The for version creates a (potentially large) list, whereas the while version does not.

The cost of allocating the list is offset against repeated calls to each, with the result that for small lists, for iterates faster, but on larger lists the while version wins out. A quick test show that the break even point on my system is around 8000 elements. YMMV

use Benchmark qw[ cmpthese ];

$h{ $_ } = 0 for 1 .. 100;
cmpthese( -3, { 
    for   => q[ $h{ $_ }++ for keys %h; ], 
    while => q[ $h{ $_ }++ while $_ = each %h; ] 
});

         Rate while   for
while 11937/s    --  -23%
for   15544/s   30%    --

$h{ $_ } = 0 for 1 .. 1000;
cmpthese( -3, { 
    for   => q[ $h{ $_ }++ for keys %h; ], 
    while => q[ $h{ $_ }++ while $_ = each %h; ] 
});

        Rate while   for
while 1165/s    --  -26%
for   1581/s   36%    --

$h{ $_ } = 0 for 1 .. 10000;
cmpthese( -3, { 
    for   => q[ $h{ $_ }++ for keys %h; ], 
    while => q[ $h{ $_ }++ while $_ = each %h; ] } );

        Rate   for while
for   80.0/s    --   -8%
while 86.9/s    9%    --

$h{ $_ } = 0 for 1 .. 5000;
cmpthese( -3, { 
    for   => q[ $h{ $_ }++ for keys %h; ], 
    while => q[ $h{ $_ }++ while $_ = each %h; ] 
});

       Rate while   for
while 207/s    --  -15%
for   243/s   17%    --

$h{ $_ } = 0 for 1 .. 8000;
cmpthese( -3, { 
    for   => q[ $h{ $_ }++ for keys %h; ], 
    while => q[ $h{ $_ }++ while $_ = each %h; ] 
});

       Rate   for while
for   119/s    --   -0%
while 119/s    0%    --
[download]