In the first case, the knowledge of the left-hand side used in evaluating the right-hand side is purely an optimisation--effectively, extended context--which doesn't change the semantic outcome of the statement.

quibble: it does change the semantic outcome. The game is judged to be worth the candle, though. Example:

use Data::Dumper;
my ($a,$b,@a);
($a,$b) = split / /, "a ";
print Dumper $a,$b;
($a,$b) = @a = split / /, "a ";
print Dumper $a,$b;
[download]

Without the optimization, the default removal of trailing empty fields leaves $b undefined.