[llvm] [GlobalISel] Improve bf16 converts with fast-math flags. (PR #200741)

[Cullen Rhodes via llvm-commits]

================
@@ -2161,9 +2161,10 @@ Register LegalizerHelper::coerceToScalar(Register Val) {
 }
 
 void LegalizerHelper::widenScalarSrc(MachineInstr &MI, LLT WideTy,
-                                     unsigned OpIdx, unsigned ExtOpcode) {
+                                     unsigned OpIdx, unsigned ExtOpcode,
+                                     unsigned Flags) {
   MachineOperand &MO = MI.getOperand(OpIdx);
-  auto ExtB = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MO});
+  auto ExtB = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MO}, Flags);
----------------
c-rhodes wrote:

is there any harm in unconditionally forwarding flags here?

https://github.com/llvm/llvm-project/pull/200741