What is interesting is that the code that you benchmarked and the code that I benchmarked, doesn't actually supply arguments to the either function, which really is the point after all.

Changing the cmpthese call to;

cmpthese(100_000, {
 shift => sub { foo(1, "foo") },
 inline => sub { bar(1, "bar") }
});
[download]

The results are the same for either method;

bruce:0:~/tmp $ ./535237.pl 
          Rate  shift inline
shift  15456/s     --    -0%
inline 15528/s     0%     --
[download]