ARMISelDAGToDAG.cpp revision ffde080ae615906545eb33dab30e7bc47c2ac838 
1//===-- ARMISelDAGToDAG.cpp - A dag to dag inst selector for ARM ----------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines an instruction selector for the ARM target.
11//
12//===----------------------------------------------------------------------===//
13
14#define DEBUG_TYPE "arm-isel"
15#include "ARM.h"
16#include "ARMAddressingModes.h"
17#include "ARMTargetMachine.h"
18#include "llvm/CallingConv.h"
19#include "llvm/Constants.h"
20#include "llvm/DerivedTypes.h"
21#include "llvm/Function.h"
22#include "llvm/Intrinsics.h"
23#include "llvm/LLVMContext.h"
24#include "llvm/CodeGen/MachineFrameInfo.h"
25#include "llvm/CodeGen/MachineFunction.h"
26#include "llvm/CodeGen/MachineInstrBuilder.h"
27#include "llvm/CodeGen/SelectionDAG.h"
28#include "llvm/CodeGen/SelectionDAGISel.h"
29#include "llvm/Target/TargetLowering.h"
30#include "llvm/Target/TargetOptions.h"
31#include "llvm/Support/CommandLine.h"
32#include "llvm/Support/Compiler.h"
33#include "llvm/Support/Debug.h"
34#include "llvm/Support/ErrorHandling.h"
35#include "llvm/Support/raw_ostream.h"
36
37using namespace llvm;
38
39static cl::opt<bool>
40DisableShifterOp("disable-shifter-op", cl::Hidden,
41  cl::desc("Disable isel of shifter-op"),
42  cl::init(false));
43
44//===--------------------------------------------------------------------===//
45/// ARMDAGToDAGISel - ARM specific code to select ARM machine
46/// instructions for SelectionDAG operations.
47///
48namespace {
49class ARMDAGToDAGISel : public SelectionDAGISel {
50  ARMBaseTargetMachine &TM;
51
52  /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
53  /// make the right decision when generating code for different targets.
54  const ARMSubtarget *Subtarget;
55
56public:
57  explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
58                           CodeGenOpt::Level OptLevel)
59    : SelectionDAGISel(tm, OptLevel), TM(tm),
60    Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
61  }
62
63  virtual const char *getPassName() const {
64    return "ARM Instruction Selection";
65  }
66
67  /// getI32Imm - Return a target constant of type i32 with the specified
68  /// value.
69  inline SDValue getI32Imm(unsigned Imm) {
70    return CurDAG->getTargetConstant(Imm, MVT::i32);
71  }
72
73  SDNode *Select(SDNode *N);
74
75  bool SelectShifterOperandReg(SDNode *Op, SDValue N, SDValue &A,
76                               SDValue &B, SDValue &C);
77  bool SelectAddrMode2(SDNode *Op, SDValue N, SDValue &Base,
78                       SDValue &Offset, SDValue &Opc);
79  bool SelectAddrMode2Offset(SDNode *Op, SDValue N,
80                             SDValue &Offset, SDValue &Opc);
81  bool SelectAddrMode3(SDNode *Op, SDValue N, SDValue &Base,
82                       SDValue &Offset, SDValue &Opc);
83  bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
84                             SDValue &Offset, SDValue &Opc);
85  bool SelectAddrMode4(SDNode *Op, SDValue N, SDValue &Addr,
86                       SDValue &Mode);
87  bool SelectAddrMode5(SDNode *Op, SDValue N, SDValue &Base,
88                       SDValue &Offset);
89  bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Align);
90
91  bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset,
92                        SDValue &Label);
93
94  bool SelectThumbAddrModeRR(SDNode *Op, SDValue N, SDValue &Base,
95                             SDValue &Offset);
96  bool SelectThumbAddrModeRI5(SDNode *Op, SDValue N, unsigned Scale,
97                              SDValue &Base, SDValue &OffImm,
98                              SDValue &Offset);
99  bool SelectThumbAddrModeS1(SDNode *Op, SDValue N, SDValue &Base,
100                             SDValue &OffImm, SDValue &Offset);
101  bool SelectThumbAddrModeS2(SDNode *Op, SDValue N, SDValue &Base,
102                             SDValue &OffImm, SDValue &Offset);
103  bool SelectThumbAddrModeS4(SDNode *Op, SDValue N, SDValue &Base,
104                             SDValue &OffImm, SDValue &Offset);
105  bool SelectThumbAddrModeSP(SDNode *Op, SDValue N, SDValue &Base,
106                             SDValue &OffImm);
107
108  bool SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
109                                 SDValue &BaseReg, SDValue &Opc);
110  bool SelectT2AddrModeImm12(SDNode *Op, SDValue N, SDValue &Base,
111                             SDValue &OffImm);
112  bool SelectT2AddrModeImm8(SDNode *Op, SDValue N, SDValue &Base,
113                            SDValue &OffImm);
114  bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
115                                 SDValue &OffImm);
116  bool SelectT2AddrModeImm8s4(SDNode *Op, SDValue N, SDValue &Base,
117                              SDValue &OffImm);
118  bool SelectT2AddrModeSoReg(SDNode *Op, SDValue N, SDValue &Base,
119                             SDValue &OffReg, SDValue &ShImm);
120
121  inline bool Pred_so_imm(SDNode *inN) const {
122    ConstantSDNode *N = cast<ConstantSDNode>(inN);
123    return ARM_AM::getSOImmVal(N->getZExtValue()) != -1;
124  }
125
126  inline bool Pred_t2_so_imm(SDNode *inN) const {
127    ConstantSDNode *N = cast<ConstantSDNode>(inN);
128    return ARM_AM::getT2SOImmVal(N->getZExtValue()) != -1;
129  }
130
131  // Include the pieces autogenerated from the target description.
132#include "ARMGenDAGISel.inc"
133
134private:
135  /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
136  /// ARM.
137  SDNode *SelectARMIndexedLoad(SDNode *N);
138  SDNode *SelectT2IndexedLoad(SDNode *N);
139
140  /// SelectVLD - Select NEON load intrinsics.  NumVecs should be
141  /// 1, 2, 3 or 4.  The opcode arrays specify the instructions used for
142  /// loads of D registers and even subregs and odd subregs of Q registers.
143  /// For NumVecs <= 2, QOpcodes1 is not used.
144  SDNode *SelectVLD(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
145                    unsigned *QOpcodes0, unsigned *QOpcodes1);
146
147  /// SelectVST - Select NEON store intrinsics.  NumVecs should
148  /// be 1, 2, 3 or 4.  The opcode arrays specify the instructions used for
149  /// stores of D registers and even subregs and odd subregs of Q registers.
150  /// For NumVecs <= 2, QOpcodes1 is not used.
151  SDNode *SelectVST(SDNode *N, unsigned NumVecs, unsigned *DOpcodes,
152                    unsigned *QOpcodes0, unsigned *QOpcodes1);
153
154  /// SelectVLDSTLane - Select NEON load/store lane intrinsics.  NumVecs should
155  /// be 2, 3 or 4.  The opcode arrays specify the instructions used for
156  /// load/store of D registers and even subregs and odd subregs of Q registers.
157  SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, unsigned NumVecs,
158                          unsigned *DOpcodes, unsigned *QOpcodes0,
159                          unsigned *QOpcodes1);
160
161  /// SelectVTBL - Select NEON VTBL and VTBX intrinsics.  NumVecs should be 2,
162  /// 3 or 4.  These are custom-selected so that a REG_SEQUENCE can be
163  /// generated to force the table registers to be consecutive.
164  SDNode *SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs, unsigned Opc);
165
166  /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
167  SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
168
169  /// SelectCMOVOp - Select CMOV instructions for ARM.
170  SDNode *SelectCMOVOp(SDNode *N);
171  SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
172                              ARMCC::CondCodes CCVal, SDValue CCR,
173                              SDValue InFlag);
174  SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
175                               ARMCC::CondCodes CCVal, SDValue CCR,
176                               SDValue InFlag);
177  SDNode *SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
178                              ARMCC::CondCodes CCVal, SDValue CCR,
179                              SDValue InFlag);
180  SDNode *SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
181                               ARMCC::CondCodes CCVal, SDValue CCR,
182                               SDValue InFlag);
183
184  SDNode *SelectConcatVector(SDNode *N);
185
186  /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
187  /// inline asm expressions.
188  virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
189                                            char ConstraintCode,
190                                            std::vector<SDValue> &OutOps);
191
192  // Form pairs of consecutive S, D, or Q registers.
193  SDNode *PairSRegs(EVT VT, SDValue V0, SDValue V1);
194  SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
195  SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
196
197  // Form sequences of 4 consecutive S, D, or Q registers.
198  SDNode *QuadSRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
199  SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
200  SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
201
202  // Form sequences of 8 consecutive D registers.
203  SDNode *OctoDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3,
204                    SDValue V4, SDValue V5, SDValue V6, SDValue V7);
205};
206}
207
208/// isInt32Immediate - This method tests to see if the node is a 32-bit constant
209/// operand. If so Imm will receive the 32-bit value.
210static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
211  if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
212    Imm = cast<ConstantSDNode>(N)->getZExtValue();
213    return true;
214  }
215  return false;
216}
217
218// isInt32Immediate - This method tests to see if a constant operand.
219// If so Imm will receive the 32 bit value.
220static bool isInt32Immediate(SDValue N, unsigned &Imm) {
221  return isInt32Immediate(N.getNode(), Imm);
222}
223
224// isOpcWithIntImmediate - This method tests to see if the node is a specific
225// opcode and that it has a immediate integer right operand.
226// If so Imm will receive the 32 bit value.
227static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
228  return N->getOpcode() == Opc &&
229         isInt32Immediate(N->getOperand(1).getNode(), Imm);
230}
231
232
233bool ARMDAGToDAGISel::SelectShifterOperandReg(SDNode *Op,
234                                              SDValue N,
235                                              SDValue &BaseReg,
236                                              SDValue &ShReg,
237                                              SDValue &Opc) {
238  if (DisableShifterOp)
239    return false;
240
241  ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
242
243  // Don't match base register only case. That is matched to a separate
244  // lower complexity pattern with explicit register operand.
245  if (ShOpcVal == ARM_AM::no_shift) return false;
246
247  BaseReg = N.getOperand(0);
248  unsigned ShImmVal = 0;
249  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
250    ShReg = CurDAG->getRegister(0, MVT::i32);
251    ShImmVal = RHS->getZExtValue() & 31;
252  } else {
253    ShReg = N.getOperand(1);
254  }
255  Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
256                                  MVT::i32);
257  return true;
258}
259
260bool ARMDAGToDAGISel::SelectAddrMode2(SDNode *Op, SDValue N,
261                                      SDValue &Base, SDValue &Offset,
262                                      SDValue &Opc) {
263  if (N.getOpcode() == ISD::MUL) {
264    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
265      // X * [3,5,9] -> X + X * [2,4,8] etc.
266      int RHSC = (int)RHS->getZExtValue();
267      if (RHSC & 1) {
268        RHSC = RHSC & ~1;
269        ARM_AM::AddrOpc AddSub = ARM_AM::add;
270        if (RHSC < 0) {
271          AddSub = ARM_AM::sub;
272          RHSC = - RHSC;
273        }
274        if (isPowerOf2_32(RHSC)) {
275          unsigned ShAmt = Log2_32(RHSC);
276          Base = Offset = N.getOperand(0);
277          Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
278                                                            ARM_AM::lsl),
279                                          MVT::i32);
280          return true;
281        }
282      }
283    }
284  }
285
286  if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
287    Base = N;
288    if (N.getOpcode() == ISD::FrameIndex) {
289      int FI = cast<FrameIndexSDNode>(N)->getIndex();
290      Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
291    } else if (N.getOpcode() == ARMISD::Wrapper &&
292               !(Subtarget->useMovt() &&
293                 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
294      Base = N.getOperand(0);
295    }
296    Offset = CurDAG->getRegister(0, MVT::i32);
297    Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
298                                                      ARM_AM::no_shift),
299                                    MVT::i32);
300    return true;
301  }
302
303  // Match simple R +/- imm12 operands.
304  if (N.getOpcode() == ISD::ADD)
305    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
306      int RHSC = (int)RHS->getZExtValue();
307      if ((RHSC >= 0 && RHSC < 0x1000) ||
308          (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
309        Base = N.getOperand(0);
310        if (Base.getOpcode() == ISD::FrameIndex) {
311          int FI = cast<FrameIndexSDNode>(Base)->getIndex();
312          Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
313        }
314        Offset = CurDAG->getRegister(0, MVT::i32);
315
316        ARM_AM::AddrOpc AddSub = ARM_AM::add;
317        if (RHSC < 0) {
318          AddSub = ARM_AM::sub;
319          RHSC = - RHSC;
320        }
321        Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
322                                                          ARM_AM::no_shift),
323                                        MVT::i32);
324        return true;
325      }
326    }
327
328  // Otherwise this is R +/- [possibly shifted] R.
329  ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
330  ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
331  unsigned ShAmt = 0;
332
333  Base   = N.getOperand(0);
334  Offset = N.getOperand(1);
335
336  if (ShOpcVal != ARM_AM::no_shift) {
337    // Check to see if the RHS of the shift is a constant, if not, we can't fold
338    // it.
339    if (ConstantSDNode *Sh =
340           dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
341      ShAmt = Sh->getZExtValue();
342      Offset = N.getOperand(1).getOperand(0);
343    } else {
344      ShOpcVal = ARM_AM::no_shift;
345    }
346  }
347
348  // Try matching (R shl C) + (R).
349  if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
350    ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
351    if (ShOpcVal != ARM_AM::no_shift) {
352      // Check to see if the RHS of the shift is a constant, if not, we can't
353      // fold it.
354      if (ConstantSDNode *Sh =
355          dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
356        ShAmt = Sh->getZExtValue();
357        Offset = N.getOperand(0).getOperand(0);
358        Base = N.getOperand(1);
359      } else {
360        ShOpcVal = ARM_AM::no_shift;
361      }
362    }
363  }
364
365  Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
366                                  MVT::i32);
367  return true;
368}
369
370bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
371                                            SDValue &Offset, SDValue &Opc) {
372  unsigned Opcode = Op->getOpcode();
373  ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
374    ? cast<LoadSDNode>(Op)->getAddressingMode()
375    : cast<StoreSDNode>(Op)->getAddressingMode();
376  ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
377    ? ARM_AM::add : ARM_AM::sub;
378  if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
379    int Val = (int)C->getZExtValue();
380    if (Val >= 0 && Val < 0x1000) { // 12 bits.
381      Offset = CurDAG->getRegister(0, MVT::i32);
382      Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
383                                                        ARM_AM::no_shift),
384                                      MVT::i32);
385      return true;
386    }
387  }
388
389  Offset = N;
390  ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
391  unsigned ShAmt = 0;
392  if (ShOpcVal != ARM_AM::no_shift) {
393    // Check to see if the RHS of the shift is a constant, if not, we can't fold
394    // it.
395    if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
396      ShAmt = Sh->getZExtValue();
397      Offset = N.getOperand(0);
398    } else {
399      ShOpcVal = ARM_AM::no_shift;
400    }
401  }
402
403  Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
404                                  MVT::i32);
405  return true;
406}
407
408
409bool ARMDAGToDAGISel::SelectAddrMode3(SDNode *Op, SDValue N,
410                                      SDValue &Base, SDValue &Offset,
411                                      SDValue &Opc) {
412  if (N.getOpcode() == ISD::SUB) {
413    // X - C  is canonicalize to X + -C, no need to handle it here.
414    Base = N.getOperand(0);
415    Offset = N.getOperand(1);
416    Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
417    return true;
418  }
419
420  if (N.getOpcode() != ISD::ADD) {
421    Base = N;
422    if (N.getOpcode() == ISD::FrameIndex) {
423      int FI = cast<FrameIndexSDNode>(N)->getIndex();
424      Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
425    }
426    Offset = CurDAG->getRegister(0, MVT::i32);
427    Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
428    return true;
429  }
430
431  // If the RHS is +/- imm8, fold into addr mode.
432  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
433    int RHSC = (int)RHS->getZExtValue();
434    if ((RHSC >= 0 && RHSC < 256) ||
435        (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
436      Base = N.getOperand(0);
437      if (Base.getOpcode() == ISD::FrameIndex) {
438        int FI = cast<FrameIndexSDNode>(Base)->getIndex();
439        Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
440      }
441      Offset = CurDAG->getRegister(0, MVT::i32);
442
443      ARM_AM::AddrOpc AddSub = ARM_AM::add;
444      if (RHSC < 0) {
445        AddSub = ARM_AM::sub;
446        RHSC = - RHSC;
447      }
448      Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
449      return true;
450    }
451  }
452
453  Base = N.getOperand(0);
454  Offset = N.getOperand(1);
455  Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
456  return true;
457}
458
459bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
460                                            SDValue &Offset, SDValue &Opc) {
461  unsigned Opcode = Op->getOpcode();
462  ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
463    ? cast<LoadSDNode>(Op)->getAddressingMode()
464    : cast<StoreSDNode>(Op)->getAddressingMode();
465  ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
466    ? ARM_AM::add : ARM_AM::sub;
467  if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
468    int Val = (int)C->getZExtValue();
469    if (Val >= 0 && Val < 256) {
470      Offset = CurDAG->getRegister(0, MVT::i32);
471      Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
472      return true;
473    }
474  }
475
476  Offset = N;
477  Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
478  return true;
479}
480
481bool ARMDAGToDAGISel::SelectAddrMode4(SDNode *Op, SDValue N,
482                                      SDValue &Addr, SDValue &Mode) {
483  Addr = N;
484  Mode = CurDAG->getTargetConstant(ARM_AM::getAM4ModeImm(ARM_AM::ia), MVT::i32);
485  return true;
486}
487
488bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N,
489                                      SDValue &Base, SDValue &Offset) {
490  if (N.getOpcode() != ISD::ADD) {
491    Base = N;
492    if (N.getOpcode() == ISD::FrameIndex) {
493      int FI = cast<FrameIndexSDNode>(N)->getIndex();
494      Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
495    } else if (N.getOpcode() == ARMISD::Wrapper &&
496               !(Subtarget->useMovt() &&
497                 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
498      Base = N.getOperand(0);
499    }
500    Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
501                                       MVT::i32);
502    return true;
503  }
504
505  // If the RHS is +/- imm8, fold into addr mode.
506  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
507    int RHSC = (int)RHS->getZExtValue();
508    if ((RHSC & 3) == 0) {  // The constant is implicitly multiplied by 4.
509      RHSC >>= 2;
510      if ((RHSC >= 0 && RHSC < 256) ||
511          (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
512        Base = N.getOperand(0);
513        if (Base.getOpcode() == ISD::FrameIndex) {
514          int FI = cast<FrameIndexSDNode>(Base)->getIndex();
515          Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
516        }
517
518        ARM_AM::AddrOpc AddSub = ARM_AM::add;
519        if (RHSC < 0) {
520          AddSub = ARM_AM::sub;
521          RHSC = - RHSC;
522        }
523        Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
524                                           MVT::i32);
525        return true;
526      }
527    }
528  }
529
530  Base = N;
531  Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
532                                     MVT::i32);
533  return true;
534}
535
536bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N,
537                                      SDValue &Addr, SDValue &Align) {
538  Addr = N;
539  // Default to no alignment.
540  Align = CurDAG->getTargetConstant(0, MVT::i32);
541  return true;
542}
543
544bool ARMDAGToDAGISel::SelectAddrModePC(SDNode *Op, SDValue N,
545                                       SDValue &Offset, SDValue &Label) {
546  if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
547    Offset = N.getOperand(0);
548    SDValue N1 = N.getOperand(1);
549    Label  = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
550                                       MVT::i32);
551    return true;
552  }
553  return false;
554}
555
556bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDNode *Op, SDValue N,
557                                            SDValue &Base, SDValue &Offset){
558  // FIXME dl should come from the parent load or store, not the address
559  if (N.getOpcode() != ISD::ADD) {
560    ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
561    if (!NC || !NC->isNullValue())
562      return false;
563
564    Base = Offset = N;
565    return true;
566  }
567
568  Base = N.getOperand(0);
569  Offset = N.getOperand(1);
570  return true;
571}
572
573bool
574ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDNode *Op, SDValue N,
575                                        unsigned Scale, SDValue &Base,
576                                        SDValue &OffImm, SDValue &Offset) {
577  if (Scale == 4) {
578    SDValue TmpBase, TmpOffImm;
579    if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
580      return false;  // We want to select tLDRspi / tSTRspi instead.
581    if (N.getOpcode() == ARMISD::Wrapper &&
582        N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
583      return false;  // We want to select tLDRpci instead.
584  }
585
586  if (N.getOpcode() != ISD::ADD) {
587    if (N.getOpcode() == ARMISD::Wrapper &&
588        !(Subtarget->useMovt() &&
589          N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
590      Base = N.getOperand(0);
591    } else
592      Base = N;
593
594    Offset = CurDAG->getRegister(0, MVT::i32);
595    OffImm = CurDAG->getTargetConstant(0, MVT::i32);
596    return true;
597  }
598
599  // Thumb does not have [sp, r] address mode.
600  RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
601  RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
602  if ((LHSR && LHSR->getReg() == ARM::SP) ||
603      (RHSR && RHSR->getReg() == ARM::SP)) {
604    Base = N;
605    Offset = CurDAG->getRegister(0, MVT::i32);
606    OffImm = CurDAG->getTargetConstant(0, MVT::i32);
607    return true;
608  }
609
610  // If the RHS is + imm5 * scale, fold into addr mode.
611  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
612    int RHSC = (int)RHS->getZExtValue();
613    if ((RHSC & (Scale-1)) == 0) {  // The constant is implicitly multiplied.
614      RHSC /= Scale;
615      if (RHSC >= 0 && RHSC < 32) {
616        Base = N.getOperand(0);
617        Offset = CurDAG->getRegister(0, MVT::i32);
618        OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
619        return true;
620      }
621    }
622  }
623
624  Base = N.getOperand(0);
625  Offset = N.getOperand(1);
626  OffImm = CurDAG->getTargetConstant(0, MVT::i32);
627  return true;
628}
629
630bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDNode *Op, SDValue N,
631                                            SDValue &Base, SDValue &OffImm,
632                                            SDValue &Offset) {
633  return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
634}
635
636bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDNode *Op, SDValue N,
637                                            SDValue &Base, SDValue &OffImm,
638                                            SDValue &Offset) {
639  return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
640}
641
642bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDNode *Op, SDValue N,
643                                            SDValue &Base, SDValue &OffImm,
644                                            SDValue &Offset) {
645  return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
646}
647
648bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDNode *Op, SDValue N,
649                                           SDValue &Base, SDValue &OffImm) {
650  if (N.getOpcode() == ISD::FrameIndex) {
651    int FI = cast<FrameIndexSDNode>(N)->getIndex();
652    Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
653    OffImm = CurDAG->getTargetConstant(0, MVT::i32);
654    return true;
655  }
656
657  if (N.getOpcode() != ISD::ADD)
658    return false;
659
660  RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
661  if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
662      (LHSR && LHSR->getReg() == ARM::SP)) {
663    // If the RHS is + imm8 * scale, fold into addr mode.
664    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
665      int RHSC = (int)RHS->getZExtValue();
666      if ((RHSC & 3) == 0) {  // The constant is implicitly multiplied.
667        RHSC >>= 2;
668        if (RHSC >= 0 && RHSC < 256) {
669          Base = N.getOperand(0);
670          if (Base.getOpcode() == ISD::FrameIndex) {
671            int FI = cast<FrameIndexSDNode>(Base)->getIndex();
672            Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
673          }
674          OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
675          return true;
676        }
677      }
678    }
679  }
680
681  return false;
682}
683
684bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDNode *Op, SDValue N,
685                                                SDValue &BaseReg,
686                                                SDValue &Opc) {
687  if (DisableShifterOp)
688    return false;
689
690  ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
691
692  // Don't match base register only case. That is matched to a separate
693  // lower complexity pattern with explicit register operand.
694  if (ShOpcVal == ARM_AM::no_shift) return false;
695
696  BaseReg = N.getOperand(0);
697  unsigned ShImmVal = 0;
698  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
699    ShImmVal = RHS->getZExtValue() & 31;
700    Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
701    return true;
702  }
703
704  return false;
705}
706
707bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDNode *Op, SDValue N,
708                                            SDValue &Base, SDValue &OffImm) {
709  // Match simple R + imm12 operands.
710
711  // Base only.
712  if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
713    if (N.getOpcode() == ISD::FrameIndex) {
714      // Match frame index...
715      int FI = cast<FrameIndexSDNode>(N)->getIndex();
716      Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
717      OffImm  = CurDAG->getTargetConstant(0, MVT::i32);
718      return true;
719    } else if (N.getOpcode() == ARMISD::Wrapper &&
720               !(Subtarget->useMovt() &&
721                 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
722      Base = N.getOperand(0);
723      if (Base.getOpcode() == ISD::TargetConstantPool)
724        return false;  // We want to select t2LDRpci instead.
725    } else
726      Base = N;
727    OffImm  = CurDAG->getTargetConstant(0, MVT::i32);
728    return true;
729  }
730
731  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
732    if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
733      // Let t2LDRi8 handle (R - imm8).
734      return false;
735
736    int RHSC = (int)RHS->getZExtValue();
737    if (N.getOpcode() == ISD::SUB)
738      RHSC = -RHSC;
739
740    if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
741      Base   = N.getOperand(0);
742      if (Base.getOpcode() == ISD::FrameIndex) {
743        int FI = cast<FrameIndexSDNode>(Base)->getIndex();
744        Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
745      }
746      OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
747      return true;
748    }
749  }
750
751  // Base only.
752  Base = N;
753  OffImm  = CurDAG->getTargetConstant(0, MVT::i32);
754  return true;
755}
756
757bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDNode *Op, SDValue N,
758                                           SDValue &Base, SDValue &OffImm) {
759  // Match simple R - imm8 operands.
760  if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
761    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
762      int RHSC = (int)RHS->getSExtValue();
763      if (N.getOpcode() == ISD::SUB)
764        RHSC = -RHSC;
765
766      if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
767        Base = N.getOperand(0);
768        if (Base.getOpcode() == ISD::FrameIndex) {
769          int FI = cast<FrameIndexSDNode>(Base)->getIndex();
770          Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
771        }
772        OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
773        return true;
774      }
775    }
776  }
777
778  return false;
779}
780
781bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
782                                                 SDValue &OffImm){
783  unsigned Opcode = Op->getOpcode();
784  ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
785    ? cast<LoadSDNode>(Op)->getAddressingMode()
786    : cast<StoreSDNode>(Op)->getAddressingMode();
787  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
788    int RHSC = (int)RHS->getZExtValue();
789    if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
790      OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
791        ? CurDAG->getTargetConstant(RHSC, MVT::i32)
792        : CurDAG->getTargetConstant(-RHSC, MVT::i32);
793      return true;
794    }
795  }
796
797  return false;
798}
799
800bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDNode *Op, SDValue N,
801                                             SDValue &Base, SDValue &OffImm) {
802  if (N.getOpcode() == ISD::ADD) {
803    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
804      int RHSC = (int)RHS->getZExtValue();
805      // 8 bits.
806      if (((RHSC & 0x3) == 0) &&
807          ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) {
808        Base   = N.getOperand(0);
809        OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
810        return true;
811      }
812    }
813  } else if (N.getOpcode() == ISD::SUB) {
814    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
815      int RHSC = (int)RHS->getZExtValue();
816      // 8 bits.
817      if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) {
818        Base   = N.getOperand(0);
819        OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
820        return true;
821      }
822    }
823  }
824
825  return false;
826}
827
828bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDNode *Op, SDValue N,
829                                            SDValue &Base,
830                                            SDValue &OffReg, SDValue &ShImm) {
831  // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
832  if (N.getOpcode() != ISD::ADD)
833    return false;
834
835  // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
836  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
837    int RHSC = (int)RHS->getZExtValue();
838    if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
839      return false;
840    else if (RHSC < 0 && RHSC >= -255) // 8 bits
841      return false;
842  }
843
844  // Look for (R + R) or (R + (R << [1,2,3])).
845  unsigned ShAmt = 0;
846  Base   = N.getOperand(0);
847  OffReg = N.getOperand(1);
848
849  // Swap if it is ((R << c) + R).
850  ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
851  if (ShOpcVal != ARM_AM::lsl) {
852    ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
853    if (ShOpcVal == ARM_AM::lsl)
854      std::swap(Base, OffReg);
855  }
856
857  if (ShOpcVal == ARM_AM::lsl) {
858    // Check to see if the RHS of the shift is a constant, if not, we can't fold
859    // it.
860    if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
861      ShAmt = Sh->getZExtValue();
862      if (ShAmt >= 4) {
863        ShAmt = 0;
864        ShOpcVal = ARM_AM::no_shift;
865      } else
866        OffReg = OffReg.getOperand(0);
867    } else {
868      ShOpcVal = ARM_AM::no_shift;
869    }
870  }
871
872  ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);
873
874  return true;
875}
876
877//===--------------------------------------------------------------------===//
878
879/// getAL - Returns a ARMCC::AL immediate node.
880static inline SDValue getAL(SelectionDAG *CurDAG) {
881  return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
882}
883
884SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) {
885  LoadSDNode *LD = cast<LoadSDNode>(N);
886  ISD::MemIndexedMode AM = LD->getAddressingMode();
887  if (AM == ISD::UNINDEXED)
888    return NULL;
889
890  EVT LoadedVT = LD->getMemoryVT();
891  SDValue Offset, AMOpc;
892  bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
893  unsigned Opcode = 0;
894  bool Match = false;
895  if (LoadedVT == MVT::i32 &&
896      SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
897    Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
898    Match = true;
899  } else if (LoadedVT == MVT::i16 &&
900             SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
901    Match = true;
902    Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
903      ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
904      : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
905  } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
906    if (LD->getExtensionType() == ISD::SEXTLOAD) {
907      if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
908        Match = true;
909        Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
910      }
911    } else {
912      if (SelectAddrMode2Offset(N, LD->getOffset(), Offset, AMOpc)) {
913        Match = true;
914        Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
915      }
916    }
917  }
918
919  if (Match) {
920    SDValue Chain = LD->getChain();
921    SDValue Base = LD->getBasePtr();
922    SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
923                     CurDAG->getRegister(0, MVT::i32), Chain };
924    return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
925                                  MVT::Other, Ops, 6);
926  }
927
928  return NULL;
929}
930
931SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) {
932  LoadSDNode *LD = cast<LoadSDNode>(N);
933  ISD::MemIndexedMode AM = LD->getAddressingMode();
934  if (AM == ISD::UNINDEXED)
935    return NULL;
936
937  EVT LoadedVT = LD->getMemoryVT();
938  bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
939  SDValue Offset;
940  bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
941  unsigned Opcode = 0;
942  bool Match = false;
943  if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
944    switch (LoadedVT.getSimpleVT().SimpleTy) {
945    case MVT::i32:
946      Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
947      break;
948    case MVT::i16:
949      if (isSExtLd)
950        Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
951      else
952        Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
953      break;
954    case MVT::i8:
955    case MVT::i1:
956      if (isSExtLd)
957        Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
958      else
959        Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
960      break;
961    default:
962      return NULL;
963    }
964    Match = true;
965  }
966
967  if (Match) {
968    SDValue Chain = LD->getChain();
969    SDValue Base = LD->getBasePtr();
970    SDValue Ops[]= { Base, Offset, getAL(CurDAG),
971                     CurDAG->getRegister(0, MVT::i32), Chain };
972    return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32,
973                                  MVT::Other, Ops, 5);
974  }
975
976  return NULL;
977}
978
979/// PairSRegs - Form a D register from a pair of S registers.
980///
981SDNode *ARMDAGToDAGISel::PairSRegs(EVT VT, SDValue V0, SDValue V1) {
982  DebugLoc dl = V0.getNode()->getDebugLoc();
983  SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
984  SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
985  const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
986  return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
987}
988
989/// PairDRegs - Form a quad register from a pair of D registers.
990///
991SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
992  DebugLoc dl = V0.getNode()->getDebugLoc();
993  SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
994  SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
995  const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
996  return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
997}
998
999/// PairQRegs - Form 4 consecutive D registers from a pair of Q registers.
1000///
1001SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) {
1002  DebugLoc dl = V0.getNode()->getDebugLoc();
1003  SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1004  SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1005  const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
1006  return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
1007}
1008
1009/// QuadSRegs - Form 4 consecutive S registers.
1010///
1011SDNode *ARMDAGToDAGISel::QuadSRegs(EVT VT, SDValue V0, SDValue V1,
1012                                   SDValue V2, SDValue V3) {
1013  DebugLoc dl = V0.getNode()->getDebugLoc();
1014  SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
1015  SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
1016  SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32);
1017  SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32);
1018  const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1019  return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1020}
1021
1022/// QuadDRegs - Form 4 consecutive D registers.
1023///
1024SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
1025                                   SDValue V2, SDValue V3) {
1026  DebugLoc dl = V0.getNode()->getDebugLoc();
1027  SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1028  SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1029  SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1030  SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1031  const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1032  return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1033}
1034
1035/// QuadQRegs - Form 4 consecutive Q registers.
1036///
1037SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1,
1038                                   SDValue V2, SDValue V3) {
1039  DebugLoc dl = V0.getNode()->getDebugLoc();
1040  SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32);
1041  SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32);
1042  SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, MVT::i32);
1043  SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, MVT::i32);
1044  const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
1045  return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
1046}
1047
1048/// OctoDRegs - Form 8 consecutive D registers.
1049///
1050SDNode *ARMDAGToDAGISel::OctoDRegs(EVT VT, SDValue V0, SDValue V1,
1051                                   SDValue V2, SDValue V3,
1052                                   SDValue V4, SDValue V5,
1053                                   SDValue V6, SDValue V7) {
1054  DebugLoc dl = V0.getNode()->getDebugLoc();
1055  SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1056  SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1057  SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32);
1058  SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32);
1059  SDValue SubReg4 = CurDAG->getTargetConstant(ARM::dsub_4, MVT::i32);
1060  SDValue SubReg5 = CurDAG->getTargetConstant(ARM::dsub_5, MVT::i32);
1061  SDValue SubReg6 = CurDAG->getTargetConstant(ARM::dsub_6, MVT::i32);
1062  SDValue SubReg7 = CurDAG->getTargetConstant(ARM::dsub_7, MVT::i32);
1063  const SDValue Ops[] ={ V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3,
1064                         V4, SubReg4, V5, SubReg5, V6, SubReg6, V7, SubReg7 };
1065  return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 16);
1066}
1067
1068/// GetNEONSubregVT - Given a type for a 128-bit NEON vector, return the type
1069/// for a 64-bit subregister of the vector.
1070static EVT GetNEONSubregVT(EVT VT) {
1071  switch (VT.getSimpleVT().SimpleTy) {
1072  default: llvm_unreachable("unhandled NEON type");
1073  case MVT::v16i8: return MVT::v8i8;
1074  case MVT::v8i16: return MVT::v4i16;
1075  case MVT::v4f32: return MVT::v2f32;
1076  case MVT::v4i32: return MVT::v2i32;
1077  case MVT::v2i64: return MVT::v1i64;
1078  }
1079}
1080
1081SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs,
1082                                   unsigned *DOpcodes, unsigned *QOpcodes0,
1083                                   unsigned *QOpcodes1) {
1084  assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
1085  DebugLoc dl = N->getDebugLoc();
1086
1087  SDValue MemAddr, Align;
1088  if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1089    return NULL;
1090
1091  SDValue Chain = N->getOperand(0);
1092  EVT VT = N->getValueType(0);
1093  bool is64BitVector = VT.is64BitVector();
1094
1095  unsigned OpcodeIndex;
1096  switch (VT.getSimpleVT().SimpleTy) {
1097  default: llvm_unreachable("unhandled vld type");
1098    // Double-register operations:
1099  case MVT::v8i8:  OpcodeIndex = 0; break;
1100  case MVT::v4i16: OpcodeIndex = 1; break;
1101  case MVT::v2f32:
1102  case MVT::v2i32: OpcodeIndex = 2; break;
1103  case MVT::v1i64: OpcodeIndex = 3; break;
1104    // Quad-register operations:
1105  case MVT::v16i8: OpcodeIndex = 0; break;
1106  case MVT::v8i16: OpcodeIndex = 1; break;
1107  case MVT::v4f32:
1108  case MVT::v4i32: OpcodeIndex = 2; break;
1109  case MVT::v2i64: OpcodeIndex = 3;
1110    assert(NumVecs == 1 && "v2i64 type only supported for VLD1");
1111    break;
1112  }
1113
1114  SDValue Pred = getAL(CurDAG);
1115  SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1116  if (is64BitVector) {
1117    unsigned Opc = DOpcodes[OpcodeIndex];
1118    const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1119    SDNode *VLd;
1120    if (NumVecs <= 2) {
1121      EVT ResTy;
1122      if (NumVecs == 1)
1123        ResTy = VT;
1124      else
1125        ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, NumVecs);
1126      VLd = CurDAG->getMachineNode(Opc, dl, ResTy, MVT::Other, Ops, 5);
1127    } else {
1128      std::vector<EVT> ResTys(NumVecs, VT);
1129      ResTys.push_back(MVT::Other);
1130      VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5);
1131    }
1132    if (NumVecs == 1)
1133      return VLd;
1134
1135    SDValue SuperReg;
1136    if (NumVecs <= 2)
1137      SuperReg = SDValue(VLd, 0);
1138    else {
1139      SDValue V0 = SDValue(VLd, 0);
1140      SDValue V1 = SDValue(VLd, 1);
1141      // Form a REG_SEQUENCE to force register allocation.
1142      SDValue V2 = SDValue(VLd, 2);
1143      // If it's a vld3, form a quad D-register but discard the last part.
1144      SDValue V3 = (NumVecs == 3)
1145          ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1146          : SDValue(VLd, 3);
1147      SuperReg = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1148    }
1149
1150    assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1151    for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1152      SDValue D = CurDAG->getTargetExtractSubreg(ARM::dsub_0+Vec,
1153                                                 dl, VT, SuperReg);
1154      ReplaceUses(SDValue(N, Vec), D);
1155    }
1156    ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, NumVecs <= 2 ? 1 : NumVecs));
1157    return NULL;
1158  }
1159
1160  EVT RegVT = GetNEONSubregVT(VT);
1161  if (NumVecs <= 2) {
1162    // Quad registers are directly supported for VLD1 and VLD2,
1163    // loading pairs of D regs.
1164    unsigned Opc = QOpcodes0[OpcodeIndex];
1165    const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain };
1166
1167    EVT ResTy;
1168    if (NumVecs == 1)
1169      ResTy = VT;
1170    else
1171      ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, 2 * NumVecs);
1172    SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTy, MVT::Other, Ops, 5);
1173
1174    // Combine the even and odd subregs to produce the result.
1175    if (NumVecs == 1)
1176      return VLd;
1177
1178    SDValue QQ = SDValue(VLd, 0);
1179    Chain = SDValue(VLd, 1);
1180
1181    SDValue Q0 = CurDAG->getTargetExtractSubreg(ARM::qsub_0, dl, VT, QQ);
1182    SDValue Q1 = CurDAG->getTargetExtractSubreg(ARM::qsub_1, dl, VT, QQ);
1183    ReplaceUses(SDValue(N, 0), Q0);
1184    ReplaceUses(SDValue(N, 1), Q1);
1185  } else {
1186    // Otherwise, quad registers are loaded with two separate instructions,
1187    // where one loads the even registers and the other loads the odd registers.
1188
1189    std::vector<EVT> ResTys(NumVecs, RegVT);
1190    ResTys.push_back(MemAddr.getValueType());
1191    ResTys.push_back(MVT::Other);
1192
1193    // Load the even subregs.
1194    unsigned Opc = QOpcodes0[OpcodeIndex];
1195    const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain };
1196    SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6);
1197    Chain = SDValue(VLdA, NumVecs+1);
1198
1199    // Load the odd subregs.
1200    Opc = QOpcodes1[OpcodeIndex];
1201    const SDValue OpsB[] = { SDValue(VLdA, NumVecs),
1202                             Align, Reg0, Pred, Reg0, Chain };
1203    SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6);
1204    Chain = SDValue(VLdB, NumVecs+1);
1205
1206    SDValue V0 = SDValue(VLdA, 0);
1207    SDValue V1 = SDValue(VLdB, 0);
1208    SDValue V2 = SDValue(VLdA, 1);
1209    SDValue V3 = SDValue(VLdB, 1);
1210    SDValue V4 = SDValue(VLdA, 2);
1211    SDValue V5 = SDValue(VLdB, 2);
1212    SDValue V6 = (NumVecs == 3)
1213      ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
1214      : SDValue(VLdA, 3);
1215    SDValue V7 = (NumVecs == 3)
1216      ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,RegVT), 0)
1217      : SDValue(VLdB, 3);
1218    SDValue RegSeq = SDValue(OctoDRegs(MVT::v8i64, V0, V1, V2, V3,
1219                                       V4, V5, V6, V7), 0);
1220
1221    // Extract out the 3 / 4 Q registers.
1222    assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1223    for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
1224      SDValue Q = CurDAG->getTargetExtractSubreg(ARM::qsub_0+Vec,
1225                                                 dl, VT, RegSeq);
1226      ReplaceUses(SDValue(N, Vec), Q);
1227    }
1228  }
1229  ReplaceUses(SDValue(N, NumVecs), Chain);
1230  return NULL;
1231}
1232
1233SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs,
1234                                   unsigned *DOpcodes, unsigned *QOpcodes0,
1235                                   unsigned *QOpcodes1) {
1236  assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range");
1237  DebugLoc dl = N->getDebugLoc();
1238
1239  SDValue MemAddr, Align;
1240  if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1241    return NULL;
1242
1243  SDValue Chain = N->getOperand(0);
1244  EVT VT = N->getOperand(3).getValueType();
1245  bool is64BitVector = VT.is64BitVector();
1246
1247  unsigned OpcodeIndex;
1248  switch (VT.getSimpleVT().SimpleTy) {
1249  default: llvm_unreachable("unhandled vst type");
1250    // Double-register operations:
1251  case MVT::v8i8:  OpcodeIndex = 0; break;
1252  case MVT::v4i16: OpcodeIndex = 1; break;
1253  case MVT::v2f32:
1254  case MVT::v2i32: OpcodeIndex = 2; break;
1255  case MVT::v1i64: OpcodeIndex = 3; break;
1256    // Quad-register operations:
1257  case MVT::v16i8: OpcodeIndex = 0; break;
1258  case MVT::v8i16: OpcodeIndex = 1; break;
1259  case MVT::v4f32:
1260  case MVT::v4i32: OpcodeIndex = 2; break;
1261  case MVT::v2i64: OpcodeIndex = 3;
1262    assert(NumVecs == 1 && "v2i64 type only supported for VST1");
1263    break;
1264  }
1265
1266  SDValue Pred = getAL(CurDAG);
1267  SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1268
1269  SmallVector<SDValue, 7> Ops;
1270  Ops.push_back(MemAddr);
1271  Ops.push_back(Align);
1272
1273  if (is64BitVector) {
1274    if (NumVecs == 1) {
1275      Ops.push_back(N->getOperand(3));
1276    } else {
1277      SDValue RegSeq;
1278      SDValue V0 = N->getOperand(0+3);
1279      SDValue V1 = N->getOperand(1+3);
1280
1281      // Form a REG_SEQUENCE to force register allocation.
1282      if (NumVecs == 2)
1283        RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1284      else {
1285        SDValue V2 = N->getOperand(2+3);
1286        // If it's a vld3, form a quad D-register and leave the last part as
1287        // an undef.
1288        SDValue V3 = (NumVecs == 3)
1289          ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1290          : N->getOperand(3+3);
1291        RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1292      }
1293      Ops.push_back(RegSeq);
1294    }
1295    Ops.push_back(Pred);
1296    Ops.push_back(Reg0); // predicate register
1297    Ops.push_back(Chain);
1298    unsigned Opc = DOpcodes[OpcodeIndex];
1299    return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 6);
1300  }
1301
1302  if (NumVecs <= 2) {
1303    // Quad registers are directly supported for VST1 and VST2.
1304    unsigned Opc = QOpcodes0[OpcodeIndex];
1305    if (NumVecs == 1) {
1306      Ops.push_back(N->getOperand(3));
1307    } else {
1308      // Form a QQ register.
1309      SDValue Q0 = N->getOperand(3);
1310      SDValue Q1 = N->getOperand(4);
1311      Ops.push_back(SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0));
1312    }
1313    Ops.push_back(Pred);
1314    Ops.push_back(Reg0); // predicate register
1315    Ops.push_back(Chain);
1316    return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 6);
1317  }
1318
1319  // Otherwise, quad registers are stored with two separate instructions,
1320  // where one stores the even registers and the other stores the odd registers.
1321
1322  // Form the QQQQ REG_SEQUENCE.
1323  SDValue V0 = N->getOperand(0+3);
1324  SDValue V1 = N->getOperand(1+3);
1325  SDValue V2 = N->getOperand(2+3);
1326  SDValue V3 = (NumVecs == 3)
1327    ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
1328    : N->getOperand(3+3);
1329  SDValue RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1330
1331  // Store the even D registers.
1332  Ops.push_back(Reg0); // post-access address offset
1333  Ops.push_back(RegSeq);
1334  Ops.push_back(Pred);
1335  Ops.push_back(Reg0); // predicate register
1336  Ops.push_back(Chain);
1337  unsigned Opc = QOpcodes0[OpcodeIndex];
1338  SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1339                                        MVT::Other, Ops.data(), 7);
1340  Chain = SDValue(VStA, 1);
1341
1342  // Store the odd D registers.
1343  Ops[0] = SDValue(VStA, 0); // MemAddr
1344  Ops[6] = Chain;
1345  Opc = QOpcodes1[OpcodeIndex];
1346  SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
1347                                        MVT::Other, Ops.data(), 7);
1348  Chain = SDValue(VStB, 1);
1349  ReplaceUses(SDValue(N, 0), Chain);
1350  return NULL;
1351}
1352
1353SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
1354                                         unsigned NumVecs, unsigned *DOpcodes,
1355                                         unsigned *QOpcodes0,
1356                                         unsigned *QOpcodes1) {
1357  assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
1358  DebugLoc dl = N->getDebugLoc();
1359
1360  SDValue MemAddr, Align;
1361  if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align))
1362    return NULL;
1363
1364  SDValue Chain = N->getOperand(0);
1365  unsigned Lane =
1366    cast<ConstantSDNode>(N->getOperand(NumVecs+3))->getZExtValue();
1367  EVT VT = IsLoad ? N->getValueType(0) : N->getOperand(3).getValueType();
1368  bool is64BitVector = VT.is64BitVector();
1369
1370  // Quad registers are handled by load/store of subregs. Find the subreg info.
1371  unsigned NumElts = 0;
1372  bool Even = false;
1373  EVT RegVT = VT;
1374  if (!is64BitVector) {
1375    RegVT = GetNEONSubregVT(VT);
1376    NumElts = RegVT.getVectorNumElements();
1377    Even = Lane < NumElts;
1378  }
1379
1380  unsigned OpcodeIndex;
1381  switch (VT.getSimpleVT().SimpleTy) {
1382  default: llvm_unreachable("unhandled vld/vst lane type");
1383    // Double-register operations:
1384  case MVT::v8i8:  OpcodeIndex = 0; break;
1385  case MVT::v4i16: OpcodeIndex = 1; break;
1386  case MVT::v2f32:
1387  case MVT::v2i32: OpcodeIndex = 2; break;
1388    // Quad-register operations:
1389  case MVT::v8i16: OpcodeIndex = 0; break;
1390  case MVT::v4f32:
1391  case MVT::v4i32: OpcodeIndex = 1; break;
1392  }
1393
1394  SDValue Pred = getAL(CurDAG);
1395  SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1396
1397  SmallVector<SDValue, 10> Ops;
1398  Ops.push_back(MemAddr);
1399  Ops.push_back(Align);
1400
1401  unsigned Opc = 0;
1402  if (is64BitVector) {
1403    Opc = DOpcodes[OpcodeIndex];
1404    SDValue RegSeq;
1405    SDValue V0 = N->getOperand(0+3);
1406    SDValue V1 = N->getOperand(1+3);
1407    if (NumVecs == 2) {
1408      RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1409    } else {
1410      SDValue V2 = N->getOperand(2+3);
1411      SDValue V3 = (NumVecs == 3)
1412        ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1413        : N->getOperand(3+3);
1414      RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1415    }
1416
1417    // Now extract the D registers back out.
1418    Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT, RegSeq));
1419    Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT, RegSeq));
1420    if (NumVecs > 2)
1421      Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT,RegSeq));
1422    if (NumVecs > 3)
1423      Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT,RegSeq));
1424  } else {
1425    // Check if this is loading the even or odd subreg of a Q register.
1426    if (Lane < NumElts) {
1427      Opc = QOpcodes0[OpcodeIndex];
1428    } else {
1429      Lane -= NumElts;
1430      Opc = QOpcodes1[OpcodeIndex];
1431    }
1432
1433    SDValue RegSeq;
1434    SDValue V0 = N->getOperand(0+3);
1435    SDValue V1 = N->getOperand(1+3);
1436    if (NumVecs == 2) {
1437      RegSeq = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0);
1438    } else {
1439      SDValue V2 = N->getOperand(2+3);
1440      SDValue V3 = (NumVecs == 3)
1441        ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1442        : N->getOperand(3+3);
1443      RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0);
1444    }
1445
1446    // Extract the subregs of the input vector.
1447    unsigned SubIdx = Even ? ARM::dsub_0 : ARM::dsub_1;
1448    for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1449      Ops.push_back(CurDAG->getTargetExtractSubreg(SubIdx+Vec*2, dl, RegVT,
1450                                                   RegSeq));
1451  }
1452  Ops.push_back(getI32Imm(Lane));
1453  Ops.push_back(Pred);
1454  Ops.push_back(Reg0);
1455  Ops.push_back(Chain);
1456
1457  if (!IsLoad)
1458    return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6);
1459
1460  std::vector<EVT> ResTys(NumVecs, RegVT);
1461  ResTys.push_back(MVT::Other);
1462  SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(),NumVecs+6);
1463
1464  // Form a REG_SEQUENCE to force register allocation.
1465  SDValue RegSeq;
1466  if (is64BitVector) {
1467    SDValue V0 = SDValue(VLdLn, 0);
1468    SDValue V1 = SDValue(VLdLn, 1);
1469    if (NumVecs == 2) {
1470      RegSeq = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0);
1471    } else {
1472      SDValue V2 = SDValue(VLdLn, 2);
1473      // If it's a vld3, form a quad D-register but discard the last part.
1474      SDValue V3 = (NumVecs == 3)
1475        ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
1476        : SDValue(VLdLn, 3);
1477      RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1478    }
1479  } else {
1480    // For 128-bit vectors, take the 64-bit results of the load and insert
1481    // them as subregs into the result.
1482    SDValue V[8];
1483    for (unsigned Vec = 0, i = 0; Vec < NumVecs; ++Vec, i+=2) {
1484      if (Even) {
1485        V[i]   = SDValue(VLdLn, Vec);
1486        V[i+1] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1487                                                dl, RegVT), 0);
1488      } else {
1489        V[i]   = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1490                                                dl, RegVT), 0);
1491        V[i+1] = SDValue(VLdLn, Vec);
1492      }
1493    }
1494    if (NumVecs == 3)
1495      V[6] = V[7] = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
1496                                                   dl, RegVT), 0);
1497
1498    if (NumVecs == 2)
1499      RegSeq = SDValue(QuadDRegs(MVT::v4i64, V[0], V[1], V[2], V[3]), 0);
1500    else
1501      RegSeq = SDValue(OctoDRegs(MVT::v8i64, V[0], V[1], V[2], V[3],
1502                                 V[4], V[5], V[6], V[7]), 0);
1503  }
1504
1505  assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
1506  assert(ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering");
1507  unsigned SubIdx = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
1508  for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
1509    ReplaceUses(SDValue(N, Vec),
1510                CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, RegSeq));
1511  ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, NumVecs));
1512  return NULL;
1513}
1514
1515SDNode *ARMDAGToDAGISel::SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs,
1516                                    unsigned Opc) {
1517  assert(NumVecs >= 2 && NumVecs <= 4 && "VTBL NumVecs out-of-range");
1518  DebugLoc dl = N->getDebugLoc();
1519  EVT VT = N->getValueType(0);
1520  unsigned FirstTblReg = IsExt ? 2 : 1;
1521
1522  // Form a REG_SEQUENCE to force register allocation.
1523  SDValue RegSeq;
1524  SDValue V0 = N->getOperand(FirstTblReg + 0);
1525  SDValue V1 = N->getOperand(FirstTblReg + 1);
1526  if (NumVecs == 2)
1527    RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0);
1528  else {
1529    SDValue V2 = N->getOperand(FirstTblReg + 2);
1530    // If it's a vtbl3, form a quad D-register and leave the last part as
1531    // an undef.
1532    SDValue V3 = (NumVecs == 3)
1533      ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
1534      : N->getOperand(FirstTblReg + 3);
1535    RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0);
1536  }
1537
1538  // Now extract the D registers back out.
1539  SmallVector<SDValue, 6> Ops;
1540  if (IsExt)
1541    Ops.push_back(N->getOperand(1));
1542  Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_0, dl, VT, RegSeq));
1543  Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_1, dl, VT, RegSeq));
1544  if (NumVecs > 2)
1545    Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_2, dl, VT, RegSeq));
1546  if (NumVecs > 3)
1547    Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::dsub_3, dl, VT, RegSeq));
1548
1549  Ops.push_back(N->getOperand(FirstTblReg + NumVecs));
1550  Ops.push_back(getAL(CurDAG)); // predicate
1551  Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // predicate register
1552  return CurDAG->getMachineNode(Opc, dl, VT, Ops.data(), Ops.size());
1553}
1554
1555SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N,
1556                                                     bool isSigned) {
1557  if (!Subtarget->hasV6T2Ops())
1558    return NULL;
1559
1560  unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
1561    : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
1562
1563
1564  // For unsigned extracts, check for a shift right and mask
1565  unsigned And_imm = 0;
1566  if (N->getOpcode() == ISD::AND) {
1567    if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
1568
1569      // The immediate is a mask of the low bits iff imm & (imm+1) == 0
1570      if (And_imm & (And_imm + 1))
1571        return NULL;
1572
1573      unsigned Srl_imm = 0;
1574      if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
1575                                Srl_imm)) {
1576        assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1577
1578        unsigned Width = CountTrailingOnes_32(And_imm);
1579        unsigned LSB = Srl_imm;
1580        SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1581        SDValue Ops[] = { N->getOperand(0).getOperand(0),
1582                          CurDAG->getTargetConstant(LSB, MVT::i32),
1583                          CurDAG->getTargetConstant(Width, MVT::i32),
1584          getAL(CurDAG), Reg0 };
1585        return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1586      }
1587    }
1588    return NULL;
1589  }
1590
1591  // Otherwise, we're looking for a shift of a shift
1592  unsigned Shl_imm = 0;
1593  if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
1594    assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
1595    unsigned Srl_imm = 0;
1596    if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
1597      assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
1598      unsigned Width = 32 - Srl_imm;
1599      int LSB = Srl_imm - Shl_imm;
1600      if (LSB < 0)
1601        return NULL;
1602      SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1603      SDValue Ops[] = { N->getOperand(0).getOperand(0),
1604                        CurDAG->getTargetConstant(LSB, MVT::i32),
1605                        CurDAG->getTargetConstant(Width, MVT::i32),
1606                        getAL(CurDAG), Reg0 };
1607      return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1608    }
1609  }
1610  return NULL;
1611}
1612
1613SDNode *ARMDAGToDAGISel::
1614SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1615                    ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1616  SDValue CPTmp0;
1617  SDValue CPTmp1;
1618  if (SelectT2ShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1)) {
1619    unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
1620    unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
1621    unsigned Opc = 0;
1622    switch (SOShOp) {
1623    case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
1624    case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
1625    case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
1626    case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
1627    default:
1628      llvm_unreachable("Unknown so_reg opcode!");
1629      break;
1630    }
1631    SDValue SOShImm =
1632      CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
1633    SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1634    SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag };
1635    return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6);
1636  }
1637  return 0;
1638}
1639
1640SDNode *ARMDAGToDAGISel::
1641SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1642                     ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1643  SDValue CPTmp0;
1644  SDValue CPTmp1;
1645  SDValue CPTmp2;
1646  if (SelectShifterOperandReg(N, TrueVal, CPTmp0, CPTmp1, CPTmp2)) {
1647    SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1648    SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag };
1649    return CurDAG->SelectNodeTo(N, ARM::MOVCCs, MVT::i32, Ops, 7);
1650  }
1651  return 0;
1652}
1653
1654SDNode *ARMDAGToDAGISel::
1655SelectT2CMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1656                    ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1657  ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1658  if (!T)
1659    return 0;
1660
1661  if (Pred_t2_so_imm(TrueVal.getNode())) {
1662    SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1663    SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1664    SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1665    return CurDAG->SelectNodeTo(N,
1666                                ARM::t2MOVCCi, MVT::i32, Ops, 5);
1667  }
1668  return 0;
1669}
1670
1671SDNode *ARMDAGToDAGISel::
1672SelectARMCMOVSoImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal,
1673                     ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) {
1674  ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal);
1675  if (!T)
1676    return 0;
1677
1678  if (Pred_so_imm(TrueVal.getNode())) {
1679    SDValue True = CurDAG->getTargetConstant(T->getZExtValue(), MVT::i32);
1680    SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32);
1681    SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag };
1682    return CurDAG->SelectNodeTo(N,
1683                                ARM::MOVCCi, MVT::i32, Ops, 5);
1684  }
1685  return 0;
1686}
1687
1688SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) {
1689  EVT VT = N->getValueType(0);
1690  SDValue FalseVal = N->getOperand(0);
1691  SDValue TrueVal  = N->getOperand(1);
1692  SDValue CC = N->getOperand(2);
1693  SDValue CCR = N->getOperand(3);
1694  SDValue InFlag = N->getOperand(4);
1695  assert(CC.getOpcode() == ISD::Constant);
1696  assert(CCR.getOpcode() == ISD::Register);
1697  ARMCC::CondCodes CCVal =
1698    (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue();
1699
1700  if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
1701    // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1702    // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
1703    // Pattern complexity = 18  cost = 1  size = 0
1704    SDValue CPTmp0;
1705    SDValue CPTmp1;
1706    SDValue CPTmp2;
1707    if (Subtarget->isThumb()) {
1708      SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal,
1709                                        CCVal, CCR, InFlag);
1710      if (!Res)
1711        Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal,
1712                               ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1713      if (Res)
1714        return Res;
1715    } else {
1716      SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal,
1717                                         CCVal, CCR, InFlag);
1718      if (!Res)
1719        Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal,
1720                               ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1721      if (Res)
1722        return Res;
1723    }
1724
1725    // Pattern: (ARMcmov:i32 GPR:i32:$false,
1726    //             (imm:i32)<<P:Pred_so_imm>>:$true,
1727    //             (imm:i32):$cc)
1728    // Emits: (MOVCCi:i32 GPR:i32:$false,
1729    //           (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
1730    // Pattern complexity = 10  cost = 1  size = 0
1731    if (Subtarget->isThumb()) {
1732      SDNode *Res = SelectT2CMOVSoImmOp(N, FalseVal, TrueVal,
1733                                        CCVal, CCR, InFlag);
1734      if (!Res)
1735        Res = SelectT2CMOVSoImmOp(N, TrueVal, FalseVal,
1736                               ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1737      if (Res)
1738        return Res;
1739    } else {
1740      SDNode *Res = SelectARMCMOVSoImmOp(N, FalseVal, TrueVal,
1741                                         CCVal, CCR, InFlag);
1742      if (!Res)
1743        Res = SelectARMCMOVSoImmOp(N, TrueVal, FalseVal,
1744                               ARMCC::getOppositeCondition(CCVal), CCR, InFlag);
1745      if (Res)
1746        return Res;
1747    }
1748  }
1749
1750  // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1751  // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1752  // Pattern complexity = 6  cost = 1  size = 0
1753  //
1754  // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1755  // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
1756  // Pattern complexity = 6  cost = 11  size = 0
1757  //
1758  // Also FCPYScc and FCPYDcc.
1759  SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32);
1760  SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag };
1761  unsigned Opc = 0;
1762  switch (VT.getSimpleVT().SimpleTy) {
1763  default: assert(false && "Illegal conditional move type!");
1764    break;
1765  case MVT::i32:
1766    Opc = Subtarget->isThumb()
1767      ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
1768      : ARM::MOVCCr;
1769    break;
1770  case MVT::f32:
1771    Opc = ARM::VMOVScc;
1772    break;
1773  case MVT::f64:
1774    Opc = ARM::VMOVDcc;
1775    break;
1776  }
1777  return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
1778}
1779
1780SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
1781  // The only time a CONCAT_VECTORS operation can have legal types is when
1782  // two 64-bit vectors are concatenated to a 128-bit vector.
1783  EVT VT = N->getValueType(0);
1784  if (!VT.is128BitVector() || N->getNumOperands() != 2)
1785    llvm_unreachable("unexpected CONCAT_VECTORS");
1786  DebugLoc dl = N->getDebugLoc();
1787  SDValue V0 = N->getOperand(0);
1788  SDValue V1 = N->getOperand(1);
1789  SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32);
1790  SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32);
1791  const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
1792  return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
1793}
1794
1795SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
1796  DebugLoc dl = N->getDebugLoc();
1797
1798  if (N->isMachineOpcode())
1799    return NULL;   // Already selected.
1800
1801  switch (N->getOpcode()) {
1802  default: break;
1803  case ISD::Constant: {
1804    unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
1805    bool UseCP = true;
1806    if (Subtarget->hasThumb2())
1807      // Thumb2-aware targets have the MOVT instruction, so all immediates can
1808      // be done with MOV + MOVT, at worst.
1809      UseCP = 0;
1810    else {
1811      if (Subtarget->isThumb()) {
1812        UseCP = (Val > 255 &&                          // MOV
1813                 ~Val > 255 &&                         // MOV + MVN
1814                 !ARM_AM::isThumbImmShiftedVal(Val));  // MOV + LSL
1815      } else
1816        UseCP = (ARM_AM::getSOImmVal(Val) == -1 &&     // MOV
1817                 ARM_AM::getSOImmVal(~Val) == -1 &&    // MVN
1818                 !ARM_AM::isSOImmTwoPartVal(Val));     // two instrs.
1819    }
1820
1821    if (UseCP) {
1822      SDValue CPIdx =
1823        CurDAG->getTargetConstantPool(ConstantInt::get(
1824                                  Type::getInt32Ty(*CurDAG->getContext()), Val),
1825                                      TLI.getPointerTy());
1826
1827      SDNode *ResNode;
1828      if (Subtarget->isThumb1Only()) {
1829        SDValue Pred = getAL(CurDAG);
1830        SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
1831        SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
1832        ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
1833                                         Ops, 4);
1834      } else {
1835        SDValue Ops[] = {
1836          CPIdx,
1837          CurDAG->getRegister(0, MVT::i32),
1838          CurDAG->getTargetConstant(0, MVT::i32),
1839          getAL(CurDAG),
1840          CurDAG->getRegister(0, MVT::i32),
1841          CurDAG->getEntryNode()
1842        };
1843        ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
1844                                       Ops, 6);
1845      }
1846      ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0));
1847      return NULL;
1848    }
1849
1850    // Other cases are autogenerated.
1851    break;
1852  }
1853  case ISD::FrameIndex: {
1854    // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
1855    int FI = cast<FrameIndexSDNode>(N)->getIndex();
1856    SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
1857    if (Subtarget->isThumb1Only()) {
1858      return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
1859                                  CurDAG->getTargetConstant(0, MVT::i32));
1860    } else {
1861      unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
1862                      ARM::t2ADDri : ARM::ADDri);
1863      SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
1864                        getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1865                        CurDAG->getRegister(0, MVT::i32) };
1866      return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
1867    }
1868  }
1869  case ISD::SRL:
1870    if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1871      return I;
1872    break;
1873  case ISD::SRA:
1874    if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true))
1875      return I;
1876    break;
1877  case ISD::MUL:
1878    if (Subtarget->isThumb1Only())
1879      break;
1880    if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1881      unsigned RHSV = C->getZExtValue();
1882      if (!RHSV) break;
1883      if (isPowerOf2_32(RHSV-1)) {  // 2^n+1?
1884        unsigned ShImm = Log2_32(RHSV-1);
1885        if (ShImm >= 32)
1886          break;
1887        SDValue V = N->getOperand(0);
1888        ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1889        SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1890        SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1891        if (Subtarget->isThumb()) {
1892          SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1893          return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
1894        } else {
1895          SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1896          return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
1897        }
1898      }
1899      if (isPowerOf2_32(RHSV+1)) {  // 2^n-1?
1900        unsigned ShImm = Log2_32(RHSV+1);
1901        if (ShImm >= 32)
1902          break;
1903        SDValue V = N->getOperand(0);
1904        ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
1905        SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
1906        SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
1907        if (Subtarget->isThumb()) {
1908          SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1909          return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 6);
1910        } else {
1911          SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
1912          return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
1913        }
1914      }
1915    }
1916    break;
1917  case ISD::AND: {
1918    // Check for unsigned bitfield extract
1919    if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false))
1920      return I;
1921
1922    // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
1923    // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
1924    // are entirely contributed by c2 and lower 16-bits are entirely contributed
1925    // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
1926    // Select it to: "movt x, ((c1 & 0xffff) >> 16)
1927    EVT VT = N->getValueType(0);
1928    if (VT != MVT::i32)
1929      break;
1930    unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
1931      ? ARM::t2MOVTi16
1932      : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
1933    if (!Opc)
1934      break;
1935    SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
1936    ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1937    if (!N1C)
1938      break;
1939    if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
1940      SDValue N2 = N0.getOperand(1);
1941      ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1942      if (!N2C)
1943        break;
1944      unsigned N1CVal = N1C->getZExtValue();
1945      unsigned N2CVal = N2C->getZExtValue();
1946      if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
1947          (N1CVal & 0xffffU) == 0xffffU &&
1948          (N2CVal & 0xffffU) == 0x0U) {
1949        SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
1950                                                  MVT::i32);
1951        SDValue Ops[] = { N0.getOperand(0), Imm16,
1952                          getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1953        return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4);
1954      }
1955    }
1956    break;
1957  }
1958  case ARMISD::VMOVRRD:
1959    return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32,
1960                                  N->getOperand(0), getAL(CurDAG),
1961                                  CurDAG->getRegister(0, MVT::i32));
1962  case ISD::UMUL_LOHI: {
1963    if (Subtarget->isThumb1Only())
1964      break;
1965    if (Subtarget->isThumb()) {
1966      SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1967                        getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1968                        CurDAG->getRegister(0, MVT::i32) };
1969      return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32,Ops,4);
1970    } else {
1971      SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1972                        getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1973                        CurDAG->getRegister(0, MVT::i32) };
1974      return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1975    }
1976  }
1977  case ISD::SMUL_LOHI: {
1978    if (Subtarget->isThumb1Only())
1979      break;
1980    if (Subtarget->isThumb()) {
1981      SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1982                        getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
1983      return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32,Ops,4);
1984    } else {
1985      SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
1986                        getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
1987                        CurDAG->getRegister(0, MVT::i32) };
1988      return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
1989    }
1990  }
1991  case ISD::LOAD: {
1992    SDNode *ResNode = 0;
1993    if (Subtarget->isThumb() && Subtarget->hasThumb2())
1994      ResNode = SelectT2IndexedLoad(N);
1995    else
1996      ResNode = SelectARMIndexedLoad(N);
1997    if (ResNode)
1998      return ResNode;
1999    // Other cases are autogenerated.
2000    break;
2001  }
2002  case ARMISD::BRCOND: {
2003    // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2004    // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2005    // Pattern complexity = 6  cost = 1  size = 0
2006
2007    // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2008    // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
2009    // Pattern complexity = 6  cost = 1  size = 0
2010
2011    // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
2012    // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
2013    // Pattern complexity = 6  cost = 1  size = 0
2014
2015    unsigned Opc = Subtarget->isThumb() ?
2016      ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
2017    SDValue Chain = N->getOperand(0);
2018    SDValue N1 = N->getOperand(1);
2019    SDValue N2 = N->getOperand(2);
2020    SDValue N3 = N->getOperand(3);
2021    SDValue InFlag = N->getOperand(4);
2022    assert(N1.getOpcode() == ISD::BasicBlock);
2023    assert(N2.getOpcode() == ISD::Constant);
2024    assert(N3.getOpcode() == ISD::Register);
2025
2026    SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2027                               cast<ConstantSDNode>(N2)->getZExtValue()),
2028                               MVT::i32);
2029    SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
2030    SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
2031                                             MVT::Flag, Ops, 5);
2032    Chain = SDValue(ResNode, 0);
2033    if (N->getNumValues() == 2) {
2034      InFlag = SDValue(ResNode, 1);
2035      ReplaceUses(SDValue(N, 1), InFlag);
2036    }
2037    ReplaceUses(SDValue(N, 0),
2038                SDValue(Chain.getNode(), Chain.getResNo()));
2039    return NULL;
2040  }
2041  case ARMISD::CMOV:
2042    return SelectCMOVOp(N);
2043  case ARMISD::CNEG: {
2044    EVT VT = N->getValueType(0);
2045    SDValue N0 = N->getOperand(0);
2046    SDValue N1 = N->getOperand(1);
2047    SDValue N2 = N->getOperand(2);
2048    SDValue N3 = N->getOperand(3);
2049    SDValue InFlag = N->getOperand(4);
2050    assert(N2.getOpcode() == ISD::Constant);
2051    assert(N3.getOpcode() == ISD::Register);
2052
2053    SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
2054                               cast<ConstantSDNode>(N2)->getZExtValue()),
2055                               MVT::i32);
2056    SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
2057    unsigned Opc = 0;
2058    switch (VT.getSimpleVT().SimpleTy) {
2059    default: assert(false && "Illegal conditional move type!");
2060      break;
2061    case MVT::f32:
2062      Opc = ARM::VNEGScc;
2063      break;
2064    case MVT::f64:
2065      Opc = ARM::VNEGDcc;
2066      break;
2067    }
2068    return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5);
2069  }
2070
2071  case ARMISD::VZIP: {
2072    unsigned Opc = 0;
2073    EVT VT = N->getValueType(0);
2074    switch (VT.getSimpleVT().SimpleTy) {
2075    default: return NULL;
2076    case MVT::v8i8:  Opc = ARM::VZIPd8; break;
2077    case MVT::v4i16: Opc = ARM::VZIPd16; break;
2078    case MVT::v2f32:
2079    case MVT::v2i32: Opc = ARM::VZIPd32; break;
2080    case MVT::v16i8: Opc = ARM::VZIPq8; break;
2081    case MVT::v8i16: Opc = ARM::VZIPq16; break;
2082    case MVT::v4f32:
2083    case MVT::v4i32: Opc = ARM::VZIPq32; break;
2084    }
2085    SDValue Pred = getAL(CurDAG);
2086    SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2087    SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2088    return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2089  }
2090  case ARMISD::VUZP: {
2091    unsigned Opc = 0;
2092    EVT VT = N->getValueType(0);
2093    switch (VT.getSimpleVT().SimpleTy) {
2094    default: return NULL;
2095    case MVT::v8i8:  Opc = ARM::VUZPd8; break;
2096    case MVT::v4i16: Opc = ARM::VUZPd16; break;
2097    case MVT::v2f32:
2098    case MVT::v2i32: Opc = ARM::VUZPd32; break;
2099    case MVT::v16i8: Opc = ARM::VUZPq8; break;
2100    case MVT::v8i16: Opc = ARM::VUZPq16; break;
2101    case MVT::v4f32:
2102    case MVT::v4i32: Opc = ARM::VUZPq32; break;
2103    }
2104    SDValue Pred = getAL(CurDAG);
2105    SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2106    SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2107    return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2108  }
2109  case ARMISD::VTRN: {
2110    unsigned Opc = 0;
2111    EVT VT = N->getValueType(0);
2112    switch (VT.getSimpleVT().SimpleTy) {
2113    default: return NULL;
2114    case MVT::v8i8:  Opc = ARM::VTRNd8; break;
2115    case MVT::v4i16: Opc = ARM::VTRNd16; break;
2116    case MVT::v2f32:
2117    case MVT::v2i32: Opc = ARM::VTRNd32; break;
2118    case MVT::v16i8: Opc = ARM::VTRNq8; break;
2119    case MVT::v8i16: Opc = ARM::VTRNq16; break;
2120    case MVT::v4f32:
2121    case MVT::v4i32: Opc = ARM::VTRNq32; break;
2122    }
2123    SDValue Pred = getAL(CurDAG);
2124    SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2125    SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
2126    return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
2127  }
2128  case ARMISD::BUILD_VECTOR: {
2129    EVT VecVT = N->getValueType(0);
2130    EVT EltVT = VecVT.getVectorElementType();
2131    unsigned NumElts = VecVT.getVectorNumElements();
2132    if (EltVT.getSimpleVT() == MVT::f64) {
2133      assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
2134      return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1));
2135    }
2136    assert(EltVT.getSimpleVT() == MVT::f32 &&
2137           "unexpected type for BUILD_VECTOR");
2138    if (NumElts == 2)
2139      return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1));
2140    assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
2141    return QuadSRegs(VecVT, N->getOperand(0), N->getOperand(1),
2142                     N->getOperand(2), N->getOperand(3));
2143  }
2144
2145  case ISD::INTRINSIC_VOID:
2146  case ISD::INTRINSIC_W_CHAIN: {
2147    unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
2148    switch (IntNo) {
2149    default:
2150      break;
2151
2152    case Intrinsic::arm_neon_vld1: {
2153      unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
2154                              ARM::VLD1d32, ARM::VLD1d64 };
2155      unsigned QOpcodes[] = { ARM::VLD1q8Pseudo, ARM::VLD1q16Pseudo,
2156                              ARM::VLD1q32Pseudo, ARM::VLD1q64Pseudo };
2157      return SelectVLD(N, 1, DOpcodes, QOpcodes, 0);
2158    }
2159
2160    case Intrinsic::arm_neon_vld2: {
2161      unsigned DOpcodes[] = { ARM::VLD2d8Pseudo, ARM::VLD2d16Pseudo,
2162                              ARM::VLD2d32Pseudo, ARM::VLD1q64Pseudo };
2163      unsigned QOpcodes[] = { ARM::VLD2q8Pseudo, ARM::VLD2q16Pseudo,
2164                              ARM::VLD2q32Pseudo };
2165      return SelectVLD(N, 2, DOpcodes, QOpcodes, 0);
2166    }
2167
2168    case Intrinsic::arm_neon_vld3: {
2169      unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16,
2170                              ARM::VLD3d32, ARM::VLD1d64T };
2171      unsigned QOpcodes0[] = { ARM::VLD3q8_UPD,
2172                               ARM::VLD3q16_UPD,
2173                               ARM::VLD3q32_UPD };
2174      unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD,
2175                               ARM::VLD3q16odd_UPD,
2176                               ARM::VLD3q32odd_UPD };
2177      return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2178    }
2179
2180    case Intrinsic::arm_neon_vld4: {
2181      unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16,
2182                              ARM::VLD4d32, ARM::VLD1d64Q };
2183      unsigned QOpcodes0[] = { ARM::VLD4q8_UPD,
2184                               ARM::VLD4q16_UPD,
2185                               ARM::VLD4q32_UPD };
2186      unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD,
2187                               ARM::VLD4q16odd_UPD,
2188                               ARM::VLD4q32odd_UPD };
2189      return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2190    }
2191
2192    case Intrinsic::arm_neon_vld2lane: {
2193      unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 };
2194      unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 };
2195      unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd };
2196      return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
2197    }
2198
2199    case Intrinsic::arm_neon_vld3lane: {
2200      unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 };
2201      unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 };
2202      unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd };
2203      return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
2204    }
2205
2206    case Intrinsic::arm_neon_vld4lane: {
2207      unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 };
2208      unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 };
2209      unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd };
2210      return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
2211    }
2212
2213    case Intrinsic::arm_neon_vst1: {
2214      unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
2215                              ARM::VST1d32, ARM::VST1d64 };
2216      unsigned QOpcodes[] = { ARM::VST1q8Pseudo, ARM::VST1q16Pseudo,
2217                              ARM::VST1q32Pseudo, ARM::VST1q64Pseudo };
2218      return SelectVST(N, 1, DOpcodes, QOpcodes, 0);
2219    }
2220
2221    case Intrinsic::arm_neon_vst2: {
2222      unsigned DOpcodes[] = { ARM::VST2d8Pseudo, ARM::VST2d16Pseudo,
2223                              ARM::VST2d32Pseudo, ARM::VST1q64Pseudo };
2224      unsigned QOpcodes[] = { ARM::VST2q8Pseudo, ARM::VST2q16Pseudo,
2225                              ARM::VST2q32Pseudo };
2226      return SelectVST(N, 2, DOpcodes, QOpcodes, 0);
2227    }
2228
2229    case Intrinsic::arm_neon_vst3: {
2230      unsigned DOpcodes[] = { ARM::VST3d8Pseudo, ARM::VST3d16Pseudo,
2231                              ARM::VST3d32Pseudo, ARM::VST1d64TPseudo };
2232      unsigned QOpcodes0[] = { ARM::VST3q8Pseudo_UPD,
2233                               ARM::VST3q16Pseudo_UPD,
2234                               ARM::VST3q32Pseudo_UPD };
2235      unsigned QOpcodes1[] = { ARM::VST3q8oddPseudo_UPD,
2236                               ARM::VST3q16oddPseudo_UPD,
2237                               ARM::VST3q32oddPseudo_UPD };
2238      return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1);
2239    }
2240
2241    case Intrinsic::arm_neon_vst4: {
2242      unsigned DOpcodes[] = { ARM::VST4d8Pseudo, ARM::VST4d16Pseudo,
2243                              ARM::VST4d32Pseudo, ARM::VST1d64QPseudo };
2244      unsigned QOpcodes0[] = { ARM::VST4q8Pseudo_UPD,
2245                               ARM::VST4q16Pseudo_UPD,
2246                               ARM::VST4q32Pseudo_UPD };
2247      unsigned QOpcodes1[] = { ARM::VST4q8oddPseudo_UPD,
2248                               ARM::VST4q16oddPseudo_UPD,
2249                               ARM::VST4q32oddPseudo_UPD };
2250      return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1);
2251    }
2252
2253    case Intrinsic::arm_neon_vst2lane: {
2254      unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 };
2255      unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 };
2256      unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd };
2257      return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1);
2258    }
2259
2260    case Intrinsic::arm_neon_vst3lane: {
2261      unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 };
2262      unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 };
2263      unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd };
2264      return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
2265    }
2266
2267    case Intrinsic::arm_neon_vst4lane: {
2268      unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 };
2269      unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 };
2270      unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd };
2271      return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
2272    }
2273    }
2274    break;
2275  }
2276
2277  case ISD::INTRINSIC_WO_CHAIN: {
2278    unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
2279    switch (IntNo) {
2280    default:
2281      break;
2282
2283    case Intrinsic::arm_neon_vtbl2:
2284      return SelectVTBL(N, false, 2, ARM::VTBL2);
2285    case Intrinsic::arm_neon_vtbl3:
2286      return SelectVTBL(N, false, 3, ARM::VTBL3);
2287    case Intrinsic::arm_neon_vtbl4:
2288      return SelectVTBL(N, false, 4, ARM::VTBL4);
2289
2290    case Intrinsic::arm_neon_vtbx2:
2291      return SelectVTBL(N, true, 2, ARM::VTBX2);
2292    case Intrinsic::arm_neon_vtbx3:
2293      return SelectVTBL(N, true, 3, ARM::VTBX3);
2294    case Intrinsic::arm_neon_vtbx4:
2295      return SelectVTBL(N, true, 4, ARM::VTBX4);
2296    }
2297    break;
2298  }
2299
2300  case ISD::CONCAT_VECTORS:
2301    return SelectConcatVector(N);
2302  }
2303
2304  return SelectCode(N);
2305}
2306
2307bool ARMDAGToDAGISel::
2308SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
2309                             std::vector<SDValue> &OutOps) {
2310  assert(ConstraintCode == 'm' && "unexpected asm memory constraint");
2311  // Require the address to be in a register.  That is safe for all ARM
2312  // variants and it is hard to do anything much smarter without knowing
2313  // how the operand is used.
2314  OutOps.push_back(Op);
2315  return false;
2316}
2317
2318/// createARMISelDag - This pass converts a legalized DAG into a
2319/// ARM-specific DAG, ready for instruction scheduling.
2320///
2321FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
2322                                     CodeGenOpt::Level OptLevel) {
2323  return new ARMDAGToDAGISel(TM, OptLevel);
2324}
2325