Okay, I tried a bunch of different things. Two sets of the same tests, one for 3 seconds, the other 6.

use Benchmark;

timethese(-3, {
    aa => sub {
        $string = "blah+blah+blah";
        $string =~ tr/+/ /;
    },
    bb => sub {
        $string = "blah+blah+blah";
        $string =~ s/\+/ /g;
    },
    cc => sub {
        $string = "blah+blah+blah+blah+blah+blah";
        $string =~ tr/+/ /;
    },
    dd => sub {
        $string = "blah+blah+blah+blah+blah+blah";
        $string =~ s/\+/ /g;
    },
    ee => sub {
        $string = "blah";
        $string =~ tr/+/ /;
    },
    ff => sub {
        $string = "blah";
        $string =~ s/\+/ /g;
    },
});

print "\n";

timethese(-6, {
    aa => sub {
        $string = "blah+blah+blah";
        $string =~ tr/+/ /;
    },
    bb => sub {
        $string = "blah+blah+blah";
        $string =~ s/\+/ /g;
    },
    cc => sub {
        $string = "blah+blah+blah+blah+blah+blah";
        $string =~ tr/+/ /;
    },
    dd => sub {
        $string = "blah+blah+blah+blah+blah+blah";
        $string =~ s/\+/ /g;
    },
    ee => sub {
        $string = "blah";
        $string =~ tr/+/ /;
    },
    ff => sub {
        $string = "blah";
        $string =~ s/\+/ /g;
    },
});
[download]

The preceding output the following:

Benchmark: running aa, bb, cc, dd, ee, ff, each for at least 3 CPU sec
+onds...
        aa:  6 wallclock secs ( 2.91 usr +  0.10 sys =  3.01 CPU) @ 42
+9021.93/s (n=1291356)
        bb:  8 wallclock secs ( 2.95 usr +  0.05 sys =  3.00 CPU) @ 20
+8165.00/s (n=624495)
        cc:  8 wallclock secs ( 2.90 usr +  0.10 sys =  3.00 CPU) @ 40
+4326.00/s (n=1212978)
        dd:  7 wallclock secs ( 2.88 usr +  0.12 sys =  3.00 CPU) @ 13
+4898.33/s (n=404695)
        ee:  8 wallclock secs ( 3.21 usr +  0.02 sys =  3.23 CPU) @ 50
+7243.96/s (n=1638398)
        ff:  6 wallclock secs ( 3.05 usr + -0.04 sys =  3.01 CPU) @ 60
+3128.24/s (n=1815416)

Benchmark: running aa, bb, cc, dd, ee, ff, each for at least 6 CPU sec
+onds...
        aa: 11 wallclock secs ( 5.87 usr +  0.14 sys =  6.01 CPU) @ 45
+1358.24/s (n=2712663)
        bb: 11 wallclock secs ( 5.99 usr +  0.01 sys =  6.00 CPU) @ 22
+2497.00/s (n=1334982)
        cc: 12 wallclock secs ( 6.12 usr +  0.01 sys =  6.13 CPU) @ 45
+0148.29/s (n=2759409)
        dd: 11 wallclock secs ( 6.39 usr +  0.02 sys =  6.41 CPU) @ 14
+9619.66/s (n=959062)
        ee: 12 wallclock secs ( 5.97 usr +  0.03 sys =  6.00 CPU) @ 52
+0242.00/s (n=3121452)
        ff:  7 wallclock secs ( 6.01 usr +  0.00 sys =  6.01 CPU) @ 60
+7347.25/s (n=3650157)
[download]

So by the looks of the runs per second statistic that tr is indeed faster than s///. What dondelelcaro mentioned about v5.005003 having a slower version of tr could be relevant though, as that is what I'm running.

Amel - f.k.a. - kel