Optimizations reduce large recursive find runtime by over 50%

[dagolden]

These commits mostly avoid Path::Class hitting File::Spec repeatedly.

In my test case of finding all .pm files in @INC (with Path::Class::Rule) I saw a run time reduction of over 50%.

[kenahoo]

I definitely like the ideas. I've cherry-picked & pushed your 9c8765a, as a separate commit. I think it's a good idea to cache the stringified version too.

But I'm not convinced we should always stringify upon creation, because some uses will create & destroy P::C objects without ever needing them as a string, and such pre-caching would be a waste in those cases. Also, if we ever start implementing the canonpath() and friends logic in P::C, we won't want it stringified, we'd probably want to work directly on the list.

[kenahoo]

I also just cherry-picked the canonpath() commit 3ffc39a.

[dagolden]

My thought on the aggressive pre-caching was that almost every method calls some perl function that forces stringification anyway. The only ones that don't are things like basename, components, and foreign representation ones.

I have a hard time imagining scenarios where people are making thousands of Path::Class objects unless they are interacting with a real filesystem. If people are doing a little bit of path manipulation, the overhead of stringification they don't need is pretty minor, whereas for people who really need it crunching through a big filesystem, the benefits of precaching are big.

We could actually go further and use the precached string in all the perl functions directly, instead of passing $self and relying on Perl to call the stringification overloading. That would be another huge optimization that just occurred to me.

[kenahoo]

Putting aside your final idea for a moment (about avoiding overloading), why is precaching actually a speedup? Doesn't it still perform stringification exactly once whether we precache or not? Precaching just seems like it calls it earlier in its life cycle.

[dagolden]

It avoids the "if" logic on each call to check if its cached. That's pretty minor compared to my point about avoiding overloading. I suspect that's what I was thinking and then stopped short for some reason.

[kenahoo]

I'm still open to these ideas, but I'm closing this pull request because it no longer applies cleanly, and because it's been a while without activity.