An essential element of this program is that, like grep, it prints out only the lines that match the pattern. You might think we'd use the -n option to suppress the default output of lines. However, what is unusual about this sed script is that it creates an input/output loop, controlling when a line is output or not.

The logic of this script is to first look for the pattern on one line and print the line if it matches. If no match is found, we read another line into the pattern space (as in previous multiline scripts). Then we copy the two-line pattern space to the hold space for safekeeping. Now the new line that was read into the pattern space previously could match the search pattern on its own, so the next match we attempt is on the second line only. Once we've determined that the pattern is not found on either the first or second lines, we remove the newline between the two lines and look for it spanning those lines.

The script is designed to accept arguments from the command line. The first argument is the search pattern. All other command-line arguments will be interpreted as filenames. Let's look at the entire script before analyzing it:

#! /bin/sh
# phrase -- search for words across lines
# $1 = search string; remaining args = filenames
search=$1
shift
for file 
do
sed '
/'"$search"'/b
N
h
s/.*\n//
/'"$search"'/b
g
s/ *\n/ /
/'"$search"'/{
g
b
}
g
D' $file 
done

The sed script tries to match the search string at three different points, each marked by the address that looks for the search pattern. The first line of the script looks for the search pattern on a line by itself:

/'"$search"'/b

If the search pattern matches the line, the branch command, without a label, transfers control to the bottom of the script where the line is printed. This makes use of sed's normal control-flow so that the next input line is read into the pattern space and control then returns to the top of the script. The branch command is used in the same way each time we try to match the pattern.

If a single input line does not match the pattern, we begin our next procedure to create a multiline pattern space. It is possible that the new line, by itself, will match the search string. It may not be apparent why this step is necessary--why not just immediately look for the pattern anywhere across two lines? The reason is that if the pattern is actually matched on the second line, we'd still output the pair of lines. In other words, the user would see the line preceding the matched line and might be confused by it. This way we output the second line by itself if that is what matches the pattern.

N
h
s/.*\n//
/'"$search"'/b

The Next command appends the next input line to the pattern space. The hold command places a copy of the two-line pattern space into the hold space. The next action will change the pattern space and we want to preserve the original intact. Before looking for the pattern, we use the substitute command to remove the previous line, up to and including the embedded newline. There are several reasons for doing it this way and not another way, so let's consider some of the alternatives. You could write a pattern that matches the search pattern only if it occurs after the embedded newline:

/\n.*'"$search"'/b

However, if a match is found, we don't want to print the entire pattern space, just the second portion of it. Using the above construct would print both lines when only the second line matches.

You might want to use the Delete command to remove the first line in the pattern space before trying to match the pattern. A side effect of the Delete command is a change in flow control that would resume execution at the top of the script. (The Delete command could conceivably be used but not without changing the logic of this script.)

So, we try to match the pattern on the second line, and if that is unsuccessful, then we try to match it across two lines:

g
s/ *\n/ /
/'"$search"'/{
g
b
}

The get command retrieves a copy of the original two-line pair from the hold space, overwriting the line we had worked with in the pattern space. The substitute command replaces the embedded newline and any spaces preceding it with a single space. Then we attempt to match the pattern. If the match is made, we don't want to print the contents of the pattern space, but rather get the duplicate from the hold space (which preserves the newline) and print it. Thus, before branching to the end of the script, the get command retrieves the copy from the hold space.

The last part of the script is executed only if the pattern has not been matched.

g
D

The get command retrieves the duplicate, that preserves the newline, from the hold space. The Delete command removes the first line in the pattern space and passes control back to the top of the script. We delete only the first part of the pattern space, instead of clearing it, because after reading another input line, it is possible to match the pattern spanning across both lines.

Here's the result when the program is run on a sample file:

$ phrase "the procedure is followed" sect3
If a pattern is followed by a \f(CW!\fP, then the procedure
is followed for all lines that do not match the pattern.
so that the procedure is followed only if there is no match.

As we mentioned at the outset, writing sed scripts is a good primer for programming. In the chapters that follow, we will be looking at the awk programming language. You will see many similarities to sed to make you comfortable but you will see a broader range of constructs for writing useful programs. As you begin trying to do more complicated tasks with sed, the scripts get so convoluted as to make them difficult to understand. One of the advantages of awk is that it handles complexity better, and once you learn the basics, awk scripts are easier to write and understand.

6.4. Advanced Flow Control Commands		7. Writing Scripts for awk