9.1.3.1 More about ‘\’ and ‘&’ with sub(), gsub(), and gensub()

CAUTION: This subsubsection has been reported to cause headaches. You might want to skip it upon first reading.

When using sub(), gsub(), or gensub(), and trying to get literal backslashes and ampersands into the replacement text, you need to remember that there are several levels of escape processing going on.

First, there is the lexical level, which is when awk reads your program and builds an internal copy of it to execute. Then there is the runtime level, which is when awk actually scans the replacement string to determine what to generate.

At both levels, awk looks for a defined set of characters that can come after a backslash. At the lexical level, it looks for the escape sequences listed in Escape Sequences. Thus, for every ‘\’ that awk processes at the runtime level, you must type two backslashes at the lexical level. When a character that is not valid for an escape sequence follows the ‘\’, BWK awk and gawk both simply remove the initial ‘\’ and put the next character into the string. Thus, for example, "a\qb" is treated as "aqb".

At the runtime level, the various functions handle sequences of ‘\’ and ‘&’ differently. The situation is (sadly) somewhat complex. Historically, the sub() and gsub() functions treated the two-character sequence ‘\&’ specially; this sequence was replaced in the generated text with a single ‘&’. Any other ‘\’ within the replacement string that did not precede an ‘&’ was passed through unchanged. This is illustrated in Table 9.1.

 You type         sub() sees          sub() generates
 ——–         ———-          —————
     \&              &            The matched text
    \\&             \&            A literal ‘&’
   \\\&             \&            A literal ‘&’
  \\\\&            \\&            A literal ‘\&’
 \\\\\&            \\&            A literal ‘\&’
\\\\\\&           \\\&            A literal ‘\\&’
    \\q             \q            A literal ‘\q’

This table shows the lexical-level processing, where an odd number of backslashes becomes an even number at the runtime level, as well as the runtime processing done by sub(). (For the sake of simplicity, the rest of the following tables only show the case of even numbers of backslashes entered at the lexical level.)

The problem with the historical approach is that there is no way to get a literal ‘\’ followed by the matched text.

Several editions of the POSIX standard attempted to fix this problem but weren’t successful. The details are irrelevant at this point in time.

At one point, the gawk maintainer submitted proposed text for a revised standard that reverts to rules that correspond more closely to the original existing practice. The proposed rules have special cases that make it possible to produce a ‘\’ preceding the matched text. This is shown in Table 9.2.

 You type         sub() sees         sub() generates
 ——–         ———-         —————
\\\\\\&           \\\&            A literal ‘\&’
  \\\\&            \\&            A literal ‘\’, followed by the matched text
    \\&             \&            A literal ‘&’
    \\q             \q            A literal ‘\q’
   \\\\             \\            \\

In a nutshell, at the runtime level, there are now three special sequences of characters (‘\\\&’, ‘\\&’, and ‘\&’) whereas historically there was only one. However, as in the historical case, any ‘\’ that is not part of one of these three sequences is not special and appears in the output literally.

gawk 3.0 and 3.1 follow these rules for sub() and gsub(). The POSIX standard took much longer to be revised than was expected. In addition, the gawk maintainer’s proposal was lost during the standardization process. The final rules are somewhat simpler. The results are similar except for one case.

The POSIX rules state that ‘\&’ in the replacement string produces a literal ‘&’, ‘\\’ produces a literal ‘\’, and ‘\’ followed by anything else is not special; the ‘\’ is placed straight into the output. These rules are presented in Table 9.3.

 You type         sub() sees         sub() generates
 ——–         ———-         —————
\\\\\\&           \\\&            A literal ‘\&’
  \\\\&            \\&            A literal ‘\’, followed by the matched text
    \\&             \&            A literal ‘&’
    \\q             \q            A literal ‘\q’
   \\\\             \\            \

The only case where the difference is noticeable is the last one: ‘\\\\’ is seen as ‘\\’ and produces ‘\’ instead of ‘\\’.

Starting with version 3.1.4, gawk followed the POSIX rules when --posix was specified (see Options). Otherwise, it continued to follow the proposed rules, as that had been its behavior for many years.

When version 4.0.0 was released, the gawk maintainer made the POSIX rules the default, breaking well over a decade’s worth of backward compatibility.⁴⁹ Needless to say, this was a bad idea, and as of version 4.0.1, gawk resumed its historical behavior, and only follows the POSIX rules when --posix is given.

The rules for gensub() are considerably simpler. At the runtime level, whenever gawk sees a ‘\’, if the following character is a digit, then the text that matched the corresponding parenthesized subexpression is placed in the generated output. Otherwise, no matter what character follows the ‘\’, it appears in the generated text and the ‘\’ does not, as shown in Table 9.4.

  You type          gensub() sees         gensub() generates
  ——–          ————-         ——————
      &                    &            The matched text
    \\&                   \&            A literal ‘&’
   \\\\                   \\            A literal ‘\’
  \\\\&                  \\&            A literal ‘\’, then the matched text
\\\\\\&                 \\\&            A literal ‘\&’
    \\q                   \q            A literal ‘q’

Because of the complexity of the lexical- and runtime-level processing and the special cases for sub() and gsub(), we recommend the use of gawk and gensub() when you have to do substitutions.