A simple benchmark indicates that the overhead isn't very much (unless the benchmark is completely flawed of course). Consider:

use strict;
use warnings;
use Benchmark qw(cmpthese);


cmpthese (
    -1, {
        BareFor => \&dofor,
        ExceptFor => \&doefor,
        ForExcept => \&dofore,
        }
    );

sub dofor {
    my $sum = 0;

    for (1 .. 1000) {
        $sum += $_;
        $sum += $_;
        $sum += $_;
    }
}


sub dofore {
    my $sum = 0;

    for (1 .. 1000) {
        eval {$sum += $_;}
    }
}


sub doefor {
    my $sum = 0;

    eval {
        for (1 .. 1000) {
            $sum += $_;
            $sum += $_;
            $sum += $_;
        }
    }
}
[download]

Prints:

            Rate   BareFor ExceptFor ForExcept
BareFor   1626/s        --       -2%      -11%
ExceptFor 1661/s        2%        --       -9%
ForExcept 1834/s       13%       10%        --
[download]

which indicates that the eval overhead is close to the cost of 2 += operations - not a whole lot in other words!