Use address-based hashing functions for Internal hash tables

[qinsoon]

mmtk/julia@a933faa introduced address-based hashing functions to the interface, and the code in object ID uses those functions.

Julia also has its internal hash table. We currently use PTRHASH_PIN and OBJHASH_PIN to pin values that are inserted into the internal hash table. We would like to use the new functions to properly deal with those hashed pointers so we don't need to pin them.

The goal is to remove PTRHASH_PIN and OBJHASH_PIN and allow those objects to be removed. However, it is unclear how much improvement we expect to see from these changes.

[d-netto]

We should start by instrumenting PTRHASH_PIN, PTR_PIN and collecting a pinning profile, to see if this is actually a bottleneck.

Will try to take a stab at it.

[qinsoon]

I looked more into this. Both keys and values may be heap pointers for htable. So simply using address-based hashing for keys (from which hash is computed) is not sufficient to deal with htables.

As for our current approach of pinning using OBJHASH_PIN, we would have to pin both keys and values in ptrhash_(put|bp).

In the long run, I think we may want to have a dedicated type for htables that may include object references, and let GC trace those htables -- we cannot do so for the current htables. I attempted to maintain a list of htables, but it was impossible for now. Some htables are created temporarily on the stack, and die without calling htable_free, we can't know when a htable becomes unused and remove it from the list.

[qinsoon]

This is the issue that caused segfault for precompilation. Precompialtion has a step called 'uniquing types'. One type may be instantiated multiple times with different package images, and objects from each package image will reference its own type. The step of uniquing types will deduplicate those types, and make sure they all reference the same type instance. The runtime uses htable to store the objects and to update their types to the unique type.

We have object pin for ptrhash_put, but we missed pin for ptrhash_bp. mmtk/julia#100 pins the keys/values in ptrhash_bp and the precompile issue is fixed.

In the long run, we still want to get rid of the pinning as stated in #240 (comment). We would want the runtime to provide a list of htables that include object references, and let MMTk either conservatively scan the htable, or properly trace the htable (and use address-based hashing for keys).