My gut feeling when I read this was that there was going to be a pretty serious impact on program speed; but assuming my Benchmarking isn't totally silly, there is no significant efficiency loss:

use strict;
use Benchmark;

timethese(100000, {
          'Using Eval'  => 'eval { &test or die }; $@ and warn("maths 
+sux:$@");',
          'using or'    => '(&test) or warn("maths sux:$@");',
});


sub test {
        #A test which always returns true
  for(1..100) {($_>0) || return 0}
  return 1;
}
boldra@trinity:~/workspace$ ./eval_test.pl
Benchmark: timing 100000 iterations of Using Eval, using or...
Using Eval: 10 wallclock secs ( 6.46 usr +  0.06 sys =  6.52 CPU)
  using or: 10 wallclock secs ( 6.28 usr +  0.11 sys =  6.39 CPU)
[download]

It does, however, look pretty messy to (IMHO) do this for every subroutine call, and it doesn't allow me to put non-fatal errors in my subroutines.