Patterns and pattern matching

A pattern is a string description. Bash uses them in various ways:

• Pathname expansion (Globbing - matching filenames)
• Pattern matching in conditional expressions
• Substring removal and search and replace in Parameter Expansion
• Pattern-based branching using the case command

The pattern description language is relatively easy. Any character that's not mentioned below matches itself. The NUL character may not occur in a pattern. If special characters are quoted, they're matched literally, i.e., without their special meaning.

Do not confuse patterns with regular expressions, because they share some symbols and do similar matching work.

The bracket expression […] mentioned above has some useful applications:

Some simple examples using normal pattern matching:

• Pattern "Hello world" matches
  • Hello world
• Pattern [Hh]"ello world" matches
  • ⇒ Hello world
  • ⇒ hello world
• Pattern Hello* matches (for example)
  • ⇒ Hello world
  • ⇒ Helloworld
  • ⇒ HelloWoRlD
  • ⇒ Hello
• Pattern Hello world[[:punct:]] matches (for example)
  • ⇒ Hello world!
  • ⇒ Hello world.
  • ⇒ Hello world+
  • ⇒ Hello world?
• Pattern [[.backslash.]]Hello[[.vertical-line.]]world[[.exclamation-mark.]] matches (using collation symbols)
  • ⇒ \Hello|world!

If you set the shell option extglob, Bash understands some powerful patterns. A <PATTERN-LIST> is one or more patterns, separated by the pipe-symbol (PATTERN|PATTERN).

Delete all but one specific file

rm -f !(survivior.txt)

* Counter-intuitively, only the [!chars] syntax for negating a character class is specified by POSIX for shell pattern matching. [^chars] is merely a commonly-supported extension. Even dash supports [^chars], but not posh.

* All of the extglob quantifiers supported by bash were supported by ksh88. The set of extglob quantifiers supported by ksh88 are identical to those supported by Bash, mksh, ksh93, and zsh.

* mksh does not support POSIX character classes. Therefore, character ranges like [0-9] are somewhat more portable than an equivalent POSIX class like [:digit:].

* Bash uses a custom runtime interpreter for pattern matching. (at least) ksh93 and zsh translate patterns into regexes and then use a regex compiler to emit and cache optimized pattern matching code. This means Bash may be an order of magnitude or more slower in cases that involve complex back-tracking (usually that means extglob quantifier nesting). You may wish to use Bash's regex support (the =~ operator) if performance is a problem, because Bash will use your C library regex implementation rather than its own pattern matcher.

TODO: describe the pattern escape bug https://gist.github.com/ormaaj/6195070

ksh93 supports some very powerful pattern matching features in addition to those described above.

* ksh93 supports arbitrary quantifiers just like ERE using the {from,to}(pattern-list) syntax. {2,4}(foo)bar matches between 2-4 "foo"'s followed by "bar". {2,}(foo)bar matches 2 or more "foo"'s followed by "bar". You can probably figure out the rest. So far, none of the other shells support this syntax.

* In ksh93, a pattern-list may be delimited by either & or |. & means "all patterns must be matched" instead of "any pattern". For example,

[[ fo0bar == @(fo[0-9]&+([[:alnum:]]))bar ]]

would be true while

[[ f00bar == @(fo[0-9]&+([[:alnum:]]))bar ]]

is false, because all members of the and-list must be satisfied. No other shell supports this so far, but you can simulate some cases in other shells using double extglob negation. The aforementioned ksh93 pattern is equivalent in Bash to:

[[ fo0bar == !(!(fo[0-9])|!(+([[:alnum:]])))bar ]]

, which is technically more portable, but ugly.

* ksh93's printf builtin can translate from shell patterns to ERE and back again using the %R and %P format specifiers respectively.

TODO: ~() (and regex), .sh.match, backrefs, special ${var/…/…} behavior, %()