It is basically the same thing as the last "thing" tested with the previous test operator but without having to re-test it (which is apparently expensive). I remember reading that in a book.

Update: I think it is called 'chaining' tests.

Update2:

To save the overhead of a low-level stat() system call, Perl caches the previously-examined inode in a special variable called simply '_'; so we can do tests like this:

if ( -e $file && -r _ && -x _) {
   # $file exists and is both readable and executable
   ...
}
[download]

It is a special filehandle used with stat, lstat, and the -X test operators. See http://perldoc.perl.org/functions/-X.html.

Answering the question:

Normal filehandle names in Perl match /[a-z_][a-z0-9_]*/, so it's a legal name, akin to a special variable like $", but a filehandle instead of a scalar. Consider print STDERR "foo\n"... no sigil.

Not answering the question but informative:

It's known as the 'magical underscore filehandle', and it exists for efficiency reasons.

Generally, OS's store a file's metadata all bunched together, in what C programmers would call a 'structure'. Whenever a piece of Perl code requests a part of a file's metadata (e.g. checking its permissions, or checking its size), a system call is made (under *nix, usually called stat() or a derivative like stat64()) that returns the file's complete metadata structure. In the example you posted this is obviously wasteful, so whenever a file test operator is used Perl keeps the returned metadata structure around in memory, in case further tests need to be performed against the same file.

You can learn a lot by using a system call tracer against your Perl code, like strace under Linux. For instance:

# file1.pl
my $size = -s $ARGV[0];
my $time = -M $ARGV[0];
[download]

$ strace perl file1.pl
[download]

# file2.pl
my $size = -s $ARGV[0];
my $time = -M _;
[download]

$ strace perl file2.pl
[download]

... and compare the number of calls to stat() or fstat64() or whatever.