xref: /freebsd/contrib/llvm-project/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp (revision 258a0d760aa8b42899a000e30f610f900a402556)
1 //===-- ARMISelDAGToDAG.cpp - A dag to dag inst selector for ARM ----------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines an instruction selector for the ARM target.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "ARM.h"
14 #include "ARMBaseInstrInfo.h"
15 #include "ARMTargetMachine.h"
16 #include "MCTargetDesc/ARMAddressingModes.h"
17 #include "Utils/ARMBaseInfo.h"
18 #include "llvm/ADT/APSInt.h"
19 #include "llvm/ADT/StringSwitch.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/SelectionDAG.h"
25 #include "llvm/CodeGen/SelectionDAGISel.h"
26 #include "llvm/CodeGen/TargetLowering.h"
27 #include "llvm/IR/CallingConv.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/IntrinsicsARM.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include <optional>
39 
40 using namespace llvm;
41 
42 #define DEBUG_TYPE "arm-isel"
43 #define PASS_NAME "ARM Instruction Selection"
44 
45 static cl::opt<bool>
46 DisableShifterOp("disable-shifter-op", cl::Hidden,
47   cl::desc("Disable isel of shifter-op"),
48   cl::init(false));
49 
50 //===--------------------------------------------------------------------===//
51 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
52 /// instructions for SelectionDAG operations.
53 ///
54 namespace {
55 
56 class ARMDAGToDAGISel : public SelectionDAGISel {
57   /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
58   /// make the right decision when generating code for different targets.
59   const ARMSubtarget *Subtarget;
60 
61 public:
62   static char ID;
63 
64   ARMDAGToDAGISel() = delete;
65 
66   explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm, CodeGenOpt::Level OptLevel)
67       : SelectionDAGISel(ID, tm, OptLevel) {}
68 
69   bool runOnMachineFunction(MachineFunction &MF) override {
70     // Reset the subtarget each time through.
71     Subtarget = &MF.getSubtarget<ARMSubtarget>();
72     SelectionDAGISel::runOnMachineFunction(MF);
73     return true;
74   }
75 
76   void PreprocessISelDAG() override;
77 
78   /// getI32Imm - Return a target constant of type i32 with the specified
79   /// value.
80   inline SDValue getI32Imm(unsigned Imm, const SDLoc &dl) {
81     return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
82   }
83 
84   void Select(SDNode *N) override;
85 
86   /// Return true as some complex patterns, like those that call
87   /// canExtractShiftFromMul can modify the DAG inplace.
88   bool ComplexPatternFuncMutatesDAG() const override { return true; }
89 
90   bool hasNoVMLxHazardUse(SDNode *N) const;
91   bool isShifterOpProfitable(const SDValue &Shift,
92                              ARM_AM::ShiftOpc ShOpcVal, unsigned ShAmt);
93   bool SelectRegShifterOperand(SDValue N, SDValue &A,
94                                SDValue &B, SDValue &C,
95                                bool CheckProfitability = true);
96   bool SelectImmShifterOperand(SDValue N, SDValue &A,
97                                SDValue &B, bool CheckProfitability = true);
98   bool SelectShiftRegShifterOperand(SDValue N, SDValue &A, SDValue &B,
99                                     SDValue &C) {
100     // Don't apply the profitability check
101     return SelectRegShifterOperand(N, A, B, C, false);
102   }
103   bool SelectShiftImmShifterOperand(SDValue N, SDValue &A, SDValue &B) {
104     // Don't apply the profitability check
105     return SelectImmShifterOperand(N, A, B, false);
106   }
107   bool SelectShiftImmShifterOperandOneUse(SDValue N, SDValue &A, SDValue &B) {
108     if (!N.hasOneUse())
109       return false;
110     return SelectImmShifterOperand(N, A, B, false);
111   }
112 
113   bool SelectAddLikeOr(SDNode *Parent, SDValue N, SDValue &Out);
114 
115   bool SelectAddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm);
116   bool SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset, SDValue &Opc);
117 
118   bool SelectCMOVPred(SDValue N, SDValue &Pred, SDValue &Reg) {
119     const ConstantSDNode *CN = cast<ConstantSDNode>(N);
120     Pred = CurDAG->getTargetConstant(CN->getZExtValue(), SDLoc(N), MVT::i32);
121     Reg = CurDAG->getRegister(ARM::CPSR, MVT::i32);
122     return true;
123   }
124 
125   bool SelectAddrMode2OffsetReg(SDNode *Op, SDValue N,
126                              SDValue &Offset, SDValue &Opc);
127   bool SelectAddrMode2OffsetImm(SDNode *Op, SDValue N,
128                              SDValue &Offset, SDValue &Opc);
129   bool SelectAddrMode2OffsetImmPre(SDNode *Op, SDValue N,
130                              SDValue &Offset, SDValue &Opc);
131   bool SelectAddrOffsetNone(SDValue N, SDValue &Base);
132   bool SelectAddrMode3(SDValue N, SDValue &Base,
133                        SDValue &Offset, SDValue &Opc);
134   bool SelectAddrMode3Offset(SDNode *Op, SDValue N,
135                              SDValue &Offset, SDValue &Opc);
136   bool IsAddressingMode5(SDValue N, SDValue &Base, SDValue &Offset, bool FP16);
137   bool SelectAddrMode5(SDValue N, SDValue &Base, SDValue &Offset);
138   bool SelectAddrMode5FP16(SDValue N, SDValue &Base, SDValue &Offset);
139   bool SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,SDValue &Align);
140   bool SelectAddrMode6Offset(SDNode *Op, SDValue N, SDValue &Offset);
141 
142   bool SelectAddrModePC(SDValue N, SDValue &Offset, SDValue &Label);
143 
144   // Thumb Addressing Modes:
145   bool SelectThumbAddrModeRR(SDValue N, SDValue &Base, SDValue &Offset);
146   bool SelectThumbAddrModeRRSext(SDValue N, SDValue &Base, SDValue &Offset);
147   bool SelectThumbAddrModeImm5S(SDValue N, unsigned Scale, SDValue &Base,
148                                 SDValue &OffImm);
149   bool SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base,
150                                  SDValue &OffImm);
151   bool SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base,
152                                  SDValue &OffImm);
153   bool SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base,
154                                  SDValue &OffImm);
155   bool SelectThumbAddrModeSP(SDValue N, SDValue &Base, SDValue &OffImm);
156   template <unsigned Shift>
157   bool SelectTAddrModeImm7(SDValue N, SDValue &Base, SDValue &OffImm);
158 
159   // Thumb 2 Addressing Modes:
160   bool SelectT2AddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm);
161   template <unsigned Shift>
162   bool SelectT2AddrModeImm8(SDValue N, SDValue &Base, SDValue &OffImm);
163   bool SelectT2AddrModeImm8(SDValue N, SDValue &Base,
164                             SDValue &OffImm);
165   bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
166                                  SDValue &OffImm);
167   template <unsigned Shift>
168   bool SelectT2AddrModeImm7Offset(SDNode *Op, SDValue N, SDValue &OffImm);
169   bool SelectT2AddrModeImm7Offset(SDNode *Op, SDValue N, SDValue &OffImm,
170                                   unsigned Shift);
171   template <unsigned Shift>
172   bool SelectT2AddrModeImm7(SDValue N, SDValue &Base, SDValue &OffImm);
173   bool SelectT2AddrModeSoReg(SDValue N, SDValue &Base,
174                              SDValue &OffReg, SDValue &ShImm);
175   bool SelectT2AddrModeExclusive(SDValue N, SDValue &Base, SDValue &OffImm);
176 
177   template<int Min, int Max>
178   bool SelectImmediateInRange(SDValue N, SDValue &OffImm);
179 
180   inline bool is_so_imm(unsigned Imm) const {
181     return ARM_AM::getSOImmVal(Imm) != -1;
182   }
183 
184   inline bool is_so_imm_not(unsigned Imm) const {
185     return ARM_AM::getSOImmVal(~Imm) != -1;
186   }
187 
188   inline bool is_t2_so_imm(unsigned Imm) const {
189     return ARM_AM::getT2SOImmVal(Imm) != -1;
190   }
191 
192   inline bool is_t2_so_imm_not(unsigned Imm) const {
193     return ARM_AM::getT2SOImmVal(~Imm) != -1;
194   }
195 
196   // Include the pieces autogenerated from the target description.
197 #include "ARMGenDAGISel.inc"
198 
199 private:
200   void transferMemOperands(SDNode *Src, SDNode *Dst);
201 
202   /// Indexed (pre/post inc/dec) load matching code for ARM.
203   bool tryARMIndexedLoad(SDNode *N);
204   bool tryT1IndexedLoad(SDNode *N);
205   bool tryT2IndexedLoad(SDNode *N);
206   bool tryMVEIndexedLoad(SDNode *N);
207   bool tryFMULFixed(SDNode *N, SDLoc dl);
208   bool tryFP_TO_INT(SDNode *N, SDLoc dl);
209   bool transformFixedFloatingPointConversion(SDNode *N, SDNode *FMul,
210                                              bool IsUnsigned,
211                                              bool FixedToFloat);
212 
213   /// SelectVLD - Select NEON load intrinsics.  NumVecs should be
214   /// 1, 2, 3 or 4.  The opcode arrays specify the instructions used for
215   /// loads of D registers and even subregs and odd subregs of Q registers.
216   /// For NumVecs <= 2, QOpcodes1 is not used.
217   void SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs,
218                  const uint16_t *DOpcodes, const uint16_t *QOpcodes0,
219                  const uint16_t *QOpcodes1);
220 
221   /// SelectVST - Select NEON store intrinsics.  NumVecs should
222   /// be 1, 2, 3 or 4.  The opcode arrays specify the instructions used for
223   /// stores of D registers and even subregs and odd subregs of Q registers.
224   /// For NumVecs <= 2, QOpcodes1 is not used.
225   void SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs,
226                  const uint16_t *DOpcodes, const uint16_t *QOpcodes0,
227                  const uint16_t *QOpcodes1);
228 
229   /// SelectVLDSTLane - Select NEON load/store lane intrinsics.  NumVecs should
230   /// be 2, 3 or 4.  The opcode arrays specify the instructions used for
231   /// load/store of D registers and Q registers.
232   void SelectVLDSTLane(SDNode *N, bool IsLoad, bool isUpdating,
233                        unsigned NumVecs, const uint16_t *DOpcodes,
234                        const uint16_t *QOpcodes);
235 
236   /// Helper functions for setting up clusters of MVE predication operands.
237   template <typename SDValueVector>
238   void AddMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc,
239                             SDValue PredicateMask);
240   template <typename SDValueVector>
241   void AddMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc,
242                             SDValue PredicateMask, SDValue Inactive);
243 
244   template <typename SDValueVector>
245   void AddEmptyMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc);
246   template <typename SDValueVector>
247   void AddEmptyMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc, EVT InactiveTy);
248 
249   /// SelectMVE_WB - Select MVE writeback load/store intrinsics.
250   void SelectMVE_WB(SDNode *N, const uint16_t *Opcodes, bool Predicated);
251 
252   /// SelectMVE_LongShift - Select MVE 64-bit scalar shift intrinsics.
253   void SelectMVE_LongShift(SDNode *N, uint16_t Opcode, bool Immediate,
254                            bool HasSaturationOperand);
255 
256   /// SelectMVE_VADCSBC - Select MVE vector add/sub-with-carry intrinsics.
257   void SelectMVE_VADCSBC(SDNode *N, uint16_t OpcodeWithCarry,
258                          uint16_t OpcodeWithNoCarry, bool Add, bool Predicated);
259 
260   /// SelectMVE_VSHLC - Select MVE intrinsics for a shift that carries between
261   /// vector lanes.
262   void SelectMVE_VSHLC(SDNode *N, bool Predicated);
263 
264   /// Select long MVE vector reductions with two vector operands
265   /// Stride is the number of vector element widths the instruction can operate
266   /// on:
267   /// 2 for long non-rounding variants, vml{a,s}ldav[a][x]: [i16, i32]
268   /// 1 for long rounding variants: vrml{a,s}ldavh[a][x]: [i32]
269   /// Stride is used when addressing the OpcodesS array which contains multiple
270   /// opcodes for each element width.
271   /// TySize is the index into the list of element types listed above
272   void SelectBaseMVE_VMLLDAV(SDNode *N, bool Predicated,
273                              const uint16_t *OpcodesS, const uint16_t *OpcodesU,
274                              size_t Stride, size_t TySize);
275 
276   /// Select a 64-bit MVE vector reduction with two vector operands
277   /// arm_mve_vmlldava_[predicated]
278   void SelectMVE_VMLLDAV(SDNode *N, bool Predicated, const uint16_t *OpcodesS,
279                          const uint16_t *OpcodesU);
280   /// Select a 72-bit MVE vector rounding reduction with two vector operands
281   /// int_arm_mve_vrmlldavha[_predicated]
282   void SelectMVE_VRMLLDAVH(SDNode *N, bool Predicated, const uint16_t *OpcodesS,
283                            const uint16_t *OpcodesU);
284 
285   /// SelectMVE_VLD - Select MVE interleaving load intrinsics. NumVecs
286   /// should be 2 or 4. The opcode array specifies the instructions
287   /// used for 8, 16 and 32-bit lane sizes respectively, and each
288   /// pointer points to a set of NumVecs sub-opcodes used for the
289   /// different stages (e.g. VLD20 versus VLD21) of each load family.
290   void SelectMVE_VLD(SDNode *N, unsigned NumVecs,
291                      const uint16_t *const *Opcodes, bool HasWriteback);
292 
293   /// SelectMVE_VxDUP - Select MVE incrementing-dup instructions. Opcodes is an
294   /// array of 3 elements for the 8, 16 and 32-bit lane sizes.
295   void SelectMVE_VxDUP(SDNode *N, const uint16_t *Opcodes,
296                        bool Wrapping, bool Predicated);
297 
298   /// Select SelectCDE_CXxD - Select CDE dual-GPR instruction (one of CX1D,
299   /// CX1DA, CX2D, CX2DA, CX3, CX3DA).
300   /// \arg \c NumExtraOps number of extra operands besides the coprocossor,
301   ///                     the accumulator and the immediate operand, i.e. 0
302   ///                     for CX1*, 1 for CX2*, 2 for CX3*
303   /// \arg \c HasAccum whether the instruction has an accumulator operand
304   void SelectCDE_CXxD(SDNode *N, uint16_t Opcode, size_t NumExtraOps,
305                       bool HasAccum);
306 
307   /// SelectVLDDup - Select NEON load-duplicate intrinsics.  NumVecs
308   /// should be 1, 2, 3 or 4.  The opcode array specifies the instructions used
309   /// for loading D registers.
310   void SelectVLDDup(SDNode *N, bool IsIntrinsic, bool isUpdating,
311                     unsigned NumVecs, const uint16_t *DOpcodes,
312                     const uint16_t *QOpcodes0 = nullptr,
313                     const uint16_t *QOpcodes1 = nullptr);
314 
315   /// Try to select SBFX/UBFX instructions for ARM.
316   bool tryV6T2BitfieldExtractOp(SDNode *N, bool isSigned);
317 
318   bool tryInsertVectorElt(SDNode *N);
319 
320   // Select special operations if node forms integer ABS pattern
321   bool tryABSOp(SDNode *N);
322 
323   bool tryReadRegister(SDNode *N);
324   bool tryWriteRegister(SDNode *N);
325 
326   bool tryInlineAsm(SDNode *N);
327 
328   void SelectCMPZ(SDNode *N, bool &SwitchEQNEToPLMI);
329 
330   void SelectCMP_SWAP(SDNode *N);
331 
332   /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
333   /// inline asm expressions.
334   bool SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
335                                     std::vector<SDValue> &OutOps) override;
336 
337   // Form pairs of consecutive R, S, D, or Q registers.
338   SDNode *createGPRPairNode(EVT VT, SDValue V0, SDValue V1);
339   SDNode *createSRegPairNode(EVT VT, SDValue V0, SDValue V1);
340   SDNode *createDRegPairNode(EVT VT, SDValue V0, SDValue V1);
341   SDNode *createQRegPairNode(EVT VT, SDValue V0, SDValue V1);
342 
343   // Form sequences of 4 consecutive S, D, or Q registers.
344   SDNode *createQuadSRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
345   SDNode *createQuadDRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
346   SDNode *createQuadQRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
347 
348   // Get the alignment operand for a NEON VLD or VST instruction.
349   SDValue GetVLDSTAlign(SDValue Align, const SDLoc &dl, unsigned NumVecs,
350                         bool is64BitVector);
351 
352   /// Checks if N is a multiplication by a constant where we can extract out a
353   /// power of two from the constant so that it can be used in a shift, but only
354   /// if it simplifies the materialization of the constant. Returns true if it
355   /// is, and assigns to PowerOfTwo the power of two that should be extracted
356   /// out and to NewMulConst the new constant to be multiplied by.
357   bool canExtractShiftFromMul(const SDValue &N, unsigned MaxShift,
358                               unsigned &PowerOfTwo, SDValue &NewMulConst) const;
359 
360   /// Replace N with M in CurDAG, in a way that also ensures that M gets
361   /// selected when N would have been selected.
362   void replaceDAGValue(const SDValue &N, SDValue M);
363 };
364 }
365 
366 char ARMDAGToDAGISel::ID = 0;
367 
368 INITIALIZE_PASS(ARMDAGToDAGISel, DEBUG_TYPE, PASS_NAME, false, false)
369 
370 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
371 /// operand. If so Imm will receive the 32-bit value.
372 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
373   if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
374     Imm = cast<ConstantSDNode>(N)->getZExtValue();
375     return true;
376   }
377   return false;
378 }
379 
380 // isInt32Immediate - This method tests to see if a constant operand.
381 // If so Imm will receive the 32 bit value.
382 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
383   return isInt32Immediate(N.getNode(), Imm);
384 }
385 
386 // isOpcWithIntImmediate - This method tests to see if the node is a specific
387 // opcode and that it has a immediate integer right operand.
388 // If so Imm will receive the 32 bit value.
389 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
390   return N->getOpcode() == Opc &&
391          isInt32Immediate(N->getOperand(1).getNode(), Imm);
392 }
393 
394 /// Check whether a particular node is a constant value representable as
395 /// (N * Scale) where (N in [\p RangeMin, \p RangeMax).
396 ///
397 /// \param ScaledConstant [out] - On success, the pre-scaled constant value.
398 static bool isScaledConstantInRange(SDValue Node, int Scale,
399                                     int RangeMin, int RangeMax,
400                                     int &ScaledConstant) {
401   assert(Scale > 0 && "Invalid scale!");
402 
403   // Check that this is a constant.
404   const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Node);
405   if (!C)
406     return false;
407 
408   ScaledConstant = (int) C->getZExtValue();
409   if ((ScaledConstant % Scale) != 0)
410     return false;
411 
412   ScaledConstant /= Scale;
413   return ScaledConstant >= RangeMin && ScaledConstant < RangeMax;
414 }
415 
416 void ARMDAGToDAGISel::PreprocessISelDAG() {
417   if (!Subtarget->hasV6T2Ops())
418     return;
419 
420   bool isThumb2 = Subtarget->isThumb();
421   // We use make_early_inc_range to avoid invalidation issues.
422   for (SDNode &N : llvm::make_early_inc_range(CurDAG->allnodes())) {
423     if (N.getOpcode() != ISD::ADD)
424       continue;
425 
426     // Look for (add X1, (and (srl X2, c1), c2)) where c2 is constant with
427     // leading zeros, followed by consecutive set bits, followed by 1 or 2
428     // trailing zeros, e.g. 1020.
429     // Transform the expression to
430     // (add X1, (shl (and (srl X2, c1), (c2>>tz)), tz)) where tz is the number
431     // of trailing zeros of c2. The left shift would be folded as an shifter
432     // operand of 'add' and the 'and' and 'srl' would become a bits extraction
433     // node (UBFX).
434 
435     SDValue N0 = N.getOperand(0);
436     SDValue N1 = N.getOperand(1);
437     unsigned And_imm = 0;
438     if (!isOpcWithIntImmediate(N1.getNode(), ISD::AND, And_imm)) {
439       if (isOpcWithIntImmediate(N0.getNode(), ISD::AND, And_imm))
440         std::swap(N0, N1);
441     }
442     if (!And_imm)
443       continue;
444 
445     // Check if the AND mask is an immediate of the form: 000.....1111111100
446     unsigned TZ = countTrailingZeros(And_imm);
447     if (TZ != 1 && TZ != 2)
448       // Be conservative here. Shifter operands aren't always free. e.g. On
449       // Swift, left shifter operand of 1 / 2 for free but others are not.
450       // e.g.
451       //  ubfx   r3, r1, #16, #8
452       //  ldr.w  r3, [r0, r3, lsl #2]
453       // vs.
454       //  mov.w  r9, #1020
455       //  and.w  r2, r9, r1, lsr #14
456       //  ldr    r2, [r0, r2]
457       continue;
458     And_imm >>= TZ;
459     if (And_imm & (And_imm + 1))
460       continue;
461 
462     // Look for (and (srl X, c1), c2).
463     SDValue Srl = N1.getOperand(0);
464     unsigned Srl_imm = 0;
465     if (!isOpcWithIntImmediate(Srl.getNode(), ISD::SRL, Srl_imm) ||
466         (Srl_imm <= 2))
467       continue;
468 
469     // Make sure first operand is not a shifter operand which would prevent
470     // folding of the left shift.
471     SDValue CPTmp0;
472     SDValue CPTmp1;
473     SDValue CPTmp2;
474     if (isThumb2) {
475       if (SelectImmShifterOperand(N0, CPTmp0, CPTmp1))
476         continue;
477     } else {
478       if (SelectImmShifterOperand(N0, CPTmp0, CPTmp1) ||
479           SelectRegShifterOperand(N0, CPTmp0, CPTmp1, CPTmp2))
480         continue;
481     }
482 
483     // Now make the transformation.
484     Srl = CurDAG->getNode(ISD::SRL, SDLoc(Srl), MVT::i32,
485                           Srl.getOperand(0),
486                           CurDAG->getConstant(Srl_imm + TZ, SDLoc(Srl),
487                                               MVT::i32));
488     N1 = CurDAG->getNode(ISD::AND, SDLoc(N1), MVT::i32,
489                          Srl,
490                          CurDAG->getConstant(And_imm, SDLoc(Srl), MVT::i32));
491     N1 = CurDAG->getNode(ISD::SHL, SDLoc(N1), MVT::i32,
492                          N1, CurDAG->getConstant(TZ, SDLoc(Srl), MVT::i32));
493     CurDAG->UpdateNodeOperands(&N, N0, N1);
494   }
495 }
496 
497 /// hasNoVMLxHazardUse - Return true if it's desirable to select a FP MLA / MLS
498 /// node. VFP / NEON fp VMLA / VMLS instructions have special RAW hazards (at
499 /// least on current ARM implementations) which should be avoidded.
500 bool ARMDAGToDAGISel::hasNoVMLxHazardUse(SDNode *N) const {
501   if (OptLevel == CodeGenOpt::None)
502     return true;
503 
504   if (!Subtarget->hasVMLxHazards())
505     return true;
506 
507   if (!N->hasOneUse())
508     return false;
509 
510   SDNode *Use = *N->use_begin();
511   if (Use->getOpcode() == ISD::CopyToReg)
512     return true;
513   if (Use->isMachineOpcode()) {
514     const ARMBaseInstrInfo *TII = static_cast<const ARMBaseInstrInfo *>(
515         CurDAG->getSubtarget().getInstrInfo());
516 
517     const MCInstrDesc &MCID = TII->get(Use->getMachineOpcode());
518     if (MCID.mayStore())
519       return true;
520     unsigned Opcode = MCID.getOpcode();
521     if (Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD)
522       return true;
523     // vmlx feeding into another vmlx. We actually want to unfold
524     // the use later in the MLxExpansion pass. e.g.
525     // vmla
526     // vmla (stall 8 cycles)
527     //
528     // vmul (5 cycles)
529     // vadd (5 cycles)
530     // vmla
531     // This adds up to about 18 - 19 cycles.
532     //
533     // vmla
534     // vmul (stall 4 cycles)
535     // vadd adds up to about 14 cycles.
536     return TII->isFpMLxInstruction(Opcode);
537   }
538 
539   return false;
540 }
541 
542 bool ARMDAGToDAGISel::isShifterOpProfitable(const SDValue &Shift,
543                                             ARM_AM::ShiftOpc ShOpcVal,
544                                             unsigned ShAmt) {
545   if (!Subtarget->isLikeA9() && !Subtarget->isSwift())
546     return true;
547   if (Shift.hasOneUse())
548     return true;
549   // R << 2 is free.
550   return ShOpcVal == ARM_AM::lsl &&
551          (ShAmt == 2 || (Subtarget->isSwift() && ShAmt == 1));
552 }
553 
554 bool ARMDAGToDAGISel::canExtractShiftFromMul(const SDValue &N,
555                                              unsigned MaxShift,
556                                              unsigned &PowerOfTwo,
557                                              SDValue &NewMulConst) const {
558   assert(N.getOpcode() == ISD::MUL);
559   assert(MaxShift > 0);
560 
561   // If the multiply is used in more than one place then changing the constant
562   // will make other uses incorrect, so don't.
563   if (!N.hasOneUse()) return false;
564   // Check if the multiply is by a constant
565   ConstantSDNode *MulConst = dyn_cast<ConstantSDNode>(N.getOperand(1));
566   if (!MulConst) return false;
567   // If the constant is used in more than one place then modifying it will mean
568   // we need to materialize two constants instead of one, which is a bad idea.
569   if (!MulConst->hasOneUse()) return false;
570   unsigned MulConstVal = MulConst->getZExtValue();
571   if (MulConstVal == 0) return false;
572 
573   // Find the largest power of 2 that MulConstVal is a multiple of
574   PowerOfTwo = MaxShift;
575   while ((MulConstVal % (1 << PowerOfTwo)) != 0) {
576     --PowerOfTwo;
577     if (PowerOfTwo == 0) return false;
578   }
579 
580   // Only optimise if the new cost is better
581   unsigned NewMulConstVal = MulConstVal / (1 << PowerOfTwo);
582   NewMulConst = CurDAG->getConstant(NewMulConstVal, SDLoc(N), MVT::i32);
583   unsigned OldCost = ConstantMaterializationCost(MulConstVal, Subtarget);
584   unsigned NewCost = ConstantMaterializationCost(NewMulConstVal, Subtarget);
585   return NewCost < OldCost;
586 }
587 
588 void ARMDAGToDAGISel::replaceDAGValue(const SDValue &N, SDValue M) {
589   CurDAG->RepositionNode(N.getNode()->getIterator(), M.getNode());
590   ReplaceUses(N, M);
591 }
592 
593 bool ARMDAGToDAGISel::SelectImmShifterOperand(SDValue N,
594                                               SDValue &BaseReg,
595                                               SDValue &Opc,
596                                               bool CheckProfitability) {
597   if (DisableShifterOp)
598     return false;
599 
600   // If N is a multiply-by-constant and it's profitable to extract a shift and
601   // use it in a shifted operand do so.
602   if (N.getOpcode() == ISD::MUL) {
603     unsigned PowerOfTwo = 0;
604     SDValue NewMulConst;
605     if (canExtractShiftFromMul(N, 31, PowerOfTwo, NewMulConst)) {
606       HandleSDNode Handle(N);
607       SDLoc Loc(N);
608       replaceDAGValue(N.getOperand(1), NewMulConst);
609       BaseReg = Handle.getValue();
610       Opc = CurDAG->getTargetConstant(
611           ARM_AM::getSORegOpc(ARM_AM::lsl, PowerOfTwo), Loc, MVT::i32);
612       return true;
613     }
614   }
615 
616   ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
617 
618   // Don't match base register only case. That is matched to a separate
619   // lower complexity pattern with explicit register operand.
620   if (ShOpcVal == ARM_AM::no_shift) return false;
621 
622   BaseReg = N.getOperand(0);
623   unsigned ShImmVal = 0;
624   ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1));
625   if (!RHS) return false;
626   ShImmVal = RHS->getZExtValue() & 31;
627   Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
628                                   SDLoc(N), MVT::i32);
629   return true;
630 }
631 
632 bool ARMDAGToDAGISel::SelectRegShifterOperand(SDValue N,
633                                               SDValue &BaseReg,
634                                               SDValue &ShReg,
635                                               SDValue &Opc,
636                                               bool CheckProfitability) {
637   if (DisableShifterOp)
638     return false;
639 
640   ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
641 
642   // Don't match base register only case. That is matched to a separate
643   // lower complexity pattern with explicit register operand.
644   if (ShOpcVal == ARM_AM::no_shift) return false;
645 
646   BaseReg = N.getOperand(0);
647   unsigned ShImmVal = 0;
648   ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1));
649   if (RHS) return false;
650 
651   ShReg = N.getOperand(1);
652   if (CheckProfitability && !isShifterOpProfitable(N, ShOpcVal, ShImmVal))
653     return false;
654   Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
655                                   SDLoc(N), MVT::i32);
656   return true;
657 }
658 
659 // Determine whether an ISD::OR's operands are suitable to turn the operation
660 // into an addition, which often has more compact encodings.
661 bool ARMDAGToDAGISel::SelectAddLikeOr(SDNode *Parent, SDValue N, SDValue &Out) {
662   assert(Parent->getOpcode() == ISD::OR && "unexpected parent");
663   Out = N;
664   return CurDAG->haveNoCommonBitsSet(N, Parent->getOperand(1));
665 }
666 
667 
668 bool ARMDAGToDAGISel::SelectAddrModeImm12(SDValue N,
669                                           SDValue &Base,
670                                           SDValue &OffImm) {
671   // Match simple R + imm12 operands.
672 
673   // Base only.
674   if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
675       !CurDAG->isBaseWithConstantOffset(N)) {
676     if (N.getOpcode() == ISD::FrameIndex) {
677       // Match frame index.
678       int FI = cast<FrameIndexSDNode>(N)->getIndex();
679       Base = CurDAG->getTargetFrameIndex(
680           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
681       OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
682       return true;
683     }
684 
685     if (N.getOpcode() == ARMISD::Wrapper &&
686         N.getOperand(0).getOpcode() != ISD::TargetGlobalAddress &&
687         N.getOperand(0).getOpcode() != ISD::TargetExternalSymbol &&
688         N.getOperand(0).getOpcode() != ISD::TargetGlobalTLSAddress) {
689       Base = N.getOperand(0);
690     } else
691       Base = N;
692     OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
693     return true;
694   }
695 
696   if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
697     int RHSC = (int)RHS->getSExtValue();
698     if (N.getOpcode() == ISD::SUB)
699       RHSC = -RHSC;
700 
701     if (RHSC > -0x1000 && RHSC < 0x1000) { // 12 bits
702       Base   = N.getOperand(0);
703       if (Base.getOpcode() == ISD::FrameIndex) {
704         int FI = cast<FrameIndexSDNode>(Base)->getIndex();
705         Base = CurDAG->getTargetFrameIndex(
706             FI, TLI->getPointerTy(CurDAG->getDataLayout()));
707       }
708       OffImm = CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32);
709       return true;
710     }
711   }
712 
713   // Base only.
714   Base = N;
715   OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
716   return true;
717 }
718 
719 
720 
721 bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset,
722                                       SDValue &Opc) {
723   if (N.getOpcode() == ISD::MUL &&
724       ((!Subtarget->isLikeA9() && !Subtarget->isSwift()) || N.hasOneUse())) {
725     if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
726       // X * [3,5,9] -> X + X * [2,4,8] etc.
727       int RHSC = (int)RHS->getZExtValue();
728       if (RHSC & 1) {
729         RHSC = RHSC & ~1;
730         ARM_AM::AddrOpc AddSub = ARM_AM::add;
731         if (RHSC < 0) {
732           AddSub = ARM_AM::sub;
733           RHSC = - RHSC;
734         }
735         if (isPowerOf2_32(RHSC)) {
736           unsigned ShAmt = Log2_32(RHSC);
737           Base = Offset = N.getOperand(0);
738           Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
739                                                             ARM_AM::lsl),
740                                           SDLoc(N), MVT::i32);
741           return true;
742         }
743       }
744     }
745   }
746 
747   if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
748       // ISD::OR that is equivalent to an ISD::ADD.
749       !CurDAG->isBaseWithConstantOffset(N))
750     return false;
751 
752   // Leave simple R +/- imm12 operands for LDRi12
753   if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::OR) {
754     int RHSC;
755     if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1,
756                                 -0x1000+1, 0x1000, RHSC)) // 12 bits.
757       return false;
758   }
759 
760   // Otherwise this is R +/- [possibly shifted] R.
761   ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::SUB ? ARM_AM::sub:ARM_AM::add;
762   ARM_AM::ShiftOpc ShOpcVal =
763     ARM_AM::getShiftOpcForNode(N.getOperand(1).getOpcode());
764   unsigned ShAmt = 0;
765 
766   Base   = N.getOperand(0);
767   Offset = N.getOperand(1);
768 
769   if (ShOpcVal != ARM_AM::no_shift) {
770     // Check to see if the RHS of the shift is a constant, if not, we can't fold
771     // it.
772     if (ConstantSDNode *Sh =
773            dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
774       ShAmt = Sh->getZExtValue();
775       if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt))
776         Offset = N.getOperand(1).getOperand(0);
777       else {
778         ShAmt = 0;
779         ShOpcVal = ARM_AM::no_shift;
780       }
781     } else {
782       ShOpcVal = ARM_AM::no_shift;
783     }
784   }
785 
786   // Try matching (R shl C) + (R).
787   if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift &&
788       !(Subtarget->isLikeA9() || Subtarget->isSwift() ||
789         N.getOperand(0).hasOneUse())) {
790     ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0).getOpcode());
791     if (ShOpcVal != ARM_AM::no_shift) {
792       // Check to see if the RHS of the shift is a constant, if not, we can't
793       // fold it.
794       if (ConstantSDNode *Sh =
795           dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
796         ShAmt = Sh->getZExtValue();
797         if (isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt)) {
798           Offset = N.getOperand(0).getOperand(0);
799           Base = N.getOperand(1);
800         } else {
801           ShAmt = 0;
802           ShOpcVal = ARM_AM::no_shift;
803         }
804       } else {
805         ShOpcVal = ARM_AM::no_shift;
806       }
807     }
808   }
809 
810   // If Offset is a multiply-by-constant and it's profitable to extract a shift
811   // and use it in a shifted operand do so.
812   if (Offset.getOpcode() == ISD::MUL && N.hasOneUse()) {
813     unsigned PowerOfTwo = 0;
814     SDValue NewMulConst;
815     if (canExtractShiftFromMul(Offset, 31, PowerOfTwo, NewMulConst)) {
816       HandleSDNode Handle(Offset);
817       replaceDAGValue(Offset.getOperand(1), NewMulConst);
818       Offset = Handle.getValue();
819       ShAmt = PowerOfTwo;
820       ShOpcVal = ARM_AM::lsl;
821     }
822   }
823 
824   Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
825                                   SDLoc(N), MVT::i32);
826   return true;
827 }
828 
829 bool ARMDAGToDAGISel::SelectAddrMode2OffsetReg(SDNode *Op, SDValue N,
830                                             SDValue &Offset, SDValue &Opc) {
831   unsigned Opcode = Op->getOpcode();
832   ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
833     ? cast<LoadSDNode>(Op)->getAddressingMode()
834     : cast<StoreSDNode>(Op)->getAddressingMode();
835   ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
836     ? ARM_AM::add : ARM_AM::sub;
837   int Val;
838   if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val))
839     return false;
840 
841   Offset = N;
842   ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
843   unsigned ShAmt = 0;
844   if (ShOpcVal != ARM_AM::no_shift) {
845     // Check to see if the RHS of the shift is a constant, if not, we can't fold
846     // it.
847     if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
848       ShAmt = Sh->getZExtValue();
849       if (isShifterOpProfitable(N, ShOpcVal, ShAmt))
850         Offset = N.getOperand(0);
851       else {
852         ShAmt = 0;
853         ShOpcVal = ARM_AM::no_shift;
854       }
855     } else {
856       ShOpcVal = ARM_AM::no_shift;
857     }
858   }
859 
860   Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
861                                   SDLoc(N), MVT::i32);
862   return true;
863 }
864 
865 bool ARMDAGToDAGISel::SelectAddrMode2OffsetImmPre(SDNode *Op, SDValue N,
866                                             SDValue &Offset, SDValue &Opc) {
867   unsigned Opcode = Op->getOpcode();
868   ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
869     ? cast<LoadSDNode>(Op)->getAddressingMode()
870     : cast<StoreSDNode>(Op)->getAddressingMode();
871   ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
872     ? ARM_AM::add : ARM_AM::sub;
873   int Val;
874   if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val)) { // 12 bits.
875     if (AddSub == ARM_AM::sub) Val *= -1;
876     Offset = CurDAG->getRegister(0, MVT::i32);
877     Opc = CurDAG->getTargetConstant(Val, SDLoc(Op), MVT::i32);
878     return true;
879   }
880 
881   return false;
882 }
883 
884 
885 bool ARMDAGToDAGISel::SelectAddrMode2OffsetImm(SDNode *Op, SDValue N,
886                                             SDValue &Offset, SDValue &Opc) {
887   unsigned Opcode = Op->getOpcode();
888   ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
889     ? cast<LoadSDNode>(Op)->getAddressingMode()
890     : cast<StoreSDNode>(Op)->getAddressingMode();
891   ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
892     ? ARM_AM::add : ARM_AM::sub;
893   int Val;
894   if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val)) { // 12 bits.
895     Offset = CurDAG->getRegister(0, MVT::i32);
896     Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
897                                                       ARM_AM::no_shift),
898                                     SDLoc(Op), MVT::i32);
899     return true;
900   }
901 
902   return false;
903 }
904 
905 bool ARMDAGToDAGISel::SelectAddrOffsetNone(SDValue N, SDValue &Base) {
906   Base = N;
907   return true;
908 }
909 
910 bool ARMDAGToDAGISel::SelectAddrMode3(SDValue N,
911                                       SDValue &Base, SDValue &Offset,
912                                       SDValue &Opc) {
913   if (N.getOpcode() == ISD::SUB) {
914     // X - C  is canonicalize to X + -C, no need to handle it here.
915     Base = N.getOperand(0);
916     Offset = N.getOperand(1);
917     Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0), SDLoc(N),
918                                     MVT::i32);
919     return true;
920   }
921 
922   if (!CurDAG->isBaseWithConstantOffset(N)) {
923     Base = N;
924     if (N.getOpcode() == ISD::FrameIndex) {
925       int FI = cast<FrameIndexSDNode>(N)->getIndex();
926       Base = CurDAG->getTargetFrameIndex(
927           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
928     }
929     Offset = CurDAG->getRegister(0, MVT::i32);
930     Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), SDLoc(N),
931                                     MVT::i32);
932     return true;
933   }
934 
935   // If the RHS is +/- imm8, fold into addr mode.
936   int RHSC;
937   if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1,
938                               -256 + 1, 256, RHSC)) { // 8 bits.
939     Base = N.getOperand(0);
940     if (Base.getOpcode() == ISD::FrameIndex) {
941       int FI = cast<FrameIndexSDNode>(Base)->getIndex();
942       Base = CurDAG->getTargetFrameIndex(
943           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
944     }
945     Offset = CurDAG->getRegister(0, MVT::i32);
946 
947     ARM_AM::AddrOpc AddSub = ARM_AM::add;
948     if (RHSC < 0) {
949       AddSub = ARM_AM::sub;
950       RHSC = -RHSC;
951     }
952     Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC), SDLoc(N),
953                                     MVT::i32);
954     return true;
955   }
956 
957   Base = N.getOperand(0);
958   Offset = N.getOperand(1);
959   Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), SDLoc(N),
960                                   MVT::i32);
961   return true;
962 }
963 
964 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
965                                             SDValue &Offset, SDValue &Opc) {
966   unsigned Opcode = Op->getOpcode();
967   ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
968     ? cast<LoadSDNode>(Op)->getAddressingMode()
969     : cast<StoreSDNode>(Op)->getAddressingMode();
970   ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
971     ? ARM_AM::add : ARM_AM::sub;
972   int Val;
973   if (isScaledConstantInRange(N, /*Scale=*/1, 0, 256, Val)) { // 12 bits.
974     Offset = CurDAG->getRegister(0, MVT::i32);
975     Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), SDLoc(Op),
976                                     MVT::i32);
977     return true;
978   }
979 
980   Offset = N;
981   Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), SDLoc(Op),
982                                   MVT::i32);
983   return true;
984 }
985 
986 bool ARMDAGToDAGISel::IsAddressingMode5(SDValue N, SDValue &Base, SDValue &Offset,
987                                         bool FP16) {
988   if (!CurDAG->isBaseWithConstantOffset(N)) {
989     Base = N;
990     if (N.getOpcode() == ISD::FrameIndex) {
991       int FI = cast<FrameIndexSDNode>(N)->getIndex();
992       Base = CurDAG->getTargetFrameIndex(
993           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
994     } else if (N.getOpcode() == ARMISD::Wrapper &&
995                N.getOperand(0).getOpcode() != ISD::TargetGlobalAddress &&
996                N.getOperand(0).getOpcode() != ISD::TargetExternalSymbol &&
997                N.getOperand(0).getOpcode() != ISD::TargetGlobalTLSAddress) {
998       Base = N.getOperand(0);
999     }
1000     Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
1001                                        SDLoc(N), MVT::i32);
1002     return true;
1003   }
1004 
1005   // If the RHS is +/- imm8, fold into addr mode.
1006   int RHSC;
1007   const int Scale = FP16 ? 2 : 4;
1008 
1009   if (isScaledConstantInRange(N.getOperand(1), Scale, -255, 256, RHSC)) {
1010     Base = N.getOperand(0);
1011     if (Base.getOpcode() == ISD::FrameIndex) {
1012       int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1013       Base = CurDAG->getTargetFrameIndex(
1014           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1015     }
1016 
1017     ARM_AM::AddrOpc AddSub = ARM_AM::add;
1018     if (RHSC < 0) {
1019       AddSub = ARM_AM::sub;
1020       RHSC = -RHSC;
1021     }
1022 
1023     if (FP16)
1024       Offset = CurDAG->getTargetConstant(ARM_AM::getAM5FP16Opc(AddSub, RHSC),
1025                                          SDLoc(N), MVT::i32);
1026     else
1027       Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
1028                                          SDLoc(N), MVT::i32);
1029 
1030     return true;
1031   }
1032 
1033   Base = N;
1034 
1035   if (FP16)
1036     Offset = CurDAG->getTargetConstant(ARM_AM::getAM5FP16Opc(ARM_AM::add, 0),
1037                                        SDLoc(N), MVT::i32);
1038   else
1039     Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
1040                                        SDLoc(N), MVT::i32);
1041 
1042   return true;
1043 }
1044 
1045 bool ARMDAGToDAGISel::SelectAddrMode5(SDValue N,
1046                                       SDValue &Base, SDValue &Offset) {
1047   return IsAddressingMode5(N, Base, Offset, /*FP16=*/ false);
1048 }
1049 
1050 bool ARMDAGToDAGISel::SelectAddrMode5FP16(SDValue N,
1051                                           SDValue &Base, SDValue &Offset) {
1052   return IsAddressingMode5(N, Base, Offset, /*FP16=*/ true);
1053 }
1054 
1055 bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,
1056                                       SDValue &Align) {
1057   Addr = N;
1058 
1059   unsigned Alignment = 0;
1060 
1061   MemSDNode *MemN = cast<MemSDNode>(Parent);
1062 
1063   if (isa<LSBaseSDNode>(MemN) ||
1064       ((MemN->getOpcode() == ARMISD::VST1_UPD ||
1065         MemN->getOpcode() == ARMISD::VLD1_UPD) &&
1066        MemN->getConstantOperandVal(MemN->getNumOperands() - 1) == 1)) {
1067     // This case occurs only for VLD1-lane/dup and VST1-lane instructions.
1068     // The maximum alignment is equal to the memory size being referenced.
1069     llvm::Align MMOAlign = MemN->getAlign();
1070     unsigned MemSize = MemN->getMemoryVT().getSizeInBits() / 8;
1071     if (MMOAlign.value() >= MemSize && MemSize > 1)
1072       Alignment = MemSize;
1073   } else {
1074     // All other uses of addrmode6 are for intrinsics.  For now just record
1075     // the raw alignment value; it will be refined later based on the legal
1076     // alignment operands for the intrinsic.
1077     Alignment = MemN->getAlign().value();
1078   }
1079 
1080   Align = CurDAG->getTargetConstant(Alignment, SDLoc(N), MVT::i32);
1081   return true;
1082 }
1083 
1084 bool ARMDAGToDAGISel::SelectAddrMode6Offset(SDNode *Op, SDValue N,
1085                                             SDValue &Offset) {
1086   LSBaseSDNode *LdSt = cast<LSBaseSDNode>(Op);
1087   ISD::MemIndexedMode AM = LdSt->getAddressingMode();
1088   if (AM != ISD::POST_INC)
1089     return false;
1090   Offset = N;
1091   if (ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N)) {
1092     if (NC->getZExtValue() * 8 == LdSt->getMemoryVT().getSizeInBits())
1093       Offset = CurDAG->getRegister(0, MVT::i32);
1094   }
1095   return true;
1096 }
1097 
1098 bool ARMDAGToDAGISel::SelectAddrModePC(SDValue N,
1099                                        SDValue &Offset, SDValue &Label) {
1100   if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
1101     Offset = N.getOperand(0);
1102     SDValue N1 = N.getOperand(1);
1103     Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
1104                                       SDLoc(N), MVT::i32);
1105     return true;
1106   }
1107 
1108   return false;
1109 }
1110 
1111 
1112 //===----------------------------------------------------------------------===//
1113 //                         Thumb Addressing Modes
1114 //===----------------------------------------------------------------------===//
1115 
1116 static bool shouldUseZeroOffsetLdSt(SDValue N) {
1117   // Negative numbers are difficult to materialise in thumb1. If we are
1118   // selecting the add of a negative, instead try to select ri with a zero
1119   // offset, so create the add node directly which will become a sub.
1120   if (N.getOpcode() != ISD::ADD)
1121     return false;
1122 
1123   // Look for an imm which is not legal for ld/st, but is legal for sub.
1124   if (auto C = dyn_cast<ConstantSDNode>(N.getOperand(1)))
1125     return C->getSExtValue() < 0 && C->getSExtValue() >= -255;
1126 
1127   return false;
1128 }
1129 
1130 bool ARMDAGToDAGISel::SelectThumbAddrModeRRSext(SDValue N, SDValue &Base,
1131                                                 SDValue &Offset) {
1132   if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(N)) {
1133     ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
1134     if (!NC || !NC->isZero())
1135       return false;
1136 
1137     Base = Offset = N;
1138     return true;
1139   }
1140 
1141   Base = N.getOperand(0);
1142   Offset = N.getOperand(1);
1143   return true;
1144 }
1145 
1146 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue N, SDValue &Base,
1147                                             SDValue &Offset) {
1148   if (shouldUseZeroOffsetLdSt(N))
1149     return false; // Select ri instead
1150   return SelectThumbAddrModeRRSext(N, Base, Offset);
1151 }
1152 
1153 bool
1154 ARMDAGToDAGISel::SelectThumbAddrModeImm5S(SDValue N, unsigned Scale,
1155                                           SDValue &Base, SDValue &OffImm) {
1156   if (shouldUseZeroOffsetLdSt(N)) {
1157     Base = N;
1158     OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1159     return true;
1160   }
1161 
1162   if (!CurDAG->isBaseWithConstantOffset(N)) {
1163     if (N.getOpcode() == ISD::ADD) {
1164       return false; // We want to select register offset instead
1165     } else if (N.getOpcode() == ARMISD::Wrapper &&
1166         N.getOperand(0).getOpcode() != ISD::TargetGlobalAddress &&
1167         N.getOperand(0).getOpcode() != ISD::TargetExternalSymbol &&
1168         N.getOperand(0).getOpcode() != ISD::TargetConstantPool &&
1169         N.getOperand(0).getOpcode() != ISD::TargetGlobalTLSAddress) {
1170       Base = N.getOperand(0);
1171     } else {
1172       Base = N;
1173     }
1174 
1175     OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1176     return true;
1177   }
1178 
1179   // If the RHS is + imm5 * scale, fold into addr mode.
1180   int RHSC;
1181   if (isScaledConstantInRange(N.getOperand(1), Scale, 0, 32, RHSC)) {
1182     Base = N.getOperand(0);
1183     OffImm = CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32);
1184     return true;
1185   }
1186 
1187   // Offset is too large, so use register offset instead.
1188   return false;
1189 }
1190 
1191 bool
1192 ARMDAGToDAGISel::SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base,
1193                                            SDValue &OffImm) {
1194   return SelectThumbAddrModeImm5S(N, 4, Base, OffImm);
1195 }
1196 
1197 bool
1198 ARMDAGToDAGISel::SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base,
1199                                            SDValue &OffImm) {
1200   return SelectThumbAddrModeImm5S(N, 2, Base, OffImm);
1201 }
1202 
1203 bool
1204 ARMDAGToDAGISel::SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base,
1205                                            SDValue &OffImm) {
1206   return SelectThumbAddrModeImm5S(N, 1, Base, OffImm);
1207 }
1208 
1209 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDValue N,
1210                                             SDValue &Base, SDValue &OffImm) {
1211   if (N.getOpcode() == ISD::FrameIndex) {
1212     int FI = cast<FrameIndexSDNode>(N)->getIndex();
1213     // Only multiples of 4 are allowed for the offset, so the frame object
1214     // alignment must be at least 4.
1215     MachineFrameInfo &MFI = MF->getFrameInfo();
1216     if (MFI.getObjectAlign(FI) < Align(4))
1217       MFI.setObjectAlignment(FI, Align(4));
1218     Base = CurDAG->getTargetFrameIndex(
1219         FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1220     OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1221     return true;
1222   }
1223 
1224   if (!CurDAG->isBaseWithConstantOffset(N))
1225     return false;
1226 
1227   if (N.getOperand(0).getOpcode() == ISD::FrameIndex) {
1228     // If the RHS is + imm8 * scale, fold into addr mode.
1229     int RHSC;
1230     if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/4, 0, 256, RHSC)) {
1231       Base = N.getOperand(0);
1232       int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1233       // Make sure the offset is inside the object, or we might fail to
1234       // allocate an emergency spill slot. (An out-of-range access is UB, but
1235       // it could show up anyway.)
1236       MachineFrameInfo &MFI = MF->getFrameInfo();
1237       if (RHSC * 4 < MFI.getObjectSize(FI)) {
1238         // For LHS+RHS to result in an offset that's a multiple of 4 the object
1239         // indexed by the LHS must be 4-byte aligned.
1240         if (!MFI.isFixedObjectIndex(FI) && MFI.getObjectAlign(FI) < Align(4))
1241           MFI.setObjectAlignment(FI, Align(4));
1242         if (MFI.getObjectAlign(FI) >= Align(4)) {
1243           Base = CurDAG->getTargetFrameIndex(
1244               FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1245           OffImm = CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32);
1246           return true;
1247         }
1248       }
1249     }
1250   }
1251 
1252   return false;
1253 }
1254 
1255 template <unsigned Shift>
1256 bool ARMDAGToDAGISel::SelectTAddrModeImm7(SDValue N, SDValue &Base,
1257                                           SDValue &OffImm) {
1258   if (N.getOpcode() == ISD::SUB || CurDAG->isBaseWithConstantOffset(N)) {
1259     int RHSC;
1260     if (isScaledConstantInRange(N.getOperand(1), 1 << Shift, -0x7f, 0x80,
1261                                 RHSC)) {
1262       Base = N.getOperand(0);
1263       if (N.getOpcode() == ISD::SUB)
1264         RHSC = -RHSC;
1265       OffImm =
1266           CurDAG->getTargetConstant(RHSC * (1 << Shift), SDLoc(N), MVT::i32);
1267       return true;
1268     }
1269   }
1270 
1271   // Base only.
1272   Base = N;
1273   OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1274   return true;
1275 }
1276 
1277 
1278 //===----------------------------------------------------------------------===//
1279 //                        Thumb 2 Addressing Modes
1280 //===----------------------------------------------------------------------===//
1281 
1282 
1283 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue N,
1284                                             SDValue &Base, SDValue &OffImm) {
1285   // Match simple R + imm12 operands.
1286 
1287   // Base only.
1288   if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
1289       !CurDAG->isBaseWithConstantOffset(N)) {
1290     if (N.getOpcode() == ISD::FrameIndex) {
1291       // Match frame index.
1292       int FI = cast<FrameIndexSDNode>(N)->getIndex();
1293       Base = CurDAG->getTargetFrameIndex(
1294           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1295       OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1296       return true;
1297     }
1298 
1299     if (N.getOpcode() == ARMISD::Wrapper &&
1300         N.getOperand(0).getOpcode() != ISD::TargetGlobalAddress &&
1301         N.getOperand(0).getOpcode() != ISD::TargetExternalSymbol &&
1302         N.getOperand(0).getOpcode() != ISD::TargetGlobalTLSAddress) {
1303       Base = N.getOperand(0);
1304       if (Base.getOpcode() == ISD::TargetConstantPool)
1305         return false;  // We want to select t2LDRpci instead.
1306     } else
1307       Base = N;
1308     OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1309     return true;
1310   }
1311 
1312   if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1313     if (SelectT2AddrModeImm8(N, Base, OffImm))
1314       // Let t2LDRi8 handle (R - imm8).
1315       return false;
1316 
1317     int RHSC = (int)RHS->getZExtValue();
1318     if (N.getOpcode() == ISD::SUB)
1319       RHSC = -RHSC;
1320 
1321     if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
1322       Base   = N.getOperand(0);
1323       if (Base.getOpcode() == ISD::FrameIndex) {
1324         int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1325         Base = CurDAG->getTargetFrameIndex(
1326             FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1327       }
1328       OffImm = CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32);
1329       return true;
1330     }
1331   }
1332 
1333   // Base only.
1334   Base = N;
1335   OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1336   return true;
1337 }
1338 
1339 template <unsigned Shift>
1340 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue N, SDValue &Base,
1341                                            SDValue &OffImm) {
1342   if (N.getOpcode() == ISD::SUB || CurDAG->isBaseWithConstantOffset(N)) {
1343     int RHSC;
1344     if (isScaledConstantInRange(N.getOperand(1), 1 << Shift, -255, 256, RHSC)) {
1345       Base = N.getOperand(0);
1346       if (Base.getOpcode() == ISD::FrameIndex) {
1347         int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1348         Base = CurDAG->getTargetFrameIndex(
1349             FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1350       }
1351 
1352       if (N.getOpcode() == ISD::SUB)
1353         RHSC = -RHSC;
1354       OffImm =
1355           CurDAG->getTargetConstant(RHSC * (1 << Shift), SDLoc(N), MVT::i32);
1356       return true;
1357     }
1358   }
1359 
1360   // Base only.
1361   Base = N;
1362   OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1363   return true;
1364 }
1365 
1366 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue N,
1367                                            SDValue &Base, SDValue &OffImm) {
1368   // Match simple R - imm8 operands.
1369   if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
1370       !CurDAG->isBaseWithConstantOffset(N))
1371     return false;
1372 
1373   if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1374     int RHSC = (int)RHS->getSExtValue();
1375     if (N.getOpcode() == ISD::SUB)
1376       RHSC = -RHSC;
1377 
1378     if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
1379       Base = N.getOperand(0);
1380       if (Base.getOpcode() == ISD::FrameIndex) {
1381         int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1382         Base = CurDAG->getTargetFrameIndex(
1383             FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1384       }
1385       OffImm = CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32);
1386       return true;
1387     }
1388   }
1389 
1390   return false;
1391 }
1392 
1393 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
1394                                                  SDValue &OffImm){
1395   unsigned Opcode = Op->getOpcode();
1396   ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
1397     ? cast<LoadSDNode>(Op)->getAddressingMode()
1398     : cast<StoreSDNode>(Op)->getAddressingMode();
1399   int RHSC;
1400   if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x100, RHSC)) { // 8 bits.
1401     OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
1402       ? CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32)
1403       : CurDAG->getTargetConstant(-RHSC, SDLoc(N), MVT::i32);
1404     return true;
1405   }
1406 
1407   return false;
1408 }
1409 
1410 template <unsigned Shift>
1411 bool ARMDAGToDAGISel::SelectT2AddrModeImm7(SDValue N, SDValue &Base,
1412                                            SDValue &OffImm) {
1413   if (N.getOpcode() == ISD::SUB || CurDAG->isBaseWithConstantOffset(N)) {
1414     int RHSC;
1415     if (isScaledConstantInRange(N.getOperand(1), 1 << Shift, -0x7f, 0x80,
1416                                 RHSC)) {
1417       Base = N.getOperand(0);
1418       if (Base.getOpcode() == ISD::FrameIndex) {
1419         int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1420         Base = CurDAG->getTargetFrameIndex(
1421             FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1422       }
1423 
1424       if (N.getOpcode() == ISD::SUB)
1425         RHSC = -RHSC;
1426       OffImm =
1427           CurDAG->getTargetConstant(RHSC * (1 << Shift), SDLoc(N), MVT::i32);
1428       return true;
1429     }
1430   }
1431 
1432   // Base only.
1433   Base = N;
1434   OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1435   return true;
1436 }
1437 
1438 template <unsigned Shift>
1439 bool ARMDAGToDAGISel::SelectT2AddrModeImm7Offset(SDNode *Op, SDValue N,
1440                                                  SDValue &OffImm) {
1441   return SelectT2AddrModeImm7Offset(Op, N, OffImm, Shift);
1442 }
1443 
1444 bool ARMDAGToDAGISel::SelectT2AddrModeImm7Offset(SDNode *Op, SDValue N,
1445                                                  SDValue &OffImm,
1446                                                  unsigned Shift) {
1447   unsigned Opcode = Op->getOpcode();
1448   ISD::MemIndexedMode AM;
1449   switch (Opcode) {
1450   case ISD::LOAD:
1451     AM = cast<LoadSDNode>(Op)->getAddressingMode();
1452     break;
1453   case ISD::STORE:
1454     AM = cast<StoreSDNode>(Op)->getAddressingMode();
1455     break;
1456   case ISD::MLOAD:
1457     AM = cast<MaskedLoadSDNode>(Op)->getAddressingMode();
1458     break;
1459   case ISD::MSTORE:
1460     AM = cast<MaskedStoreSDNode>(Op)->getAddressingMode();
1461     break;
1462   default:
1463     llvm_unreachable("Unexpected Opcode for Imm7Offset");
1464   }
1465 
1466   int RHSC;
1467   // 7 bit constant, shifted by Shift.
1468   if (isScaledConstantInRange(N, 1 << Shift, 0, 0x80, RHSC)) {
1469     OffImm =
1470         ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
1471             ? CurDAG->getTargetConstant(RHSC * (1 << Shift), SDLoc(N), MVT::i32)
1472             : CurDAG->getTargetConstant(-RHSC * (1 << Shift), SDLoc(N),
1473                                         MVT::i32);
1474     return true;
1475   }
1476   return false;
1477 }
1478 
1479 template <int Min, int Max>
1480 bool ARMDAGToDAGISel::SelectImmediateInRange(SDValue N, SDValue &OffImm) {
1481   int Val;
1482   if (isScaledConstantInRange(N, 1, Min, Max, Val)) {
1483     OffImm = CurDAG->getTargetConstant(Val, SDLoc(N), MVT::i32);
1484     return true;
1485   }
1486   return false;
1487 }
1488 
1489 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue N,
1490                                             SDValue &Base,
1491                                             SDValue &OffReg, SDValue &ShImm) {
1492   // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
1493   if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(N))
1494     return false;
1495 
1496   // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
1497   if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1498     int RHSC = (int)RHS->getZExtValue();
1499     if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
1500       return false;
1501     else if (RHSC < 0 && RHSC >= -255) // 8 bits
1502       return false;
1503   }
1504 
1505   // Look for (R + R) or (R + (R << [1,2,3])).
1506   unsigned ShAmt = 0;
1507   Base   = N.getOperand(0);
1508   OffReg = N.getOperand(1);
1509 
1510   // Swap if it is ((R << c) + R).
1511   ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg.getOpcode());
1512   if (ShOpcVal != ARM_AM::lsl) {
1513     ShOpcVal = ARM_AM::getShiftOpcForNode(Base.getOpcode());
1514     if (ShOpcVal == ARM_AM::lsl)
1515       std::swap(Base, OffReg);
1516   }
1517 
1518   if (ShOpcVal == ARM_AM::lsl) {
1519     // Check to see if the RHS of the shift is a constant, if not, we can't fold
1520     // it.
1521     if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
1522       ShAmt = Sh->getZExtValue();
1523       if (ShAmt < 4 && isShifterOpProfitable(OffReg, ShOpcVal, ShAmt))
1524         OffReg = OffReg.getOperand(0);
1525       else {
1526         ShAmt = 0;
1527       }
1528     }
1529   }
1530 
1531   // If OffReg is a multiply-by-constant and it's profitable to extract a shift
1532   // and use it in a shifted operand do so.
1533   if (OffReg.getOpcode() == ISD::MUL && N.hasOneUse()) {
1534     unsigned PowerOfTwo = 0;
1535     SDValue NewMulConst;
1536     if (canExtractShiftFromMul(OffReg, 3, PowerOfTwo, NewMulConst)) {
1537       HandleSDNode Handle(OffReg);
1538       replaceDAGValue(OffReg.getOperand(1), NewMulConst);
1539       OffReg = Handle.getValue();
1540       ShAmt = PowerOfTwo;
1541     }
1542   }
1543 
1544   ShImm = CurDAG->getTargetConstant(ShAmt, SDLoc(N), MVT::i32);
1545 
1546   return true;
1547 }
1548 
1549 bool ARMDAGToDAGISel::SelectT2AddrModeExclusive(SDValue N, SDValue &Base,
1550                                                 SDValue &OffImm) {
1551   // This *must* succeed since it's used for the irreplaceable ldrex and strex
1552   // instructions.
1553   Base = N;
1554   OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1555 
1556   if (N.getOpcode() != ISD::ADD || !CurDAG->isBaseWithConstantOffset(N))
1557     return true;
1558 
1559   ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1));
1560   if (!RHS)
1561     return true;
1562 
1563   uint32_t RHSC = (int)RHS->getZExtValue();
1564   if (RHSC > 1020 || RHSC % 4 != 0)
1565     return true;
1566 
1567   Base = N.getOperand(0);
1568   if (Base.getOpcode() == ISD::FrameIndex) {
1569     int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1570     Base = CurDAG->getTargetFrameIndex(
1571         FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1572   }
1573 
1574   OffImm = CurDAG->getTargetConstant(RHSC/4, SDLoc(N), MVT::i32);
1575   return true;
1576 }
1577 
1578 //===--------------------------------------------------------------------===//
1579 
1580 /// getAL - Returns a ARMCC::AL immediate node.
1581 static inline SDValue getAL(SelectionDAG *CurDAG, const SDLoc &dl) {
1582   return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, dl, MVT::i32);
1583 }
1584 
1585 void ARMDAGToDAGISel::transferMemOperands(SDNode *N, SDNode *Result) {
1586   MachineMemOperand *MemOp = cast<MemSDNode>(N)->getMemOperand();
1587   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Result), {MemOp});
1588 }
1589 
1590 bool ARMDAGToDAGISel::tryARMIndexedLoad(SDNode *N) {
1591   LoadSDNode *LD = cast<LoadSDNode>(N);
1592   ISD::MemIndexedMode AM = LD->getAddressingMode();
1593   if (AM == ISD::UNINDEXED)
1594     return false;
1595 
1596   EVT LoadedVT = LD->getMemoryVT();
1597   SDValue Offset, AMOpc;
1598   bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1599   unsigned Opcode = 0;
1600   bool Match = false;
1601   if (LoadedVT == MVT::i32 && isPre &&
1602       SelectAddrMode2OffsetImmPre(N, LD->getOffset(), Offset, AMOpc)) {
1603     Opcode = ARM::LDR_PRE_IMM;
1604     Match = true;
1605   } else if (LoadedVT == MVT::i32 && !isPre &&
1606       SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) {
1607     Opcode = ARM::LDR_POST_IMM;
1608     Match = true;
1609   } else if (LoadedVT == MVT::i32 &&
1610       SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) {
1611     Opcode = isPre ? ARM::LDR_PRE_REG : ARM::LDR_POST_REG;
1612     Match = true;
1613 
1614   } else if (LoadedVT == MVT::i16 &&
1615              SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
1616     Match = true;
1617     Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
1618       ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
1619       : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
1620   } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
1621     if (LD->getExtensionType() == ISD::SEXTLOAD) {
1622       if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
1623         Match = true;
1624         Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
1625       }
1626     } else {
1627       if (isPre &&
1628           SelectAddrMode2OffsetImmPre(N, LD->getOffset(), Offset, AMOpc)) {
1629         Match = true;
1630         Opcode = ARM::LDRB_PRE_IMM;
1631       } else if (!isPre &&
1632                   SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) {
1633         Match = true;
1634         Opcode = ARM::LDRB_POST_IMM;
1635       } else if (SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) {
1636         Match = true;
1637         Opcode = isPre ? ARM::LDRB_PRE_REG : ARM::LDRB_POST_REG;
1638       }
1639     }
1640   }
1641 
1642   if (Match) {
1643     if (Opcode == ARM::LDR_PRE_IMM || Opcode == ARM::LDRB_PRE_IMM) {
1644       SDValue Chain = LD->getChain();
1645       SDValue Base = LD->getBasePtr();
1646       SDValue Ops[]= { Base, AMOpc, getAL(CurDAG, SDLoc(N)),
1647                        CurDAG->getRegister(0, MVT::i32), Chain };
1648       SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, MVT::i32,
1649                                            MVT::Other, Ops);
1650       transferMemOperands(N, New);
1651       ReplaceNode(N, New);
1652       return true;
1653     } else {
1654       SDValue Chain = LD->getChain();
1655       SDValue Base = LD->getBasePtr();
1656       SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG, SDLoc(N)),
1657                        CurDAG->getRegister(0, MVT::i32), Chain };
1658       SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, MVT::i32,
1659                                            MVT::Other, Ops);
1660       transferMemOperands(N, New);
1661       ReplaceNode(N, New);
1662       return true;
1663     }
1664   }
1665 
1666   return false;
1667 }
1668 
1669 bool ARMDAGToDAGISel::tryT1IndexedLoad(SDNode *N) {
1670   LoadSDNode *LD = cast<LoadSDNode>(N);
1671   EVT LoadedVT = LD->getMemoryVT();
1672   ISD::MemIndexedMode AM = LD->getAddressingMode();
1673   if (AM != ISD::POST_INC || LD->getExtensionType() != ISD::NON_EXTLOAD ||
1674       LoadedVT.getSimpleVT().SimpleTy != MVT::i32)
1675     return false;
1676 
1677   auto *COffs = dyn_cast<ConstantSDNode>(LD->getOffset());
1678   if (!COffs || COffs->getZExtValue() != 4)
1679     return false;
1680 
1681   // A T1 post-indexed load is just a single register LDM: LDM r0!, {r1}.
1682   // The encoding of LDM is not how the rest of ISel expects a post-inc load to
1683   // look however, so we use a pseudo here and switch it for a tLDMIA_UPD after
1684   // ISel.
1685   SDValue Chain = LD->getChain();
1686   SDValue Base = LD->getBasePtr();
1687   SDValue Ops[]= { Base, getAL(CurDAG, SDLoc(N)),
1688                    CurDAG->getRegister(0, MVT::i32), Chain };
1689   SDNode *New = CurDAG->getMachineNode(ARM::tLDR_postidx, SDLoc(N), MVT::i32,
1690                                        MVT::i32, MVT::Other, Ops);
1691   transferMemOperands(N, New);
1692   ReplaceNode(N, New);
1693   return true;
1694 }
1695 
1696 bool ARMDAGToDAGISel::tryT2IndexedLoad(SDNode *N) {
1697   LoadSDNode *LD = cast<LoadSDNode>(N);
1698   ISD::MemIndexedMode AM = LD->getAddressingMode();
1699   if (AM == ISD::UNINDEXED)
1700     return false;
1701 
1702   EVT LoadedVT = LD->getMemoryVT();
1703   bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
1704   SDValue Offset;
1705   bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1706   unsigned Opcode = 0;
1707   bool Match = false;
1708   if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
1709     switch (LoadedVT.getSimpleVT().SimpleTy) {
1710     case MVT::i32:
1711       Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
1712       break;
1713     case MVT::i16:
1714       if (isSExtLd)
1715         Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
1716       else
1717         Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
1718       break;
1719     case MVT::i8:
1720     case MVT::i1:
1721       if (isSExtLd)
1722         Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
1723       else
1724         Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
1725       break;
1726     default:
1727       return false;
1728     }
1729     Match = true;
1730   }
1731 
1732   if (Match) {
1733     SDValue Chain = LD->getChain();
1734     SDValue Base = LD->getBasePtr();
1735     SDValue Ops[]= { Base, Offset, getAL(CurDAG, SDLoc(N)),
1736                      CurDAG->getRegister(0, MVT::i32), Chain };
1737     SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, MVT::i32,
1738                                          MVT::Other, Ops);
1739     transferMemOperands(N, New);
1740     ReplaceNode(N, New);
1741     return true;
1742   }
1743 
1744   return false;
1745 }
1746 
1747 bool ARMDAGToDAGISel::tryMVEIndexedLoad(SDNode *N) {
1748   EVT LoadedVT;
1749   unsigned Opcode = 0;
1750   bool isSExtLd, isPre;
1751   Align Alignment;
1752   ARMVCC::VPTCodes Pred;
1753   SDValue PredReg;
1754   SDValue Chain, Base, Offset;
1755 
1756   if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1757     ISD::MemIndexedMode AM = LD->getAddressingMode();
1758     if (AM == ISD::UNINDEXED)
1759       return false;
1760     LoadedVT = LD->getMemoryVT();
1761     if (!LoadedVT.isVector())
1762       return false;
1763 
1764     Chain = LD->getChain();
1765     Base = LD->getBasePtr();
1766     Offset = LD->getOffset();
1767     Alignment = LD->getAlign();
1768     isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
1769     isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1770     Pred = ARMVCC::None;
1771     PredReg = CurDAG->getRegister(0, MVT::i32);
1772   } else if (MaskedLoadSDNode *LD = dyn_cast<MaskedLoadSDNode>(N)) {
1773     ISD::MemIndexedMode AM = LD->getAddressingMode();
1774     if (AM == ISD::UNINDEXED)
1775       return false;
1776     LoadedVT = LD->getMemoryVT();
1777     if (!LoadedVT.isVector())
1778       return false;
1779 
1780     Chain = LD->getChain();
1781     Base = LD->getBasePtr();
1782     Offset = LD->getOffset();
1783     Alignment = LD->getAlign();
1784     isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
1785     isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1786     Pred = ARMVCC::Then;
1787     PredReg = LD->getMask();
1788   } else
1789     llvm_unreachable("Expected a Load or a Masked Load!");
1790 
1791   // We allow LE non-masked loads to change the type (for example use a vldrb.8
1792   // as opposed to a vldrw.32). This can allow extra addressing modes or
1793   // alignments for what is otherwise an equivalent instruction.
1794   bool CanChangeType = Subtarget->isLittle() && !isa<MaskedLoadSDNode>(N);
1795 
1796   SDValue NewOffset;
1797   if (Alignment >= Align(2) && LoadedVT == MVT::v4i16 &&
1798       SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 1)) {
1799     if (isSExtLd)
1800       Opcode = isPre ? ARM::MVE_VLDRHS32_pre : ARM::MVE_VLDRHS32_post;
1801     else
1802       Opcode = isPre ? ARM::MVE_VLDRHU32_pre : ARM::MVE_VLDRHU32_post;
1803   } else if (LoadedVT == MVT::v8i8 &&
1804              SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 0)) {
1805     if (isSExtLd)
1806       Opcode = isPre ? ARM::MVE_VLDRBS16_pre : ARM::MVE_VLDRBS16_post;
1807     else
1808       Opcode = isPre ? ARM::MVE_VLDRBU16_pre : ARM::MVE_VLDRBU16_post;
1809   } else if (LoadedVT == MVT::v4i8 &&
1810              SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 0)) {
1811     if (isSExtLd)
1812       Opcode = isPre ? ARM::MVE_VLDRBS32_pre : ARM::MVE_VLDRBS32_post;
1813     else
1814       Opcode = isPre ? ARM::MVE_VLDRBU32_pre : ARM::MVE_VLDRBU32_post;
1815   } else if (Alignment >= Align(4) &&
1816              (CanChangeType || LoadedVT == MVT::v4i32 ||
1817               LoadedVT == MVT::v4f32) &&
1818              SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 2))
1819     Opcode = isPre ? ARM::MVE_VLDRWU32_pre : ARM::MVE_VLDRWU32_post;
1820   else if (Alignment >= Align(2) &&
1821            (CanChangeType || LoadedVT == MVT::v8i16 ||
1822             LoadedVT == MVT::v8f16) &&
1823            SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 1))
1824     Opcode = isPre ? ARM::MVE_VLDRHU16_pre : ARM::MVE_VLDRHU16_post;
1825   else if ((CanChangeType || LoadedVT == MVT::v16i8) &&
1826            SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 0))
1827     Opcode = isPre ? ARM::MVE_VLDRBU8_pre : ARM::MVE_VLDRBU8_post;
1828   else
1829     return false;
1830 
1831   SDValue Ops[] = {Base,
1832                    NewOffset,
1833                    CurDAG->getTargetConstant(Pred, SDLoc(N), MVT::i32),
1834                    PredReg,
1835                    CurDAG->getRegister(0, MVT::i32), // tp_reg
1836                    Chain};
1837   SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32,
1838                                        N->getValueType(0), MVT::Other, Ops);
1839   transferMemOperands(N, New);
1840   ReplaceUses(SDValue(N, 0), SDValue(New, 1));
1841   ReplaceUses(SDValue(N, 1), SDValue(New, 0));
1842   ReplaceUses(SDValue(N, 2), SDValue(New, 2));
1843   CurDAG->RemoveDeadNode(N);
1844   return true;
1845 }
1846 
1847 /// Form a GPRPair pseudo register from a pair of GPR regs.
1848 SDNode *ARMDAGToDAGISel::createGPRPairNode(EVT VT, SDValue V0, SDValue V1) {
1849   SDLoc dl(V0.getNode());
1850   SDValue RegClass =
1851     CurDAG->getTargetConstant(ARM::GPRPairRegClassID, dl, MVT::i32);
1852   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::gsub_0, dl, MVT::i32);
1853   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::gsub_1, dl, MVT::i32);
1854   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1855   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1856 }
1857 
1858 /// Form a D register from a pair of S registers.
1859 SDNode *ARMDAGToDAGISel::createSRegPairNode(EVT VT, SDValue V0, SDValue V1) {
1860   SDLoc dl(V0.getNode());
1861   SDValue RegClass =
1862     CurDAG->getTargetConstant(ARM::DPR_VFP2RegClassID, dl, MVT::i32);
1863   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, dl, MVT::i32);
1864   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, dl, MVT::i32);
1865   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1866   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1867 }
1868 
1869 /// Form a quad register from a pair of D registers.
1870 SDNode *ARMDAGToDAGISel::createDRegPairNode(EVT VT, SDValue V0, SDValue V1) {
1871   SDLoc dl(V0.getNode());
1872   SDValue RegClass = CurDAG->getTargetConstant(ARM::QPRRegClassID, dl,
1873                                                MVT::i32);
1874   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, dl, MVT::i32);
1875   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, dl, MVT::i32);
1876   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1877   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1878 }
1879 
1880 /// Form 4 consecutive D registers from a pair of Q registers.
1881 SDNode *ARMDAGToDAGISel::createQRegPairNode(EVT VT, SDValue V0, SDValue V1) {
1882   SDLoc dl(V0.getNode());
1883   SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, dl,
1884                                                MVT::i32);
1885   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, dl, MVT::i32);
1886   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, dl, MVT::i32);
1887   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1888   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1889 }
1890 
1891 /// Form 4 consecutive S registers.
1892 SDNode *ARMDAGToDAGISel::createQuadSRegsNode(EVT VT, SDValue V0, SDValue V1,
1893                                    SDValue V2, SDValue V3) {
1894   SDLoc dl(V0.getNode());
1895   SDValue RegClass =
1896     CurDAG->getTargetConstant(ARM::QPR_VFP2RegClassID, dl, MVT::i32);
1897   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, dl, MVT::i32);
1898   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, dl, MVT::i32);
1899   SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, dl, MVT::i32);
1900   SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, dl, MVT::i32);
1901   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1902                                     V2, SubReg2, V3, SubReg3 };
1903   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1904 }
1905 
1906 /// Form 4 consecutive D registers.
1907 SDNode *ARMDAGToDAGISel::createQuadDRegsNode(EVT VT, SDValue V0, SDValue V1,
1908                                    SDValue V2, SDValue V3) {
1909   SDLoc dl(V0.getNode());
1910   SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, dl,
1911                                                MVT::i32);
1912   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, dl, MVT::i32);
1913   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, dl, MVT::i32);
1914   SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, dl, MVT::i32);
1915   SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, dl, MVT::i32);
1916   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1917                                     V2, SubReg2, V3, SubReg3 };
1918   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1919 }
1920 
1921 /// Form 4 consecutive Q registers.
1922 SDNode *ARMDAGToDAGISel::createQuadQRegsNode(EVT VT, SDValue V0, SDValue V1,
1923                                    SDValue V2, SDValue V3) {
1924   SDLoc dl(V0.getNode());
1925   SDValue RegClass = CurDAG->getTargetConstant(ARM::QQQQPRRegClassID, dl,
1926                                                MVT::i32);
1927   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, dl, MVT::i32);
1928   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, dl, MVT::i32);
1929   SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, dl, MVT::i32);
1930   SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, dl, MVT::i32);
1931   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1932                                     V2, SubReg2, V3, SubReg3 };
1933   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1934 }
1935 
1936 /// GetVLDSTAlign - Get the alignment (in bytes) for the alignment operand
1937 /// of a NEON VLD or VST instruction.  The supported values depend on the
1938 /// number of registers being loaded.
1939 SDValue ARMDAGToDAGISel::GetVLDSTAlign(SDValue Align, const SDLoc &dl,
1940                                        unsigned NumVecs, bool is64BitVector) {
1941   unsigned NumRegs = NumVecs;
1942   if (!is64BitVector && NumVecs < 3)
1943     NumRegs *= 2;
1944 
1945   unsigned Alignment = cast<ConstantSDNode>(Align)->getZExtValue();
1946   if (Alignment >= 32 && NumRegs == 4)
1947     Alignment = 32;
1948   else if (Alignment >= 16 && (NumRegs == 2 || NumRegs == 4))
1949     Alignment = 16;
1950   else if (Alignment >= 8)
1951     Alignment = 8;
1952   else
1953     Alignment = 0;
1954 
1955   return CurDAG->getTargetConstant(Alignment, dl, MVT::i32);
1956 }
1957 
1958 static bool isVLDfixed(unsigned Opc)
1959 {
1960   switch (Opc) {
1961   default: return false;
1962   case ARM::VLD1d8wb_fixed : return true;
1963   case ARM::VLD1d16wb_fixed : return true;
1964   case ARM::VLD1d64Qwb_fixed : return true;
1965   case ARM::VLD1d32wb_fixed : return true;
1966   case ARM::VLD1d64wb_fixed : return true;
1967   case ARM::VLD1d8TPseudoWB_fixed : return true;
1968   case ARM::VLD1d16TPseudoWB_fixed : return true;
1969   case ARM::VLD1d32TPseudoWB_fixed : return true;
1970   case ARM::VLD1d64TPseudoWB_fixed : return true;
1971   case ARM::VLD1d8QPseudoWB_fixed : return true;
1972   case ARM::VLD1d16QPseudoWB_fixed : return true;
1973   case ARM::VLD1d32QPseudoWB_fixed : return true;
1974   case ARM::VLD1d64QPseudoWB_fixed : return true;
1975   case ARM::VLD1q8wb_fixed : return true;
1976   case ARM::VLD1q16wb_fixed : return true;
1977   case ARM::VLD1q32wb_fixed : return true;
1978   case ARM::VLD1q64wb_fixed : return true;
1979   case ARM::VLD1DUPd8wb_fixed : return true;
1980   case ARM::VLD1DUPd16wb_fixed : return true;
1981   case ARM::VLD1DUPd32wb_fixed : return true;
1982   case ARM::VLD1DUPq8wb_fixed : return true;
1983   case ARM::VLD1DUPq16wb_fixed : return true;
1984   case ARM::VLD1DUPq32wb_fixed : return true;
1985   case ARM::VLD2d8wb_fixed : return true;
1986   case ARM::VLD2d16wb_fixed : return true;
1987   case ARM::VLD2d32wb_fixed : return true;
1988   case ARM::VLD2q8PseudoWB_fixed : return true;
1989   case ARM::VLD2q16PseudoWB_fixed : return true;
1990   case ARM::VLD2q32PseudoWB_fixed : return true;
1991   case ARM::VLD2DUPd8wb_fixed : return true;
1992   case ARM::VLD2DUPd16wb_fixed : return true;
1993   case ARM::VLD2DUPd32wb_fixed : return true;
1994   case ARM::VLD2DUPq8OddPseudoWB_fixed: return true;
1995   case ARM::VLD2DUPq16OddPseudoWB_fixed: return true;
1996   case ARM::VLD2DUPq32OddPseudoWB_fixed: return true;
1997   }
1998 }
1999 
2000 static bool isVSTfixed(unsigned Opc)
2001 {
2002   switch (Opc) {
2003   default: return false;
2004   case ARM::VST1d8wb_fixed : return true;
2005   case ARM::VST1d16wb_fixed : return true;
2006   case ARM::VST1d32wb_fixed : return true;
2007   case ARM::VST1d64wb_fixed : return true;
2008   case ARM::VST1q8wb_fixed : return true;
2009   case ARM::VST1q16wb_fixed : return true;
2010   case ARM::VST1q32wb_fixed : return true;
2011   case ARM::VST1q64wb_fixed : return true;
2012   case ARM::VST1d8TPseudoWB_fixed : return true;
2013   case ARM::VST1d16TPseudoWB_fixed : return true;
2014   case ARM::VST1d32TPseudoWB_fixed : return true;
2015   case ARM::VST1d64TPseudoWB_fixed : return true;
2016   case ARM::VST1d8QPseudoWB_fixed : return true;
2017   case ARM::VST1d16QPseudoWB_fixed : return true;
2018   case ARM::VST1d32QPseudoWB_fixed : return true;
2019   case ARM::VST1d64QPseudoWB_fixed : return true;
2020   case ARM::VST2d8wb_fixed : return true;
2021   case ARM::VST2d16wb_fixed : return true;
2022   case ARM::VST2d32wb_fixed : return true;
2023   case ARM::VST2q8PseudoWB_fixed : return true;
2024   case ARM::VST2q16PseudoWB_fixed : return true;
2025   case ARM::VST2q32PseudoWB_fixed : return true;
2026   }
2027 }
2028 
2029 // Get the register stride update opcode of a VLD/VST instruction that
2030 // is otherwise equivalent to the given fixed stride updating instruction.
2031 static unsigned getVLDSTRegisterUpdateOpcode(unsigned Opc) {
2032   assert((isVLDfixed(Opc) || isVSTfixed(Opc))
2033     && "Incorrect fixed stride updating instruction.");
2034   switch (Opc) {
2035   default: break;
2036   case ARM::VLD1d8wb_fixed: return ARM::VLD1d8wb_register;
2037   case ARM::VLD1d16wb_fixed: return ARM::VLD1d16wb_register;
2038   case ARM::VLD1d32wb_fixed: return ARM::VLD1d32wb_register;
2039   case ARM::VLD1d64wb_fixed: return ARM::VLD1d64wb_register;
2040   case ARM::VLD1q8wb_fixed: return ARM::VLD1q8wb_register;
2041   case ARM::VLD1q16wb_fixed: return ARM::VLD1q16wb_register;
2042   case ARM::VLD1q32wb_fixed: return ARM::VLD1q32wb_register;
2043   case ARM::VLD1q64wb_fixed: return ARM::VLD1q64wb_register;
2044   case ARM::VLD1d64Twb_fixed: return ARM::VLD1d64Twb_register;
2045   case ARM::VLD1d64Qwb_fixed: return ARM::VLD1d64Qwb_register;
2046   case ARM::VLD1d8TPseudoWB_fixed: return ARM::VLD1d8TPseudoWB_register;
2047   case ARM::VLD1d16TPseudoWB_fixed: return ARM::VLD1d16TPseudoWB_register;
2048   case ARM::VLD1d32TPseudoWB_fixed: return ARM::VLD1d32TPseudoWB_register;
2049   case ARM::VLD1d64TPseudoWB_fixed: return ARM::VLD1d64TPseudoWB_register;
2050   case ARM::VLD1d8QPseudoWB_fixed: return ARM::VLD1d8QPseudoWB_register;
2051   case ARM::VLD1d16QPseudoWB_fixed: return ARM::VLD1d16QPseudoWB_register;
2052   case ARM::VLD1d32QPseudoWB_fixed: return ARM::VLD1d32QPseudoWB_register;
2053   case ARM::VLD1d64QPseudoWB_fixed: return ARM::VLD1d64QPseudoWB_register;
2054   case ARM::VLD1DUPd8wb_fixed : return ARM::VLD1DUPd8wb_register;
2055   case ARM::VLD1DUPd16wb_fixed : return ARM::VLD1DUPd16wb_register;
2056   case ARM::VLD1DUPd32wb_fixed : return ARM::VLD1DUPd32wb_register;
2057   case ARM::VLD1DUPq8wb_fixed : return ARM::VLD1DUPq8wb_register;
2058   case ARM::VLD1DUPq16wb_fixed : return ARM::VLD1DUPq16wb_register;
2059   case ARM::VLD1DUPq32wb_fixed : return ARM::VLD1DUPq32wb_register;
2060   case ARM::VLD2DUPq8OddPseudoWB_fixed: return ARM::VLD2DUPq8OddPseudoWB_register;
2061   case ARM::VLD2DUPq16OddPseudoWB_fixed: return ARM::VLD2DUPq16OddPseudoWB_register;
2062   case ARM::VLD2DUPq32OddPseudoWB_fixed: return ARM::VLD2DUPq32OddPseudoWB_register;
2063 
2064   case ARM::VST1d8wb_fixed: return ARM::VST1d8wb_register;
2065   case ARM::VST1d16wb_fixed: return ARM::VST1d16wb_register;
2066   case ARM::VST1d32wb_fixed: return ARM::VST1d32wb_register;
2067   case ARM::VST1d64wb_fixed: return ARM::VST1d64wb_register;
2068   case ARM::VST1q8wb_fixed: return ARM::VST1q8wb_register;
2069   case ARM::VST1q16wb_fixed: return ARM::VST1q16wb_register;
2070   case ARM::VST1q32wb_fixed: return ARM::VST1q32wb_register;
2071   case ARM::VST1q64wb_fixed: return ARM::VST1q64wb_register;
2072   case ARM::VST1d8TPseudoWB_fixed: return ARM::VST1d8TPseudoWB_register;
2073   case ARM::VST1d16TPseudoWB_fixed: return ARM::VST1d16TPseudoWB_register;
2074   case ARM::VST1d32TPseudoWB_fixed: return ARM::VST1d32TPseudoWB_register;
2075   case ARM::VST1d64TPseudoWB_fixed: return ARM::VST1d64TPseudoWB_register;
2076   case ARM::VST1d8QPseudoWB_fixed: return ARM::VST1d8QPseudoWB_register;
2077   case ARM::VST1d16QPseudoWB_fixed: return ARM::VST1d16QPseudoWB_register;
2078   case ARM::VST1d32QPseudoWB_fixed: return ARM::VST1d32QPseudoWB_register;
2079   case ARM::VST1d64QPseudoWB_fixed: return ARM::VST1d64QPseudoWB_register;
2080 
2081   case ARM::VLD2d8wb_fixed: return ARM::VLD2d8wb_register;
2082   case ARM::VLD2d16wb_fixed: return ARM::VLD2d16wb_register;
2083   case ARM::VLD2d32wb_fixed: return ARM::VLD2d32wb_register;
2084   case ARM::VLD2q8PseudoWB_fixed: return ARM::VLD2q8PseudoWB_register;
2085   case ARM::VLD2q16PseudoWB_fixed: return ARM::VLD2q16PseudoWB_register;
2086   case ARM::VLD2q32PseudoWB_fixed: return ARM::VLD2q32PseudoWB_register;
2087 
2088   case ARM::VST2d8wb_fixed: return ARM::VST2d8wb_register;
2089   case ARM::VST2d16wb_fixed: return ARM::VST2d16wb_register;
2090   case ARM::VST2d32wb_fixed: return ARM::VST2d32wb_register;
2091   case ARM::VST2q8PseudoWB_fixed: return ARM::VST2q8PseudoWB_register;
2092   case ARM::VST2q16PseudoWB_fixed: return ARM::VST2q16PseudoWB_register;
2093   case ARM::VST2q32PseudoWB_fixed: return ARM::VST2q32PseudoWB_register;
2094 
2095   case ARM::VLD2DUPd8wb_fixed: return ARM::VLD2DUPd8wb_register;
2096   case ARM::VLD2DUPd16wb_fixed: return ARM::VLD2DUPd16wb_register;
2097   case ARM::VLD2DUPd32wb_fixed: return ARM::VLD2DUPd32wb_register;
2098   }
2099   return Opc; // If not one we handle, return it unchanged.
2100 }
2101 
2102 /// Returns true if the given increment is a Constant known to be equal to the
2103 /// access size performed by a NEON load/store. This means the "[rN]!" form can
2104 /// be used.
2105 static bool isPerfectIncrement(SDValue Inc, EVT VecTy, unsigned NumVecs) {
2106   auto C = dyn_cast<ConstantSDNode>(Inc);
2107   return C && C->getZExtValue() == VecTy.getSizeInBits() / 8 * NumVecs;
2108 }
2109 
2110 void ARMDAGToDAGISel::SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs,
2111                                 const uint16_t *DOpcodes,
2112                                 const uint16_t *QOpcodes0,
2113                                 const uint16_t *QOpcodes1) {
2114   assert(Subtarget->hasNEON());
2115   assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
2116   SDLoc dl(N);
2117 
2118   SDValue MemAddr, Align;
2119   bool IsIntrinsic = !isUpdating;  // By coincidence, all supported updating
2120                                    // nodes are not intrinsics.
2121   unsigned AddrOpIdx = IsIntrinsic ? 2 : 1;
2122   if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
2123     return;
2124 
2125   SDValue Chain = N->getOperand(0);
2126   EVT VT = N->getValueType(0);
2127   bool is64BitVector = VT.is64BitVector();
2128   Align = GetVLDSTAlign(Align, dl, NumVecs, is64BitVector);
2129 
2130   unsigned OpcodeIndex;
2131   switch (VT.getSimpleVT().SimpleTy) {
2132   default: llvm_unreachable("unhandled vld type");
2133     // Double-register operations:
2134   case MVT::v8i8:  OpcodeIndex = 0; break;
2135   case MVT::v4f16:
2136   case MVT::v4bf16:
2137   case MVT::v4i16: OpcodeIndex = 1; break;
2138   case MVT::v2f32:
2139   case MVT::v2i32: OpcodeIndex = 2; break;
2140   case MVT::v1i64: OpcodeIndex = 3; break;
2141     // Quad-register operations:
2142   case MVT::v16i8: OpcodeIndex = 0; break;
2143   case MVT::v8f16:
2144   case MVT::v8bf16:
2145   case MVT::v8i16: OpcodeIndex = 1; break;
2146   case MVT::v4f32:
2147   case MVT::v4i32: OpcodeIndex = 2; break;
2148   case MVT::v2f64:
2149   case MVT::v2i64: OpcodeIndex = 3; break;
2150   }
2151 
2152   EVT ResTy;
2153   if (NumVecs == 1)
2154     ResTy = VT;
2155   else {
2156     unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
2157     if (!is64BitVector)
2158       ResTyElts *= 2;
2159     ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, ResTyElts);
2160   }
2161   std::vector<EVT> ResTys;
2162   ResTys.push_back(ResTy);
2163   if (isUpdating)
2164     ResTys.push_back(MVT::i32);
2165   ResTys.push_back(MVT::Other);
2166 
2167   SDValue Pred = getAL(CurDAG, dl);
2168   SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2169   SDNode *VLd;
2170   SmallVector<SDValue, 7> Ops;
2171 
2172   // Double registers and VLD1/VLD2 quad registers are directly supported.
2173   if (is64BitVector || NumVecs <= 2) {
2174     unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
2175                     QOpcodes0[OpcodeIndex]);
2176     Ops.push_back(MemAddr);
2177     Ops.push_back(Align);
2178     if (isUpdating) {
2179       SDValue Inc = N->getOperand(AddrOpIdx + 1);
2180       bool IsImmUpdate = isPerfectIncrement(Inc, VT, NumVecs);
2181       if (!IsImmUpdate) {
2182         // We use a VLD1 for v1i64 even if the pseudo says vld2/3/4, so
2183         // check for the opcode rather than the number of vector elements.
2184         if (isVLDfixed(Opc))
2185           Opc = getVLDSTRegisterUpdateOpcode(Opc);
2186         Ops.push_back(Inc);
2187       // VLD1/VLD2 fixed increment does not need Reg0 so only include it in
2188       // the operands if not such an opcode.
2189       } else if (!isVLDfixed(Opc))
2190         Ops.push_back(Reg0);
2191     }
2192     Ops.push_back(Pred);
2193     Ops.push_back(Reg0);
2194     Ops.push_back(Chain);
2195     VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
2196 
2197   } else {
2198     // Otherwise, quad registers are loaded with two separate instructions,
2199     // where one loads the even registers and the other loads the odd registers.
2200     EVT AddrTy = MemAddr.getValueType();
2201 
2202     // Load the even subregs.  This is always an updating load, so that it
2203     // provides the address to the second load for the odd subregs.
2204     SDValue ImplDef =
2205       SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, ResTy), 0);
2206     const SDValue OpsA[] = { MemAddr, Align, Reg0, ImplDef, Pred, Reg0, Chain };
2207     SDNode *VLdA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl,
2208                                           ResTy, AddrTy, MVT::Other, OpsA);
2209     Chain = SDValue(VLdA, 2);
2210 
2211     // Load the odd subregs.
2212     Ops.push_back(SDValue(VLdA, 1));
2213     Ops.push_back(Align);
2214     if (isUpdating) {
2215       SDValue Inc = N->getOperand(AddrOpIdx + 1);
2216       assert(isa<ConstantSDNode>(Inc.getNode()) &&
2217              "only constant post-increment update allowed for VLD3/4");
2218       (void)Inc;
2219       Ops.push_back(Reg0);
2220     }
2221     Ops.push_back(SDValue(VLdA, 0));
2222     Ops.push_back(Pred);
2223     Ops.push_back(Reg0);
2224     Ops.push_back(Chain);
2225     VLd = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys, Ops);
2226   }
2227 
2228   // Transfer memoperands.
2229   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2230   CurDAG->setNodeMemRefs(cast<MachineSDNode>(VLd), {MemOp});
2231 
2232   if (NumVecs == 1) {
2233     ReplaceNode(N, VLd);
2234     return;
2235   }
2236 
2237   // Extract out the subregisters.
2238   SDValue SuperReg = SDValue(VLd, 0);
2239   static_assert(ARM::dsub_7 == ARM::dsub_0 + 7 &&
2240                     ARM::qsub_3 == ARM::qsub_0 + 3,
2241                 "Unexpected subreg numbering");
2242   unsigned Sub0 = (is64BitVector ? ARM::dsub_0 : ARM::qsub_0);
2243   for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
2244     ReplaceUses(SDValue(N, Vec),
2245                 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg));
2246   ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, 1));
2247   if (isUpdating)
2248     ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLd, 2));
2249   CurDAG->RemoveDeadNode(N);
2250 }
2251 
2252 void ARMDAGToDAGISel::SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs,
2253                                 const uint16_t *DOpcodes,
2254                                 const uint16_t *QOpcodes0,
2255                                 const uint16_t *QOpcodes1) {
2256   assert(Subtarget->hasNEON());
2257   assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range");
2258   SDLoc dl(N);
2259 
2260   SDValue MemAddr, Align;
2261   bool IsIntrinsic = !isUpdating;  // By coincidence, all supported updating
2262                                    // nodes are not intrinsics.
2263   unsigned AddrOpIdx = IsIntrinsic ? 2 : 1;
2264   unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1)
2265   if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
2266     return;
2267 
2268   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2269 
2270   SDValue Chain = N->getOperand(0);
2271   EVT VT = N->getOperand(Vec0Idx).getValueType();
2272   bool is64BitVector = VT.is64BitVector();
2273   Align = GetVLDSTAlign(Align, dl, NumVecs, is64BitVector);
2274 
2275   unsigned OpcodeIndex;
2276   switch (VT.getSimpleVT().SimpleTy) {
2277   default: llvm_unreachable("unhandled vst type");
2278     // Double-register operations:
2279   case MVT::v8i8:  OpcodeIndex = 0; break;
2280   case MVT::v4f16:
2281   case MVT::v4bf16:
2282   case MVT::v4i16: OpcodeIndex = 1; break;
2283   case MVT::v2f32:
2284   case MVT::v2i32: OpcodeIndex = 2; break;
2285   case MVT::v1i64: OpcodeIndex = 3; break;
2286     // Quad-register operations:
2287   case MVT::v16i8: OpcodeIndex = 0; break;
2288   case MVT::v8f16:
2289   case MVT::v8bf16:
2290   case MVT::v8i16: OpcodeIndex = 1; break;
2291   case MVT::v4f32:
2292   case MVT::v4i32: OpcodeIndex = 2; break;
2293   case MVT::v2f64:
2294   case MVT::v2i64: OpcodeIndex = 3; break;
2295   }
2296 
2297   std::vector<EVT> ResTys;
2298   if (isUpdating)
2299     ResTys.push_back(MVT::i32);
2300   ResTys.push_back(MVT::Other);
2301 
2302   SDValue Pred = getAL(CurDAG, dl);
2303   SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2304   SmallVector<SDValue, 7> Ops;
2305 
2306   // Double registers and VST1/VST2 quad registers are directly supported.
2307   if (is64BitVector || NumVecs <= 2) {
2308     SDValue SrcReg;
2309     if (NumVecs == 1) {
2310       SrcReg = N->getOperand(Vec0Idx);
2311     } else if (is64BitVector) {
2312       // Form a REG_SEQUENCE to force register allocation.
2313       SDValue V0 = N->getOperand(Vec0Idx + 0);
2314       SDValue V1 = N->getOperand(Vec0Idx + 1);
2315       if (NumVecs == 2)
2316         SrcReg = SDValue(createDRegPairNode(MVT::v2i64, V0, V1), 0);
2317       else {
2318         SDValue V2 = N->getOperand(Vec0Idx + 2);
2319         // If it's a vst3, form a quad D-register and leave the last part as
2320         // an undef.
2321         SDValue V3 = (NumVecs == 3)
2322           ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
2323           : N->getOperand(Vec0Idx + 3);
2324         SrcReg = SDValue(createQuadDRegsNode(MVT::v4i64, V0, V1, V2, V3), 0);
2325       }
2326     } else {
2327       // Form a QQ register.
2328       SDValue Q0 = N->getOperand(Vec0Idx);
2329       SDValue Q1 = N->getOperand(Vec0Idx + 1);
2330       SrcReg = SDValue(createQRegPairNode(MVT::v4i64, Q0, Q1), 0);
2331     }
2332 
2333     unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
2334                     QOpcodes0[OpcodeIndex]);
2335     Ops.push_back(MemAddr);
2336     Ops.push_back(Align);
2337     if (isUpdating) {
2338       SDValue Inc = N->getOperand(AddrOpIdx + 1);
2339       bool IsImmUpdate = isPerfectIncrement(Inc, VT, NumVecs);
2340       if (!IsImmUpdate) {
2341         // We use a VST1 for v1i64 even if the pseudo says VST2/3/4, so
2342         // check for the opcode rather than the number of vector elements.
2343         if (isVSTfixed(Opc))
2344           Opc = getVLDSTRegisterUpdateOpcode(Opc);
2345         Ops.push_back(Inc);
2346       }
2347       // VST1/VST2 fixed increment does not need Reg0 so only include it in
2348       // the operands if not such an opcode.
2349       else if (!isVSTfixed(Opc))
2350         Ops.push_back(Reg0);
2351     }
2352     Ops.push_back(SrcReg);
2353     Ops.push_back(Pred);
2354     Ops.push_back(Reg0);
2355     Ops.push_back(Chain);
2356     SDNode *VSt = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
2357 
2358     // Transfer memoperands.
2359     CurDAG->setNodeMemRefs(cast<MachineSDNode>(VSt), {MemOp});
2360 
2361     ReplaceNode(N, VSt);
2362     return;
2363   }
2364 
2365   // Otherwise, quad registers are stored with two separate instructions,
2366   // where one stores the even registers and the other stores the odd registers.
2367 
2368   // Form the QQQQ REG_SEQUENCE.
2369   SDValue V0 = N->getOperand(Vec0Idx + 0);
2370   SDValue V1 = N->getOperand(Vec0Idx + 1);
2371   SDValue V2 = N->getOperand(Vec0Idx + 2);
2372   SDValue V3 = (NumVecs == 3)
2373     ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
2374     : N->getOperand(Vec0Idx + 3);
2375   SDValue RegSeq = SDValue(createQuadQRegsNode(MVT::v8i64, V0, V1, V2, V3), 0);
2376 
2377   // Store the even D registers.  This is always an updating store, so that it
2378   // provides the address to the second store for the odd subregs.
2379   const SDValue OpsA[] = { MemAddr, Align, Reg0, RegSeq, Pred, Reg0, Chain };
2380   SDNode *VStA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl,
2381                                         MemAddr.getValueType(),
2382                                         MVT::Other, OpsA);
2383   CurDAG->setNodeMemRefs(cast<MachineSDNode>(VStA), {MemOp});
2384   Chain = SDValue(VStA, 1);
2385 
2386   // Store the odd D registers.
2387   Ops.push_back(SDValue(VStA, 0));
2388   Ops.push_back(Align);
2389   if (isUpdating) {
2390     SDValue Inc = N->getOperand(AddrOpIdx + 1);
2391     assert(isa<ConstantSDNode>(Inc.getNode()) &&
2392            "only constant post-increment update allowed for VST3/4");
2393     (void)Inc;
2394     Ops.push_back(Reg0);
2395   }
2396   Ops.push_back(RegSeq);
2397   Ops.push_back(Pred);
2398   Ops.push_back(Reg0);
2399   Ops.push_back(Chain);
2400   SDNode *VStB = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys,
2401                                         Ops);
2402   CurDAG->setNodeMemRefs(cast<MachineSDNode>(VStB), {MemOp});
2403   ReplaceNode(N, VStB);
2404 }
2405 
2406 void ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad, bool isUpdating,
2407                                       unsigned NumVecs,
2408                                       const uint16_t *DOpcodes,
2409                                       const uint16_t *QOpcodes) {
2410   assert(Subtarget->hasNEON());
2411   assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
2412   SDLoc dl(N);
2413 
2414   SDValue MemAddr, Align;
2415   bool IsIntrinsic = !isUpdating;  // By coincidence, all supported updating
2416                                    // nodes are not intrinsics.
2417   unsigned AddrOpIdx = IsIntrinsic ? 2 : 1;
2418   unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1)
2419   if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
2420     return;
2421 
2422   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2423 
2424   SDValue Chain = N->getOperand(0);
2425   unsigned Lane =
2426     cast<ConstantSDNode>(N->getOperand(Vec0Idx + NumVecs))->getZExtValue();
2427   EVT VT = N->getOperand(Vec0Idx).getValueType();
2428   bool is64BitVector = VT.is64BitVector();
2429 
2430   unsigned Alignment = 0;
2431   if (NumVecs != 3) {
2432     Alignment = cast<ConstantSDNode>(Align)->getZExtValue();
2433     unsigned NumBytes = NumVecs * VT.getScalarSizeInBits() / 8;
2434     if (Alignment > NumBytes)
2435       Alignment = NumBytes;
2436     if (Alignment < 8 && Alignment < NumBytes)
2437       Alignment = 0;
2438     // Alignment must be a power of two; make sure of that.
2439     Alignment = (Alignment & -Alignment);
2440     if (Alignment == 1)
2441       Alignment = 0;
2442   }
2443   Align = CurDAG->getTargetConstant(Alignment, dl, MVT::i32);
2444 
2445   unsigned OpcodeIndex;
2446   switch (VT.getSimpleVT().SimpleTy) {
2447   default: llvm_unreachable("unhandled vld/vst lane type");
2448     // Double-register operations:
2449   case MVT::v8i8:  OpcodeIndex = 0; break;
2450   case MVT::v4f16:
2451   case MVT::v4bf16:
2452   case MVT::v4i16: OpcodeIndex = 1; break;
2453   case MVT::v2f32:
2454   case MVT::v2i32: OpcodeIndex = 2; break;
2455     // Quad-register operations:
2456   case MVT::v8f16:
2457   case MVT::v8bf16:
2458   case MVT::v8i16: OpcodeIndex = 0; break;
2459   case MVT::v4f32:
2460   case MVT::v4i32: OpcodeIndex = 1; break;
2461   }
2462 
2463   std::vector<EVT> ResTys;
2464   if (IsLoad) {
2465     unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
2466     if (!is64BitVector)
2467       ResTyElts *= 2;
2468     ResTys.push_back(EVT::getVectorVT(*CurDAG->getContext(),
2469                                       MVT::i64, ResTyElts));
2470   }
2471   if (isUpdating)
2472     ResTys.push_back(MVT::i32);
2473   ResTys.push_back(MVT::Other);
2474 
2475   SDValue Pred = getAL(CurDAG, dl);
2476   SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2477 
2478   SmallVector<SDValue, 8> Ops;
2479   Ops.push_back(MemAddr);
2480   Ops.push_back(Align);
2481   if (isUpdating) {
2482     SDValue Inc = N->getOperand(AddrOpIdx + 1);
2483     bool IsImmUpdate =
2484         isPerfectIncrement(Inc, VT.getVectorElementType(), NumVecs);
2485     Ops.push_back(IsImmUpdate ? Reg0 : Inc);
2486   }
2487 
2488   SDValue SuperReg;
2489   SDValue V0 = N->getOperand(Vec0Idx + 0);
2490   SDValue V1 = N->getOperand(Vec0Idx + 1);
2491   if (NumVecs == 2) {
2492     if (is64BitVector)
2493       SuperReg = SDValue(createDRegPairNode(MVT::v2i64, V0, V1), 0);
2494     else
2495       SuperReg = SDValue(createQRegPairNode(MVT::v4i64, V0, V1), 0);
2496   } else {
2497     SDValue V2 = N->getOperand(Vec0Idx + 2);
2498     SDValue V3 = (NumVecs == 3)
2499       ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
2500       : N->getOperand(Vec0Idx + 3);
2501     if (is64BitVector)
2502       SuperReg = SDValue(createQuadDRegsNode(MVT::v4i64, V0, V1, V2, V3), 0);
2503     else
2504       SuperReg = SDValue(createQuadQRegsNode(MVT::v8i64, V0, V1, V2, V3), 0);
2505   }
2506   Ops.push_back(SuperReg);
2507   Ops.push_back(getI32Imm(Lane, dl));
2508   Ops.push_back(Pred);
2509   Ops.push_back(Reg0);
2510   Ops.push_back(Chain);
2511 
2512   unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
2513                                   QOpcodes[OpcodeIndex]);
2514   SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
2515   CurDAG->setNodeMemRefs(cast<MachineSDNode>(VLdLn), {MemOp});
2516   if (!IsLoad) {
2517     ReplaceNode(N, VLdLn);
2518     return;
2519   }
2520 
2521   // Extract the subregisters.
2522   SuperReg = SDValue(VLdLn, 0);
2523   static_assert(ARM::dsub_7 == ARM::dsub_0 + 7 &&
2524                     ARM::qsub_3 == ARM::qsub_0 + 3,
2525                 "Unexpected subreg numbering");
2526   unsigned Sub0 = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
2527   for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
2528     ReplaceUses(SDValue(N, Vec),
2529                 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg));
2530   ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, 1));
2531   if (isUpdating)
2532     ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdLn, 2));
2533   CurDAG->RemoveDeadNode(N);
2534 }
2535 
2536 template <typename SDValueVector>
2537 void ARMDAGToDAGISel::AddMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc,
2538                                            SDValue PredicateMask) {
2539   Ops.push_back(CurDAG->getTargetConstant(ARMVCC::Then, Loc, MVT::i32));
2540   Ops.push_back(PredicateMask);
2541   Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // tp_reg
2542 }
2543 
2544 template <typename SDValueVector>
2545 void ARMDAGToDAGISel::AddMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc,
2546                                            SDValue PredicateMask,
2547                                            SDValue Inactive) {
2548   Ops.push_back(CurDAG->getTargetConstant(ARMVCC::Then, Loc, MVT::i32));
2549   Ops.push_back(PredicateMask);
2550   Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // tp_reg
2551   Ops.push_back(Inactive);
2552 }
2553 
2554 template <typename SDValueVector>
2555 void ARMDAGToDAGISel::AddEmptyMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc) {
2556   Ops.push_back(CurDAG->getTargetConstant(ARMVCC::None, Loc, MVT::i32));
2557   Ops.push_back(CurDAG->getRegister(0, MVT::i32));
2558   Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // tp_reg
2559 }
2560 
2561 template <typename SDValueVector>
2562 void ARMDAGToDAGISel::AddEmptyMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc,
2563                                                 EVT InactiveTy) {
2564   Ops.push_back(CurDAG->getTargetConstant(ARMVCC::None, Loc, MVT::i32));
2565   Ops.push_back(CurDAG->getRegister(0, MVT::i32));
2566   Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // tp_reg
2567   Ops.push_back(SDValue(
2568       CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, Loc, InactiveTy), 0));
2569 }
2570 
2571 void ARMDAGToDAGISel::SelectMVE_WB(SDNode *N, const uint16_t *Opcodes,
2572                                    bool Predicated) {
2573   SDLoc Loc(N);
2574   SmallVector<SDValue, 8> Ops;
2575 
2576   uint16_t Opcode;
2577   switch (N->getValueType(1).getVectorElementType().getSizeInBits()) {
2578   case 32:
2579     Opcode = Opcodes[0];
2580     break;
2581   case 64:
2582     Opcode = Opcodes[1];
2583     break;
2584   default:
2585     llvm_unreachable("bad vector element size in SelectMVE_WB");
2586   }
2587 
2588   Ops.push_back(N->getOperand(2)); // vector of base addresses
2589 
2590   int32_t ImmValue = cast<ConstantSDNode>(N->getOperand(3))->getZExtValue();
2591   Ops.push_back(getI32Imm(ImmValue, Loc)); // immediate offset
2592 
2593   if (Predicated)
2594     AddMVEPredicateToOps(Ops, Loc, N->getOperand(4));
2595   else
2596     AddEmptyMVEPredicateToOps(Ops, Loc);
2597 
2598   Ops.push_back(N->getOperand(0)); // chain
2599 
2600   SmallVector<EVT, 8> VTs;
2601   VTs.push_back(N->getValueType(1));
2602   VTs.push_back(N->getValueType(0));
2603   VTs.push_back(N->getValueType(2));
2604 
2605   SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), VTs, Ops);
2606   ReplaceUses(SDValue(N, 0), SDValue(New, 1));
2607   ReplaceUses(SDValue(N, 1), SDValue(New, 0));
2608   ReplaceUses(SDValue(N, 2), SDValue(New, 2));
2609   transferMemOperands(N, New);
2610   CurDAG->RemoveDeadNode(N);
2611 }
2612 
2613 void ARMDAGToDAGISel::SelectMVE_LongShift(SDNode *N, uint16_t Opcode,
2614                                           bool Immediate,
2615                                           bool HasSaturationOperand) {
2616   SDLoc Loc(N);
2617   SmallVector<SDValue, 8> Ops;
2618 
2619   // Two 32-bit halves of the value to be shifted
2620   Ops.push_back(N->getOperand(1));
2621   Ops.push_back(N->getOperand(2));
2622 
2623   // The shift count
2624   if (Immediate) {
2625     int32_t ImmValue = cast<ConstantSDNode>(N->getOperand(3))->getZExtValue();
2626     Ops.push_back(getI32Imm(ImmValue, Loc)); // immediate shift count
2627   } else {
2628     Ops.push_back(N->getOperand(3));
2629   }
2630 
2631   // The immediate saturation operand, if any
2632   if (HasSaturationOperand) {
2633     int32_t SatOp = cast<ConstantSDNode>(N->getOperand(4))->getZExtValue();
2634     int SatBit = (SatOp == 64 ? 0 : 1);
2635     Ops.push_back(getI32Imm(SatBit, Loc));
2636   }
2637 
2638   // MVE scalar shifts are IT-predicable, so include the standard
2639   // predicate arguments.
2640   Ops.push_back(getAL(CurDAG, Loc));
2641   Ops.push_back(CurDAG->getRegister(0, MVT::i32));
2642 
2643   CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), ArrayRef(Ops));
2644 }
2645 
2646 void ARMDAGToDAGISel::SelectMVE_VADCSBC(SDNode *N, uint16_t OpcodeWithCarry,
2647                                         uint16_t OpcodeWithNoCarry,
2648                                         bool Add, bool Predicated) {
2649   SDLoc Loc(N);
2650   SmallVector<SDValue, 8> Ops;
2651   uint16_t Opcode;
2652 
2653   unsigned FirstInputOp = Predicated ? 2 : 1;
2654 
2655   // Two input vectors and the input carry flag
2656   Ops.push_back(N->getOperand(FirstInputOp));
2657   Ops.push_back(N->getOperand(FirstInputOp + 1));
2658   SDValue CarryIn = N->getOperand(FirstInputOp + 2);
2659   ConstantSDNode *CarryInConstant = dyn_cast<ConstantSDNode>(CarryIn);
2660   uint32_t CarryMask = 1 << 29;
2661   uint32_t CarryExpected = Add ? 0 : CarryMask;
2662   if (CarryInConstant &&
2663       (CarryInConstant->getZExtValue() & CarryMask) == CarryExpected) {
2664     Opcode = OpcodeWithNoCarry;
2665   } else {
2666     Ops.push_back(CarryIn);
2667     Opcode = OpcodeWithCarry;
2668   }
2669 
2670   if (Predicated)
2671     AddMVEPredicateToOps(Ops, Loc,
2672                          N->getOperand(FirstInputOp + 3),  // predicate
2673                          N->getOperand(FirstInputOp - 1)); // inactive
2674   else
2675     AddEmptyMVEPredicateToOps(Ops, Loc, N->getValueType(0));
2676 
2677   CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), ArrayRef(Ops));
2678 }
2679 
2680 void ARMDAGToDAGISel::SelectMVE_VSHLC(SDNode *N, bool Predicated) {
2681   SDLoc Loc(N);
2682   SmallVector<SDValue, 8> Ops;
2683 
2684   // One vector input, followed by a 32-bit word of bits to shift in
2685   // and then an immediate shift count
2686   Ops.push_back(N->getOperand(1));
2687   Ops.push_back(N->getOperand(2));
2688   int32_t ImmValue = cast<ConstantSDNode>(N->getOperand(3))->getZExtValue();
2689   Ops.push_back(getI32Imm(ImmValue, Loc)); // immediate shift count
2690 
2691   if (Predicated)
2692     AddMVEPredicateToOps(Ops, Loc, N->getOperand(4));
2693   else
2694     AddEmptyMVEPredicateToOps(Ops, Loc);
2695 
2696   CurDAG->SelectNodeTo(N, ARM::MVE_VSHLC, N->getVTList(), ArrayRef(Ops));
2697 }
2698 
2699 static bool SDValueToConstBool(SDValue SDVal) {
2700   assert(isa<ConstantSDNode>(SDVal) && "expected a compile-time constant");
2701   ConstantSDNode *SDValConstant = dyn_cast<ConstantSDNode>(SDVal);
2702   uint64_t Value = SDValConstant->getZExtValue();
2703   assert((Value == 0 || Value == 1) && "expected value 0 or 1");
2704   return Value;
2705 }
2706 
2707 void ARMDAGToDAGISel::SelectBaseMVE_VMLLDAV(SDNode *N, bool Predicated,
2708                                             const uint16_t *OpcodesS,
2709                                             const uint16_t *OpcodesU,
2710                                             size_t Stride, size_t TySize) {
2711   assert(TySize < Stride && "Invalid TySize");
2712   bool IsUnsigned = SDValueToConstBool(N->getOperand(1));
2713   bool IsSub = SDValueToConstBool(N->getOperand(2));
2714   bool IsExchange = SDValueToConstBool(N->getOperand(3));
2715   if (IsUnsigned) {
2716     assert(!IsSub &&
2717            "Unsigned versions of vmlsldav[a]/vrmlsldavh[a] do not exist");
2718     assert(!IsExchange &&
2719            "Unsigned versions of vmlaldav[a]x/vrmlaldavh[a]x do not exist");
2720   }
2721 
2722   auto OpIsZero = [N](size_t OpNo) {
2723     if (ConstantSDNode *OpConst = dyn_cast<ConstantSDNode>(N->getOperand(OpNo)))
2724       if (OpConst->getZExtValue() == 0)
2725         return true;
2726     return false;
2727   };
2728 
2729   // If the input accumulator value is not zero, select an instruction with
2730   // accumulator, otherwise select an instruction without accumulator
2731   bool IsAccum = !(OpIsZero(4) && OpIsZero(5));
2732 
2733   const uint16_t *Opcodes = IsUnsigned ? OpcodesU : OpcodesS;
2734   if (IsSub)
2735     Opcodes += 4 * Stride;
2736   if (IsExchange)
2737     Opcodes += 2 * Stride;
2738   if (IsAccum)
2739     Opcodes += Stride;
2740   uint16_t Opcode = Opcodes[TySize];
2741 
2742   SDLoc Loc(N);
2743   SmallVector<SDValue, 8> Ops;
2744   // Push the accumulator operands, if they are used
2745   if (IsAccum) {
2746     Ops.push_back(N->getOperand(4));
2747     Ops.push_back(N->getOperand(5));
2748   }
2749   // Push the two vector operands
2750   Ops.push_back(N->getOperand(6));
2751   Ops.push_back(N->getOperand(7));
2752 
2753   if (Predicated)
2754     AddMVEPredicateToOps(Ops, Loc, N->getOperand(8));
2755   else
2756     AddEmptyMVEPredicateToOps(Ops, Loc);
2757 
2758   CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), ArrayRef(Ops));
2759 }
2760 
2761 void ARMDAGToDAGISel::SelectMVE_VMLLDAV(SDNode *N, bool Predicated,
2762                                         const uint16_t *OpcodesS,
2763                                         const uint16_t *OpcodesU) {
2764   EVT VecTy = N->getOperand(6).getValueType();
2765   size_t SizeIndex;
2766   switch (VecTy.getVectorElementType().getSizeInBits()) {
2767   case 16:
2768     SizeIndex = 0;
2769     break;
2770   case 32:
2771     SizeIndex = 1;
2772     break;
2773   default:
2774     llvm_unreachable("bad vector element size");
2775   }
2776 
2777   SelectBaseMVE_VMLLDAV(N, Predicated, OpcodesS, OpcodesU, 2, SizeIndex);
2778 }
2779 
2780 void ARMDAGToDAGISel::SelectMVE_VRMLLDAVH(SDNode *N, bool Predicated,
2781                                           const uint16_t *OpcodesS,
2782                                           const uint16_t *OpcodesU) {
2783   assert(
2784       N->getOperand(6).getValueType().getVectorElementType().getSizeInBits() ==
2785           32 &&
2786       "bad vector element size");
2787   SelectBaseMVE_VMLLDAV(N, Predicated, OpcodesS, OpcodesU, 1, 0);
2788 }
2789 
2790 void ARMDAGToDAGISel::SelectMVE_VLD(SDNode *N, unsigned NumVecs,
2791                                     const uint16_t *const *Opcodes,
2792                                     bool HasWriteback) {
2793   EVT VT = N->getValueType(0);
2794   SDLoc Loc(N);
2795 
2796   const uint16_t *OurOpcodes;
2797   switch (VT.getVectorElementType().getSizeInBits()) {
2798   case 8:
2799     OurOpcodes = Opcodes[0];
2800     break;
2801   case 16:
2802     OurOpcodes = Opcodes[1];
2803     break;
2804   case 32:
2805     OurOpcodes = Opcodes[2];
2806     break;
2807   default:
2808     llvm_unreachable("bad vector element size in SelectMVE_VLD");
2809   }
2810 
2811   EVT DataTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, NumVecs * 2);
2812   SmallVector<EVT, 4> ResultTys = {DataTy, MVT::Other};
2813   unsigned PtrOperand = HasWriteback ? 1 : 2;
2814 
2815   auto Data = SDValue(
2816       CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, Loc, DataTy), 0);
2817   SDValue Chain = N->getOperand(0);
2818   // Add a MVE_VLDn instruction for each Vec, except the last
2819   for (unsigned Stage = 0; Stage < NumVecs - 1; ++Stage) {
2820     SDValue Ops[] = {Data, N->getOperand(PtrOperand), Chain};
2821     auto LoadInst =
2822         CurDAG->getMachineNode(OurOpcodes[Stage], Loc, ResultTys, Ops);
2823     Data = SDValue(LoadInst, 0);
2824     Chain = SDValue(LoadInst, 1);
2825     transferMemOperands(N, LoadInst);
2826   }
2827   // The last may need a writeback on it
2828   if (HasWriteback)
2829     ResultTys = {DataTy, MVT::i32, MVT::Other};
2830   SDValue Ops[] = {Data, N->getOperand(PtrOperand), Chain};
2831   auto LoadInst =
2832       CurDAG->getMachineNode(OurOpcodes[NumVecs - 1], Loc, ResultTys, Ops);
2833   transferMemOperands(N, LoadInst);
2834 
2835   unsigned i;
2836   for (i = 0; i < NumVecs; i++)
2837     ReplaceUses(SDValue(N, i),
2838                 CurDAG->getTargetExtractSubreg(ARM::qsub_0 + i, Loc, VT,
2839                                                SDValue(LoadInst, 0)));
2840   if (HasWriteback)
2841     ReplaceUses(SDValue(N, i++), SDValue(LoadInst, 1));
2842   ReplaceUses(SDValue(N, i), SDValue(LoadInst, HasWriteback ? 2 : 1));
2843   CurDAG->RemoveDeadNode(N);
2844 }
2845 
2846 void ARMDAGToDAGISel::SelectMVE_VxDUP(SDNode *N, const uint16_t *Opcodes,
2847                                       bool Wrapping, bool Predicated) {
2848   EVT VT = N->getValueType(0);
2849   SDLoc Loc(N);
2850 
2851   uint16_t Opcode;
2852   switch (VT.getScalarSizeInBits()) {
2853   case 8:
2854     Opcode = Opcodes[0];
2855     break;
2856   case 16:
2857     Opcode = Opcodes[1];
2858     break;
2859   case 32:
2860     Opcode = Opcodes[2];
2861     break;
2862   default:
2863     llvm_unreachable("bad vector element size in SelectMVE_VxDUP");
2864   }
2865 
2866   SmallVector<SDValue, 8> Ops;
2867   unsigned OpIdx = 1;
2868 
2869   SDValue Inactive;
2870   if (Predicated)
2871     Inactive = N->getOperand(OpIdx++);
2872 
2873   Ops.push_back(N->getOperand(OpIdx++));     // base
2874   if (Wrapping)
2875     Ops.push_back(N->getOperand(OpIdx++));   // limit
2876 
2877   SDValue ImmOp = N->getOperand(OpIdx++);    // step
2878   int ImmValue = cast<ConstantSDNode>(ImmOp)->getZExtValue();
2879   Ops.push_back(getI32Imm(ImmValue, Loc));
2880 
2881   if (Predicated)
2882     AddMVEPredicateToOps(Ops, Loc, N->getOperand(OpIdx), Inactive);
2883   else
2884     AddEmptyMVEPredicateToOps(Ops, Loc, N->getValueType(0));
2885 
2886   CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), ArrayRef(Ops));
2887 }
2888 
2889 void ARMDAGToDAGISel::SelectCDE_CXxD(SDNode *N, uint16_t Opcode,
2890                                      size_t NumExtraOps, bool HasAccum) {
2891   bool IsBigEndian = CurDAG->getDataLayout().isBigEndian();
2892   SDLoc Loc(N);
2893   SmallVector<SDValue, 8> Ops;
2894 
2895   unsigned OpIdx = 1;
2896 
2897   // Convert and append the immediate operand designating the coprocessor.
2898   SDValue ImmCorpoc = N->getOperand(OpIdx++);
2899   uint32_t ImmCoprocVal = cast<ConstantSDNode>(ImmCorpoc)->getZExtValue();
2900   Ops.push_back(getI32Imm(ImmCoprocVal, Loc));
2901 
2902   // For accumulating variants copy the low and high order parts of the
2903   // accumulator into a register pair and add it to the operand vector.
2904   if (HasAccum) {
2905     SDValue AccLo = N->getOperand(OpIdx++);
2906     SDValue AccHi = N->getOperand(OpIdx++);
2907     if (IsBigEndian)
2908       std::swap(AccLo, AccHi);
2909     Ops.push_back(SDValue(createGPRPairNode(MVT::Untyped, AccLo, AccHi), 0));
2910   }
2911 
2912   // Copy extra operands as-is.
2913   for (size_t I = 0; I < NumExtraOps; I++)
2914     Ops.push_back(N->getOperand(OpIdx++));
2915 
2916   // Convert and append the immediate operand
2917   SDValue Imm = N->getOperand(OpIdx);
2918   uint32_t ImmVal = cast<ConstantSDNode>(Imm)->getZExtValue();
2919   Ops.push_back(getI32Imm(ImmVal, Loc));
2920 
2921   // Accumulating variants are IT-predicable, add predicate operands.
2922   if (HasAccum) {
2923     SDValue Pred = getAL(CurDAG, Loc);
2924     SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2925     Ops.push_back(Pred);
2926     Ops.push_back(PredReg);
2927   }
2928 
2929   // Create the CDE intruction
2930   SDNode *InstrNode = CurDAG->getMachineNode(Opcode, Loc, MVT::Untyped, Ops);
2931   SDValue ResultPair = SDValue(InstrNode, 0);
2932 
2933   // The original intrinsic had two outputs, and the output of the dual-register
2934   // CDE instruction is a register pair. We need to extract the two subregisters
2935   // and replace all uses of the original outputs with the extracted
2936   // subregisters.
2937   uint16_t SubRegs[2] = {ARM::gsub_0, ARM::gsub_1};
2938   if (IsBigEndian)
2939     std::swap(SubRegs[0], SubRegs[1]);
2940 
2941   for (size_t ResIdx = 0; ResIdx < 2; ResIdx++) {
2942     if (SDValue(N, ResIdx).use_empty())
2943       continue;
2944     SDValue SubReg = CurDAG->getTargetExtractSubreg(SubRegs[ResIdx], Loc,
2945                                                     MVT::i32, ResultPair);
2946     ReplaceUses(SDValue(N, ResIdx), SubReg);
2947   }
2948 
2949   CurDAG->RemoveDeadNode(N);
2950 }
2951 
2952 void ARMDAGToDAGISel::SelectVLDDup(SDNode *N, bool IsIntrinsic,
2953                                    bool isUpdating, unsigned NumVecs,
2954                                    const uint16_t *DOpcodes,
2955                                    const uint16_t *QOpcodes0,
2956                                    const uint16_t *QOpcodes1) {
2957   assert(Subtarget->hasNEON());
2958   assert(NumVecs >= 1 && NumVecs <= 4 && "VLDDup NumVecs out-of-range");
2959   SDLoc dl(N);
2960 
2961   SDValue MemAddr, Align;
2962   unsigned AddrOpIdx = IsIntrinsic ? 2 : 1;
2963   if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
2964     return;
2965 
2966   SDValue Chain = N->getOperand(0);
2967   EVT VT = N->getValueType(0);
2968   bool is64BitVector = VT.is64BitVector();
2969 
2970   unsigned Alignment = 0;
2971   if (NumVecs != 3) {
2972     Alignment = cast<ConstantSDNode>(Align)->getZExtValue();
2973     unsigned NumBytes = NumVecs * VT.getScalarSizeInBits() / 8;
2974     if (Alignment > NumBytes)
2975       Alignment = NumBytes;
2976     if (Alignment < 8 && Alignment < NumBytes)
2977       Alignment = 0;
2978     // Alignment must be a power of two; make sure of that.
2979     Alignment = (Alignment & -Alignment);
2980     if (Alignment == 1)
2981       Alignment = 0;
2982   }
2983   Align = CurDAG->getTargetConstant(Alignment, dl, MVT::i32);
2984 
2985   unsigned OpcodeIndex;
2986   switch (VT.getSimpleVT().SimpleTy) {
2987   default: llvm_unreachable("unhandled vld-dup type");
2988   case MVT::v8i8:
2989   case MVT::v16i8: OpcodeIndex = 0; break;
2990   case MVT::v4i16:
2991   case MVT::v8i16:
2992   case MVT::v4f16:
2993   case MVT::v8f16:
2994   case MVT::v4bf16:
2995   case MVT::v8bf16:
2996                   OpcodeIndex = 1; break;
2997   case MVT::v2f32:
2998   case MVT::v2i32:
2999   case MVT::v4f32:
3000   case MVT::v4i32: OpcodeIndex = 2; break;
3001   case MVT::v1f64:
3002   case MVT::v1i64: OpcodeIndex = 3; break;
3003   }
3004 
3005   unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
3006   if (!is64BitVector)
3007     ResTyElts *= 2;
3008   EVT ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, ResTyElts);
3009 
3010   std::vector<EVT> ResTys;
3011   ResTys.push_back(ResTy);
3012   if (isUpdating)
3013     ResTys.push_back(MVT::i32);
3014   ResTys.push_back(MVT::Other);
3015 
3016   SDValue Pred = getAL(CurDAG, dl);
3017   SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3018 
3019   SmallVector<SDValue, 6> Ops;
3020   Ops.push_back(MemAddr);
3021   Ops.push_back(Align);
3022   unsigned Opc = is64BitVector    ? DOpcodes[OpcodeIndex]
3023                  : (NumVecs == 1) ? QOpcodes0[OpcodeIndex]
3024                                   : QOpcodes1[OpcodeIndex];
3025   if (isUpdating) {
3026     SDValue Inc = N->getOperand(2);
3027     bool IsImmUpdate =
3028         isPerfectIncrement(Inc, VT.getVectorElementType(), NumVecs);
3029     if (IsImmUpdate) {
3030       if (!isVLDfixed(Opc))
3031         Ops.push_back(Reg0);
3032     } else {
3033       if (isVLDfixed(Opc))
3034         Opc = getVLDSTRegisterUpdateOpcode(Opc);
3035       Ops.push_back(Inc);
3036     }
3037   }
3038   if (is64BitVector || NumVecs == 1) {
3039     // Double registers and VLD1 quad registers are directly supported.
3040   } else if (NumVecs == 2) {
3041     const SDValue OpsA[] = {MemAddr, Align, Pred, Reg0, Chain};
3042     SDNode *VLdA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl, ResTy,
3043                                           MVT::Other, OpsA);
3044     Chain = SDValue(VLdA, 1);
3045   } else {
3046     SDValue ImplDef = SDValue(
3047         CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, ResTy), 0);
3048     const SDValue OpsA[] = {MemAddr, Align, ImplDef, Pred, Reg0, Chain};
3049     SDNode *VLdA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl, ResTy,
3050                                           MVT::Other, OpsA);
3051     Ops.push_back(SDValue(VLdA, 0));
3052     Chain = SDValue(VLdA, 1);
3053   }
3054 
3055   Ops.push_back(Pred);
3056   Ops.push_back(Reg0);
3057   Ops.push_back(Chain);
3058 
3059   SDNode *VLdDup = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
3060 
3061   // Transfer memoperands.
3062   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
3063   CurDAG->setNodeMemRefs(cast<MachineSDNode>(VLdDup), {MemOp});
3064 
3065   // Extract the subregisters.
3066   if (NumVecs == 1) {
3067     ReplaceUses(SDValue(N, 0), SDValue(VLdDup, 0));
3068   } else {
3069     SDValue SuperReg = SDValue(VLdDup, 0);
3070     static_assert(ARM::dsub_7 == ARM::dsub_0 + 7, "Unexpected subreg numbering");
3071     unsigned SubIdx = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
3072     for (unsigned Vec = 0; Vec != NumVecs; ++Vec) {
3073       ReplaceUses(SDValue(N, Vec),
3074                   CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, SuperReg));
3075     }
3076   }
3077   ReplaceUses(SDValue(N, NumVecs), SDValue(VLdDup, 1));
3078   if (isUpdating)
3079     ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdDup, 2));
3080   CurDAG->RemoveDeadNode(N);
3081 }
3082 
3083 bool ARMDAGToDAGISel::tryInsertVectorElt(SDNode *N) {
3084   if (!Subtarget->hasMVEIntegerOps())
3085     return false;
3086 
3087   SDLoc dl(N);
3088 
3089   // We are trying to use VMOV/VMOVX/VINS to more efficiently lower insert and
3090   // extracts of v8f16 and v8i16 vectors. Check that we have two adjacent
3091   // inserts of the correct type:
3092   SDValue Ins1 = SDValue(N, 0);
3093   SDValue Ins2 = N->getOperand(0);
3094   EVT VT = Ins1.getValueType();
3095   if (Ins2.getOpcode() != ISD::INSERT_VECTOR_ELT || !Ins2.hasOneUse() ||
3096       !isa<ConstantSDNode>(Ins1.getOperand(2)) ||
3097       !isa<ConstantSDNode>(Ins2.getOperand(2)) ||
3098       (VT != MVT::v8f16 && VT != MVT::v8i16) || (Ins2.getValueType() != VT))
3099     return false;
3100 
3101   unsigned Lane1 = Ins1.getConstantOperandVal(2);
3102   unsigned Lane2 = Ins2.getConstantOperandVal(2);
3103   if (Lane2 % 2 != 0 || Lane1 != Lane2 + 1)
3104     return false;
3105 
3106   // If the inserted values will be able to use T/B already, leave it to the
3107   // existing tablegen patterns. For example VCVTT/VCVTB.
3108   SDValue Val1 = Ins1.getOperand(1);
3109   SDValue Val2 = Ins2.getOperand(1);
3110   if (Val1.getOpcode() == ISD::FP_ROUND || Val2.getOpcode() == ISD::FP_ROUND)
3111     return false;
3112 
3113   // Check if the inserted values are both extracts.
3114   if ((Val1.getOpcode() == ISD::EXTRACT_VECTOR_ELT ||
3115        Val1.getOpcode() == ARMISD::VGETLANEu) &&
3116       (Val2.getOpcode() == ISD::EXTRACT_VECTOR_ELT ||
3117        Val2.getOpcode() == ARMISD::VGETLANEu) &&
3118       isa<ConstantSDNode>(Val1.getOperand(1)) &&
3119       isa<ConstantSDNode>(Val2.getOperand(1)) &&
3120       (Val1.getOperand(0).getValueType() == MVT::v8f16 ||
3121        Val1.getOperand(0).getValueType() == MVT::v8i16) &&
3122       (Val2.getOperand(0).getValueType() == MVT::v8f16 ||
3123        Val2.getOperand(0).getValueType() == MVT::v8i16)) {
3124     unsigned ExtractLane1 = Val1.getConstantOperandVal(1);
3125     unsigned ExtractLane2 = Val2.getConstantOperandVal(1);
3126 
3127     // If the two extracted lanes are from the same place and adjacent, this
3128     // simplifies into a f32 lane move.
3129     if (Val1.getOperand(0) == Val2.getOperand(0) && ExtractLane2 % 2 == 0 &&
3130         ExtractLane1 == ExtractLane2 + 1) {
3131       SDValue NewExt = CurDAG->getTargetExtractSubreg(
3132           ARM::ssub_0 + ExtractLane2 / 2, dl, MVT::f32, Val1.getOperand(0));
3133       SDValue NewIns = CurDAG->getTargetInsertSubreg(
3134           ARM::ssub_0 + Lane2 / 2, dl, VT, Ins2.getOperand(0),
3135           NewExt);
3136       ReplaceUses(Ins1, NewIns);
3137       return true;
3138     }
3139 
3140     // Else v8i16 pattern of an extract and an insert, with a optional vmovx for
3141     // extracting odd lanes.
3142     if (VT == MVT::v8i16 && Subtarget->hasFullFP16()) {
3143       SDValue Inp1 = CurDAG->getTargetExtractSubreg(
3144           ARM::ssub_0 + ExtractLane1 / 2, dl, MVT::f32, Val1.getOperand(0));
3145       SDValue Inp2 = CurDAG->getTargetExtractSubreg(
3146           ARM::ssub_0 + ExtractLane2 / 2, dl, MVT::f32, Val2.getOperand(0));
3147       if (ExtractLane1 % 2 != 0)
3148         Inp1 = SDValue(CurDAG->getMachineNode(ARM::VMOVH, dl, MVT::f32, Inp1), 0);
3149       if (ExtractLane2 % 2 != 0)
3150         Inp2 = SDValue(CurDAG->getMachineNode(ARM::VMOVH, dl, MVT::f32, Inp2), 0);
3151       SDNode *VINS = CurDAG->getMachineNode(ARM::VINSH, dl, MVT::f32, Inp2, Inp1);
3152       SDValue NewIns =
3153           CurDAG->getTargetInsertSubreg(ARM::ssub_0 + Lane2 / 2, dl, MVT::v4f32,
3154                                         Ins2.getOperand(0), SDValue(VINS, 0));
3155       ReplaceUses(Ins1, NewIns);
3156       return true;
3157     }
3158   }
3159 
3160   // The inserted values are not extracted - if they are f16 then insert them
3161   // directly using a VINS.
3162   if (VT == MVT::v8f16 && Subtarget->hasFullFP16()) {
3163     SDNode *VINS = CurDAG->getMachineNode(ARM::VINSH, dl, MVT::f32, Val2, Val1);
3164     SDValue NewIns =
3165         CurDAG->getTargetInsertSubreg(ARM::ssub_0 + Lane2 / 2, dl, MVT::v4f32,
3166                                       Ins2.getOperand(0), SDValue(VINS, 0));
3167     ReplaceUses(Ins1, NewIns);
3168     return true;
3169   }
3170 
3171   return false;
3172 }
3173 
3174 bool ARMDAGToDAGISel::transformFixedFloatingPointConversion(SDNode *N,
3175                                                             SDNode *FMul,
3176                                                             bool IsUnsigned,
3177                                                             bool FixedToFloat) {
3178   auto Type = N->getValueType(0);
3179   unsigned ScalarBits = Type.getScalarSizeInBits();
3180   if (ScalarBits > 32)
3181     return false;
3182 
3183   SDNodeFlags FMulFlags = FMul->getFlags();
3184   // The fixed-point vcvt and vcvt+vmul are not always equivalent if inf is
3185   // allowed in 16 bit unsigned floats
3186   if (ScalarBits == 16 && !FMulFlags.hasNoInfs() && IsUnsigned)
3187     return false;
3188 
3189   SDValue ImmNode = FMul->getOperand(1);
3190   SDValue VecVal = FMul->getOperand(0);
3191   if (VecVal->getOpcode() == ISD::UINT_TO_FP ||
3192       VecVal->getOpcode() == ISD::SINT_TO_FP)
3193     VecVal = VecVal->getOperand(0);
3194 
3195   if (VecVal.getValueType().getScalarSizeInBits() != ScalarBits)
3196     return false;
3197 
3198   if (ImmNode.getOpcode() == ISD::BITCAST) {
3199     if (ImmNode.getValueType().getScalarSizeInBits() != ScalarBits)
3200       return false;
3201     ImmNode = ImmNode.getOperand(0);
3202   }
3203 
3204   if (ImmNode.getValueType().getScalarSizeInBits() != ScalarBits)
3205     return false;
3206 
3207   APFloat ImmAPF(0.0f);
3208   switch (ImmNode.getOpcode()) {
3209   case ARMISD::VMOVIMM:
3210   case ARMISD::VDUP: {
3211     if (!isa<ConstantSDNode>(ImmNode.getOperand(0)))
3212       return false;
3213     unsigned Imm = ImmNode.getConstantOperandVal(0);
3214     if (ImmNode.getOpcode() == ARMISD::VMOVIMM)
3215       Imm = ARM_AM::decodeVMOVModImm(Imm, ScalarBits);
3216     ImmAPF =
3217         APFloat(ScalarBits == 32 ? APFloat::IEEEsingle() : APFloat::IEEEhalf(),
3218                 APInt(ScalarBits, Imm));
3219     break;
3220   }
3221   case ARMISD::VMOVFPIMM: {
3222     ImmAPF = APFloat(ARM_AM::getFPImmFloat(ImmNode.getConstantOperandVal(0)));
3223     break;
3224   }
3225   default:
3226     return false;
3227   }
3228 
3229   // Where n is the number of fractional bits, multiplying by 2^n will convert
3230   // from float to fixed and multiplying by 2^-n will convert from fixed to
3231   // float. Taking log2 of the factor (after taking the inverse in the case of
3232   // float to fixed) will give n.
3233   APFloat ToConvert = ImmAPF;
3234   if (FixedToFloat) {
3235     if (!ImmAPF.getExactInverse(&ToConvert))
3236       return false;
3237   }
3238   APSInt Converted(64, false);
3239   bool IsExact;
3240   ToConvert.convertToInteger(Converted, llvm::RoundingMode::NearestTiesToEven,
3241                              &IsExact);
3242   if (!IsExact || !Converted.isPowerOf2())
3243     return false;
3244 
3245   unsigned FracBits = Converted.logBase2();
3246   if (FracBits > ScalarBits)
3247     return false;
3248 
3249   SmallVector<SDValue, 3> Ops{
3250       VecVal, CurDAG->getConstant(FracBits, SDLoc(N), MVT::i32)};
3251   AddEmptyMVEPredicateToOps(Ops, SDLoc(N), Type);
3252 
3253   unsigned int Opcode;
3254   switch (ScalarBits) {
3255   case 16:
3256     if (FixedToFloat)
3257       Opcode = IsUnsigned ? ARM::MVE_VCVTf16u16_fix : ARM::MVE_VCVTf16s16_fix;
3258     else
3259       Opcode = IsUnsigned ? ARM::MVE_VCVTu16f16_fix : ARM::MVE_VCVTs16f16_fix;
3260     break;
3261   case 32:
3262     if (FixedToFloat)
3263       Opcode = IsUnsigned ? ARM::MVE_VCVTf32u32_fix : ARM::MVE_VCVTf32s32_fix;
3264     else
3265       Opcode = IsUnsigned ? ARM::MVE_VCVTu32f32_fix : ARM::MVE_VCVTs32f32_fix;
3266     break;
3267   default:
3268     llvm_unreachable("unexpected number of scalar bits");
3269     break;
3270   }
3271 
3272   ReplaceNode(N, CurDAG->getMachineNode(Opcode, SDLoc(N), Type, Ops));
3273   return true;
3274 }
3275 
3276 bool ARMDAGToDAGISel::tryFP_TO_INT(SDNode *N, SDLoc dl) {
3277   // Transform a floating-point to fixed-point conversion to a VCVT
3278   if (!Subtarget->hasMVEFloatOps())
3279     return false;
3280   EVT Type = N->getValueType(0);
3281   if (!Type.isVector())
3282     return false;
3283   unsigned int ScalarBits = Type.getScalarSizeInBits();
3284 
3285   bool IsUnsigned = N->getOpcode() == ISD::FP_TO_UINT ||
3286                     N->getOpcode() == ISD::FP_TO_UINT_SAT;
3287   SDNode *Node = N->getOperand(0).getNode();
3288 
3289   // floating-point to fixed-point with one fractional bit gets turned into an
3290   // FP_TO_[U|S]INT(FADD (x, x)) rather than an FP_TO_[U|S]INT(FMUL (x, y))
3291   if (Node->getOpcode() == ISD::FADD) {
3292     if (Node->getOperand(0) != Node->getOperand(1))
3293       return false;
3294     SDNodeFlags Flags = Node->getFlags();
3295     // The fixed-point vcvt and vcvt+vmul are not always equivalent if inf is
3296     // allowed in 16 bit unsigned floats
3297     if (ScalarBits == 16 && !Flags.hasNoInfs() && IsUnsigned)
3298       return false;
3299 
3300     unsigned Opcode;
3301     switch (ScalarBits) {
3302     case 16:
3303       Opcode = IsUnsigned ? ARM::MVE_VCVTu16f16_fix : ARM::MVE_VCVTs16f16_fix;
3304       break;
3305     case 32:
3306       Opcode = IsUnsigned ? ARM::MVE_VCVTu32f32_fix : ARM::MVE_VCVTs32f32_fix;
3307       break;
3308     }
3309     SmallVector<SDValue, 3> Ops{Node->getOperand(0),
3310                                 CurDAG->getConstant(1, dl, MVT::i32)};
3311     AddEmptyMVEPredicateToOps(Ops, dl, Type);
3312 
3313     ReplaceNode(N, CurDAG->getMachineNode(Opcode, dl, Type, Ops));
3314     return true;
3315   }
3316 
3317   if (Node->getOpcode() != ISD::FMUL)
3318     return false;
3319 
3320   return transformFixedFloatingPointConversion(N, Node, IsUnsigned, false);
3321 }
3322 
3323 bool ARMDAGToDAGISel::tryFMULFixed(SDNode *N, SDLoc dl) {
3324   // Transform a fixed-point to floating-point conversion to a VCVT
3325   if (!Subtarget->hasMVEFloatOps())
3326     return false;
3327   auto Type = N->getValueType(0);
3328   if (!Type.isVector())
3329     return false;
3330 
3331   auto LHS = N->getOperand(0);
3332   if (LHS.getOpcode() != ISD::SINT_TO_FP && LHS.getOpcode() != ISD::UINT_TO_FP)
3333     return false;
3334 
3335   return transformFixedFloatingPointConversion(
3336       N, N, LHS.getOpcode() == ISD::UINT_TO_FP, true);
3337 }
3338 
3339 bool ARMDAGToDAGISel::tryV6T2BitfieldExtractOp(SDNode *N, bool isSigned) {
3340   if (!Subtarget->hasV6T2Ops())
3341     return false;
3342 
3343   unsigned Opc = isSigned
3344     ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
3345     : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
3346   SDLoc dl(N);
3347 
3348   // For unsigned extracts, check for a shift right and mask
3349   unsigned And_imm = 0;
3350   if (N->getOpcode() == ISD::AND) {
3351     if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
3352 
3353       // The immediate is a mask of the low bits iff imm & (imm+1) == 0
3354       if (And_imm & (And_imm + 1))
3355         return false;
3356 
3357       unsigned Srl_imm = 0;
3358       if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
3359                                 Srl_imm)) {
3360         assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
3361 
3362         // Mask off the unnecessary bits of the AND immediate; normally
3363         // DAGCombine will do this, but that might not happen if
3364         // targetShrinkDemandedConstant chooses a different immediate.
3365         And_imm &= -1U >> Srl_imm;
3366 
3367         // Note: The width operand is encoded as width-1.
3368         unsigned Width = countTrailingOnes(And_imm) - 1;
3369         unsigned LSB = Srl_imm;
3370 
3371         SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3372 
3373         if ((LSB + Width + 1) == N->getValueType(0).getSizeInBits()) {
3374           // It's cheaper to use a right shift to extract the top bits.
3375           if (Subtarget->isThumb()) {
3376             Opc = isSigned ? ARM::t2ASRri : ARM::t2LSRri;
3377             SDValue Ops[] = { N->getOperand(0).getOperand(0),
3378                               CurDAG->getTargetConstant(LSB, dl, MVT::i32),
3379                               getAL(CurDAG, dl), Reg0, Reg0 };
3380             CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3381             return true;
3382           }
3383 
3384           // ARM models shift instructions as MOVsi with shifter operand.
3385           ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(ISD::SRL);
3386           SDValue ShOpc =
3387             CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, LSB), dl,
3388                                       MVT::i32);
3389           SDValue Ops[] = { N->getOperand(0).getOperand(0), ShOpc,
3390                             getAL(CurDAG, dl), Reg0, Reg0 };
3391           CurDAG->SelectNodeTo(N, ARM::MOVsi, MVT::i32, Ops);
3392           return true;
3393         }
3394 
3395         assert(LSB + Width + 1 <= 32 && "Shouldn't create an invalid ubfx");
3396         SDValue Ops[] = { N->getOperand(0).getOperand(0),
3397                           CurDAG->getTargetConstant(LSB, dl, MVT::i32),
3398                           CurDAG->getTargetConstant(Width, dl, MVT::i32),
3399                           getAL(CurDAG, dl), Reg0 };
3400         CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3401         return true;
3402       }
3403     }
3404     return false;
3405   }
3406 
3407   // Otherwise, we're looking for a shift of a shift
3408   unsigned Shl_imm = 0;
3409   if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
3410     assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
3411     unsigned Srl_imm = 0;
3412     if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
3413       assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
3414       // Note: The width operand is encoded as width-1.
3415       unsigned Width = 32 - Srl_imm - 1;
3416       int LSB = Srl_imm - Shl_imm;
3417       if (LSB < 0)
3418         return false;
3419       SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3420       assert(LSB + Width + 1 <= 32 && "Shouldn't create an invalid ubfx");
3421       SDValue Ops[] = { N->getOperand(0).getOperand(0),
3422                         CurDAG->getTargetConstant(LSB, dl, MVT::i32),
3423                         CurDAG->getTargetConstant(Width, dl, MVT::i32),
3424                         getAL(CurDAG, dl), Reg0 };
3425       CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3426       return true;
3427     }
3428   }
3429 
3430   // Or we are looking for a shift of an and, with a mask operand
3431   if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, And_imm) &&
3432       isShiftedMask_32(And_imm)) {
3433     unsigned Srl_imm = 0;
3434     unsigned LSB = countTrailingZeros(And_imm);
3435     // Shift must be the same as the ands lsb
3436     if (isInt32Immediate(N->getOperand(1), Srl_imm) && Srl_imm == LSB) {
3437       assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
3438       unsigned MSB = 31 - countLeadingZeros(And_imm);
3439       // Note: The width operand is encoded as width-1.
3440       unsigned Width = MSB - LSB;
3441       SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3442       assert(Srl_imm + Width + 1 <= 32 && "Shouldn't create an invalid ubfx");
3443       SDValue Ops[] = { N->getOperand(0).getOperand(0),
3444                         CurDAG->getTargetConstant(Srl_imm, dl, MVT::i32),
3445                         CurDAG->getTargetConstant(Width, dl, MVT::i32),
3446                         getAL(CurDAG, dl), Reg0 };
3447       CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3448       return true;
3449     }
3450   }
3451 
3452   if (N->getOpcode() == ISD::SIGN_EXTEND_INREG) {
3453     unsigned Width = cast<VTSDNode>(N->getOperand(1))->getVT().getSizeInBits();
3454     unsigned LSB = 0;
3455     if (!isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL, LSB) &&
3456         !isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRA, LSB))
3457       return false;
3458 
3459     if (LSB + Width > 32)
3460       return false;
3461 
3462     SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3463     assert(LSB + Width <= 32 && "Shouldn't create an invalid ubfx");
3464     SDValue Ops[] = { N->getOperand(0).getOperand(0),
3465                       CurDAG->getTargetConstant(LSB, dl, MVT::i32),
3466                       CurDAG->getTargetConstant(Width - 1, dl, MVT::i32),
3467                       getAL(CurDAG, dl), Reg0 };
3468     CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3469     return true;
3470   }
3471 
3472   return false;
3473 }
3474 
3475 /// Target-specific DAG combining for ISD::SUB.
3476 /// Target-independent combining lowers SELECT_CC nodes of the form
3477 /// select_cc setg[ge] X,  0,  X, -X
3478 /// select_cc setgt    X, -1,  X, -X
3479 /// select_cc setl[te] X,  0, -X,  X
3480 /// select_cc setlt    X,  1, -X,  X
3481 /// which represent Integer ABS into:
3482 /// Y = sra (X, size(X)-1); sub (xor (X, Y), Y)
3483 /// ARM instruction selection detects the latter and matches it to
3484 /// ARM::ABS or ARM::t2ABS machine node.
3485 bool ARMDAGToDAGISel::tryABSOp(SDNode *N){
3486   SDValue SUBSrc0 = N->getOperand(0);
3487   SDValue SUBSrc1 = N->getOperand(1);
3488   EVT VT = N->getValueType(0);
3489 
3490   if (Subtarget->isThumb1Only())
3491     return false;
3492 
3493   if (SUBSrc0.getOpcode() != ISD::XOR || SUBSrc1.getOpcode() != ISD::SRA)
3494     return false;
3495 
3496   SDValue XORSrc0 = SUBSrc0.getOperand(0);
3497   SDValue XORSrc1 = SUBSrc0.getOperand(1);
3498   SDValue SRASrc0 = SUBSrc1.getOperand(0);
3499   SDValue SRASrc1 = SUBSrc1.getOperand(1);
3500   ConstantSDNode *SRAConstant =  dyn_cast<ConstantSDNode>(SRASrc1);
3501   EVT XType = SRASrc0.getValueType();
3502   unsigned Size = XType.getSizeInBits() - 1;
3503 
3504   if (XORSrc1 == SUBSrc1 && XORSrc0 == SRASrc0 && XType.isInteger() &&
3505       SRAConstant != nullptr && Size == SRAConstant->getZExtValue()) {
3506     unsigned Opcode = Subtarget->isThumb2() ? ARM::t2ABS : ARM::ABS;
3507     CurDAG->SelectNodeTo(N, Opcode, VT, XORSrc0);
3508     return true;
3509   }
3510 
3511   return false;
3512 }
3513 
3514 /// We've got special pseudo-instructions for these
3515 void ARMDAGToDAGISel::SelectCMP_SWAP(SDNode *N) {
3516   unsigned Opcode;
3517   EVT MemTy = cast<MemSDNode>(N)->getMemoryVT();
3518   if (MemTy == MVT::i8)
3519     Opcode = Subtarget->isThumb() ? ARM::tCMP_SWAP_8 : ARM::CMP_SWAP_8;
3520   else if (MemTy == MVT::i16)
3521     Opcode = Subtarget->isThumb() ? ARM::tCMP_SWAP_16 : ARM::CMP_SWAP_16;
3522   else if (MemTy == MVT::i32)
3523     Opcode = Subtarget->isThumb() ? ARM::tCMP_SWAP_32 : ARM::CMP_SWAP_32;
3524   else
3525     llvm_unreachable("Unknown AtomicCmpSwap type");
3526 
3527   SDValue Ops[] = {N->getOperand(1), N->getOperand(2), N->getOperand(3),
3528                    N->getOperand(0)};
3529   SDNode *CmpSwap = CurDAG->getMachineNode(
3530       Opcode, SDLoc(N),
3531       CurDAG->getVTList(MVT::i32, MVT::i32, MVT::Other), Ops);
3532 
3533   MachineMemOperand *MemOp = cast<MemSDNode>(N)->getMemOperand();
3534   CurDAG->setNodeMemRefs(cast<MachineSDNode>(CmpSwap), {MemOp});
3535 
3536   ReplaceUses(SDValue(N, 0), SDValue(CmpSwap, 0));
3537   ReplaceUses(SDValue(N, 1), SDValue(CmpSwap, 2));
3538   CurDAG->RemoveDeadNode(N);
3539 }
3540 
3541 static std::optional<std::pair<unsigned, unsigned>>
3542 getContiguousRangeOfSetBits(const APInt &A) {
3543   unsigned FirstOne = A.getBitWidth() - A.countLeadingZeros() - 1;
3544   unsigned LastOne = A.countTrailingZeros();
3545   if (A.countPopulation() != (FirstOne - LastOne + 1))
3546     return std::nullopt;
3547   return std::make_pair(FirstOne, LastOne);
3548 }
3549 
3550 void ARMDAGToDAGISel::SelectCMPZ(SDNode *N, bool &SwitchEQNEToPLMI) {
3551   assert(N->getOpcode() == ARMISD::CMPZ);
3552   SwitchEQNEToPLMI = false;
3553 
3554   if (!Subtarget->isThumb())
3555     // FIXME: Work out whether it is profitable to do this in A32 mode - LSL and
3556     // LSR don't exist as standalone instructions - they need the barrel shifter.
3557     return;
3558 
3559   // select (cmpz (and X, C), #0) -> (LSLS X) or (LSRS X) or (LSRS (LSLS X))
3560   SDValue And = N->getOperand(0);
3561   if (!And->hasOneUse())
3562     return;
3563 
3564   SDValue Zero = N->getOperand(1);
3565   if (!isa<ConstantSDNode>(Zero) || !cast<ConstantSDNode>(Zero)->isZero() ||
3566       And->getOpcode() != ISD::AND)
3567     return;
3568   SDValue X = And.getOperand(0);
3569   auto C = dyn_cast<ConstantSDNode>(And.getOperand(1));
3570 
3571   if (!C)
3572     return;
3573   auto Range = getContiguousRangeOfSetBits(C->getAPIntValue());
3574   if (!Range)
3575     return;
3576 
3577   // There are several ways to lower this:
3578   SDNode *NewN;
3579   SDLoc dl(N);
3580 
3581   auto EmitShift = [&](unsigned Opc, SDValue Src, unsigned Imm) -> SDNode* {
3582     if (Subtarget->isThumb2()) {
3583       Opc = (Opc == ARM::tLSLri) ? ARM::t2LSLri : ARM::t2LSRri;
3584       SDValue Ops[] = { Src, CurDAG->getTargetConstant(Imm, dl, MVT::i32),
3585                         getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32),
3586                         CurDAG->getRegister(0, MVT::i32) };
3587       return CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops);
3588     } else {
3589       SDValue Ops[] = {CurDAG->getRegister(ARM::CPSR, MVT::i32), Src,
3590                        CurDAG->getTargetConstant(Imm, dl, MVT::i32),
3591                        getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32)};
3592       return CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops);
3593     }
3594   };
3595 
3596   if (Range->second == 0) {
3597     //  1. Mask includes the LSB -> Simply shift the top N bits off
3598     NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
3599     ReplaceNode(And.getNode(), NewN);
3600   } else if (Range->first == 31) {
3601     //  2. Mask includes the MSB -> Simply shift the bottom N bits off
3602     NewN = EmitShift(ARM::tLSRri, X, Range->second);
3603     ReplaceNode(And.getNode(), NewN);
3604   } else if (Range->first == Range->second) {
3605     //  3. Only one bit is set. We can shift this into the sign bit and use a
3606     //     PL/MI comparison.
3607     NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
3608     ReplaceNode(And.getNode(), NewN);
3609 
3610     SwitchEQNEToPLMI = true;
3611   } else if (!Subtarget->hasV6T2Ops()) {
3612     //  4. Do a double shift to clear bottom and top bits, but only in
3613     //     thumb-1 mode as in thumb-2 we can use UBFX.
3614     NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
3615     NewN = EmitShift(ARM::tLSRri, SDValue(NewN, 0),
3616                      Range->second + (31 - Range->first));
3617     ReplaceNode(And.getNode(), NewN);
3618   }
3619 }
3620 
3621 static unsigned getVectorShuffleOpcode(EVT VT, unsigned Opc64[3],
3622                                        unsigned Opc128[3]) {
3623   assert((VT.is64BitVector() || VT.is128BitVector()) &&
3624          "Unexpected vector shuffle length");
3625   switch (VT.getScalarSizeInBits()) {
3626   default:
3627     llvm_unreachable("Unexpected vector shuffle element size");
3628   case 8:
3629     return VT.is64BitVector() ? Opc64[0] : Opc128[0];
3630   case 16:
3631     return VT.is64BitVector() ? Opc64[1] : Opc128[1];
3632   case 32:
3633     return VT.is64BitVector() ? Opc64[2] : Opc128[2];
3634   }
3635 }
3636 
3637 void ARMDAGToDAGISel::Select(SDNode *N) {
3638   SDLoc dl(N);
3639 
3640   if (N->isMachineOpcode()) {
3641     N->setNodeId(-1);
3642     return;   // Already selected.
3643   }
3644 
3645   switch (N->getOpcode()) {
3646   default: break;
3647   case ISD::STORE: {
3648     // For Thumb1, match an sp-relative store in C++. This is a little
3649     // unfortunate, but I don't think I can make the chain check work
3650     // otherwise.  (The chain of the store has to be the same as the chain
3651     // of the CopyFromReg, or else we can't replace the CopyFromReg with
3652     // a direct reference to "SP".)
3653     //
3654     // This is only necessary on Thumb1 because Thumb1 sp-relative stores use
3655     // a different addressing mode from other four-byte stores.
3656     //
3657     // This pattern usually comes up with call arguments.
3658     StoreSDNode *ST = cast<StoreSDNode>(N);
3659     SDValue Ptr = ST->getBasePtr();
3660     if (Subtarget->isThumb1Only() && ST->isUnindexed()) {
3661       int RHSC = 0;
3662       if (Ptr.getOpcode() == ISD::ADD &&
3663           isScaledConstantInRange(Ptr.getOperand(1), /*Scale=*/4, 0, 256, RHSC))
3664         Ptr = Ptr.getOperand(0);
3665 
3666       if (Ptr.getOpcode() == ISD::CopyFromReg &&
3667           cast<RegisterSDNode>(Ptr.getOperand(1))->getReg() == ARM::SP &&
3668           Ptr.getOperand(0) == ST->getChain()) {
3669         SDValue Ops[] = {ST->getValue(),
3670                          CurDAG->getRegister(ARM::SP, MVT::i32),
3671                          CurDAG->getTargetConstant(RHSC, dl, MVT::i32),
3672                          getAL(CurDAG, dl),
3673                          CurDAG->getRegister(0, MVT::i32),
3674                          ST->getChain()};
3675         MachineSDNode *ResNode =
3676             CurDAG->getMachineNode(ARM::tSTRspi, dl, MVT::Other, Ops);
3677         MachineMemOperand *MemOp = ST->getMemOperand();
3678         CurDAG->setNodeMemRefs(cast<MachineSDNode>(ResNode), {MemOp});
3679         ReplaceNode(N, ResNode);
3680         return;
3681       }
3682     }
3683     break;
3684   }
3685   case ISD::WRITE_REGISTER:
3686     if (tryWriteRegister(N))
3687       return;
3688     break;
3689   case ISD::READ_REGISTER:
3690     if (tryReadRegister(N))
3691       return;
3692     break;
3693   case ISD::INLINEASM:
3694   case ISD::INLINEASM_BR:
3695     if (tryInlineAsm(N))
3696       return;
3697     break;
3698   case ISD::SUB:
3699     // Select special operations if SUB node forms integer ABS pattern
3700     if (tryABSOp(N))
3701       return;
3702     // Other cases are autogenerated.
3703     break;
3704   case ISD::Constant: {
3705     unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
3706     // If we can't materialize the constant we need to use a literal pool
3707     if (ConstantMaterializationCost(Val, Subtarget) > 2) {
3708       SDValue CPIdx = CurDAG->getTargetConstantPool(
3709           ConstantInt::get(Type::getInt32Ty(*CurDAG->getContext()), Val),
3710           TLI->getPointerTy(CurDAG->getDataLayout()));
3711 
3712       SDNode *ResNode;
3713       if (Subtarget->isThumb()) {
3714         SDValue Ops[] = {
3715           CPIdx,
3716           getAL(CurDAG, dl),
3717           CurDAG->getRegister(0, MVT::i32),
3718           CurDAG->getEntryNode()
3719         };
3720         ResNode = CurDAG->getMachineNode(ARM::tLDRpci, dl, MVT::i32, MVT::Other,
3721                                          Ops);
3722       } else {
3723         SDValue Ops[] = {
3724           CPIdx,
3725           CurDAG->getTargetConstant(0, dl, MVT::i32),
3726           getAL(CurDAG, dl),
3727           CurDAG->getRegister(0, MVT::i32),
3728           CurDAG->getEntryNode()
3729         };
3730         ResNode = CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
3731                                          Ops);
3732       }
3733       // Annotate the Node with memory operand information so that MachineInstr
3734       // queries work properly. This e.g. gives the register allocation the
3735       // required information for rematerialization.
3736       MachineFunction& MF = CurDAG->getMachineFunction();
3737       MachineMemOperand *MemOp =
3738           MF.getMachineMemOperand(MachinePointerInfo::getConstantPool(MF),
3739                                   MachineMemOperand::MOLoad, 4, Align(4));
3740 
3741       CurDAG->setNodeMemRefs(cast<MachineSDNode>(ResNode), {MemOp});
3742 
3743       ReplaceNode(N, ResNode);
3744       return;
3745     }
3746 
3747     // Other cases are autogenerated.
3748     break;
3749   }
3750   case ISD::FrameIndex: {
3751     // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
3752     int FI = cast<FrameIndexSDNode>(N)->getIndex();
3753     SDValue TFI = CurDAG->getTargetFrameIndex(
3754         FI, TLI->getPointerTy(CurDAG->getDataLayout()));
3755     if (Subtarget->isThumb1Only()) {
3756       // Set the alignment of the frame object to 4, to avoid having to generate
3757       // more than one ADD
3758       MachineFrameInfo &MFI = MF->getFrameInfo();
3759       if (MFI.getObjectAlign(FI) < Align(4))
3760         MFI.setObjectAlignment(FI, Align(4));
3761       CurDAG->SelectNodeTo(N, ARM::tADDframe, MVT::i32, TFI,
3762                            CurDAG->getTargetConstant(0, dl, MVT::i32));
3763       return;
3764     } else {
3765       unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
3766                       ARM::t2ADDri : ARM::ADDri);
3767       SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, dl, MVT::i32),
3768                         getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32),
3769                         CurDAG->getRegister(0, MVT::i32) };
3770       CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3771       return;
3772     }
3773   }
3774   case ISD::INSERT_VECTOR_ELT: {
3775     if (tryInsertVectorElt(N))
3776       return;
3777     break;
3778   }
3779   case ISD::SRL:
3780     if (tryV6T2BitfieldExtractOp(N, false))
3781       return;
3782     break;
3783   case ISD::SIGN_EXTEND_INREG:
3784   case ISD::SRA:
3785     if (tryV6T2BitfieldExtractOp(N, true))
3786       return;
3787     break;
3788   case ISD::FP_TO_UINT:
3789   case ISD::FP_TO_SINT:
3790   case ISD::FP_TO_UINT_SAT:
3791   case ISD::FP_TO_SINT_SAT:
3792     if (tryFP_TO_INT(N, dl))
3793       return;
3794     break;
3795   case ISD::FMUL:
3796     if (tryFMULFixed(N, dl))
3797       return;
3798     break;
3799   case ISD::MUL:
3800     if (Subtarget->isThumb1Only())
3801       break;
3802     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
3803       unsigned RHSV = C->getZExtValue();
3804       if (!RHSV) break;
3805       if (isPowerOf2_32(RHSV-1)) {  // 2^n+1?
3806         unsigned ShImm = Log2_32(RHSV-1);
3807         if (ShImm >= 32)
3808           break;
3809         SDValue V = N->getOperand(0);
3810         ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
3811         SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, dl, MVT::i32);
3812         SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3813         if (Subtarget->isThumb()) {
3814           SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG, dl), Reg0, Reg0 };
3815           CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops);
3816           return;
3817         } else {
3818           SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG, dl), Reg0,
3819                             Reg0 };
3820           CurDAG->SelectNodeTo(N, ARM::ADDrsi, MVT::i32, Ops);
3821           return;
3822         }
3823       }
3824       if (isPowerOf2_32(RHSV+1)) {  // 2^n-1?
3825         unsigned ShImm = Log2_32(RHSV+1);
3826         if (ShImm >= 32)
3827           break;
3828         SDValue V = N->getOperand(0);
3829         ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
3830         SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, dl, MVT::i32);
3831         SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3832         if (Subtarget->isThumb()) {
3833           SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG, dl), Reg0, Reg0 };
3834           CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops);
3835           return;
3836         } else {
3837           SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG, dl), Reg0,
3838                             Reg0 };
3839           CurDAG->SelectNodeTo(N, ARM::RSBrsi, MVT::i32, Ops);
3840           return;
3841         }
3842       }
3843     }
3844     break;
3845   case ISD::AND: {
3846     // Check for unsigned bitfield extract
3847     if (tryV6T2BitfieldExtractOp(N, false))
3848       return;
3849 
3850     // If an immediate is used in an AND node, it is possible that the immediate
3851     // can be more optimally materialized when negated. If this is the case we
3852     // can negate the immediate and use a BIC instead.
3853     auto *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1));
3854     if (N1C && N1C->hasOneUse() && Subtarget->isThumb()) {
3855       uint32_t Imm = (uint32_t) N1C->getZExtValue();
3856 
3857       // In Thumb2 mode, an AND can take a 12-bit immediate. If this
3858       // immediate can be negated and fit in the immediate operand of
3859       // a t2BIC, don't do any manual transform here as this can be
3860       // handled by the generic ISel machinery.
3861       bool PreferImmediateEncoding =
3862         Subtarget->hasThumb2() && (is_t2_so_imm(Imm) || is_t2_so_imm_not(Imm));
3863       if (!PreferImmediateEncoding &&
3864           ConstantMaterializationCost(Imm, Subtarget) >
3865               ConstantMaterializationCost(~Imm, Subtarget)) {
3866         // The current immediate costs more to materialize than a negated
3867         // immediate, so negate the immediate and use a BIC.
3868         SDValue NewImm =
3869           CurDAG->getConstant(~N1C->getZExtValue(), dl, MVT::i32);
3870         // If the new constant didn't exist before, reposition it in the topological
3871         // ordering so it is just before N. Otherwise, don't touch its location.
3872         if (NewImm->getNodeId() == -1)
3873           CurDAG->RepositionNode(N->getIterator(), NewImm.getNode());
3874 
3875         if (!Subtarget->hasThumb2()) {
3876           SDValue Ops[] = {CurDAG->getRegister(ARM::CPSR, MVT::i32),
3877                            N->getOperand(0), NewImm, getAL(CurDAG, dl),
3878                            CurDAG->getRegister(0, MVT::i32)};
3879           ReplaceNode(N, CurDAG->getMachineNode(ARM::tBIC, dl, MVT::i32, Ops));
3880           return;
3881         } else {
3882           SDValue Ops[] = {N->getOperand(0), NewImm, getAL(CurDAG, dl),
3883                            CurDAG->getRegister(0, MVT::i32),
3884                            CurDAG->getRegister(0, MVT::i32)};
3885           ReplaceNode(N,
3886                       CurDAG->getMachineNode(ARM::t2BICrr, dl, MVT::i32, Ops));
3887           return;
3888         }
3889       }
3890     }
3891 
3892     // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
3893     // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
3894     // are entirely contributed by c2 and lower 16-bits are entirely contributed
3895     // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
3896     // Select it to: "movt x, ((c1 & 0xffff) >> 16)
3897     EVT VT = N->getValueType(0);
3898     if (VT != MVT::i32)
3899       break;
3900     unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
3901       ? ARM::t2MOVTi16
3902       : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
3903     if (!Opc)
3904       break;
3905     SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3906     N1C = dyn_cast<ConstantSDNode>(N1);
3907     if (!N1C)
3908       break;
3909     if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
3910       SDValue N2 = N0.getOperand(1);
3911       ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
3912       if (!N2C)
3913         break;
3914       unsigned N1CVal = N1C->getZExtValue();
3915       unsigned N2CVal = N2C->getZExtValue();
3916       if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
3917           (N1CVal & 0xffffU) == 0xffffU &&
3918           (N2CVal & 0xffffU) == 0x0U) {
3919         SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
3920                                                   dl, MVT::i32);
3921         SDValue Ops[] = { N0.getOperand(0), Imm16,
3922                           getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32) };
3923         ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, VT, Ops));
3924         return;
3925       }
3926     }
3927 
3928     break;
3929   }
3930   case ARMISD::UMAAL: {
3931     unsigned Opc = Subtarget->isThumb() ? ARM::t2UMAAL : ARM::UMAAL;
3932     SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
3933                       N->getOperand(2), N->getOperand(3),
3934                       getAL(CurDAG, dl),
3935                       CurDAG->getRegister(0, MVT::i32) };
3936     ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, MVT::i32, MVT::i32, Ops));
3937     return;
3938   }
3939   case ARMISD::UMLAL:{
3940     if (Subtarget->isThumb()) {
3941       SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
3942                         N->getOperand(3), getAL(CurDAG, dl),
3943                         CurDAG->getRegister(0, MVT::i32)};
3944       ReplaceNode(
3945           N, CurDAG->getMachineNode(ARM::t2UMLAL, dl, MVT::i32, MVT::i32, Ops));
3946       return;
3947     }else{
3948       SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
3949                         N->getOperand(3), getAL(CurDAG, dl),
3950                         CurDAG->getRegister(0, MVT::i32),
3951                         CurDAG->getRegister(0, MVT::i32) };
3952       ReplaceNode(N, CurDAG->getMachineNode(
3953                          Subtarget->hasV6Ops() ? ARM::UMLAL : ARM::UMLALv5, dl,
3954                          MVT::i32, MVT::i32, Ops));
3955       return;
3956     }
3957   }
3958   case ARMISD::SMLAL:{
3959     if (Subtarget->isThumb()) {
3960       SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
3961                         N->getOperand(3), getAL(CurDAG, dl),
3962                         CurDAG->getRegister(0, MVT::i32)};
3963       ReplaceNode(
3964           N, CurDAG->getMachineNode(ARM::t2SMLAL, dl, MVT::i32, MVT::i32, Ops));
3965       return;
3966     }else{
3967       SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
3968                         N->getOperand(3), getAL(CurDAG, dl),
3969                         CurDAG->getRegister(0, MVT::i32),
3970                         CurDAG->getRegister(0, MVT::i32) };
3971       ReplaceNode(N, CurDAG->getMachineNode(
3972                          Subtarget->hasV6Ops() ? ARM::SMLAL : ARM::SMLALv5, dl,
3973                          MVT::i32, MVT::i32, Ops));
3974       return;
3975     }
3976   }
3977   case ARMISD::SUBE: {
3978     if (!Subtarget->hasV6Ops() || !Subtarget->hasDSP())
3979       break;
3980     // Look for a pattern to match SMMLS
3981     // (sube a, (smul_loHi a, b), (subc 0, (smul_LOhi(a, b))))
3982     if (N->getOperand(1).getOpcode() != ISD::SMUL_LOHI ||
3983         N->getOperand(2).getOpcode() != ARMISD::SUBC ||
3984         !SDValue(N, 1).use_empty())
3985       break;
3986 
3987     if (Subtarget->isThumb())
3988       assert(Subtarget->hasThumb2() &&
3989              "This pattern should not be generated for Thumb");
3990 
3991     SDValue SmulLoHi = N->getOperand(1);
3992     SDValue Subc = N->getOperand(2);
3993     auto *Zero = dyn_cast<ConstantSDNode>(Subc.getOperand(0));
3994 
3995     if (!Zero || Zero->getZExtValue() != 0 ||
3996         Subc.getOperand(1) != SmulLoHi.getValue(0) ||
3997         N->getOperand(1) != SmulLoHi.getValue(1) ||
3998         N->getOperand(2) != Subc.getValue(1))
3999       break;
4000 
4001     unsigned Opc = Subtarget->isThumb2() ? ARM::t2SMMLS : ARM::SMMLS;
4002     SDValue Ops[] = { SmulLoHi.getOperand(0), SmulLoHi.getOperand(1),
4003                       N->getOperand(0), getAL(CurDAG, dl),
4004                       CurDAG->getRegister(0, MVT::i32) };
4005     ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops));
4006     return;
4007   }
4008   case ISD::LOAD: {
4009     if (Subtarget->hasMVEIntegerOps() && tryMVEIndexedLoad(N))
4010       return;
4011     if (Subtarget->isThumb() && Subtarget->hasThumb2()) {
4012       if (tryT2IndexedLoad(N))
4013         return;
4014     } else if (Subtarget->isThumb()) {
4015       if (tryT1IndexedLoad(N))
4016         return;
4017     } else if (tryARMIndexedLoad(N))
4018       return;
4019     // Other cases are autogenerated.
4020     break;
4021   }
4022   case ISD::MLOAD:
4023     if (Subtarget->hasMVEIntegerOps() && tryMVEIndexedLoad(N))
4024       return;
4025     // Other cases are autogenerated.
4026     break;
4027   case ARMISD::WLSSETUP: {
4028     SDNode *New = CurDAG->getMachineNode(ARM::t2WhileLoopSetup, dl, MVT::i32,
4029                                          N->getOperand(0));
4030     ReplaceUses(N, New);
4031     CurDAG->RemoveDeadNode(N);
4032     return;
4033   }
4034   case ARMISD::WLS: {
4035     SDNode *New = CurDAG->getMachineNode(ARM::t2WhileLoopStart, dl, MVT::Other,
4036                                          N->getOperand(1), N->getOperand(2),
4037                                          N->getOperand(0));
4038     ReplaceUses(N, New);
4039     CurDAG->RemoveDeadNode(N);
4040     return;
4041   }
4042   case ARMISD::LE: {
4043     SDValue Ops[] = { N->getOperand(1),
4044                       N->getOperand(2),
4045                       N->getOperand(0) };
4046     unsigned Opc = ARM::t2LoopEnd;
4047     SDNode *New = CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops);
4048     ReplaceUses(N, New);
4049     CurDAG->RemoveDeadNode(N);
4050     return;
4051   }
4052   case ARMISD::LDRD: {
4053     if (Subtarget->isThumb2())
4054       break; // TableGen handles isel in this case.
4055     SDValue Base, RegOffset, ImmOffset;
4056     const SDValue &Chain = N->getOperand(0);
4057     const SDValue &Addr = N->getOperand(1);
4058     SelectAddrMode3(Addr, Base, RegOffset, ImmOffset);
4059     if (RegOffset != CurDAG->getRegister(0, MVT::i32)) {
4060       // The register-offset variant of LDRD mandates that the register
4061       // allocated to RegOffset is not reused in any of the remaining operands.
4062       // This restriction is currently not enforced. Therefore emitting this
4063       // variant is explicitly avoided.
4064       Base = Addr;
4065       RegOffset = CurDAG->getRegister(0, MVT::i32);
4066     }
4067     SDValue Ops[] = {Base, RegOffset, ImmOffset, Chain};
4068     SDNode *New = CurDAG->getMachineNode(ARM::LOADDUAL, dl,
4069                                          {MVT::Untyped, MVT::Other}, Ops);
4070     SDValue Lo = CurDAG->getTargetExtractSubreg(ARM::gsub_0, dl, MVT::i32,
4071                                                 SDValue(New, 0));
4072     SDValue Hi = CurDAG->getTargetExtractSubreg(ARM::gsub_1, dl, MVT::i32,
4073                                                 SDValue(New, 0));
4074     transferMemOperands(N, New);
4075     ReplaceUses(SDValue(N, 0), Lo);
4076     ReplaceUses(SDValue(N, 1), Hi);
4077     ReplaceUses(SDValue(N, 2), SDValue(New, 1));
4078     CurDAG->RemoveDeadNode(N);
4079     return;
4080   }
4081   case ARMISD::STRD: {
4082     if (Subtarget->isThumb2())
4083       break; // TableGen handles isel in this case.
4084     SDValue Base, RegOffset, ImmOffset;
4085     const SDValue &Chain = N->getOperand(0);
4086     const SDValue &Addr = N->getOperand(3);
4087     SelectAddrMode3(Addr, Base, RegOffset, ImmOffset);
4088     if (RegOffset != CurDAG->getRegister(0, MVT::i32)) {
4089       // The register-offset variant of STRD mandates that the register
4090       // allocated to RegOffset is not reused in any of the remaining operands.
4091       // This restriction is currently not enforced. Therefore emitting this
4092       // variant is explicitly avoided.
4093       Base = Addr;
4094       RegOffset = CurDAG->getRegister(0, MVT::i32);
4095     }
4096     SDNode *RegPair =
4097         createGPRPairNode(MVT::Untyped, N->getOperand(1), N->getOperand(2));
4098     SDValue Ops[] = {SDValue(RegPair, 0), Base, RegOffset, ImmOffset, Chain};
4099     SDNode *New = CurDAG->getMachineNode(ARM::STOREDUAL, dl, MVT::Other, Ops);
4100     transferMemOperands(N, New);
4101     ReplaceUses(SDValue(N, 0), SDValue(New, 0));
4102     CurDAG->RemoveDeadNode(N);
4103     return;
4104   }
4105   case ARMISD::LOOP_DEC: {
4106     SDValue Ops[] = { N->getOperand(1),
4107                       N->getOperand(2),
4108                       N->getOperand(0) };
4109     SDNode *Dec =
4110       CurDAG->getMachineNode(ARM::t2LoopDec, dl,
4111                              CurDAG->getVTList(MVT::i32, MVT::Other), Ops);
4112     ReplaceUses(N, Dec);
4113     CurDAG->RemoveDeadNode(N);
4114     return;
4115   }
4116   case ARMISD::BRCOND: {
4117     // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
4118     // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
4119     // Pattern complexity = 6  cost = 1  size = 0
4120 
4121     // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
4122     // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
4123     // Pattern complexity = 6  cost = 1  size = 0
4124 
4125     // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
4126     // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
4127     // Pattern complexity = 6  cost = 1  size = 0
4128 
4129     unsigned Opc = Subtarget->isThumb() ?
4130       ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
4131     SDValue Chain = N->getOperand(0);
4132     SDValue N1 = N->getOperand(1);
4133     SDValue N2 = N->getOperand(2);
4134     SDValue N3 = N->getOperand(3);
4135     SDValue InFlag = N->getOperand(4);
4136     assert(N1.getOpcode() == ISD::BasicBlock);
4137     assert(N2.getOpcode() == ISD::Constant);
4138     assert(N3.getOpcode() == ISD::Register);
4139 
4140     unsigned CC = (unsigned) cast<ConstantSDNode>(N2)->getZExtValue();
4141 
4142     if (InFlag.getOpcode() == ARMISD::CMPZ) {
4143       if (InFlag.getOperand(0).getOpcode() == ISD::INTRINSIC_W_CHAIN) {
4144         SDValue Int = InFlag.getOperand(0);
4145         uint64_t ID = cast<ConstantSDNode>(Int->getOperand(1))->getZExtValue();
4146 
4147         // Handle low-overhead loops.
4148         if (ID == Intrinsic::loop_decrement_reg) {
4149           SDValue Elements = Int.getOperand(2);
4150           SDValue Size = CurDAG->getTargetConstant(
4151             cast<ConstantSDNode>(Int.getOperand(3))->getZExtValue(), dl,
4152                                  MVT::i32);
4153 
4154           SDValue Args[] = { Elements, Size, Int.getOperand(0) };
4155           SDNode *LoopDec =
4156             CurDAG->getMachineNode(ARM::t2LoopDec, dl,
4157                                    CurDAG->getVTList(MVT::i32, MVT::Other),
4158                                    Args);
4159           ReplaceUses(Int.getNode(), LoopDec);
4160 
4161           SDValue EndArgs[] = { SDValue(LoopDec, 0), N1, Chain };
4162           SDNode *LoopEnd =
4163             CurDAG->getMachineNode(ARM::t2LoopEnd, dl, MVT::Other, EndArgs);
4164 
4165           ReplaceUses(N, LoopEnd);
4166           CurDAG->RemoveDeadNode(N);
4167           CurDAG->RemoveDeadNode(InFlag.getNode());
4168           CurDAG->RemoveDeadNode(Int.getNode());
4169           return;
4170         }
4171       }
4172 
4173       bool SwitchEQNEToPLMI;
4174       SelectCMPZ(InFlag.getNode(), SwitchEQNEToPLMI);
4175       InFlag = N->getOperand(4);
4176 
4177       if (SwitchEQNEToPLMI) {
4178         switch ((ARMCC::CondCodes)CC) {
4179         default: llvm_unreachable("CMPZ must be either NE or EQ!");
4180         case ARMCC::NE:
4181           CC = (unsigned)ARMCC::MI;
4182           break;
4183         case ARMCC::EQ:
4184           CC = (unsigned)ARMCC::PL;
4185           break;
4186         }
4187       }
4188     }
4189 
4190     SDValue Tmp2 = CurDAG->getTargetConstant(CC, dl, MVT::i32);
4191     SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
4192     SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
4193                                              MVT::Glue, Ops);
4194     Chain = SDValue(ResNode, 0);
4195     if (N->getNumValues() == 2) {
4196       InFlag = SDValue(ResNode, 1);
4197       ReplaceUses(SDValue(N, 1), InFlag);
4198     }
4199     ReplaceUses(SDValue(N, 0),
4200                 SDValue(Chain.getNode(), Chain.getResNo()));
4201     CurDAG->RemoveDeadNode(N);
4202     return;
4203   }
4204 
4205   case ARMISD::CMPZ: {
4206     // select (CMPZ X, #-C) -> (CMPZ (ADDS X, #C), #0)
4207     //   This allows us to avoid materializing the expensive negative constant.
4208     //   The CMPZ #0 is useless and will be peepholed away but we need to keep it
4209     //   for its glue output.
4210     SDValue X = N->getOperand(0);
4211     auto *C = dyn_cast<ConstantSDNode>(N->getOperand(1).getNode());
4212     if (C && C->getSExtValue() < 0 && Subtarget->isThumb()) {
4213       int64_t Addend = -C->getSExtValue();
4214 
4215       SDNode *Add = nullptr;
4216       // ADDS can be better than CMN if the immediate fits in a
4217       // 16-bit ADDS, which means either [0,256) for tADDi8 or [0,8) for tADDi3.
4218       // Outside that range we can just use a CMN which is 32-bit but has a
4219       // 12-bit immediate range.
4220       if (Addend < 1<<8) {
4221         if (Subtarget->isThumb2()) {
4222           SDValue Ops[] = { X, CurDAG->getTargetConstant(Addend, dl, MVT::i32),
4223                             getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32),
4224                             CurDAG->getRegister(0, MVT::i32) };
4225           Add = CurDAG->getMachineNode(ARM::t2ADDri, dl, MVT::i32, Ops);
4226         } else {
4227           unsigned Opc = (Addend < 1<<3) ? ARM::tADDi3 : ARM::tADDi8;
4228           SDValue Ops[] = {CurDAG->getRegister(ARM::CPSR, MVT::i32), X,
4229                            CurDAG->getTargetConstant(Addend, dl, MVT::i32),
4230                            getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32)};
4231           Add = CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops);
4232         }
4233       }
4234       if (Add) {
4235         SDValue Ops2[] = {SDValue(Add, 0), CurDAG->getConstant(0, dl, MVT::i32)};
4236         CurDAG->MorphNodeTo(N, ARMISD::CMPZ, CurDAG->getVTList(MVT::Glue), Ops2);
4237       }
4238     }
4239     // Other cases are autogenerated.
4240     break;
4241   }
4242 
4243   case ARMISD::CMOV: {
4244     SDValue InFlag = N->getOperand(4);
4245 
4246     if (InFlag.getOpcode() == ARMISD::CMPZ) {
4247       bool SwitchEQNEToPLMI;
4248       SelectCMPZ(InFlag.getNode(), SwitchEQNEToPLMI);
4249 
4250       if (SwitchEQNEToPLMI) {
4251         SDValue ARMcc = N->getOperand(2);
4252         ARMCC::CondCodes CC =
4253           (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
4254 
4255         switch (CC) {
4256         default: llvm_unreachable("CMPZ must be either NE or EQ!");
4257         case ARMCC::NE:
4258           CC = ARMCC::MI;
4259           break;
4260         case ARMCC::EQ:
4261           CC = ARMCC::PL;
4262           break;
4263         }
4264         SDValue NewARMcc = CurDAG->getConstant((unsigned)CC, dl, MVT::i32);
4265         SDValue Ops[] = {N->getOperand(0), N->getOperand(1), NewARMcc,
4266                          N->getOperand(3), N->getOperand(4)};
4267         CurDAG->MorphNodeTo(N, ARMISD::CMOV, N->getVTList(), Ops);
4268       }
4269 
4270     }
4271     // Other cases are autogenerated.
4272     break;
4273   }
4274   case ARMISD::VZIP: {
4275     EVT VT = N->getValueType(0);
4276     // vzip.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm.
4277     unsigned Opc64[] = {ARM::VZIPd8, ARM::VZIPd16, ARM::VTRNd32};
4278     unsigned Opc128[] = {ARM::VZIPq8, ARM::VZIPq16, ARM::VZIPq32};
4279     unsigned Opc = getVectorShuffleOpcode(VT, Opc64, Opc128);
4280     SDValue Pred = getAL(CurDAG, dl);
4281     SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
4282     SDValue Ops[] = {N->getOperand(0), N->getOperand(1), Pred, PredReg};
4283     ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, VT, VT, Ops));
4284     return;
4285   }
4286   case ARMISD::VUZP: {
4287     EVT VT = N->getValueType(0);
4288     // vuzp.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm.
4289     unsigned Opc64[] = {ARM::VUZPd8, ARM::VUZPd16, ARM::VTRNd32};
4290     unsigned Opc128[] = {ARM::VUZPq8, ARM::VUZPq16, ARM::VUZPq32};
4291     unsigned Opc = getVectorShuffleOpcode(VT, Opc64, Opc128);
4292     SDValue Pred = getAL(CurDAG, dl);
4293     SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
4294     SDValue Ops[] = {N->getOperand(0), N->getOperand(1), Pred, PredReg};
4295     ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, VT, VT, Ops));
4296     return;
4297   }
4298   case ARMISD::VTRN: {
4299     EVT VT = N->getValueType(0);
4300     unsigned Opc64[] = {ARM::VTRNd8, ARM::VTRNd16, ARM::VTRNd32};
4301     unsigned Opc128[] = {ARM::VTRNq8, ARM::VTRNq16, ARM::VTRNq32};
4302     unsigned Opc = getVectorShuffleOpcode(VT, Opc64, Opc128);
4303     SDValue Pred = getAL(CurDAG, dl);
4304     SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
4305     SDValue Ops[] = {N->getOperand(0), N->getOperand(1), Pred, PredReg};
4306     ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, VT, VT, Ops));
4307     return;
4308   }
4309   case ARMISD::BUILD_VECTOR: {
4310     EVT VecVT = N->getValueType(0);
4311     EVT EltVT = VecVT.getVectorElementType();
4312     unsigned NumElts = VecVT.getVectorNumElements();
4313     if (EltVT == MVT::f64) {
4314       assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
4315       ReplaceNode(
4316           N, createDRegPairNode(VecVT, N->getOperand(0), N->getOperand(1)));
4317       return;
4318     }
4319     assert(EltVT == MVT::f32 && "unexpected type for BUILD_VECTOR");
4320     if (NumElts == 2) {
4321       ReplaceNode(
4322           N, createSRegPairNode(VecVT, N->getOperand(0), N->getOperand(1)));
4323       return;
4324     }
4325     assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
4326     ReplaceNode(N,
4327                 createQuadSRegsNode(VecVT, N->getOperand(0), N->getOperand(1),
4328                                     N->getOperand(2), N->getOperand(3)));
4329     return;
4330   }
4331 
4332   case ARMISD::VLD1DUP: {
4333     static const uint16_t DOpcodes[] = { ARM::VLD1DUPd8, ARM::VLD1DUPd16,
4334                                          ARM::VLD1DUPd32 };
4335     static const uint16_t QOpcodes[] = { ARM::VLD1DUPq8, ARM::VLD1DUPq16,
4336                                          ARM::VLD1DUPq32 };
4337     SelectVLDDup(N, /* IsIntrinsic= */ false, false, 1, DOpcodes, QOpcodes);
4338     return;
4339   }
4340 
4341   case ARMISD::VLD2DUP: {
4342     static const uint16_t Opcodes[] = { ARM::VLD2DUPd8, ARM::VLD2DUPd16,
4343                                         ARM::VLD2DUPd32 };
4344     SelectVLDDup(N, /* IsIntrinsic= */ false, false, 2, Opcodes);
4345     return;
4346   }
4347 
4348   case ARMISD::VLD3DUP: {
4349     static const uint16_t Opcodes[] = { ARM::VLD3DUPd8Pseudo,
4350                                         ARM::VLD3DUPd16Pseudo,
4351                                         ARM::VLD3DUPd32Pseudo };
4352     SelectVLDDup(N, /* IsIntrinsic= */ false, false, 3, Opcodes);
4353     return;
4354   }
4355 
4356   case ARMISD::VLD4DUP: {
4357     static const uint16_t Opcodes[] = { ARM::VLD4DUPd8Pseudo,
4358                                         ARM::VLD4DUPd16Pseudo,
4359                                         ARM::VLD4DUPd32Pseudo };
4360     SelectVLDDup(N, /* IsIntrinsic= */ false, false, 4, Opcodes);
4361     return;
4362   }
4363 
4364   case ARMISD::VLD1DUP_UPD: {
4365     static const uint16_t DOpcodes[] = { ARM::VLD1DUPd8wb_fixed,
4366                                          ARM::VLD1DUPd16wb_fixed,
4367                                          ARM::VLD1DUPd32wb_fixed };
4368     static const uint16_t QOpcodes[] = { ARM::VLD1DUPq8wb_fixed,
4369                                          ARM::VLD1DUPq16wb_fixed,
4370                                          ARM::VLD1DUPq32wb_fixed };
4371     SelectVLDDup(N, /* IsIntrinsic= */ false, true, 1, DOpcodes, QOpcodes);
4372     return;
4373   }
4374 
4375   case ARMISD::VLD2DUP_UPD: {
4376     static const uint16_t DOpcodes[] = { ARM::VLD2DUPd8wb_fixed,
4377                                          ARM::VLD2DUPd16wb_fixed,
4378                                          ARM::VLD2DUPd32wb_fixed,
4379                                          ARM::VLD1q64wb_fixed };
4380     static const uint16_t QOpcodes0[] = { ARM::VLD2DUPq8EvenPseudo,
4381                                           ARM::VLD2DUPq16EvenPseudo,
4382                                           ARM::VLD2DUPq32EvenPseudo };
4383     static const uint16_t QOpcodes1[] = { ARM::VLD2DUPq8OddPseudoWB_fixed,
4384                                           ARM::VLD2DUPq16OddPseudoWB_fixed,
4385                                           ARM::VLD2DUPq32OddPseudoWB_fixed };
4386     SelectVLDDup(N, /* IsIntrinsic= */ false, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
4387     return;
4388   }
4389 
4390   case ARMISD::VLD3DUP_UPD: {
4391     static const uint16_t DOpcodes[] = { ARM::VLD3DUPd8Pseudo_UPD,
4392                                          ARM::VLD3DUPd16Pseudo_UPD,
4393                                          ARM::VLD3DUPd32Pseudo_UPD,
4394                                          ARM::VLD1d64TPseudoWB_fixed };
4395     static const uint16_t QOpcodes0[] = { ARM::VLD3DUPq8EvenPseudo,
4396                                           ARM::VLD3DUPq16EvenPseudo,
4397                                           ARM::VLD3DUPq32EvenPseudo };
4398     static const uint16_t QOpcodes1[] = { ARM::VLD3DUPq8OddPseudo_UPD,
4399                                           ARM::VLD3DUPq16OddPseudo_UPD,
4400                                           ARM::VLD3DUPq32OddPseudo_UPD };
4401     SelectVLDDup(N, /* IsIntrinsic= */ false, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
4402     return;
4403   }
4404 
4405   case ARMISD::VLD4DUP_UPD: {
4406     static const uint16_t DOpcodes[] = { ARM::VLD4DUPd8Pseudo_UPD,
4407                                          ARM::VLD4DUPd16Pseudo_UPD,
4408                                          ARM::VLD4DUPd32Pseudo_UPD,
4409                                          ARM::VLD1d64QPseudoWB_fixed };
4410     static const uint16_t QOpcodes0[] = { ARM::VLD4DUPq8EvenPseudo,
4411                                           ARM::VLD4DUPq16EvenPseudo,
4412                                           ARM::VLD4DUPq32EvenPseudo };
4413     static const uint16_t QOpcodes1[] = { ARM::VLD4DUPq8OddPseudo_UPD,
4414                                           ARM::VLD4DUPq16OddPseudo_UPD,
4415                                           ARM::VLD4DUPq32OddPseudo_UPD };
4416     SelectVLDDup(N, /* IsIntrinsic= */ false, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
4417     return;
4418   }
4419 
4420   case ARMISD::VLD1_UPD: {
4421     static const uint16_t DOpcodes[] = { ARM::VLD1d8wb_fixed,
4422                                          ARM::VLD1d16wb_fixed,
4423                                          ARM::VLD1d32wb_fixed,
4424                                          ARM::VLD1d64wb_fixed };
4425     static const uint16_t QOpcodes[] = { ARM::VLD1q8wb_fixed,
4426                                          ARM::VLD1q16wb_fixed,
4427                                          ARM::VLD1q32wb_fixed,
4428                                          ARM::VLD1q64wb_fixed };
4429     SelectVLD(N, true, 1, DOpcodes, QOpcodes, nullptr);
4430     return;
4431   }
4432 
4433   case ARMISD::VLD2_UPD: {
4434     if (Subtarget->hasNEON()) {
4435       static const uint16_t DOpcodes[] = {
4436           ARM::VLD2d8wb_fixed, ARM::VLD2d16wb_fixed, ARM::VLD2d32wb_fixed,
4437           ARM::VLD1q64wb_fixed};
4438       static const uint16_t QOpcodes[] = {ARM::VLD2q8PseudoWB_fixed,
4439                                           ARM::VLD2q16PseudoWB_fixed,
4440                                           ARM::VLD2q32PseudoWB_fixed};
4441       SelectVLD(N, true, 2, DOpcodes, QOpcodes, nullptr);
4442     } else {
4443       static const uint16_t Opcodes8[] = {ARM::MVE_VLD20_8,
4444                                           ARM::MVE_VLD21_8_wb};
4445       static const uint16_t Opcodes16[] = {ARM::MVE_VLD20_16,
4446                                            ARM::MVE_VLD21_16_wb};
4447       static const uint16_t Opcodes32[] = {ARM::MVE_VLD20_32,
4448                                            ARM::MVE_VLD21_32_wb};
4449       static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32};
4450       SelectMVE_VLD(N, 2, Opcodes, true);
4451     }
4452     return;
4453   }
4454 
4455   case ARMISD::VLD3_UPD: {
4456     static const uint16_t DOpcodes[] = { ARM::VLD3d8Pseudo_UPD,
4457                                          ARM::VLD3d16Pseudo_UPD,
4458                                          ARM::VLD3d32Pseudo_UPD,
4459                                          ARM::VLD1d64TPseudoWB_fixed};
4460     static const uint16_t QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD,
4461                                           ARM::VLD3q16Pseudo_UPD,
4462                                           ARM::VLD3q32Pseudo_UPD };
4463     static const uint16_t QOpcodes1[] = { ARM::VLD3q8oddPseudo_UPD,
4464                                           ARM::VLD3q16oddPseudo_UPD,
4465                                           ARM::VLD3q32oddPseudo_UPD };
4466     SelectVLD(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
4467     return;
4468   }
4469 
4470   case ARMISD::VLD4_UPD: {
4471     if (Subtarget->hasNEON()) {
4472       static const uint16_t DOpcodes[] = {
4473           ARM::VLD4d8Pseudo_UPD, ARM::VLD4d16Pseudo_UPD, ARM::VLD4d32Pseudo_UPD,
4474           ARM::VLD1d64QPseudoWB_fixed};
4475       static const uint16_t QOpcodes0[] = {ARM::VLD4q8Pseudo_UPD,
4476                                            ARM::VLD4q16Pseudo_UPD,
4477                                            ARM::VLD4q32Pseudo_UPD};
4478       static const uint16_t QOpcodes1[] = {ARM::VLD4q8oddPseudo_UPD,
4479                                            ARM::VLD4q16oddPseudo_UPD,
4480                                            ARM::VLD4q32oddPseudo_UPD};
4481       SelectVLD(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
4482     } else {
4483       static const uint16_t Opcodes8[] = {ARM::MVE_VLD40_8, ARM::MVE_VLD41_8,
4484                                           ARM::MVE_VLD42_8,
4485                                           ARM::MVE_VLD43_8_wb};
4486       static const uint16_t Opcodes16[] = {ARM::MVE_VLD40_16, ARM::MVE_VLD41_16,
4487                                            ARM::MVE_VLD42_16,
4488                                            ARM::MVE_VLD43_16_wb};
4489       static const uint16_t Opcodes32[] = {ARM::MVE_VLD40_32, ARM::MVE_VLD41_32,
4490                                            ARM::MVE_VLD42_32,
4491                                            ARM::MVE_VLD43_32_wb};
4492       static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32};
4493       SelectMVE_VLD(N, 4, Opcodes, true);
4494     }
4495     return;
4496   }
4497 
4498   case ARMISD::VLD1x2_UPD: {
4499     if (Subtarget->hasNEON()) {
4500       static const uint16_t DOpcodes[] = {
4501           ARM::VLD1q8wb_fixed, ARM::VLD1q16wb_fixed, ARM::VLD1q32wb_fixed,
4502           ARM::VLD1q64wb_fixed};
4503       static const uint16_t QOpcodes[] = {
4504           ARM::VLD1d8QPseudoWB_fixed, ARM::VLD1d16QPseudoWB_fixed,
4505           ARM::VLD1d32QPseudoWB_fixed, ARM::VLD1d64QPseudoWB_fixed};
4506       SelectVLD(N, true, 2, DOpcodes, QOpcodes, nullptr);
4507       return;
4508     }
4509     break;
4510   }
4511 
4512   case ARMISD::VLD1x3_UPD: {
4513     if (Subtarget->hasNEON()) {
4514       static const uint16_t DOpcodes[] = {
4515           ARM::VLD1d8TPseudoWB_fixed, ARM::VLD1d16TPseudoWB_fixed,
4516           ARM::VLD1d32TPseudoWB_fixed, ARM::VLD1d64TPseudoWB_fixed};
4517       static const uint16_t QOpcodes0[] = {
4518           ARM::VLD1q8LowTPseudo_UPD, ARM::VLD1q16LowTPseudo_UPD,
4519           ARM::VLD1q32LowTPseudo_UPD, ARM::VLD1q64LowTPseudo_UPD};
4520       static const uint16_t QOpcodes1[] = {
4521           ARM::VLD1q8HighTPseudo_UPD, ARM::VLD1q16HighTPseudo_UPD,
4522           ARM::VLD1q32HighTPseudo_UPD, ARM::VLD1q64HighTPseudo_UPD};
4523       SelectVLD(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
4524       return;
4525     }
4526     break;
4527   }
4528 
4529   case ARMISD::VLD1x4_UPD: {
4530     if (Subtarget->hasNEON()) {
4531       static const uint16_t DOpcodes[] = {
4532           ARM::VLD1d8QPseudoWB_fixed, ARM::VLD1d16QPseudoWB_fixed,
4533           ARM::VLD1d32QPseudoWB_fixed, ARM::VLD1d64QPseudoWB_fixed};
4534       static const uint16_t QOpcodes0[] = {
4535           ARM::VLD1q8LowQPseudo_UPD, ARM::VLD1q16LowQPseudo_UPD,
4536           ARM::VLD1q32LowQPseudo_UPD, ARM::VLD1q64LowQPseudo_UPD};
4537       static const uint16_t QOpcodes1[] = {
4538           ARM::VLD1q8HighQPseudo_UPD, ARM::VLD1q16HighQPseudo_UPD,
4539           ARM::VLD1q32HighQPseudo_UPD, ARM::VLD1q64HighQPseudo_UPD};
4540       SelectVLD(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
4541       return;
4542     }
4543     break;
4544   }
4545 
4546   case ARMISD::VLD2LN_UPD: {
4547     static const uint16_t DOpcodes[] = { ARM::VLD2LNd8Pseudo_UPD,
4548                                          ARM::VLD2LNd16Pseudo_UPD,
4549                                          ARM::VLD2LNd32Pseudo_UPD };
4550     static const uint16_t QOpcodes[] = { ARM::VLD2LNq16Pseudo_UPD,
4551                                          ARM::VLD2LNq32Pseudo_UPD };
4552     SelectVLDSTLane(N, true, true, 2, DOpcodes, QOpcodes);
4553     return;
4554   }
4555 
4556   case ARMISD::VLD3LN_UPD: {
4557     static const uint16_t DOpcodes[] = { ARM::VLD3LNd8Pseudo_UPD,
4558                                          ARM::VLD3LNd16Pseudo_UPD,
4559                                          ARM::VLD3LNd32Pseudo_UPD };
4560     static const uint16_t QOpcodes[] = { ARM::VLD3LNq16Pseudo_UPD,
4561                                          ARM::VLD3LNq32Pseudo_UPD };
4562     SelectVLDSTLane(N, true, true, 3, DOpcodes, QOpcodes);
4563     return;
4564   }
4565 
4566   case ARMISD::VLD4LN_UPD: {
4567     static const uint16_t DOpcodes[] = { ARM::VLD4LNd8Pseudo_UPD,
4568                                          ARM::VLD4LNd16Pseudo_UPD,
4569                                          ARM::VLD4LNd32Pseudo_UPD };
4570     static const uint16_t QOpcodes[] = { ARM::VLD4LNq16Pseudo_UPD,
4571                                          ARM::VLD4LNq32Pseudo_UPD };
4572     SelectVLDSTLane(N, true, true, 4, DOpcodes, QOpcodes);
4573     return;
4574   }
4575 
4576   case ARMISD::VST1_UPD: {
4577     static const uint16_t DOpcodes[] = { ARM::VST1d8wb_fixed,
4578                                          ARM::VST1d16wb_fixed,
4579                                          ARM::VST1d32wb_fixed,
4580                                          ARM::VST1d64wb_fixed };
4581     static const uint16_t QOpcodes[] = { ARM::VST1q8wb_fixed,
4582                                          ARM::VST1q16wb_fixed,
4583                                          ARM::VST1q32wb_fixed,
4584                                          ARM::VST1q64wb_fixed };
4585     SelectVST(N, true, 1, DOpcodes, QOpcodes, nullptr);
4586     return;
4587   }
4588 
4589   case ARMISD::VST2_UPD: {
4590     if (Subtarget->hasNEON()) {
4591       static const uint16_t DOpcodes[] = {
4592           ARM::VST2d8wb_fixed, ARM::VST2d16wb_fixed, ARM::VST2d32wb_fixed,
4593           ARM::VST1q64wb_fixed};
4594       static const uint16_t QOpcodes[] = {ARM::VST2q8PseudoWB_fixed,
4595                                           ARM::VST2q16PseudoWB_fixed,
4596                                           ARM::VST2q32PseudoWB_fixed};
4597       SelectVST(N, true, 2, DOpcodes, QOpcodes, nullptr);
4598       return;
4599     }
4600     break;
4601   }
4602 
4603   case ARMISD::VST3_UPD: {
4604     static const uint16_t DOpcodes[] = { ARM::VST3d8Pseudo_UPD,
4605                                          ARM::VST3d16Pseudo_UPD,
4606                                          ARM::VST3d32Pseudo_UPD,
4607                                          ARM::VST1d64TPseudoWB_fixed};
4608     static const uint16_t QOpcodes0[] = { ARM::VST3q8Pseudo_UPD,
4609                                           ARM::VST3q16Pseudo_UPD,
4610                                           ARM::VST3q32Pseudo_UPD };
4611     static const uint16_t QOpcodes1[] = { ARM::VST3q8oddPseudo_UPD,
4612                                           ARM::VST3q16oddPseudo_UPD,
4613                                           ARM::VST3q32oddPseudo_UPD };
4614     SelectVST(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
4615     return;
4616   }
4617 
4618   case ARMISD::VST4_UPD: {
4619     if (Subtarget->hasNEON()) {
4620       static const uint16_t DOpcodes[] = {
4621           ARM::VST4d8Pseudo_UPD, ARM::VST4d16Pseudo_UPD, ARM::VST4d32Pseudo_UPD,
4622           ARM::VST1d64QPseudoWB_fixed};
4623       static const uint16_t QOpcodes0[] = {ARM::VST4q8Pseudo_UPD,
4624                                            ARM::VST4q16Pseudo_UPD,
4625                                            ARM::VST4q32Pseudo_UPD};
4626       static const uint16_t QOpcodes1[] = {ARM::VST4q8oddPseudo_UPD,
4627                                            ARM::VST4q16oddPseudo_UPD,
4628                                            ARM::VST4q32oddPseudo_UPD};
4629       SelectVST(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
4630       return;
4631     }
4632     break;
4633   }
4634 
4635   case ARMISD::VST1x2_UPD: {
4636     if (Subtarget->hasNEON()) {
4637       static const uint16_t DOpcodes[] = { ARM::VST1q8wb_fixed,
4638                                            ARM::VST1q16wb_fixed,
4639                                            ARM::VST1q32wb_fixed,
4640                                            ARM::VST1q64wb_fixed};
4641       static const uint16_t QOpcodes[] = { ARM::VST1d8QPseudoWB_fixed,
4642                                            ARM::VST1d16QPseudoWB_fixed,
4643                                            ARM::VST1d32QPseudoWB_fixed,
4644                                            ARM::VST1d64QPseudoWB_fixed };
4645       SelectVST(N, true, 2, DOpcodes, QOpcodes, nullptr);
4646       return;
4647     }
4648     break;
4649   }
4650 
4651   case ARMISD::VST1x3_UPD: {
4652     if (Subtarget->hasNEON()) {
4653       static const uint16_t DOpcodes[] = { ARM::VST1d8TPseudoWB_fixed,
4654                                            ARM::VST1d16TPseudoWB_fixed,
4655                                            ARM::VST1d32TPseudoWB_fixed,
4656                                            ARM::VST1d64TPseudoWB_fixed };
4657       static const uint16_t QOpcodes0[] = { ARM::VST1q8LowTPseudo_UPD,
4658                                             ARM::VST1q16LowTPseudo_UPD,
4659                                             ARM::VST1q32LowTPseudo_UPD,
4660                                             ARM::VST1q64LowTPseudo_UPD };
4661       static const uint16_t QOpcodes1[] = { ARM::VST1q8HighTPseudo_UPD,
4662                                             ARM::VST1q16HighTPseudo_UPD,
4663                                             ARM::VST1q32HighTPseudo_UPD,
4664                                             ARM::VST1q64HighTPseudo_UPD };
4665       SelectVST(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
4666       return;
4667     }
4668     break;
4669   }
4670 
4671   case ARMISD::VST1x4_UPD: {
4672     if (Subtarget->hasNEON()) {
4673       static const uint16_t DOpcodes[] = { ARM::VST1d8QPseudoWB_fixed,
4674                                            ARM::VST1d16QPseudoWB_fixed,
4675                                            ARM::VST1d32QPseudoWB_fixed,
4676                                            ARM::VST1d64QPseudoWB_fixed };
4677       static const uint16_t QOpcodes0[] = { ARM::VST1q8LowQPseudo_UPD,
4678                                             ARM::VST1q16LowQPseudo_UPD,
4679                                             ARM::VST1q32LowQPseudo_UPD,
4680                                             ARM::VST1q64LowQPseudo_UPD };
4681       static const uint16_t QOpcodes1[] = { ARM::VST1q8HighQPseudo_UPD,
4682                                             ARM::VST1q16HighQPseudo_UPD,
4683                                             ARM::VST1q32HighQPseudo_UPD,
4684                                             ARM::VST1q64HighQPseudo_UPD };
4685       SelectVST(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
4686       return;
4687     }
4688     break;
4689   }
4690   case ARMISD::VST2LN_UPD: {
4691     static const uint16_t DOpcodes[] = { ARM::VST2LNd8Pseudo_UPD,
4692                                          ARM::VST2LNd16Pseudo_UPD,
4693                                          ARM::VST2LNd32Pseudo_UPD };
4694     static const uint16_t QOpcodes[] = { ARM::VST2LNq16Pseudo_UPD,
4695                                          ARM::VST2LNq32Pseudo_UPD };
4696     SelectVLDSTLane(N, false, true, 2, DOpcodes, QOpcodes);
4697     return;
4698   }
4699 
4700   case ARMISD::VST3LN_UPD: {
4701     static const uint16_t DOpcodes[] = { ARM::VST3LNd8Pseudo_UPD,
4702                                          ARM::VST3LNd16Pseudo_UPD,
4703                                          ARM::VST3LNd32Pseudo_UPD };
4704     static const uint16_t QOpcodes[] = { ARM::VST3LNq16Pseudo_UPD,
4705                                          ARM::VST3LNq32Pseudo_UPD };
4706     SelectVLDSTLane(N, false, true, 3, DOpcodes, QOpcodes);
4707     return;
4708   }
4709 
4710   case ARMISD::VST4LN_UPD: {
4711     static const uint16_t DOpcodes[] = { ARM::VST4LNd8Pseudo_UPD,
4712                                          ARM::VST4LNd16Pseudo_UPD,
4713                                          ARM::VST4LNd32Pseudo_UPD };
4714     static const uint16_t QOpcodes[] = { ARM::VST4LNq16Pseudo_UPD,
4715                                          ARM::VST4LNq32Pseudo_UPD };
4716     SelectVLDSTLane(N, false, true, 4, DOpcodes, QOpcodes);
4717     return;
4718   }
4719 
4720   case ISD::INTRINSIC_VOID:
4721   case ISD::INTRINSIC_W_CHAIN: {
4722     unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
4723     switch (IntNo) {
4724     default:
4725       break;
4726 
4727     case Intrinsic::arm_mrrc:
4728     case Intrinsic::arm_mrrc2: {
4729       SDLoc dl(N);
4730       SDValue Chain = N->getOperand(0);
4731       unsigned Opc;
4732 
4733       if (Subtarget->isThumb())
4734         Opc = (IntNo == Intrinsic::arm_mrrc ? ARM::t2MRRC : ARM::t2MRRC2);
4735       else
4736         Opc = (IntNo == Intrinsic::arm_mrrc ? ARM::MRRC : ARM::MRRC2);
4737 
4738       SmallVector<SDValue, 5> Ops;
4739       Ops.push_back(getI32Imm(cast<ConstantSDNode>(N->getOperand(2))->getZExtValue(), dl)); /* coproc */
4740       Ops.push_back(getI32Imm(cast<ConstantSDNode>(N->getOperand(3))->getZExtValue(), dl)); /* opc */
4741       Ops.push_back(getI32Imm(cast<ConstantSDNode>(N->getOperand(4))->getZExtValue(), dl)); /* CRm */
4742 
4743       // The mrrc2 instruction in ARM doesn't allow predicates, the top 4 bits of the encoded
4744       // instruction will always be '1111' but it is possible in assembly language to specify
4745       // AL as a predicate to mrrc2 but it doesn't make any difference to the encoded instruction.
4746       if (Opc != ARM::MRRC2) {
4747         Ops.push_back(getAL(CurDAG, dl));
4748         Ops.push_back(CurDAG->getRegister(0, MVT::i32));
4749       }
4750 
4751       Ops.push_back(Chain);
4752 
4753       // Writes to two registers.
4754       const EVT RetType[] = {MVT::i32, MVT::i32, MVT::Other};
4755 
4756       ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, RetType, Ops));
4757       return;
4758     }
4759     case Intrinsic::arm_ldaexd:
4760     case Intrinsic::arm_ldrexd: {
4761       SDLoc dl(N);
4762       SDValue Chain = N->getOperand(0);
4763       SDValue MemAddr = N->getOperand(2);
4764       bool isThumb = Subtarget->isThumb() && Subtarget->hasV8MBaselineOps();
4765 
4766       bool IsAcquire = IntNo == Intrinsic::arm_ldaexd;
4767       unsigned NewOpc = isThumb ? (IsAcquire ? ARM::t2LDAEXD : ARM::t2LDREXD)
4768                                 : (IsAcquire ? ARM::LDAEXD : ARM::LDREXD);
4769 
4770       // arm_ldrexd returns a i64 value in {i32, i32}
4771       std::vector<EVT> ResTys;
4772       if (isThumb) {
4773         ResTys.push_back(MVT::i32);
4774         ResTys.push_back(MVT::i32);
4775       } else
4776         ResTys.push_back(MVT::Untyped);
4777       ResTys.push_back(MVT::Other);
4778 
4779       // Place arguments in the right order.
4780       SDValue Ops[] = {MemAddr, getAL(CurDAG, dl),
4781                        CurDAG->getRegister(0, MVT::i32), Chain};
4782       SDNode *Ld = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops);
4783       // Transfer memoperands.
4784       MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
4785       CurDAG->setNodeMemRefs(cast<MachineSDNode>(Ld), {MemOp});
4786 
4787       // Remap uses.
4788       SDValue OutChain = isThumb ? SDValue(Ld, 2) : SDValue(Ld, 1);
4789       if (!SDValue(N, 0).use_empty()) {
4790         SDValue Result;
4791         if (isThumb)
4792           Result = SDValue(Ld, 0);
4793         else {
4794           SDValue SubRegIdx =
4795             CurDAG->getTargetConstant(ARM::gsub_0, dl, MVT::i32);
4796           SDNode *ResNode = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
4797               dl, MVT::i32, SDValue(Ld, 0), SubRegIdx);
4798           Result = SDValue(ResNode,0);
4799         }
4800         ReplaceUses(SDValue(N, 0), Result);
4801       }
4802       if (!SDValue(N, 1).use_empty()) {
4803         SDValue Result;
4804         if (isThumb)
4805           Result = SDValue(Ld, 1);
4806         else {
4807           SDValue SubRegIdx =
4808             CurDAG->getTargetConstant(ARM::gsub_1, dl, MVT::i32);
4809           SDNode *ResNode = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
4810               dl, MVT::i32, SDValue(Ld, 0), SubRegIdx);
4811           Result = SDValue(ResNode,0);
4812         }
4813         ReplaceUses(SDValue(N, 1), Result);
4814       }
4815       ReplaceUses(SDValue(N, 2), OutChain);
4816       CurDAG->RemoveDeadNode(N);
4817       return;
4818     }
4819     case Intrinsic::arm_stlexd:
4820     case Intrinsic::arm_strexd: {
4821       SDLoc dl(N);
4822       SDValue Chain = N->getOperand(0);
4823       SDValue Val0 = N->getOperand(2);
4824       SDValue Val1 = N->getOperand(3);
4825       SDValue MemAddr = N->getOperand(4);
4826 
4827       // Store exclusive double return a i32 value which is the return status
4828       // of the issued store.
4829       const EVT ResTys[] = {MVT::i32, MVT::Other};
4830 
4831       bool isThumb = Subtarget->isThumb() && Subtarget->hasThumb2();
4832       // Place arguments in the right order.
4833       SmallVector<SDValue, 7> Ops;
4834       if (isThumb) {
4835         Ops.push_back(Val0);
4836         Ops.push_back(Val1);
4837       } else
4838         // arm_strexd uses GPRPair.
4839         Ops.push_back(SDValue(createGPRPairNode(MVT::Untyped, Val0, Val1), 0));
4840       Ops.push_back(MemAddr);
4841       Ops.push_back(getAL(CurDAG, dl));
4842       Ops.push_back(CurDAG->getRegister(0, MVT::i32));
4843       Ops.push_back(Chain);
4844 
4845       bool IsRelease = IntNo == Intrinsic::arm_stlexd;
4846       unsigned NewOpc = isThumb ? (IsRelease ? ARM::t2STLEXD : ARM::t2STREXD)
4847                                 : (IsRelease ? ARM::STLEXD : ARM::STREXD);
4848 
4849       SDNode *St = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops);
4850       // Transfer memoperands.
4851       MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
4852       CurDAG->setNodeMemRefs(cast<MachineSDNode>(St), {MemOp});
4853 
4854       ReplaceNode(N, St);
4855       return;
4856     }
4857 
4858     case Intrinsic::arm_neon_vld1: {
4859       static const uint16_t DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
4860                                            ARM::VLD1d32, ARM::VLD1d64 };
4861       static const uint16_t QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
4862                                            ARM::VLD1q32, ARM::VLD1q64};
4863       SelectVLD(N, false, 1, DOpcodes, QOpcodes, nullptr);
4864       return;
4865     }
4866 
4867     case Intrinsic::arm_neon_vld1x2: {
4868       static const uint16_t DOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
4869                                            ARM::VLD1q32, ARM::VLD1q64 };
4870       static const uint16_t QOpcodes[] = { ARM::VLD1d8QPseudo,
4871                                            ARM::VLD1d16QPseudo,
4872                                            ARM::VLD1d32QPseudo,
4873                                            ARM::VLD1d64QPseudo };
4874       SelectVLD(N, false, 2, DOpcodes, QOpcodes, nullptr);
4875       return;
4876     }
4877 
4878     case Intrinsic::arm_neon_vld1x3: {
4879       static const uint16_t DOpcodes[] = { ARM::VLD1d8TPseudo,
4880                                            ARM::VLD1d16TPseudo,
4881                                            ARM::VLD1d32TPseudo,
4882                                            ARM::VLD1d64TPseudo };
4883       static const uint16_t QOpcodes0[] = { ARM::VLD1q8LowTPseudo_UPD,
4884                                             ARM::VLD1q16LowTPseudo_UPD,
4885                                             ARM::VLD1q32LowTPseudo_UPD,
4886                                             ARM::VLD1q64LowTPseudo_UPD };
4887       static const uint16_t QOpcodes1[] = { ARM::VLD1q8HighTPseudo,
4888                                             ARM::VLD1q16HighTPseudo,
4889                                             ARM::VLD1q32HighTPseudo,
4890                                             ARM::VLD1q64HighTPseudo };
4891       SelectVLD(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
4892       return;
4893     }
4894 
4895     case Intrinsic::arm_neon_vld1x4: {
4896       static const uint16_t DOpcodes[] = { ARM::VLD1d8QPseudo,
4897                                            ARM::VLD1d16QPseudo,
4898                                            ARM::VLD1d32QPseudo,
4899                                            ARM::VLD1d64QPseudo };
4900       static const uint16_t QOpcodes0[] = { ARM::VLD1q8LowQPseudo_UPD,
4901                                             ARM::VLD1q16LowQPseudo_UPD,
4902                                             ARM::VLD1q32LowQPseudo_UPD,
4903                                             ARM::VLD1q64LowQPseudo_UPD };
4904       static const uint16_t QOpcodes1[] = { ARM::VLD1q8HighQPseudo,
4905                                             ARM::VLD1q16HighQPseudo,
4906                                             ARM::VLD1q32HighQPseudo,
4907                                             ARM::VLD1q64HighQPseudo };
4908       SelectVLD(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
4909       return;
4910     }
4911 
4912     case Intrinsic::arm_neon_vld2: {
4913       static const uint16_t DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
4914                                            ARM::VLD2d32, ARM::VLD1q64 };
4915       static const uint16_t QOpcodes[] = { ARM::VLD2q8Pseudo, ARM::VLD2q16Pseudo,
4916                                            ARM::VLD2q32Pseudo };
4917       SelectVLD(N, false, 2, DOpcodes, QOpcodes, nullptr);
4918       return;
4919     }
4920 
4921     case Intrinsic::arm_neon_vld3: {
4922       static const uint16_t DOpcodes[] = { ARM::VLD3d8Pseudo,
4923                                            ARM::VLD3d16Pseudo,
4924                                            ARM::VLD3d32Pseudo,
4925                                            ARM::VLD1d64TPseudo };
4926       static const uint16_t QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD,
4927                                             ARM::VLD3q16Pseudo_UPD,
4928                                             ARM::VLD3q32Pseudo_UPD };
4929       static const uint16_t QOpcodes1[] = { ARM::VLD3q8oddPseudo,
4930                                             ARM::VLD3q16oddPseudo,
4931                                             ARM::VLD3q32oddPseudo };
4932       SelectVLD(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
4933       return;
4934     }
4935 
4936     case Intrinsic::arm_neon_vld4: {
4937       static const uint16_t DOpcodes[] = { ARM::VLD4d8Pseudo,
4938                                            ARM::VLD4d16Pseudo,
4939                                            ARM::VLD4d32Pseudo,
4940                                            ARM::VLD1d64QPseudo };
4941       static const uint16_t QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD,
4942                                             ARM::VLD4q16Pseudo_UPD,
4943                                             ARM::VLD4q32Pseudo_UPD };
4944       static const uint16_t QOpcodes1[] = { ARM::VLD4q8oddPseudo,
4945                                             ARM::VLD4q16oddPseudo,
4946                                             ARM::VLD4q32oddPseudo };
4947       SelectVLD(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
4948       return;
4949     }
4950 
4951     case Intrinsic::arm_neon_vld2dup: {
4952       static const uint16_t DOpcodes[] = { ARM::VLD2DUPd8, ARM::VLD2DUPd16,
4953                                            ARM::VLD2DUPd32, ARM::VLD1q64 };
4954       static const uint16_t QOpcodes0[] = { ARM::VLD2DUPq8EvenPseudo,
4955                                             ARM::VLD2DUPq16EvenPseudo,
4956                                             ARM::VLD2DUPq32EvenPseudo };
4957       static const uint16_t QOpcodes1[] = { ARM::VLD2DUPq8OddPseudo,
4958                                             ARM::VLD2DUPq16OddPseudo,
4959                                             ARM::VLD2DUPq32OddPseudo };
4960       SelectVLDDup(N, /* IsIntrinsic= */ true, false, 2,
4961                    DOpcodes, QOpcodes0, QOpcodes1);
4962       return;
4963     }
4964 
4965     case Intrinsic::arm_neon_vld3dup: {
4966       static const uint16_t DOpcodes[] = { ARM::VLD3DUPd8Pseudo,
4967                                            ARM::VLD3DUPd16Pseudo,
4968                                            ARM::VLD3DUPd32Pseudo,
4969                                            ARM::VLD1d64TPseudo };
4970       static const uint16_t QOpcodes0[] = { ARM::VLD3DUPq8EvenPseudo,
4971                                             ARM::VLD3DUPq16EvenPseudo,
4972                                             ARM::VLD3DUPq32EvenPseudo };
4973       static const uint16_t QOpcodes1[] = { ARM::VLD3DUPq8OddPseudo,
4974                                             ARM::VLD3DUPq16OddPseudo,
4975                                             ARM::VLD3DUPq32OddPseudo };
4976       SelectVLDDup(N, /* IsIntrinsic= */ true, false, 3,
4977                    DOpcodes, QOpcodes0, QOpcodes1);
4978       return;
4979     }
4980 
4981     case Intrinsic::arm_neon_vld4dup: {
4982       static const uint16_t DOpcodes[] = { ARM::VLD4DUPd8Pseudo,
4983                                            ARM::VLD4DUPd16Pseudo,
4984                                            ARM::VLD4DUPd32Pseudo,
4985                                            ARM::VLD1d64QPseudo };
4986       static const uint16_t QOpcodes0[] = { ARM::VLD4DUPq8EvenPseudo,
4987                                             ARM::VLD4DUPq16EvenPseudo,
4988                                             ARM::VLD4DUPq32EvenPseudo };
4989       static const uint16_t QOpcodes1[] = { ARM::VLD4DUPq8OddPseudo,
4990                                             ARM::VLD4DUPq16OddPseudo,
4991                                             ARM::VLD4DUPq32OddPseudo };
4992       SelectVLDDup(N, /* IsIntrinsic= */ true, false, 4,
4993                    DOpcodes, QOpcodes0, QOpcodes1);
4994       return;
4995     }
4996 
4997     case Intrinsic::arm_neon_vld2lane: {
4998       static const uint16_t DOpcodes[] = { ARM::VLD2LNd8Pseudo,
4999                                            ARM::VLD2LNd16Pseudo,
5000                                            ARM::VLD2LNd32Pseudo };
5001       static const uint16_t QOpcodes[] = { ARM::VLD2LNq16Pseudo,
5002                                            ARM::VLD2LNq32Pseudo };
5003       SelectVLDSTLane(N, true, false, 2, DOpcodes, QOpcodes);
5004       return;
5005     }
5006 
5007     case Intrinsic::arm_neon_vld3lane: {
5008       static const uint16_t DOpcodes[] = { ARM::VLD3LNd8Pseudo,
5009                                            ARM::VLD3LNd16Pseudo,
5010                                            ARM::VLD3LNd32Pseudo };
5011       static const uint16_t QOpcodes[] = { ARM::VLD3LNq16Pseudo,
5012                                            ARM::VLD3LNq32Pseudo };
5013       SelectVLDSTLane(N, true, false, 3, DOpcodes, QOpcodes);
5014       return;
5015     }
5016 
5017     case Intrinsic::arm_neon_vld4lane: {
5018       static const uint16_t DOpcodes[] = { ARM::VLD4LNd8Pseudo,
5019                                            ARM::VLD4LNd16Pseudo,
5020                                            ARM::VLD4LNd32Pseudo };
5021       static const uint16_t QOpcodes[] = { ARM::VLD4LNq16Pseudo,
5022                                            ARM::VLD4LNq32Pseudo };
5023       SelectVLDSTLane(N, true, false, 4, DOpcodes, QOpcodes);
5024       return;
5025     }
5026 
5027     case Intrinsic::arm_neon_vst1: {
5028       static const uint16_t DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
5029                                            ARM::VST1d32, ARM::VST1d64 };
5030       static const uint16_t QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
5031                                            ARM::VST1q32, ARM::VST1q64 };
5032       SelectVST(N, false, 1, DOpcodes, QOpcodes, nullptr);
5033       return;
5034     }
5035 
5036     case Intrinsic::arm_neon_vst1x2: {
5037       static const uint16_t DOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
5038                                            ARM::VST1q32, ARM::VST1q64 };
5039       static const uint16_t QOpcodes[] = { ARM::VST1d8QPseudo,
5040                                            ARM::VST1d16QPseudo,
5041                                            ARM::VST1d32QPseudo,
5042                                            ARM::VST1d64QPseudo };
5043       SelectVST(N, false, 2, DOpcodes, QOpcodes, nullptr);
5044       return;
5045     }
5046 
5047     case Intrinsic::arm_neon_vst1x3: {
5048       static const uint16_t DOpcodes[] = { ARM::VST1d8TPseudo,
5049                                            ARM::VST1d16TPseudo,
5050                                            ARM::VST1d32TPseudo,
5051                                            ARM::VST1d64TPseudo };
5052       static const uint16_t QOpcodes0[] = { ARM::VST1q8LowTPseudo_UPD,
5053                                             ARM::VST1q16LowTPseudo_UPD,
5054                                             ARM::VST1q32LowTPseudo_UPD,
5055                                             ARM::VST1q64LowTPseudo_UPD };
5056       static const uint16_t QOpcodes1[] = { ARM::VST1q8HighTPseudo,
5057                                             ARM::VST1q16HighTPseudo,
5058                                             ARM::VST1q32HighTPseudo,
5059                                             ARM::VST1q64HighTPseudo };
5060       SelectVST(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
5061       return;
5062     }
5063 
5064     case Intrinsic::arm_neon_vst1x4: {
5065       static const uint16_t DOpcodes[] = { ARM::VST1d8QPseudo,
5066                                            ARM::VST1d16QPseudo,
5067                                            ARM::VST1d32QPseudo,
5068                                            ARM::VST1d64QPseudo };
5069       static const uint16_t QOpcodes0[] = { ARM::VST1q8LowQPseudo_UPD,
5070                                             ARM::VST1q16LowQPseudo_UPD,
5071                                             ARM::VST1q32LowQPseudo_UPD,
5072                                             ARM::VST1q64LowQPseudo_UPD };
5073       static const uint16_t QOpcodes1[] = { ARM::VST1q8HighQPseudo,
5074                                             ARM::VST1q16HighQPseudo,
5075                                             ARM::VST1q32HighQPseudo,
5076                                             ARM::VST1q64HighQPseudo };
5077       SelectVST(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
5078       return;
5079     }
5080 
5081     case Intrinsic::arm_neon_vst2: {
5082       static const uint16_t DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
5083                                            ARM::VST2d32, ARM::VST1q64 };
5084       static const uint16_t QOpcodes[] = { ARM::VST2q8Pseudo, ARM::VST2q16Pseudo,
5085                                            ARM::VST2q32Pseudo };
5086       SelectVST(N, false, 2, DOpcodes, QOpcodes, nullptr);
5087       return;
5088     }
5089 
5090     case Intrinsic::arm_neon_vst3: {
5091       static const uint16_t DOpcodes[] = { ARM::VST3d8Pseudo,
5092                                            ARM::VST3d16Pseudo,
5093                                            ARM::VST3d32Pseudo,
5094                                            ARM::VST1d64TPseudo };
5095       static const uint16_t QOpcodes0[] = { ARM::VST3q8Pseudo_UPD,
5096                                             ARM::VST3q16Pseudo_UPD,
5097                                             ARM::VST3q32Pseudo_UPD };
5098       static const uint16_t QOpcodes1[] = { ARM::VST3q8oddPseudo,
5099                                             ARM::VST3q16oddPseudo,
5100                                             ARM::VST3q32oddPseudo };
5101       SelectVST(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
5102       return;
5103     }
5104 
5105     case Intrinsic::arm_neon_vst4: {
5106       static const uint16_t DOpcodes[] = { ARM::VST4d8Pseudo,
5107                                            ARM::VST4d16Pseudo,
5108                                            ARM::VST4d32Pseudo,
5109                                            ARM::VST1d64QPseudo };
5110       static const uint16_t QOpcodes0[] = { ARM::VST4q8Pseudo_UPD,
5111                                             ARM::VST4q16Pseudo_UPD,
5112                                             ARM::VST4q32Pseudo_UPD };
5113       static const uint16_t QOpcodes1[] = { ARM::VST4q8oddPseudo,
5114                                             ARM::VST4q16oddPseudo,
5115                                             ARM::VST4q32oddPseudo };
5116       SelectVST(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
5117       return;
5118     }
5119 
5120     case Intrinsic::arm_neon_vst2lane: {
5121       static const uint16_t DOpcodes[] = { ARM::VST2LNd8Pseudo,
5122                                            ARM::VST2LNd16Pseudo,
5123                                            ARM::VST2LNd32Pseudo };
5124       static const uint16_t QOpcodes[] = { ARM::VST2LNq16Pseudo,
5125                                            ARM::VST2LNq32Pseudo };
5126       SelectVLDSTLane(N, false, false, 2, DOpcodes, QOpcodes);
5127       return;
5128     }
5129 
5130     case Intrinsic::arm_neon_vst3lane: {
5131       static const uint16_t DOpcodes[] = { ARM::VST3LNd8Pseudo,
5132                                            ARM::VST3LNd16Pseudo,
5133                                            ARM::VST3LNd32Pseudo };
5134       static const uint16_t QOpcodes[] = { ARM::VST3LNq16Pseudo,
5135                                            ARM::VST3LNq32Pseudo };
5136       SelectVLDSTLane(N, false, false, 3, DOpcodes, QOpcodes);
5137       return;
5138     }
5139 
5140     case Intrinsic::arm_neon_vst4lane: {
5141       static const uint16_t DOpcodes[] = { ARM::VST4LNd8Pseudo,
5142                                            ARM::VST4LNd16Pseudo,
5143                                            ARM::VST4LNd32Pseudo };
5144       static const uint16_t QOpcodes[] = { ARM::VST4LNq16Pseudo,
5145                                            ARM::VST4LNq32Pseudo };
5146       SelectVLDSTLane(N, false, false, 4, DOpcodes, QOpcodes);
5147       return;
5148     }
5149 
5150     case Intrinsic::arm_mve_vldr_gather_base_wb:
5151     case Intrinsic::arm_mve_vldr_gather_base_wb_predicated: {
5152       static const uint16_t Opcodes[] = {ARM::MVE_VLDRWU32_qi_pre,
5153                                          ARM::MVE_VLDRDU64_qi_pre};
5154       SelectMVE_WB(N, Opcodes,
5155                    IntNo == Intrinsic::arm_mve_vldr_gather_base_wb_predicated);
5156       return;
5157     }
5158 
5159     case Intrinsic::arm_mve_vld2q: {
5160       static const uint16_t Opcodes8[] = {ARM::MVE_VLD20_8, ARM::MVE_VLD21_8};
5161       static const uint16_t Opcodes16[] = {ARM::MVE_VLD20_16,
5162                                            ARM::MVE_VLD21_16};
5163       static const uint16_t Opcodes32[] = {ARM::MVE_VLD20_32,
5164                                            ARM::MVE_VLD21_32};
5165       static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32};
5166       SelectMVE_VLD(N, 2, Opcodes, false);
5167       return;
5168     }
5169 
5170     case Intrinsic::arm_mve_vld4q: {
5171       static const uint16_t Opcodes8[] = {ARM::MVE_VLD40_8, ARM::MVE_VLD41_8,
5172                                           ARM::MVE_VLD42_8, ARM::MVE_VLD43_8};
5173       static const uint16_t Opcodes16[] = {ARM::MVE_VLD40_16, ARM::MVE_VLD41_16,
5174                                            ARM::MVE_VLD42_16,
5175                                            ARM::MVE_VLD43_16};
5176       static const uint16_t Opcodes32[] = {ARM::MVE_VLD40_32, ARM::MVE_VLD41_32,
5177                                            ARM::MVE_VLD42_32,
5178                                            ARM::MVE_VLD43_32};
5179       static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32};
5180       SelectMVE_VLD(N, 4, Opcodes, false);
5181       return;
5182     }
5183     }
5184     break;
5185   }
5186 
5187   case ISD::INTRINSIC_WO_CHAIN: {
5188     unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
5189     switch (IntNo) {
5190     default:
5191       break;
5192 
5193     // Scalar f32 -> bf16
5194     case Intrinsic::arm_neon_vcvtbfp2bf: {
5195       SDLoc dl(N);
5196       const SDValue &Src = N->getOperand(1);
5197       llvm::EVT DestTy = N->getValueType(0);
5198       SDValue Pred = getAL(CurDAG, dl);
5199       SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
5200       SDValue Ops[] = { Src, Src, Pred, Reg0 };
5201       CurDAG->SelectNodeTo(N, ARM::BF16_VCVTB, DestTy, Ops);
5202       return;
5203     }
5204 
5205     // Vector v4f32 -> v4bf16
5206     case Intrinsic::arm_neon_vcvtfp2bf: {
5207       SDLoc dl(N);
5208       const SDValue &Src = N->getOperand(1);
5209       SDValue Pred = getAL(CurDAG, dl);
5210       SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
5211       SDValue Ops[] = { Src, Pred, Reg0 };
5212       CurDAG->SelectNodeTo(N, ARM::BF16_VCVT, MVT::v4bf16, Ops);
5213       return;
5214     }
5215 
5216     case Intrinsic::arm_mve_urshrl:
5217       SelectMVE_LongShift(N, ARM::MVE_URSHRL, true, false);
5218       return;
5219     case Intrinsic::arm_mve_uqshll:
5220       SelectMVE_LongShift(N, ARM::MVE_UQSHLL, true, false);
5221       return;
5222     case Intrinsic::arm_mve_srshrl:
5223       SelectMVE_LongShift(N, ARM::MVE_SRSHRL, true, false);
5224       return;
5225     case Intrinsic::arm_mve_sqshll:
5226       SelectMVE_LongShift(N, ARM::MVE_SQSHLL, true, false);
5227       return;
5228     case Intrinsic::arm_mve_uqrshll:
5229       SelectMVE_LongShift(N, ARM::MVE_UQRSHLL, false, true);
5230       return;
5231     case Intrinsic::arm_mve_sqrshrl:
5232       SelectMVE_LongShift(N, ARM::MVE_SQRSHRL, false, true);
5233       return;
5234 
5235     case Intrinsic::arm_mve_vadc:
5236     case Intrinsic::arm_mve_vadc_predicated:
5237       SelectMVE_VADCSBC(N, ARM::MVE_VADC, ARM::MVE_VADCI, true,
5238                         IntNo == Intrinsic::arm_mve_vadc_predicated);
5239       return;
5240     case Intrinsic::arm_mve_vsbc:
5241     case Intrinsic::arm_mve_vsbc_predicated:
5242       SelectMVE_VADCSBC(N, ARM::MVE_VSBC, ARM::MVE_VSBCI, true,
5243                         IntNo == Intrinsic::arm_mve_vsbc_predicated);
5244       return;
5245     case Intrinsic::arm_mve_vshlc:
5246     case Intrinsic::arm_mve_vshlc_predicated:
5247       SelectMVE_VSHLC(N, IntNo == Intrinsic::arm_mve_vshlc_predicated);
5248       return;
5249 
5250     case Intrinsic::arm_mve_vmlldava:
5251     case Intrinsic::arm_mve_vmlldava_predicated: {
5252       static const uint16_t OpcodesU[] = {
5253           ARM::MVE_VMLALDAVu16,   ARM::MVE_VMLALDAVu32,
5254           ARM::MVE_VMLALDAVau16,  ARM::MVE_VMLALDAVau32,
5255       };
5256       static const uint16_t OpcodesS[] = {
5257           ARM::MVE_VMLALDAVs16,   ARM::MVE_VMLALDAVs32,
5258           ARM::MVE_VMLALDAVas16,  ARM::MVE_VMLALDAVas32,
5259           ARM::MVE_VMLALDAVxs16,  ARM::MVE_VMLALDAVxs32,
5260           ARM::MVE_VMLALDAVaxs16, ARM::MVE_VMLALDAVaxs32,
5261           ARM::MVE_VMLSLDAVs16,   ARM::MVE_VMLSLDAVs32,
5262           ARM::MVE_VMLSLDAVas16,  ARM::MVE_VMLSLDAVas32,
5263           ARM::MVE_VMLSLDAVxs16,  ARM::MVE_VMLSLDAVxs32,
5264           ARM::MVE_VMLSLDAVaxs16, ARM::MVE_VMLSLDAVaxs32,
5265       };
5266       SelectMVE_VMLLDAV(N, IntNo == Intrinsic::arm_mve_vmlldava_predicated,
5267                         OpcodesS, OpcodesU);
5268       return;
5269     }
5270 
5271     case Intrinsic::arm_mve_vrmlldavha:
5272     case Intrinsic::arm_mve_vrmlldavha_predicated: {
5273       static const uint16_t OpcodesU[] = {
5274           ARM::MVE_VRMLALDAVHu32,  ARM::MVE_VRMLALDAVHau32,
5275       };
5276       static const uint16_t OpcodesS[] = {
5277           ARM::MVE_VRMLALDAVHs32,  ARM::MVE_VRMLALDAVHas32,
5278           ARM::MVE_VRMLALDAVHxs32, ARM::MVE_VRMLALDAVHaxs32,
5279           ARM::MVE_VRMLSLDAVHs32,  ARM::MVE_VRMLSLDAVHas32,
5280           ARM::MVE_VRMLSLDAVHxs32, ARM::MVE_VRMLSLDAVHaxs32,
5281       };
5282       SelectMVE_VRMLLDAVH(N, IntNo == Intrinsic::arm_mve_vrmlldavha_predicated,
5283                           OpcodesS, OpcodesU);
5284       return;
5285     }
5286 
5287     case Intrinsic::arm_mve_vidup:
5288     case Intrinsic::arm_mve_vidup_predicated: {
5289       static const uint16_t Opcodes[] = {
5290           ARM::MVE_VIDUPu8, ARM::MVE_VIDUPu16, ARM::MVE_VIDUPu32,
5291       };
5292       SelectMVE_VxDUP(N, Opcodes, false,
5293                       IntNo == Intrinsic::arm_mve_vidup_predicated);
5294       return;
5295     }
5296 
5297     case Intrinsic::arm_mve_vddup:
5298     case Intrinsic::arm_mve_vddup_predicated: {
5299       static const uint16_t Opcodes[] = {
5300           ARM::MVE_VDDUPu8, ARM::MVE_VDDUPu16, ARM::MVE_VDDUPu32,
5301       };
5302       SelectMVE_VxDUP(N, Opcodes, false,
5303                       IntNo == Intrinsic::arm_mve_vddup_predicated);
5304       return;
5305     }
5306 
5307     case Intrinsic::arm_mve_viwdup:
5308     case Intrinsic::arm_mve_viwdup_predicated: {
5309       static const uint16_t Opcodes[] = {
5310           ARM::MVE_VIWDUPu8, ARM::MVE_VIWDUPu16, ARM::MVE_VIWDUPu32,
5311       };
5312       SelectMVE_VxDUP(N, Opcodes, true,
5313                       IntNo == Intrinsic::arm_mve_viwdup_predicated);
5314       return;
5315     }
5316 
5317     case Intrinsic::arm_mve_vdwdup:
5318     case Intrinsic::arm_mve_vdwdup_predicated: {
5319       static const uint16_t Opcodes[] = {
5320           ARM::MVE_VDWDUPu8, ARM::MVE_VDWDUPu16, ARM::MVE_VDWDUPu32,
5321       };
5322       SelectMVE_VxDUP(N, Opcodes, true,
5323                       IntNo == Intrinsic::arm_mve_vdwdup_predicated);
5324       return;
5325     }
5326 
5327     case Intrinsic::arm_cde_cx1d:
5328     case Intrinsic::arm_cde_cx1da:
5329     case Intrinsic::arm_cde_cx2d:
5330     case Intrinsic::arm_cde_cx2da:
5331     case Intrinsic::arm_cde_cx3d:
5332     case Intrinsic::arm_cde_cx3da: {
5333       bool HasAccum = IntNo == Intrinsic::arm_cde_cx1da ||
5334                       IntNo == Intrinsic::arm_cde_cx2da ||
5335                       IntNo == Intrinsic::arm_cde_cx3da;
5336       size_t NumExtraOps;
5337       uint16_t Opcode;
5338       switch (IntNo) {
5339       case Intrinsic::arm_cde_cx1d:
5340       case Intrinsic::arm_cde_cx1da:
5341         NumExtraOps = 0;
5342         Opcode = HasAccum ? ARM::CDE_CX1DA : ARM::CDE_CX1D;
5343         break;
5344       case Intrinsic::arm_cde_cx2d:
5345       case Intrinsic::arm_cde_cx2da:
5346         NumExtraOps = 1;
5347         Opcode = HasAccum ? ARM::CDE_CX2DA : ARM::CDE_CX2D;
5348         break;
5349       case Intrinsic::arm_cde_cx3d:
5350       case Intrinsic::arm_cde_cx3da:
5351         NumExtraOps = 2;
5352         Opcode = HasAccum ? ARM::CDE_CX3DA : ARM::CDE_CX3D;
5353         break;
5354       default:
5355         llvm_unreachable("Unexpected opcode");
5356       }
5357       SelectCDE_CXxD(N, Opcode, NumExtraOps, HasAccum);
5358       return;
5359     }
5360     }
5361     break;
5362   }
5363 
5364   case ISD::ATOMIC_CMP_SWAP:
5365     SelectCMP_SWAP(N);
5366     return;
5367   }
5368 
5369   SelectCode(N);
5370 }
5371 
5372 // Inspect a register string of the form
5373 // cp<coprocessor>:<opc1>:c<CRn>:c<CRm>:<opc2> (32bit) or
5374 // cp<coprocessor>:<opc1>:c<CRm> (64bit) inspect the fields of the string
5375 // and obtain the integer operands from them, adding these operands to the
5376 // provided vector.
5377 static void getIntOperandsFromRegisterString(StringRef RegString,
5378                                              SelectionDAG *CurDAG,
5379                                              const SDLoc &DL,
5380                                              std::vector<SDValue> &Ops) {
5381   SmallVector<StringRef, 5> Fields;
5382   RegString.split(Fields, ':');
5383 
5384   if (Fields.size() > 1) {
5385     bool AllIntFields = true;
5386 
5387     for (StringRef Field : Fields) {
5388       // Need to trim out leading 'cp' characters and get the integer field.
5389       unsigned IntField;
5390       AllIntFields &= !Field.trim("CPcp").getAsInteger(10, IntField);
5391       Ops.push_back(CurDAG->getTargetConstant(IntField, DL, MVT::i32));
5392     }
5393 
5394     assert(AllIntFields &&
5395             "Unexpected non-integer value in special register string.");
5396     (void)AllIntFields;
5397   }
5398 }
5399 
5400 // Maps a Banked Register string to its mask value. The mask value returned is
5401 // for use in the MRSbanked / MSRbanked instruction nodes as the Banked Register
5402 // mask operand, which expresses which register is to be used, e.g. r8, and in
5403 // which mode it is to be used, e.g. usr. Returns -1 to signify that the string
5404 // was invalid.
5405 static inline int getBankedRegisterMask(StringRef RegString) {
5406   auto TheReg = ARMBankedReg::lookupBankedRegByName(RegString.lower());
5407   if (!TheReg)
5408      return -1;
5409   return TheReg->Encoding;
5410 }
5411 
5412 // The flags here are common to those allowed for apsr in the A class cores and
5413 // those allowed for the special registers in the M class cores. Returns a
5414 // value representing which flags were present, -1 if invalid.
5415 static inline int getMClassFlagsMask(StringRef Flags) {
5416   return StringSwitch<int>(Flags)
5417           .Case("", 0x2) // no flags means nzcvq for psr registers, and 0x2 is
5418                          // correct when flags are not permitted
5419           .Case("g", 0x1)
5420           .Case("nzcvq", 0x2)
5421           .Case("nzcvqg", 0x3)
5422           .Default(-1);
5423 }
5424 
5425 // Maps MClass special registers string to its value for use in the
5426 // t2MRS_M/t2MSR_M instruction nodes as the SYSm value operand.
5427 // Returns -1 to signify that the string was invalid.
5428 static int getMClassRegisterMask(StringRef Reg, const ARMSubtarget *Subtarget) {
5429   auto TheReg = ARMSysReg::lookupMClassSysRegByName(Reg);
5430   const FeatureBitset &FeatureBits = Subtarget->getFeatureBits();
5431   if (!TheReg || !TheReg->hasRequiredFeatures(FeatureBits))
5432     return -1;
5433   return (int)(TheReg->Encoding & 0xFFF); // SYSm value
5434 }
5435 
5436 static int getARClassRegisterMask(StringRef Reg, StringRef Flags) {
5437   // The mask operand contains the special register (R Bit) in bit 4, whether
5438   // the register is spsr (R bit is 1) or one of cpsr/apsr (R bit is 0), and
5439   // bits 3-0 contains the fields to be accessed in the special register, set by
5440   // the flags provided with the register.
5441   int Mask = 0;
5442   if (Reg == "apsr") {
5443     // The flags permitted for apsr are the same flags that are allowed in
5444     // M class registers. We get the flag value and then shift the flags into
5445     // the correct place to combine with the mask.
5446     Mask = getMClassFlagsMask(Flags);
5447     if (Mask == -1)
5448       return -1;
5449     return Mask << 2;
5450   }
5451 
5452   if (Reg != "cpsr" && Reg != "spsr") {
5453     return -1;
5454   }
5455 
5456   // This is the same as if the flags were "fc"
5457   if (Flags.empty() || Flags == "all")
5458     return Mask | 0x9;
5459 
5460   // Inspect the supplied flags string and set the bits in the mask for
5461   // the relevant and valid flags allowed for cpsr and spsr.
5462   for (char Flag : Flags) {
5463     int FlagVal;
5464     switch (Flag) {
5465       case 'c':
5466         FlagVal = 0x1;
5467         break;
5468       case 'x':
5469         FlagVal = 0x2;
5470         break;
5471       case 's':
5472         FlagVal = 0x4;
5473         break;
5474       case 'f':
5475         FlagVal = 0x8;
5476         break;
5477       default:
5478         FlagVal = 0;
5479     }
5480 
5481     // This avoids allowing strings where the same flag bit appears twice.
5482     if (!FlagVal || (Mask & FlagVal))
5483       return -1;
5484     Mask |= FlagVal;
5485   }
5486 
5487   // If the register is spsr then we need to set the R bit.
5488   if (Reg == "spsr")
5489     Mask |= 0x10;
5490 
5491   return Mask;
5492 }
5493 
5494 // Lower the read_register intrinsic to ARM specific DAG nodes
5495 // using the supplied metadata string to select the instruction node to use
5496 // and the registers/masks to construct as operands for the node.
5497 bool ARMDAGToDAGISel::tryReadRegister(SDNode *N){
5498   const auto *MD = cast<MDNodeSDNode>(N->getOperand(1));
5499   const auto *RegString = cast<MDString>(MD->getMD()->getOperand(0));
5500   bool IsThumb2 = Subtarget->isThumb2();
5501   SDLoc DL(N);
5502 
5503   std::vector<SDValue> Ops;
5504   getIntOperandsFromRegisterString(RegString->getString(), CurDAG, DL, Ops);
5505 
5506   if (!Ops.empty()) {
5507     // If the special register string was constructed of fields (as defined
5508     // in the ACLE) then need to lower to MRC node (32 bit) or
5509     // MRRC node(64 bit), we can make the distinction based on the number of
5510     // operands we have.
5511     unsigned Opcode;
5512     SmallVector<EVT, 3> ResTypes;
5513     if (Ops.size() == 5){
5514       Opcode = IsThumb2 ? ARM::t2MRC : ARM::MRC;
5515       ResTypes.append({ MVT::i32, MVT::Other });
5516     } else {
5517       assert(Ops.size() == 3 &&
5518               "Invalid number of fields in special register string.");
5519       Opcode = IsThumb2 ? ARM::t2MRRC : ARM::MRRC;
5520       ResTypes.append({ MVT::i32, MVT::i32, MVT::Other });
5521     }
5522 
5523     Ops.push_back(getAL(CurDAG, DL));
5524     Ops.push_back(CurDAG->getRegister(0, MVT::i32));
5525     Ops.push_back(N->getOperand(0));
5526     ReplaceNode(N, CurDAG->getMachineNode(Opcode, DL, ResTypes, Ops));
5527     return true;
5528   }
5529 
5530   std::string SpecialReg = RegString->getString().lower();
5531 
5532   int BankedReg = getBankedRegisterMask(SpecialReg);
5533   if (BankedReg != -1) {
5534     Ops = { CurDAG->getTargetConstant(BankedReg, DL, MVT::i32),
5535             getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5536             N->getOperand(0) };
5537     ReplaceNode(
5538         N, CurDAG->getMachineNode(IsThumb2 ? ARM::t2MRSbanked : ARM::MRSbanked,
5539                                   DL, MVT::i32, MVT::Other, Ops));
5540     return true;
5541   }
5542 
5543   // The VFP registers are read by creating SelectionDAG nodes with opcodes
5544   // corresponding to the register that is being read from. So we switch on the
5545   // string to find which opcode we need to use.
5546   unsigned Opcode = StringSwitch<unsigned>(SpecialReg)
5547                     .Case("fpscr", ARM::VMRS)
5548                     .Case("fpexc", ARM::VMRS_FPEXC)
5549                     .Case("fpsid", ARM::VMRS_FPSID)
5550                     .Case("mvfr0", ARM::VMRS_MVFR0)
5551                     .Case("mvfr1", ARM::VMRS_MVFR1)
5552                     .Case("mvfr2", ARM::VMRS_MVFR2)
5553                     .Case("fpinst", ARM::VMRS_FPINST)
5554                     .Case("fpinst2", ARM::VMRS_FPINST2)
5555                     .Default(0);
5556 
5557   // If an opcode was found then we can lower the read to a VFP instruction.
5558   if (Opcode) {
5559     if (!Subtarget->hasVFP2Base())
5560       return false;
5561     if (Opcode == ARM::VMRS_MVFR2 && !Subtarget->hasFPARMv8Base())
5562       return false;
5563 
5564     Ops = { getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5565             N->getOperand(0) };
5566     ReplaceNode(N,
5567                 CurDAG->getMachineNode(Opcode, DL, MVT::i32, MVT::Other, Ops));
5568     return true;
5569   }
5570 
5571   // If the target is M Class then need to validate that the register string
5572   // is an acceptable value, so check that a mask can be constructed from the
5573   // string.
5574   if (Subtarget->isMClass()) {
5575     int SYSmValue = getMClassRegisterMask(SpecialReg, Subtarget);
5576     if (SYSmValue == -1)
5577       return false;
5578 
5579     SDValue Ops[] = { CurDAG->getTargetConstant(SYSmValue, DL, MVT::i32),
5580                       getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5581                       N->getOperand(0) };
5582     ReplaceNode(
5583         N, CurDAG->getMachineNode(ARM::t2MRS_M, DL, MVT::i32, MVT::Other, Ops));
5584     return true;
5585   }
5586 
5587   // Here we know the target is not M Class so we need to check if it is one
5588   // of the remaining possible values which are apsr, cpsr or spsr.
5589   if (SpecialReg == "apsr" || SpecialReg == "cpsr") {
5590     Ops = { getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5591             N->getOperand(0) };
5592     ReplaceNode(N, CurDAG->getMachineNode(IsThumb2 ? ARM::t2MRS_AR : ARM::MRS,
5593                                           DL, MVT::i32, MVT::Other, Ops));
5594     return true;
5595   }
5596 
5597   if (SpecialReg == "spsr") {
5598     Ops = { getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5599             N->getOperand(0) };
5600     ReplaceNode(
5601         N, CurDAG->getMachineNode(IsThumb2 ? ARM::t2MRSsys_AR : ARM::MRSsys, DL,
5602                                   MVT::i32, MVT::Other, Ops));
5603     return true;
5604   }
5605 
5606   return false;
5607 }
5608 
5609 // Lower the write_register intrinsic to ARM specific DAG nodes
5610 // using the supplied metadata string to select the instruction node to use
5611 // and the registers/masks to use in the nodes
5612 bool ARMDAGToDAGISel::tryWriteRegister(SDNode *N){
5613   const auto *MD = cast<MDNodeSDNode>(N->getOperand(1));
5614   const auto *RegString = cast<MDString>(MD->getMD()->getOperand(0));
5615   bool IsThumb2 = Subtarget->isThumb2();
5616   SDLoc DL(N);
5617 
5618   std::vector<SDValue> Ops;
5619   getIntOperandsFromRegisterString(RegString->getString(), CurDAG, DL, Ops);
5620 
5621   if (!Ops.empty()) {
5622     // If the special register string was constructed of fields (as defined
5623     // in the ACLE) then need to lower to MCR node (32 bit) or
5624     // MCRR node(64 bit), we can make the distinction based on the number of
5625     // operands we have.
5626     unsigned Opcode;
5627     if (Ops.size() == 5) {
5628       Opcode = IsThumb2 ? ARM::t2MCR : ARM::MCR;
5629       Ops.insert(Ops.begin()+2, N->getOperand(2));
5630     } else {
5631       assert(Ops.size() == 3 &&
5632               "Invalid number of fields in special register string.");
5633       Opcode = IsThumb2 ? ARM::t2MCRR : ARM::MCRR;
5634       SDValue WriteValue[] = { N->getOperand(2), N->getOperand(3) };
5635       Ops.insert(Ops.begin()+2, WriteValue, WriteValue+2);
5636     }
5637 
5638     Ops.push_back(getAL(CurDAG, DL));
5639     Ops.push_back(CurDAG->getRegister(0, MVT::i32));
5640     Ops.push_back(N->getOperand(0));
5641 
5642     ReplaceNode(N, CurDAG->getMachineNode(Opcode, DL, MVT::Other, Ops));
5643     return true;
5644   }
5645 
5646   std::string SpecialReg = RegString->getString().lower();
5647   int BankedReg = getBankedRegisterMask(SpecialReg);
5648   if (BankedReg != -1) {
5649     Ops = { CurDAG->getTargetConstant(BankedReg, DL, MVT::i32), N->getOperand(2),
5650             getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5651             N->getOperand(0) };
5652     ReplaceNode(
5653         N, CurDAG->getMachineNode(IsThumb2 ? ARM::t2MSRbanked : ARM::MSRbanked,
5654                                   DL, MVT::Other, Ops));
5655     return true;
5656   }
5657 
5658   // The VFP registers are written to by creating SelectionDAG nodes with
5659   // opcodes corresponding to the register that is being written. So we switch
5660   // on the string to find which opcode we need to use.
5661   unsigned Opcode = StringSwitch<unsigned>(SpecialReg)
5662                     .Case("fpscr", ARM::VMSR)
5663                     .Case("fpexc", ARM::VMSR_FPEXC)
5664                     .Case("fpsid", ARM::VMSR_FPSID)
5665                     .Case("fpinst", ARM::VMSR_FPINST)
5666                     .Case("fpinst2", ARM::VMSR_FPINST2)
5667                     .Default(0);
5668 
5669   if (Opcode) {
5670     if (!Subtarget->hasVFP2Base())
5671       return false;
5672     Ops = { N->getOperand(2), getAL(CurDAG, DL),
5673             CurDAG->getRegister(0, MVT::i32), N->getOperand(0) };
5674     ReplaceNode(N, CurDAG->getMachineNode(Opcode, DL, MVT::Other, Ops));
5675     return true;
5676   }
5677 
5678   std::pair<StringRef, StringRef> Fields;
5679   Fields = StringRef(SpecialReg).rsplit('_');
5680   std::string Reg = Fields.first.str();
5681   StringRef Flags = Fields.second;
5682 
5683   // If the target was M Class then need to validate the special register value
5684   // and retrieve the mask for use in the instruction node.
5685   if (Subtarget->isMClass()) {
5686     int SYSmValue = getMClassRegisterMask(SpecialReg, Subtarget);
5687     if (SYSmValue == -1)
5688       return false;
5689 
5690     SDValue Ops[] = { CurDAG->getTargetConstant(SYSmValue, DL, MVT::i32),
5691                       N->getOperand(2), getAL(CurDAG, DL),
5692                       CurDAG->getRegister(0, MVT::i32), N->getOperand(0) };
5693     ReplaceNode(N, CurDAG->getMachineNode(ARM::t2MSR_M, DL, MVT::Other, Ops));
5694     return true;
5695   }
5696 
5697   // We then check to see if a valid mask can be constructed for one of the
5698   // register string values permitted for the A and R class cores. These values
5699   // are apsr, spsr and cpsr; these are also valid on older cores.
5700   int Mask = getARClassRegisterMask(Reg, Flags);
5701   if (Mask != -1) {
5702     Ops = { CurDAG->getTargetConstant(Mask, DL, MVT::i32), N->getOperand(2),
5703             getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5704             N->getOperand(0) };
5705     ReplaceNode(N, CurDAG->getMachineNode(IsThumb2 ? ARM::t2MSR_AR : ARM::MSR,
5706                                           DL, MVT::Other, Ops));
5707     return true;
5708   }
5709 
5710   return false;
5711 }
5712 
5713 bool ARMDAGToDAGISel::tryInlineAsm(SDNode *N){
5714   std::vector<SDValue> AsmNodeOperands;
5715   unsigned Flag, Kind;
5716   bool Changed = false;
5717   unsigned NumOps = N->getNumOperands();
5718 
5719   // Normally, i64 data is bounded to two arbitrary GRPs for "%r" constraint.
5720   // However, some instrstions (e.g. ldrexd/strexd in ARM mode) require
5721   // (even/even+1) GPRs and use %n and %Hn to refer to the individual regs
5722   // respectively. Since there is no constraint to explicitly specify a
5723   // reg pair, we use GPRPair reg class for "%r" for 64-bit data. For Thumb,
5724   // the 64-bit data may be referred by H, Q, R modifiers, so we still pack
5725   // them into a GPRPair.
5726 
5727   SDLoc dl(N);
5728   SDValue Glue = N->getGluedNode() ? N->getOperand(NumOps - 1) : SDValue();
5729 
5730   SmallVector<bool, 8> OpChanged;
5731   // Glue node will be appended late.
5732   for(unsigned i = 0, e = N->getGluedNode() ? NumOps - 1 : NumOps; i < e; ++i) {
5733     SDValue op = N->getOperand(i);
5734     AsmNodeOperands.push_back(op);
5735 
5736     if (i < InlineAsm::Op_FirstOperand)
5737       continue;
5738 
5739     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(i))) {
5740       Flag = C->getZExtValue();
5741       Kind = InlineAsm::getKind(Flag);
5742     }
5743     else
5744       continue;
5745 
5746     // Immediate operands to inline asm in the SelectionDAG are modeled with
5747     // two operands. The first is a constant of value InlineAsm::Kind_Imm, and
5748     // the second is a constant with the value of the immediate. If we get here
5749     // and we have a Kind_Imm, skip the next operand, and continue.
5750     if (Kind == InlineAsm::Kind_Imm) {
5751       SDValue op = N->getOperand(++i);
5752       AsmNodeOperands.push_back(op);
5753       continue;
5754     }
5755 
5756     unsigned NumRegs = InlineAsm::getNumOperandRegisters(Flag);
5757     if (NumRegs)
5758       OpChanged.push_back(false);
5759 
5760     unsigned DefIdx = 0;
5761     bool IsTiedToChangedOp = false;
5762     // If it's a use that is tied with a previous def, it has no
5763     // reg class constraint.
5764     if (Changed && InlineAsm::isUseOperandTiedToDef(Flag, DefIdx))
5765       IsTiedToChangedOp = OpChanged[DefIdx];
5766 
5767     // Memory operands to inline asm in the SelectionDAG are modeled with two
5768     // operands: a constant of value InlineAsm::Kind_Mem followed by the input
5769     // operand. If we get here and we have a Kind_Mem, skip the next operand (so
5770     // it doesn't get misinterpreted), and continue. We do this here because
5771     // it's important to update the OpChanged array correctly before moving on.
5772     if (Kind == InlineAsm::Kind_Mem) {
5773       SDValue op = N->getOperand(++i);
5774       AsmNodeOperands.push_back(op);
5775       continue;
5776     }
5777 
5778     if (Kind != InlineAsm::Kind_RegUse && Kind != InlineAsm::Kind_RegDef
5779         && Kind != InlineAsm::Kind_RegDefEarlyClobber)
5780       continue;
5781 
5782     unsigned RC;
5783     bool HasRC = InlineAsm::hasRegClassConstraint(Flag, RC);
5784     if ((!IsTiedToChangedOp && (!HasRC || RC != ARM::GPRRegClassID))
5785         || NumRegs != 2)
5786       continue;
5787 
5788     assert((i+2 < NumOps) && "Invalid number of operands in inline asm");
5789     SDValue V0 = N->getOperand(i+1);
5790     SDValue V1 = N->getOperand(i+2);
5791     Register Reg0 = cast<RegisterSDNode>(V0)->getReg();
5792     Register Reg1 = cast<RegisterSDNode>(V1)->getReg();
5793     SDValue PairedReg;
5794     MachineRegisterInfo &MRI = MF->getRegInfo();
5795 
5796     if (Kind == InlineAsm::Kind_RegDef ||
5797         Kind == InlineAsm::Kind_RegDefEarlyClobber) {
5798       // Replace the two GPRs with 1 GPRPair and copy values from GPRPair to
5799       // the original GPRs.
5800 
5801       Register GPVR = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5802       PairedReg = CurDAG->getRegister(GPVR, MVT::Untyped);
5803       SDValue Chain = SDValue(N,0);
5804 
5805       SDNode *GU = N->getGluedUser();
5806       SDValue RegCopy = CurDAG->getCopyFromReg(Chain, dl, GPVR, MVT::Untyped,
5807                                                Chain.getValue(1));
5808 
5809       // Extract values from a GPRPair reg and copy to the original GPR reg.
5810       SDValue Sub0 = CurDAG->getTargetExtractSubreg(ARM::gsub_0, dl, MVT::i32,
5811                                                     RegCopy);
5812       SDValue Sub1 = CurDAG->getTargetExtractSubreg(ARM::gsub_1, dl, MVT::i32,
5813                                                     RegCopy);
5814       SDValue T0 = CurDAG->getCopyToReg(Sub0, dl, Reg0, Sub0,
5815                                         RegCopy.getValue(1));
5816       SDValue T1 = CurDAG->getCopyToReg(Sub1, dl, Reg1, Sub1, T0.getValue(1));
5817 
5818       // Update the original glue user.
5819       std::vector<SDValue> Ops(GU->op_begin(), GU->op_end()-1);
5820       Ops.push_back(T1.getValue(1));
5821       CurDAG->UpdateNodeOperands(GU, Ops);
5822     }
5823     else {
5824       // For Kind  == InlineAsm::Kind_RegUse, we first copy two GPRs into a
5825       // GPRPair and then pass the GPRPair to the inline asm.
5826       SDValue Chain = AsmNodeOperands[InlineAsm::Op_InputChain];
5827 
5828       // As REG_SEQ doesn't take RegisterSDNode, we copy them first.
5829       SDValue T0 = CurDAG->getCopyFromReg(Chain, dl, Reg0, MVT::i32,
5830                                           Chain.getValue(1));
5831       SDValue T1 = CurDAG->getCopyFromReg(Chain, dl, Reg1, MVT::i32,
5832                                           T0.getValue(1));
5833       SDValue Pair = SDValue(createGPRPairNode(MVT::Untyped, T0, T1), 0);
5834 
5835       // Copy REG_SEQ into a GPRPair-typed VR and replace the original two
5836       // i32 VRs of inline asm with it.
5837       Register GPVR = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5838       PairedReg = CurDAG->getRegister(GPVR, MVT::Untyped);
5839       Chain = CurDAG->getCopyToReg(T1, dl, GPVR, Pair, T1.getValue(1));
5840 
5841       AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
5842       Glue = Chain.getValue(1);
5843     }
5844 
5845     Changed = true;
5846 
5847     if(PairedReg.getNode()) {
5848       OpChanged[OpChanged.size() -1 ] = true;
5849       Flag = InlineAsm::getFlagWord(Kind, 1 /* RegNum*/);
5850       if (IsTiedToChangedOp)
5851         Flag = InlineAsm::getFlagWordForMatchingOp(Flag, DefIdx);
5852       else
5853         Flag = InlineAsm::getFlagWordForRegClass(Flag, ARM::GPRPairRegClassID);
5854       // Replace the current flag.
5855       AsmNodeOperands[AsmNodeOperands.size() -1] = CurDAG->getTargetConstant(
5856           Flag, dl, MVT::i32);
5857       // Add the new register node and skip the original two GPRs.
5858       AsmNodeOperands.push_back(PairedReg);
5859       // Skip the next two GPRs.
5860       i += 2;
5861     }
5862   }
5863 
5864   if (Glue.getNode())
5865     AsmNodeOperands.push_back(Glue);
5866   if (!Changed)
5867     return false;
5868 
5869   SDValue New = CurDAG->getNode(N->getOpcode(), SDLoc(N),
5870       CurDAG->getVTList(MVT::Other, MVT::Glue), AsmNodeOperands);
5871   New->setNodeId(-1);
5872   ReplaceNode(N, New.getNode());
5873   return true;
5874 }
5875 
5876 
5877 bool ARMDAGToDAGISel::
5878 SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
5879                              std::vector<SDValue> &OutOps) {
5880   switch(ConstraintID) {
5881   default:
5882     llvm_unreachable("Unexpected asm memory constraint");
5883   case InlineAsm::Constraint_m:
5884   case InlineAsm::Constraint_o:
5885   case InlineAsm::Constraint_Q:
5886   case InlineAsm::Constraint_Um:
5887   case InlineAsm::Constraint_Un:
5888   case InlineAsm::Constraint_Uq:
5889   case InlineAsm::Constraint_Us:
5890   case InlineAsm::Constraint_Ut:
5891   case InlineAsm::Constraint_Uv:
5892   case InlineAsm::Constraint_Uy:
5893     // Require the address to be in a register.  That is safe for all ARM
5894     // variants and it is hard to do anything much smarter without knowing
5895     // how the operand is used.
5896     OutOps.push_back(Op);
5897     return false;
5898   }
5899   return true;
5900 }
5901 
5902 /// createARMISelDag - This pass converts a legalized DAG into a
5903 /// ARM-specific DAG, ready for instruction scheduling.
5904 ///
5905 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
5906                                      CodeGenOpt::Level OptLevel) {
5907   return new ARMDAGToDAGISel(TM, OptLevel);
5908 }
5909