Allow me some ramblings... Try this code as an experiment:

$a = "ab cdef  gh  ijklmn";
$b = "xyz";
while($a =~ /\G(\w\w)/g) {
    print "1: match in \$a: $1\n";
    if($b =~ /\G(\w)/g) {
        print "2: match in \$b: $1\n";
    } else {
        print "2: no match in \$b - replacing \$a\n";
        $a = "AB CD EF";
    }
    if($a =~ /\G(\s*)/g && length $1) { # The match itself never fails
        print "3: skipping whitespace in \$a\n";
    }
}
[download]

It prints:

1: match in $a: ab
2: match in $b: x
3: skipping whitespace in $a
1: match in $a: cd
2: match in $b: y
1: match in $a: ef
2: match in $b: z
3: skipping whitespace in $a
1: match in $a: gh
2: no match in $b - replacing $a
1: match in $a: AB
2: match in $b: x
3: skipping whitespace in $a
1: match in $a: CD
2: match in $b: y
3: skipping whitespace in $a
1: match in $a: EF
2: match in $b: z

Look how two regexes (1 and 3) match on $a, sharing the same pos pointer, and another one (2) matching on $b, using its own pointer, independently of the other.

What can one conclude from this?

• Each match starts at position 0, at first (re your question)
• It continues from there; when a match fails, it resets the pointer to 0 (unless you used the /c modifier). If you try to match again, it'll start again from the top.
• Each scalar has its own associated position pointer, not one per regex.
• If you change the scalar midway, it resets its associated pointer.