[llvm] AMDGPU: Fix selection failure on bf16 inverse sqrt (PR #172044)

[Shilei Tian via llvm-commits]

================
@@ -16508,7 +16508,9 @@ SDValue SITargetLowering::performFDivCombine(SDNode *N,
   SelectionDAG &DAG = DCI.DAG;
   SDLoc SL(N);
   EVT VT = N->getValueType(0);
-  if ((VT != MVT::f16 && VT != MVT::bf16) || !Subtarget->has16BitInsts())
+
+  // fsqrt legality correlates to rsq availability.
----------------
shiltian wrote:

not sure if this is a good check tho

https://github.com/llvm/llvm-project/pull/172044