[llvm] [PowerPC][CodeGen] Expand ISD::AssertNoFPClass for ppc_fp128 (PR #152357)

[Amy Kwan via llvm-commits]

================
@@ -1966,6 +1967,14 @@ void DAGTypeLegalizer::ExpandFloatRes_FNEG(SDNode *N, SDValue &Lo,
   Hi = DAG.getNode(ISD::FNEG, dl, Hi.getValueType(), Hi);
 }
 
+void DAGTypeLegalizer::ExpandFloatRes_AssertNoFPClass(SDNode *N, SDValue &Lo,
+                                                      SDValue &Hi) {
+  SDLoc dl(N);
+  GetExpandedFloat(N->getOperand(0), Lo, Hi);
+  Lo = DAG.getNode(ISD::AssertNoFPClass, dl, Lo.getValueType(), Lo);
+  Hi = DAG.getNode(ISD::AssertNoFPClass, dl, Hi.getValueType(), Hi);
----------------
amy-kwan wrote:

Ah, yeah, I noticed I have a typo here.

As Roland mentioned, the class comes from the first element, which is what I meant:
```
SDValue PPCTargetLowering::LowerIS_FPCLASS(SDValue Op,
                                           SelectionDAG &DAG) const {
  assert(Subtarget.hasP9Vector() && "Test data class requires Power9");
  SDValue LHS = Op.getOperand(0);
  uint64_t RHSC = Op.getConstantOperandVal(1);
  SDLoc Dl(Op);
  FPClassTest Category = static_cast<FPClassTest>(RHSC);
  if (LHS.getValueType() == MVT::ppcf128) {
    // The higher part determines the value class.
    LHS = DAG.getNode(ISD::EXTRACT_ELEMENT, Dl, MVT::f64, LHS,
                      DAG.getConstant(1, Dl, MVT::i32));
  }

  return getDataClassTest(LHS, Category, Dl, DAG, Subtarget);
}
```

Unless I am misunderstanding, based on the two comments, propagating the information from element 1 is correct but it sounds like should just drop the operation for now.

@arsenm Dumb question, sorry - but what exactly does dropping the operation look like in this case? I assume the Hi and Lo would not be set then?

https://github.com/llvm/llvm-project/pull/152357