[llvm] [RISCV] Fold FP32->BF16->FP32 (PR #69687)

[Shao-Ce SUN via llvm-commits]

================
@@ -2158,6 +2158,16 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
     }
     break;
   }
+  case RISCVISD::FP_EXTEND_BF16: {
+    SDValue V = Node->getOperand(0);
+    // fold (fp_extend_bf16 (fp_round_bf16 op)) -> op
----------------
sunshaoce wrote:

Although doing so would improve some accuracy, it's not quite appropriate. Can we add it to fast math?

https://github.com/llvm/llvm-project/pull/69687