A.8 Regexp Ranges and Locales: A Long Sad Story

This section describes the confusing history of ranges within regular expressions and their interactions with locales, and how this affected different versions of gawk.

The original Unix tools that worked with regular expressions defined character ranges (such as ‘[a-z]’) to match any character between the first character in the range and the last character in the range, inclusive. Ordering was based on the numeric value of each character in the machine’s native character set. Thus, on ASCII-based systems, ‘[a-z]’ matched all the lowercase letters, and only the lowercase letters, as the numeric values for the letters from ‘a’ through ‘z’ were contiguous. (On an EBCDIC system, the range ‘[a-z]’ includes additional nonalphabetic characters as well.)

Almost all introductory Unix literature explained range expressions as working in this fashion, and in particular, would teach that the “correct” way to match lowercase letters was with ‘[a-z]’, and that ‘[A-Z]’ was the “correct” way to match uppercase letters. And indeed, this was true.116

The 1992 POSIX standard introduced the idea of locales (see Where You Are Makes a Difference). Because many locales include other letters besides the plain 26 letters of the English alphabet, the POSIX standard added character classes (see Using Bracket Expressions) as a way to match different kinds of characters besides the traditional ones in the ASCII character set.

However, the standard changed the interpretation of range expressions. In the "C" and "POSIX" locales, a range expression like ‘[a-dx-z]’ is still equivalent to ‘[abcdxyz]’, as in ASCII. But outside those locales, the ordering was defined to be based on collation order.

What does that mean? In many locales, ‘A’ and ‘a’ are both less than ‘B’. In other words, these locales sort characters in dictionary order, and ‘[a-dx-z]’ is typically not equivalent to ‘[abcdxyz]’; instead, it might be equivalent to ‘[ABCXYabcdxyz]’, for example.

This point needs to be emphasized: much literature teaches that you should use ‘[a-z]’ to match a lowercase character. But on systems with non-ASCII locales, this also matches all of the uppercase characters except ‘A’ or ‘Z’! This was a continuous cause of confusion, even well into the twenty-first century.

To demonstrate these issues, the following example uses the sub() function, which does text replacement (see String-Manipulation Functions). Here, the intent is to remove trailing uppercase characters:

$ echo something1234abc | gawk-3.1.8 '{ sub("[A-Z]*$", ""); print }'
-| something1234a

This output is unexpected, as the ‘bc’ at the end of ‘something1234abc’ should not normally match ‘[A-Z]*’. This result is due to the locale setting (and thus you may not see it on your system).

Similar considerations apply to other ranges. For example, ‘["-/]’ is perfectly valid in ASCII, but is not valid in many Unicode locales, such as en_US.UTF-8.

Early versions of gawk used regexp matching code that was not locale-aware, so ranges had their traditional interpretation.

When gawk switched to using locale-aware regexp matchers, the problems began; especially as both GNU/Linux and commercial Unix vendors started implementing non-ASCII locales, and making them the default. Perhaps the most frequently asked question became something like, “Why does ‘[A-Z]’ match lowercase letters?!?”

This situation existed for close to 10 years, if not more, and the gawk maintainer grew weary of trying to explain that gawk was being nicely standards-compliant, and that the issue was in the user’s locale. During the development of version 4.0, he modified gawk to always treat ranges in the original, pre-POSIX fashion, unless --posix was used (see Command-Line Options).117

Fortunately, shortly before the final release of gawk 4.0, the maintainer learned that the 2008 standard had changed the definition of ranges, such that outside the "C" and "POSIX" locales, the meaning of range expressions was undefined.118

By using this lovely technical term, the standard gives license to implementers to implement ranges in whatever way they choose. The gawk maintainer chose to apply the pre-POSIX meaning both with the default regexp matching and when --traditional or --posix are used. In all cases gawk remains POSIX-compliant.


Footnotes

(116)

And Life was good.

(117)

And thus was born the Campaign for Rational Range Interpretation (or RRI). A number of GNU tools have already implemented this change, or will soon. Thanks to Karl Berry for coining the phrase “Rational Range Interpretation.”

(118)

See the standard and its rationale.