Optimize Cache lookup

[mateofio]

No description provided.

[mateofio]

This one also makes a measurable improvement in Frozen triggers. 2% of frame time is spent in std::map lookups.

Without this PR

With this PR

[mateofio]

Adding constexpr for Spec yields more small improvements.

[mateofio]

After removing the async check.

[Ghabry]

Could you discuss the advantage of "hash key via string concat" vs "hash key as tuple"?

[mateofio]

The real win here is switching from tree based map to hash based unordered_map.

As for the difference between building a composite string vs hashing tuple elements, that would need to be profiled to get a definite answer.

Unfortunately both idioms require string copies, but it's possible the tuple version uses shorter strings more and so we could trigger the small buffer optimization more often and avoid calls to malloc(). Offset that with multiple calls to your hash function instead of hashing one buffer in a single call.

The standard library doesn't support arbitrary hashing of tuples, for that we would need something like boost::hash_combine. That's the reason I did the composite string approach.

[carstene1ns]

Looks good to me.