Graph vs. Linear IR: Improving our Intermediate Representation

[shreyas-omkar]

Graph vs. Linear IR: Improving our Intermediate Representation

Context

EdgeFlow currently uses a custom graph-based Intermediate Representation (IR) designed for lightweight edge and embedded targets.

As the project grows, we need to evaluate whether:

• We continue hardening and optimizing the custom IR, or
• We migrate (partially or fully) to a standardized framework like MLIR (Multi-Level IR).

Discussion Points

• What optimizations are easier in a graph-based IR vs a linear IR?
• What is the real cost of adopting MLIR for embedded-focused compilers?
• How much flexibility do we lose or gain with a standard IR?
• Is a hybrid approach viable?

Goal

Debate the long-term trade-offs of refactoring the compiler core versus maintaining a purpose-built IR optimized for edge deployments.