[Mlir-commits] [mlir] [mlir] [arith] Fold trunc(extremum(ext(lhs), ext(rhs))) (PR #214658)

[Chuanqi Xu]

ChuanqiXu9 wrote:

> I don't think this rewrite is sound in the current form, repro with counterexamples: https://gist.github.com/kuhar/0b26ec9569195ea1d18bf0da62e27734

Thanks for pointing this out. Learned a lot. After spending the whole morning for this, I realized that if we want make this sound, we have to add `fastmath<nnan>` to the original example to make it work:

```
func.func @narrowExtremaOfExtf(%arg0: f16, %arg1: f16)
    -> f16 {
  %lhs = arith.extf %arg0 : f16 to f32
  %rhs = arith.extf %arg1 : f16 to f32
  %maximum = arith.maximumf %lhs, %rhs fastmath<nnan> : f32
  %maximumTrunc = arith.truncf %maximum : f32 to f16
  return %maximumTrunc : f16
}
```

this makes it less useful but still useful if the end users use fastmath mode. And it can be even better if we can introduce a new mode to allow qNaN but not sNaN.

https://github.com/llvm/llvm-project/pull/214658