[AMDGPU] Implement llvm.lrint intrinsic lowering (#98931)

This patch enabled the target-independent lowering of llvm.lrint via
GlobalISel.
For SelectionDAG, the instrinsic is custom lowered for AMDGPU.

Author: sgundapa

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

@@ -3900,6 +3900,17 @@ LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT LowerHintTy) {
     changeOpcode(MI, TargetOpcode::G_INTRINSIC_ROUNDEVEN);
     return Legalized;
   }
+  case TargetOpcode::G_INTRINSIC_LRINT:
+  case TargetOpcode::G_INTRINSIC_LLRINT: {
+    Register DstReg = MI.getOperand(0).getReg();
+    Register SrcReg = MI.getOperand(1).getReg();
+    LLT SrcTy = MRI.getType(SrcReg);
+    auto Round =
+        MIRBuilder.buildInstr(TargetOpcode::G_FRINT, {SrcTy}, {SrcReg});
+    MIRBuilder.buildFPTOSI(DstReg, Round);
+    MI.eraseFromParent();
+    return Legalized;
+  }
   case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: {
     auto [OldValRes, SuccessRes, Addr, CmpVal, NewVal] = MI.getFirst5Regs();
     Register NewOldValRes = MRI.cloneVirtualRegister(OldValRes);
@@ -4755,6 +4766,8 @@ LegalizerHelper::fewerElementsVector(MachineInstr &MI, unsigned TypeIdx,
   case G_FCEIL:
   case G_FFLOOR:
   case G_FRINT:
+  case G_INTRINSIC_LRINT:
+  case G_INTRINSIC_LLRINT:
   case G_INTRINSIC_ROUND:
   case G_INTRINSIC_ROUNDEVEN:
   case G_INTRINSIC_TRUNC:

llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

@@ -4336,6 +4336,16 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
     // targets where it is not needed.
     Results.push_back(Node->getOperand(0));
     break;
+  case ISD::LRINT:
+  case ISD::LLRINT: {
+    SDValue Arg = Node->getOperand(0);
+    EVT ArgVT = Arg.getValueType();
+    EVT ResVT = Node->getValueType(0);
+    SDLoc dl(Node);
+    SDValue RoundNode = DAG.getNode(ISD::FRINT, dl, ArgVT, Arg);
+    Results.push_back(DAG.getNode(ISD::FP_TO_SINT, dl, ResVT, RoundNode));
+    break;
+  }
   case ISD::GLOBAL_OFFSET_TABLE:
   case ISD::GlobalAddress:
   case ISD::GlobalTLSAddress:

llvm/lib/CodeGen/TargetLoweringBase.cpp

@@ -824,15 +824,15 @@ void TargetLoweringBase::initActions() {
                      Expand);
 
   // These library functions default to expand.
-  setOperationAction(
-      {ISD::FCBRT, ISD::FLOG,   ISD::FLOG2,  ISD::FLOG10,     ISD::FEXP,
-       ISD::FEXP2, ISD::FEXP10, ISD::FFLOOR, ISD::FNEARBYINT, ISD::FCEIL,
-       ISD::FRINT, ISD::FTRUNC, ISD::LRINT,  ISD::LLRINT,     ISD::FROUNDEVEN,
-       ISD::FTAN,  ISD::FACOS,  ISD::FASIN,  ISD::FATAN,      ISD::FCOSH,
-       ISD::FSINH, ISD::FTANH},
-      {MVT::f32, MVT::f64, MVT::f128}, Expand);
-
-  setOperationAction({ISD::LROUND, ISD::LLROUND},
+  setOperationAction({ISD::FCBRT,      ISD::FLOG,  ISD::FLOG2,  ISD::FLOG10,
+                      ISD::FEXP,       ISD::FEXP2, ISD::FEXP10, ISD::FFLOOR,
+                      ISD::FNEARBYINT, ISD::FCEIL, ISD::FRINT,  ISD::FTRUNC,
+                      ISD::FROUNDEVEN, ISD::FTAN,  ISD::FACOS,  ISD::FASIN,
+                      ISD::FATAN,      ISD::FCOSH, ISD::FSINH,  ISD::FTANH},
+                     {MVT::f32, MVT::f64, MVT::f128}, Expand);
+
+  // FIXME: Query RuntimeLibCalls to make the decision.
+  setOperationAction({ISD::LRINT, ISD::LLRINT, ISD::LROUND, ISD::LLROUND},
                      {MVT::f32, MVT::f64, MVT::f128}, LibCall);
 
   setOperationAction({ISD::FTAN, ISD::FACOS, ISD::FASIN, ISD::FATAN, ISD::FCOSH,

llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp

@@ -408,6 +408,9 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
 
   setOperationAction(ISD::FRINT, {MVT::f16, MVT::f32, MVT::f64}, Custom);
 
+  setOperationAction({ISD::LRINT, ISD::LLRINT}, {MVT::f16, MVT::f32, MVT::f64},
+                     Expand);
+
   setOperationAction(ISD::FREM, {MVT::f16, MVT::f32, MVT::f64}, Custom);
 
   if (Subtarget->has16BitInsts())

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

@@ -1146,6 +1146,11 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
       .scalarize(0)
       .lower();
 
+  getActionDefinitionsBuilder({G_INTRINSIC_LRINT, G_INTRINSIC_LLRINT})
+      .clampScalar(0, S16, S64)
+      .scalarize(0)
+      .lower();
+
   if (ST.has16BitInsts()) {
     getActionDefinitionsBuilder(
         {G_INTRINSIC_TRUNC, G_FCEIL, G_INTRINSIC_ROUNDEVEN})