Unpack Leaf

[treeowl]

Currently,

data Leaf k v = L !k v
data HashMap k v
    = ...
    | Leaf !Hash !(Leaf k v)
    ...

We never change the hash without changing the key, and therefore also the Leaf value. So as far as I can tell, we only get a benefit from the indirection from HashMap to Leaf when a key-value pair moves between a Leaf node and a Collision node, which we assume is rare. The rest of the time, it's purely a waste of time and space. We should try using UNPACK and check that nothing regresses unexpectedly.

[treeowl]

Specifically, unpacking makes each leaf node take four words, whereas each of them currently takes six. The extra indirection is probably hurting us significantly as well. The only thing we have to watch out for is useless boxing. If we see any of that, we'll have to hunt it down and destroy it.

[treeowl]

Sigh. In my benchmarks (which, as usual, I don't really trust), this doesn't seem to work out very well. But I don't understand why. I'll put it together as a PR in case someone else wants to try to figure it out.

[treeowl]

I'm thinking a lot of the trouble is likely caused by indirect effects. In particular, I believe we're inlining too much in general, which is probably blowing up the instruction cache.

[treeowl]

UGH! Silly me... We have -funbox-strict-fields, which already does this. Let's switch to doing it manually, because it's clearer that way.