Jeff Niu

@silvasean made an attempt to write a backward analysis in TorchMLIR at some point, but I didn't hear much so I assumed he managed to make it work. The framework...

> So with a different class, how will you not duplicate a lot of code (it's not that much, but still mostly duplicated)? Some code can probably be factored out,...

I'm fine with that approach as long as it can be done without pulling out too much of the implementation code out into the public header files (maybe by explicitly...

Can you post a draft patch to phabricator? It might be easier to look at there

I think you can ignore the interprocedural bit for now. It makes things a little too complicated. The predecessor state can be augment to include the other direction, but this...

SGTM

Thanks for filing! We haven't implemented compile-time interpretation of LLVM intrinsics, and it's not clear that's a worthwhile endeavor. That said, there are other ways we can make this work.

A decorator would be useful from a library design perspective to clearly communicate to users what functions are not intended to run at compile time. However, the language can't rely...

We can inject fold/interpret hooks into the LLVMDialect. We're not bound by upstream in this regard. I suspect there will be resistance to adding intrinsics folders upstream anyways.

I have been able to reproduce this only in the playground environment.