[llvm] [X86][SSE] Improve ISD::FP_TO_*INT_SAT vector lowering (PR #199416)

[Simon Pilgrim via llvm-commits]

================
@@ -22828,8 +22822,89 @@ X86TargetLowering::LowerFP_TO_INT_SAT(SDValue Op, SelectionDAG &DAG) const {
     SDValue Res = DAG.getNode(IsSigned ? X86ISD::CVTTP2IBS : X86ISD::CVTTP2IUBS,
                               dl, VecI16VT, Src);
     return DAG.getNode(ISD::TRUNCATE, dl, DstVT, Res);
-  }
+  } else if (DstVT == MVT::v4i32 && Subtarget.hasSSE2()) {
+    unsigned SatWidth = SatVT.getScalarSizeInBits();
+    assert(SatWidth <= 32 &&
+           "Expected saturation width no wider than result element");
+
+    if (SatWidth == 32) {
+      if (IsSigned) {
+        SDValue Cvt = DAG.getNode(ISD::FP_TO_SINT, dl, DstVT, Src);
----------------
RKSimon wrote:

Don't you have to use X86ISD::CVTTP2SI to guarantee this will be handled properly?

https://github.com/llvm/llvm-project/pull/199416