LLIR

LLIR is an optimizing backend which analytically generates probability graphs to maximize fallthrough contrast between branches. LLIR outputs LLVM and has a flexible input format so it can be used as a backend for many languages.

Documentation

API documentation can be generated through npm run docs and is also available online.

LLIR 101

The LLIR compiler is described in a paper (todo add link). It works around generating symmetric constraint graphs, all 'functions' having explicit Domain and Codomain nodes which can be simultaneously constrained.

When generating code with LLIR you will generally need to do the following.

1. Choose a model for how your program will be stored in memory. By default this is MemoryStore.
2. Create an ExecutionGraph with your memory store.
3. Create a Subgraph which contains your side-effect entry & exit points. LLIR offers LLVM intrinsics so this is not necessary.
4. Populate your ExecutionGraph
5. Create a generator (typically Generator.LL but you can use a different type of generator)
6. Create a Churn object which will essentially take say how you want to generate the code. E.g. to a stream, or JIT-friendly formats. Generally you'd want AutoChurn or StreamChurn