mirod,
I can't share the actual data (work) but I think the following might make things a little more clear. If not, then I will live happily with the solution that I am currently constructing.

Mock up of the log file that I am working with:

2011-04-28 13:25:47 INFO [main:114] <Message><Tag attribute="value">An
+swer</Tag></Message>
2011-04-28 13:45:12 DEBUG [Populate::List:31] <Message><Tag attribute=
+"value">Answer</Tag></Message>
[download]

In other words, a Log4J standard log where the log entry is an XML document. I am parsing the log similar to the code below:

while (<$fh>) {
    chomp;
    my ($date, $time, $log_lvl, $trace, $xml) = split ' ', $_, 5;
}
[download]

For each XML document, I need to convert it to a perl data structure and do something with it. That would look something like:

my $twig = XML::Twig->new();
while (<$fh>) {
    chomp;
    my ($date, $time, $log_lvl, $trace, $xml) = split ' ', $_, 5;
    my %data_structure;
    $twig->parse($xml);
    # Build up %data_structure using $twig
}
[download]

I could easily change this code to be "elegant" as such:

while (<$fh>) {
    chomp;
    my ($date, $time, $log_lvl, $trace, $xml) = split ' ', $_, 5;
    my $data_structure = extract_data($xml);
}

sub extract_data {
    my ($xml) = @_;
    my $data = {};
    my $twig = XML::Twig->new(
        twig_handlers => {
            Message => sub { handle_message(@_, $data) }
        }
    );
    $twig->parse($xml);
    return $data;
}

sub handle_message {
    # ...
}
[download]

There is absolutely nothing wrong with this and I haven't profiled it to see that it isn't fast enough but that is my concern. I would like to inline as much as possible. So now that I have laid it out there I realize if it were someone else asking this question I would tell them to quit being falsely lazy, write it in a clear maintainable way and profile it and only worry about performance if it was unacceptable.