Register proteus pass late

[koparasy]

Registers pass late as discussed in #79 .

Pending tasks:

• Verify correctness in systems and architectures
• Verify constant threadDim/BlockDim specialization ( PR Specializes module based on BlockDim.* values and GridDim.* values. #42 )

[davidbeckingsale]

Is there concern this will break any of our analysis/transformations? e.g. lambda specialization stuff looks for very specific patterns

[koparasy]

Is there concern this will break any of our analysis/transformations? e.g. lambda specialization stuff looks for very specific patterns

Thnx for pointing this out. You are right! It may, I have not checked the patterns we are currently "catching". Do we have tests for these patterns?

[davidbeckingsale]

Test programs yes, but no IR checking

[koparasy]

this is a low priority PR, and I am still testing things out. Deciding on which stage we are going to run the pipeline is important and we should do it asap. In terms of performance, it makes more sense to register it late.

[koparasy]

I will take that back. It is possible to run the annotation analysis early in the pipeline, and the registering of the LLVM IR files late in the pipeline. This may give us both capabilities.

[davidbeckingsale]

If the GitLab passes then I'm not worried :)

[koparasy]

Passed :)

[ggeorgakoudis]

An issue with late registration is that we may miss JIT annotated functions that have been inlined.

We don't test for that, e.g., marking always_inline a JIT-annotated function.

[koparasy]

That's a good point