The Ko evaluator and weaver pipelines comprise multiple steps, the first few of which are shared. Both pipelines begin by parsing the Ko source code and producing libraries of named circuit wiring diagrams, including resolutions for all references across circuits and packages.
This stage is the first place, where dataflow-based (and still type-free) automatic program rewriting and analysis can be performed.
The evaluator and weaver take different paths from here (described in the following sections). However, they both reuse the same dataflow engine which guides the execution of custom logic over circuit programs.
The evaluator executes the wired circuit program optimistically “on-the-fly”, before verifying that the entire program is type-safe. It simulates the flow of values as the program execution unfolds and, in an efficient manner, accummulates type information about the whole program as execution unfolds. Type inconsistencies are caught as they occur and result in a Ko stack trace and an error message.
The weaving process comprises two stages: type-inference and code generation.
In the type-inference stage, the weaver executes a Ko program by flowing types (instead of values) across the circuit wires connecting the output and input of transformations in the program.
As a result, it produces an Intermediate Representation (IR) of the Ko program which is represented as a list of circuits (together with their wiring diagrams) as well as concrete deep-type annotations on all arguments, return values and internal wires of each circuit.
This IR can then be used to implement straightforward code-generators that convert the IR representation (type annotated circuits) into fully-statically typed code in any general-purpose programming language, like Go, Java, C++, etc.
Our weaver (still in development) currently supports code generation of Go code. After code-generation completes, a direct invocation of the Go compiler produces the final binary. This is handled by the Ko compiler for convenience and encapsulation.