Improve Support for Dataflow Errors to Work with the IDE

[iamrecursion]

Summary

In order to allow APIs to be better designed around broken value semantics, we need to make some deep changes to the interpreter to support these in a more effective way. At the same time, we need to add a way of propagating panic information through nodes while continuing to execute in the IDE.

Value

Errors will be handled in a fashion that is amenable to both dataflow-oriented and normal programming.

Specification

• Add support for traces to broken values. This trace should contain the trace of where the value has been created, with each element being a trace.
• Increase the number of type checks to handle broken values properly in more places. This will slow the compiler down but must include:
  • Atom construction, which should produce a broken value.
  • Method resolution, which should be identity on a broken value.
  • Polyglot arguments evaluation, where the polyglot function should never be called.
  • Function calls should leave things alone except for if the function itself is an error.
• Extend the builtins DSL to handle broken values by performing typechecks in code. These should be profiled. Will need to infer to avoid these typechecks when an argument is a runtime error.
• Properly create broken values in more places (e.g. division by zero).
• Add stdlib tests for these cases.
• Create a Sentry value type that contains the panic and the stack trace. This should be an Exception.
• Add typechecks to the same places as for broken values that look for the sentry value and re-throw the sentry from the new location.

Acceptance Criteria & Test Cases

• We handle all of this properly.

[wdanilo]

I would add benchmarking as criteria with comparison to the previous version n set of benchmarks - these benchmarks (comparisons) should be part of the PR + demo day.