[llvm] [AMDGPU] Widen fminnum/fmaxnum bf16 to v2bf16 (PR #215042)

via llvm-commits llvm-commits at lists.llvm.org
Tue Aug 11 08:50:42 PDT 2026


================
@@ -7709,6 +7708,7 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::ABS:
   case ISD::FABS:
   case ISD::FNEG:
+  case ISD::FCANONICALIZE:
----------------
LU-JOHN wrote:

FCANONICALIZE was previously handled with:
```
    if (Op.getValueType() == MVT::bf16)
      return lowerScalarBF16FCanonicalize(Op, DAG);
    return splitUnaryVectorOp(Op, DAG);
```
Since `bf16` to `v2bf16` is supported by the `Promote` action now the call to `lowerScalarBF16FCanonicalize` was removed.  Moving the case label to line 7711 just joins `FCANONICALIZE` with other labels with the same:
`return splitUnaryVectorOp(Op, DAG);
`
body.

https://github.com/llvm/llvm-project/pull/215042


More information about the llvm-commits mailing list