get_cache_key: replace sorted() with immutable Map

[matthiasdiener]

sorted() is currently the function with the highest impact on compile performance in a mirgecom-y3prediction rhs compile on M1 (i.e., it is the top entry in cProfile when sorting by tottime, just above islpy's casting wrapper). This PR reduces the number of calls to sorted by 80%.

Other options might be:

• use frozenset
• use hash(immutables.Map(...))
• use frozendict

[inducer]

Wait... this change reduces compile time by 80%?

[matthiasdiener]

Wait... this change reduces compile time by 80%?

No, it reduces the number of calls to sorted by 80% (7.5 million to 1 Million). Overall compile time is improved by ~5%.

[inducer]

Sounds good. Every little bit helps! :slightly_smiling_face:

[inducer]

(And thanks!)

[inducer]

One additional item is that, in many situations, there won't be a **kwargs component to deal with, and @optimize_mapper could specialize that away. For expensive mappers, this may be a good idea.