Let's test that hypothesis:

use Time::HiRes qw(gettimeofday tv_interval);

my @array = (1..100);
# say "@array"; # 0.332776 s
# say @array;   # 0.052496 s
# #<nothing>    # 0.052361 s

my $t0 = [gettimeofday];

for (1..10000)
{
   my @array2 = @array;
}

say tv_interval($t0);
__END__
SV = IV(0xe47dd8) at 0xe47de8
  REFCNT = 1
  FLAGS = (IOK,pIOK)
  IV = 1
SV = PVIV(0xe512d8) at 0xe47de8
  REFCNT = 1
  FLAGS = (IOK,POK,pIOK,pPOK)
  IV = 1
  PV = 0xe56050 "1"\0
  CUR = 1
  LEN = 16
[download]

So yes, interpolating the array into the string adds a PV (string) representation to the scalar.

So what causes the performance gain, though small, going from this:

use Time::HiRes qw(gettimeofday tv_interval);

my @array = (1..100);

for my $element (@array)
{
   $element .= '';
}
for my $element (@array)
{
   $element += 0;
}

my $t0 = [gettimeofday];

for (1..10000)
{
   my @array2 = @array;
}

say tv_interval($t0);
[download]

To this:

use Time::HiRes qw(gettimeofday tv_interval);

my @array = (1..100);

my $t0 = [gettimeofday];

for (1..10000)
{
   my @array2 = @array;
}

say tv_interval($t0);
[download]

That makes sense, though I am curious if perl bothers to copy both internal scalar representations. I ran my test again preceded by the below. I force each element into numeric context again after interpolating. I get almost all the performance back with this approach, but not quite all of it.

for my $element (@array)
{
   $element += 0;
}
[download]