[Mlir-commits] [mlir] [mlir] Add FP software implementation lowering pass: `arith-to-apfloat` (PR #166618)

[Renato Golin]

https://github.com/rengolin commented:

Overall, this seems ok, but it's a minor (almost toy) implementation of soft float for MLIR.

This is nice, because we can rely on LLVM to do that for us, but only as much as it supports the types (not all) and we have the operations (not all). 

So, I'm curious as to what's the inspiration for this PR, and what's the ultimate goal. Hopefully not to just leave it as is.

I also imagine this is one of those passes that are done right at the end of the pipeline, in case nothing else covers float semantics. How can we show it interoperates well with the existing LLVM lowering passes?

Finally, I also imagine this is for CPU only, for debug purposes only, etc. So the impact of the code should be small. It currently is, but the question is: what if we add a proper soft-float implementation, what would be the runtime footprint, given it will be mostly unused? 

https://github.com/llvm/llvm-project/pull/166618