What is more efficient:
    for (1..1e6) { ... }
or
    @k=1..1e6; for (@k) {...} ?

Yes, I see your point, but what I was getting at in asking about real efficiency is that it is my gut feeling (unsupported by any benchmarking: there is no real application to benchmark) that if you have a hash with up to, say, about a million keys, the time to copy those keys into a separate array (as in the example code) will be trivial in comparison the time needed to acutally do with those key/value pairs (as returned by the custom each) whatever it is you want to do with them.

If you have more than 10 million keys in a hash, you're probably on the verge of moving everything into a database anyway.

The 10-100 1-10 million key range would seem (again, my gut feeling here) to be where the question of run-time efficiency would come into play, but why cross that bridge before you come to it? (Or are you already standing on that bridge?)