What if we skipped checking the prototype and set up both @_, and $a and $b?

sub reduce(&@) {
   my $f = shift;
   my $a = shift;
   my $b;

   our @args;
   (local *_, local *args) = (sub { \@_ }->($a, $b), \@_);

   no strict 'refs';
   my $caller = caller();
   local *{$caller."::a"} = \$a;
   local *{$caller."::b"} = \$b;

   $b = shift(@args), $a = &$f while @args;
   return $a;
}
[download]

Turns out it's negligeably faster for moderately long lists.

          Rate ike1_p ike1_n ike2_p ike1_i ike2_i ike2_n
ike1_p 6.34/ms     --    -2%    -2%    -4%    -6%     -7%
ike1_n 6.46/ms     2%     --    -1%    -3%    -4%     -5%
ike2_p 6.50/ms     3%     1%     --    -2%    -3%     -4%
ike1_i 6.63/ms     5%     3%     2%     --    -1%     -2%
ike2_i 6.72/ms     6%     4%     3%     1%     --     -1%
ike2_n 6.79/ms     7%     5%     4%     2%     1%      --

(I)nline:        reduce { $a * $b }     1..100;
(P)rototype:     reduce \&prototyped,   1..100;
(N)o prototype:  reduce \&unprototyped, 1..100;
[download]

Full Results

Benchmark code