[Mlir-commits] [mlir] [mlir][arith] Fold scaling_extf and scaling_truncf with constant operands (PR #215123)

[Hung Kuan Tseng]

Tim096 wrote:

@tgymnich thanks for the suggestion, taken. If you have views on the substance
as well, the two decisions I would most like checked are: folding only when the
scale is already `f8E8M0FNU` (so that the folder does not settle #215295), and
the NaN special case, which returns a NaN of the result semantics directly rather
than widening the scale, so that the fold still succeeds when widening the input
would be lossy.

Premerge is green and the branch is mergeable; the exhaustive sweep in the
description was re-run after #214919 landed.


https://github.com/llvm/llvm-project/pull/215123