This patch is the root of the GHash rework, all other diff will be based on it:

Reduce average load from 3.0 to 0.75

This is the big performance booster part, e.g. makes tracing a dyntopo stroke between 25% and 30% faster.

Not much to say about it, aside that it obviously increase memory footprint (about 25% - 30% too).

Add optional shrinking

I.e. ghashes/gsets can now shrink their buckets array when you remove enough entries. This remains optional and OFF by default.

Add code to use masking instead of modulo

Buckets indices are obtained from hashes by “reducing” the hash value into the valid bucket range. This can be done either by bit-masking, or using modulo operation.
The former is quicker, but requires real hashes, while the later is slower (average 10% impact on ghash operations) but can also be used as a 'fake' hashing on raw values, like e.g. indices.

In Blender currently not all ghash usages actually hash their keys, so we stick to modulo for now (masking is ifdef’ed out), we may however investigate the benefits of switching to masking with systematic very basic hashing later…

Add various missing API helpers

I.e. a way to deep-copy a ghash/gset, a way to (re-)reserve entries (i.e. manually grow or shrink the ghash after its creation), and an 'add' function to ghash matching gset one (to have a way to 'add or nop' for ghashes as well).

Various code refactoring

• Get rid of the 'hack' regarding ghash size when used as gset (it’s simpler and safer to have two structs defined here, and cast pointers as needed).
• Various re-shuffle and factorization in low-level internal code.
• Some work on hashing helpers, introducing some murmur2a-based hashing too.