[SDag][ARM][RISCV] Allow lowering CTPOP into a libcall

llvm/include/llvm/CodeGen/BasicTTIImpl.h

@@ -2503,7 +2503,7 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
         return (LT.first * 2);
       else
         return (LT.first * 1);
-    } else if (!TLI->isOperationExpand(ISD, LT.second)) {
+    } else if (TLI->isOperationCustom(ISD, LT.second)) {
       // If the operation is custom lowered then assume
       // that the code is twice as expensive.
       return (LT.first * 2);

llvm/include/llvm/IR/RuntimeLibcalls.def

@@ -85,6 +85,9 @@ HANDLE_LIBCALL(NEG_I64, "__negdi2")
 HANDLE_LIBCALL(CTLZ_I32, "__clzsi2")
 HANDLE_LIBCALL(CTLZ_I64, "__clzdi2")
 HANDLE_LIBCALL(CTLZ_I128, "__clzti2")
+HANDLE_LIBCALL(CTPOP_I32, "__popcountsi2")
+HANDLE_LIBCALL(CTPOP_I64, "__popcountdi2")
+HANDLE_LIBCALL(CTPOP_I128, "__popcountti2")
 
 // Floating-point
 HANDLE_LIBCALL(ADD_F32, "__addsf3")

llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp

@@ -129,7 +129,8 @@ class SelectionDAGLegalize {
                                      ArrayRef<int> Mask) const;
 
   std::pair<SDValue, SDValue> ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
-                        TargetLowering::ArgListTy &&Args, bool isSigned);
+                                            TargetLowering::ArgListTy &&Args,
+                                            bool IsSigned, EVT RetVT);
   std::pair<SDValue, SDValue> ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
 
   void ExpandFPLibCall(SDNode *Node, RTLIB::Libcall LC,
@@ -150,6 +151,9 @@ class SelectionDAGLegalize {
                           RTLIB::Libcall Call_F80, RTLIB::Libcall Call_F128,
                           RTLIB::Libcall Call_PPCF128,
                           SmallVectorImpl<SDValue> &Results);
+  SDValue ExpandBitCountingLibCall(SDNode *Node, RTLIB::Libcall CallI32,
+                                   RTLIB::Libcall CallI64,
+                                   RTLIB::Libcall CallI128);
   void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
   void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
 
@@ -2114,9 +2118,10 @@ SDValue SelectionDAGLegalize::ExpandSPLAT_VECTOR(SDNode *Node) {
 // register, return the lo part and set the hi part to the by-reg argument in
 // the first.  If it does fit into a single register, return the result and
 // leave the Hi part unset.
-std::pair<SDValue, SDValue> SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
-                                            TargetLowering::ArgListTy &&Args,
-                                            bool isSigned) {
+std::pair<SDValue, SDValue>
+SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
+                                    TargetLowering::ArgListTy &&Args,
+                                    bool IsSigned, EVT RetVT) {
   EVT CodePtrTy = TLI.getPointerTy(DAG.getDataLayout());
   SDValue Callee;
   if (const char *LibcallName = TLI.getLibcallName(LC))
@@ -2127,7 +2132,6 @@ std::pair<SDValue, SDValue> SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall L
                                 Node->getOperationName(&DAG));
   }
 
-  EVT RetVT = Node->getValueType(0);
   Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
 
   // By default, the input chain to this libcall is the entry node of the
@@ -2147,7 +2151,7 @@ std::pair<SDValue, SDValue> SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall L
     InChain = TCChain;
 
   TargetLowering::CallLoweringInfo CLI(DAG);
-  bool signExtend = TLI.shouldSignExtendTypeInLibCall(RetTy, isSigned);
+  bool signExtend = TLI.shouldSignExtendTypeInLibCall(RetTy, IsSigned);
   CLI.setDebugLoc(SDLoc(Node))
       .setChain(InChain)
       .setLibCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee,
@@ -2183,7 +2187,8 @@ std::pair<SDValue, SDValue> SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall L
     Args.push_back(Entry);
   }
 
-  return ExpandLibCall(LC, Node, std::move(Args), isSigned);
+  return ExpandLibCall(LC, Node, std::move(Args), isSigned,
+                       Node->getValueType(0));
 }
 
 void SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
@@ -2259,6 +2264,50 @@ void SelectionDAGLegalize::ExpandArgFPLibCall(SDNode* Node,
   ExpandFPLibCall(Node, LC, Results);
 }
 
+SDValue SelectionDAGLegalize::ExpandBitCountingLibCall(
+    SDNode *Node, RTLIB::Libcall CallI32, RTLIB::Libcall CallI64,
+    RTLIB::Libcall CallI128) {
+  RTLIB::Libcall LC;
+  switch (Node->getSimpleValueType(0).SimpleTy) {
+  default:
+    llvm_unreachable("Unexpected request for libcall!");
+  case MVT::i32:
+    LC = CallI32;
+    break;
+  case MVT::i64:
+    LC = CallI64;
+    break;
+  case MVT::i128:
+    LC = CallI128;
+    break;
+  }
+
+  // Bit-counting libcalls have one unsigned argument and return `int`.
+  // Note that `int` may be illegal on this target; ExpandLibCall will
+  // take care of promoting it to a legal type.
+  SDValue Op = Node->getOperand(0);
+  EVT IntVT =
+      EVT::getIntegerVT(*DAG.getContext(), DAG.getLibInfo().getIntSize());
+
+  TargetLowering::ArgListEntry Arg;
+  EVT ArgVT = Op.getValueType();
+  Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
+  Arg.Node = Op;
+  Arg.Ty = ArgTy;
+  Arg.IsSExt = TLI.shouldSignExtendTypeInLibCall(ArgTy, /*IsSigned=*/false);
+  Arg.IsZExt = !Arg.IsSExt;
+
+  SDValue Res = ExpandLibCall(LC, Node, TargetLowering::ArgListTy{Arg},
+                              /*IsSigned=*/true, IntVT)
+                    .first;
+
+  // If ExpandLibCall created a tail call, the result was already
+  // of the correct type. Otherwise, we need to sign extend it.
+  if (Res.getValueType() != MVT::Other)
+    Res = DAG.getSExtOrTrunc(Res, SDLoc(Node), Node->getValueType(0));
+  return Res;
+}
+
 /// Issue libcalls to __{u}divmod to compute div / rem pairs.
 void
 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
@@ -4993,19 +5042,12 @@ void SelectionDAGLegalize::ConvertNodeToLibcall(SDNode *Node) {
                                        RTLIB::MUL_I64, RTLIB::MUL_I128));
     break;
   case ISD::CTLZ_ZERO_UNDEF:
-    switch (Node->getSimpleValueType(0).SimpleTy) {
-    default:
-      llvm_unreachable("LibCall explicitly requested, but not available");
-    case MVT::i32:
-      Results.push_back(ExpandLibCall(RTLIB::CTLZ_I32, Node, false).first);
-      break;
-    case MVT::i64:
-      Results.push_back(ExpandLibCall(RTLIB::CTLZ_I64, Node, false).first);
-      break;
-    case MVT::i128:
-      Results.push_back(ExpandLibCall(RTLIB::CTLZ_I128, Node, false).first);
-      break;
-    }
+    Results.push_back(ExpandBitCountingLibCall(
+        Node, RTLIB::CTLZ_I32, RTLIB::CTLZ_I64, RTLIB::CTLZ_I128));
+    break;
+  case ISD::CTPOP:
+    Results.push_back(ExpandBitCountingLibCall(
+        Node, RTLIB::CTPOP_I32, RTLIB::CTPOP_I64, RTLIB::CTPOP_I128));
     break;
   case ISD::RESET_FPENV: {
     // It is legalized to call 'fesetenv(FE_DFL_ENV)'. On most targets

llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp

@@ -4012,15 +4012,35 @@ void DAGTypeLegalizer::ExpandIntRes_ABD(SDNode *N, SDValue &Lo, SDValue &Hi) {
   SplitInteger(Result, Lo, Hi);
 }
 
-void DAGTypeLegalizer::ExpandIntRes_CTPOP(SDNode *N,
-                                          SDValue &Lo, SDValue &Hi) {
-  SDLoc dl(N);
+void DAGTypeLegalizer::ExpandIntRes_CTPOP(SDNode *N, SDValue &Lo, SDValue &Hi) {
+  SDValue Op = N->getOperand(0);
+  EVT VT = N->getValueType(0);
+  SDLoc DL(N);
+
+  if (TLI.getOperationAction(ISD::CTPOP, VT) == TargetLoweringBase::LibCall) {
+    RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
+    if (VT == MVT::i32)
+      LC = RTLIB::CTPOP_I32;
+    else if (VT == MVT::i64)
+      LC = RTLIB::CTPOP_I64;
+    else if (VT == MVT::i128)
+      LC = RTLIB::CTPOP_I128;
+    assert(LC != RTLIB::UNKNOWN_LIBCALL && TLI.getLibcallName(LC) &&
+           "LibCall explicitly requested, but not available");
+    TargetLowering::MakeLibCallOptions CallOptions;
+    EVT IntVT =
+        EVT::getIntegerVT(*DAG.getContext(), DAG.getLibInfo().getIntSize());
+    SDValue Res = TLI.makeLibCall(DAG, LC, IntVT, Op, CallOptions, DL).first;
+    SplitInteger(DAG.getSExtOrTrunc(Res, DL, VT), Lo, Hi);
+    return;
+  }
+
   // ctpop(HiLo) -> ctpop(Hi)+ctpop(Lo)
-  GetExpandedInteger(N->getOperand(0), Lo, Hi);
+  GetExpandedInteger(Op, Lo, Hi);
   EVT NVT = Lo.getValueType();
-  Lo = DAG.getNode(ISD::ADD, dl, NVT, DAG.getNode(ISD::CTPOP, dl, NVT, Lo),
-                   DAG.getNode(ISD::CTPOP, dl, NVT, Hi));
-  Hi = DAG.getConstant(0, dl, NVT);
+  Lo = DAG.getNode(ISD::ADD, DL, NVT, DAG.getNode(ISD::CTPOP, DL, NVT, Lo),
+                   DAG.getNode(ISD::CTPOP, DL, NVT, Hi));
+  Hi = DAG.getConstant(0, DL, NVT);
 }
 
 void DAGTypeLegalizer::ExpandIntRes_CTTZ(SDNode *N,

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

@@ -9390,8 +9390,9 @@ SDValue TargetLowering::expandCTTZ(SDNode *Node, SelectionDAG &DAG) const {
                         !isOperationLegalOrCustomOrPromote(ISD::XOR, VT)))
     return SDValue();
 
-  // Emit Table Lookup if ISD::CTLZ and ISD::CTPOP are not legal.
-  if (!VT.isVector() && isOperationExpand(ISD::CTPOP, VT) &&
+  // Emit Table Lookup if ISD::CTPOP used in the fallback path below is going
+  // to be expanded or converted to a libcall.
+  if (!VT.isVector() && !isOperationLegalOrCustomOrPromote(ISD::CTPOP, VT) &&
       !isOperationLegal(ISD::CTLZ, VT))
     if (SDValue V = CTTZTableLookup(Node, DAG, dl, VT, Op, NumBitsPerElt))
       return V;

llvm/lib/Target/ARM/ARMISelLowering.cpp

@@ -1221,7 +1221,8 @@ ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::ROTR, VT, Expand);
   }
   setOperationAction(ISD::CTTZ,  MVT::i32, Custom);
-  setOperationAction(ISD::CTPOP, MVT::i32, Expand);
+  setOperationAction(ISD::CTPOP, MVT::i32, LibCall);
+  setOperationAction(ISD::CTPOP, MVT::i64, LibCall);
   if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only()) {
     setOperationAction(ISD::CTLZ, MVT::i32, Expand);
     setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, LibCall);

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -395,7 +395,12 @@ RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
     if (Subtarget.is64Bit())
       setOperationAction({ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF}, MVT::i32, Custom);
   } else {
-    setOperationAction({ISD::CTTZ, ISD::CTPOP}, XLenVT, Expand);
+    setOperationAction(ISD::CTTZ, XLenVT, Expand);
+    if (Subtarget.is64Bit())
+      setOperationAction(ISD::CTPOP, MVT::i128, LibCall);
+    else
+      setOperationAction(ISD::CTPOP, MVT::i32, LibCall);
+    setOperationAction(ISD::CTPOP, MVT::i64, LibCall);
   }
 
   if (Subtarget.hasStdExtZbb() || Subtarget.hasVendorXTHeadBb() ||

llvm/test/CodeGen/ARM/popcnt.ll

@@ -324,76 +324,22 @@ define i32 @ctpop16(i16 %x) nounwind readnone {
 define i32 @ctpop32(i32 %x) nounwind readnone {
 ; CHECK-LABEL: ctpop32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    ldr r1, .LCPI22_0
-; CHECK-NEXT:    ldr r2, .LCPI22_3
-; CHECK-NEXT:    and r1, r1, r0, lsr #1
-; CHECK-NEXT:    ldr r12, .LCPI22_1
-; CHECK-NEXT:    sub r0, r0, r1
-; CHECK-NEXT:    ldr r3, .LCPI22_2
-; CHECK-NEXT:    and r1, r0, r2
-; CHECK-NEXT:    and r0, r2, r0, lsr #2
-; CHECK-NEXT:    add r0, r1, r0
-; CHECK-NEXT:    add r0, r0, r0, lsr #4
-; CHECK-NEXT:    and r0, r0, r12
-; CHECK-NEXT:    mul r1, r0, r3
-; CHECK-NEXT:    lsr r0, r1, #24
-; CHECK-NEXT:    mov pc, lr
-; CHECK-NEXT:    .p2align 2
-; CHECK-NEXT:  @ %bb.1:
-; CHECK-NEXT:  .LCPI22_0:
-; CHECK-NEXT:    .long 1431655765 @ 0x55555555
-; CHECK-NEXT:  .LCPI22_1:
-; CHECK-NEXT:    .long 252645135 @ 0xf0f0f0f
-; CHECK-NEXT:  .LCPI22_2:
-; CHECK-NEXT:    .long 16843009 @ 0x1010101
-; CHECK-NEXT:  .LCPI22_3:
-; CHECK-NEXT:    .long 858993459 @ 0x33333333
+; CHECK-NEXT:    b __popcountsi2
   %count = tail call i32 @llvm.ctpop.i32(i32 %x)
   ret i32 %count
 }
 
-define i32 @ctpop64(i64 %x) nounwind readnone {
+define i64 @ctpop64(i64 %x) nounwind readnone {
 ; CHECK-LABEL: ctpop64:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    .save {r4, lr}
-; CHECK-NEXT:    push {r4, lr}
-; CHECK-NEXT:    ldr r2, .LCPI23_0
-; CHECK-NEXT:    ldr r3, .LCPI23_3
-; CHECK-NEXT:    and r4, r2, r0, lsr #1
-; CHECK-NEXT:    and r2, r2, r1, lsr #1
-; CHECK-NEXT:    sub r0, r0, r4
-; CHECK-NEXT:    sub r1, r1, r2
-; CHECK-NEXT:    and r4, r0, r3
-; CHECK-NEXT:    and r2, r1, r3
-; CHECK-NEXT:    and r0, r3, r0, lsr #2
-; CHECK-NEXT:    and r1, r3, r1, lsr #2
-; CHECK-NEXT:    add r0, r4, r0
-; CHECK-NEXT:    ldr lr, .LCPI23_1
-; CHECK-NEXT:    add r1, r2, r1
-; CHECK-NEXT:    ldr r12, .LCPI23_2
-; CHECK-NEXT:    add r0, r0, r0, lsr #4
-; CHECK-NEXT:    and r0, r0, lr
-; CHECK-NEXT:    add r1, r1, r1, lsr #4
-; CHECK-NEXT:    mul r2, r0, r12
-; CHECK-NEXT:    and r0, r1, lr
-; CHECK-NEXT:    mul r1, r0, r12
-; CHECK-NEXT:    lsr r0, r2, #24
-; CHECK-NEXT:    add r0, r0, r1, lsr #24
-; CHECK-NEXT:    pop {r4, lr}
+; CHECK-NEXT:    .save {r11, lr}
+; CHECK-NEXT:    push {r11, lr}
+; CHECK-NEXT:    bl __popcountdi2
+; CHECK-NEXT:    asr r1, r0, #31
+; CHECK-NEXT:    pop {r11, lr}
 ; CHECK-NEXT:    mov pc, lr
-; CHECK-NEXT:    .p2align 2
-; CHECK-NEXT:  @ %bb.1:
-; CHECK-NEXT:  .LCPI23_0:
-; CHECK-NEXT:    .long 1431655765 @ 0x55555555
-; CHECK-NEXT:  .LCPI23_1:
-; CHECK-NEXT:    .long 252645135 @ 0xf0f0f0f
-; CHECK-NEXT:  .LCPI23_2:
-; CHECK-NEXT:    .long 16843009 @ 0x1010101
-; CHECK-NEXT:  .LCPI23_3:
-; CHECK-NEXT:    .long 858993459 @ 0x33333333
   %count = tail call i64 @llvm.ctpop.i64(i64 %x)
-  %conv = trunc i64 %count to i32
-  ret i32 %conv
+  ret i64 %count
 }
 
 define i32 @ctpop_eq_one(i64 %x) nounwind readnone {