xref: /freebsd/contrib/llvm-project/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp (revision 1db9f3b21e39176dd5b67cf8ac378633b172463e)
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, CodeGenOptLevel 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,
335                                     InlineAsm::ConstraintCode ConstraintID,
336                                     std::vector<SDValue> &OutOps) override;
337 
338   // Form pairs of consecutive R, S, D, or Q registers.
339   SDNode *createGPRPairNode(EVT VT, SDValue V0, SDValue V1);
340   SDNode *createSRegPairNode(EVT VT, SDValue V0, SDValue V1);
341   SDNode *createDRegPairNode(EVT VT, SDValue V0, SDValue V1);
342   SDNode *createQRegPairNode(EVT VT, SDValue V0, SDValue V1);
343 
344   // Form sequences of 4 consecutive S, D, or Q registers.
345   SDNode *createQuadSRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
346   SDNode *createQuadDRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
347   SDNode *createQuadQRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
348 
349   // Get the alignment operand for a NEON VLD or VST instruction.
350   SDValue GetVLDSTAlign(SDValue Align, const SDLoc &dl, unsigned NumVecs,
351                         bool is64BitVector);
352 
353   /// Checks if N is a multiplication by a constant where we can extract out a
354   /// power of two from the constant so that it can be used in a shift, but only
355   /// if it simplifies the materialization of the constant. Returns true if it
356   /// is, and assigns to PowerOfTwo the power of two that should be extracted
357   /// out and to NewMulConst the new constant to be multiplied by.
358   bool canExtractShiftFromMul(const SDValue &N, unsigned MaxShift,
359                               unsigned &PowerOfTwo, SDValue &NewMulConst) const;
360 
361   /// Replace N with M in CurDAG, in a way that also ensures that M gets
362   /// selected when N would have been selected.
363   void replaceDAGValue(const SDValue &N, SDValue M);
364 };
365 }
366 
367 char ARMDAGToDAGISel::ID = 0;
368 
369 INITIALIZE_PASS(ARMDAGToDAGISel, DEBUG_TYPE, PASS_NAME, false, false)
370 
371 /// isInt32Immediate - This method tests to see if the node is a 32-bit constant
372 /// operand. If so Imm will receive the 32-bit value.
373 static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
374   if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
375     Imm = N->getAsZExtVal();
376     return true;
377   }
378   return false;
379 }
380 
381 // isInt32Immediate - This method tests to see if a constant operand.
382 // If so Imm will receive the 32 bit value.
383 static bool isInt32Immediate(SDValue N, unsigned &Imm) {
384   return isInt32Immediate(N.getNode(), Imm);
385 }
386 
387 // isOpcWithIntImmediate - This method tests to see if the node is a specific
388 // opcode and that it has a immediate integer right operand.
389 // If so Imm will receive the 32 bit value.
390 static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
391   return N->getOpcode() == Opc &&
392          isInt32Immediate(N->getOperand(1).getNode(), Imm);
393 }
394 
395 /// Check whether a particular node is a constant value representable as
396 /// (N * Scale) where (N in [\p RangeMin, \p RangeMax).
397 ///
398 /// \param ScaledConstant [out] - On success, the pre-scaled constant value.
399 static bool isScaledConstantInRange(SDValue Node, int Scale,
400                                     int RangeMin, int RangeMax,
401                                     int &ScaledConstant) {
402   assert(Scale > 0 && "Invalid scale!");
403 
404   // Check that this is a constant.
405   const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Node);
406   if (!C)
407     return false;
408 
409   ScaledConstant = (int) C->getZExtValue();
410   if ((ScaledConstant % Scale) != 0)
411     return false;
412 
413   ScaledConstant /= Scale;
414   return ScaledConstant >= RangeMin && ScaledConstant < RangeMax;
415 }
416 
417 void ARMDAGToDAGISel::PreprocessISelDAG() {
418   if (!Subtarget->hasV6T2Ops())
419     return;
420 
421   bool isThumb2 = Subtarget->isThumb();
422   // We use make_early_inc_range to avoid invalidation issues.
423   for (SDNode &N : llvm::make_early_inc_range(CurDAG->allnodes())) {
424     if (N.getOpcode() != ISD::ADD)
425       continue;
426 
427     // Look for (add X1, (and (srl X2, c1), c2)) where c2 is constant with
428     // leading zeros, followed by consecutive set bits, followed by 1 or 2
429     // trailing zeros, e.g. 1020.
430     // Transform the expression to
431     // (add X1, (shl (and (srl X2, c1), (c2>>tz)), tz)) where tz is the number
432     // of trailing zeros of c2. The left shift would be folded as an shifter
433     // operand of 'add' and the 'and' and 'srl' would become a bits extraction
434     // node (UBFX).
435 
436     SDValue N0 = N.getOperand(0);
437     SDValue N1 = N.getOperand(1);
438     unsigned And_imm = 0;
439     if (!isOpcWithIntImmediate(N1.getNode(), ISD::AND, And_imm)) {
440       if (isOpcWithIntImmediate(N0.getNode(), ISD::AND, And_imm))
441         std::swap(N0, N1);
442     }
443     if (!And_imm)
444       continue;
445 
446     // Check if the AND mask is an immediate of the form: 000.....1111111100
447     unsigned TZ = llvm::countr_zero(And_imm);
448     if (TZ != 1 && TZ != 2)
449       // Be conservative here. Shifter operands aren't always free. e.g. On
450       // Swift, left shifter operand of 1 / 2 for free but others are not.
451       // e.g.
452       //  ubfx   r3, r1, #16, #8
453       //  ldr.w  r3, [r0, r3, lsl #2]
454       // vs.
455       //  mov.w  r9, #1020
456       //  and.w  r2, r9, r1, lsr #14
457       //  ldr    r2, [r0, r2]
458       continue;
459     And_imm >>= TZ;
460     if (And_imm & (And_imm + 1))
461       continue;
462 
463     // Look for (and (srl X, c1), c2).
464     SDValue Srl = N1.getOperand(0);
465     unsigned Srl_imm = 0;
466     if (!isOpcWithIntImmediate(Srl.getNode(), ISD::SRL, Srl_imm) ||
467         (Srl_imm <= 2))
468       continue;
469 
470     // Make sure first operand is not a shifter operand which would prevent
471     // folding of the left shift.
472     SDValue CPTmp0;
473     SDValue CPTmp1;
474     SDValue CPTmp2;
475     if (isThumb2) {
476       if (SelectImmShifterOperand(N0, CPTmp0, CPTmp1))
477         continue;
478     } else {
479       if (SelectImmShifterOperand(N0, CPTmp0, CPTmp1) ||
480           SelectRegShifterOperand(N0, CPTmp0, CPTmp1, CPTmp2))
481         continue;
482     }
483 
484     // Now make the transformation.
485     Srl = CurDAG->getNode(ISD::SRL, SDLoc(Srl), MVT::i32,
486                           Srl.getOperand(0),
487                           CurDAG->getConstant(Srl_imm + TZ, SDLoc(Srl),
488                                               MVT::i32));
489     N1 = CurDAG->getNode(ISD::AND, SDLoc(N1), MVT::i32,
490                          Srl,
491                          CurDAG->getConstant(And_imm, SDLoc(Srl), MVT::i32));
492     N1 = CurDAG->getNode(ISD::SHL, SDLoc(N1), MVT::i32,
493                          N1, CurDAG->getConstant(TZ, SDLoc(Srl), MVT::i32));
494     CurDAG->UpdateNodeOperands(&N, N0, N1);
495   }
496 }
497 
498 /// hasNoVMLxHazardUse - Return true if it's desirable to select a FP MLA / MLS
499 /// node. VFP / NEON fp VMLA / VMLS instructions have special RAW hazards (at
500 /// least on current ARM implementations) which should be avoidded.
501 bool ARMDAGToDAGISel::hasNoVMLxHazardUse(SDNode *N) const {
502   if (OptLevel == CodeGenOptLevel::None)
503     return true;
504 
505   if (!Subtarget->hasVMLxHazards())
506     return true;
507 
508   if (!N->hasOneUse())
509     return false;
510 
511   SDNode *Use = *N->use_begin();
512   if (Use->getOpcode() == ISD::CopyToReg)
513     return true;
514   if (Use->isMachineOpcode()) {
515     const ARMBaseInstrInfo *TII = static_cast<const ARMBaseInstrInfo *>(
516         CurDAG->getSubtarget().getInstrInfo());
517 
518     const MCInstrDesc &MCID = TII->get(Use->getMachineOpcode());
519     if (MCID.mayStore())
520       return true;
521     unsigned Opcode = MCID.getOpcode();
522     if (Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD)
523       return true;
524     // vmlx feeding into another vmlx. We actually want to unfold
525     // the use later in the MLxExpansion pass. e.g.
526     // vmla
527     // vmla (stall 8 cycles)
528     //
529     // vmul (5 cycles)
530     // vadd (5 cycles)
531     // vmla
532     // This adds up to about 18 - 19 cycles.
533     //
534     // vmla
535     // vmul (stall 4 cycles)
536     // vadd adds up to about 14 cycles.
537     return TII->isFpMLxInstruction(Opcode);
538   }
539 
540   return false;
541 }
542 
543 bool ARMDAGToDAGISel::isShifterOpProfitable(const SDValue &Shift,
544                                             ARM_AM::ShiftOpc ShOpcVal,
545                                             unsigned ShAmt) {
546   if (!Subtarget->isLikeA9() && !Subtarget->isSwift())
547     return true;
548   if (Shift.hasOneUse())
549     return true;
550   // R << 2 is free.
551   return ShOpcVal == ARM_AM::lsl &&
552          (ShAmt == 2 || (Subtarget->isSwift() && ShAmt == 1));
553 }
554 
555 bool ARMDAGToDAGISel::canExtractShiftFromMul(const SDValue &N,
556                                              unsigned MaxShift,
557                                              unsigned &PowerOfTwo,
558                                              SDValue &NewMulConst) const {
559   assert(N.getOpcode() == ISD::MUL);
560   assert(MaxShift > 0);
561 
562   // If the multiply is used in more than one place then changing the constant
563   // will make other uses incorrect, so don't.
564   if (!N.hasOneUse()) return false;
565   // Check if the multiply is by a constant
566   ConstantSDNode *MulConst = dyn_cast<ConstantSDNode>(N.getOperand(1));
567   if (!MulConst) return false;
568   // If the constant is used in more than one place then modifying it will mean
569   // we need to materialize two constants instead of one, which is a bad idea.
570   if (!MulConst->hasOneUse()) return false;
571   unsigned MulConstVal = MulConst->getZExtValue();
572   if (MulConstVal == 0) return false;
573 
574   // Find the largest power of 2 that MulConstVal is a multiple of
575   PowerOfTwo = MaxShift;
576   while ((MulConstVal % (1 << PowerOfTwo)) != 0) {
577     --PowerOfTwo;
578     if (PowerOfTwo == 0) return false;
579   }
580 
581   // Only optimise if the new cost is better
582   unsigned NewMulConstVal = MulConstVal / (1 << PowerOfTwo);
583   NewMulConst = CurDAG->getConstant(NewMulConstVal, SDLoc(N), MVT::i32);
584   unsigned OldCost = ConstantMaterializationCost(MulConstVal, Subtarget);
585   unsigned NewCost = ConstantMaterializationCost(NewMulConstVal, Subtarget);
586   return NewCost < OldCost;
587 }
588 
589 void ARMDAGToDAGISel::replaceDAGValue(const SDValue &N, SDValue M) {
590   CurDAG->RepositionNode(N.getNode()->getIterator(), M.getNode());
591   ReplaceUses(N, M);
592 }
593 
594 bool ARMDAGToDAGISel::SelectImmShifterOperand(SDValue N,
595                                               SDValue &BaseReg,
596                                               SDValue &Opc,
597                                               bool CheckProfitability) {
598   if (DisableShifterOp)
599     return false;
600 
601   // If N is a multiply-by-constant and it's profitable to extract a shift and
602   // use it in a shifted operand do so.
603   if (N.getOpcode() == ISD::MUL) {
604     unsigned PowerOfTwo = 0;
605     SDValue NewMulConst;
606     if (canExtractShiftFromMul(N, 31, PowerOfTwo, NewMulConst)) {
607       HandleSDNode Handle(N);
608       SDLoc Loc(N);
609       replaceDAGValue(N.getOperand(1), NewMulConst);
610       BaseReg = Handle.getValue();
611       Opc = CurDAG->getTargetConstant(
612           ARM_AM::getSORegOpc(ARM_AM::lsl, PowerOfTwo), Loc, MVT::i32);
613       return true;
614     }
615   }
616 
617   ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
618 
619   // Don't match base register only case. That is matched to a separate
620   // lower complexity pattern with explicit register operand.
621   if (ShOpcVal == ARM_AM::no_shift) return false;
622 
623   BaseReg = N.getOperand(0);
624   unsigned ShImmVal = 0;
625   ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1));
626   if (!RHS) return false;
627   ShImmVal = RHS->getZExtValue() & 31;
628   Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
629                                   SDLoc(N), MVT::i32);
630   return true;
631 }
632 
633 bool ARMDAGToDAGISel::SelectRegShifterOperand(SDValue N,
634                                               SDValue &BaseReg,
635                                               SDValue &ShReg,
636                                               SDValue &Opc,
637                                               bool CheckProfitability) {
638   if (DisableShifterOp)
639     return false;
640 
641   ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
642 
643   // Don't match base register only case. That is matched to a separate
644   // lower complexity pattern with explicit register operand.
645   if (ShOpcVal == ARM_AM::no_shift) return false;
646 
647   BaseReg = N.getOperand(0);
648   unsigned ShImmVal = 0;
649   ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1));
650   if (RHS) return false;
651 
652   ShReg = N.getOperand(1);
653   if (CheckProfitability && !isShifterOpProfitable(N, ShOpcVal, ShImmVal))
654     return false;
655   Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
656                                   SDLoc(N), MVT::i32);
657   return true;
658 }
659 
660 // Determine whether an ISD::OR's operands are suitable to turn the operation
661 // into an addition, which often has more compact encodings.
662 bool ARMDAGToDAGISel::SelectAddLikeOr(SDNode *Parent, SDValue N, SDValue &Out) {
663   assert(Parent->getOpcode() == ISD::OR && "unexpected parent");
664   Out = N;
665   return CurDAG->haveNoCommonBitsSet(N, Parent->getOperand(1));
666 }
667 
668 
669 bool ARMDAGToDAGISel::SelectAddrModeImm12(SDValue N,
670                                           SDValue &Base,
671                                           SDValue &OffImm) {
672   // Match simple R + imm12 operands.
673 
674   // Base only.
675   if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
676       !CurDAG->isBaseWithConstantOffset(N)) {
677     if (N.getOpcode() == ISD::FrameIndex) {
678       // Match frame index.
679       int FI = cast<FrameIndexSDNode>(N)->getIndex();
680       Base = CurDAG->getTargetFrameIndex(
681           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
682       OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
683       return true;
684     }
685 
686     if (N.getOpcode() == ARMISD::Wrapper &&
687         N.getOperand(0).getOpcode() != ISD::TargetGlobalAddress &&
688         N.getOperand(0).getOpcode() != ISD::TargetExternalSymbol &&
689         N.getOperand(0).getOpcode() != ISD::TargetGlobalTLSAddress) {
690       Base = N.getOperand(0);
691     } else
692       Base = N;
693     OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
694     return true;
695   }
696 
697   if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
698     int RHSC = (int)RHS->getSExtValue();
699     if (N.getOpcode() == ISD::SUB)
700       RHSC = -RHSC;
701 
702     if (RHSC > -0x1000 && RHSC < 0x1000) { // 12 bits
703       Base   = N.getOperand(0);
704       if (Base.getOpcode() == ISD::FrameIndex) {
705         int FI = cast<FrameIndexSDNode>(Base)->getIndex();
706         Base = CurDAG->getTargetFrameIndex(
707             FI, TLI->getPointerTy(CurDAG->getDataLayout()));
708       }
709       OffImm = CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32);
710       return true;
711     }
712   }
713 
714   // Base only.
715   Base = N;
716   OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
717   return true;
718 }
719 
720 
721 
722 bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset,
723                                       SDValue &Opc) {
724   if (N.getOpcode() == ISD::MUL &&
725       ((!Subtarget->isLikeA9() && !Subtarget->isSwift()) || N.hasOneUse())) {
726     if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
727       // X * [3,5,9] -> X + X * [2,4,8] etc.
728       int RHSC = (int)RHS->getZExtValue();
729       if (RHSC & 1) {
730         RHSC = RHSC & ~1;
731         ARM_AM::AddrOpc AddSub = ARM_AM::add;
732         if (RHSC < 0) {
733           AddSub = ARM_AM::sub;
734           RHSC = - RHSC;
735         }
736         if (isPowerOf2_32(RHSC)) {
737           unsigned ShAmt = Log2_32(RHSC);
738           Base = Offset = N.getOperand(0);
739           Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
740                                                             ARM_AM::lsl),
741                                           SDLoc(N), MVT::i32);
742           return true;
743         }
744       }
745     }
746   }
747 
748   if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
749       // ISD::OR that is equivalent to an ISD::ADD.
750       !CurDAG->isBaseWithConstantOffset(N))
751     return false;
752 
753   // Leave simple R +/- imm12 operands for LDRi12
754   if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::OR) {
755     int RHSC;
756     if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1,
757                                 -0x1000+1, 0x1000, RHSC)) // 12 bits.
758       return false;
759   }
760 
761   // Otherwise this is R +/- [possibly shifted] R.
762   ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::SUB ? ARM_AM::sub:ARM_AM::add;
763   ARM_AM::ShiftOpc ShOpcVal =
764     ARM_AM::getShiftOpcForNode(N.getOperand(1).getOpcode());
765   unsigned ShAmt = 0;
766 
767   Base   = N.getOperand(0);
768   Offset = N.getOperand(1);
769 
770   if (ShOpcVal != ARM_AM::no_shift) {
771     // Check to see if the RHS of the shift is a constant, if not, we can't fold
772     // it.
773     if (ConstantSDNode *Sh =
774            dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
775       ShAmt = Sh->getZExtValue();
776       if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt))
777         Offset = N.getOperand(1).getOperand(0);
778       else {
779         ShAmt = 0;
780         ShOpcVal = ARM_AM::no_shift;
781       }
782     } else {
783       ShOpcVal = ARM_AM::no_shift;
784     }
785   }
786 
787   // Try matching (R shl C) + (R).
788   if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift &&
789       !(Subtarget->isLikeA9() || Subtarget->isSwift() ||
790         N.getOperand(0).hasOneUse())) {
791     ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0).getOpcode());
792     if (ShOpcVal != ARM_AM::no_shift) {
793       // Check to see if the RHS of the shift is a constant, if not, we can't
794       // fold it.
795       if (ConstantSDNode *Sh =
796           dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
797         ShAmt = Sh->getZExtValue();
798         if (isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt)) {
799           Offset = N.getOperand(0).getOperand(0);
800           Base = N.getOperand(1);
801         } else {
802           ShAmt = 0;
803           ShOpcVal = ARM_AM::no_shift;
804         }
805       } else {
806         ShOpcVal = ARM_AM::no_shift;
807       }
808     }
809   }
810 
811   // If Offset is a multiply-by-constant and it's profitable to extract a shift
812   // and use it in a shifted operand do so.
813   if (Offset.getOpcode() == ISD::MUL && N.hasOneUse()) {
814     unsigned PowerOfTwo = 0;
815     SDValue NewMulConst;
816     if (canExtractShiftFromMul(Offset, 31, PowerOfTwo, NewMulConst)) {
817       HandleSDNode Handle(Offset);
818       replaceDAGValue(Offset.getOperand(1), NewMulConst);
819       Offset = Handle.getValue();
820       ShAmt = PowerOfTwo;
821       ShOpcVal = ARM_AM::lsl;
822     }
823   }
824 
825   Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
826                                   SDLoc(N), MVT::i32);
827   return true;
828 }
829 
830 bool ARMDAGToDAGISel::SelectAddrMode2OffsetReg(SDNode *Op, SDValue N,
831                                             SDValue &Offset, SDValue &Opc) {
832   unsigned Opcode = Op->getOpcode();
833   ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
834     ? cast<LoadSDNode>(Op)->getAddressingMode()
835     : cast<StoreSDNode>(Op)->getAddressingMode();
836   ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
837     ? ARM_AM::add : ARM_AM::sub;
838   int Val;
839   if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val))
840     return false;
841 
842   Offset = N;
843   ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode());
844   unsigned ShAmt = 0;
845   if (ShOpcVal != ARM_AM::no_shift) {
846     // Check to see if the RHS of the shift is a constant, if not, we can't fold
847     // it.
848     if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
849       ShAmt = Sh->getZExtValue();
850       if (isShifterOpProfitable(N, ShOpcVal, ShAmt))
851         Offset = N.getOperand(0);
852       else {
853         ShAmt = 0;
854         ShOpcVal = ARM_AM::no_shift;
855       }
856     } else {
857       ShOpcVal = ARM_AM::no_shift;
858     }
859   }
860 
861   Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
862                                   SDLoc(N), MVT::i32);
863   return true;
864 }
865 
866 bool ARMDAGToDAGISel::SelectAddrMode2OffsetImmPre(SDNode *Op, SDValue N,
867                                             SDValue &Offset, SDValue &Opc) {
868   unsigned Opcode = Op->getOpcode();
869   ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
870     ? cast<LoadSDNode>(Op)->getAddressingMode()
871     : cast<StoreSDNode>(Op)->getAddressingMode();
872   ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
873     ? ARM_AM::add : ARM_AM::sub;
874   int Val;
875   if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val)) { // 12 bits.
876     if (AddSub == ARM_AM::sub) Val *= -1;
877     Offset = CurDAG->getRegister(0, MVT::i32);
878     Opc = CurDAG->getTargetConstant(Val, SDLoc(Op), MVT::i32);
879     return true;
880   }
881 
882   return false;
883 }
884 
885 
886 bool ARMDAGToDAGISel::SelectAddrMode2OffsetImm(SDNode *Op, SDValue N,
887                                             SDValue &Offset, SDValue &Opc) {
888   unsigned Opcode = Op->getOpcode();
889   ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
890     ? cast<LoadSDNode>(Op)->getAddressingMode()
891     : cast<StoreSDNode>(Op)->getAddressingMode();
892   ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
893     ? ARM_AM::add : ARM_AM::sub;
894   int Val;
895   if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val)) { // 12 bits.
896     Offset = CurDAG->getRegister(0, MVT::i32);
897     Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
898                                                       ARM_AM::no_shift),
899                                     SDLoc(Op), MVT::i32);
900     return true;
901   }
902 
903   return false;
904 }
905 
906 bool ARMDAGToDAGISel::SelectAddrOffsetNone(SDValue N, SDValue &Base) {
907   Base = N;
908   return true;
909 }
910 
911 bool ARMDAGToDAGISel::SelectAddrMode3(SDValue N,
912                                       SDValue &Base, SDValue &Offset,
913                                       SDValue &Opc) {
914   if (N.getOpcode() == ISD::SUB) {
915     // X - C  is canonicalize to X + -C, no need to handle it here.
916     Base = N.getOperand(0);
917     Offset = N.getOperand(1);
918     Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0), SDLoc(N),
919                                     MVT::i32);
920     return true;
921   }
922 
923   if (!CurDAG->isBaseWithConstantOffset(N)) {
924     Base = N;
925     if (N.getOpcode() == ISD::FrameIndex) {
926       int FI = cast<FrameIndexSDNode>(N)->getIndex();
927       Base = CurDAG->getTargetFrameIndex(
928           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
929     }
930     Offset = CurDAG->getRegister(0, MVT::i32);
931     Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), SDLoc(N),
932                                     MVT::i32);
933     return true;
934   }
935 
936   // If the RHS is +/- imm8, fold into addr mode.
937   int RHSC;
938   if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1,
939                               -256 + 1, 256, RHSC)) { // 8 bits.
940     Base = N.getOperand(0);
941     if (Base.getOpcode() == ISD::FrameIndex) {
942       int FI = cast<FrameIndexSDNode>(Base)->getIndex();
943       Base = CurDAG->getTargetFrameIndex(
944           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
945     }
946     Offset = CurDAG->getRegister(0, MVT::i32);
947 
948     ARM_AM::AddrOpc AddSub = ARM_AM::add;
949     if (RHSC < 0) {
950       AddSub = ARM_AM::sub;
951       RHSC = -RHSC;
952     }
953     Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC), SDLoc(N),
954                                     MVT::i32);
955     return true;
956   }
957 
958   Base = N.getOperand(0);
959   Offset = N.getOperand(1);
960   Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), SDLoc(N),
961                                   MVT::i32);
962   return true;
963 }
964 
965 bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N,
966                                             SDValue &Offset, SDValue &Opc) {
967   unsigned Opcode = Op->getOpcode();
968   ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
969     ? cast<LoadSDNode>(Op)->getAddressingMode()
970     : cast<StoreSDNode>(Op)->getAddressingMode();
971   ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
972     ? ARM_AM::add : ARM_AM::sub;
973   int Val;
974   if (isScaledConstantInRange(N, /*Scale=*/1, 0, 256, Val)) { // 12 bits.
975     Offset = CurDAG->getRegister(0, MVT::i32);
976     Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), SDLoc(Op),
977                                     MVT::i32);
978     return true;
979   }
980 
981   Offset = N;
982   Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), SDLoc(Op),
983                                   MVT::i32);
984   return true;
985 }
986 
987 bool ARMDAGToDAGISel::IsAddressingMode5(SDValue N, SDValue &Base, SDValue &Offset,
988                                         bool FP16) {
989   if (!CurDAG->isBaseWithConstantOffset(N)) {
990     Base = N;
991     if (N.getOpcode() == ISD::FrameIndex) {
992       int FI = cast<FrameIndexSDNode>(N)->getIndex();
993       Base = CurDAG->getTargetFrameIndex(
994           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
995     } else if (N.getOpcode() == ARMISD::Wrapper &&
996                N.getOperand(0).getOpcode() != ISD::TargetGlobalAddress &&
997                N.getOperand(0).getOpcode() != ISD::TargetExternalSymbol &&
998                N.getOperand(0).getOpcode() != ISD::TargetGlobalTLSAddress) {
999       Base = N.getOperand(0);
1000     }
1001     Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
1002                                        SDLoc(N), MVT::i32);
1003     return true;
1004   }
1005 
1006   // If the RHS is +/- imm8, fold into addr mode.
1007   int RHSC;
1008   const int Scale = FP16 ? 2 : 4;
1009 
1010   if (isScaledConstantInRange(N.getOperand(1), Scale, -255, 256, RHSC)) {
1011     Base = N.getOperand(0);
1012     if (Base.getOpcode() == ISD::FrameIndex) {
1013       int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1014       Base = CurDAG->getTargetFrameIndex(
1015           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1016     }
1017 
1018     ARM_AM::AddrOpc AddSub = ARM_AM::add;
1019     if (RHSC < 0) {
1020       AddSub = ARM_AM::sub;
1021       RHSC = -RHSC;
1022     }
1023 
1024     if (FP16)
1025       Offset = CurDAG->getTargetConstant(ARM_AM::getAM5FP16Opc(AddSub, RHSC),
1026                                          SDLoc(N), MVT::i32);
1027     else
1028       Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
1029                                          SDLoc(N), MVT::i32);
1030 
1031     return true;
1032   }
1033 
1034   Base = N;
1035 
1036   if (FP16)
1037     Offset = CurDAG->getTargetConstant(ARM_AM::getAM5FP16Opc(ARM_AM::add, 0),
1038                                        SDLoc(N), MVT::i32);
1039   else
1040     Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
1041                                        SDLoc(N), MVT::i32);
1042 
1043   return true;
1044 }
1045 
1046 bool ARMDAGToDAGISel::SelectAddrMode5(SDValue N,
1047                                       SDValue &Base, SDValue &Offset) {
1048   return IsAddressingMode5(N, Base, Offset, /*FP16=*/ false);
1049 }
1050 
1051 bool ARMDAGToDAGISel::SelectAddrMode5FP16(SDValue N,
1052                                           SDValue &Base, SDValue &Offset) {
1053   return IsAddressingMode5(N, Base, Offset, /*FP16=*/ true);
1054 }
1055 
1056 bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,
1057                                       SDValue &Align) {
1058   Addr = N;
1059 
1060   unsigned Alignment = 0;
1061 
1062   MemSDNode *MemN = cast<MemSDNode>(Parent);
1063 
1064   if (isa<LSBaseSDNode>(MemN) ||
1065       ((MemN->getOpcode() == ARMISD::VST1_UPD ||
1066         MemN->getOpcode() == ARMISD::VLD1_UPD) &&
1067        MemN->getConstantOperandVal(MemN->getNumOperands() - 1) == 1)) {
1068     // This case occurs only for VLD1-lane/dup and VST1-lane instructions.
1069     // The maximum alignment is equal to the memory size being referenced.
1070     llvm::Align MMOAlign = MemN->getAlign();
1071     unsigned MemSize = MemN->getMemoryVT().getSizeInBits() / 8;
1072     if (MMOAlign.value() >= MemSize && MemSize > 1)
1073       Alignment = MemSize;
1074   } else {
1075     // All other uses of addrmode6 are for intrinsics.  For now just record
1076     // the raw alignment value; it will be refined later based on the legal
1077     // alignment operands for the intrinsic.
1078     Alignment = MemN->getAlign().value();
1079   }
1080 
1081   Align = CurDAG->getTargetConstant(Alignment, SDLoc(N), MVT::i32);
1082   return true;
1083 }
1084 
1085 bool ARMDAGToDAGISel::SelectAddrMode6Offset(SDNode *Op, SDValue N,
1086                                             SDValue &Offset) {
1087   LSBaseSDNode *LdSt = cast<LSBaseSDNode>(Op);
1088   ISD::MemIndexedMode AM = LdSt->getAddressingMode();
1089   if (AM != ISD::POST_INC)
1090     return false;
1091   Offset = N;
1092   if (ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N)) {
1093     if (NC->getZExtValue() * 8 == LdSt->getMemoryVT().getSizeInBits())
1094       Offset = CurDAG->getRegister(0, MVT::i32);
1095   }
1096   return true;
1097 }
1098 
1099 bool ARMDAGToDAGISel::SelectAddrModePC(SDValue N,
1100                                        SDValue &Offset, SDValue &Label) {
1101   if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
1102     Offset = N.getOperand(0);
1103     SDValue N1 = N.getOperand(1);
1104     Label = CurDAG->getTargetConstant(N1->getAsZExtVal(), 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     if (!isNullConstant(N))
1134       return false;
1135 
1136     Base = Offset = N;
1137     return true;
1138   }
1139 
1140   Base = N.getOperand(0);
1141   Offset = N.getOperand(1);
1142   return true;
1143 }
1144 
1145 bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue N, SDValue &Base,
1146                                             SDValue &Offset) {
1147   if (shouldUseZeroOffsetLdSt(N))
1148     return false; // Select ri instead
1149   return SelectThumbAddrModeRRSext(N, Base, Offset);
1150 }
1151 
1152 bool
1153 ARMDAGToDAGISel::SelectThumbAddrModeImm5S(SDValue N, unsigned Scale,
1154                                           SDValue &Base, SDValue &OffImm) {
1155   if (shouldUseZeroOffsetLdSt(N)) {
1156     Base = N;
1157     OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1158     return true;
1159   }
1160 
1161   if (!CurDAG->isBaseWithConstantOffset(N)) {
1162     if (N.getOpcode() == ISD::ADD) {
1163       return false; // We want to select register offset instead
1164     } else if (N.getOpcode() == ARMISD::Wrapper &&
1165         N.getOperand(0).getOpcode() != ISD::TargetGlobalAddress &&
1166         N.getOperand(0).getOpcode() != ISD::TargetExternalSymbol &&
1167         N.getOperand(0).getOpcode() != ISD::TargetConstantPool &&
1168         N.getOperand(0).getOpcode() != ISD::TargetGlobalTLSAddress) {
1169       Base = N.getOperand(0);
1170     } else {
1171       Base = N;
1172     }
1173 
1174     OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1175     return true;
1176   }
1177 
1178   // If the RHS is + imm5 * scale, fold into addr mode.
1179   int RHSC;
1180   if (isScaledConstantInRange(N.getOperand(1), Scale, 0, 32, RHSC)) {
1181     Base = N.getOperand(0);
1182     OffImm = CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32);
1183     return true;
1184   }
1185 
1186   // Offset is too large, so use register offset instead.
1187   return false;
1188 }
1189 
1190 bool
1191 ARMDAGToDAGISel::SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base,
1192                                            SDValue &OffImm) {
1193   return SelectThumbAddrModeImm5S(N, 4, Base, OffImm);
1194 }
1195 
1196 bool
1197 ARMDAGToDAGISel::SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base,
1198                                            SDValue &OffImm) {
1199   return SelectThumbAddrModeImm5S(N, 2, Base, OffImm);
1200 }
1201 
1202 bool
1203 ARMDAGToDAGISel::SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base,
1204                                            SDValue &OffImm) {
1205   return SelectThumbAddrModeImm5S(N, 1, Base, OffImm);
1206 }
1207 
1208 bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDValue N,
1209                                             SDValue &Base, SDValue &OffImm) {
1210   if (N.getOpcode() == ISD::FrameIndex) {
1211     int FI = cast<FrameIndexSDNode>(N)->getIndex();
1212     // Only multiples of 4 are allowed for the offset, so the frame object
1213     // alignment must be at least 4.
1214     MachineFrameInfo &MFI = MF->getFrameInfo();
1215     if (MFI.getObjectAlign(FI) < Align(4))
1216       MFI.setObjectAlignment(FI, Align(4));
1217     Base = CurDAG->getTargetFrameIndex(
1218         FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1219     OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1220     return true;
1221   }
1222 
1223   if (!CurDAG->isBaseWithConstantOffset(N))
1224     return false;
1225 
1226   if (N.getOperand(0).getOpcode() == ISD::FrameIndex) {
1227     // If the RHS is + imm8 * scale, fold into addr mode.
1228     int RHSC;
1229     if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/4, 0, 256, RHSC)) {
1230       Base = N.getOperand(0);
1231       int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1232       // Make sure the offset is inside the object, or we might fail to
1233       // allocate an emergency spill slot. (An out-of-range access is UB, but
1234       // it could show up anyway.)
1235       MachineFrameInfo &MFI = MF->getFrameInfo();
1236       if (RHSC * 4 < MFI.getObjectSize(FI)) {
1237         // For LHS+RHS to result in an offset that's a multiple of 4 the object
1238         // indexed by the LHS must be 4-byte aligned.
1239         if (!MFI.isFixedObjectIndex(FI) && MFI.getObjectAlign(FI) < Align(4))
1240           MFI.setObjectAlignment(FI, Align(4));
1241         if (MFI.getObjectAlign(FI) >= Align(4)) {
1242           Base = CurDAG->getTargetFrameIndex(
1243               FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1244           OffImm = CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32);
1245           return true;
1246         }
1247       }
1248     }
1249   }
1250 
1251   return false;
1252 }
1253 
1254 template <unsigned Shift>
1255 bool ARMDAGToDAGISel::SelectTAddrModeImm7(SDValue N, SDValue &Base,
1256                                           SDValue &OffImm) {
1257   if (N.getOpcode() == ISD::SUB || CurDAG->isBaseWithConstantOffset(N)) {
1258     int RHSC;
1259     if (isScaledConstantInRange(N.getOperand(1), 1 << Shift, -0x7f, 0x80,
1260                                 RHSC)) {
1261       Base = N.getOperand(0);
1262       if (N.getOpcode() == ISD::SUB)
1263         RHSC = -RHSC;
1264       OffImm =
1265           CurDAG->getTargetConstant(RHSC * (1 << Shift), SDLoc(N), MVT::i32);
1266       return true;
1267     }
1268   }
1269 
1270   // Base only.
1271   Base = N;
1272   OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1273   return true;
1274 }
1275 
1276 
1277 //===----------------------------------------------------------------------===//
1278 //                        Thumb 2 Addressing Modes
1279 //===----------------------------------------------------------------------===//
1280 
1281 
1282 bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue N,
1283                                             SDValue &Base, SDValue &OffImm) {
1284   // Match simple R + imm12 operands.
1285 
1286   // Base only.
1287   if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
1288       !CurDAG->isBaseWithConstantOffset(N)) {
1289     if (N.getOpcode() == ISD::FrameIndex) {
1290       // Match frame index.
1291       int FI = cast<FrameIndexSDNode>(N)->getIndex();
1292       Base = CurDAG->getTargetFrameIndex(
1293           FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1294       OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1295       return true;
1296     }
1297 
1298     if (N.getOpcode() == ARMISD::Wrapper &&
1299         N.getOperand(0).getOpcode() != ISD::TargetGlobalAddress &&
1300         N.getOperand(0).getOpcode() != ISD::TargetExternalSymbol &&
1301         N.getOperand(0).getOpcode() != ISD::TargetGlobalTLSAddress) {
1302       Base = N.getOperand(0);
1303       if (Base.getOpcode() == ISD::TargetConstantPool)
1304         return false;  // We want to select t2LDRpci instead.
1305     } else
1306       Base = N;
1307     OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1308     return true;
1309   }
1310 
1311   if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1312     if (SelectT2AddrModeImm8(N, Base, OffImm))
1313       // Let t2LDRi8 handle (R - imm8).
1314       return false;
1315 
1316     int RHSC = (int)RHS->getZExtValue();
1317     if (N.getOpcode() == ISD::SUB)
1318       RHSC = -RHSC;
1319 
1320     if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
1321       Base   = N.getOperand(0);
1322       if (Base.getOpcode() == ISD::FrameIndex) {
1323         int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1324         Base = CurDAG->getTargetFrameIndex(
1325             FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1326       }
1327       OffImm = CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32);
1328       return true;
1329     }
1330   }
1331 
1332   // Base only.
1333   Base = N;
1334   OffImm  = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1335   return true;
1336 }
1337 
1338 template <unsigned Shift>
1339 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue N, SDValue &Base,
1340                                            SDValue &OffImm) {
1341   if (N.getOpcode() == ISD::SUB || CurDAG->isBaseWithConstantOffset(N)) {
1342     int RHSC;
1343     if (isScaledConstantInRange(N.getOperand(1), 1 << Shift, -255, 256, RHSC)) {
1344       Base = N.getOperand(0);
1345       if (Base.getOpcode() == ISD::FrameIndex) {
1346         int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1347         Base = CurDAG->getTargetFrameIndex(
1348             FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1349       }
1350 
1351       if (N.getOpcode() == ISD::SUB)
1352         RHSC = -RHSC;
1353       OffImm =
1354           CurDAG->getTargetConstant(RHSC * (1 << Shift), SDLoc(N), MVT::i32);
1355       return true;
1356     }
1357   }
1358 
1359   // Base only.
1360   Base = N;
1361   OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1362   return true;
1363 }
1364 
1365 bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue N,
1366                                            SDValue &Base, SDValue &OffImm) {
1367   // Match simple R - imm8 operands.
1368   if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
1369       !CurDAG->isBaseWithConstantOffset(N))
1370     return false;
1371 
1372   if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1373     int RHSC = (int)RHS->getSExtValue();
1374     if (N.getOpcode() == ISD::SUB)
1375       RHSC = -RHSC;
1376 
1377     if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
1378       Base = N.getOperand(0);
1379       if (Base.getOpcode() == ISD::FrameIndex) {
1380         int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1381         Base = CurDAG->getTargetFrameIndex(
1382             FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1383       }
1384       OffImm = CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32);
1385       return true;
1386     }
1387   }
1388 
1389   return false;
1390 }
1391 
1392 bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N,
1393                                                  SDValue &OffImm){
1394   unsigned Opcode = Op->getOpcode();
1395   ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
1396     ? cast<LoadSDNode>(Op)->getAddressingMode()
1397     : cast<StoreSDNode>(Op)->getAddressingMode();
1398   int RHSC;
1399   if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x100, RHSC)) { // 8 bits.
1400     OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
1401       ? CurDAG->getTargetConstant(RHSC, SDLoc(N), MVT::i32)
1402       : CurDAG->getTargetConstant(-RHSC, SDLoc(N), MVT::i32);
1403     return true;
1404   }
1405 
1406   return false;
1407 }
1408 
1409 template <unsigned Shift>
1410 bool ARMDAGToDAGISel::SelectT2AddrModeImm7(SDValue N, SDValue &Base,
1411                                            SDValue &OffImm) {
1412   if (N.getOpcode() == ISD::SUB || CurDAG->isBaseWithConstantOffset(N)) {
1413     int RHSC;
1414     if (isScaledConstantInRange(N.getOperand(1), 1 << Shift, -0x7f, 0x80,
1415                                 RHSC)) {
1416       Base = N.getOperand(0);
1417       if (Base.getOpcode() == ISD::FrameIndex) {
1418         int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1419         Base = CurDAG->getTargetFrameIndex(
1420             FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1421       }
1422 
1423       if (N.getOpcode() == ISD::SUB)
1424         RHSC = -RHSC;
1425       OffImm =
1426           CurDAG->getTargetConstant(RHSC * (1 << Shift), SDLoc(N), MVT::i32);
1427       return true;
1428     }
1429   }
1430 
1431   // Base only.
1432   Base = N;
1433   OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1434   return true;
1435 }
1436 
1437 template <unsigned Shift>
1438 bool ARMDAGToDAGISel::SelectT2AddrModeImm7Offset(SDNode *Op, SDValue N,
1439                                                  SDValue &OffImm) {
1440   return SelectT2AddrModeImm7Offset(Op, N, OffImm, Shift);
1441 }
1442 
1443 bool ARMDAGToDAGISel::SelectT2AddrModeImm7Offset(SDNode *Op, SDValue N,
1444                                                  SDValue &OffImm,
1445                                                  unsigned Shift) {
1446   unsigned Opcode = Op->getOpcode();
1447   ISD::MemIndexedMode AM;
1448   switch (Opcode) {
1449   case ISD::LOAD:
1450     AM = cast<LoadSDNode>(Op)->getAddressingMode();
1451     break;
1452   case ISD::STORE:
1453     AM = cast<StoreSDNode>(Op)->getAddressingMode();
1454     break;
1455   case ISD::MLOAD:
1456     AM = cast<MaskedLoadSDNode>(Op)->getAddressingMode();
1457     break;
1458   case ISD::MSTORE:
1459     AM = cast<MaskedStoreSDNode>(Op)->getAddressingMode();
1460     break;
1461   default:
1462     llvm_unreachable("Unexpected Opcode for Imm7Offset");
1463   }
1464 
1465   int RHSC;
1466   // 7 bit constant, shifted by Shift.
1467   if (isScaledConstantInRange(N, 1 << Shift, 0, 0x80, RHSC)) {
1468     OffImm =
1469         ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
1470             ? CurDAG->getTargetConstant(RHSC * (1 << Shift), SDLoc(N), MVT::i32)
1471             : CurDAG->getTargetConstant(-RHSC * (1 << Shift), SDLoc(N),
1472                                         MVT::i32);
1473     return true;
1474   }
1475   return false;
1476 }
1477 
1478 template <int Min, int Max>
1479 bool ARMDAGToDAGISel::SelectImmediateInRange(SDValue N, SDValue &OffImm) {
1480   int Val;
1481   if (isScaledConstantInRange(N, 1, Min, Max, Val)) {
1482     OffImm = CurDAG->getTargetConstant(Val, SDLoc(N), MVT::i32);
1483     return true;
1484   }
1485   return false;
1486 }
1487 
1488 bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue N,
1489                                             SDValue &Base,
1490                                             SDValue &OffReg, SDValue &ShImm) {
1491   // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
1492   if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(N))
1493     return false;
1494 
1495   // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
1496   if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1497     int RHSC = (int)RHS->getZExtValue();
1498     if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
1499       return false;
1500     else if (RHSC < 0 && RHSC >= -255) // 8 bits
1501       return false;
1502   }
1503 
1504   // Look for (R + R) or (R + (R << [1,2,3])).
1505   unsigned ShAmt = 0;
1506   Base   = N.getOperand(0);
1507   OffReg = N.getOperand(1);
1508 
1509   // Swap if it is ((R << c) + R).
1510   ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg.getOpcode());
1511   if (ShOpcVal != ARM_AM::lsl) {
1512     ShOpcVal = ARM_AM::getShiftOpcForNode(Base.getOpcode());
1513     if (ShOpcVal == ARM_AM::lsl)
1514       std::swap(Base, OffReg);
1515   }
1516 
1517   if (ShOpcVal == ARM_AM::lsl) {
1518     // Check to see if the RHS of the shift is a constant, if not, we can't fold
1519     // it.
1520     if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
1521       ShAmt = Sh->getZExtValue();
1522       if (ShAmt < 4 && isShifterOpProfitable(OffReg, ShOpcVal, ShAmt))
1523         OffReg = OffReg.getOperand(0);
1524       else {
1525         ShAmt = 0;
1526       }
1527     }
1528   }
1529 
1530   // If OffReg is a multiply-by-constant and it's profitable to extract a shift
1531   // and use it in a shifted operand do so.
1532   if (OffReg.getOpcode() == ISD::MUL && N.hasOneUse()) {
1533     unsigned PowerOfTwo = 0;
1534     SDValue NewMulConst;
1535     if (canExtractShiftFromMul(OffReg, 3, PowerOfTwo, NewMulConst)) {
1536       HandleSDNode Handle(OffReg);
1537       replaceDAGValue(OffReg.getOperand(1), NewMulConst);
1538       OffReg = Handle.getValue();
1539       ShAmt = PowerOfTwo;
1540     }
1541   }
1542 
1543   ShImm = CurDAG->getTargetConstant(ShAmt, SDLoc(N), MVT::i32);
1544 
1545   return true;
1546 }
1547 
1548 bool ARMDAGToDAGISel::SelectT2AddrModeExclusive(SDValue N, SDValue &Base,
1549                                                 SDValue &OffImm) {
1550   // This *must* succeed since it's used for the irreplaceable ldrex and strex
1551   // instructions.
1552   Base = N;
1553   OffImm = CurDAG->getTargetConstant(0, SDLoc(N), MVT::i32);
1554 
1555   if (N.getOpcode() != ISD::ADD || !CurDAG->isBaseWithConstantOffset(N))
1556     return true;
1557 
1558   ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1));
1559   if (!RHS)
1560     return true;
1561 
1562   uint32_t RHSC = (int)RHS->getZExtValue();
1563   if (RHSC > 1020 || RHSC % 4 != 0)
1564     return true;
1565 
1566   Base = N.getOperand(0);
1567   if (Base.getOpcode() == ISD::FrameIndex) {
1568     int FI = cast<FrameIndexSDNode>(Base)->getIndex();
1569     Base = CurDAG->getTargetFrameIndex(
1570         FI, TLI->getPointerTy(CurDAG->getDataLayout()));
1571   }
1572 
1573   OffImm = CurDAG->getTargetConstant(RHSC/4, SDLoc(N), MVT::i32);
1574   return true;
1575 }
1576 
1577 //===--------------------------------------------------------------------===//
1578 
1579 /// getAL - Returns a ARMCC::AL immediate node.
1580 static inline SDValue getAL(SelectionDAG *CurDAG, const SDLoc &dl) {
1581   return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, dl, MVT::i32);
1582 }
1583 
1584 void ARMDAGToDAGISel::transferMemOperands(SDNode *N, SDNode *Result) {
1585   MachineMemOperand *MemOp = cast<MemSDNode>(N)->getMemOperand();
1586   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Result), {MemOp});
1587 }
1588 
1589 bool ARMDAGToDAGISel::tryARMIndexedLoad(SDNode *N) {
1590   LoadSDNode *LD = cast<LoadSDNode>(N);
1591   ISD::MemIndexedMode AM = LD->getAddressingMode();
1592   if (AM == ISD::UNINDEXED)
1593     return false;
1594 
1595   EVT LoadedVT = LD->getMemoryVT();
1596   SDValue Offset, AMOpc;
1597   bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1598   unsigned Opcode = 0;
1599   bool Match = false;
1600   if (LoadedVT == MVT::i32 && isPre &&
1601       SelectAddrMode2OffsetImmPre(N, LD->getOffset(), Offset, AMOpc)) {
1602     Opcode = ARM::LDR_PRE_IMM;
1603     Match = true;
1604   } else if (LoadedVT == MVT::i32 && !isPre &&
1605       SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) {
1606     Opcode = ARM::LDR_POST_IMM;
1607     Match = true;
1608   } else if (LoadedVT == MVT::i32 &&
1609       SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) {
1610     Opcode = isPre ? ARM::LDR_PRE_REG : ARM::LDR_POST_REG;
1611     Match = true;
1612 
1613   } else if (LoadedVT == MVT::i16 &&
1614              SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
1615     Match = true;
1616     Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
1617       ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
1618       : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
1619   } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
1620     if (LD->getExtensionType() == ISD::SEXTLOAD) {
1621       if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) {
1622         Match = true;
1623         Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
1624       }
1625     } else {
1626       if (isPre &&
1627           SelectAddrMode2OffsetImmPre(N, LD->getOffset(), Offset, AMOpc)) {
1628         Match = true;
1629         Opcode = ARM::LDRB_PRE_IMM;
1630       } else if (!isPre &&
1631                   SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) {
1632         Match = true;
1633         Opcode = ARM::LDRB_POST_IMM;
1634       } else if (SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) {
1635         Match = true;
1636         Opcode = isPre ? ARM::LDRB_PRE_REG : ARM::LDRB_POST_REG;
1637       }
1638     }
1639   }
1640 
1641   if (Match) {
1642     if (Opcode == ARM::LDR_PRE_IMM || Opcode == ARM::LDRB_PRE_IMM) {
1643       SDValue Chain = LD->getChain();
1644       SDValue Base = LD->getBasePtr();
1645       SDValue Ops[]= { Base, AMOpc, getAL(CurDAG, SDLoc(N)),
1646                        CurDAG->getRegister(0, MVT::i32), Chain };
1647       SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, MVT::i32,
1648                                            MVT::Other, Ops);
1649       transferMemOperands(N, New);
1650       ReplaceNode(N, New);
1651       return true;
1652     } else {
1653       SDValue Chain = LD->getChain();
1654       SDValue Base = LD->getBasePtr();
1655       SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG, SDLoc(N)),
1656                        CurDAG->getRegister(0, MVT::i32), Chain };
1657       SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, MVT::i32,
1658                                            MVT::Other, Ops);
1659       transferMemOperands(N, New);
1660       ReplaceNode(N, New);
1661       return true;
1662     }
1663   }
1664 
1665   return false;
1666 }
1667 
1668 bool ARMDAGToDAGISel::tryT1IndexedLoad(SDNode *N) {
1669   LoadSDNode *LD = cast<LoadSDNode>(N);
1670   EVT LoadedVT = LD->getMemoryVT();
1671   ISD::MemIndexedMode AM = LD->getAddressingMode();
1672   if (AM != ISD::POST_INC || LD->getExtensionType() != ISD::NON_EXTLOAD ||
1673       LoadedVT.getSimpleVT().SimpleTy != MVT::i32)
1674     return false;
1675 
1676   auto *COffs = dyn_cast<ConstantSDNode>(LD->getOffset());
1677   if (!COffs || COffs->getZExtValue() != 4)
1678     return false;
1679 
1680   // A T1 post-indexed load is just a single register LDM: LDM r0!, {r1}.
1681   // The encoding of LDM is not how the rest of ISel expects a post-inc load to
1682   // look however, so we use a pseudo here and switch it for a tLDMIA_UPD after
1683   // ISel.
1684   SDValue Chain = LD->getChain();
1685   SDValue Base = LD->getBasePtr();
1686   SDValue Ops[]= { Base, getAL(CurDAG, SDLoc(N)),
1687                    CurDAG->getRegister(0, MVT::i32), Chain };
1688   SDNode *New = CurDAG->getMachineNode(ARM::tLDR_postidx, SDLoc(N), MVT::i32,
1689                                        MVT::i32, MVT::Other, Ops);
1690   transferMemOperands(N, New);
1691   ReplaceNode(N, New);
1692   return true;
1693 }
1694 
1695 bool ARMDAGToDAGISel::tryT2IndexedLoad(SDNode *N) {
1696   LoadSDNode *LD = cast<LoadSDNode>(N);
1697   ISD::MemIndexedMode AM = LD->getAddressingMode();
1698   if (AM == ISD::UNINDEXED)
1699     return false;
1700 
1701   EVT LoadedVT = LD->getMemoryVT();
1702   bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
1703   SDValue Offset;
1704   bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1705   unsigned Opcode = 0;
1706   bool Match = false;
1707   if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) {
1708     switch (LoadedVT.getSimpleVT().SimpleTy) {
1709     case MVT::i32:
1710       Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
1711       break;
1712     case MVT::i16:
1713       if (isSExtLd)
1714         Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
1715       else
1716         Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
1717       break;
1718     case MVT::i8:
1719     case MVT::i1:
1720       if (isSExtLd)
1721         Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
1722       else
1723         Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
1724       break;
1725     default:
1726       return false;
1727     }
1728     Match = true;
1729   }
1730 
1731   if (Match) {
1732     SDValue Chain = LD->getChain();
1733     SDValue Base = LD->getBasePtr();
1734     SDValue Ops[]= { Base, Offset, getAL(CurDAG, SDLoc(N)),
1735                      CurDAG->getRegister(0, MVT::i32), Chain };
1736     SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, MVT::i32,
1737                                          MVT::Other, Ops);
1738     transferMemOperands(N, New);
1739     ReplaceNode(N, New);
1740     return true;
1741   }
1742 
1743   return false;
1744 }
1745 
1746 bool ARMDAGToDAGISel::tryMVEIndexedLoad(SDNode *N) {
1747   EVT LoadedVT;
1748   unsigned Opcode = 0;
1749   bool isSExtLd, isPre;
1750   Align Alignment;
1751   ARMVCC::VPTCodes Pred;
1752   SDValue PredReg;
1753   SDValue Chain, Base, Offset;
1754 
1755   if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1756     ISD::MemIndexedMode AM = LD->getAddressingMode();
1757     if (AM == ISD::UNINDEXED)
1758       return false;
1759     LoadedVT = LD->getMemoryVT();
1760     if (!LoadedVT.isVector())
1761       return false;
1762 
1763     Chain = LD->getChain();
1764     Base = LD->getBasePtr();
1765     Offset = LD->getOffset();
1766     Alignment = LD->getAlign();
1767     isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
1768     isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1769     Pred = ARMVCC::None;
1770     PredReg = CurDAG->getRegister(0, MVT::i32);
1771   } else if (MaskedLoadSDNode *LD = dyn_cast<MaskedLoadSDNode>(N)) {
1772     ISD::MemIndexedMode AM = LD->getAddressingMode();
1773     if (AM == ISD::UNINDEXED)
1774       return false;
1775     LoadedVT = LD->getMemoryVT();
1776     if (!LoadedVT.isVector())
1777       return false;
1778 
1779     Chain = LD->getChain();
1780     Base = LD->getBasePtr();
1781     Offset = LD->getOffset();
1782     Alignment = LD->getAlign();
1783     isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
1784     isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
1785     Pred = ARMVCC::Then;
1786     PredReg = LD->getMask();
1787   } else
1788     llvm_unreachable("Expected a Load or a Masked Load!");
1789 
1790   // We allow LE non-masked loads to change the type (for example use a vldrb.8
1791   // as opposed to a vldrw.32). This can allow extra addressing modes or
1792   // alignments for what is otherwise an equivalent instruction.
1793   bool CanChangeType = Subtarget->isLittle() && !isa<MaskedLoadSDNode>(N);
1794 
1795   SDValue NewOffset;
1796   if (Alignment >= Align(2) && LoadedVT == MVT::v4i16 &&
1797       SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 1)) {
1798     if (isSExtLd)
1799       Opcode = isPre ? ARM::MVE_VLDRHS32_pre : ARM::MVE_VLDRHS32_post;
1800     else
1801       Opcode = isPre ? ARM::MVE_VLDRHU32_pre : ARM::MVE_VLDRHU32_post;
1802   } else if (LoadedVT == MVT::v8i8 &&
1803              SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 0)) {
1804     if (isSExtLd)
1805       Opcode = isPre ? ARM::MVE_VLDRBS16_pre : ARM::MVE_VLDRBS16_post;
1806     else
1807       Opcode = isPre ? ARM::MVE_VLDRBU16_pre : ARM::MVE_VLDRBU16_post;
1808   } else if (LoadedVT == MVT::v4i8 &&
1809              SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 0)) {
1810     if (isSExtLd)
1811       Opcode = isPre ? ARM::MVE_VLDRBS32_pre : ARM::MVE_VLDRBS32_post;
1812     else
1813       Opcode = isPre ? ARM::MVE_VLDRBU32_pre : ARM::MVE_VLDRBU32_post;
1814   } else if (Alignment >= Align(4) &&
1815              (CanChangeType || LoadedVT == MVT::v4i32 ||
1816               LoadedVT == MVT::v4f32) &&
1817              SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 2))
1818     Opcode = isPre ? ARM::MVE_VLDRWU32_pre : ARM::MVE_VLDRWU32_post;
1819   else if (Alignment >= Align(2) &&
1820            (CanChangeType || LoadedVT == MVT::v8i16 ||
1821             LoadedVT == MVT::v8f16) &&
1822            SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 1))
1823     Opcode = isPre ? ARM::MVE_VLDRHU16_pre : ARM::MVE_VLDRHU16_post;
1824   else if ((CanChangeType || LoadedVT == MVT::v16i8) &&
1825            SelectT2AddrModeImm7Offset(N, Offset, NewOffset, 0))
1826     Opcode = isPre ? ARM::MVE_VLDRBU8_pre : ARM::MVE_VLDRBU8_post;
1827   else
1828     return false;
1829 
1830   SDValue Ops[] = {Base,
1831                    NewOffset,
1832                    CurDAG->getTargetConstant(Pred, SDLoc(N), MVT::i32),
1833                    PredReg,
1834                    CurDAG->getRegister(0, MVT::i32), // tp_reg
1835                    Chain};
1836   SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32,
1837                                        N->getValueType(0), MVT::Other, Ops);
1838   transferMemOperands(N, New);
1839   ReplaceUses(SDValue(N, 0), SDValue(New, 1));
1840   ReplaceUses(SDValue(N, 1), SDValue(New, 0));
1841   ReplaceUses(SDValue(N, 2), SDValue(New, 2));
1842   CurDAG->RemoveDeadNode(N);
1843   return true;
1844 }
1845 
1846 /// Form a GPRPair pseudo register from a pair of GPR regs.
1847 SDNode *ARMDAGToDAGISel::createGPRPairNode(EVT VT, SDValue V0, SDValue V1) {
1848   SDLoc dl(V0.getNode());
1849   SDValue RegClass =
1850     CurDAG->getTargetConstant(ARM::GPRPairRegClassID, dl, MVT::i32);
1851   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::gsub_0, dl, MVT::i32);
1852   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::gsub_1, dl, MVT::i32);
1853   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1854   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1855 }
1856 
1857 /// Form a D register from a pair of S registers.
1858 SDNode *ARMDAGToDAGISel::createSRegPairNode(EVT VT, SDValue V0, SDValue V1) {
1859   SDLoc dl(V0.getNode());
1860   SDValue RegClass =
1861     CurDAG->getTargetConstant(ARM::DPR_VFP2RegClassID, dl, MVT::i32);
1862   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, dl, MVT::i32);
1863   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, dl, MVT::i32);
1864   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1865   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1866 }
1867 
1868 /// Form a quad register from a pair of D registers.
1869 SDNode *ARMDAGToDAGISel::createDRegPairNode(EVT VT, SDValue V0, SDValue V1) {
1870   SDLoc dl(V0.getNode());
1871   SDValue RegClass = CurDAG->getTargetConstant(ARM::QPRRegClassID, dl,
1872                                                MVT::i32);
1873   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, dl, MVT::i32);
1874   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, dl, MVT::i32);
1875   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1876   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1877 }
1878 
1879 /// Form 4 consecutive D registers from a pair of Q registers.
1880 SDNode *ARMDAGToDAGISel::createQRegPairNode(EVT VT, SDValue V0, SDValue V1) {
1881   SDLoc dl(V0.getNode());
1882   SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, dl,
1883                                                MVT::i32);
1884   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, dl, MVT::i32);
1885   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, dl, MVT::i32);
1886   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 };
1887   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1888 }
1889 
1890 /// Form 4 consecutive S registers.
1891 SDNode *ARMDAGToDAGISel::createQuadSRegsNode(EVT VT, SDValue V0, SDValue V1,
1892                                    SDValue V2, SDValue V3) {
1893   SDLoc dl(V0.getNode());
1894   SDValue RegClass =
1895     CurDAG->getTargetConstant(ARM::QPR_VFP2RegClassID, dl, MVT::i32);
1896   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, dl, MVT::i32);
1897   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, dl, MVT::i32);
1898   SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, dl, MVT::i32);
1899   SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, dl, MVT::i32);
1900   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1901                                     V2, SubReg2, V3, SubReg3 };
1902   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1903 }
1904 
1905 /// Form 4 consecutive D registers.
1906 SDNode *ARMDAGToDAGISel::createQuadDRegsNode(EVT VT, SDValue V0, SDValue V1,
1907                                    SDValue V2, SDValue V3) {
1908   SDLoc dl(V0.getNode());
1909   SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, dl,
1910                                                MVT::i32);
1911   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, dl, MVT::i32);
1912   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, dl, MVT::i32);
1913   SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, dl, MVT::i32);
1914   SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, dl, MVT::i32);
1915   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1916                                     V2, SubReg2, V3, SubReg3 };
1917   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1918 }
1919 
1920 /// Form 4 consecutive Q registers.
1921 SDNode *ARMDAGToDAGISel::createQuadQRegsNode(EVT VT, SDValue V0, SDValue V1,
1922                                    SDValue V2, SDValue V3) {
1923   SDLoc dl(V0.getNode());
1924   SDValue RegClass = CurDAG->getTargetConstant(ARM::QQQQPRRegClassID, dl,
1925                                                MVT::i32);
1926   SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, dl, MVT::i32);
1927   SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, dl, MVT::i32);
1928   SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, dl, MVT::i32);
1929   SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, dl, MVT::i32);
1930   const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1,
1931                                     V2, SubReg2, V3, SubReg3 };
1932   return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops);
1933 }
1934 
1935 /// GetVLDSTAlign - Get the alignment (in bytes) for the alignment operand
1936 /// of a NEON VLD or VST instruction.  The supported values depend on the
1937 /// number of registers being loaded.
1938 SDValue ARMDAGToDAGISel::GetVLDSTAlign(SDValue Align, const SDLoc &dl,
1939                                        unsigned NumVecs, bool is64BitVector) {
1940   unsigned NumRegs = NumVecs;
1941   if (!is64BitVector && NumVecs < 3)
1942     NumRegs *= 2;
1943 
1944   unsigned Alignment = Align->getAsZExtVal();
1945   if (Alignment >= 32 && NumRegs == 4)
1946     Alignment = 32;
1947   else if (Alignment >= 16 && (NumRegs == 2 || NumRegs == 4))
1948     Alignment = 16;
1949   else if (Alignment >= 8)
1950     Alignment = 8;
1951   else
1952     Alignment = 0;
1953 
1954   return CurDAG->getTargetConstant(Alignment, dl, MVT::i32);
1955 }
1956 
1957 static bool isVLDfixed(unsigned Opc)
1958 {
1959   switch (Opc) {
1960   default: return false;
1961   case ARM::VLD1d8wb_fixed : return true;
1962   case ARM::VLD1d16wb_fixed : return true;
1963   case ARM::VLD1d64Qwb_fixed : return true;
1964   case ARM::VLD1d32wb_fixed : return true;
1965   case ARM::VLD1d64wb_fixed : return true;
1966   case ARM::VLD1d8TPseudoWB_fixed : return true;
1967   case ARM::VLD1d16TPseudoWB_fixed : return true;
1968   case ARM::VLD1d32TPseudoWB_fixed : return true;
1969   case ARM::VLD1d64TPseudoWB_fixed : return true;
1970   case ARM::VLD1d8QPseudoWB_fixed : return true;
1971   case ARM::VLD1d16QPseudoWB_fixed : return true;
1972   case ARM::VLD1d32QPseudoWB_fixed : return true;
1973   case ARM::VLD1d64QPseudoWB_fixed : return true;
1974   case ARM::VLD1q8wb_fixed : return true;
1975   case ARM::VLD1q16wb_fixed : return true;
1976   case ARM::VLD1q32wb_fixed : return true;
1977   case ARM::VLD1q64wb_fixed : return true;
1978   case ARM::VLD1DUPd8wb_fixed : return true;
1979   case ARM::VLD1DUPd16wb_fixed : return true;
1980   case ARM::VLD1DUPd32wb_fixed : return true;
1981   case ARM::VLD1DUPq8wb_fixed : return true;
1982   case ARM::VLD1DUPq16wb_fixed : return true;
1983   case ARM::VLD1DUPq32wb_fixed : return true;
1984   case ARM::VLD2d8wb_fixed : return true;
1985   case ARM::VLD2d16wb_fixed : return true;
1986   case ARM::VLD2d32wb_fixed : return true;
1987   case ARM::VLD2q8PseudoWB_fixed : return true;
1988   case ARM::VLD2q16PseudoWB_fixed : return true;
1989   case ARM::VLD2q32PseudoWB_fixed : return true;
1990   case ARM::VLD2DUPd8wb_fixed : return true;
1991   case ARM::VLD2DUPd16wb_fixed : return true;
1992   case ARM::VLD2DUPd32wb_fixed : return true;
1993   case ARM::VLD2DUPq8OddPseudoWB_fixed: return true;
1994   case ARM::VLD2DUPq16OddPseudoWB_fixed: return true;
1995   case ARM::VLD2DUPq32OddPseudoWB_fixed: return true;
1996   }
1997 }
1998 
1999 static bool isVSTfixed(unsigned Opc)
2000 {
2001   switch (Opc) {
2002   default: return false;
2003   case ARM::VST1d8wb_fixed : return true;
2004   case ARM::VST1d16wb_fixed : return true;
2005   case ARM::VST1d32wb_fixed : return true;
2006   case ARM::VST1d64wb_fixed : return true;
2007   case ARM::VST1q8wb_fixed : return true;
2008   case ARM::VST1q16wb_fixed : return true;
2009   case ARM::VST1q32wb_fixed : return true;
2010   case ARM::VST1q64wb_fixed : return true;
2011   case ARM::VST1d8TPseudoWB_fixed : return true;
2012   case ARM::VST1d16TPseudoWB_fixed : return true;
2013   case ARM::VST1d32TPseudoWB_fixed : return true;
2014   case ARM::VST1d64TPseudoWB_fixed : return true;
2015   case ARM::VST1d8QPseudoWB_fixed : return true;
2016   case ARM::VST1d16QPseudoWB_fixed : return true;
2017   case ARM::VST1d32QPseudoWB_fixed : return true;
2018   case ARM::VST1d64QPseudoWB_fixed : return true;
2019   case ARM::VST2d8wb_fixed : return true;
2020   case ARM::VST2d16wb_fixed : return true;
2021   case ARM::VST2d32wb_fixed : return true;
2022   case ARM::VST2q8PseudoWB_fixed : return true;
2023   case ARM::VST2q16PseudoWB_fixed : return true;
2024   case ARM::VST2q32PseudoWB_fixed : return true;
2025   }
2026 }
2027 
2028 // Get the register stride update opcode of a VLD/VST instruction that
2029 // is otherwise equivalent to the given fixed stride updating instruction.
2030 static unsigned getVLDSTRegisterUpdateOpcode(unsigned Opc) {
2031   assert((isVLDfixed(Opc) || isVSTfixed(Opc))
2032     && "Incorrect fixed stride updating instruction.");
2033   switch (Opc) {
2034   default: break;
2035   case ARM::VLD1d8wb_fixed: return ARM::VLD1d8wb_register;
2036   case ARM::VLD1d16wb_fixed: return ARM::VLD1d16wb_register;
2037   case ARM::VLD1d32wb_fixed: return ARM::VLD1d32wb_register;
2038   case ARM::VLD1d64wb_fixed: return ARM::VLD1d64wb_register;
2039   case ARM::VLD1q8wb_fixed: return ARM::VLD1q8wb_register;
2040   case ARM::VLD1q16wb_fixed: return ARM::VLD1q16wb_register;
2041   case ARM::VLD1q32wb_fixed: return ARM::VLD1q32wb_register;
2042   case ARM::VLD1q64wb_fixed: return ARM::VLD1q64wb_register;
2043   case ARM::VLD1d64Twb_fixed: return ARM::VLD1d64Twb_register;
2044   case ARM::VLD1d64Qwb_fixed: return ARM::VLD1d64Qwb_register;
2045   case ARM::VLD1d8TPseudoWB_fixed: return ARM::VLD1d8TPseudoWB_register;
2046   case ARM::VLD1d16TPseudoWB_fixed: return ARM::VLD1d16TPseudoWB_register;
2047   case ARM::VLD1d32TPseudoWB_fixed: return ARM::VLD1d32TPseudoWB_register;
2048   case ARM::VLD1d64TPseudoWB_fixed: return ARM::VLD1d64TPseudoWB_register;
2049   case ARM::VLD1d8QPseudoWB_fixed: return ARM::VLD1d8QPseudoWB_register;
2050   case ARM::VLD1d16QPseudoWB_fixed: return ARM::VLD1d16QPseudoWB_register;
2051   case ARM::VLD1d32QPseudoWB_fixed: return ARM::VLD1d32QPseudoWB_register;
2052   case ARM::VLD1d64QPseudoWB_fixed: return ARM::VLD1d64QPseudoWB_register;
2053   case ARM::VLD1DUPd8wb_fixed : return ARM::VLD1DUPd8wb_register;
2054   case ARM::VLD1DUPd16wb_fixed : return ARM::VLD1DUPd16wb_register;
2055   case ARM::VLD1DUPd32wb_fixed : return ARM::VLD1DUPd32wb_register;
2056   case ARM::VLD1DUPq8wb_fixed : return ARM::VLD1DUPq8wb_register;
2057   case ARM::VLD1DUPq16wb_fixed : return ARM::VLD1DUPq16wb_register;
2058   case ARM::VLD1DUPq32wb_fixed : return ARM::VLD1DUPq32wb_register;
2059   case ARM::VLD2DUPq8OddPseudoWB_fixed: return ARM::VLD2DUPq8OddPseudoWB_register;
2060   case ARM::VLD2DUPq16OddPseudoWB_fixed: return ARM::VLD2DUPq16OddPseudoWB_register;
2061   case ARM::VLD2DUPq32OddPseudoWB_fixed: return ARM::VLD2DUPq32OddPseudoWB_register;
2062 
2063   case ARM::VST1d8wb_fixed: return ARM::VST1d8wb_register;
2064   case ARM::VST1d16wb_fixed: return ARM::VST1d16wb_register;
2065   case ARM::VST1d32wb_fixed: return ARM::VST1d32wb_register;
2066   case ARM::VST1d64wb_fixed: return ARM::VST1d64wb_register;
2067   case ARM::VST1q8wb_fixed: return ARM::VST1q8wb_register;
2068   case ARM::VST1q16wb_fixed: return ARM::VST1q16wb_register;
2069   case ARM::VST1q32wb_fixed: return ARM::VST1q32wb_register;
2070   case ARM::VST1q64wb_fixed: return ARM::VST1q64wb_register;
2071   case ARM::VST1d8TPseudoWB_fixed: return ARM::VST1d8TPseudoWB_register;
2072   case ARM::VST1d16TPseudoWB_fixed: return ARM::VST1d16TPseudoWB_register;
2073   case ARM::VST1d32TPseudoWB_fixed: return ARM::VST1d32TPseudoWB_register;
2074   case ARM::VST1d64TPseudoWB_fixed: return ARM::VST1d64TPseudoWB_register;
2075   case ARM::VST1d8QPseudoWB_fixed: return ARM::VST1d8QPseudoWB_register;
2076   case ARM::VST1d16QPseudoWB_fixed: return ARM::VST1d16QPseudoWB_register;
2077   case ARM::VST1d32QPseudoWB_fixed: return ARM::VST1d32QPseudoWB_register;
2078   case ARM::VST1d64QPseudoWB_fixed: return ARM::VST1d64QPseudoWB_register;
2079 
2080   case ARM::VLD2d8wb_fixed: return ARM::VLD2d8wb_register;
2081   case ARM::VLD2d16wb_fixed: return ARM::VLD2d16wb_register;
2082   case ARM::VLD2d32wb_fixed: return ARM::VLD2d32wb_register;
2083   case ARM::VLD2q8PseudoWB_fixed: return ARM::VLD2q8PseudoWB_register;
2084   case ARM::VLD2q16PseudoWB_fixed: return ARM::VLD2q16PseudoWB_register;
2085   case ARM::VLD2q32PseudoWB_fixed: return ARM::VLD2q32PseudoWB_register;
2086 
2087   case ARM::VST2d8wb_fixed: return ARM::VST2d8wb_register;
2088   case ARM::VST2d16wb_fixed: return ARM::VST2d16wb_register;
2089   case ARM::VST2d32wb_fixed: return ARM::VST2d32wb_register;
2090   case ARM::VST2q8PseudoWB_fixed: return ARM::VST2q8PseudoWB_register;
2091   case ARM::VST2q16PseudoWB_fixed: return ARM::VST2q16PseudoWB_register;
2092   case ARM::VST2q32PseudoWB_fixed: return ARM::VST2q32PseudoWB_register;
2093 
2094   case ARM::VLD2DUPd8wb_fixed: return ARM::VLD2DUPd8wb_register;
2095   case ARM::VLD2DUPd16wb_fixed: return ARM::VLD2DUPd16wb_register;
2096   case ARM::VLD2DUPd32wb_fixed: return ARM::VLD2DUPd32wb_register;
2097   }
2098   return Opc; // If not one we handle, return it unchanged.
2099 }
2100 
2101 /// Returns true if the given increment is a Constant known to be equal to the
2102 /// access size performed by a NEON load/store. This means the "[rN]!" form can
2103 /// be used.
2104 static bool isPerfectIncrement(SDValue Inc, EVT VecTy, unsigned NumVecs) {
2105   auto C = dyn_cast<ConstantSDNode>(Inc);
2106   return C && C->getZExtValue() == VecTy.getSizeInBits() / 8 * NumVecs;
2107 }
2108 
2109 void ARMDAGToDAGISel::SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs,
2110                                 const uint16_t *DOpcodes,
2111                                 const uint16_t *QOpcodes0,
2112                                 const uint16_t *QOpcodes1) {
2113   assert(Subtarget->hasNEON());
2114   assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");
2115   SDLoc dl(N);
2116 
2117   SDValue MemAddr, Align;
2118   bool IsIntrinsic = !isUpdating;  // By coincidence, all supported updating
2119                                    // nodes are not intrinsics.
2120   unsigned AddrOpIdx = IsIntrinsic ? 2 : 1;
2121   if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
2122     return;
2123 
2124   SDValue Chain = N->getOperand(0);
2125   EVT VT = N->getValueType(0);
2126   bool is64BitVector = VT.is64BitVector();
2127   Align = GetVLDSTAlign(Align, dl, NumVecs, is64BitVector);
2128 
2129   unsigned OpcodeIndex;
2130   switch (VT.getSimpleVT().SimpleTy) {
2131   default: llvm_unreachable("unhandled vld type");
2132     // Double-register operations:
2133   case MVT::v8i8:  OpcodeIndex = 0; break;
2134   case MVT::v4f16:
2135   case MVT::v4bf16:
2136   case MVT::v4i16: OpcodeIndex = 1; break;
2137   case MVT::v2f32:
2138   case MVT::v2i32: OpcodeIndex = 2; break;
2139   case MVT::v1i64: OpcodeIndex = 3; break;
2140     // Quad-register operations:
2141   case MVT::v16i8: OpcodeIndex = 0; break;
2142   case MVT::v8f16:
2143   case MVT::v8bf16:
2144   case MVT::v8i16: OpcodeIndex = 1; break;
2145   case MVT::v4f32:
2146   case MVT::v4i32: OpcodeIndex = 2; break;
2147   case MVT::v2f64:
2148   case MVT::v2i64: OpcodeIndex = 3; break;
2149   }
2150 
2151   EVT ResTy;
2152   if (NumVecs == 1)
2153     ResTy = VT;
2154   else {
2155     unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
2156     if (!is64BitVector)
2157       ResTyElts *= 2;
2158     ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, ResTyElts);
2159   }
2160   std::vector<EVT> ResTys;
2161   ResTys.push_back(ResTy);
2162   if (isUpdating)
2163     ResTys.push_back(MVT::i32);
2164   ResTys.push_back(MVT::Other);
2165 
2166   SDValue Pred = getAL(CurDAG, dl);
2167   SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2168   SDNode *VLd;
2169   SmallVector<SDValue, 7> Ops;
2170 
2171   // Double registers and VLD1/VLD2 quad registers are directly supported.
2172   if (is64BitVector || NumVecs <= 2) {
2173     unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
2174                     QOpcodes0[OpcodeIndex]);
2175     Ops.push_back(MemAddr);
2176     Ops.push_back(Align);
2177     if (isUpdating) {
2178       SDValue Inc = N->getOperand(AddrOpIdx + 1);
2179       bool IsImmUpdate = isPerfectIncrement(Inc, VT, NumVecs);
2180       if (!IsImmUpdate) {
2181         // We use a VLD1 for v1i64 even if the pseudo says vld2/3/4, so
2182         // check for the opcode rather than the number of vector elements.
2183         if (isVLDfixed(Opc))
2184           Opc = getVLDSTRegisterUpdateOpcode(Opc);
2185         Ops.push_back(Inc);
2186       // VLD1/VLD2 fixed increment does not need Reg0 so only include it in
2187       // the operands if not such an opcode.
2188       } else if (!isVLDfixed(Opc))
2189         Ops.push_back(Reg0);
2190     }
2191     Ops.push_back(Pred);
2192     Ops.push_back(Reg0);
2193     Ops.push_back(Chain);
2194     VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
2195 
2196   } else {
2197     // Otherwise, quad registers are loaded with two separate instructions,
2198     // where one loads the even registers and the other loads the odd registers.
2199     EVT AddrTy = MemAddr.getValueType();
2200 
2201     // Load the even subregs.  This is always an updating load, so that it
2202     // provides the address to the second load for the odd subregs.
2203     SDValue ImplDef =
2204       SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, ResTy), 0);
2205     const SDValue OpsA[] = { MemAddr, Align, Reg0, ImplDef, Pred, Reg0, Chain };
2206     SDNode *VLdA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl,
2207                                           ResTy, AddrTy, MVT::Other, OpsA);
2208     Chain = SDValue(VLdA, 2);
2209 
2210     // Load the odd subregs.
2211     Ops.push_back(SDValue(VLdA, 1));
2212     Ops.push_back(Align);
2213     if (isUpdating) {
2214       SDValue Inc = N->getOperand(AddrOpIdx + 1);
2215       assert(isa<ConstantSDNode>(Inc.getNode()) &&
2216              "only constant post-increment update allowed for VLD3/4");
2217       (void)Inc;
2218       Ops.push_back(Reg0);
2219     }
2220     Ops.push_back(SDValue(VLdA, 0));
2221     Ops.push_back(Pred);
2222     Ops.push_back(Reg0);
2223     Ops.push_back(Chain);
2224     VLd = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys, Ops);
2225   }
2226 
2227   // Transfer memoperands.
2228   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2229   CurDAG->setNodeMemRefs(cast<MachineSDNode>(VLd), {MemOp});
2230 
2231   if (NumVecs == 1) {
2232     ReplaceNode(N, VLd);
2233     return;
2234   }
2235 
2236   // Extract out the subregisters.
2237   SDValue SuperReg = SDValue(VLd, 0);
2238   static_assert(ARM::dsub_7 == ARM::dsub_0 + 7 &&
2239                     ARM::qsub_3 == ARM::qsub_0 + 3,
2240                 "Unexpected subreg numbering");
2241   unsigned Sub0 = (is64BitVector ? ARM::dsub_0 : ARM::qsub_0);
2242   for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
2243     ReplaceUses(SDValue(N, Vec),
2244                 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg));
2245   ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, 1));
2246   if (isUpdating)
2247     ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLd, 2));
2248   CurDAG->RemoveDeadNode(N);
2249 }
2250 
2251 void ARMDAGToDAGISel::SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs,
2252                                 const uint16_t *DOpcodes,
2253                                 const uint16_t *QOpcodes0,
2254                                 const uint16_t *QOpcodes1) {
2255   assert(Subtarget->hasNEON());
2256   assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range");
2257   SDLoc dl(N);
2258 
2259   SDValue MemAddr, Align;
2260   bool IsIntrinsic = !isUpdating;  // By coincidence, all supported updating
2261                                    // nodes are not intrinsics.
2262   unsigned AddrOpIdx = IsIntrinsic ? 2 : 1;
2263   unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1)
2264   if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
2265     return;
2266 
2267   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2268 
2269   SDValue Chain = N->getOperand(0);
2270   EVT VT = N->getOperand(Vec0Idx).getValueType();
2271   bool is64BitVector = VT.is64BitVector();
2272   Align = GetVLDSTAlign(Align, dl, NumVecs, is64BitVector);
2273 
2274   unsigned OpcodeIndex;
2275   switch (VT.getSimpleVT().SimpleTy) {
2276   default: llvm_unreachable("unhandled vst type");
2277     // Double-register operations:
2278   case MVT::v8i8:  OpcodeIndex = 0; break;
2279   case MVT::v4f16:
2280   case MVT::v4bf16:
2281   case MVT::v4i16: OpcodeIndex = 1; break;
2282   case MVT::v2f32:
2283   case MVT::v2i32: OpcodeIndex = 2; break;
2284   case MVT::v1i64: OpcodeIndex = 3; break;
2285     // Quad-register operations:
2286   case MVT::v16i8: OpcodeIndex = 0; break;
2287   case MVT::v8f16:
2288   case MVT::v8bf16:
2289   case MVT::v8i16: OpcodeIndex = 1; break;
2290   case MVT::v4f32:
2291   case MVT::v4i32: OpcodeIndex = 2; break;
2292   case MVT::v2f64:
2293   case MVT::v2i64: OpcodeIndex = 3; break;
2294   }
2295 
2296   std::vector<EVT> ResTys;
2297   if (isUpdating)
2298     ResTys.push_back(MVT::i32);
2299   ResTys.push_back(MVT::Other);
2300 
2301   SDValue Pred = getAL(CurDAG, dl);
2302   SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2303   SmallVector<SDValue, 7> Ops;
2304 
2305   // Double registers and VST1/VST2 quad registers are directly supported.
2306   if (is64BitVector || NumVecs <= 2) {
2307     SDValue SrcReg;
2308     if (NumVecs == 1) {
2309       SrcReg = N->getOperand(Vec0Idx);
2310     } else if (is64BitVector) {
2311       // Form a REG_SEQUENCE to force register allocation.
2312       SDValue V0 = N->getOperand(Vec0Idx + 0);
2313       SDValue V1 = N->getOperand(Vec0Idx + 1);
2314       if (NumVecs == 2)
2315         SrcReg = SDValue(createDRegPairNode(MVT::v2i64, V0, V1), 0);
2316       else {
2317         SDValue V2 = N->getOperand(Vec0Idx + 2);
2318         // If it's a vst3, form a quad D-register and leave the last part as
2319         // an undef.
2320         SDValue V3 = (NumVecs == 3)
2321           ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0)
2322           : N->getOperand(Vec0Idx + 3);
2323         SrcReg = SDValue(createQuadDRegsNode(MVT::v4i64, V0, V1, V2, V3), 0);
2324       }
2325     } else {
2326       // Form a QQ register.
2327       SDValue Q0 = N->getOperand(Vec0Idx);
2328       SDValue Q1 = N->getOperand(Vec0Idx + 1);
2329       SrcReg = SDValue(createQRegPairNode(MVT::v4i64, Q0, Q1), 0);
2330     }
2331 
2332     unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
2333                     QOpcodes0[OpcodeIndex]);
2334     Ops.push_back(MemAddr);
2335     Ops.push_back(Align);
2336     if (isUpdating) {
2337       SDValue Inc = N->getOperand(AddrOpIdx + 1);
2338       bool IsImmUpdate = isPerfectIncrement(Inc, VT, NumVecs);
2339       if (!IsImmUpdate) {
2340         // We use a VST1 for v1i64 even if the pseudo says VST2/3/4, so
2341         // check for the opcode rather than the number of vector elements.
2342         if (isVSTfixed(Opc))
2343           Opc = getVLDSTRegisterUpdateOpcode(Opc);
2344         Ops.push_back(Inc);
2345       }
2346       // VST1/VST2 fixed increment does not need Reg0 so only include it in
2347       // the operands if not such an opcode.
2348       else if (!isVSTfixed(Opc))
2349         Ops.push_back(Reg0);
2350     }
2351     Ops.push_back(SrcReg);
2352     Ops.push_back(Pred);
2353     Ops.push_back(Reg0);
2354     Ops.push_back(Chain);
2355     SDNode *VSt = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
2356 
2357     // Transfer memoperands.
2358     CurDAG->setNodeMemRefs(cast<MachineSDNode>(VSt), {MemOp});
2359 
2360     ReplaceNode(N, VSt);
2361     return;
2362   }
2363 
2364   // Otherwise, quad registers are stored with two separate instructions,
2365   // where one stores the even registers and the other stores the odd registers.
2366 
2367   // Form the QQQQ REG_SEQUENCE.
2368   SDValue V0 = N->getOperand(Vec0Idx + 0);
2369   SDValue V1 = N->getOperand(Vec0Idx + 1);
2370   SDValue V2 = N->getOperand(Vec0Idx + 2);
2371   SDValue V3 = (NumVecs == 3)
2372     ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
2373     : N->getOperand(Vec0Idx + 3);
2374   SDValue RegSeq = SDValue(createQuadQRegsNode(MVT::v8i64, V0, V1, V2, V3), 0);
2375 
2376   // Store the even D registers.  This is always an updating store, so that it
2377   // provides the address to the second store for the odd subregs.
2378   const SDValue OpsA[] = { MemAddr, Align, Reg0, RegSeq, Pred, Reg0, Chain };
2379   SDNode *VStA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl,
2380                                         MemAddr.getValueType(),
2381                                         MVT::Other, OpsA);
2382   CurDAG->setNodeMemRefs(cast<MachineSDNode>(VStA), {MemOp});
2383   Chain = SDValue(VStA, 1);
2384 
2385   // Store the odd D registers.
2386   Ops.push_back(SDValue(VStA, 0));
2387   Ops.push_back(Align);
2388   if (isUpdating) {
2389     SDValue Inc = N->getOperand(AddrOpIdx + 1);
2390     assert(isa<ConstantSDNode>(Inc.getNode()) &&
2391            "only constant post-increment update allowed for VST3/4");
2392     (void)Inc;
2393     Ops.push_back(Reg0);
2394   }
2395   Ops.push_back(RegSeq);
2396   Ops.push_back(Pred);
2397   Ops.push_back(Reg0);
2398   Ops.push_back(Chain);
2399   SDNode *VStB = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys,
2400                                         Ops);
2401   CurDAG->setNodeMemRefs(cast<MachineSDNode>(VStB), {MemOp});
2402   ReplaceNode(N, VStB);
2403 }
2404 
2405 void ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad, bool isUpdating,
2406                                       unsigned NumVecs,
2407                                       const uint16_t *DOpcodes,
2408                                       const uint16_t *QOpcodes) {
2409   assert(Subtarget->hasNEON());
2410   assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
2411   SDLoc dl(N);
2412 
2413   SDValue MemAddr, Align;
2414   bool IsIntrinsic = !isUpdating;  // By coincidence, all supported updating
2415                                    // nodes are not intrinsics.
2416   unsigned AddrOpIdx = IsIntrinsic ? 2 : 1;
2417   unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1)
2418   if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
2419     return;
2420 
2421   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
2422 
2423   SDValue Chain = N->getOperand(0);
2424   unsigned Lane = N->getConstantOperandVal(Vec0Idx + NumVecs);
2425   EVT VT = N->getOperand(Vec0Idx).getValueType();
2426   bool is64BitVector = VT.is64BitVector();
2427 
2428   unsigned Alignment = 0;
2429   if (NumVecs != 3) {
2430     Alignment = Align->getAsZExtVal();
2431     unsigned NumBytes = NumVecs * VT.getScalarSizeInBits() / 8;
2432     if (Alignment > NumBytes)
2433       Alignment = NumBytes;
2434     if (Alignment < 8 && Alignment < NumBytes)
2435       Alignment = 0;
2436     // Alignment must be a power of two; make sure of that.
2437     Alignment = (Alignment & -Alignment);
2438     if (Alignment == 1)
2439       Alignment = 0;
2440   }
2441   Align = CurDAG->getTargetConstant(Alignment, dl, MVT::i32);
2442 
2443   unsigned OpcodeIndex;
2444   switch (VT.getSimpleVT().SimpleTy) {
2445   default: llvm_unreachable("unhandled vld/vst lane type");
2446     // Double-register operations:
2447   case MVT::v8i8:  OpcodeIndex = 0; break;
2448   case MVT::v4f16:
2449   case MVT::v4bf16:
2450   case MVT::v4i16: OpcodeIndex = 1; break;
2451   case MVT::v2f32:
2452   case MVT::v2i32: OpcodeIndex = 2; break;
2453     // Quad-register operations:
2454   case MVT::v8f16:
2455   case MVT::v8bf16:
2456   case MVT::v8i16: OpcodeIndex = 0; break;
2457   case MVT::v4f32:
2458   case MVT::v4i32: OpcodeIndex = 1; break;
2459   }
2460 
2461   std::vector<EVT> ResTys;
2462   if (IsLoad) {
2463     unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
2464     if (!is64BitVector)
2465       ResTyElts *= 2;
2466     ResTys.push_back(EVT::getVectorVT(*CurDAG->getContext(),
2467                                       MVT::i64, ResTyElts));
2468   }
2469   if (isUpdating)
2470     ResTys.push_back(MVT::i32);
2471   ResTys.push_back(MVT::Other);
2472 
2473   SDValue Pred = getAL(CurDAG, dl);
2474   SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
2475 
2476   SmallVector<SDValue, 8> Ops;
2477   Ops.push_back(MemAddr);
2478   Ops.push_back(Align);
2479   if (isUpdating) {
2480     SDValue Inc = N->getOperand(AddrOpIdx + 1);
2481     bool IsImmUpdate =
2482         isPerfectIncrement(Inc, VT.getVectorElementType(), NumVecs);
2483     Ops.push_back(IsImmUpdate ? Reg0 : Inc);
2484   }
2485 
2486   SDValue SuperReg;
2487   SDValue V0 = N->getOperand(Vec0Idx + 0);
2488   SDValue V1 = N->getOperand(Vec0Idx + 1);
2489   if (NumVecs == 2) {
2490     if (is64BitVector)
2491       SuperReg = SDValue(createDRegPairNode(MVT::v2i64, V0, V1), 0);
2492     else
2493       SuperReg = SDValue(createQRegPairNode(MVT::v4i64, V0, V1), 0);
2494   } else {
2495     SDValue V2 = N->getOperand(Vec0Idx + 2);
2496     SDValue V3 = (NumVecs == 3)
2497       ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0)
2498       : N->getOperand(Vec0Idx + 3);
2499     if (is64BitVector)
2500       SuperReg = SDValue(createQuadDRegsNode(MVT::v4i64, V0, V1, V2, V3), 0);
2501     else
2502       SuperReg = SDValue(createQuadQRegsNode(MVT::v8i64, V0, V1, V2, V3), 0);
2503   }
2504   Ops.push_back(SuperReg);
2505   Ops.push_back(getI32Imm(Lane, dl));
2506   Ops.push_back(Pred);
2507   Ops.push_back(Reg0);
2508   Ops.push_back(Chain);
2509 
2510   unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] :
2511                                   QOpcodes[OpcodeIndex]);
2512   SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
2513   CurDAG->setNodeMemRefs(cast<MachineSDNode>(VLdLn), {MemOp});
2514   if (!IsLoad) {
2515     ReplaceNode(N, VLdLn);
2516     return;
2517   }
2518 
2519   // Extract the subregisters.
2520   SuperReg = SDValue(VLdLn, 0);
2521   static_assert(ARM::dsub_7 == ARM::dsub_0 + 7 &&
2522                     ARM::qsub_3 == ARM::qsub_0 + 3,
2523                 "Unexpected subreg numbering");
2524   unsigned Sub0 = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
2525   for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
2526     ReplaceUses(SDValue(N, Vec),
2527                 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg));
2528   ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, 1));
2529   if (isUpdating)
2530     ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdLn, 2));
2531   CurDAG->RemoveDeadNode(N);
2532 }
2533 
2534 template <typename SDValueVector>
2535 void ARMDAGToDAGISel::AddMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc,
2536                                            SDValue PredicateMask) {
2537   Ops.push_back(CurDAG->getTargetConstant(ARMVCC::Then, Loc, MVT::i32));
2538   Ops.push_back(PredicateMask);
2539   Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // tp_reg
2540 }
2541 
2542 template <typename SDValueVector>
2543 void ARMDAGToDAGISel::AddMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc,
2544                                            SDValue PredicateMask,
2545                                            SDValue Inactive) {
2546   Ops.push_back(CurDAG->getTargetConstant(ARMVCC::Then, Loc, MVT::i32));
2547   Ops.push_back(PredicateMask);
2548   Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // tp_reg
2549   Ops.push_back(Inactive);
2550 }
2551 
2552 template <typename SDValueVector>
2553 void ARMDAGToDAGISel::AddEmptyMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc) {
2554   Ops.push_back(CurDAG->getTargetConstant(ARMVCC::None, Loc, MVT::i32));
2555   Ops.push_back(CurDAG->getRegister(0, MVT::i32));
2556   Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // tp_reg
2557 }
2558 
2559 template <typename SDValueVector>
2560 void ARMDAGToDAGISel::AddEmptyMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc,
2561                                                 EVT InactiveTy) {
2562   Ops.push_back(CurDAG->getTargetConstant(ARMVCC::None, Loc, MVT::i32));
2563   Ops.push_back(CurDAG->getRegister(0, MVT::i32));
2564   Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // tp_reg
2565   Ops.push_back(SDValue(
2566       CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, Loc, InactiveTy), 0));
2567 }
2568 
2569 void ARMDAGToDAGISel::SelectMVE_WB(SDNode *N, const uint16_t *Opcodes,
2570                                    bool Predicated) {
2571   SDLoc Loc(N);
2572   SmallVector<SDValue, 8> Ops;
2573 
2574   uint16_t Opcode;
2575   switch (N->getValueType(1).getVectorElementType().getSizeInBits()) {
2576   case 32:
2577     Opcode = Opcodes[0];
2578     break;
2579   case 64:
2580     Opcode = Opcodes[1];
2581     break;
2582   default:
2583     llvm_unreachable("bad vector element size in SelectMVE_WB");
2584   }
2585 
2586   Ops.push_back(N->getOperand(2)); // vector of base addresses
2587 
2588   int32_t ImmValue = N->getConstantOperandVal(3);
2589   Ops.push_back(getI32Imm(ImmValue, Loc)); // immediate offset
2590 
2591   if (Predicated)
2592     AddMVEPredicateToOps(Ops, Loc, N->getOperand(4));
2593   else
2594     AddEmptyMVEPredicateToOps(Ops, Loc);
2595 
2596   Ops.push_back(N->getOperand(0)); // chain
2597 
2598   SmallVector<EVT, 8> VTs;
2599   VTs.push_back(N->getValueType(1));
2600   VTs.push_back(N->getValueType(0));
2601   VTs.push_back(N->getValueType(2));
2602 
2603   SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), VTs, Ops);
2604   ReplaceUses(SDValue(N, 0), SDValue(New, 1));
2605   ReplaceUses(SDValue(N, 1), SDValue(New, 0));
2606   ReplaceUses(SDValue(N, 2), SDValue(New, 2));
2607   transferMemOperands(N, New);
2608   CurDAG->RemoveDeadNode(N);
2609 }
2610 
2611 void ARMDAGToDAGISel::SelectMVE_LongShift(SDNode *N, uint16_t Opcode,
2612                                           bool Immediate,
2613                                           bool HasSaturationOperand) {
2614   SDLoc Loc(N);
2615   SmallVector<SDValue, 8> Ops;
2616 
2617   // Two 32-bit halves of the value to be shifted
2618   Ops.push_back(N->getOperand(1));
2619   Ops.push_back(N->getOperand(2));
2620 
2621   // The shift count
2622   if (Immediate) {
2623     int32_t ImmValue = N->getConstantOperandVal(3);
2624     Ops.push_back(getI32Imm(ImmValue, Loc)); // immediate shift count
2625   } else {
2626     Ops.push_back(N->getOperand(3));
2627   }
2628 
2629   // The immediate saturation operand, if any
2630   if (HasSaturationOperand) {
2631     int32_t SatOp = N->getConstantOperandVal(4);
2632     int SatBit = (SatOp == 64 ? 0 : 1);
2633     Ops.push_back(getI32Imm(SatBit, Loc));
2634   }
2635 
2636   // MVE scalar shifts are IT-predicable, so include the standard
2637   // predicate arguments.
2638   Ops.push_back(getAL(CurDAG, Loc));
2639   Ops.push_back(CurDAG->getRegister(0, MVT::i32));
2640 
2641   CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), ArrayRef(Ops));
2642 }
2643 
2644 void ARMDAGToDAGISel::SelectMVE_VADCSBC(SDNode *N, uint16_t OpcodeWithCarry,
2645                                         uint16_t OpcodeWithNoCarry,
2646                                         bool Add, bool Predicated) {
2647   SDLoc Loc(N);
2648   SmallVector<SDValue, 8> Ops;
2649   uint16_t Opcode;
2650 
2651   unsigned FirstInputOp = Predicated ? 2 : 1;
2652 
2653   // Two input vectors and the input carry flag
2654   Ops.push_back(N->getOperand(FirstInputOp));
2655   Ops.push_back(N->getOperand(FirstInputOp + 1));
2656   SDValue CarryIn = N->getOperand(FirstInputOp + 2);
2657   ConstantSDNode *CarryInConstant = dyn_cast<ConstantSDNode>(CarryIn);
2658   uint32_t CarryMask = 1 << 29;
2659   uint32_t CarryExpected = Add ? 0 : CarryMask;
2660   if (CarryInConstant &&
2661       (CarryInConstant->getZExtValue() & CarryMask) == CarryExpected) {
2662     Opcode = OpcodeWithNoCarry;
2663   } else {
2664     Ops.push_back(CarryIn);
2665     Opcode = OpcodeWithCarry;
2666   }
2667 
2668   if (Predicated)
2669     AddMVEPredicateToOps(Ops, Loc,
2670                          N->getOperand(FirstInputOp + 3),  // predicate
2671                          N->getOperand(FirstInputOp - 1)); // inactive
2672   else
2673     AddEmptyMVEPredicateToOps(Ops, Loc, N->getValueType(0));
2674 
2675   CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), ArrayRef(Ops));
2676 }
2677 
2678 void ARMDAGToDAGISel::SelectMVE_VSHLC(SDNode *N, bool Predicated) {
2679   SDLoc Loc(N);
2680   SmallVector<SDValue, 8> Ops;
2681 
2682   // One vector input, followed by a 32-bit word of bits to shift in
2683   // and then an immediate shift count
2684   Ops.push_back(N->getOperand(1));
2685   Ops.push_back(N->getOperand(2));
2686   int32_t ImmValue = N->getConstantOperandVal(3);
2687   Ops.push_back(getI32Imm(ImmValue, Loc)); // immediate shift count
2688 
2689   if (Predicated)
2690     AddMVEPredicateToOps(Ops, Loc, N->getOperand(4));
2691   else
2692     AddEmptyMVEPredicateToOps(Ops, Loc);
2693 
2694   CurDAG->SelectNodeTo(N, ARM::MVE_VSHLC, N->getVTList(), ArrayRef(Ops));
2695 }
2696 
2697 static bool SDValueToConstBool(SDValue SDVal) {
2698   assert(isa<ConstantSDNode>(SDVal) && "expected a compile-time constant");
2699   ConstantSDNode *SDValConstant = dyn_cast<ConstantSDNode>(SDVal);
2700   uint64_t Value = SDValConstant->getZExtValue();
2701   assert((Value == 0 || Value == 1) && "expected value 0 or 1");
2702   return Value;
2703 }
2704 
2705 void ARMDAGToDAGISel::SelectBaseMVE_VMLLDAV(SDNode *N, bool Predicated,
2706                                             const uint16_t *OpcodesS,
2707                                             const uint16_t *OpcodesU,
2708                                             size_t Stride, size_t TySize) {
2709   assert(TySize < Stride && "Invalid TySize");
2710   bool IsUnsigned = SDValueToConstBool(N->getOperand(1));
2711   bool IsSub = SDValueToConstBool(N->getOperand(2));
2712   bool IsExchange = SDValueToConstBool(N->getOperand(3));
2713   if (IsUnsigned) {
2714     assert(!IsSub &&
2715            "Unsigned versions of vmlsldav[a]/vrmlsldavh[a] do not exist");
2716     assert(!IsExchange &&
2717            "Unsigned versions of vmlaldav[a]x/vrmlaldavh[a]x do not exist");
2718   }
2719 
2720   auto OpIsZero = [N](size_t OpNo) {
2721     return isNullConstant(N->getOperand(OpNo));
2722   };
2723 
2724   // If the input accumulator value is not zero, select an instruction with
2725   // accumulator, otherwise select an instruction without accumulator
2726   bool IsAccum = !(OpIsZero(4) && OpIsZero(5));
2727 
2728   const uint16_t *Opcodes = IsUnsigned ? OpcodesU : OpcodesS;
2729   if (IsSub)
2730     Opcodes += 4 * Stride;
2731   if (IsExchange)
2732     Opcodes += 2 * Stride;
2733   if (IsAccum)
2734     Opcodes += Stride;
2735   uint16_t Opcode = Opcodes[TySize];
2736 
2737   SDLoc Loc(N);
2738   SmallVector<SDValue, 8> Ops;
2739   // Push the accumulator operands, if they are used
2740   if (IsAccum) {
2741     Ops.push_back(N->getOperand(4));
2742     Ops.push_back(N->getOperand(5));
2743   }
2744   // Push the two vector operands
2745   Ops.push_back(N->getOperand(6));
2746   Ops.push_back(N->getOperand(7));
2747 
2748   if (Predicated)
2749     AddMVEPredicateToOps(Ops, Loc, N->getOperand(8));
2750   else
2751     AddEmptyMVEPredicateToOps(Ops, Loc);
2752 
2753   CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), ArrayRef(Ops));
2754 }
2755 
2756 void ARMDAGToDAGISel::SelectMVE_VMLLDAV(SDNode *N, bool Predicated,
2757                                         const uint16_t *OpcodesS,
2758                                         const uint16_t *OpcodesU) {
2759   EVT VecTy = N->getOperand(6).getValueType();
2760   size_t SizeIndex;
2761   switch (VecTy.getVectorElementType().getSizeInBits()) {
2762   case 16:
2763     SizeIndex = 0;
2764     break;
2765   case 32:
2766     SizeIndex = 1;
2767     break;
2768   default:
2769     llvm_unreachable("bad vector element size");
2770   }
2771 
2772   SelectBaseMVE_VMLLDAV(N, Predicated, OpcodesS, OpcodesU, 2, SizeIndex);
2773 }
2774 
2775 void ARMDAGToDAGISel::SelectMVE_VRMLLDAVH(SDNode *N, bool Predicated,
2776                                           const uint16_t *OpcodesS,
2777                                           const uint16_t *OpcodesU) {
2778   assert(
2779       N->getOperand(6).getValueType().getVectorElementType().getSizeInBits() ==
2780           32 &&
2781       "bad vector element size");
2782   SelectBaseMVE_VMLLDAV(N, Predicated, OpcodesS, OpcodesU, 1, 0);
2783 }
2784 
2785 void ARMDAGToDAGISel::SelectMVE_VLD(SDNode *N, unsigned NumVecs,
2786                                     const uint16_t *const *Opcodes,
2787                                     bool HasWriteback) {
2788   EVT VT = N->getValueType(0);
2789   SDLoc Loc(N);
2790 
2791   const uint16_t *OurOpcodes;
2792   switch (VT.getVectorElementType().getSizeInBits()) {
2793   case 8:
2794     OurOpcodes = Opcodes[0];
2795     break;
2796   case 16:
2797     OurOpcodes = Opcodes[1];
2798     break;
2799   case 32:
2800     OurOpcodes = Opcodes[2];
2801     break;
2802   default:
2803     llvm_unreachable("bad vector element size in SelectMVE_VLD");
2804   }
2805 
2806   EVT DataTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, NumVecs * 2);
2807   SmallVector<EVT, 4> ResultTys = {DataTy, MVT::Other};
2808   unsigned PtrOperand = HasWriteback ? 1 : 2;
2809 
2810   auto Data = SDValue(
2811       CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, Loc, DataTy), 0);
2812   SDValue Chain = N->getOperand(0);
2813   // Add a MVE_VLDn instruction for each Vec, except the last
2814   for (unsigned Stage = 0; Stage < NumVecs - 1; ++Stage) {
2815     SDValue Ops[] = {Data, N->getOperand(PtrOperand), Chain};
2816     auto LoadInst =
2817         CurDAG->getMachineNode(OurOpcodes[Stage], Loc, ResultTys, Ops);
2818     Data = SDValue(LoadInst, 0);
2819     Chain = SDValue(LoadInst, 1);
2820     transferMemOperands(N, LoadInst);
2821   }
2822   // The last may need a writeback on it
2823   if (HasWriteback)
2824     ResultTys = {DataTy, MVT::i32, MVT::Other};
2825   SDValue Ops[] = {Data, N->getOperand(PtrOperand), Chain};
2826   auto LoadInst =
2827       CurDAG->getMachineNode(OurOpcodes[NumVecs - 1], Loc, ResultTys, Ops);
2828   transferMemOperands(N, LoadInst);
2829 
2830   unsigned i;
2831   for (i = 0; i < NumVecs; i++)
2832     ReplaceUses(SDValue(N, i),
2833                 CurDAG->getTargetExtractSubreg(ARM::qsub_0 + i, Loc, VT,
2834                                                SDValue(LoadInst, 0)));
2835   if (HasWriteback)
2836     ReplaceUses(SDValue(N, i++), SDValue(LoadInst, 1));
2837   ReplaceUses(SDValue(N, i), SDValue(LoadInst, HasWriteback ? 2 : 1));
2838   CurDAG->RemoveDeadNode(N);
2839 }
2840 
2841 void ARMDAGToDAGISel::SelectMVE_VxDUP(SDNode *N, const uint16_t *Opcodes,
2842                                       bool Wrapping, bool Predicated) {
2843   EVT VT = N->getValueType(0);
2844   SDLoc Loc(N);
2845 
2846   uint16_t Opcode;
2847   switch (VT.getScalarSizeInBits()) {
2848   case 8:
2849     Opcode = Opcodes[0];
2850     break;
2851   case 16:
2852     Opcode = Opcodes[1];
2853     break;
2854   case 32:
2855     Opcode = Opcodes[2];
2856     break;
2857   default:
2858     llvm_unreachable("bad vector element size in SelectMVE_VxDUP");
2859   }
2860 
2861   SmallVector<SDValue, 8> Ops;
2862   unsigned OpIdx = 1;
2863 
2864   SDValue Inactive;
2865   if (Predicated)
2866     Inactive = N->getOperand(OpIdx++);
2867 
2868   Ops.push_back(N->getOperand(OpIdx++));     // base
2869   if (Wrapping)
2870     Ops.push_back(N->getOperand(OpIdx++));   // limit
2871 
2872   SDValue ImmOp = N->getOperand(OpIdx++);    // step
2873   int ImmValue = ImmOp->getAsZExtVal();
2874   Ops.push_back(getI32Imm(ImmValue, Loc));
2875 
2876   if (Predicated)
2877     AddMVEPredicateToOps(Ops, Loc, N->getOperand(OpIdx), Inactive);
2878   else
2879     AddEmptyMVEPredicateToOps(Ops, Loc, N->getValueType(0));
2880 
2881   CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), ArrayRef(Ops));
2882 }
2883 
2884 void ARMDAGToDAGISel::SelectCDE_CXxD(SDNode *N, uint16_t Opcode,
2885                                      size_t NumExtraOps, bool HasAccum) {
2886   bool IsBigEndian = CurDAG->getDataLayout().isBigEndian();
2887   SDLoc Loc(N);
2888   SmallVector<SDValue, 8> Ops;
2889 
2890   unsigned OpIdx = 1;
2891 
2892   // Convert and append the immediate operand designating the coprocessor.
2893   SDValue ImmCorpoc = N->getOperand(OpIdx++);
2894   uint32_t ImmCoprocVal = ImmCorpoc->getAsZExtVal();
2895   Ops.push_back(getI32Imm(ImmCoprocVal, Loc));
2896 
2897   // For accumulating variants copy the low and high order parts of the
2898   // accumulator into a register pair and add it to the operand vector.
2899   if (HasAccum) {
2900     SDValue AccLo = N->getOperand(OpIdx++);
2901     SDValue AccHi = N->getOperand(OpIdx++);
2902     if (IsBigEndian)
2903       std::swap(AccLo, AccHi);
2904     Ops.push_back(SDValue(createGPRPairNode(MVT::Untyped, AccLo, AccHi), 0));
2905   }
2906 
2907   // Copy extra operands as-is.
2908   for (size_t I = 0; I < NumExtraOps; I++)
2909     Ops.push_back(N->getOperand(OpIdx++));
2910 
2911   // Convert and append the immediate operand
2912   SDValue Imm = N->getOperand(OpIdx);
2913   uint32_t ImmVal = Imm->getAsZExtVal();
2914   Ops.push_back(getI32Imm(ImmVal, Loc));
2915 
2916   // Accumulating variants are IT-predicable, add predicate operands.
2917   if (HasAccum) {
2918     SDValue Pred = getAL(CurDAG, Loc);
2919     SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
2920     Ops.push_back(Pred);
2921     Ops.push_back(PredReg);
2922   }
2923 
2924   // Create the CDE intruction
2925   SDNode *InstrNode = CurDAG->getMachineNode(Opcode, Loc, MVT::Untyped, Ops);
2926   SDValue ResultPair = SDValue(InstrNode, 0);
2927 
2928   // The original intrinsic had two outputs, and the output of the dual-register
2929   // CDE instruction is a register pair. We need to extract the two subregisters
2930   // and replace all uses of the original outputs with the extracted
2931   // subregisters.
2932   uint16_t SubRegs[2] = {ARM::gsub_0, ARM::gsub_1};
2933   if (IsBigEndian)
2934     std::swap(SubRegs[0], SubRegs[1]);
2935 
2936   for (size_t ResIdx = 0; ResIdx < 2; ResIdx++) {
2937     if (SDValue(N, ResIdx).use_empty())
2938       continue;
2939     SDValue SubReg = CurDAG->getTargetExtractSubreg(SubRegs[ResIdx], Loc,
2940                                                     MVT::i32, ResultPair);
2941     ReplaceUses(SDValue(N, ResIdx), SubReg);
2942   }
2943 
2944   CurDAG->RemoveDeadNode(N);
2945 }
2946 
2947 void ARMDAGToDAGISel::SelectVLDDup(SDNode *N, bool IsIntrinsic,
2948                                    bool isUpdating, unsigned NumVecs,
2949                                    const uint16_t *DOpcodes,
2950                                    const uint16_t *QOpcodes0,
2951                                    const uint16_t *QOpcodes1) {
2952   assert(Subtarget->hasNEON());
2953   assert(NumVecs >= 1 && NumVecs <= 4 && "VLDDup NumVecs out-of-range");
2954   SDLoc dl(N);
2955 
2956   SDValue MemAddr, Align;
2957   unsigned AddrOpIdx = IsIntrinsic ? 2 : 1;
2958   if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align))
2959     return;
2960 
2961   SDValue Chain = N->getOperand(0);
2962   EVT VT = N->getValueType(0);
2963   bool is64BitVector = VT.is64BitVector();
2964 
2965   unsigned Alignment = 0;
2966   if (NumVecs != 3) {
2967     Alignment = Align->getAsZExtVal();
2968     unsigned NumBytes = NumVecs * VT.getScalarSizeInBits() / 8;
2969     if (Alignment > NumBytes)
2970       Alignment = NumBytes;
2971     if (Alignment < 8 && Alignment < NumBytes)
2972       Alignment = 0;
2973     // Alignment must be a power of two; make sure of that.
2974     Alignment = (Alignment & -Alignment);
2975     if (Alignment == 1)
2976       Alignment = 0;
2977   }
2978   Align = CurDAG->getTargetConstant(Alignment, dl, MVT::i32);
2979 
2980   unsigned OpcodeIndex;
2981   switch (VT.getSimpleVT().SimpleTy) {
2982   default: llvm_unreachable("unhandled vld-dup type");
2983   case MVT::v8i8:
2984   case MVT::v16i8: OpcodeIndex = 0; break;
2985   case MVT::v4i16:
2986   case MVT::v8i16:
2987   case MVT::v4f16:
2988   case MVT::v8f16:
2989   case MVT::v4bf16:
2990   case MVT::v8bf16:
2991                   OpcodeIndex = 1; break;
2992   case MVT::v2f32:
2993   case MVT::v2i32:
2994   case MVT::v4f32:
2995   case MVT::v4i32: OpcodeIndex = 2; break;
2996   case MVT::v1f64:
2997   case MVT::v1i64: OpcodeIndex = 3; break;
2998   }
2999 
3000   unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs;
3001   if (!is64BitVector)
3002     ResTyElts *= 2;
3003   EVT ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, ResTyElts);
3004 
3005   std::vector<EVT> ResTys;
3006   ResTys.push_back(ResTy);
3007   if (isUpdating)
3008     ResTys.push_back(MVT::i32);
3009   ResTys.push_back(MVT::Other);
3010 
3011   SDValue Pred = getAL(CurDAG, dl);
3012   SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3013 
3014   SmallVector<SDValue, 6> Ops;
3015   Ops.push_back(MemAddr);
3016   Ops.push_back(Align);
3017   unsigned Opc = is64BitVector    ? DOpcodes[OpcodeIndex]
3018                  : (NumVecs == 1) ? QOpcodes0[OpcodeIndex]
3019                                   : QOpcodes1[OpcodeIndex];
3020   if (isUpdating) {
3021     SDValue Inc = N->getOperand(2);
3022     bool IsImmUpdate =
3023         isPerfectIncrement(Inc, VT.getVectorElementType(), NumVecs);
3024     if (IsImmUpdate) {
3025       if (!isVLDfixed(Opc))
3026         Ops.push_back(Reg0);
3027     } else {
3028       if (isVLDfixed(Opc))
3029         Opc = getVLDSTRegisterUpdateOpcode(Opc);
3030       Ops.push_back(Inc);
3031     }
3032   }
3033   if (is64BitVector || NumVecs == 1) {
3034     // Double registers and VLD1 quad registers are directly supported.
3035   } else if (NumVecs == 2) {
3036     const SDValue OpsA[] = {MemAddr, Align, Pred, Reg0, Chain};
3037     SDNode *VLdA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl, ResTy,
3038                                           MVT::Other, OpsA);
3039     Chain = SDValue(VLdA, 1);
3040   } else {
3041     SDValue ImplDef = SDValue(
3042         CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, ResTy), 0);
3043     const SDValue OpsA[] = {MemAddr, Align, ImplDef, Pred, Reg0, Chain};
3044     SDNode *VLdA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl, ResTy,
3045                                           MVT::Other, OpsA);
3046     Ops.push_back(SDValue(VLdA, 0));
3047     Chain = SDValue(VLdA, 1);
3048   }
3049 
3050   Ops.push_back(Pred);
3051   Ops.push_back(Reg0);
3052   Ops.push_back(Chain);
3053 
3054   SDNode *VLdDup = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
3055 
3056   // Transfer memoperands.
3057   MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
3058   CurDAG->setNodeMemRefs(cast<MachineSDNode>(VLdDup), {MemOp});
3059 
3060   // Extract the subregisters.
3061   if (NumVecs == 1) {
3062     ReplaceUses(SDValue(N, 0), SDValue(VLdDup, 0));
3063   } else {
3064     SDValue SuperReg = SDValue(VLdDup, 0);
3065     static_assert(ARM::dsub_7 == ARM::dsub_0 + 7, "Unexpected subreg numbering");
3066     unsigned SubIdx = is64BitVector ? ARM::dsub_0 : ARM::qsub_0;
3067     for (unsigned Vec = 0; Vec != NumVecs; ++Vec) {
3068       ReplaceUses(SDValue(N, Vec),
3069                   CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, SuperReg));
3070     }
3071   }
3072   ReplaceUses(SDValue(N, NumVecs), SDValue(VLdDup, 1));
3073   if (isUpdating)
3074     ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdDup, 2));
3075   CurDAG->RemoveDeadNode(N);
3076 }
3077 
3078 bool ARMDAGToDAGISel::tryInsertVectorElt(SDNode *N) {
3079   if (!Subtarget->hasMVEIntegerOps())
3080     return false;
3081 
3082   SDLoc dl(N);
3083 
3084   // We are trying to use VMOV/VMOVX/VINS to more efficiently lower insert and
3085   // extracts of v8f16 and v8i16 vectors. Check that we have two adjacent
3086   // inserts of the correct type:
3087   SDValue Ins1 = SDValue(N, 0);
3088   SDValue Ins2 = N->getOperand(0);
3089   EVT VT = Ins1.getValueType();
3090   if (Ins2.getOpcode() != ISD::INSERT_VECTOR_ELT || !Ins2.hasOneUse() ||
3091       !isa<ConstantSDNode>(Ins1.getOperand(2)) ||
3092       !isa<ConstantSDNode>(Ins2.getOperand(2)) ||
3093       (VT != MVT::v8f16 && VT != MVT::v8i16) || (Ins2.getValueType() != VT))
3094     return false;
3095 
3096   unsigned Lane1 = Ins1.getConstantOperandVal(2);
3097   unsigned Lane2 = Ins2.getConstantOperandVal(2);
3098   if (Lane2 % 2 != 0 || Lane1 != Lane2 + 1)
3099     return false;
3100 
3101   // If the inserted values will be able to use T/B already, leave it to the
3102   // existing tablegen patterns. For example VCVTT/VCVTB.
3103   SDValue Val1 = Ins1.getOperand(1);
3104   SDValue Val2 = Ins2.getOperand(1);
3105   if (Val1.getOpcode() == ISD::FP_ROUND || Val2.getOpcode() == ISD::FP_ROUND)
3106     return false;
3107 
3108   // Check if the inserted values are both extracts.
3109   if ((Val1.getOpcode() == ISD::EXTRACT_VECTOR_ELT ||
3110        Val1.getOpcode() == ARMISD::VGETLANEu) &&
3111       (Val2.getOpcode() == ISD::EXTRACT_VECTOR_ELT ||
3112        Val2.getOpcode() == ARMISD::VGETLANEu) &&
3113       isa<ConstantSDNode>(Val1.getOperand(1)) &&
3114       isa<ConstantSDNode>(Val2.getOperand(1)) &&
3115       (Val1.getOperand(0).getValueType() == MVT::v8f16 ||
3116        Val1.getOperand(0).getValueType() == MVT::v8i16) &&
3117       (Val2.getOperand(0).getValueType() == MVT::v8f16 ||
3118        Val2.getOperand(0).getValueType() == MVT::v8i16)) {
3119     unsigned ExtractLane1 = Val1.getConstantOperandVal(1);
3120     unsigned ExtractLane2 = Val2.getConstantOperandVal(1);
3121 
3122     // If the two extracted lanes are from the same place and adjacent, this
3123     // simplifies into a f32 lane move.
3124     if (Val1.getOperand(0) == Val2.getOperand(0) && ExtractLane2 % 2 == 0 &&
3125         ExtractLane1 == ExtractLane2 + 1) {
3126       SDValue NewExt = CurDAG->getTargetExtractSubreg(
3127           ARM::ssub_0 + ExtractLane2 / 2, dl, MVT::f32, Val1.getOperand(0));
3128       SDValue NewIns = CurDAG->getTargetInsertSubreg(
3129           ARM::ssub_0 + Lane2 / 2, dl, VT, Ins2.getOperand(0),
3130           NewExt);
3131       ReplaceUses(Ins1, NewIns);
3132       return true;
3133     }
3134 
3135     // Else v8i16 pattern of an extract and an insert, with a optional vmovx for
3136     // extracting odd lanes.
3137     if (VT == MVT::v8i16 && Subtarget->hasFullFP16()) {
3138       SDValue Inp1 = CurDAG->getTargetExtractSubreg(
3139           ARM::ssub_0 + ExtractLane1 / 2, dl, MVT::f32, Val1.getOperand(0));
3140       SDValue Inp2 = CurDAG->getTargetExtractSubreg(
3141           ARM::ssub_0 + ExtractLane2 / 2, dl, MVT::f32, Val2.getOperand(0));
3142       if (ExtractLane1 % 2 != 0)
3143         Inp1 = SDValue(CurDAG->getMachineNode(ARM::VMOVH, dl, MVT::f32, Inp1), 0);
3144       if (ExtractLane2 % 2 != 0)
3145         Inp2 = SDValue(CurDAG->getMachineNode(ARM::VMOVH, dl, MVT::f32, Inp2), 0);
3146       SDNode *VINS = CurDAG->getMachineNode(ARM::VINSH, dl, MVT::f32, Inp2, Inp1);
3147       SDValue NewIns =
3148           CurDAG->getTargetInsertSubreg(ARM::ssub_0 + Lane2 / 2, dl, MVT::v4f32,
3149                                         Ins2.getOperand(0), SDValue(VINS, 0));
3150       ReplaceUses(Ins1, NewIns);
3151       return true;
3152     }
3153   }
3154 
3155   // The inserted values are not extracted - if they are f16 then insert them
3156   // directly using a VINS.
3157   if (VT == MVT::v8f16 && Subtarget->hasFullFP16()) {
3158     SDNode *VINS = CurDAG->getMachineNode(ARM::VINSH, dl, MVT::f32, Val2, Val1);
3159     SDValue NewIns =
3160         CurDAG->getTargetInsertSubreg(ARM::ssub_0 + Lane2 / 2, dl, MVT::v4f32,
3161                                       Ins2.getOperand(0), SDValue(VINS, 0));
3162     ReplaceUses(Ins1, NewIns);
3163     return true;
3164   }
3165 
3166   return false;
3167 }
3168 
3169 bool ARMDAGToDAGISel::transformFixedFloatingPointConversion(SDNode *N,
3170                                                             SDNode *FMul,
3171                                                             bool IsUnsigned,
3172                                                             bool FixedToFloat) {
3173   auto Type = N->getValueType(0);
3174   unsigned ScalarBits = Type.getScalarSizeInBits();
3175   if (ScalarBits > 32)
3176     return false;
3177 
3178   SDNodeFlags FMulFlags = FMul->getFlags();
3179   // The fixed-point vcvt and vcvt+vmul are not always equivalent if inf is
3180   // allowed in 16 bit unsigned floats
3181   if (ScalarBits == 16 && !FMulFlags.hasNoInfs() && IsUnsigned)
3182     return false;
3183 
3184   SDValue ImmNode = FMul->getOperand(1);
3185   SDValue VecVal = FMul->getOperand(0);
3186   if (VecVal->getOpcode() == ISD::UINT_TO_FP ||
3187       VecVal->getOpcode() == ISD::SINT_TO_FP)
3188     VecVal = VecVal->getOperand(0);
3189 
3190   if (VecVal.getValueType().getScalarSizeInBits() != ScalarBits)
3191     return false;
3192 
3193   if (ImmNode.getOpcode() == ISD::BITCAST) {
3194     if (ImmNode.getValueType().getScalarSizeInBits() != ScalarBits)
3195       return false;
3196     ImmNode = ImmNode.getOperand(0);
3197   }
3198 
3199   if (ImmNode.getValueType().getScalarSizeInBits() != ScalarBits)
3200     return false;
3201 
3202   APFloat ImmAPF(0.0f);
3203   switch (ImmNode.getOpcode()) {
3204   case ARMISD::VMOVIMM:
3205   case ARMISD::VDUP: {
3206     if (!isa<ConstantSDNode>(ImmNode.getOperand(0)))
3207       return false;
3208     unsigned Imm = ImmNode.getConstantOperandVal(0);
3209     if (ImmNode.getOpcode() == ARMISD::VMOVIMM)
3210       Imm = ARM_AM::decodeVMOVModImm(Imm, ScalarBits);
3211     ImmAPF =
3212         APFloat(ScalarBits == 32 ? APFloat::IEEEsingle() : APFloat::IEEEhalf(),
3213                 APInt(ScalarBits, Imm));
3214     break;
3215   }
3216   case ARMISD::VMOVFPIMM: {
3217     ImmAPF = APFloat(ARM_AM::getFPImmFloat(ImmNode.getConstantOperandVal(0)));
3218     break;
3219   }
3220   default:
3221     return false;
3222   }
3223 
3224   // Where n is the number of fractional bits, multiplying by 2^n will convert
3225   // from float to fixed and multiplying by 2^-n will convert from fixed to
3226   // float. Taking log2 of the factor (after taking the inverse in the case of
3227   // float to fixed) will give n.
3228   APFloat ToConvert = ImmAPF;
3229   if (FixedToFloat) {
3230     if (!ImmAPF.getExactInverse(&ToConvert))
3231       return false;
3232   }
3233   APSInt Converted(64, false);
3234   bool IsExact;
3235   ToConvert.convertToInteger(Converted, llvm::RoundingMode::NearestTiesToEven,
3236                              &IsExact);
3237   if (!IsExact || !Converted.isPowerOf2())
3238     return false;
3239 
3240   unsigned FracBits = Converted.logBase2();
3241   if (FracBits > ScalarBits)
3242     return false;
3243 
3244   SmallVector<SDValue, 3> Ops{
3245       VecVal, CurDAG->getConstant(FracBits, SDLoc(N), MVT::i32)};
3246   AddEmptyMVEPredicateToOps(Ops, SDLoc(N), Type);
3247 
3248   unsigned int Opcode;
3249   switch (ScalarBits) {
3250   case 16:
3251     if (FixedToFloat)
3252       Opcode = IsUnsigned ? ARM::MVE_VCVTf16u16_fix : ARM::MVE_VCVTf16s16_fix;
3253     else
3254       Opcode = IsUnsigned ? ARM::MVE_VCVTu16f16_fix : ARM::MVE_VCVTs16f16_fix;
3255     break;
3256   case 32:
3257     if (FixedToFloat)
3258       Opcode = IsUnsigned ? ARM::MVE_VCVTf32u32_fix : ARM::MVE_VCVTf32s32_fix;
3259     else
3260       Opcode = IsUnsigned ? ARM::MVE_VCVTu32f32_fix : ARM::MVE_VCVTs32f32_fix;
3261     break;
3262   default:
3263     llvm_unreachable("unexpected number of scalar bits");
3264     break;
3265   }
3266 
3267   ReplaceNode(N, CurDAG->getMachineNode(Opcode, SDLoc(N), Type, Ops));
3268   return true;
3269 }
3270 
3271 bool ARMDAGToDAGISel::tryFP_TO_INT(SDNode *N, SDLoc dl) {
3272   // Transform a floating-point to fixed-point conversion to a VCVT
3273   if (!Subtarget->hasMVEFloatOps())
3274     return false;
3275   EVT Type = N->getValueType(0);
3276   if (!Type.isVector())
3277     return false;
3278   unsigned int ScalarBits = Type.getScalarSizeInBits();
3279 
3280   bool IsUnsigned = N->getOpcode() == ISD::FP_TO_UINT ||
3281                     N->getOpcode() == ISD::FP_TO_UINT_SAT;
3282   SDNode *Node = N->getOperand(0).getNode();
3283 
3284   // floating-point to fixed-point with one fractional bit gets turned into an
3285   // FP_TO_[U|S]INT(FADD (x, x)) rather than an FP_TO_[U|S]INT(FMUL (x, y))
3286   if (Node->getOpcode() == ISD::FADD) {
3287     if (Node->getOperand(0) != Node->getOperand(1))
3288       return false;
3289     SDNodeFlags Flags = Node->getFlags();
3290     // The fixed-point vcvt and vcvt+vmul are not always equivalent if inf is
3291     // allowed in 16 bit unsigned floats
3292     if (ScalarBits == 16 && !Flags.hasNoInfs() && IsUnsigned)
3293       return false;
3294 
3295     unsigned Opcode;
3296     switch (ScalarBits) {
3297     case 16:
3298       Opcode = IsUnsigned ? ARM::MVE_VCVTu16f16_fix : ARM::MVE_VCVTs16f16_fix;
3299       break;
3300     case 32:
3301       Opcode = IsUnsigned ? ARM::MVE_VCVTu32f32_fix : ARM::MVE_VCVTs32f32_fix;
3302       break;
3303     }
3304     SmallVector<SDValue, 3> Ops{Node->getOperand(0),
3305                                 CurDAG->getConstant(1, dl, MVT::i32)};
3306     AddEmptyMVEPredicateToOps(Ops, dl, Type);
3307 
3308     ReplaceNode(N, CurDAG->getMachineNode(Opcode, dl, Type, Ops));
3309     return true;
3310   }
3311 
3312   if (Node->getOpcode() != ISD::FMUL)
3313     return false;
3314 
3315   return transformFixedFloatingPointConversion(N, Node, IsUnsigned, false);
3316 }
3317 
3318 bool ARMDAGToDAGISel::tryFMULFixed(SDNode *N, SDLoc dl) {
3319   // Transform a fixed-point to floating-point conversion to a VCVT
3320   if (!Subtarget->hasMVEFloatOps())
3321     return false;
3322   auto Type = N->getValueType(0);
3323   if (!Type.isVector())
3324     return false;
3325 
3326   auto LHS = N->getOperand(0);
3327   if (LHS.getOpcode() != ISD::SINT_TO_FP && LHS.getOpcode() != ISD::UINT_TO_FP)
3328     return false;
3329 
3330   return transformFixedFloatingPointConversion(
3331       N, N, LHS.getOpcode() == ISD::UINT_TO_FP, true);
3332 }
3333 
3334 bool ARMDAGToDAGISel::tryV6T2BitfieldExtractOp(SDNode *N, bool isSigned) {
3335   if (!Subtarget->hasV6T2Ops())
3336     return false;
3337 
3338   unsigned Opc = isSigned
3339     ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX)
3340     : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX);
3341   SDLoc dl(N);
3342 
3343   // For unsigned extracts, check for a shift right and mask
3344   unsigned And_imm = 0;
3345   if (N->getOpcode() == ISD::AND) {
3346     if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) {
3347 
3348       // The immediate is a mask of the low bits iff imm & (imm+1) == 0
3349       if (And_imm & (And_imm + 1))
3350         return false;
3351 
3352       unsigned Srl_imm = 0;
3353       if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL,
3354                                 Srl_imm)) {
3355         assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
3356 
3357         // Mask off the unnecessary bits of the AND immediate; normally
3358         // DAGCombine will do this, but that might not happen if
3359         // targetShrinkDemandedConstant chooses a different immediate.
3360         And_imm &= -1U >> Srl_imm;
3361 
3362         // Note: The width operand is encoded as width-1.
3363         unsigned Width = llvm::countr_one(And_imm) - 1;
3364         unsigned LSB = Srl_imm;
3365 
3366         SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3367 
3368         if ((LSB + Width + 1) == N->getValueType(0).getSizeInBits()) {
3369           // It's cheaper to use a right shift to extract the top bits.
3370           if (Subtarget->isThumb()) {
3371             Opc = isSigned ? ARM::t2ASRri : ARM::t2LSRri;
3372             SDValue Ops[] = { N->getOperand(0).getOperand(0),
3373                               CurDAG->getTargetConstant(LSB, dl, MVT::i32),
3374                               getAL(CurDAG, dl), Reg0, Reg0 };
3375             CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3376             return true;
3377           }
3378 
3379           // ARM models shift instructions as MOVsi with shifter operand.
3380           ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(ISD::SRL);
3381           SDValue ShOpc =
3382             CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, LSB), dl,
3383                                       MVT::i32);
3384           SDValue Ops[] = { N->getOperand(0).getOperand(0), ShOpc,
3385                             getAL(CurDAG, dl), Reg0, Reg0 };
3386           CurDAG->SelectNodeTo(N, ARM::MOVsi, MVT::i32, Ops);
3387           return true;
3388         }
3389 
3390         assert(LSB + Width + 1 <= 32 && "Shouldn't create an invalid ubfx");
3391         SDValue Ops[] = { N->getOperand(0).getOperand(0),
3392                           CurDAG->getTargetConstant(LSB, dl, MVT::i32),
3393                           CurDAG->getTargetConstant(Width, dl, MVT::i32),
3394                           getAL(CurDAG, dl), Reg0 };
3395         CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3396         return true;
3397       }
3398     }
3399     return false;
3400   }
3401 
3402   // Otherwise, we're looking for a shift of a shift
3403   unsigned Shl_imm = 0;
3404   if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) {
3405     assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
3406     unsigned Srl_imm = 0;
3407     if (isInt32Immediate(N->getOperand(1), Srl_imm)) {
3408       assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
3409       // Note: The width operand is encoded as width-1.
3410       unsigned Width = 32 - Srl_imm - 1;
3411       int LSB = Srl_imm - Shl_imm;
3412       if (LSB < 0)
3413         return false;
3414       SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3415       assert(LSB + Width + 1 <= 32 && "Shouldn't create an invalid ubfx");
3416       SDValue Ops[] = { N->getOperand(0).getOperand(0),
3417                         CurDAG->getTargetConstant(LSB, dl, MVT::i32),
3418                         CurDAG->getTargetConstant(Width, dl, MVT::i32),
3419                         getAL(CurDAG, dl), Reg0 };
3420       CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3421       return true;
3422     }
3423   }
3424 
3425   // Or we are looking for a shift of an and, with a mask operand
3426   if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, And_imm) &&
3427       isShiftedMask_32(And_imm)) {
3428     unsigned Srl_imm = 0;
3429     unsigned LSB = llvm::countr_zero(And_imm);
3430     // Shift must be the same as the ands lsb
3431     if (isInt32Immediate(N->getOperand(1), Srl_imm) && Srl_imm == LSB) {
3432       assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
3433       unsigned MSB = llvm::Log2_32(And_imm);
3434       // Note: The width operand is encoded as width-1.
3435       unsigned Width = MSB - LSB;
3436       SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3437       assert(Srl_imm + Width + 1 <= 32 && "Shouldn't create an invalid ubfx");
3438       SDValue Ops[] = { N->getOperand(0).getOperand(0),
3439                         CurDAG->getTargetConstant(Srl_imm, dl, MVT::i32),
3440                         CurDAG->getTargetConstant(Width, dl, MVT::i32),
3441                         getAL(CurDAG, dl), Reg0 };
3442       CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3443       return true;
3444     }
3445   }
3446 
3447   if (N->getOpcode() == ISD::SIGN_EXTEND_INREG) {
3448     unsigned Width = cast<VTSDNode>(N->getOperand(1))->getVT().getSizeInBits();
3449     unsigned LSB = 0;
3450     if (!isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL, LSB) &&
3451         !isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRA, LSB))
3452       return false;
3453 
3454     if (LSB + Width > 32)
3455       return false;
3456 
3457     SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3458     assert(LSB + Width <= 32 && "Shouldn't create an invalid ubfx");
3459     SDValue Ops[] = { N->getOperand(0).getOperand(0),
3460                       CurDAG->getTargetConstant(LSB, dl, MVT::i32),
3461                       CurDAG->getTargetConstant(Width - 1, dl, MVT::i32),
3462                       getAL(CurDAG, dl), Reg0 };
3463     CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3464     return true;
3465   }
3466 
3467   return false;
3468 }
3469 
3470 /// Target-specific DAG combining for ISD::SUB.
3471 /// Target-independent combining lowers SELECT_CC nodes of the form
3472 /// select_cc setg[ge] X,  0,  X, -X
3473 /// select_cc setgt    X, -1,  X, -X
3474 /// select_cc setl[te] X,  0, -X,  X
3475 /// select_cc setlt    X,  1, -X,  X
3476 /// which represent Integer ABS into:
3477 /// Y = sra (X, size(X)-1); sub (xor (X, Y), Y)
3478 /// ARM instruction selection detects the latter and matches it to
3479 /// ARM::ABS or ARM::t2ABS machine node.
3480 bool ARMDAGToDAGISel::tryABSOp(SDNode *N){
3481   SDValue SUBSrc0 = N->getOperand(0);
3482   SDValue SUBSrc1 = N->getOperand(1);
3483   EVT VT = N->getValueType(0);
3484 
3485   if (Subtarget->isThumb1Only())
3486     return false;
3487 
3488   if (SUBSrc0.getOpcode() != ISD::XOR || SUBSrc1.getOpcode() != ISD::SRA)
3489     return false;
3490 
3491   SDValue XORSrc0 = SUBSrc0.getOperand(0);
3492   SDValue XORSrc1 = SUBSrc0.getOperand(1);
3493   SDValue SRASrc0 = SUBSrc1.getOperand(0);
3494   SDValue SRASrc1 = SUBSrc1.getOperand(1);
3495   ConstantSDNode *SRAConstant =  dyn_cast<ConstantSDNode>(SRASrc1);
3496   EVT XType = SRASrc0.getValueType();
3497   unsigned Size = XType.getSizeInBits() - 1;
3498 
3499   if (XORSrc1 == SUBSrc1 && XORSrc0 == SRASrc0 && XType.isInteger() &&
3500       SRAConstant != nullptr && Size == SRAConstant->getZExtValue()) {
3501     unsigned Opcode = Subtarget->isThumb2() ? ARM::t2ABS : ARM::ABS;
3502     CurDAG->SelectNodeTo(N, Opcode, VT, XORSrc0);
3503     return true;
3504   }
3505 
3506   return false;
3507 }
3508 
3509 /// We've got special pseudo-instructions for these
3510 void ARMDAGToDAGISel::SelectCMP_SWAP(SDNode *N) {
3511   unsigned Opcode;
3512   EVT MemTy = cast<MemSDNode>(N)->getMemoryVT();
3513   if (MemTy == MVT::i8)
3514     Opcode = Subtarget->isThumb() ? ARM::tCMP_SWAP_8 : ARM::CMP_SWAP_8;
3515   else if (MemTy == MVT::i16)
3516     Opcode = Subtarget->isThumb() ? ARM::tCMP_SWAP_16 : ARM::CMP_SWAP_16;
3517   else if (MemTy == MVT::i32)
3518     Opcode = Subtarget->isThumb() ? ARM::tCMP_SWAP_32 : ARM::CMP_SWAP_32;
3519   else
3520     llvm_unreachable("Unknown AtomicCmpSwap type");
3521 
3522   SDValue Ops[] = {N->getOperand(1), N->getOperand(2), N->getOperand(3),
3523                    N->getOperand(0)};
3524   SDNode *CmpSwap = CurDAG->getMachineNode(
3525       Opcode, SDLoc(N),
3526       CurDAG->getVTList(MVT::i32, MVT::i32, MVT::Other), Ops);
3527 
3528   MachineMemOperand *MemOp = cast<MemSDNode>(N)->getMemOperand();
3529   CurDAG->setNodeMemRefs(cast<MachineSDNode>(CmpSwap), {MemOp});
3530 
3531   ReplaceUses(SDValue(N, 0), SDValue(CmpSwap, 0));
3532   ReplaceUses(SDValue(N, 1), SDValue(CmpSwap, 2));
3533   CurDAG->RemoveDeadNode(N);
3534 }
3535 
3536 static std::optional<std::pair<unsigned, unsigned>>
3537 getContiguousRangeOfSetBits(const APInt &A) {
3538   unsigned FirstOne = A.getBitWidth() - A.countl_zero() - 1;
3539   unsigned LastOne = A.countr_zero();
3540   if (A.popcount() != (FirstOne - LastOne + 1))
3541     return std::nullopt;
3542   return std::make_pair(FirstOne, LastOne);
3543 }
3544 
3545 void ARMDAGToDAGISel::SelectCMPZ(SDNode *N, bool &SwitchEQNEToPLMI) {
3546   assert(N->getOpcode() == ARMISD::CMPZ);
3547   SwitchEQNEToPLMI = false;
3548 
3549   if (!Subtarget->isThumb())
3550     // FIXME: Work out whether it is profitable to do this in A32 mode - LSL and
3551     // LSR don't exist as standalone instructions - they need the barrel shifter.
3552     return;
3553 
3554   // select (cmpz (and X, C), #0) -> (LSLS X) or (LSRS X) or (LSRS (LSLS X))
3555   SDValue And = N->getOperand(0);
3556   if (!And->hasOneUse())
3557     return;
3558 
3559   SDValue Zero = N->getOperand(1);
3560   if (!isNullConstant(Zero) || And->getOpcode() != ISD::AND)
3561     return;
3562   SDValue X = And.getOperand(0);
3563   auto C = dyn_cast<ConstantSDNode>(And.getOperand(1));
3564 
3565   if (!C)
3566     return;
3567   auto Range = getContiguousRangeOfSetBits(C->getAPIntValue());
3568   if (!Range)
3569     return;
3570 
3571   // There are several ways to lower this:
3572   SDNode *NewN;
3573   SDLoc dl(N);
3574 
3575   auto EmitShift = [&](unsigned Opc, SDValue Src, unsigned Imm) -> SDNode* {
3576     if (Subtarget->isThumb2()) {
3577       Opc = (Opc == ARM::tLSLri) ? ARM::t2LSLri : ARM::t2LSRri;
3578       SDValue Ops[] = { Src, CurDAG->getTargetConstant(Imm, dl, MVT::i32),
3579                         getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32),
3580                         CurDAG->getRegister(0, MVT::i32) };
3581       return CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops);
3582     } else {
3583       SDValue Ops[] = {CurDAG->getRegister(ARM::CPSR, MVT::i32), Src,
3584                        CurDAG->getTargetConstant(Imm, dl, MVT::i32),
3585                        getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32)};
3586       return CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops);
3587     }
3588   };
3589 
3590   if (Range->second == 0) {
3591     //  1. Mask includes the LSB -> Simply shift the top N bits off
3592     NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
3593     ReplaceNode(And.getNode(), NewN);
3594   } else if (Range->first == 31) {
3595     //  2. Mask includes the MSB -> Simply shift the bottom N bits off
3596     NewN = EmitShift(ARM::tLSRri, X, Range->second);
3597     ReplaceNode(And.getNode(), NewN);
3598   } else if (Range->first == Range->second) {
3599     //  3. Only one bit is set. We can shift this into the sign bit and use a
3600     //     PL/MI comparison.
3601     NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
3602     ReplaceNode(And.getNode(), NewN);
3603 
3604     SwitchEQNEToPLMI = true;
3605   } else if (!Subtarget->hasV6T2Ops()) {
3606     //  4. Do a double shift to clear bottom and top bits, but only in
3607     //     thumb-1 mode as in thumb-2 we can use UBFX.
3608     NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
3609     NewN = EmitShift(ARM::tLSRri, SDValue(NewN, 0),
3610                      Range->second + (31 - Range->first));
3611     ReplaceNode(And.getNode(), NewN);
3612   }
3613 }
3614 
3615 static unsigned getVectorShuffleOpcode(EVT VT, unsigned Opc64[3],
3616                                        unsigned Opc128[3]) {
3617   assert((VT.is64BitVector() || VT.is128BitVector()) &&
3618          "Unexpected vector shuffle length");
3619   switch (VT.getScalarSizeInBits()) {
3620   default:
3621     llvm_unreachable("Unexpected vector shuffle element size");
3622   case 8:
3623     return VT.is64BitVector() ? Opc64[0] : Opc128[0];
3624   case 16:
3625     return VT.is64BitVector() ? Opc64[1] : Opc128[1];
3626   case 32:
3627     return VT.is64BitVector() ? Opc64[2] : Opc128[2];
3628   }
3629 }
3630 
3631 void ARMDAGToDAGISel::Select(SDNode *N) {
3632   SDLoc dl(N);
3633 
3634   if (N->isMachineOpcode()) {
3635     N->setNodeId(-1);
3636     return;   // Already selected.
3637   }
3638 
3639   switch (N->getOpcode()) {
3640   default: break;
3641   case ISD::STORE: {
3642     // For Thumb1, match an sp-relative store in C++. This is a little
3643     // unfortunate, but I don't think I can make the chain check work
3644     // otherwise.  (The chain of the store has to be the same as the chain
3645     // of the CopyFromReg, or else we can't replace the CopyFromReg with
3646     // a direct reference to "SP".)
3647     //
3648     // This is only necessary on Thumb1 because Thumb1 sp-relative stores use
3649     // a different addressing mode from other four-byte stores.
3650     //
3651     // This pattern usually comes up with call arguments.
3652     StoreSDNode *ST = cast<StoreSDNode>(N);
3653     SDValue Ptr = ST->getBasePtr();
3654     if (Subtarget->isThumb1Only() && ST->isUnindexed()) {
3655       int RHSC = 0;
3656       if (Ptr.getOpcode() == ISD::ADD &&
3657           isScaledConstantInRange(Ptr.getOperand(1), /*Scale=*/4, 0, 256, RHSC))
3658         Ptr = Ptr.getOperand(0);
3659 
3660       if (Ptr.getOpcode() == ISD::CopyFromReg &&
3661           cast<RegisterSDNode>(Ptr.getOperand(1))->getReg() == ARM::SP &&
3662           Ptr.getOperand(0) == ST->getChain()) {
3663         SDValue Ops[] = {ST->getValue(),
3664                          CurDAG->getRegister(ARM::SP, MVT::i32),
3665                          CurDAG->getTargetConstant(RHSC, dl, MVT::i32),
3666                          getAL(CurDAG, dl),
3667                          CurDAG->getRegister(0, MVT::i32),
3668                          ST->getChain()};
3669         MachineSDNode *ResNode =
3670             CurDAG->getMachineNode(ARM::tSTRspi, dl, MVT::Other, Ops);
3671         MachineMemOperand *MemOp = ST->getMemOperand();
3672         CurDAG->setNodeMemRefs(cast<MachineSDNode>(ResNode), {MemOp});
3673         ReplaceNode(N, ResNode);
3674         return;
3675       }
3676     }
3677     break;
3678   }
3679   case ISD::WRITE_REGISTER:
3680     if (tryWriteRegister(N))
3681       return;
3682     break;
3683   case ISD::READ_REGISTER:
3684     if (tryReadRegister(N))
3685       return;
3686     break;
3687   case ISD::INLINEASM:
3688   case ISD::INLINEASM_BR:
3689     if (tryInlineAsm(N))
3690       return;
3691     break;
3692   case ISD::SUB:
3693     // Select special operations if SUB node forms integer ABS pattern
3694     if (tryABSOp(N))
3695       return;
3696     // Other cases are autogenerated.
3697     break;
3698   case ISD::Constant: {
3699     unsigned Val = N->getAsZExtVal();
3700     // If we can't materialize the constant we need to use a literal pool
3701     if (ConstantMaterializationCost(Val, Subtarget) > 2 &&
3702         !Subtarget->genExecuteOnly()) {
3703       SDValue CPIdx = CurDAG->getTargetConstantPool(
3704           ConstantInt::get(Type::getInt32Ty(*CurDAG->getContext()), Val),
3705           TLI->getPointerTy(CurDAG->getDataLayout()));
3706 
3707       SDNode *ResNode;
3708       if (Subtarget->isThumb()) {
3709         SDValue Ops[] = {
3710           CPIdx,
3711           getAL(CurDAG, dl),
3712           CurDAG->getRegister(0, MVT::i32),
3713           CurDAG->getEntryNode()
3714         };
3715         ResNode = CurDAG->getMachineNode(ARM::tLDRpci, dl, MVT::i32, MVT::Other,
3716                                          Ops);
3717       } else {
3718         SDValue Ops[] = {
3719           CPIdx,
3720           CurDAG->getTargetConstant(0, dl, MVT::i32),
3721           getAL(CurDAG, dl),
3722           CurDAG->getRegister(0, MVT::i32),
3723           CurDAG->getEntryNode()
3724         };
3725         ResNode = CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
3726                                          Ops);
3727       }
3728       // Annotate the Node with memory operand information so that MachineInstr
3729       // queries work properly. This e.g. gives the register allocation the
3730       // required information for rematerialization.
3731       MachineFunction& MF = CurDAG->getMachineFunction();
3732       MachineMemOperand *MemOp =
3733           MF.getMachineMemOperand(MachinePointerInfo::getConstantPool(MF),
3734                                   MachineMemOperand::MOLoad, 4, Align(4));
3735 
3736       CurDAG->setNodeMemRefs(cast<MachineSDNode>(ResNode), {MemOp});
3737 
3738       ReplaceNode(N, ResNode);
3739       return;
3740     }
3741 
3742     // Other cases are autogenerated.
3743     break;
3744   }
3745   case ISD::FrameIndex: {
3746     // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
3747     int FI = cast<FrameIndexSDNode>(N)->getIndex();
3748     SDValue TFI = CurDAG->getTargetFrameIndex(
3749         FI, TLI->getPointerTy(CurDAG->getDataLayout()));
3750     if (Subtarget->isThumb1Only()) {
3751       // Set the alignment of the frame object to 4, to avoid having to generate
3752       // more than one ADD
3753       MachineFrameInfo &MFI = MF->getFrameInfo();
3754       if (MFI.getObjectAlign(FI) < Align(4))
3755         MFI.setObjectAlignment(FI, Align(4));
3756       CurDAG->SelectNodeTo(N, ARM::tADDframe, MVT::i32, TFI,
3757                            CurDAG->getTargetConstant(0, dl, MVT::i32));
3758       return;
3759     } else {
3760       unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
3761                       ARM::t2ADDri : ARM::ADDri);
3762       SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, dl, MVT::i32),
3763                         getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32),
3764                         CurDAG->getRegister(0, MVT::i32) };
3765       CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops);
3766       return;
3767     }
3768   }
3769   case ISD::INSERT_VECTOR_ELT: {
3770     if (tryInsertVectorElt(N))
3771       return;
3772     break;
3773   }
3774   case ISD::SRL:
3775     if (tryV6T2BitfieldExtractOp(N, false))
3776       return;
3777     break;
3778   case ISD::SIGN_EXTEND_INREG:
3779   case ISD::SRA:
3780     if (tryV6T2BitfieldExtractOp(N, true))
3781       return;
3782     break;
3783   case ISD::FP_TO_UINT:
3784   case ISD::FP_TO_SINT:
3785   case ISD::FP_TO_UINT_SAT:
3786   case ISD::FP_TO_SINT_SAT:
3787     if (tryFP_TO_INT(N, dl))
3788       return;
3789     break;
3790   case ISD::FMUL:
3791     if (tryFMULFixed(N, dl))
3792       return;
3793     break;
3794   case ISD::MUL:
3795     if (Subtarget->isThumb1Only())
3796       break;
3797     if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
3798       unsigned RHSV = C->getZExtValue();
3799       if (!RHSV) break;
3800       if (isPowerOf2_32(RHSV-1)) {  // 2^n+1?
3801         unsigned ShImm = Log2_32(RHSV-1);
3802         if (ShImm >= 32)
3803           break;
3804         SDValue V = N->getOperand(0);
3805         ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
3806         SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, dl, MVT::i32);
3807         SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3808         if (Subtarget->isThumb()) {
3809           SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG, dl), Reg0, Reg0 };
3810           CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops);
3811           return;
3812         } else {
3813           SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG, dl), Reg0,
3814                             Reg0 };
3815           CurDAG->SelectNodeTo(N, ARM::ADDrsi, MVT::i32, Ops);
3816           return;
3817         }
3818       }
3819       if (isPowerOf2_32(RHSV+1)) {  // 2^n-1?
3820         unsigned ShImm = Log2_32(RHSV+1);
3821         if (ShImm >= 32)
3822           break;
3823         SDValue V = N->getOperand(0);
3824         ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
3825         SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, dl, MVT::i32);
3826         SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
3827         if (Subtarget->isThumb()) {
3828           SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG, dl), Reg0, Reg0 };
3829           CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops);
3830           return;
3831         } else {
3832           SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG, dl), Reg0,
3833                             Reg0 };
3834           CurDAG->SelectNodeTo(N, ARM::RSBrsi, MVT::i32, Ops);
3835           return;
3836         }
3837       }
3838     }
3839     break;
3840   case ISD::AND: {
3841     // Check for unsigned bitfield extract
3842     if (tryV6T2BitfieldExtractOp(N, false))
3843       return;
3844 
3845     // If an immediate is used in an AND node, it is possible that the immediate
3846     // can be more optimally materialized when negated. If this is the case we
3847     // can negate the immediate and use a BIC instead.
3848     auto *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1));
3849     if (N1C && N1C->hasOneUse() && Subtarget->isThumb()) {
3850       uint32_t Imm = (uint32_t) N1C->getZExtValue();
3851 
3852       // In Thumb2 mode, an AND can take a 12-bit immediate. If this
3853       // immediate can be negated and fit in the immediate operand of
3854       // a t2BIC, don't do any manual transform here as this can be
3855       // handled by the generic ISel machinery.
3856       bool PreferImmediateEncoding =
3857         Subtarget->hasThumb2() && (is_t2_so_imm(Imm) || is_t2_so_imm_not(Imm));
3858       if (!PreferImmediateEncoding &&
3859           ConstantMaterializationCost(Imm, Subtarget) >
3860               ConstantMaterializationCost(~Imm, Subtarget)) {
3861         // The current immediate costs more to materialize than a negated
3862         // immediate, so negate the immediate and use a BIC.
3863         SDValue NewImm =
3864           CurDAG->getConstant(~N1C->getZExtValue(), dl, MVT::i32);
3865         // If the new constant didn't exist before, reposition it in the topological
3866         // ordering so it is just before N. Otherwise, don't touch its location.
3867         if (NewImm->getNodeId() == -1)
3868           CurDAG->RepositionNode(N->getIterator(), NewImm.getNode());
3869 
3870         if (!Subtarget->hasThumb2()) {
3871           SDValue Ops[] = {CurDAG->getRegister(ARM::CPSR, MVT::i32),
3872                            N->getOperand(0), NewImm, getAL(CurDAG, dl),
3873                            CurDAG->getRegister(0, MVT::i32)};
3874           ReplaceNode(N, CurDAG->getMachineNode(ARM::tBIC, dl, MVT::i32, Ops));
3875           return;
3876         } else {
3877           SDValue Ops[] = {N->getOperand(0), NewImm, getAL(CurDAG, dl),
3878                            CurDAG->getRegister(0, MVT::i32),
3879                            CurDAG->getRegister(0, MVT::i32)};
3880           ReplaceNode(N,
3881                       CurDAG->getMachineNode(ARM::t2BICrr, dl, MVT::i32, Ops));
3882           return;
3883         }
3884       }
3885     }
3886 
3887     // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits
3888     // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits
3889     // are entirely contributed by c2 and lower 16-bits are entirely contributed
3890     // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)).
3891     // Select it to: "movt x, ((c1 & 0xffff) >> 16)
3892     EVT VT = N->getValueType(0);
3893     if (VT != MVT::i32)
3894       break;
3895     unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2())
3896       ? ARM::t2MOVTi16
3897       : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0);
3898     if (!Opc)
3899       break;
3900     SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3901     N1C = dyn_cast<ConstantSDNode>(N1);
3902     if (!N1C)
3903       break;
3904     if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) {
3905       SDValue N2 = N0.getOperand(1);
3906       ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
3907       if (!N2C)
3908         break;
3909       unsigned N1CVal = N1C->getZExtValue();
3910       unsigned N2CVal = N2C->getZExtValue();
3911       if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) &&
3912           (N1CVal & 0xffffU) == 0xffffU &&
3913           (N2CVal & 0xffffU) == 0x0U) {
3914         SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16,
3915                                                   dl, MVT::i32);
3916         SDValue Ops[] = { N0.getOperand(0), Imm16,
3917                           getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32) };
3918         ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, VT, Ops));
3919         return;
3920       }
3921     }
3922 
3923     break;
3924   }
3925   case ARMISD::UMAAL: {
3926     unsigned Opc = Subtarget->isThumb() ? ARM::t2UMAAL : ARM::UMAAL;
3927     SDValue Ops[] = { N->getOperand(0), N->getOperand(1),
3928                       N->getOperand(2), N->getOperand(3),
3929                       getAL(CurDAG, dl),
3930                       CurDAG->getRegister(0, MVT::i32) };
3931     ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, MVT::i32, MVT::i32, Ops));
3932     return;
3933   }
3934   case ARMISD::UMLAL:{
3935     if (Subtarget->isThumb()) {
3936       SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
3937                         N->getOperand(3), getAL(CurDAG, dl),
3938                         CurDAG->getRegister(0, MVT::i32)};
3939       ReplaceNode(
3940           N, CurDAG->getMachineNode(ARM::t2UMLAL, dl, MVT::i32, MVT::i32, Ops));
3941       return;
3942     }else{
3943       SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
3944                         N->getOperand(3), getAL(CurDAG, dl),
3945                         CurDAG->getRegister(0, MVT::i32),
3946                         CurDAG->getRegister(0, MVT::i32) };
3947       ReplaceNode(N, CurDAG->getMachineNode(
3948                          Subtarget->hasV6Ops() ? ARM::UMLAL : ARM::UMLALv5, dl,
3949                          MVT::i32, MVT::i32, Ops));
3950       return;
3951     }
3952   }
3953   case ARMISD::SMLAL:{
3954     if (Subtarget->isThumb()) {
3955       SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
3956                         N->getOperand(3), getAL(CurDAG, dl),
3957                         CurDAG->getRegister(0, MVT::i32)};
3958       ReplaceNode(
3959           N, CurDAG->getMachineNode(ARM::t2SMLAL, dl, MVT::i32, MVT::i32, Ops));
3960       return;
3961     }else{
3962       SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
3963                         N->getOperand(3), getAL(CurDAG, dl),
3964                         CurDAG->getRegister(0, MVT::i32),
3965                         CurDAG->getRegister(0, MVT::i32) };
3966       ReplaceNode(N, CurDAG->getMachineNode(
3967                          Subtarget->hasV6Ops() ? ARM::SMLAL : ARM::SMLALv5, dl,
3968                          MVT::i32, MVT::i32, Ops));
3969       return;
3970     }
3971   }
3972   case ARMISD::SUBE: {
3973     if (!Subtarget->hasV6Ops() || !Subtarget->hasDSP())
3974       break;
3975     // Look for a pattern to match SMMLS
3976     // (sube a, (smul_loHi a, b), (subc 0, (smul_LOhi(a, b))))
3977     if (N->getOperand(1).getOpcode() != ISD::SMUL_LOHI ||
3978         N->getOperand(2).getOpcode() != ARMISD::SUBC ||
3979         !SDValue(N, 1).use_empty())
3980       break;
3981 
3982     if (Subtarget->isThumb())
3983       assert(Subtarget->hasThumb2() &&
3984              "This pattern should not be generated for Thumb");
3985 
3986     SDValue SmulLoHi = N->getOperand(1);
3987     SDValue Subc = N->getOperand(2);
3988     SDValue Zero = Subc.getOperand(0);
3989 
3990     if (!isNullConstant(Zero) || Subc.getOperand(1) != SmulLoHi.getValue(0) ||
3991         N->getOperand(1) != SmulLoHi.getValue(1) ||
3992         N->getOperand(2) != Subc.getValue(1))
3993       break;
3994 
3995     unsigned Opc = Subtarget->isThumb2() ? ARM::t2SMMLS : ARM::SMMLS;
3996     SDValue Ops[] = { SmulLoHi.getOperand(0), SmulLoHi.getOperand(1),
3997                       N->getOperand(0), getAL(CurDAG, dl),
3998                       CurDAG->getRegister(0, MVT::i32) };
3999     ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops));
4000     return;
4001   }
4002   case ISD::LOAD: {
4003     if (Subtarget->hasMVEIntegerOps() && tryMVEIndexedLoad(N))
4004       return;
4005     if (Subtarget->isThumb() && Subtarget->hasThumb2()) {
4006       if (tryT2IndexedLoad(N))
4007         return;
4008     } else if (Subtarget->isThumb()) {
4009       if (tryT1IndexedLoad(N))
4010         return;
4011     } else if (tryARMIndexedLoad(N))
4012       return;
4013     // Other cases are autogenerated.
4014     break;
4015   }
4016   case ISD::MLOAD:
4017     if (Subtarget->hasMVEIntegerOps() && tryMVEIndexedLoad(N))
4018       return;
4019     // Other cases are autogenerated.
4020     break;
4021   case ARMISD::WLSSETUP: {
4022     SDNode *New = CurDAG->getMachineNode(ARM::t2WhileLoopSetup, dl, MVT::i32,
4023                                          N->getOperand(0));
4024     ReplaceUses(N, New);
4025     CurDAG->RemoveDeadNode(N);
4026     return;
4027   }
4028   case ARMISD::WLS: {
4029     SDNode *New = CurDAG->getMachineNode(ARM::t2WhileLoopStart, dl, MVT::Other,
4030                                          N->getOperand(1), N->getOperand(2),
4031                                          N->getOperand(0));
4032     ReplaceUses(N, New);
4033     CurDAG->RemoveDeadNode(N);
4034     return;
4035   }
4036   case ARMISD::LE: {
4037     SDValue Ops[] = { N->getOperand(1),
4038                       N->getOperand(2),
4039                       N->getOperand(0) };
4040     unsigned Opc = ARM::t2LoopEnd;
4041     SDNode *New = CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops);
4042     ReplaceUses(N, New);
4043     CurDAG->RemoveDeadNode(N);
4044     return;
4045   }
4046   case ARMISD::LDRD: {
4047     if (Subtarget->isThumb2())
4048       break; // TableGen handles isel in this case.
4049     SDValue Base, RegOffset, ImmOffset;
4050     const SDValue &Chain = N->getOperand(0);
4051     const SDValue &Addr = N->getOperand(1);
4052     SelectAddrMode3(Addr, Base, RegOffset, ImmOffset);
4053     if (RegOffset != CurDAG->getRegister(0, MVT::i32)) {
4054       // The register-offset variant of LDRD mandates that the register
4055       // allocated to RegOffset is not reused in any of the remaining operands.
4056       // This restriction is currently not enforced. Therefore emitting this
4057       // variant is explicitly avoided.
4058       Base = Addr;
4059       RegOffset = CurDAG->getRegister(0, MVT::i32);
4060     }
4061     SDValue Ops[] = {Base, RegOffset, ImmOffset, Chain};
4062     SDNode *New = CurDAG->getMachineNode(ARM::LOADDUAL, dl,
4063                                          {MVT::Untyped, MVT::Other}, Ops);
4064     SDValue Lo = CurDAG->getTargetExtractSubreg(ARM::gsub_0, dl, MVT::i32,
4065                                                 SDValue(New, 0));
4066     SDValue Hi = CurDAG->getTargetExtractSubreg(ARM::gsub_1, dl, MVT::i32,
4067                                                 SDValue(New, 0));
4068     transferMemOperands(N, New);
4069     ReplaceUses(SDValue(N, 0), Lo);
4070     ReplaceUses(SDValue(N, 1), Hi);
4071     ReplaceUses(SDValue(N, 2), SDValue(New, 1));
4072     CurDAG->RemoveDeadNode(N);
4073     return;
4074   }
4075   case ARMISD::STRD: {
4076     if (Subtarget->isThumb2())
4077       break; // TableGen handles isel in this case.
4078     SDValue Base, RegOffset, ImmOffset;
4079     const SDValue &Chain = N->getOperand(0);
4080     const SDValue &Addr = N->getOperand(3);
4081     SelectAddrMode3(Addr, Base, RegOffset, ImmOffset);
4082     if (RegOffset != CurDAG->getRegister(0, MVT::i32)) {
4083       // The register-offset variant of STRD mandates that the register
4084       // allocated to RegOffset is not reused in any of the remaining operands.
4085       // This restriction is currently not enforced. Therefore emitting this
4086       // variant is explicitly avoided.
4087       Base = Addr;
4088       RegOffset = CurDAG->getRegister(0, MVT::i32);
4089     }
4090     SDNode *RegPair =
4091         createGPRPairNode(MVT::Untyped, N->getOperand(1), N->getOperand(2));
4092     SDValue Ops[] = {SDValue(RegPair, 0), Base, RegOffset, ImmOffset, Chain};
4093     SDNode *New = CurDAG->getMachineNode(ARM::STOREDUAL, dl, MVT::Other, Ops);
4094     transferMemOperands(N, New);
4095     ReplaceUses(SDValue(N, 0), SDValue(New, 0));
4096     CurDAG->RemoveDeadNode(N);
4097     return;
4098   }
4099   case ARMISD::LOOP_DEC: {
4100     SDValue Ops[] = { N->getOperand(1),
4101                       N->getOperand(2),
4102                       N->getOperand(0) };
4103     SDNode *Dec =
4104       CurDAG->getMachineNode(ARM::t2LoopDec, dl,
4105                              CurDAG->getVTList(MVT::i32, MVT::Other), Ops);
4106     ReplaceUses(N, Dec);
4107     CurDAG->RemoveDeadNode(N);
4108     return;
4109   }
4110   case ARMISD::BRCOND: {
4111     // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
4112     // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
4113     // Pattern complexity = 6  cost = 1  size = 0
4114 
4115     // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
4116     // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
4117     // Pattern complexity = 6  cost = 1  size = 0
4118 
4119     // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
4120     // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
4121     // Pattern complexity = 6  cost = 1  size = 0
4122 
4123     unsigned Opc = Subtarget->isThumb() ?
4124       ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
4125     SDValue Chain = N->getOperand(0);
4126     SDValue N1 = N->getOperand(1);
4127     SDValue N2 = N->getOperand(2);
4128     SDValue N3 = N->getOperand(3);
4129     SDValue InGlue = N->getOperand(4);
4130     assert(N1.getOpcode() == ISD::BasicBlock);
4131     assert(N2.getOpcode() == ISD::Constant);
4132     assert(N3.getOpcode() == ISD::Register);
4133 
4134     unsigned CC = (unsigned)N2->getAsZExtVal();
4135 
4136     if (InGlue.getOpcode() == ARMISD::CMPZ) {
4137       if (InGlue.getOperand(0).getOpcode() == ISD::INTRINSIC_W_CHAIN) {
4138         SDValue Int = InGlue.getOperand(0);
4139         uint64_t ID = Int->getConstantOperandVal(1);
4140 
4141         // Handle low-overhead loops.
4142         if (ID == Intrinsic::loop_decrement_reg) {
4143           SDValue Elements = Int.getOperand(2);
4144           SDValue Size = CurDAG->getTargetConstant(Int.getConstantOperandVal(3),
4145                                                    dl, MVT::i32);
4146 
4147           SDValue Args[] = { Elements, Size, Int.getOperand(0) };
4148           SDNode *LoopDec =
4149             CurDAG->getMachineNode(ARM::t2LoopDec, dl,
4150                                    CurDAG->getVTList(MVT::i32, MVT::Other),
4151                                    Args);
4152           ReplaceUses(Int.getNode(), LoopDec);
4153 
4154           SDValue EndArgs[] = { SDValue(LoopDec, 0), N1, Chain };
4155           SDNode *LoopEnd =
4156             CurDAG->getMachineNode(ARM::t2LoopEnd, dl, MVT::Other, EndArgs);
4157 
4158           ReplaceUses(N, LoopEnd);
4159           CurDAG->RemoveDeadNode(N);
4160           CurDAG->RemoveDeadNode(InGlue.getNode());
4161           CurDAG->RemoveDeadNode(Int.getNode());
4162           return;
4163         }
4164       }
4165 
4166       bool SwitchEQNEToPLMI;
4167       SelectCMPZ(InGlue.getNode(), SwitchEQNEToPLMI);
4168       InGlue = N->getOperand(4);
4169 
4170       if (SwitchEQNEToPLMI) {
4171         switch ((ARMCC::CondCodes)CC) {
4172         default: llvm_unreachable("CMPZ must be either NE or EQ!");
4173         case ARMCC::NE:
4174           CC = (unsigned)ARMCC::MI;
4175           break;
4176         case ARMCC::EQ:
4177           CC = (unsigned)ARMCC::PL;
4178           break;
4179         }
4180       }
4181     }
4182 
4183     SDValue Tmp2 = CurDAG->getTargetConstant(CC, dl, MVT::i32);
4184     SDValue Ops[] = { N1, Tmp2, N3, Chain, InGlue };
4185     SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
4186                                              MVT::Glue, Ops);
4187     Chain = SDValue(ResNode, 0);
4188     if (N->getNumValues() == 2) {
4189       InGlue = SDValue(ResNode, 1);
4190       ReplaceUses(SDValue(N, 1), InGlue);
4191     }
4192     ReplaceUses(SDValue(N, 0),
4193                 SDValue(Chain.getNode(), Chain.getResNo()));
4194     CurDAG->RemoveDeadNode(N);
4195     return;
4196   }
4197 
4198   case ARMISD::CMPZ: {
4199     // select (CMPZ X, #-C) -> (CMPZ (ADDS X, #C), #0)
4200     //   This allows us to avoid materializing the expensive negative constant.
4201     //   The CMPZ #0 is useless and will be peepholed away but we need to keep it
4202     //   for its glue output.
4203     SDValue X = N->getOperand(0);
4204     auto *C = dyn_cast<ConstantSDNode>(N->getOperand(1).getNode());
4205     if (C && C->getSExtValue() < 0 && Subtarget->isThumb()) {
4206       int64_t Addend = -C->getSExtValue();
4207 
4208       SDNode *Add = nullptr;
4209       // ADDS can be better than CMN if the immediate fits in a
4210       // 16-bit ADDS, which means either [0,256) for tADDi8 or [0,8) for tADDi3.
4211       // Outside that range we can just use a CMN which is 32-bit but has a
4212       // 12-bit immediate range.
4213       if (Addend < 1<<8) {
4214         if (Subtarget->isThumb2()) {
4215           SDValue Ops[] = { X, CurDAG->getTargetConstant(Addend, dl, MVT::i32),
4216                             getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32),
4217                             CurDAG->getRegister(0, MVT::i32) };
4218           Add = CurDAG->getMachineNode(ARM::t2ADDri, dl, MVT::i32, Ops);
4219         } else {
4220           unsigned Opc = (Addend < 1<<3) ? ARM::tADDi3 : ARM::tADDi8;
4221           SDValue Ops[] = {CurDAG->getRegister(ARM::CPSR, MVT::i32), X,
4222                            CurDAG->getTargetConstant(Addend, dl, MVT::i32),
4223                            getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32)};
4224           Add = CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops);
4225         }
4226       }
4227       if (Add) {
4228         SDValue Ops2[] = {SDValue(Add, 0), CurDAG->getConstant(0, dl, MVT::i32)};
4229         CurDAG->MorphNodeTo(N, ARMISD::CMPZ, CurDAG->getVTList(MVT::Glue), Ops2);
4230       }
4231     }
4232     // Other cases are autogenerated.
4233     break;
4234   }
4235 
4236   case ARMISD::CMOV: {
4237     SDValue InGlue = N->getOperand(4);
4238 
4239     if (InGlue.getOpcode() == ARMISD::CMPZ) {
4240       bool SwitchEQNEToPLMI;
4241       SelectCMPZ(InGlue.getNode(), SwitchEQNEToPLMI);
4242 
4243       if (SwitchEQNEToPLMI) {
4244         SDValue ARMcc = N->getOperand(2);
4245         ARMCC::CondCodes CC = (ARMCC::CondCodes)ARMcc->getAsZExtVal();
4246 
4247         switch (CC) {
4248         default: llvm_unreachable("CMPZ must be either NE or EQ!");
4249         case ARMCC::NE:
4250           CC = ARMCC::MI;
4251           break;
4252         case ARMCC::EQ:
4253           CC = ARMCC::PL;
4254           break;
4255         }
4256         SDValue NewARMcc = CurDAG->getConstant((unsigned)CC, dl, MVT::i32);
4257         SDValue Ops[] = {N->getOperand(0), N->getOperand(1), NewARMcc,
4258                          N->getOperand(3), N->getOperand(4)};
4259         CurDAG->MorphNodeTo(N, ARMISD::CMOV, N->getVTList(), Ops);
4260       }
4261 
4262     }
4263     // Other cases are autogenerated.
4264     break;
4265   }
4266   case ARMISD::VZIP: {
4267     EVT VT = N->getValueType(0);
4268     // vzip.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm.
4269     unsigned Opc64[] = {ARM::VZIPd8, ARM::VZIPd16, ARM::VTRNd32};
4270     unsigned Opc128[] = {ARM::VZIPq8, ARM::VZIPq16, ARM::VZIPq32};
4271     unsigned Opc = getVectorShuffleOpcode(VT, Opc64, Opc128);
4272     SDValue Pred = getAL(CurDAG, dl);
4273     SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
4274     SDValue Ops[] = {N->getOperand(0), N->getOperand(1), Pred, PredReg};
4275     ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, VT, VT, Ops));
4276     return;
4277   }
4278   case ARMISD::VUZP: {
4279     EVT VT = N->getValueType(0);
4280     // vuzp.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm.
4281     unsigned Opc64[] = {ARM::VUZPd8, ARM::VUZPd16, ARM::VTRNd32};
4282     unsigned Opc128[] = {ARM::VUZPq8, ARM::VUZPq16, ARM::VUZPq32};
4283     unsigned Opc = getVectorShuffleOpcode(VT, Opc64, Opc128);
4284     SDValue Pred = getAL(CurDAG, dl);
4285     SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
4286     SDValue Ops[] = {N->getOperand(0), N->getOperand(1), Pred, PredReg};
4287     ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, VT, VT, Ops));
4288     return;
4289   }
4290   case ARMISD::VTRN: {
4291     EVT VT = N->getValueType(0);
4292     unsigned Opc64[] = {ARM::VTRNd8, ARM::VTRNd16, ARM::VTRNd32};
4293     unsigned Opc128[] = {ARM::VTRNq8, ARM::VTRNq16, ARM::VTRNq32};
4294     unsigned Opc = getVectorShuffleOpcode(VT, Opc64, Opc128);
4295     SDValue Pred = getAL(CurDAG, dl);
4296     SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
4297     SDValue Ops[] = {N->getOperand(0), N->getOperand(1), Pred, PredReg};
4298     ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, VT, VT, Ops));
4299     return;
4300   }
4301   case ARMISD::BUILD_VECTOR: {
4302     EVT VecVT = N->getValueType(0);
4303     EVT EltVT = VecVT.getVectorElementType();
4304     unsigned NumElts = VecVT.getVectorNumElements();
4305     if (EltVT == MVT::f64) {
4306       assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
4307       ReplaceNode(
4308           N, createDRegPairNode(VecVT, N->getOperand(0), N->getOperand(1)));
4309       return;
4310     }
4311     assert(EltVT == MVT::f32 && "unexpected type for BUILD_VECTOR");
4312     if (NumElts == 2) {
4313       ReplaceNode(
4314           N, createSRegPairNode(VecVT, N->getOperand(0), N->getOperand(1)));
4315       return;
4316     }
4317     assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
4318     ReplaceNode(N,
4319                 createQuadSRegsNode(VecVT, N->getOperand(0), N->getOperand(1),
4320                                     N->getOperand(2), N->getOperand(3)));
4321     return;
4322   }
4323 
4324   case ARMISD::VLD1DUP: {
4325     static const uint16_t DOpcodes[] = { ARM::VLD1DUPd8, ARM::VLD1DUPd16,
4326                                          ARM::VLD1DUPd32 };
4327     static const uint16_t QOpcodes[] = { ARM::VLD1DUPq8, ARM::VLD1DUPq16,
4328                                          ARM::VLD1DUPq32 };
4329     SelectVLDDup(N, /* IsIntrinsic= */ false, false, 1, DOpcodes, QOpcodes);
4330     return;
4331   }
4332 
4333   case ARMISD::VLD2DUP: {
4334     static const uint16_t Opcodes[] = { ARM::VLD2DUPd8, ARM::VLD2DUPd16,
4335                                         ARM::VLD2DUPd32 };
4336     SelectVLDDup(N, /* IsIntrinsic= */ false, false, 2, Opcodes);
4337     return;
4338   }
4339 
4340   case ARMISD::VLD3DUP: {
4341     static const uint16_t Opcodes[] = { ARM::VLD3DUPd8Pseudo,
4342                                         ARM::VLD3DUPd16Pseudo,
4343                                         ARM::VLD3DUPd32Pseudo };
4344     SelectVLDDup(N, /* IsIntrinsic= */ false, false, 3, Opcodes);
4345     return;
4346   }
4347 
4348   case ARMISD::VLD4DUP: {
4349     static const uint16_t Opcodes[] = { ARM::VLD4DUPd8Pseudo,
4350                                         ARM::VLD4DUPd16Pseudo,
4351                                         ARM::VLD4DUPd32Pseudo };
4352     SelectVLDDup(N, /* IsIntrinsic= */ false, false, 4, Opcodes);
4353     return;
4354   }
4355 
4356   case ARMISD::VLD1DUP_UPD: {
4357     static const uint16_t DOpcodes[] = { ARM::VLD1DUPd8wb_fixed,
4358                                          ARM::VLD1DUPd16wb_fixed,
4359                                          ARM::VLD1DUPd32wb_fixed };
4360     static const uint16_t QOpcodes[] = { ARM::VLD1DUPq8wb_fixed,
4361                                          ARM::VLD1DUPq16wb_fixed,
4362                                          ARM::VLD1DUPq32wb_fixed };
4363     SelectVLDDup(N, /* IsIntrinsic= */ false, true, 1, DOpcodes, QOpcodes);
4364     return;
4365   }
4366 
4367   case ARMISD::VLD2DUP_UPD: {
4368     static const uint16_t DOpcodes[] = { ARM::VLD2DUPd8wb_fixed,
4369                                          ARM::VLD2DUPd16wb_fixed,
4370                                          ARM::VLD2DUPd32wb_fixed,
4371                                          ARM::VLD1q64wb_fixed };
4372     static const uint16_t QOpcodes0[] = { ARM::VLD2DUPq8EvenPseudo,
4373                                           ARM::VLD2DUPq16EvenPseudo,
4374                                           ARM::VLD2DUPq32EvenPseudo };
4375     static const uint16_t QOpcodes1[] = { ARM::VLD2DUPq8OddPseudoWB_fixed,
4376                                           ARM::VLD2DUPq16OddPseudoWB_fixed,
4377                                           ARM::VLD2DUPq32OddPseudoWB_fixed };
4378     SelectVLDDup(N, /* IsIntrinsic= */ false, true, 2, DOpcodes, QOpcodes0, QOpcodes1);
4379     return;
4380   }
4381 
4382   case ARMISD::VLD3DUP_UPD: {
4383     static const uint16_t DOpcodes[] = { ARM::VLD3DUPd8Pseudo_UPD,
4384                                          ARM::VLD3DUPd16Pseudo_UPD,
4385                                          ARM::VLD3DUPd32Pseudo_UPD,
4386                                          ARM::VLD1d64TPseudoWB_fixed };
4387     static const uint16_t QOpcodes0[] = { ARM::VLD3DUPq8EvenPseudo,
4388                                           ARM::VLD3DUPq16EvenPseudo,
4389                                           ARM::VLD3DUPq32EvenPseudo };
4390     static const uint16_t QOpcodes1[] = { ARM::VLD3DUPq8OddPseudo_UPD,
4391                                           ARM::VLD3DUPq16OddPseudo_UPD,
4392                                           ARM::VLD3DUPq32OddPseudo_UPD };
4393     SelectVLDDup(N, /* IsIntrinsic= */ false, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
4394     return;
4395   }
4396 
4397   case ARMISD::VLD4DUP_UPD: {
4398     static const uint16_t DOpcodes[] = { ARM::VLD4DUPd8Pseudo_UPD,
4399                                          ARM::VLD4DUPd16Pseudo_UPD,
4400                                          ARM::VLD4DUPd32Pseudo_UPD,
4401                                          ARM::VLD1d64QPseudoWB_fixed };
4402     static const uint16_t QOpcodes0[] = { ARM::VLD4DUPq8EvenPseudo,
4403                                           ARM::VLD4DUPq16EvenPseudo,
4404                                           ARM::VLD4DUPq32EvenPseudo };
4405     static const uint16_t QOpcodes1[] = { ARM::VLD4DUPq8OddPseudo_UPD,
4406                                           ARM::VLD4DUPq16OddPseudo_UPD,
4407                                           ARM::VLD4DUPq32OddPseudo_UPD };
4408     SelectVLDDup(N, /* IsIntrinsic= */ false, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
4409     return;
4410   }
4411 
4412   case ARMISD::VLD1_UPD: {
4413     static const uint16_t DOpcodes[] = { ARM::VLD1d8wb_fixed,
4414                                          ARM::VLD1d16wb_fixed,
4415                                          ARM::VLD1d32wb_fixed,
4416                                          ARM::VLD1d64wb_fixed };
4417     static const uint16_t QOpcodes[] = { ARM::VLD1q8wb_fixed,
4418                                          ARM::VLD1q16wb_fixed,
4419                                          ARM::VLD1q32wb_fixed,
4420                                          ARM::VLD1q64wb_fixed };
4421     SelectVLD(N, true, 1, DOpcodes, QOpcodes, nullptr);
4422     return;
4423   }
4424 
4425   case ARMISD::VLD2_UPD: {
4426     if (Subtarget->hasNEON()) {
4427       static const uint16_t DOpcodes[] = {
4428           ARM::VLD2d8wb_fixed, ARM::VLD2d16wb_fixed, ARM::VLD2d32wb_fixed,
4429           ARM::VLD1q64wb_fixed};
4430       static const uint16_t QOpcodes[] = {ARM::VLD2q8PseudoWB_fixed,
4431                                           ARM::VLD2q16PseudoWB_fixed,
4432                                           ARM::VLD2q32PseudoWB_fixed};
4433       SelectVLD(N, true, 2, DOpcodes, QOpcodes, nullptr);
4434     } else {
4435       static const uint16_t Opcodes8[] = {ARM::MVE_VLD20_8,
4436                                           ARM::MVE_VLD21_8_wb};
4437       static const uint16_t Opcodes16[] = {ARM::MVE_VLD20_16,
4438                                            ARM::MVE_VLD21_16_wb};
4439       static const uint16_t Opcodes32[] = {ARM::MVE_VLD20_32,
4440                                            ARM::MVE_VLD21_32_wb};
4441       static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32};
4442       SelectMVE_VLD(N, 2, Opcodes, true);
4443     }
4444     return;
4445   }
4446 
4447   case ARMISD::VLD3_UPD: {
4448     static const uint16_t DOpcodes[] = { ARM::VLD3d8Pseudo_UPD,
4449                                          ARM::VLD3d16Pseudo_UPD,
4450                                          ARM::VLD3d32Pseudo_UPD,
4451                                          ARM::VLD1d64TPseudoWB_fixed};
4452     static const uint16_t QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD,
4453                                           ARM::VLD3q16Pseudo_UPD,
4454                                           ARM::VLD3q32Pseudo_UPD };
4455     static const uint16_t QOpcodes1[] = { ARM::VLD3q8oddPseudo_UPD,
4456                                           ARM::VLD3q16oddPseudo_UPD,
4457                                           ARM::VLD3q32oddPseudo_UPD };
4458     SelectVLD(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
4459     return;
4460   }
4461 
4462   case ARMISD::VLD4_UPD: {
4463     if (Subtarget->hasNEON()) {
4464       static const uint16_t DOpcodes[] = {
4465           ARM::VLD4d8Pseudo_UPD, ARM::VLD4d16Pseudo_UPD, ARM::VLD4d32Pseudo_UPD,
4466           ARM::VLD1d64QPseudoWB_fixed};
4467       static const uint16_t QOpcodes0[] = {ARM::VLD4q8Pseudo_UPD,
4468                                            ARM::VLD4q16Pseudo_UPD,
4469                                            ARM::VLD4q32Pseudo_UPD};
4470       static const uint16_t QOpcodes1[] = {ARM::VLD4q8oddPseudo_UPD,
4471                                            ARM::VLD4q16oddPseudo_UPD,
4472                                            ARM::VLD4q32oddPseudo_UPD};
4473       SelectVLD(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
4474     } else {
4475       static const uint16_t Opcodes8[] = {ARM::MVE_VLD40_8, ARM::MVE_VLD41_8,
4476                                           ARM::MVE_VLD42_8,
4477                                           ARM::MVE_VLD43_8_wb};
4478       static const uint16_t Opcodes16[] = {ARM::MVE_VLD40_16, ARM::MVE_VLD41_16,
4479                                            ARM::MVE_VLD42_16,
4480                                            ARM::MVE_VLD43_16_wb};
4481       static const uint16_t Opcodes32[] = {ARM::MVE_VLD40_32, ARM::MVE_VLD41_32,
4482                                            ARM::MVE_VLD42_32,
4483                                            ARM::MVE_VLD43_32_wb};
4484       static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32};
4485       SelectMVE_VLD(N, 4, Opcodes, true);
4486     }
4487     return;
4488   }
4489 
4490   case ARMISD::VLD1x2_UPD: {
4491     if (Subtarget->hasNEON()) {
4492       static const uint16_t DOpcodes[] = {
4493           ARM::VLD1q8wb_fixed, ARM::VLD1q16wb_fixed, ARM::VLD1q32wb_fixed,
4494           ARM::VLD1q64wb_fixed};
4495       static const uint16_t QOpcodes[] = {
4496           ARM::VLD1d8QPseudoWB_fixed, ARM::VLD1d16QPseudoWB_fixed,
4497           ARM::VLD1d32QPseudoWB_fixed, ARM::VLD1d64QPseudoWB_fixed};
4498       SelectVLD(N, true, 2, DOpcodes, QOpcodes, nullptr);
4499       return;
4500     }
4501     break;
4502   }
4503 
4504   case ARMISD::VLD1x3_UPD: {
4505     if (Subtarget->hasNEON()) {
4506       static const uint16_t DOpcodes[] = {
4507           ARM::VLD1d8TPseudoWB_fixed, ARM::VLD1d16TPseudoWB_fixed,
4508           ARM::VLD1d32TPseudoWB_fixed, ARM::VLD1d64TPseudoWB_fixed};
4509       static const uint16_t QOpcodes0[] = {
4510           ARM::VLD1q8LowTPseudo_UPD, ARM::VLD1q16LowTPseudo_UPD,
4511           ARM::VLD1q32LowTPseudo_UPD, ARM::VLD1q64LowTPseudo_UPD};
4512       static const uint16_t QOpcodes1[] = {
4513           ARM::VLD1q8HighTPseudo_UPD, ARM::VLD1q16HighTPseudo_UPD,
4514           ARM::VLD1q32HighTPseudo_UPD, ARM::VLD1q64HighTPseudo_UPD};
4515       SelectVLD(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
4516       return;
4517     }
4518     break;
4519   }
4520 
4521   case ARMISD::VLD1x4_UPD: {
4522     if (Subtarget->hasNEON()) {
4523       static const uint16_t DOpcodes[] = {
4524           ARM::VLD1d8QPseudoWB_fixed, ARM::VLD1d16QPseudoWB_fixed,
4525           ARM::VLD1d32QPseudoWB_fixed, ARM::VLD1d64QPseudoWB_fixed};
4526       static const uint16_t QOpcodes0[] = {
4527           ARM::VLD1q8LowQPseudo_UPD, ARM::VLD1q16LowQPseudo_UPD,
4528           ARM::VLD1q32LowQPseudo_UPD, ARM::VLD1q64LowQPseudo_UPD};
4529       static const uint16_t QOpcodes1[] = {
4530           ARM::VLD1q8HighQPseudo_UPD, ARM::VLD1q16HighQPseudo_UPD,
4531           ARM::VLD1q32HighQPseudo_UPD, ARM::VLD1q64HighQPseudo_UPD};
4532       SelectVLD(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
4533       return;
4534     }
4535     break;
4536   }
4537 
4538   case ARMISD::VLD2LN_UPD: {
4539     static const uint16_t DOpcodes[] = { ARM::VLD2LNd8Pseudo_UPD,
4540                                          ARM::VLD2LNd16Pseudo_UPD,
4541                                          ARM::VLD2LNd32Pseudo_UPD };
4542     static const uint16_t QOpcodes[] = { ARM::VLD2LNq16Pseudo_UPD,
4543                                          ARM::VLD2LNq32Pseudo_UPD };
4544     SelectVLDSTLane(N, true, true, 2, DOpcodes, QOpcodes);
4545     return;
4546   }
4547 
4548   case ARMISD::VLD3LN_UPD: {
4549     static const uint16_t DOpcodes[] = { ARM::VLD3LNd8Pseudo_UPD,
4550                                          ARM::VLD3LNd16Pseudo_UPD,
4551                                          ARM::VLD3LNd32Pseudo_UPD };
4552     static const uint16_t QOpcodes[] = { ARM::VLD3LNq16Pseudo_UPD,
4553                                          ARM::VLD3LNq32Pseudo_UPD };
4554     SelectVLDSTLane(N, true, true, 3, DOpcodes, QOpcodes);
4555     return;
4556   }
4557 
4558   case ARMISD::VLD4LN_UPD: {
4559     static const uint16_t DOpcodes[] = { ARM::VLD4LNd8Pseudo_UPD,
4560                                          ARM::VLD4LNd16Pseudo_UPD,
4561                                          ARM::VLD4LNd32Pseudo_UPD };
4562     static const uint16_t QOpcodes[] = { ARM::VLD4LNq16Pseudo_UPD,
4563                                          ARM::VLD4LNq32Pseudo_UPD };
4564     SelectVLDSTLane(N, true, true, 4, DOpcodes, QOpcodes);
4565     return;
4566   }
4567 
4568   case ARMISD::VST1_UPD: {
4569     static const uint16_t DOpcodes[] = { ARM::VST1d8wb_fixed,
4570                                          ARM::VST1d16wb_fixed,
4571                                          ARM::VST1d32wb_fixed,
4572                                          ARM::VST1d64wb_fixed };
4573     static const uint16_t QOpcodes[] = { ARM::VST1q8wb_fixed,
4574                                          ARM::VST1q16wb_fixed,
4575                                          ARM::VST1q32wb_fixed,
4576                                          ARM::VST1q64wb_fixed };
4577     SelectVST(N, true, 1, DOpcodes, QOpcodes, nullptr);
4578     return;
4579   }
4580 
4581   case ARMISD::VST2_UPD: {
4582     if (Subtarget->hasNEON()) {
4583       static const uint16_t DOpcodes[] = {
4584           ARM::VST2d8wb_fixed, ARM::VST2d16wb_fixed, ARM::VST2d32wb_fixed,
4585           ARM::VST1q64wb_fixed};
4586       static const uint16_t QOpcodes[] = {ARM::VST2q8PseudoWB_fixed,
4587                                           ARM::VST2q16PseudoWB_fixed,
4588                                           ARM::VST2q32PseudoWB_fixed};
4589       SelectVST(N, true, 2, DOpcodes, QOpcodes, nullptr);
4590       return;
4591     }
4592     break;
4593   }
4594 
4595   case ARMISD::VST3_UPD: {
4596     static const uint16_t DOpcodes[] = { ARM::VST3d8Pseudo_UPD,
4597                                          ARM::VST3d16Pseudo_UPD,
4598                                          ARM::VST3d32Pseudo_UPD,
4599                                          ARM::VST1d64TPseudoWB_fixed};
4600     static const uint16_t QOpcodes0[] = { ARM::VST3q8Pseudo_UPD,
4601                                           ARM::VST3q16Pseudo_UPD,
4602                                           ARM::VST3q32Pseudo_UPD };
4603     static const uint16_t QOpcodes1[] = { ARM::VST3q8oddPseudo_UPD,
4604                                           ARM::VST3q16oddPseudo_UPD,
4605                                           ARM::VST3q32oddPseudo_UPD };
4606     SelectVST(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
4607     return;
4608   }
4609 
4610   case ARMISD::VST4_UPD: {
4611     if (Subtarget->hasNEON()) {
4612       static const uint16_t DOpcodes[] = {
4613           ARM::VST4d8Pseudo_UPD, ARM::VST4d16Pseudo_UPD, ARM::VST4d32Pseudo_UPD,
4614           ARM::VST1d64QPseudoWB_fixed};
4615       static const uint16_t QOpcodes0[] = {ARM::VST4q8Pseudo_UPD,
4616                                            ARM::VST4q16Pseudo_UPD,
4617                                            ARM::VST4q32Pseudo_UPD};
4618       static const uint16_t QOpcodes1[] = {ARM::VST4q8oddPseudo_UPD,
4619                                            ARM::VST4q16oddPseudo_UPD,
4620                                            ARM::VST4q32oddPseudo_UPD};
4621       SelectVST(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
4622       return;
4623     }
4624     break;
4625   }
4626 
4627   case ARMISD::VST1x2_UPD: {
4628     if (Subtarget->hasNEON()) {
4629       static const uint16_t DOpcodes[] = { ARM::VST1q8wb_fixed,
4630                                            ARM::VST1q16wb_fixed,
4631                                            ARM::VST1q32wb_fixed,
4632                                            ARM::VST1q64wb_fixed};
4633       static const uint16_t QOpcodes[] = { ARM::VST1d8QPseudoWB_fixed,
4634                                            ARM::VST1d16QPseudoWB_fixed,
4635                                            ARM::VST1d32QPseudoWB_fixed,
4636                                            ARM::VST1d64QPseudoWB_fixed };
4637       SelectVST(N, true, 2, DOpcodes, QOpcodes, nullptr);
4638       return;
4639     }
4640     break;
4641   }
4642 
4643   case ARMISD::VST1x3_UPD: {
4644     if (Subtarget->hasNEON()) {
4645       static const uint16_t DOpcodes[] = { ARM::VST1d8TPseudoWB_fixed,
4646                                            ARM::VST1d16TPseudoWB_fixed,
4647                                            ARM::VST1d32TPseudoWB_fixed,
4648                                            ARM::VST1d64TPseudoWB_fixed };
4649       static const uint16_t QOpcodes0[] = { ARM::VST1q8LowTPseudo_UPD,
4650                                             ARM::VST1q16LowTPseudo_UPD,
4651                                             ARM::VST1q32LowTPseudo_UPD,
4652                                             ARM::VST1q64LowTPseudo_UPD };
4653       static const uint16_t QOpcodes1[] = { ARM::VST1q8HighTPseudo_UPD,
4654                                             ARM::VST1q16HighTPseudo_UPD,
4655                                             ARM::VST1q32HighTPseudo_UPD,
4656                                             ARM::VST1q64HighTPseudo_UPD };
4657       SelectVST(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1);
4658       return;
4659     }
4660     break;
4661   }
4662 
4663   case ARMISD::VST1x4_UPD: {
4664     if (Subtarget->hasNEON()) {
4665       static const uint16_t DOpcodes[] = { ARM::VST1d8QPseudoWB_fixed,
4666                                            ARM::VST1d16QPseudoWB_fixed,
4667                                            ARM::VST1d32QPseudoWB_fixed,
4668                                            ARM::VST1d64QPseudoWB_fixed };
4669       static const uint16_t QOpcodes0[] = { ARM::VST1q8LowQPseudo_UPD,
4670                                             ARM::VST1q16LowQPseudo_UPD,
4671                                             ARM::VST1q32LowQPseudo_UPD,
4672                                             ARM::VST1q64LowQPseudo_UPD };
4673       static const uint16_t QOpcodes1[] = { ARM::VST1q8HighQPseudo_UPD,
4674                                             ARM::VST1q16HighQPseudo_UPD,
4675                                             ARM::VST1q32HighQPseudo_UPD,
4676                                             ARM::VST1q64HighQPseudo_UPD };
4677       SelectVST(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1);
4678       return;
4679     }
4680     break;
4681   }
4682   case ARMISD::VST2LN_UPD: {
4683     static const uint16_t DOpcodes[] = { ARM::VST2LNd8Pseudo_UPD,
4684                                          ARM::VST2LNd16Pseudo_UPD,
4685                                          ARM::VST2LNd32Pseudo_UPD };
4686     static const uint16_t QOpcodes[] = { ARM::VST2LNq16Pseudo_UPD,
4687                                          ARM::VST2LNq32Pseudo_UPD };
4688     SelectVLDSTLane(N, false, true, 2, DOpcodes, QOpcodes);
4689     return;
4690   }
4691 
4692   case ARMISD::VST3LN_UPD: {
4693     static const uint16_t DOpcodes[] = { ARM::VST3LNd8Pseudo_UPD,
4694                                          ARM::VST3LNd16Pseudo_UPD,
4695                                          ARM::VST3LNd32Pseudo_UPD };
4696     static const uint16_t QOpcodes[] = { ARM::VST3LNq16Pseudo_UPD,
4697                                          ARM::VST3LNq32Pseudo_UPD };
4698     SelectVLDSTLane(N, false, true, 3, DOpcodes, QOpcodes);
4699     return;
4700   }
4701 
4702   case ARMISD::VST4LN_UPD: {
4703     static const uint16_t DOpcodes[] = { ARM::VST4LNd8Pseudo_UPD,
4704                                          ARM::VST4LNd16Pseudo_UPD,
4705                                          ARM::VST4LNd32Pseudo_UPD };
4706     static const uint16_t QOpcodes[] = { ARM::VST4LNq16Pseudo_UPD,
4707                                          ARM::VST4LNq32Pseudo_UPD };
4708     SelectVLDSTLane(N, false, true, 4, DOpcodes, QOpcodes);
4709     return;
4710   }
4711 
4712   case ISD::INTRINSIC_VOID:
4713   case ISD::INTRINSIC_W_CHAIN: {
4714     unsigned IntNo = N->getConstantOperandVal(1);
4715     switch (IntNo) {
4716     default:
4717       break;
4718 
4719     case Intrinsic::arm_mrrc:
4720     case Intrinsic::arm_mrrc2: {
4721       SDLoc dl(N);
4722       SDValue Chain = N->getOperand(0);
4723       unsigned Opc;
4724 
4725       if (Subtarget->isThumb())
4726         Opc = (IntNo == Intrinsic::arm_mrrc ? ARM::t2MRRC : ARM::t2MRRC2);
4727       else
4728         Opc = (IntNo == Intrinsic::arm_mrrc ? ARM::MRRC : ARM::MRRC2);
4729 
4730       SmallVector<SDValue, 5> Ops;
4731       Ops.push_back(getI32Imm(N->getConstantOperandVal(2), dl)); /* coproc */
4732       Ops.push_back(getI32Imm(N->getConstantOperandVal(3), dl)); /* opc */
4733       Ops.push_back(getI32Imm(N->getConstantOperandVal(4), dl)); /* CRm */
4734 
4735       // The mrrc2 instruction in ARM doesn't allow predicates, the top 4 bits of the encoded
4736       // instruction will always be '1111' but it is possible in assembly language to specify
4737       // AL as a predicate to mrrc2 but it doesn't make any difference to the encoded instruction.
4738       if (Opc != ARM::MRRC2) {
4739         Ops.push_back(getAL(CurDAG, dl));
4740         Ops.push_back(CurDAG->getRegister(0, MVT::i32));
4741       }
4742 
4743       Ops.push_back(Chain);
4744 
4745       // Writes to two registers.
4746       const EVT RetType[] = {MVT::i32, MVT::i32, MVT::Other};
4747 
4748       ReplaceNode(N, CurDAG->getMachineNode(Opc, dl, RetType, Ops));
4749       return;
4750     }
4751     case Intrinsic::arm_ldaexd:
4752     case Intrinsic::arm_ldrexd: {
4753       SDLoc dl(N);
4754       SDValue Chain = N->getOperand(0);
4755       SDValue MemAddr = N->getOperand(2);
4756       bool isThumb = Subtarget->isThumb() && Subtarget->hasV8MBaselineOps();
4757 
4758       bool IsAcquire = IntNo == Intrinsic::arm_ldaexd;
4759       unsigned NewOpc = isThumb ? (IsAcquire ? ARM::t2LDAEXD : ARM::t2LDREXD)
4760                                 : (IsAcquire ? ARM::LDAEXD : ARM::LDREXD);
4761 
4762       // arm_ldrexd returns a i64 value in {i32, i32}
4763       std::vector<EVT> ResTys;
4764       if (isThumb) {
4765         ResTys.push_back(MVT::i32);
4766         ResTys.push_back(MVT::i32);
4767       } else
4768         ResTys.push_back(MVT::Untyped);
4769       ResTys.push_back(MVT::Other);
4770 
4771       // Place arguments in the right order.
4772       SDValue Ops[] = {MemAddr, getAL(CurDAG, dl),
4773                        CurDAG->getRegister(0, MVT::i32), Chain};
4774       SDNode *Ld = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops);
4775       // Transfer memoperands.
4776       MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
4777       CurDAG->setNodeMemRefs(cast<MachineSDNode>(Ld), {MemOp});
4778 
4779       // Remap uses.
4780       SDValue OutChain = isThumb ? SDValue(Ld, 2) : SDValue(Ld, 1);
4781       if (!SDValue(N, 0).use_empty()) {
4782         SDValue Result;
4783         if (isThumb)
4784           Result = SDValue(Ld, 0);
4785         else {
4786           SDValue SubRegIdx =
4787             CurDAG->getTargetConstant(ARM::gsub_0, dl, MVT::i32);
4788           SDNode *ResNode = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
4789               dl, MVT::i32, SDValue(Ld, 0), SubRegIdx);
4790           Result = SDValue(ResNode,0);
4791         }
4792         ReplaceUses(SDValue(N, 0), Result);
4793       }
4794       if (!SDValue(N, 1).use_empty()) {
4795         SDValue Result;
4796         if (isThumb)
4797           Result = SDValue(Ld, 1);
4798         else {
4799           SDValue SubRegIdx =
4800             CurDAG->getTargetConstant(ARM::gsub_1, dl, MVT::i32);
4801           SDNode *ResNode = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
4802               dl, MVT::i32, SDValue(Ld, 0), SubRegIdx);
4803           Result = SDValue(ResNode,0);
4804         }
4805         ReplaceUses(SDValue(N, 1), Result);
4806       }
4807       ReplaceUses(SDValue(N, 2), OutChain);
4808       CurDAG->RemoveDeadNode(N);
4809       return;
4810     }
4811     case Intrinsic::arm_stlexd:
4812     case Intrinsic::arm_strexd: {
4813       SDLoc dl(N);
4814       SDValue Chain = N->getOperand(0);
4815       SDValue Val0 = N->getOperand(2);
4816       SDValue Val1 = N->getOperand(3);
4817       SDValue MemAddr = N->getOperand(4);
4818 
4819       // Store exclusive double return a i32 value which is the return status
4820       // of the issued store.
4821       const EVT ResTys[] = {MVT::i32, MVT::Other};
4822 
4823       bool isThumb = Subtarget->isThumb() && Subtarget->hasThumb2();
4824       // Place arguments in the right order.
4825       SmallVector<SDValue, 7> Ops;
4826       if (isThumb) {
4827         Ops.push_back(Val0);
4828         Ops.push_back(Val1);
4829       } else
4830         // arm_strexd uses GPRPair.
4831         Ops.push_back(SDValue(createGPRPairNode(MVT::Untyped, Val0, Val1), 0));
4832       Ops.push_back(MemAddr);
4833       Ops.push_back(getAL(CurDAG, dl));
4834       Ops.push_back(CurDAG->getRegister(0, MVT::i32));
4835       Ops.push_back(Chain);
4836 
4837       bool IsRelease = IntNo == Intrinsic::arm_stlexd;
4838       unsigned NewOpc = isThumb ? (IsRelease ? ARM::t2STLEXD : ARM::t2STREXD)
4839                                 : (IsRelease ? ARM::STLEXD : ARM::STREXD);
4840 
4841       SDNode *St = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops);
4842       // Transfer memoperands.
4843       MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(N)->getMemOperand();
4844       CurDAG->setNodeMemRefs(cast<MachineSDNode>(St), {MemOp});
4845 
4846       ReplaceNode(N, St);
4847       return;
4848     }
4849 
4850     case Intrinsic::arm_neon_vld1: {
4851       static const uint16_t DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16,
4852                                            ARM::VLD1d32, ARM::VLD1d64 };
4853       static const uint16_t QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
4854                                            ARM::VLD1q32, ARM::VLD1q64};
4855       SelectVLD(N, false, 1, DOpcodes, QOpcodes, nullptr);
4856       return;
4857     }
4858 
4859     case Intrinsic::arm_neon_vld1x2: {
4860       static const uint16_t DOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16,
4861                                            ARM::VLD1q32, ARM::VLD1q64 };
4862       static const uint16_t QOpcodes[] = { ARM::VLD1d8QPseudo,
4863                                            ARM::VLD1d16QPseudo,
4864                                            ARM::VLD1d32QPseudo,
4865                                            ARM::VLD1d64QPseudo };
4866       SelectVLD(N, false, 2, DOpcodes, QOpcodes, nullptr);
4867       return;
4868     }
4869 
4870     case Intrinsic::arm_neon_vld1x3: {
4871       static const uint16_t DOpcodes[] = { ARM::VLD1d8TPseudo,
4872                                            ARM::VLD1d16TPseudo,
4873                                            ARM::VLD1d32TPseudo,
4874                                            ARM::VLD1d64TPseudo };
4875       static const uint16_t QOpcodes0[] = { ARM::VLD1q8LowTPseudo_UPD,
4876                                             ARM::VLD1q16LowTPseudo_UPD,
4877                                             ARM::VLD1q32LowTPseudo_UPD,
4878                                             ARM::VLD1q64LowTPseudo_UPD };
4879       static const uint16_t QOpcodes1[] = { ARM::VLD1q8HighTPseudo,
4880                                             ARM::VLD1q16HighTPseudo,
4881                                             ARM::VLD1q32HighTPseudo,
4882                                             ARM::VLD1q64HighTPseudo };
4883       SelectVLD(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
4884       return;
4885     }
4886 
4887     case Intrinsic::arm_neon_vld1x4: {
4888       static const uint16_t DOpcodes[] = { ARM::VLD1d8QPseudo,
4889                                            ARM::VLD1d16QPseudo,
4890                                            ARM::VLD1d32QPseudo,
4891                                            ARM::VLD1d64QPseudo };
4892       static const uint16_t QOpcodes0[] = { ARM::VLD1q8LowQPseudo_UPD,
4893                                             ARM::VLD1q16LowQPseudo_UPD,
4894                                             ARM::VLD1q32LowQPseudo_UPD,
4895                                             ARM::VLD1q64LowQPseudo_UPD };
4896       static const uint16_t QOpcodes1[] = { ARM::VLD1q8HighQPseudo,
4897                                             ARM::VLD1q16HighQPseudo,
4898                                             ARM::VLD1q32HighQPseudo,
4899                                             ARM::VLD1q64HighQPseudo };
4900       SelectVLD(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
4901       return;
4902     }
4903 
4904     case Intrinsic::arm_neon_vld2: {
4905       static const uint16_t DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16,
4906                                            ARM::VLD2d32, ARM::VLD1q64 };
4907       static const uint16_t QOpcodes[] = { ARM::VLD2q8Pseudo, ARM::VLD2q16Pseudo,
4908                                            ARM::VLD2q32Pseudo };
4909       SelectVLD(N, false, 2, DOpcodes, QOpcodes, nullptr);
4910       return;
4911     }
4912 
4913     case Intrinsic::arm_neon_vld3: {
4914       static const uint16_t DOpcodes[] = { ARM::VLD3d8Pseudo,
4915                                            ARM::VLD3d16Pseudo,
4916                                            ARM::VLD3d32Pseudo,
4917                                            ARM::VLD1d64TPseudo };
4918       static const uint16_t QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD,
4919                                             ARM::VLD3q16Pseudo_UPD,
4920                                             ARM::VLD3q32Pseudo_UPD };
4921       static const uint16_t QOpcodes1[] = { ARM::VLD3q8oddPseudo,
4922                                             ARM::VLD3q16oddPseudo,
4923                                             ARM::VLD3q32oddPseudo };
4924       SelectVLD(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
4925       return;
4926     }
4927 
4928     case Intrinsic::arm_neon_vld4: {
4929       static const uint16_t DOpcodes[] = { ARM::VLD4d8Pseudo,
4930                                            ARM::VLD4d16Pseudo,
4931                                            ARM::VLD4d32Pseudo,
4932                                            ARM::VLD1d64QPseudo };
4933       static const uint16_t QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD,
4934                                             ARM::VLD4q16Pseudo_UPD,
4935                                             ARM::VLD4q32Pseudo_UPD };
4936       static const uint16_t QOpcodes1[] = { ARM::VLD4q8oddPseudo,
4937                                             ARM::VLD4q16oddPseudo,
4938                                             ARM::VLD4q32oddPseudo };
4939       SelectVLD(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
4940       return;
4941     }
4942 
4943     case Intrinsic::arm_neon_vld2dup: {
4944       static const uint16_t DOpcodes[] = { ARM::VLD2DUPd8, ARM::VLD2DUPd16,
4945                                            ARM::VLD2DUPd32, ARM::VLD1q64 };
4946       static const uint16_t QOpcodes0[] = { ARM::VLD2DUPq8EvenPseudo,
4947                                             ARM::VLD2DUPq16EvenPseudo,
4948                                             ARM::VLD2DUPq32EvenPseudo };
4949       static const uint16_t QOpcodes1[] = { ARM::VLD2DUPq8OddPseudo,
4950                                             ARM::VLD2DUPq16OddPseudo,
4951                                             ARM::VLD2DUPq32OddPseudo };
4952       SelectVLDDup(N, /* IsIntrinsic= */ true, false, 2,
4953                    DOpcodes, QOpcodes0, QOpcodes1);
4954       return;
4955     }
4956 
4957     case Intrinsic::arm_neon_vld3dup: {
4958       static const uint16_t DOpcodes[] = { ARM::VLD3DUPd8Pseudo,
4959                                            ARM::VLD3DUPd16Pseudo,
4960                                            ARM::VLD3DUPd32Pseudo,
4961                                            ARM::VLD1d64TPseudo };
4962       static const uint16_t QOpcodes0[] = { ARM::VLD3DUPq8EvenPseudo,
4963                                             ARM::VLD3DUPq16EvenPseudo,
4964                                             ARM::VLD3DUPq32EvenPseudo };
4965       static const uint16_t QOpcodes1[] = { ARM::VLD3DUPq8OddPseudo,
4966                                             ARM::VLD3DUPq16OddPseudo,
4967                                             ARM::VLD3DUPq32OddPseudo };
4968       SelectVLDDup(N, /* IsIntrinsic= */ true, false, 3,
4969                    DOpcodes, QOpcodes0, QOpcodes1);
4970       return;
4971     }
4972 
4973     case Intrinsic::arm_neon_vld4dup: {
4974       static const uint16_t DOpcodes[] = { ARM::VLD4DUPd8Pseudo,
4975                                            ARM::VLD4DUPd16Pseudo,
4976                                            ARM::VLD4DUPd32Pseudo,
4977                                            ARM::VLD1d64QPseudo };
4978       static const uint16_t QOpcodes0[] = { ARM::VLD4DUPq8EvenPseudo,
4979                                             ARM::VLD4DUPq16EvenPseudo,
4980                                             ARM::VLD4DUPq32EvenPseudo };
4981       static const uint16_t QOpcodes1[] = { ARM::VLD4DUPq8OddPseudo,
4982                                             ARM::VLD4DUPq16OddPseudo,
4983                                             ARM::VLD4DUPq32OddPseudo };
4984       SelectVLDDup(N, /* IsIntrinsic= */ true, false, 4,
4985                    DOpcodes, QOpcodes0, QOpcodes1);
4986       return;
4987     }
4988 
4989     case Intrinsic::arm_neon_vld2lane: {
4990       static const uint16_t DOpcodes[] = { ARM::VLD2LNd8Pseudo,
4991                                            ARM::VLD2LNd16Pseudo,
4992                                            ARM::VLD2LNd32Pseudo };
4993       static const uint16_t QOpcodes[] = { ARM::VLD2LNq16Pseudo,
4994                                            ARM::VLD2LNq32Pseudo };
4995       SelectVLDSTLane(N, true, false, 2, DOpcodes, QOpcodes);
4996       return;
4997     }
4998 
4999     case Intrinsic::arm_neon_vld3lane: {
5000       static const uint16_t DOpcodes[] = { ARM::VLD3LNd8Pseudo,
5001                                            ARM::VLD3LNd16Pseudo,
5002                                            ARM::VLD3LNd32Pseudo };
5003       static const uint16_t QOpcodes[] = { ARM::VLD3LNq16Pseudo,
5004                                            ARM::VLD3LNq32Pseudo };
5005       SelectVLDSTLane(N, true, false, 3, DOpcodes, QOpcodes);
5006       return;
5007     }
5008 
5009     case Intrinsic::arm_neon_vld4lane: {
5010       static const uint16_t DOpcodes[] = { ARM::VLD4LNd8Pseudo,
5011                                            ARM::VLD4LNd16Pseudo,
5012                                            ARM::VLD4LNd32Pseudo };
5013       static const uint16_t QOpcodes[] = { ARM::VLD4LNq16Pseudo,
5014                                            ARM::VLD4LNq32Pseudo };
5015       SelectVLDSTLane(N, true, false, 4, DOpcodes, QOpcodes);
5016       return;
5017     }
5018 
5019     case Intrinsic::arm_neon_vst1: {
5020       static const uint16_t DOpcodes[] = { ARM::VST1d8, ARM::VST1d16,
5021                                            ARM::VST1d32, ARM::VST1d64 };
5022       static const uint16_t QOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
5023                                            ARM::VST1q32, ARM::VST1q64 };
5024       SelectVST(N, false, 1, DOpcodes, QOpcodes, nullptr);
5025       return;
5026     }
5027 
5028     case Intrinsic::arm_neon_vst1x2: {
5029       static const uint16_t DOpcodes[] = { ARM::VST1q8, ARM::VST1q16,
5030                                            ARM::VST1q32, ARM::VST1q64 };
5031       static const uint16_t QOpcodes[] = { ARM::VST1d8QPseudo,
5032                                            ARM::VST1d16QPseudo,
5033                                            ARM::VST1d32QPseudo,
5034                                            ARM::VST1d64QPseudo };
5035       SelectVST(N, false, 2, DOpcodes, QOpcodes, nullptr);
5036       return;
5037     }
5038 
5039     case Intrinsic::arm_neon_vst1x3: {
5040       static const uint16_t DOpcodes[] = { ARM::VST1d8TPseudo,
5041                                            ARM::VST1d16TPseudo,
5042                                            ARM::VST1d32TPseudo,
5043                                            ARM::VST1d64TPseudo };
5044       static const uint16_t QOpcodes0[] = { ARM::VST1q8LowTPseudo_UPD,
5045                                             ARM::VST1q16LowTPseudo_UPD,
5046                                             ARM::VST1q32LowTPseudo_UPD,
5047                                             ARM::VST1q64LowTPseudo_UPD };
5048       static const uint16_t QOpcodes1[] = { ARM::VST1q8HighTPseudo,
5049                                             ARM::VST1q16HighTPseudo,
5050                                             ARM::VST1q32HighTPseudo,
5051                                             ARM::VST1q64HighTPseudo };
5052       SelectVST(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
5053       return;
5054     }
5055 
5056     case Intrinsic::arm_neon_vst1x4: {
5057       static const uint16_t DOpcodes[] = { ARM::VST1d8QPseudo,
5058                                            ARM::VST1d16QPseudo,
5059                                            ARM::VST1d32QPseudo,
5060                                            ARM::VST1d64QPseudo };
5061       static const uint16_t QOpcodes0[] = { ARM::VST1q8LowQPseudo_UPD,
5062                                             ARM::VST1q16LowQPseudo_UPD,
5063                                             ARM::VST1q32LowQPseudo_UPD,
5064                                             ARM::VST1q64LowQPseudo_UPD };
5065       static const uint16_t QOpcodes1[] = { ARM::VST1q8HighQPseudo,
5066                                             ARM::VST1q16HighQPseudo,
5067                                             ARM::VST1q32HighQPseudo,
5068                                             ARM::VST1q64HighQPseudo };
5069       SelectVST(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
5070       return;
5071     }
5072 
5073     case Intrinsic::arm_neon_vst2: {
5074       static const uint16_t DOpcodes[] = { ARM::VST2d8, ARM::VST2d16,
5075                                            ARM::VST2d32, ARM::VST1q64 };
5076       static const uint16_t QOpcodes[] = { ARM::VST2q8Pseudo, ARM::VST2q16Pseudo,
5077                                            ARM::VST2q32Pseudo };
5078       SelectVST(N, false, 2, DOpcodes, QOpcodes, nullptr);
5079       return;
5080     }
5081 
5082     case Intrinsic::arm_neon_vst3: {
5083       static const uint16_t DOpcodes[] = { ARM::VST3d8Pseudo,
5084                                            ARM::VST3d16Pseudo,
5085                                            ARM::VST3d32Pseudo,
5086                                            ARM::VST1d64TPseudo };
5087       static const uint16_t QOpcodes0[] = { ARM::VST3q8Pseudo_UPD,
5088                                             ARM::VST3q16Pseudo_UPD,
5089                                             ARM::VST3q32Pseudo_UPD };
5090       static const uint16_t QOpcodes1[] = { ARM::VST3q8oddPseudo,
5091                                             ARM::VST3q16oddPseudo,
5092                                             ARM::VST3q32oddPseudo };
5093       SelectVST(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1);
5094       return;
5095     }
5096 
5097     case Intrinsic::arm_neon_vst4: {
5098       static const uint16_t DOpcodes[] = { ARM::VST4d8Pseudo,
5099                                            ARM::VST4d16Pseudo,
5100                                            ARM::VST4d32Pseudo,
5101                                            ARM::VST1d64QPseudo };
5102       static const uint16_t QOpcodes0[] = { ARM::VST4q8Pseudo_UPD,
5103                                             ARM::VST4q16Pseudo_UPD,
5104                                             ARM::VST4q32Pseudo_UPD };
5105       static const uint16_t QOpcodes1[] = { ARM::VST4q8oddPseudo,
5106                                             ARM::VST4q16oddPseudo,
5107                                             ARM::VST4q32oddPseudo };
5108       SelectVST(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1);
5109       return;
5110     }
5111 
5112     case Intrinsic::arm_neon_vst2lane: {
5113       static const uint16_t DOpcodes[] = { ARM::VST2LNd8Pseudo,
5114                                            ARM::VST2LNd16Pseudo,
5115                                            ARM::VST2LNd32Pseudo };
5116       static const uint16_t QOpcodes[] = { ARM::VST2LNq16Pseudo,
5117                                            ARM::VST2LNq32Pseudo };
5118       SelectVLDSTLane(N, false, false, 2, DOpcodes, QOpcodes);
5119       return;
5120     }
5121 
5122     case Intrinsic::arm_neon_vst3lane: {
5123       static const uint16_t DOpcodes[] = { ARM::VST3LNd8Pseudo,
5124                                            ARM::VST3LNd16Pseudo,
5125                                            ARM::VST3LNd32Pseudo };
5126       static const uint16_t QOpcodes[] = { ARM::VST3LNq16Pseudo,
5127                                            ARM::VST3LNq32Pseudo };
5128       SelectVLDSTLane(N, false, false, 3, DOpcodes, QOpcodes);
5129       return;
5130     }
5131 
5132     case Intrinsic::arm_neon_vst4lane: {
5133       static const uint16_t DOpcodes[] = { ARM::VST4LNd8Pseudo,
5134                                            ARM::VST4LNd16Pseudo,
5135                                            ARM::VST4LNd32Pseudo };
5136       static const uint16_t QOpcodes[] = { ARM::VST4LNq16Pseudo,
5137                                            ARM::VST4LNq32Pseudo };
5138       SelectVLDSTLane(N, false, false, 4, DOpcodes, QOpcodes);
5139       return;
5140     }
5141 
5142     case Intrinsic::arm_mve_vldr_gather_base_wb:
5143     case Intrinsic::arm_mve_vldr_gather_base_wb_predicated: {
5144       static const uint16_t Opcodes[] = {ARM::MVE_VLDRWU32_qi_pre,
5145                                          ARM::MVE_VLDRDU64_qi_pre};
5146       SelectMVE_WB(N, Opcodes,
5147                    IntNo == Intrinsic::arm_mve_vldr_gather_base_wb_predicated);
5148       return;
5149     }
5150 
5151     case Intrinsic::arm_mve_vld2q: {
5152       static const uint16_t Opcodes8[] = {ARM::MVE_VLD20_8, ARM::MVE_VLD21_8};
5153       static const uint16_t Opcodes16[] = {ARM::MVE_VLD20_16,
5154                                            ARM::MVE_VLD21_16};
5155       static const uint16_t Opcodes32[] = {ARM::MVE_VLD20_32,
5156                                            ARM::MVE_VLD21_32};
5157       static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32};
5158       SelectMVE_VLD(N, 2, Opcodes, false);
5159       return;
5160     }
5161 
5162     case Intrinsic::arm_mve_vld4q: {
5163       static const uint16_t Opcodes8[] = {ARM::MVE_VLD40_8, ARM::MVE_VLD41_8,
5164                                           ARM::MVE_VLD42_8, ARM::MVE_VLD43_8};
5165       static const uint16_t Opcodes16[] = {ARM::MVE_VLD40_16, ARM::MVE_VLD41_16,
5166                                            ARM::MVE_VLD42_16,
5167                                            ARM::MVE_VLD43_16};
5168       static const uint16_t Opcodes32[] = {ARM::MVE_VLD40_32, ARM::MVE_VLD41_32,
5169                                            ARM::MVE_VLD42_32,
5170                                            ARM::MVE_VLD43_32};
5171       static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32};
5172       SelectMVE_VLD(N, 4, Opcodes, false);
5173       return;
5174     }
5175     }
5176     break;
5177   }
5178 
5179   case ISD::INTRINSIC_WO_CHAIN: {
5180     unsigned IntNo = N->getConstantOperandVal(0);
5181     switch (IntNo) {
5182     default:
5183       break;
5184 
5185     // Scalar f32 -> bf16
5186     case Intrinsic::arm_neon_vcvtbfp2bf: {
5187       SDLoc dl(N);
5188       const SDValue &Src = N->getOperand(1);
5189       llvm::EVT DestTy = N->getValueType(0);
5190       SDValue Pred = getAL(CurDAG, dl);
5191       SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
5192       SDValue Ops[] = { Src, Src, Pred, Reg0 };
5193       CurDAG->SelectNodeTo(N, ARM::BF16_VCVTB, DestTy, Ops);
5194       return;
5195     }
5196 
5197     // Vector v4f32 -> v4bf16
5198     case Intrinsic::arm_neon_vcvtfp2bf: {
5199       SDLoc dl(N);
5200       const SDValue &Src = N->getOperand(1);
5201       SDValue Pred = getAL(CurDAG, dl);
5202       SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
5203       SDValue Ops[] = { Src, Pred, Reg0 };
5204       CurDAG->SelectNodeTo(N, ARM::BF16_VCVT, MVT::v4bf16, Ops);
5205       return;
5206     }
5207 
5208     case Intrinsic::arm_mve_urshrl:
5209       SelectMVE_LongShift(N, ARM::MVE_URSHRL, true, false);
5210       return;
5211     case Intrinsic::arm_mve_uqshll:
5212       SelectMVE_LongShift(N, ARM::MVE_UQSHLL, true, false);
5213       return;
5214     case Intrinsic::arm_mve_srshrl:
5215       SelectMVE_LongShift(N, ARM::MVE_SRSHRL, true, false);
5216       return;
5217     case Intrinsic::arm_mve_sqshll:
5218       SelectMVE_LongShift(N, ARM::MVE_SQSHLL, true, false);
5219       return;
5220     case Intrinsic::arm_mve_uqrshll:
5221       SelectMVE_LongShift(N, ARM::MVE_UQRSHLL, false, true);
5222       return;
5223     case Intrinsic::arm_mve_sqrshrl:
5224       SelectMVE_LongShift(N, ARM::MVE_SQRSHRL, false, true);
5225       return;
5226 
5227     case Intrinsic::arm_mve_vadc:
5228     case Intrinsic::arm_mve_vadc_predicated:
5229       SelectMVE_VADCSBC(N, ARM::MVE_VADC, ARM::MVE_VADCI, true,
5230                         IntNo == Intrinsic::arm_mve_vadc_predicated);
5231       return;
5232     case Intrinsic::arm_mve_vsbc:
5233     case Intrinsic::arm_mve_vsbc_predicated:
5234       SelectMVE_VADCSBC(N, ARM::MVE_VSBC, ARM::MVE_VSBCI, true,
5235                         IntNo == Intrinsic::arm_mve_vsbc_predicated);
5236       return;
5237     case Intrinsic::arm_mve_vshlc:
5238     case Intrinsic::arm_mve_vshlc_predicated:
5239       SelectMVE_VSHLC(N, IntNo == Intrinsic::arm_mve_vshlc_predicated);
5240       return;
5241 
5242     case Intrinsic::arm_mve_vmlldava:
5243     case Intrinsic::arm_mve_vmlldava_predicated: {
5244       static const uint16_t OpcodesU[] = {
5245           ARM::MVE_VMLALDAVu16,   ARM::MVE_VMLALDAVu32,
5246           ARM::MVE_VMLALDAVau16,  ARM::MVE_VMLALDAVau32,
5247       };
5248       static const uint16_t OpcodesS[] = {
5249           ARM::MVE_VMLALDAVs16,   ARM::MVE_VMLALDAVs32,
5250           ARM::MVE_VMLALDAVas16,  ARM::MVE_VMLALDAVas32,
5251           ARM::MVE_VMLALDAVxs16,  ARM::MVE_VMLALDAVxs32,
5252           ARM::MVE_VMLALDAVaxs16, ARM::MVE_VMLALDAVaxs32,
5253           ARM::MVE_VMLSLDAVs16,   ARM::MVE_VMLSLDAVs32,
5254           ARM::MVE_VMLSLDAVas16,  ARM::MVE_VMLSLDAVas32,
5255           ARM::MVE_VMLSLDAVxs16,  ARM::MVE_VMLSLDAVxs32,
5256           ARM::MVE_VMLSLDAVaxs16, ARM::MVE_VMLSLDAVaxs32,
5257       };
5258       SelectMVE_VMLLDAV(N, IntNo == Intrinsic::arm_mve_vmlldava_predicated,
5259                         OpcodesS, OpcodesU);
5260       return;
5261     }
5262 
5263     case Intrinsic::arm_mve_vrmlldavha:
5264     case Intrinsic::arm_mve_vrmlldavha_predicated: {
5265       static const uint16_t OpcodesU[] = {
5266           ARM::MVE_VRMLALDAVHu32,  ARM::MVE_VRMLALDAVHau32,
5267       };
5268       static const uint16_t OpcodesS[] = {
5269           ARM::MVE_VRMLALDAVHs32,  ARM::MVE_VRMLALDAVHas32,
5270           ARM::MVE_VRMLALDAVHxs32, ARM::MVE_VRMLALDAVHaxs32,
5271           ARM::MVE_VRMLSLDAVHs32,  ARM::MVE_VRMLSLDAVHas32,
5272           ARM::MVE_VRMLSLDAVHxs32, ARM::MVE_VRMLSLDAVHaxs32,
5273       };
5274       SelectMVE_VRMLLDAVH(N, IntNo == Intrinsic::arm_mve_vrmlldavha_predicated,
5275                           OpcodesS, OpcodesU);
5276       return;
5277     }
5278 
5279     case Intrinsic::arm_mve_vidup:
5280     case Intrinsic::arm_mve_vidup_predicated: {
5281       static const uint16_t Opcodes[] = {
5282           ARM::MVE_VIDUPu8, ARM::MVE_VIDUPu16, ARM::MVE_VIDUPu32,
5283       };
5284       SelectMVE_VxDUP(N, Opcodes, false,
5285                       IntNo == Intrinsic::arm_mve_vidup_predicated);
5286       return;
5287     }
5288 
5289     case Intrinsic::arm_mve_vddup:
5290     case Intrinsic::arm_mve_vddup_predicated: {
5291       static const uint16_t Opcodes[] = {
5292           ARM::MVE_VDDUPu8, ARM::MVE_VDDUPu16, ARM::MVE_VDDUPu32,
5293       };
5294       SelectMVE_VxDUP(N, Opcodes, false,
5295                       IntNo == Intrinsic::arm_mve_vddup_predicated);
5296       return;
5297     }
5298 
5299     case Intrinsic::arm_mve_viwdup:
5300     case Intrinsic::arm_mve_viwdup_predicated: {
5301       static const uint16_t Opcodes[] = {
5302           ARM::MVE_VIWDUPu8, ARM::MVE_VIWDUPu16, ARM::MVE_VIWDUPu32,
5303       };
5304       SelectMVE_VxDUP(N, Opcodes, true,
5305                       IntNo == Intrinsic::arm_mve_viwdup_predicated);
5306       return;
5307     }
5308 
5309     case Intrinsic::arm_mve_vdwdup:
5310     case Intrinsic::arm_mve_vdwdup_predicated: {
5311       static const uint16_t Opcodes[] = {
5312           ARM::MVE_VDWDUPu8, ARM::MVE_VDWDUPu16, ARM::MVE_VDWDUPu32,
5313       };
5314       SelectMVE_VxDUP(N, Opcodes, true,
5315                       IntNo == Intrinsic::arm_mve_vdwdup_predicated);
5316       return;
5317     }
5318 
5319     case Intrinsic::arm_cde_cx1d:
5320     case Intrinsic::arm_cde_cx1da:
5321     case Intrinsic::arm_cde_cx2d:
5322     case Intrinsic::arm_cde_cx2da:
5323     case Intrinsic::arm_cde_cx3d:
5324     case Intrinsic::arm_cde_cx3da: {
5325       bool HasAccum = IntNo == Intrinsic::arm_cde_cx1da ||
5326                       IntNo == Intrinsic::arm_cde_cx2da ||
5327                       IntNo == Intrinsic::arm_cde_cx3da;
5328       size_t NumExtraOps;
5329       uint16_t Opcode;
5330       switch (IntNo) {
5331       case Intrinsic::arm_cde_cx1d:
5332       case Intrinsic::arm_cde_cx1da:
5333         NumExtraOps = 0;
5334         Opcode = HasAccum ? ARM::CDE_CX1DA : ARM::CDE_CX1D;
5335         break;
5336       case Intrinsic::arm_cde_cx2d:
5337       case Intrinsic::arm_cde_cx2da:
5338         NumExtraOps = 1;
5339         Opcode = HasAccum ? ARM::CDE_CX2DA : ARM::CDE_CX2D;
5340         break;
5341       case Intrinsic::arm_cde_cx3d:
5342       case Intrinsic::arm_cde_cx3da:
5343         NumExtraOps = 2;
5344         Opcode = HasAccum ? ARM::CDE_CX3DA : ARM::CDE_CX3D;
5345         break;
5346       default:
5347         llvm_unreachable("Unexpected opcode");
5348       }
5349       SelectCDE_CXxD(N, Opcode, NumExtraOps, HasAccum);
5350       return;
5351     }
5352     }
5353     break;
5354   }
5355 
5356   case ISD::ATOMIC_CMP_SWAP:
5357     SelectCMP_SWAP(N);
5358     return;
5359   }
5360 
5361   SelectCode(N);
5362 }
5363 
5364 // Inspect a register string of the form
5365 // cp<coprocessor>:<opc1>:c<CRn>:c<CRm>:<opc2> (32bit) or
5366 // cp<coprocessor>:<opc1>:c<CRm> (64bit) inspect the fields of the string
5367 // and obtain the integer operands from them, adding these operands to the
5368 // provided vector.
5369 static void getIntOperandsFromRegisterString(StringRef RegString,
5370                                              SelectionDAG *CurDAG,
5371                                              const SDLoc &DL,
5372                                              std::vector<SDValue> &Ops) {
5373   SmallVector<StringRef, 5> Fields;
5374   RegString.split(Fields, ':');
5375 
5376   if (Fields.size() > 1) {
5377     bool AllIntFields = true;
5378 
5379     for (StringRef Field : Fields) {
5380       // Need to trim out leading 'cp' characters and get the integer field.
5381       unsigned IntField;
5382       AllIntFields &= !Field.trim("CPcp").getAsInteger(10, IntField);
5383       Ops.push_back(CurDAG->getTargetConstant(IntField, DL, MVT::i32));
5384     }
5385 
5386     assert(AllIntFields &&
5387             "Unexpected non-integer value in special register string.");
5388     (void)AllIntFields;
5389   }
5390 }
5391 
5392 // Maps a Banked Register string to its mask value. The mask value returned is
5393 // for use in the MRSbanked / MSRbanked instruction nodes as the Banked Register
5394 // mask operand, which expresses which register is to be used, e.g. r8, and in
5395 // which mode it is to be used, e.g. usr. Returns -1 to signify that the string
5396 // was invalid.
5397 static inline int getBankedRegisterMask(StringRef RegString) {
5398   auto TheReg = ARMBankedReg::lookupBankedRegByName(RegString.lower());
5399   if (!TheReg)
5400      return -1;
5401   return TheReg->Encoding;
5402 }
5403 
5404 // The flags here are common to those allowed for apsr in the A class cores and
5405 // those allowed for the special registers in the M class cores. Returns a
5406 // value representing which flags were present, -1 if invalid.
5407 static inline int getMClassFlagsMask(StringRef Flags) {
5408   return StringSwitch<int>(Flags)
5409           .Case("", 0x2) // no flags means nzcvq for psr registers, and 0x2 is
5410                          // correct when flags are not permitted
5411           .Case("g", 0x1)
5412           .Case("nzcvq", 0x2)
5413           .Case("nzcvqg", 0x3)
5414           .Default(-1);
5415 }
5416 
5417 // Maps MClass special registers string to its value for use in the
5418 // t2MRS_M/t2MSR_M instruction nodes as the SYSm value operand.
5419 // Returns -1 to signify that the string was invalid.
5420 static int getMClassRegisterMask(StringRef Reg, const ARMSubtarget *Subtarget) {
5421   auto TheReg = ARMSysReg::lookupMClassSysRegByName(Reg);
5422   const FeatureBitset &FeatureBits = Subtarget->getFeatureBits();
5423   if (!TheReg || !TheReg->hasRequiredFeatures(FeatureBits))
5424     return -1;
5425   return (int)(TheReg->Encoding & 0xFFF); // SYSm value
5426 }
5427 
5428 static int getARClassRegisterMask(StringRef Reg, StringRef Flags) {
5429   // The mask operand contains the special register (R Bit) in bit 4, whether
5430   // the register is spsr (R bit is 1) or one of cpsr/apsr (R bit is 0), and
5431   // bits 3-0 contains the fields to be accessed in the special register, set by
5432   // the flags provided with the register.
5433   int Mask = 0;
5434   if (Reg == "apsr") {
5435     // The flags permitted for apsr are the same flags that are allowed in
5436     // M class registers. We get the flag value and then shift the flags into
5437     // the correct place to combine with the mask.
5438     Mask = getMClassFlagsMask(Flags);
5439     if (Mask == -1)
5440       return -1;
5441     return Mask << 2;
5442   }
5443 
5444   if (Reg != "cpsr" && Reg != "spsr") {
5445     return -1;
5446   }
5447 
5448   // This is the same as if the flags were "fc"
5449   if (Flags.empty() || Flags == "all")
5450     return Mask | 0x9;
5451 
5452   // Inspect the supplied flags string and set the bits in the mask for
5453   // the relevant and valid flags allowed for cpsr and spsr.
5454   for (char Flag : Flags) {
5455     int FlagVal;
5456     switch (Flag) {
5457       case 'c':
5458         FlagVal = 0x1;
5459         break;
5460       case 'x':
5461         FlagVal = 0x2;
5462         break;
5463       case 's':
5464         FlagVal = 0x4;
5465         break;
5466       case 'f':
5467         FlagVal = 0x8;
5468         break;
5469       default:
5470         FlagVal = 0;
5471     }
5472 
5473     // This avoids allowing strings where the same flag bit appears twice.
5474     if (!FlagVal || (Mask & FlagVal))
5475       return -1;
5476     Mask |= FlagVal;
5477   }
5478 
5479   // If the register is spsr then we need to set the R bit.
5480   if (Reg == "spsr")
5481     Mask |= 0x10;
5482 
5483   return Mask;
5484 }
5485 
5486 // Lower the read_register intrinsic to ARM specific DAG nodes
5487 // using the supplied metadata string to select the instruction node to use
5488 // and the registers/masks to construct as operands for the node.
5489 bool ARMDAGToDAGISel::tryReadRegister(SDNode *N){
5490   const auto *MD = cast<MDNodeSDNode>(N->getOperand(1));
5491   const auto *RegString = cast<MDString>(MD->getMD()->getOperand(0));
5492   bool IsThumb2 = Subtarget->isThumb2();
5493   SDLoc DL(N);
5494 
5495   std::vector<SDValue> Ops;
5496   getIntOperandsFromRegisterString(RegString->getString(), CurDAG, DL, Ops);
5497 
5498   if (!Ops.empty()) {
5499     // If the special register string was constructed of fields (as defined
5500     // in the ACLE) then need to lower to MRC node (32 bit) or
5501     // MRRC node(64 bit), we can make the distinction based on the number of
5502     // operands we have.
5503     unsigned Opcode;
5504     SmallVector<EVT, 3> ResTypes;
5505     if (Ops.size() == 5){
5506       Opcode = IsThumb2 ? ARM::t2MRC : ARM::MRC;
5507       ResTypes.append({ MVT::i32, MVT::Other });
5508     } else {
5509       assert(Ops.size() == 3 &&
5510               "Invalid number of fields in special register string.");
5511       Opcode = IsThumb2 ? ARM::t2MRRC : ARM::MRRC;
5512       ResTypes.append({ MVT::i32, MVT::i32, MVT::Other });
5513     }
5514 
5515     Ops.push_back(getAL(CurDAG, DL));
5516     Ops.push_back(CurDAG->getRegister(0, MVT::i32));
5517     Ops.push_back(N->getOperand(0));
5518     ReplaceNode(N, CurDAG->getMachineNode(Opcode, DL, ResTypes, Ops));
5519     return true;
5520   }
5521 
5522   std::string SpecialReg = RegString->getString().lower();
5523 
5524   int BankedReg = getBankedRegisterMask(SpecialReg);
5525   if (BankedReg != -1) {
5526     Ops = { CurDAG->getTargetConstant(BankedReg, DL, MVT::i32),
5527             getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5528             N->getOperand(0) };
5529     ReplaceNode(
5530         N, CurDAG->getMachineNode(IsThumb2 ? ARM::t2MRSbanked : ARM::MRSbanked,
5531                                   DL, MVT::i32, MVT::Other, Ops));
5532     return true;
5533   }
5534 
5535   // The VFP registers are read by creating SelectionDAG nodes with opcodes
5536   // corresponding to the register that is being read from. So we switch on the
5537   // string to find which opcode we need to use.
5538   unsigned Opcode = StringSwitch<unsigned>(SpecialReg)
5539                     .Case("fpscr", ARM::VMRS)
5540                     .Case("fpexc", ARM::VMRS_FPEXC)
5541                     .Case("fpsid", ARM::VMRS_FPSID)
5542                     .Case("mvfr0", ARM::VMRS_MVFR0)
5543                     .Case("mvfr1", ARM::VMRS_MVFR1)
5544                     .Case("mvfr2", ARM::VMRS_MVFR2)
5545                     .Case("fpinst", ARM::VMRS_FPINST)
5546                     .Case("fpinst2", ARM::VMRS_FPINST2)
5547                     .Default(0);
5548 
5549   // If an opcode was found then we can lower the read to a VFP instruction.
5550   if (Opcode) {
5551     if (!Subtarget->hasVFP2Base())
5552       return false;
5553     if (Opcode == ARM::VMRS_MVFR2 && !Subtarget->hasFPARMv8Base())
5554       return false;
5555 
5556     Ops = { getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5557             N->getOperand(0) };
5558     ReplaceNode(N,
5559                 CurDAG->getMachineNode(Opcode, DL, MVT::i32, MVT::Other, Ops));
5560     return true;
5561   }
5562 
5563   // If the target is M Class then need to validate that the register string
5564   // is an acceptable value, so check that a mask can be constructed from the
5565   // string.
5566   if (Subtarget->isMClass()) {
5567     int SYSmValue = getMClassRegisterMask(SpecialReg, Subtarget);
5568     if (SYSmValue == -1)
5569       return false;
5570 
5571     SDValue Ops[] = { CurDAG->getTargetConstant(SYSmValue, DL, MVT::i32),
5572                       getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5573                       N->getOperand(0) };
5574     ReplaceNode(
5575         N, CurDAG->getMachineNode(ARM::t2MRS_M, DL, MVT::i32, MVT::Other, Ops));
5576     return true;
5577   }
5578 
5579   // Here we know the target is not M Class so we need to check if it is one
5580   // of the remaining possible values which are apsr, cpsr or spsr.
5581   if (SpecialReg == "apsr" || SpecialReg == "cpsr") {
5582     Ops = { getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5583             N->getOperand(0) };
5584     ReplaceNode(N, CurDAG->getMachineNode(IsThumb2 ? ARM::t2MRS_AR : ARM::MRS,
5585                                           DL, MVT::i32, MVT::Other, Ops));
5586     return true;
5587   }
5588 
5589   if (SpecialReg == "spsr") {
5590     Ops = { getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5591             N->getOperand(0) };
5592     ReplaceNode(
5593         N, CurDAG->getMachineNode(IsThumb2 ? ARM::t2MRSsys_AR : ARM::MRSsys, DL,
5594                                   MVT::i32, MVT::Other, Ops));
5595     return true;
5596   }
5597 
5598   return false;
5599 }
5600 
5601 // Lower the write_register intrinsic to ARM specific DAG nodes
5602 // using the supplied metadata string to select the instruction node to use
5603 // and the registers/masks to use in the nodes
5604 bool ARMDAGToDAGISel::tryWriteRegister(SDNode *N){
5605   const auto *MD = cast<MDNodeSDNode>(N->getOperand(1));
5606   const auto *RegString = cast<MDString>(MD->getMD()->getOperand(0));
5607   bool IsThumb2 = Subtarget->isThumb2();
5608   SDLoc DL(N);
5609 
5610   std::vector<SDValue> Ops;
5611   getIntOperandsFromRegisterString(RegString->getString(), CurDAG, DL, Ops);
5612 
5613   if (!Ops.empty()) {
5614     // If the special register string was constructed of fields (as defined
5615     // in the ACLE) then need to lower to MCR node (32 bit) or
5616     // MCRR node(64 bit), we can make the distinction based on the number of
5617     // operands we have.
5618     unsigned Opcode;
5619     if (Ops.size() == 5) {
5620       Opcode = IsThumb2 ? ARM::t2MCR : ARM::MCR;
5621       Ops.insert(Ops.begin()+2, N->getOperand(2));
5622     } else {
5623       assert(Ops.size() == 3 &&
5624               "Invalid number of fields in special register string.");
5625       Opcode = IsThumb2 ? ARM::t2MCRR : ARM::MCRR;
5626       SDValue WriteValue[] = { N->getOperand(2), N->getOperand(3) };
5627       Ops.insert(Ops.begin()+2, WriteValue, WriteValue+2);
5628     }
5629 
5630     Ops.push_back(getAL(CurDAG, DL));
5631     Ops.push_back(CurDAG->getRegister(0, MVT::i32));
5632     Ops.push_back(N->getOperand(0));
5633 
5634     ReplaceNode(N, CurDAG->getMachineNode(Opcode, DL, MVT::Other, Ops));
5635     return true;
5636   }
5637 
5638   std::string SpecialReg = RegString->getString().lower();
5639   int BankedReg = getBankedRegisterMask(SpecialReg);
5640   if (BankedReg != -1) {
5641     Ops = { CurDAG->getTargetConstant(BankedReg, DL, MVT::i32), N->getOperand(2),
5642             getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5643             N->getOperand(0) };
5644     ReplaceNode(
5645         N, CurDAG->getMachineNode(IsThumb2 ? ARM::t2MSRbanked : ARM::MSRbanked,
5646                                   DL, MVT::Other, Ops));
5647     return true;
5648   }
5649 
5650   // The VFP registers are written to by creating SelectionDAG nodes with
5651   // opcodes corresponding to the register that is being written. So we switch
5652   // on the string to find which opcode we need to use.
5653   unsigned Opcode = StringSwitch<unsigned>(SpecialReg)
5654                     .Case("fpscr", ARM::VMSR)
5655                     .Case("fpexc", ARM::VMSR_FPEXC)
5656                     .Case("fpsid", ARM::VMSR_FPSID)
5657                     .Case("fpinst", ARM::VMSR_FPINST)
5658                     .Case("fpinst2", ARM::VMSR_FPINST2)
5659                     .Default(0);
5660 
5661   if (Opcode) {
5662     if (!Subtarget->hasVFP2Base())
5663       return false;
5664     Ops = { N->getOperand(2), getAL(CurDAG, DL),
5665             CurDAG->getRegister(0, MVT::i32), N->getOperand(0) };
5666     ReplaceNode(N, CurDAG->getMachineNode(Opcode, DL, MVT::Other, Ops));
5667     return true;
5668   }
5669 
5670   std::pair<StringRef, StringRef> Fields;
5671   Fields = StringRef(SpecialReg).rsplit('_');
5672   std::string Reg = Fields.first.str();
5673   StringRef Flags = Fields.second;
5674 
5675   // If the target was M Class then need to validate the special register value
5676   // and retrieve the mask for use in the instruction node.
5677   if (Subtarget->isMClass()) {
5678     int SYSmValue = getMClassRegisterMask(SpecialReg, Subtarget);
5679     if (SYSmValue == -1)
5680       return false;
5681 
5682     SDValue Ops[] = { CurDAG->getTargetConstant(SYSmValue, DL, MVT::i32),
5683                       N->getOperand(2), getAL(CurDAG, DL),
5684                       CurDAG->getRegister(0, MVT::i32), N->getOperand(0) };
5685     ReplaceNode(N, CurDAG->getMachineNode(ARM::t2MSR_M, DL, MVT::Other, Ops));
5686     return true;
5687   }
5688 
5689   // We then check to see if a valid mask can be constructed for one of the
5690   // register string values permitted for the A and R class cores. These values
5691   // are apsr, spsr and cpsr; these are also valid on older cores.
5692   int Mask = getARClassRegisterMask(Reg, Flags);
5693   if (Mask != -1) {
5694     Ops = { CurDAG->getTargetConstant(Mask, DL, MVT::i32), N->getOperand(2),
5695             getAL(CurDAG, DL), CurDAG->getRegister(0, MVT::i32),
5696             N->getOperand(0) };
5697     ReplaceNode(N, CurDAG->getMachineNode(IsThumb2 ? ARM::t2MSR_AR : ARM::MSR,
5698                                           DL, MVT::Other, Ops));
5699     return true;
5700   }
5701 
5702   return false;
5703 }
5704 
5705 bool ARMDAGToDAGISel::tryInlineAsm(SDNode *N){
5706   std::vector<SDValue> AsmNodeOperands;
5707   InlineAsm::Flag Flag;
5708   bool Changed = false;
5709   unsigned NumOps = N->getNumOperands();
5710 
5711   // Normally, i64 data is bounded to two arbitrary GRPs for "%r" constraint.
5712   // However, some instrstions (e.g. ldrexd/strexd in ARM mode) require
5713   // (even/even+1) GPRs and use %n and %Hn to refer to the individual regs
5714   // respectively. Since there is no constraint to explicitly specify a
5715   // reg pair, we use GPRPair reg class for "%r" for 64-bit data. For Thumb,
5716   // the 64-bit data may be referred by H, Q, R modifiers, so we still pack
5717   // them into a GPRPair.
5718 
5719   SDLoc dl(N);
5720   SDValue Glue = N->getGluedNode() ? N->getOperand(NumOps - 1) : SDValue();
5721 
5722   SmallVector<bool, 8> OpChanged;
5723   // Glue node will be appended late.
5724   for(unsigned i = 0, e = N->getGluedNode() ? NumOps - 1 : NumOps; i < e; ++i) {
5725     SDValue op = N->getOperand(i);
5726     AsmNodeOperands.push_back(op);
5727 
5728     if (i < InlineAsm::Op_FirstOperand)
5729       continue;
5730 
5731     if (const auto *C = dyn_cast<ConstantSDNode>(N->getOperand(i)))
5732       Flag = InlineAsm::Flag(C->getZExtValue());
5733     else
5734       continue;
5735 
5736     // Immediate operands to inline asm in the SelectionDAG are modeled with
5737     // two operands. The first is a constant of value InlineAsm::Kind::Imm, and
5738     // the second is a constant with the value of the immediate. If we get here
5739     // and we have a Kind::Imm, skip the next operand, and continue.
5740     if (Flag.isImmKind()) {
5741       SDValue op = N->getOperand(++i);
5742       AsmNodeOperands.push_back(op);
5743       continue;
5744     }
5745 
5746     const unsigned NumRegs = Flag.getNumOperandRegisters();
5747     if (NumRegs)
5748       OpChanged.push_back(false);
5749 
5750     unsigned DefIdx = 0;
5751     bool IsTiedToChangedOp = false;
5752     // If it's a use that is tied with a previous def, it has no
5753     // reg class constraint.
5754     if (Changed && Flag.isUseOperandTiedToDef(DefIdx))
5755       IsTiedToChangedOp = OpChanged[DefIdx];
5756 
5757     // Memory operands to inline asm in the SelectionDAG are modeled with two
5758     // operands: a constant of value InlineAsm::Kind::Mem followed by the input
5759     // operand. If we get here and we have a Kind::Mem, skip the next operand
5760     // (so it doesn't get misinterpreted), and continue. We do this here because
5761     // it's important to update the OpChanged array correctly before moving on.
5762     if (Flag.isMemKind()) {
5763       SDValue op = N->getOperand(++i);
5764       AsmNodeOperands.push_back(op);
5765       continue;
5766     }
5767 
5768     if (!Flag.isRegUseKind() && !Flag.isRegDefKind() &&
5769         !Flag.isRegDefEarlyClobberKind())
5770       continue;
5771 
5772     unsigned RC;
5773     const bool HasRC = Flag.hasRegClassConstraint(RC);
5774     if ((!IsTiedToChangedOp && (!HasRC || RC != ARM::GPRRegClassID))
5775         || NumRegs != 2)
5776       continue;
5777 
5778     assert((i+2 < NumOps) && "Invalid number of operands in inline asm");
5779     SDValue V0 = N->getOperand(i+1);
5780     SDValue V1 = N->getOperand(i+2);
5781     Register Reg0 = cast<RegisterSDNode>(V0)->getReg();
5782     Register Reg1 = cast<RegisterSDNode>(V1)->getReg();
5783     SDValue PairedReg;
5784     MachineRegisterInfo &MRI = MF->getRegInfo();
5785 
5786     if (Flag.isRegDefKind() || Flag.isRegDefEarlyClobberKind()) {
5787       // Replace the two GPRs with 1 GPRPair and copy values from GPRPair to
5788       // the original GPRs.
5789 
5790       Register GPVR = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5791       PairedReg = CurDAG->getRegister(GPVR, MVT::Untyped);
5792       SDValue Chain = SDValue(N,0);
5793 
5794       SDNode *GU = N->getGluedUser();
5795       SDValue RegCopy = CurDAG->getCopyFromReg(Chain, dl, GPVR, MVT::Untyped,
5796                                                Chain.getValue(1));
5797 
5798       // Extract values from a GPRPair reg and copy to the original GPR reg.
5799       SDValue Sub0 = CurDAG->getTargetExtractSubreg(ARM::gsub_0, dl, MVT::i32,
5800                                                     RegCopy);
5801       SDValue Sub1 = CurDAG->getTargetExtractSubreg(ARM::gsub_1, dl, MVT::i32,
5802                                                     RegCopy);
5803       SDValue T0 = CurDAG->getCopyToReg(Sub0, dl, Reg0, Sub0,
5804                                         RegCopy.getValue(1));
5805       SDValue T1 = CurDAG->getCopyToReg(Sub1, dl, Reg1, Sub1, T0.getValue(1));
5806 
5807       // Update the original glue user.
5808       std::vector<SDValue> Ops(GU->op_begin(), GU->op_end()-1);
5809       Ops.push_back(T1.getValue(1));
5810       CurDAG->UpdateNodeOperands(GU, Ops);
5811     } else {
5812       // For Kind  == InlineAsm::Kind::RegUse, we first copy two GPRs into a
5813       // GPRPair and then pass the GPRPair to the inline asm.
5814       SDValue Chain = AsmNodeOperands[InlineAsm::Op_InputChain];
5815 
5816       // As REG_SEQ doesn't take RegisterSDNode, we copy them first.
5817       SDValue T0 = CurDAG->getCopyFromReg(Chain, dl, Reg0, MVT::i32,
5818                                           Chain.getValue(1));
5819       SDValue T1 = CurDAG->getCopyFromReg(Chain, dl, Reg1, MVT::i32,
5820                                           T0.getValue(1));
5821       SDValue Pair = SDValue(createGPRPairNode(MVT::Untyped, T0, T1), 0);
5822 
5823       // Copy REG_SEQ into a GPRPair-typed VR and replace the original two
5824       // i32 VRs of inline asm with it.
5825       Register GPVR = MRI.createVirtualRegister(&ARM::GPRPairRegClass);
5826       PairedReg = CurDAG->getRegister(GPVR, MVT::Untyped);
5827       Chain = CurDAG->getCopyToReg(T1, dl, GPVR, Pair, T1.getValue(1));
5828 
5829       AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
5830       Glue = Chain.getValue(1);
5831     }
5832 
5833     Changed = true;
5834 
5835     if(PairedReg.getNode()) {
5836       OpChanged[OpChanged.size() -1 ] = true;
5837       Flag = InlineAsm::Flag(Flag.getKind(), 1 /* RegNum*/);
5838       if (IsTiedToChangedOp)
5839         Flag.setMatchingOp(DefIdx);
5840       else
5841         Flag.setRegClass(ARM::GPRPairRegClassID);
5842       // Replace the current flag.
5843       AsmNodeOperands[AsmNodeOperands.size() -1] = CurDAG->getTargetConstant(
5844           Flag, dl, MVT::i32);
5845       // Add the new register node and skip the original two GPRs.
5846       AsmNodeOperands.push_back(PairedReg);
5847       // Skip the next two GPRs.
5848       i += 2;
5849     }
5850   }
5851 
5852   if (Glue.getNode())
5853     AsmNodeOperands.push_back(Glue);
5854   if (!Changed)
5855     return false;
5856 
5857   SDValue New = CurDAG->getNode(N->getOpcode(), SDLoc(N),
5858       CurDAG->getVTList(MVT::Other, MVT::Glue), AsmNodeOperands);
5859   New->setNodeId(-1);
5860   ReplaceNode(N, New.getNode());
5861   return true;
5862 }
5863 
5864 bool ARMDAGToDAGISel::SelectInlineAsmMemoryOperand(
5865     const SDValue &Op, InlineAsm::ConstraintCode ConstraintID,
5866     std::vector<SDValue> &OutOps) {
5867   switch(ConstraintID) {
5868   default:
5869     llvm_unreachable("Unexpected asm memory constraint");
5870   case InlineAsm::ConstraintCode::m:
5871   case InlineAsm::ConstraintCode::o:
5872   case InlineAsm::ConstraintCode::Q:
5873   case InlineAsm::ConstraintCode::Um:
5874   case InlineAsm::ConstraintCode::Un:
5875   case InlineAsm::ConstraintCode::Uq:
5876   case InlineAsm::ConstraintCode::Us:
5877   case InlineAsm::ConstraintCode::Ut:
5878   case InlineAsm::ConstraintCode::Uv:
5879   case InlineAsm::ConstraintCode::Uy:
5880     // Require the address to be in a register.  That is safe for all ARM
5881     // variants and it is hard to do anything much smarter without knowing
5882     // how the operand is used.
5883     OutOps.push_back(Op);
5884     return false;
5885   }
5886   return true;
5887 }
5888 
5889 /// createARMISelDag - This pass converts a legalized DAG into a
5890 /// ARM-specific DAG, ready for instruction scheduling.
5891 ///
5892 FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM,
5893                                      CodeGenOptLevel OptLevel) {
5894   return new ARMDAGToDAGISel(TM, OptLevel);
5895 }
5896