the point is rather that at run time the variable doesn't exist any more in your example.

Exactly. And the reason it doesn't exist is because it wasn't captured. That would have kept it alive. You can see the capture increasing the ref count in the following snippet:

$ perl -MDevel::Peek -e'
   sub f {
      my $x;
      Dump($x);
      my @a = map { sub { $x } } 1..3;
      Dump($x);
   }
   f()
'
SV = NULL(0x0) at 0x817bc90
  REFCNT = 1
  FLAGS = (PADMY)
SV = NULL(0x0) at 0x817bc90
  REFCNT = 4
  FLAGS = (PADMY)
[download]

Consider:

use strict;

{
my $a = \$a;
print "a=$a\n";

sub f {
  my $var = shift;
  print $var, "=", eval "\$$var", "\n";
};
}

&f("a");
[download]

$a references itself, so it will never be garbage-collected, yet it is not accessible in the eval.

Indeed it has nothing (or little) to do with garbage collection.

IMHO it's the dynamic scope (i.e. the scope local has), eval can access the surrounding lexpads of f's closure only as long as there is a call chain connecting them, because the "capture" of $a is missing!

use strict;
 
{
my $a = \$a;
print "a=$a\n";
 
sub f {
  my $var = shift;
  print $var, "=", eval "\$$var", "\n";
};
 
i(); # works
 
}
 
#&f("a");
  
sub g { f("a") }
sub h { g() }
sub i { h() }
[download]

Cheers Rolf