December 1996), with the following exceptions:

• The += operator wasn’t added until Bash 3.1.

  • As a work-around, use array[${#array[*]}]=word to append a single element.

• Accessing subset arrays of the arguments array is inconsistent if pos=0 in ${@:pos:len}.

  • In Bash 2.x and 3.x, it works as expected, except that argument 0 is silently missing. For example ${@:0:3} gives arguments 1 and 2; where ${@:1:3} gives arguments 1, 2, and 3. This means that if pos=0, then only len-1 arguments are given back.
  • In Bash 4.0, argument 0 can be accessed, but if pos=0, then it only gives back len-1 arguments. So, ${@:0:3} gives arguments 0 and 1.
  • In Bash 4.1 and higher, it works in the way described in the main part of this document.

Now, Bash 1.x doesn’t have arrays at all. $@ and $* work, but using : to select a range of elements from them doesn’t. Good thing most boxes have been updated since 1996!