Revision 2 not available (showing current revision instead)

AllAboutGetline

Editor's notes: The article was originally posted on the comp.lang.awk newsgroup, and can still be found there at this url: all about getline Note that the article, besides giving a very complete overview about getline usage, also reflects a specific point of view about it (that of the author), so in the article you will find comments about "good" or "correct" and "bad" or "wrong" uses of getline. Of course, although the author is a well-known and authoritative voice in the awk community, everyone is free to have his own view (see the periodic debates about getline on comp.lang.awk for an example). So my advice is to use the article primarily as a technical reference, and as a help to develop your own opinion and style with getline.

Getline Summary, November 11th 2006

The following summary, composed by Ed Morton to address the recurring issue of getline (mis)use, was based primarily on information from the book "Effective Awk Programming", Third Edition By Arnold Robbins; (http://www.oreilly.com/catalog/awkprog3) with review and additional input from many of the comp.lang.awk regulars, including Steve Calfee, Martin Cohen, Manuel Collado, Jürgen Kahrs, Kenny McCormack, Janis Papanagnou, Anton Treuenfels, Thomas Weidenfeller, John LaBadie and Edward Rosten.

Getline


getline is fine when used correctly (see below for a list of those cases), but it's best avoided by default because:

  1. It allows people to stick to their preconceived ideas of how to program rather than learning the easier way that awk was designed to read input. It's like C programmers continuing to do procedural programming in C++ rather than learning the new paradigm and the supporting language constructs.
  2. It has many insidious caveats that come back to bite you either immediately or in future. The succeeding discussion captures some of those and explains when getline IS appropriate.

As the book "Effective Awk Programming", Third Edition By Arnold Robbins; (http://www.oreilly.com/catalog/awkprog3) which provides much of the source for this discussion says:

"The getline command is used in several different ways and should not be used by beginners. ... come back and study the getline command after you have reviewed the rest ... and have a good knowledge of how awk works."

Variants


The following summarises the eight variants of getline applications, listing which variables are set by each one:

VariantVariables Set
getline$0, ${1...NF}, NF, FNR, NR, FILENAME
getline varvar, FNR, NR, FILENAME
getline < file$0, ${1...NF}, NF
getline var < filevar
command | getline$0, ${1...NF}, NF
command | getline varvar
command |& getline$0, ${1...NF}, NF
command |& getline varvar

The "command |& ..." variants are GNU awk (gawk) extensions. gawk also populates the ERRNO builtin variable if getline fails.

Although calling getline is very rarely the right approach (see below), if you need to do it the safest ways to invoke getline are:

    if/while ( (getline var < file) > 0)
    if/while ( (command | getline var) > 0)
    if/while ( (command |& getline var) > 0)

since those do not affect any of the builtin variables and they allow you to correctly test for getline succeeding or failing. If you need the input record split into separate fields, just call "split()" to do that.

Caveats


Users of getline have to be aware of the following non-obvious effects of using it:

a) Normally FILENAME is not set within a BEGIN section, but a non-redirected call to getline will set it.

b) Calling "getline < FILENAME" is NOT the same as calling "getline". The second form will read the next record from FILENAME while the first form will read the first record again.

c) Calling getline without a var to be set will update $0 and $NF so they will have a different value for subsequent processing than they had for prior processing in the same condition/action block.

d) Many of the getline variants above set some but not all of the builtin variables, so you need to be very careful that it's setting the ones you need/expect it to.

e) According to POSIX,

getline < expression

is ambiguous if expression contains unparenthesized operators other than "$"; for example,

getline < dir "/" file

is ambiguous because the concatenation operator is not parenthesized. You should write it as

getline < (dir "/" file)

if you want your program to be portable to other awk implementations.

f) In POSIX-compliant awks (e.g. gawk --posix) a failure of getline (e.g. trying to read from a non-readable file) will be fatal to the program, otherwise it won't.

g) Unredirected getline can defeat the simple and usual rule to handle input file transitions:

FNR==1 { ... start of file actions ... }

File transitions can occur at getlines, so FNR==1 needs to also be checked after each unredirected (from a specific file name) getline. e.g. if you want to print the first line of each of these files:

$ cat file1
a
b
$ cat file2
c
d

you'd normally do:

$ awk 'FNR==1{print}' file1 file2
a
c

but if a "getline" snuck in, it could have the unexpected consequence of skipping the test for FNR==1 and so not printing the first line of the second file.

$ awk 'FNR==1{print}/b/{getline}' file1 file2
a

h) Using getline in the BEGIN section to skip lines makes your program difficult to apply to multiple files. e.g. with data like...

some header line
----------------
data line 1
data line 2
...
data line 10000

you may consider using...

BEGIN { getline header; getline }
{ whatever_using_header_and_data_on_the_line() }

instead of...

FNR == 1 { header = $0 }
FNR < 3 { next }
{ whatever_using_header_and_data_on_the_line() }

but the getline version would not work on multiple files since the BEGIN section would only be executed once, before the first file is processed, whereas the non-getline version would work as-is. This is one example of the common case where the getline command itself isn't directly causing the problem, but the type of design you can end up with if you select a getline approach is not ideal.

Applications


getline is an appropriate solution for the following:

a) Reading from a pipe, e.g.:

command = "ls"
while ( (command | getline var) > 0) {
  print var
}
close(command)

b) Reading from a coprocess, e.g.:

command = "LC_ALL=C sort"
n = split("abcdefghijklmnopqrstuvwxyz", a, "")

for (i = n; i > 0; i--)
  print a[i] |& command
close(command, "to")

while ((command |& getline var) > 0)
  print "got", var
close(command)

c) In the BEGIN section, reading some initial data that's referenced during processing multiple subsequent input files, e.g.:

BEGIN {
  while ( (getline var < ARGV[1]) > 0) {
    data[var]++
  }
  close(ARGV[1])
  ARGV[1]=""
}
$0 in data

d) Recursive-descent parsing of an input file or files, e.g.:

awk 'function read(file) {
       while ( (getline < file) > 0) {
         if ($1 == "include") {
           read($2)
         } else {
           print > ARGV[2]
         }
       }
       close(file)
     }
     BEGIN{
       read(ARGV[1])
       ARGV[1]=""
       close(ARGV[2])
     }1' file1 tmp

In all other cases, it's clearest, simplest, less error-prone, and easiest to maintain to let awks normal text-processing read the records. In the case of "c", whether to use the BEGIN+getline approach or just collect the data within the awk condition/action part after testing for the first file is largely a style choice.

"a" above calls the UNIX command "ls" to list the current directory contents, then prints the result one line at a time.

"b" above writes the letters of the alphabet in reverse order, one per line, down the two-way pipe to the UNIX "sort" command. It then closes the write end of the pipe, so that sort receives an end-of-file indication. This causes sort to sort the data and write the sorted data back to the gawk program. Once all of the data has been read, gawk terminates the coprocess and exits. This is particularly necessary in order to use the UNIX "sort" utility as part of a coprocess since sort must read all of its input data before it can produce any output. The sort program does not receive an end-of-file indication until gawk closes the write end of the pipe. Other programs can be invoked as just:

command = "program"
do {
  print data |& command
  command |& getline var
} while (data left to process)
close(command)

Not that calling close() with a second argument is also gawk-specific.

"c" above reads every record of the first file passed as an argument to awk into an array and then for every subsequent file passed as an argument will print every record from that file that matches any of the records that appeared in the first file (and so are stored in the "data" array). This could alternatively have been implemented as:

# fails if first file is empty
NR==FNR{ data[$0]++; next }
$0 in data

or:

FILENAME==ARGV[1] { data[$0]++; next }
$0 in data

or:

FILENAME=="specificFileName" { data[$0]++; next }
$0 in data

or (gawk only):

ARGIND==1 { data[$0]++; next }
$0 in data

"d" above not only expands all the lines that say "include subfile", but by writing the result to a tmp file, resetting ARGV[1] (the highest level input file) and not resetting ARGV[2] (the tmp file), it then lets awk do any normal record parsing on the result of the expansion since that's now stored in the tmp file. If you don't need that, just do the "print" to stdout and remove any other references to a tmp file or ARGV[2]. In this case, since it's convenient to use $1 and $2, and no other part of the program references any builtin variables, getline was used without populating an explicit variable. This method is limited in its recursion depth to the total number of open files the OS permits at one time.

Tips


The following tips may help if, after reading the above, you discover you have an appropriate application for getline or if you're looking for an alternative solution to using getline:

a) If you need to distinguish between a normal EOF or some read or opening error, you have to use gawks ERRNO variable or code it as:

if/while ( (e = (getline var < file)) > 0) { ... }
close(file)
if(e < 0) some_error_handling

b) Don't forget to close() any file you open for reading. The common idiom for getline and other methods of opening files/streams is:

cmd="some command"
do something with cmd
close(cmd)

c) A common misapplication of getline is to just skip a few lines of an input file. The following discusses how to do that without using getline with all that implies as discussed above. This discussion builds on the common awk idiom to "decrement a variable to zero" by putting the decrement of the variable as the second term in an "and" clause with the first part being the variable itself, so the decrement only occurs if the variable is non-zero:

i) Print the Nth record after some pattern:

awk 'c&&!--c;/pattern/{c=N}' file

ii) Print every record except the Nth record after some pattern:

awk 'c&&!--c{next}/pattern/{c=N}' file

iii) Print the N records after some pattern:

awk 'c&&c--;/pattern/{c=N}' file

iiii) Print every record except the N records after some pattern:

awk 'c&&c--{next}/pattern/{c=N}' file

Editor's note: see [1] for some notes about the above code

In this example there are no blank lines and the output is all aligned with the left hand column and you want to print $0 for the second record following the record that contains some pattern, e.g. the number 3:

$ cat file
line 1
line 2
line 3
line 4
line 5
line 6
line 7
line 8
$ awk '/3/{getline;getline;print}' file
line 5

That works Just fine. Now let's see the concise way to do it without getline:

$ awk 'c&&!--c;/3/{c=2}' file
line 5

It's not quite so obvious at a glance what that does, but it uses an idiom that most awk programmers could do well to learn and it is briefer and avoids all those getline caveats.

Now let's say we want to print the 5th line after the pattern instead of the 2nd line. Then we'd have:

$ awk '/3/{getline;getline;getline;getline;getline;print}' file
line 8
$ awk 'c&&!--c;/3/{c=5}' file
line 8

i.e. we have to add a whole series of additional getline calls to the getline version, as opposed to just changing the counter from 2 to 5 for the non-getline version. In reality, you'd probably completely rewrite the getline version to use a loop:

$ awk '/3/{for (c=1;c<=5;c++) getline; print}' file
line 8

Still not as concise as the non-getline version, has all the getline caveats and required a redesign of the code just to change a counter.

Now let's say we also have to print the word "Eureka" if the number 4 appears in the input file. With the getline verion, you now have to do something like:

$ awk '/3/{for (c=1;c<=5;c++) { getline; if ($0 ~ /4/) print "Eureka!" }
print}' file
Eureka!
line 8

whereas with the non-getline version you just have to do:

$ awk 'c&&!--c;/3/{c=5}/4/{print "Eureka!"}' file
Eureka!
line 8

i.e. with the getline version, you have to work around the fact that you're now processing records outside of the normal awk work-loop, whereas with the non-getline version you just have to drop your test for "4" into the normal place and let awks normal record processing deal with it like it always does.

Actually, if you look closely at the above you'll notice we just unintentionally introduced a bug in the getline version. Consider what would happen in both versions if 3 and 4 appear on the same line. The non-getline version would behave correctly, but to fix the getline version, you'd need to duplicate the condition somewhere, e.g. perhaps something like this:

$ awk '/3/{for (c=1;c<=5;c++) { if ($0 ~ /4/) print "Eureka!"; getline }
if ($0 ~ /4/) print "Eureka!"; print}' file
Eureka!
line 8

Now consider how the above would behave when there aren't 5 lines left in the input file or when the last line of the file contains both a 3 and a 4. i.e. there are still design questions to be answered and bugs that will appear at the limits of the input space.

Ignoring those bugs since this is not intended as a discussion on debugging getline programs, let's say you no longer need to print the 5th record after the number 3 but still have to do the Eureka on 4. With the getline version, you'd strip out the test for 3 and the getline stuff to be left with:

$ awk '{if ($0 ~ /4/) print "Eureka!"}' file
Eureka!

which you'd then presumably rewrite as:

$ awk '/4/{print "Eureka!"}' file
Eureka!

which is what you get just by removing everything involving the test for 3 and counter in the non-getline version (i.e. "c&&!--c;/3/{c=5}"}:

$ awk '/4/{print "Eureka!"}' file
Eureka!

i.e. again, one small requirement change required a complete redesign of the getline code, but just the absolute minimum necessary tweak to the non-getline version.

So, what you see above in the getline case was significant redesign required for every tiny requirement change, much larger amounts of handwritten code required, insidious bugs introduced during development and challenging design questions at the limits of your input space, whereas the non-getline version always had less code, was much easier to modify as requirements changed, and was much more obvious, predictable, and correct in how it would behave at the limits of the input space.