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

[Jacques Pienaar]

https://github.com/jpienaar commented:

It makes sense to me that we are using APFloat here to provide this debugging cover. APFloat is also what is used in some of the folders, and if there are bugs we should discover it as that benefits everyone. Same with improvements to its performance - I mean folks could of course still add a different pass in same form that used eigen instead (or some could already be overloading apfloat to use it). So basically I don't see performance as being main key nor do I see this as limiting that, here its about support & consistent numerics.

I'd have imagined trunc and ext would also be part of initial coverage, but perhaps that already works (although messing up rounding is easy) or perhaps as you are focusing not on heterogeneous floating point computations (yet).

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