Make extensibleState primarily keyed by TypeRep instead of type names

[liskin]

Description

We've been using the String we get out of show . typeOf as key in extensibleState, but that has a somewhat serious bug: it shows unqualified type names, so if two modules use the same type name, their extensible states will be stored in one place and get overwritten all the time.

To fix this, the extensibleState map is now primarily keyed by the TypeRep themselves, with fallback to String for not yet deserialized data. XMonad.Core now exports showExtType which serializes type names qualified, and this is used in writeStateToFile.

A simpler fix would be to just change the serialization of type names in XMonad.Util.ExtensibleState, but I'm afraid that might slows things down: Most types used here will start with "XMonad.", and that's a lot of useless linked-list pointer jumping.

Fixes: xmonad/xmonad-contrib#94
Related: xmonad/xmonad-contrib#600

---

I must admit I don't really like the code, especially the xmonad-contrib part. There's the assumption that a Right key maps to a Right value, and Left key to Left value, but enforcing that in the types probably means adding a another field to the XState.

Checklist

• I've read CONTRIBUTING.md

• I've confirmed these changes don't belong in xmonad-contrib instead

• I updated the CHANGES.md file

[slotThe]

A simpler fix would be to just change the serialization of type names in XMonad.Util.ExtensibleState, but I'm afraid that might slows things down: Most types used here will start with "XMonad.", and that's a lot of useless linked-list pointer jumping.

As a great man once told me: have you benchmarked this? We still use String all over the place, so I can't imagine that a linked list of length 7 has that much of a performance impact

[liskin]

I didn't, true. It just feels wrong, and a couple other people suggested using Fingerprint in the past, so I went ahead and tried implementing it. It may indeed be negligible in the grand scheme of things (although that would suggest other things are also terribly inefficient). I'll try to benchmark this somehow in the coming days.

[geekosaur]

Turns out Fingerprint (and therefore the internal representation of TypeRep) is not stable across compiles, much less GHC versions, so we would need to use the String anyway unless you deal with this during serialization and deserialization. I'd just go ahead and use the String, possibly with "XMonad." prefix removed (although this is a problem if we're passed a type that isn't defined by xmonad).

[geekosaur]

LGTM