$xml =~ s{(?:(<!\[CDATA\[.*?\]\]>)|(<[^\s?!>]+))}{ 
  $1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} }seg;

# or ... sligthly more efficient for XMLs with most tags without the dots 

$xml =~ s{(?:(<!\[CDATA\[.*?\]\]>)|(<[^\s?!>\.]+\.[^\s?!>]+))}{ 
  $1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} }seg;

# and now even more efficient thanks to moving the &lt; outside the or

$xml =~ s{<(?:(!\[CDATA\[.*?\]\]>)|(<[^\s?!>\.]+\.[^\s?!>]+))}{
  '<'. ($1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} ) }seg;


##</code><code>##

my $XML = <<'__XML__';
<?xml version="1.0"?>
<TOP>
  <SUB>
    <![CDATA[<SOME.OTHER.TYPE>BLAH</SOME.OTHER.TYPE> ]]>
    <THIS>STUFF</THIS>
    <SOME.TYPE>T</SOME.TYPE>
    <SOME.OTHER.TYPE>BLAH</SOME.OTHER.TYPE>
    <![CDATA[<SOME.OTHER.TYPE>BLAH</SOME.OTHER.TYPE> ]]>
  </SUB>
</TOP>
__XML__

use Benchmark;

sub one {
	my $xml = $XML;

	$xml =~ s{(?:(<!\[CDATA\[.*?\]\]>)|(<[^\s?!>]+))}{ $1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} }seg;
	$xml =~ s{(?:(<!\[CDATA\[.*?\]\]>)|(<[^\s?!>]+))}{ $1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} }seg;
	$xml =~ s{(?:(<!\[CDATA\[.*?\]\]>)|(<[^\s?!>]+))}{ $1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} }seg;
	return;
}

sub two {
	my $xml = $XML;

	$xml =~ s{(?:(<!\[CDATA\[.*?\]\]>)|(<[^\s?!>\.]+\.[^\s?!>]+))}{ $1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} }seg;
	$xml =~ s{(?:(<!\[CDATA\[.*?\]\]>)|(<[^\s?!>\.]+\.[^\s?!>]+))}{ $1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} }seg;
	$xml =~ s{(?:(<!\[CDATA\[.*?\]\]>)|(<[^\s?!>\.]+\.[^\s?!>]+))}{ $1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} }seg;
	return;
}

sub three {
	my $xml = $XML;

	$xml =~ s{<(?:(!\[CDATA\[.*?\]\]>)|(<[^\s?!>\.]+\.[^\s?!>]+))}{ '<'. ($1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} ) }seg;
	$xml =~ s{<(?:(!\[CDATA\[.*?\]\]>)|(<[^\s?!>\.]+\.[^\s?!>]+))}{ '<'. ($1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} ) }seg;
	$xml =~ s{<(?:(!\[CDATA\[.*?\]\]>)|(<[^\s?!>\.]+\.[^\s?!>]+))}{ '<'. ($1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} ) }seg;
	return;
}

sub four {
	my $xml = $XML;

	$xml =~ s{<(?:(!\[CDATA\[.*?\]\]>)|(<[^\s?!>]+))}{ '<'. ($1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} ) }seg;
	$xml =~ s{<(?:(!\[CDATA\[.*?\]\]>)|(<[^\s?!>]+))}{ '<'. ($1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} ) }seg;
	$xml =~ s{<(?:(!\[CDATA\[.*?\]\]>)|(<[^\s?!>]+))}{ '<'. ($1 or do {(my $tagname = $2 ) =~ tr/.//d;$tagname;} ) }seg;
	return;
}

timethese( 100000, {
	one => \&one,
	two => \&two,
	three => \&three,
	four => \&four,
});
__END__

Benchmark: timing 100000 iterations of four, one, three, two...
      four:  2 wallclock secs ( 1.89 usr +  0.00 sys =  1.89 CPU) @ 52882.07/s (n=100000)
       one: 10 wallclock secs ( 9.53 usr +  0.00 sys =  9.53 CPU) @ 10492.08/s (n=100000)
     three:  2 wallclock secs ( 1.86 usr +  0.00 sys =  1.86 CPU) @ 53792.36/s (n=100000)
       two:  6 wallclock secs ( 6.27 usr +  0.00 sys =  6.27 CPU) @ 15959.14/s (n=100000)