I think tail call elimination aka optimization is wrong here , because it's not optimized in Perl (you noted it's slow)

One can call it tail-call (without elimination)

TCE in languages like LISP optimize calls to jumped and can replace the necessity for loop constructs.

> by assigning it to $_[0]->{state}), but it never ends up using it.

It does, but only once as Boolean test.

> goto &sub is slower

Really? I expected the same...

But I kind of remember that the implementation is awkwardly cleaning the frame right after it was created...

> I think we could use a loop

I had a similar idea, but I would return names not references, and store current-state inside the loop. This would improve readability a lot and allow to easily trace/log the execution.

Will add tested code tomorrow

  ... To be continued ...

update

use v5.14;
use warnings;


my ($state,$last);

sub initial_state {
    say "*** Initializing";
    return "some_state";
}

sub some_state {
  my $in = shift;
  if ( not $in ) {
    return "some_state";
  } else {
    return "final_state";
  }
}

sub final_state {
  say "*** Finalizing";
  return undef;
}


$state = "initial_state";

my @input = (0,0,0,1);
my $log = 1;
while ( $state ) {
  $last  = $state;
  my $in = shift @input;
  no strict 'refs';
  $state   = $state->($in);
  say "$last \t--($in)-> \t$state" if $log && $state;
}
[download]

-->

*** Initializing
initial_state     --(0)->     some_state
some_state     --(0)->     some_state
some_state     --(0)->     some_state
some_state     --(1)->     final_state
*** Finalizing
[download]