Benchmarking eval of strings, since you want to include compile time:

#!/usr/bin/perl -w
use strict;

my $foo = <<'BAR';
my $test = 1;
#ifdef UNDEF
    $test = 0;
#endif
print "\$test => $test\n";
BAR

my $bar = <<'BAZ';
use constant UNDEF => 0;
my $test = 1;
if (UNDEF) {
   $test = 0;
}
print "\$test => $test\n";
BAZ

use Benchmark;

=pod

this is still wrong
timethese (100000, {
              cpp => sub { eval "perl -P -e '$foo'";},
              prl => sub { eval "perl -e '$bar'";}
             }
       );
=cut

timethese (1000, {
              cpp => sub { system "perl -P -e '$foo'";},
              prl => sub { system "perl -e '$bar'";}
             }
       );
[download]

Results are:

=pod
$ perl bmp.pl
Benchmark: timing 1000 iterations of cpp, prl...
       cpp: 88 wallclock secs ( 0.08 usr  0.54 sys + 52.56 cusr 29.98 
+csys = 83.16 CPU) @ 1612.90/s (n=1000)
       prl: 49 wallclock secs ( 0.09 usr  0.56 sys + 33.78 cusr 12.19 
+csys = 46.62 CPU) @ 1538.46/s (n=1000)
=cut
[download]

Instead of forking a new interpreter each time, we can change the tests to:

timethese (10000, {
              cpp => sub { eval "$foo"; },
              prl => sub { eval "$bar"; }});
=pod
$ perl -P bmp.pl
Benchmark: timing 10000 iterations of cpp, prl...
       cpp:  1 wallclock secs ( 0.50 usr +  0.02 sys =  0.52 CPU) @ 19
+230.77/s (n=10000)
       prl:  2 wallclock secs ( 2.40 usr +  0.03 sys =  2.43 CPU) @ 41
+15.23/s (n=10000)
=cut
[download]

Pretty much a wash, at least for your code.

Update: Corrected code, was only firing up an interpreter before. U²: Still had it wrong, showing two ways of doing it with two very different results. It looks like the interpreter with preprocessor is slower to start, but compilation is indeed faster by a pretty wide margin. Your mileage will vary. ++ariels for catching my blunders so kindly.

After Compline,
Zaxo