XLA as a backend in custom deep learning framework

[adendek]

Dear experts,

I have one question regarding xla. I checked xla tutorials and I couldn't find any hint there.
I've been implementing my deep learning framework from scratch, which idea is to be as "understandable" as possible. Right now, it supports a simplified object-oriented version of autograd and some functionalities to create neural networks.

I am wondering how hard would be to add XLA as a GPU/TPU backend.
Could you please give me a hint on that?