In this sort of situation, you'll want to aim for the code inside the 5,000,000 iterations to be as minimal as possible.

I ran a few commandline tests comparing the smartmatch with a regex. The regex was 5-10 times faster. Here's a typical run:

$ time perl -Mstrict -Mwarnings -E '
my @x = ((q{AXXX}) x 4000, qw{BXXX CXXX});
my $y = q{BXXX};
my $c = 0;
for (1 .. 5000) {       
    $y ~~ @x && ++$c;
}                
say qq{count=$c};
'                    
count=5000

real    0m1.128s
user    0m1.122s
sys    0m0.004s

$ time perl -Mstrict -Mwarnings -e '
my @x = ((q{AXXX}) x 4000, qw{BXXX CXXX});
my $y = q{BXXX};
my $c = 0;
my $z = join q{|} => @x;
for (1 .. 5000) {
    $z =~ m{\b$y\b} && ++$c;
}                
print qq{count=$c\n};
'
count=5000

real    0m0.142s
user    0m0.138s
sys    0m0.003s
[download]

See also: Benchmark