Sorry for the delay in responding further to this, and thanks to everyone for their input. The fix I have committed for now was to move the .* match out into a separate regex from the = [a-zA-Z] part and this works okay, but I would prefer something slightly less ugly.

Here is some stand-alone code that demonstrates the regression, although it doesn't show quite as dramatic a slow-up as my original:

#!env perl

$l = 'x' x 50 . "\n";
$x = $l x 50;
$p = "=foo bar\n";
$_= ($x . $p) x 500 . $x;

$nx = 0;

while (1) {
    if (m/\G ( = [a-zA-Z] .* ) \n/xgc) {
        $pod .= $1;
    }
    elsif (m/\G x+ \n/xgc) {
        # match xxx lines
        $nx++;
    }
    else {
        last;
    }
}
[download]

My results show this taking 3-4 times as long under 5.20.0 as under 5.18.0:

woz$ perlbrew use 5.18.0
woz$ time perl re.pl 

real    0m0.035s
user    0m0.026s
sys    0m0.004s
woz$ perlbrew use 5.20.0
woz$ time perl re.pl 

real    0m0.128s
user    0m0.120s
sys    0m0.005s
[download]

- Andrew