Multi library support

[EvanMachusak]

• Refinements to signature resolution:

  • During generic inference, we will only infer T to be System.Any if there are no other compatible inferences. Even in cases where inferring T to be some type U with U <> System.Any whose cost is higher than inferring T to be System.Any, we would infer T = U. This prevents type inference from aggressively picking Any in too many cases.
  • Generic inference now considers model conversion functions, e.g. ToInterval(FHIR.Period) Interval<DateTime>, so that for functions like end of(argument Interval<T>) T, we can infer T to be System.DateTime for end of period where period is a FHIR.Period.
  • There are no model conversions that create lists currently, but this pattern would need to be implemented for list types also; that is not done in this PR.
  • Coercion cost computation now considers list and interval conversions immediately after ExactMatch analysis. The cost to coerce List<U> to List<V> is computed as the cost to coerce U to V; this same logic applies to interval point types as well.
  • Compatible has been split into LessCompatible and MoreCompatible, where MoreCompatible is preferential to LessCompatible. Coercions from Any to List or Interval types are considered LessCompatible than coercions to other values.
  • Coercion from List<U> to List<V> is now implemented through a query pattern.
  • List promotion and demotion now considers list "degree":
  • Non-list types have a degree of 0
  • List<T> has a degree of 1
  • List<List<T>> has a degree of 2
  • Only where the degree of the list differs by 1 can a promotion or demotion be performed. While it's possible to recursively use the query pattern to go from List<List<List<T>>> to T during a ListDemotion event, this behavior should be controlled by a currently non-existent option.
  • Model-defined conversion functions (e.g., in FHIRHelpers will now be considered during ImplicitConversion coercion costs, and those coercions will be implemented by creating a FunctionRef to those conversion functions.

• Added multi-library support:

  • Definitions can now be resolved across libraries
  • Library providers (memory, file system) were added; more testing is needed on these
  • A rudimentary StreamInspector class that attempts to detect and disambiguate CQL, JSON ELM, and XML ELM was implemented to identify libraries regardless of specific file naming conventions; this would also be useful in networking use cases.

• Implemented HL7 XML tests

  • Created a dynamic data test that reads HL7 XML test files from a test folder, compiles the expression & expectation ELM, executes, and compares the results:
  • 1472 passed
  • 50 skipped
  • Removed HL7Tests from the CoreTests project as they are now redundant.

[ewoutkramer]

Sounds like a ton of good improvements.