[Tracking Issue] Refactor AOT executor codegen into IRModule -> IRModule form

[mbaret]

The AOT executor codegen has accumulated significant technical debt and dead code during its development, resulting in a component that is difficult to understand, test and extend. Refactoring it will improve code-quality and make it easier to introduce new features to expand the capabilities of AOT code generation.

This refactor aims to observe the following principles:

• Break the logic into passes with a single responsibility.
• Introduce tests for the internal logic of AOT (not just end-to-end integration testing).
• Remove side-channels and reduce the usage of 'magic attributes' to pass information.
• De-duplicate logic where possible.

PRs

• P1. Introduce a standalone pass to lower a Relay main to TIR [AOT] Add AOTLowerMain pass to lower a Relay main into TIR #12550
• P2. Introduce an analysis pass to create the 'function metadata' [AOT] Add CreateFunctionMetadata analysis pass #13095
• P3. Introduce an analysis pass to create the 'executor metadata' [AOT] Add CreateExecutorMetadata analysis pass #13250
• P4. Refactor AOTExecutorCodegenModule to use the new passes [AOT] Refactor AOTExecutorCodegenModule using new AOT passes #12549
• P5. Remove the 'StorageRewrite' memory planning strategy (USMP only)
• P?: Move context-specific passes out of the executor codegen (AnnotateUsedMemory/ConvertForLoopsToSerial)
• P?: Remove 'device_contexts' module attribute

[ekalda]

@areusch @mbaret is there a plan to continue this work? It looks like it's stuck in an awkward place where there is some duplicated code and a set of functions that are not used outside of the tests :)

[mbaret]

It's not my intention to continue with this refactor, I've instead been working on a Relax version of AOT. To avoid duplication, I could migrate the necessary passes to the unity branch and delete them from main. Let me know if this is the preferred course of action or whether there's any community interest in picking this refactor up.

[ekalda]

Yes, I don't think the efforts to prototype Relax should justify bringing technical dept into TVM, so it would be better to remove the duplicated code. I don't think it is blocking anybody though, so it's not urgent.