Update: as pointed out by sgifford, the repeat operator isn't getting constant-folded, so most of this is wrong - it is the pad entry for the subexpression mentioned in the last two paragraphs that is grabbing the memory.
End update

The 10MB string in $x is freed by the undef $x, but there is another copy of the string attached to the optree (the compiled code), since constant folding will replace

  my $x = "A" x 10_000_000;
[download]

with

  my $x = "AAAAA..[10_000_000 long]..";
[download]

You can test this by putting the code in an anonymous sub, and then freeing it:

  our $x;
  my $code = sub { $x = "A" x 10_000_000 };
  showmem("start");
  &$code;
  showmem("allocated");
  undef $x;
  showmem("var freed");
  undef $code;
  showmem("code freed");
[download]

Running that here shows VMSize at each step of:

start: 2896 kB
allocated: 22436 kB
var freed: 12668 kB
code freed: 2900 kB

However, I feel that it is quite rare to have big constants like this in real code, so the simplistic approach of "fold anything that's constant" is still probably the right thing for perl to do.

Unfortunately you cannot sidestep this just by replacing the constants with variables, since then a different aspect kicks in: perl's compiler assigns an entry on the pad for the results of subexpressions, and this intermediate target then holds the extra copy of the string.

I'm not sure whether any of the pad-peeking CPAN modules show these unnamed pad entries, but the -DX debugging flag will help at least to verify their existence, eg:

  perl -DXt -we 'my $l = 100; my $a = "A" x $l;'
[download]

Hugo