How very odd!

My testing on Win2k, ActivePerl 5.8.8 shows that the case of the data doesn't affect the timing, and that old_way is consistently just shy of 50% faster than with_tr.

Upper case data
===============
            Rate with_tr old_way
with_tr 183850/s      --    -32%
old_way 272269/s     48%      --

Lower case data
===============
            Rate with_tr old_way
with_tr 182262/s      --    -33%
old_way 270854/s     49%      --
[download]

Benchmark code:

Of course, inlining the code saves a lot of time:

              Rate   with_tr   old_way with_tr_i old_way_i
with_tr   180473/s        --      -32%      -39%      -52%
old_way   265506/s       47%        --      -10%      -29%
with_tr_i 296058/s       64%       12%        --      -20%
old_way_i 372377/s      106%       40%       26%        --
[download]

Benchmark code:

...

my $old_way_inline = <<'__EOI__';
   my @rv = split( /\./, $string, 2 );
__EOI__

my $with_tr_inline = <<'__EOI__';
   my( $id, $session ) = split( /\./, $string, 2 );
   $id =~ tr/a-z/A-Z/;
   my @rv = ( $id, $session );
__EOI__

{
   our $string = $upper_id_string;

   cmpthese($TIMES, {
      old_way   => 'use strict; use warnings; our $string; my @rv = ol
+d_way($string); 1;',
      with_tr   => 'use strict; use warnings; our $string; my @rv = wi
+th_tr($string); 1;',
      old_way_i => 'use strict; use warnings; our $string; ' . $old_wa
+y_inline . ' 1;',
      with_tr_i => 'use strict; use warnings; our $string; ' . $with_t
+r_inline . ' 1;',
   }),
}
[download]