[Mlir-commits] [mlir] [mlir][arith] Expand the scaling ops with the value of the scale (PR #217892)

[Hung-Kuan Tseng]

Tim096 wrote:

Thanks, that settles it for the case this patch adds: an `f8E5M3` scale reaches the instruction through the `extf` to `f32` in `ArithToAMDGPU`, and bits 31:23 of that `f32` are its sign and exponent, so the mantissa is dropped there while the generic expansion keeps it. That is a point for option 1 above -- take the instruction only for an `f8E8M0FNU` scale or an explicit `toward_zero` truncation, and emit the multiply or divide otherwise. I would do that as the follow-up rather than here, unless you would rather see it in this PR.

Force-pushed to `f2ab962`: the `SCHECK-SAME` nit, and a rebase past #216653, which added its converters at the same spot in `ExpandOps.cpp`; the resolution is both blocks in sequence. `check-mlir` is green on the new head.

https://github.com/llvm/llvm-project/pull/217892