In general, I don't think you can get there from here. Consider the transform  s/foo/fishstick/g. How do you transform the HTML fo<b>o</b>?

Assuming you constrain the replacement to have the same length as the original, something like this would do the job:

   use Regexp::Common;

   while( $html =~ s/($RE{balanced}{-parens=>'<>'})// )
   {
      $tags{$-[0]} .= $1;
   }

   $html =~ s/foo/bar/g;

   foreach my $point (sort {$b<=>$a} keys (%tags))
   {
      substr($html, $point, 0 ) = $tags{$point};
   }
[download]

For the pathological case of a tag with an attribute containing a '>' -- at this point you know as well as I do that you're into a full HTML parser:

   use HTML::Parser;

   # Remove the s///g from this one to leave tags alone
   # Alternately, specify additional methods to alter only
   # specific token types
   sub tagpush {$_ = shift; s/foo/bar/g; $tags{length($html)} .= $_ ;}
   sub txtpush { $html .= "@_";  }

   my $p = HTML::Parser->new(unbroken_text => 1,
        text_h    => [ \&txtpush, "text" ],
        default_h => [ \&tagpush, "text" ],
       );

   my $file = shift || usage();
   $p->parse_file($file) || die "Can't open file $file: $!\n";

   $html =~ s/foo/bar/g;

   foreach my $point (sort {$b<=>$a} keys (%tags))
   {
      substr($html, $point, 0 ) = $tags{$point};
   }
[download]

This last once handles cases like f<!-- -->oo, f<b>oo</b> and <img alt=">foo"> properly as well, which a token parser will not catch.