[LoongArch][DAG] Custom lowering for vector SETCC operations (#177904)

### Summary
This PR resolves https://github.com/llvm/llvm-project/issues/177863

llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp

diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index 6cb0983..bbf199f 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -411,7 +411,7 @@
       setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
       setOperationAction(ISD::INSERT_SUBVECTOR, VT, Legal);
 
-      setOperationAction(ISD::SETCC, VT, Legal);
+      setOperationAction(ISD::SETCC, VT, Custom);
       setOperationAction(ISD::VSELECT, VT, Legal);
       setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
     }
@@ -613,6 +613,8 @@
     return lowerVECREDUCE(Op, DAG);
   case ISD::ConstantFP:
     return lowerConstantFP(Op, DAG);
+  case ISD::SETCC:
+    return lowerSETCC(Op, DAG);
   }
   return SDValue();
 }
@@ -725,6 +727,35 @@
   return SDValue();
 }
 
+// Ensure SETCC result and operand have the same bit width; isel does not
+// support mismatched widths.
+SDValue LoongArchTargetLowering::lowerSETCC(SDValue Op,
+                                            SelectionDAG &DAG) const {
+  SDLoc DL(Op);
+  EVT ResultVT = Op.getValueType();
+  EVT OperandVT = Op.getOperand(0).getValueType();
+
+  EVT SetCCResultVT =
+      getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), OperandVT);
+
+  if (ResultVT == SetCCResultVT)
+    return Op;
+
+  assert(Op.getOperand(0).getValueType() == Op.getOperand(1).getValueType() &&
+         "SETCC operands must have the same type!");
+
+  SDValue SetCCNode =
+      DAG.getNode(ISD::SETCC, DL, SetCCResultVT, Op.getOperand(0),
+                  Op.getOperand(1), Op.getOperand(2));
+
+  if (ResultVT.bitsGT(SetCCResultVT))
+    SetCCNode = DAG.getNode(ISD::SIGN_EXTEND, DL, ResultVT, SetCCNode);
+  else if (ResultVT.bitsLT(SetCCResultVT))
+    SetCCNode = DAG.getNode(ISD::TRUNCATE, DL, ResultVT, SetCCNode);
+
+  return SetCCNode;
+}
+
 // Lower vecreduce_add using vhaddw instructions.
 // For Example:
 //  call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %a)

llvm/lib/Target/LoongArch/LoongArchISelLowering.h

diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.h b/llvm/lib/Target/LoongArch/LoongArchISelLowering.h
index 74d2b7d..126ea05 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.h
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.h
@@ -239,6 +239,7 @@
   SDValue lowerVECREDUCE_ADD(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerConstantFP(SDValue Op, SelectionDAG &DAG) const;
+  SDValue lowerSETCC(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerRotate(SDValue Op, SelectionDAG &DAG) const;
 
   bool isFPImmLegal(const APFloat &Imm, EVT VT,

llvm/test/CodeGen/LoongArch/pr177863.ll

diff --git a/llvm/test/CodeGen/LoongArch/pr177863.ll b/llvm/test/CodeGen/LoongArch/pr177863.ll
new file mode 100644
index 0000000..8edbd33
--- /dev/null
+++ b/llvm/test/CodeGen/LoongArch/pr177863.ll
@@ -0,0 +1,41 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+
+; RUN: llc --mtriple=loongarch32 -mattr=+lasx --verify-machineinstrs < %s | FileCheck %s --check-prefix=LA32
+; RUN: llc --mtriple=loongarch64 -mattr=+lasx --verify-machineinstrs < %s | FileCheck %s --check-prefix=LA64
+
+define <4 x i1> @test(<4 x i64> %shuffle2, <4 x i64> %shuffle4) {
+; LA32-LABEL: test:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    xvseq.d $xr0, $xr1, $xr0
+; LA32-NEXT:    xvpickve2gr.w $a0, $xr0, 0
+; LA32-NEXT:    vinsgr2vr.w $vr1, $a0, 0
+; LA32-NEXT:    xvpickve2gr.w $a0, $xr0, 2
+; LA32-NEXT:    vinsgr2vr.w $vr1, $a0, 1
+; LA32-NEXT:    xvpickve2gr.w $a0, $xr0, 4
+; LA32-NEXT:    vinsgr2vr.w $vr1, $a0, 2
+; LA32-NEXT:    xvpickve2gr.w $a0, $xr0, 6
+; LA32-NEXT:    vinsgr2vr.w $vr1, $a0, 3
+; LA32-NEXT:    vrepli.b $vr0, -1
+; LA32-NEXT:    vxor.v $vr0, $vr1, $vr0
+; LA32-NEXT:    ret
+;
+; LA64-LABEL: test:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    xvseq.d $xr0, $xr1, $xr0
+; LA64-NEXT:    xvpickve2gr.d $a0, $xr0, 0
+; LA64-NEXT:    vinsgr2vr.w $vr1, $a0, 0
+; LA64-NEXT:    xvpickve2gr.d $a0, $xr0, 1
+; LA64-NEXT:    vinsgr2vr.w $vr1, $a0, 1
+; LA64-NEXT:    xvpickve2gr.d $a0, $xr0, 2
+; LA64-NEXT:    vinsgr2vr.w $vr1, $a0, 2
+; LA64-NEXT:    xvpickve2gr.d $a0, $xr0, 3
+; LA64-NEXT:    vinsgr2vr.w $vr1, $a0, 3
+; LA64-NEXT:    vrepli.b $vr0, -1
+; LA64-NEXT:    vxor.v $vr0, $vr1, $vr0
+; LA64-NEXT:    ret
+entry:
+  %conv5 = trunc nuw <4 x i64> %shuffle4 to <4 x i32>
+  %conv3 = trunc nuw <4 x i64> %shuffle2 to <4 x i32>
+  %cmp = icmp ne <4 x i32> %conv5, %conv3
+  ret <4 x i1> %cmp
+}