Next: , Previous: , Up: Conditionals   [Contents][Index]

6.3 Recursion in m4

There is no direct support for loops in m4, but macros can be recursive. There is no limit on the number of recursion levels, other than those enforced by your hardware and operating system.

Loops can be programmed using recursion and the conditionals described previously.

There is a builtin macro, shift, which can, among other things, be used for iterating through the actual arguments to a macro:

Builtin: shift (arg1, …)

Takes any number of arguments, and expands to all its arguments except arg1, separated by commas, with each argument quoted.

The macro shift is recognized only with parameters.

shift(`foo', `bar', `baz')

An example of the use of shift is this macro:

Composite: reverse (…)

Takes any number of arguments, and reverses their order.

It is implemented as:

define(`reverse', `ifelse(`$#', `0', , `$#', `1', ``$1'',
                          `reverse(shift($@)), `$1'')')
reverse(`foo', `bar', `gnats', `and gnus')
⇒and gnus, gnats, bar, foo

While not a very interesting macro, it does show how simple loops can be made with shift, ifelse and recursion. It also shows that shift is usually used with ‘$@’. Another example of this is an implementation of a short-circuiting conditional operator.

Composite: cond (test-1, string-1, equal-1, [test-2], [string-2], [equal-2], …, [not-equal])

Similar to ifelse, where an equal comparison between the first two strings results in the third, otherwise the first three arguments are discarded and the process repeats. The difference is that each test-<n> is expanded only when it is encountered. This means that every third argument to cond is normally given one more level of quoting than the corresponding argument to ifelse.

Here is the implementation of cond, along with a demonstration of how it can short-circuit the side effects in side. Notice how all the unquoted side effects happen regardless of how many comparisons are made with ifelse, compared with only the relevant effects with cond.

`ifelse(`$#', `1', `$1',
        `ifelse($1, `$2', `$3',
define(`side', `define(`counter', incr(counter))$1')dnl
`define(`counter', `0')dnl
ifelse(side(`$1'), `yes', `one comparison: ',
       side(`$1'), `no', `two comparisons: ',
       side(`$1'), `maybe', `three comparisons: ',
       `side(`default answer: ')')counter')dnl
`define(`counter', `0')dnl
cond(`side(`$1')', `yes', `one comparison: ',
     `side(`$1')', `no', `two comparisons: ',
     `side(`$1')', `maybe', `three comparisons: ',
     `side(`default answer: ')')counter')dnl
⇒one comparison: 3
⇒two comparisons: 3
⇒three comparisons: 3
example1(`feeling rather indecisive today')
⇒default answer: 4
⇒one comparison: 1
⇒two comparisons: 2
⇒three comparisons: 3
example2(`feeling rather indecisive today')
⇒default answer: 4

Another common task that requires iteration is joining a list of arguments into a single string.

Composite: join ([separator], [args…])
Composite: joinall ([separator], [args…])

Generate a single-quoted string, consisting of each arg separated by separator. While joinall always outputs a separator between arguments, join avoids the separator for an empty arg.

Here are some examples of its usage, based on the implementation m4-1.4.19/examples/join.m4 distributed in this package:

$ m4 -I examples
join,join(`-'),join(`-', `'),join(`-', `', `')
joinall,joinall(`-'),joinall(`-', `'),joinall(`-', `', `')
join(`-', `1')
join(`-', `1', `2', `3')
join(`', `1', `2', `3')
join(`-', `', `1', `', `', `2', `')
joinall(`-', `', `1', `', `', `2', `')
join(`,', `1', `2', `3')
define(`nargs', `$#')dnl
nargs(join(`,', `1', `2', `3'))

Examining the implementation shows some interesting points about several m4 programming idioms.

$ m4 -I examples
⇒# join(sep, args) - join each non-empty ARG into a single
⇒# string, with each element separated by SEP
⇒`ifelse(`$#', `2', ``$2'',
⇒  `ifelse(`$2', `', `', ``$2'_')$0(`$1', shift(shift($@)))')')
⇒`ifelse(`$#$2', `2', `',
⇒  `ifelse(`$2', `', `', ``$1$2'')$0(`$1', shift(shift($@)))')')
⇒# joinall(sep, args) - join each ARG, including empty ones,
⇒# into a single string, with each element separated by SEP
⇒define(`joinall', ``$2'_$0(`$1', shift($@))')
⇒`ifelse(`$#', `2', `', ``$1$3'$0(`$1', shift(shift($@)))')')

First, notice that this implementation creates helper macros _join and _joinall. This division of labor makes it easier to output the correct number of separator instances: join and joinall are responsible for the first argument, without a separator, while _join and _joinall are responsible for all remaining arguments, always outputting a separator when outputting an argument.

Next, observe how join decides to iterate to itself, because the first arg was empty, or to output the argument and swap over to _join. If the argument is non-empty, then the nested ifelse results in an unquoted ‘_’, which is concatenated with the ‘$0’ to form the next macro name to invoke. The joinall implementation is simpler since it does not have to suppress empty arg; it always executes once then defers to _joinall.

Another important idiom is the idea that separator is reused for each iteration. Each iteration has one less argument, but rather than discarding ‘$1’ by iterating with $0(shift($@)), the macro discards ‘$2’ by using $0(`$1', shift(shift($@))).

Next, notice that it is possible to compare more than one condition in a single ifelse test. The test of ‘$#$2’ against ‘2’ allows _join to iterate for two separate reasons—either there are still more than two arguments, or there are exactly two arguments but the last argument is not empty.

Finally, notice that these macros require exactly two arguments to terminate recursion, but that they still correctly result in empty output when given no args (i.e., zero or one macro argument). On the first pass when there are too few arguments, the shift results in no output, but leaves an empty string to serve as the required second argument for the second pass. Put another way, ‘`$1', shift($@)’ is not the same as ‘$@’, since only the former guarantees at least two arguments.

Sometimes, a recursive algorithm requires adding quotes to each element, or treating multiple arguments as a single element:

Composite: quote (…)
Composite: dquote (…)
Composite: dquote_elt (…)

Takes any number of arguments, and adds quoting. With quote, only one level of quoting is added, effectively removing whitespace after commas and turning multiple arguments into a single string. With dquote, two levels of quoting are added, one around each element, and one around the list. And with dquote_elt, two levels of quoting are added around each element.

An actual implementation of these three macros is distributed as m4-1.4.19/examples/quote.m4 in this package. First, let’s examine their usage:

$ m4 -I examples
-quote(`1', `2')-dquote(`1', `2')-dquote_elt(`1', `2')-
define(`n', `$#')dnl
-n(quote(`1', `2'))-n(dquote(`1', `2'))-n(dquote_elt(`1', `2'))-
dquote(dquote_elt(`1', `2'))
dquote_elt(dquote(`1', `2'))

The last two lines show that when given two arguments, dquote results in one string, while dquote_elt results in two. Now, examine the implementation. Note that quote and dquote_elt make decisions based on their number of arguments, so that when called without arguments, they result in nothing instead of a quoted empty string; this is so that it is possible to distinguish between no arguments and an empty first argument. dquote, on the other hand, results in a string no matter what, since it is still possible to tell whether it was invoked without arguments based on the resulting string.

$ m4 -I examples
⇒# quote(args) - convert args to single-quoted string
⇒define(`quote', `ifelse(`$#', `0', `', ``$*'')')
⇒# dquote(args) - convert args to quoted list of quoted strings
⇒define(`dquote', ``$@'')
⇒# dquote_elt(args) - convert args to list of double-quoted strings
⇒define(`dquote_elt', `ifelse(`$#', `0', `', `$#', `1', ```$1''',
⇒                             ```$1'',$0(shift($@))')')

It is worth pointing out that ‘quote(args)’ is more efficient than ‘joinall(`,', args)’ for producing the same output.

One more useful macro based on shift allows portably selecting an arbitrary argument (usually greater than the ninth argument), without relying on the GNU extension of multi-digit arguments (see Arguments).

Composite: argn (n, …)

Expands to argument n out of the remaining arguments. n must be a positive number. Usually invoked as ‘argn(`n',$@)’.

It is implemented as:

define(`argn', `ifelse(`$1', 1, ``$2'',
  `argn(decr(`$1'), shift(shift($@)))')')
argn(`1', `a')
define(`foo', `argn(`11', $@)')
foo(`a', `b', `c', `d', `e', `f', `g', `h', `i', `j', `k', `l')

Next: , Previous: , Up: Conditionals   [Contents][Index]