[llvm] [X86] Combine FRINT + FP_TO_SINT to LRINT (PR #126477)

[Phoebe Wang via llvm-commits]

================
@@ -56380,6 +56381,17 @@ static SDValue combineSIntToFP(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+static SDValue combineFPToSInt(SDNode *N, SelectionDAG &DAG,
+                               const X86Subtarget &Subtarget) {
+  EVT VT = N->getValueType(0);
+  SDValue Src = N->getOperand(0);
+  if (Src.getOpcode() == ISD::FRINT && VT.getScalarType() == MVT::i32 &&
+      Src.hasOneUse())
+    return DAG.getNode(ISD::LRINT, SDLoc(N), VT, Src.getOperand(0));
----------------
phoebewang wrote:

> Do we need to check soft float and no-sse+no-x87 to avoid creating libcalls with the wrong register? Maybe isTypeLegal for the FP type is enough?

I don't think so. It is different with #126217, 1) we don't combine TRUNC this time, so we won't eliminate ZEXT; 2) we based on the fact the out of range result is UB. We don't care the result in this case.

> Do we need to check f16 is legal?

I think it doesn't matter. 1) if the f16 is not legal, the ABI is broken, it's UB to the compiler; 2) we don't have f16 math library so far, the lowering for f16 FRINT/LRINT is not supported for now.

https://github.com/llvm/llvm-project/pull/126477