Why is nullglob not default?

Question

In most shells nullglob isn't the default. That means, for example, if you run this command

ls *

in an empty directory, it will expand the * glob to a literal *, instead to an empty list of arguments. There are ways to change that behaviour, so that * in an empty directory will return an empty list of arguments, which would seem more intuitive.

So, is there a reason why nullglob is disabled by default? If so, what is that reason?

Answer 1

The nullglob option¹ would not be ideal in a number of cases. And ls is a good example:

ls *.txt

Or its more correct equivalent:

ls -- *.txt

(to list files whose name ends in .txt except for those of type directory where it lists their contents instead; maybe you meant to use the -d option to avoid that special treatment of directory files).

With nullglob on would run ls with no argument which is treated as ls -- . (list the contents of the current directory) if no files match, which is probably worse than calling ls with a literal *.txt as argument².

You'd have similar problems with most text utilities:

grep foo *.txt

Would look for foo on stdin if there's no txt file.

A more sensible default, and the one of csh, tcsh, zsh or fish 2.3+ (and of early Unix shells) is to cancel the command altogether if the glob doesn't match.

bash (since version 3) has a failglob option for that (interesting to this discussion, since contrary to ash, AT&T ksh or zsh, bash doesn't support local scopes for options³, that option when enabled globally does break a few things like the bash-completion functions).

Note that csh and tcsh are slightly different from zsh, fish or bash -O failglob in cases like:

ls -- *.txt *.html

Where you need all the globs to not-match for the command to be cancelled. For instance, if there's one txt file and no html file, that becomes:

ls -- file.txt

You can get that behaviour with zsh with set -o cshnullglob though a more sensible way to do it in zsh would be to use a glob like:

ls -- *.(txt|html)

In zsh and ksh93, you can also apply nullglob on a per-glob basis, which is a lot saner approach than modifying a global setting:

files=( *.txt(N)  ) # zsh
files=( ~(N)*.txt ) # ksh93

would create an empty array if there's no txt file instead of failing the command with an error (or making it an array with one *.txt literal argument with other shells).

Versions of fish prior to 2.3 would work like bash -O nullglob but give a warning when interactive when a glob has no match. Since 2.3, it works like zsh except for globs used in for, set or count.

Now, on the history note, the behaviour was actually broken by the Bourne shell. In prior versions of Unix, globbing was done via the /etc/glob helper and that helper behaved like csh: it would fail the command if none of the globs matched any file and remove the globs with no match otherwise.

So the situation we're in today is due to a bad decision made in the Bourne shell.

Note that the Bourne shell (and the C shell) came with another new Unix feature: the environment. That meant variable expansion (it's predecessor only had the $1, $2... positional parameters). The Bourne shell also introduced command substitution.

Another poor design decision of the Bourne shell was to perform globbing (and splitting) upon the expansion of variables and command substitution (possibly for backward compatibility with the Thompson shell where echo $1 would still invoke /etc/glob if $1 contained wildcards (it was more like pre-processor macro expansion there, as in the expanded value was parsed again as shell code)).

Failing globs that don't match would mean for instance that:

pattern='a.*b'
grep $pattern file

would fail the command (unless there are some a.whateverb files in the current directory). csh (which also performs globbing upon variable expansion) does fail the command in that case (and I'd argue it's better than leaving a dormant bug there, even if it's not as good as not doing globbing at all like in rc/ zsh / fish...).

---

^{¹ added in 2.0 in 1996 with the introduction of the shopt builtin, named after zsh's equivalent option, though bash had the allow_null_glob_expansion variable for that in earlier versions}

^{² for which ls would likely report an error that the *.txt file doesn't exist, unless it has been created in the interval, or the current directory happens to be searchable but not readable and that file or directory exists. Try after mkdir -p '*.txt/wtf'; chmod a=,u=wx . for instance}

^{³ version 4.4 saw some improvement on that front in that options set by set -o could be made local to functions with local - like in the Almquist shell, but that doesn't work for bash's second set of options, the ones set with shopt}

Answer 2

This design incorporates a safe approach, considering the use of nullglob as the default for commands that yield void results when provided with void arguments. However, this default behavior may not be ideal for commands like ls. While nullglob proves useful in scripts, it is not entirely safe unless ls is replaced with a function, as demonstrated below:

myls() { (($#)) && ls "$@" ; }

This ensures the desired outcome when employing nullglob:

shopt -s nullglob
myls empty_directory/*

When designing a language default, the priority is to adopt an option that minimizes unexpected behavior. Opting for exceptions over unpredictable actions is preferred.

It's essential to note that introducing nullglob into a complex script with unpredictable input variations may render it unstable. In practice, it could be considered a seemingly good idea that, in reality, requires additional time for preventive debugging.