Use the "isCompare" machine instruction attribute instead of calling the

[oota-llvm.git] / lib / Target / ARM / ARMISelDAGToDAG.cpp

diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 9128889b14bcbb1158bd71feea6739dc425df34e..f4012c71d7e9392f9f3988fcdff15dd95cbebf15 100644 (file)
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -36,6 +36,11 @@
 
 using namespace llvm;
 
+static cl::opt<bool>
+DisableShifterOp("disable-shifter-op", cl::Hidden,
+  cl::desc("Disable isel of shifter-op"),
+  cl::init(false));
+
 //===--------------------------------------------------------------------===//
 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
 /// instructions for SelectionDAG operations.
@@ -143,6 +148,11 @@ private:
                           unsigned *DOpcodes, unsigned *QOpcodes0,
                           unsigned *QOpcodes1);
 
+  /// SelectVTBL - Select NEON VTBL and VTBX intrinsics.  NumVecs should be 2,
+  /// 3 or 4.  These are custom-selected so that a REG_SEQUENCE can be
+  /// generated to force the table registers to be consecutive.
+  SDNode *SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs, unsigned Opc);
+
   /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
   SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
 
@@ -215,6 +225,9 @@ bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
                                               SDValue &BaseReg,
                                               SDValue &ShReg,
                                               SDValue &Opc) {
+  if (DisableShifterOp)
+    return false;
+
   ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
 
   // Don't match base register only case. That is matched to a separate
@@ -533,10 +546,9 @@ bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
                                             SDValue &Base, SDValue &Offset){
   // FIXME dl should come from the parent load or store, not the address
-  DebugLoc dl = Op->getDebugLoc();
   if (N.getOpcode() != ISD::ADD) {
     ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
-    if (!NC || NC->getZExtValue() != 0)
+    if (!NC || !NC->isNullValue())
       return false;
 
     Base = Offset = N;
@@ -662,6 +674,9 @@ bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
 bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
                                                 SDValue &BaseReg,
                                                 SDValue &Opc) {
+  if (DisableShifterOp)
+    return false;
+
   ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
 
   // Don't match base register only case. That is matched to a separate
@@ -1198,7 +1213,7 @@ SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
 SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
                                    unsigned *DOpcodes, unsigned *QOpcodes0,
                                    unsigned *QOpcodes1) {
-  assert(NumVecs >=1 && NumVecs <= 4 && "VST NumVecs out-of-range");
+  assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range");
   DebugLoc dl = N->getDebugLoc();
 
   SDValue MemAddr, Align;
@@ -1382,13 +1397,11 @@ SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
 
   // Quad registers are handled by load/store of subregs. Find the subreg info.
   unsigned NumElts = 0;
-  int SubregIdx = 0;
   bool Even = false;
   EVT RegVT = VT;
   if (!is64BitVector) {
     RegVT = GetNEONSubregVT(VT);
     NumElts = RegVT.getVectorNumElements();
-    SubregIdx = (Lane < NumElts) ? ARM::dsub_0 : ARM::dsub_1;
     Even = Lane < NumElts;
   }
 
@@ -1527,6 +1540,46 @@ SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
   return NULL;
 }
 
+SDNode *ARMDAGToDAGISel::SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs,
+                                    unsigned Opc) {
+  assert(NumVecs >= 2 && NumVecs <= 4 && "VTBL NumVecs out-of-range");
+  DebugLoc dl = N->getDebugLoc();
+  EVT VT = N->getValueType(0);
+  unsigned FirstTblReg = IsExt ? 2 : 1;
+
+  // Form a REG_SEQUENCE to force register allocation.
+  SDValue RegSeq;
+  SDValue V0 = N->getOperand(FirstTblReg + 0);
+  SDValue V1 = N->getOperand(FirstTblReg + 1);
+  if (NumVecs == 2)
+    RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0);
+  else {
+    SDValue V2 = N->getOperand(FirstTblReg + 2);
+    // If it's a vtbl3, form a quad D-register and leave the last part as 
+    // an undef.
+    SDValue V3 = (NumVecs == 3)
+      ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
+      : N->getOperand(FirstTblReg + 3);
+    RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
+  }
+
+  // Now extract the D registers back out.
+  SmallVector<SDValue, 6> Ops;
+  if (IsExt)
+    Ops.push_back(N->getOperand(1));
+  Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT, RegSeq));
+  Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT, RegSeq));
+  if (NumVecs > 2)
+    Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT, RegSeq));
+  if (NumVecs > 3)
+    Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT, RegSeq));
+
+  Ops.push_back(N->getOperand(FirstTblReg + NumVecs));
+  Ops.push_back(getAL(CurDAG)); // predicate
+  Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // predicate register
+  return CurDAG->getMachineNode(Opc, dl, VT, Ops.data(), Ops.size());
+}
+
 SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
                                                      bool isSigned) {
   if (!Subtarget->hasV6T2Ops())
@@ -2284,6 +2337,29 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
     break;
   }
 
+  case ISD::INTRINSIC_WO_CHAIN: {
+    unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
+    switch (IntNo) {
+    default:
+      break;
+
+    case Intrinsic::arm_neon_vtbl2:
+      return SelectVTBL(N, false, 2, ARM::VTBL2);
+    case Intrinsic::arm_neon_vtbl3:
+      return SelectVTBL(N, false, 3, ARM::VTBL3);
+    case Intrinsic::arm_neon_vtbl4:
+      return SelectVTBL(N, false, 4, ARM::VTBL4);
+
+    case Intrinsic::arm_neon_vtbx2:
+      return SelectVTBL(N, true, 2, ARM::VTBX2);
+    case Intrinsic::arm_neon_vtbx3:
+      return SelectVTBL(N, true, 3, ARM::VTBX3);
+    case Intrinsic::arm_neon_vtbx4:
+      return SelectVTBL(N, true, 4, ARM::VTBX4);
+    }
+    break;
+  }
+
   case ISD::CONCAT_VECTORS:
     return SelectConcatVector(N);
   }