ARMISelLowering.h revision 268c743a3ba44ada364938bc5ff9b1be219df54f
1//===-- ARMISelLowering.h - ARM DAG Lowering Interface ----------*- C++ -*-===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines the interfaces that ARM uses to lower LLVM code into a
11// selection DAG.
12//
13//===----------------------------------------------------------------------===//
14
15#ifndef ARMISELLOWERING_H
16#define ARMISELLOWERING_H
17
18#include "ARM.h"
19#include "ARMSubtarget.h"
20#include "llvm/CodeGen/CallingConvLower.h"
21#include "llvm/CodeGen/FastISel.h"
22#include "llvm/CodeGen/SelectionDAG.h"
23#include "llvm/Target/TargetLowering.h"
24#include "llvm/Target/TargetRegisterInfo.h"
25#include <vector>
26
27namespace llvm {
28  class ARMConstantPoolValue;
29
30  namespace ARMISD {
31    // ARM Specific DAG Nodes
32    enum NodeType {
33      // Start the numbering where the builtin ops and target ops leave off.
34      FIRST_NUMBER = ISD::BUILTIN_OP_END,
35
36      Wrapper,      // Wrapper - A wrapper node for TargetConstantPool,
37                    // TargetExternalSymbol, and TargetGlobalAddress.
38      WrapperDYN,   // WrapperDYN - A wrapper node for TargetGlobalAddress in
39                    // DYN mode.
40      WrapperPIC,   // WrapperPIC - A wrapper node for TargetGlobalAddress in
41                    // PIC mode.
42      WrapperJT,    // WrapperJT - A wrapper node for TargetJumpTable
43
44      // Add pseudo op to model memcpy for struct byval.
45      COPY_STRUCT_BYVAL,
46
47      CALL,         // Function call.
48      CALL_PRED,    // Function call that's predicable.
49      CALL_NOLINK,  // Function call with branch not branch-and-link.
50      tCALL,        // Thumb function call.
51      BRCOND,       // Conditional branch.
52      BR_JT,        // Jumptable branch.
53      BR2_JT,       // Jumptable branch (2 level - jumptable entry is a jump).
54      RET_FLAG,     // Return with a flag operand.
55
56      PIC_ADD,      // Add with a PC operand and a PIC label.
57
58      CMP,          // ARM compare instructions.
59      CMN,          // ARM CMN instructions.
60      CMPZ,         // ARM compare that sets only Z flag.
61      CMPFP,        // ARM VFP compare instruction, sets FPSCR.
62      CMPFPw0,      // ARM VFP compare against zero instruction, sets FPSCR.
63      FMSTAT,       // ARM fmstat instruction.
64
65      CMOV,         // ARM conditional move instructions.
66
67      BCC_i64,
68
69      RBIT,         // ARM bitreverse instruction
70
71      FTOSI,        // FP to sint within a FP register.
72      FTOUI,        // FP to uint within a FP register.
73      SITOF,        // sint to FP within a FP register.
74      UITOF,        // uint to FP within a FP register.
75
76      SRL_FLAG,     // V,Flag = srl_flag X -> srl X, 1 + save carry out.
77      SRA_FLAG,     // V,Flag = sra_flag X -> sra X, 1 + save carry out.
78      RRX,          // V = RRX X, Flag     -> srl X, 1 + shift in carry flag.
79
80      ADDC,         // Add with carry
81      ADDE,         // Add using carry
82      SUBC,         // Sub with carry
83      SUBE,         // Sub using carry
84
85      VMOVRRD,      // double to two gprs.
86      VMOVDRR,      // Two gprs to double.
87
88      EH_SJLJ_SETJMP,         // SjLj exception handling setjmp.
89      EH_SJLJ_LONGJMP,        // SjLj exception handling longjmp.
90
91      TC_RETURN,    // Tail call return pseudo.
92
93      THREAD_POINTER,
94
95      DYN_ALLOC,    // Dynamic allocation on the stack.
96
97      MEMBARRIER_MCR, // Memory barrier (MCR)
98
99      PRELOAD,      // Preload
100
101      VCEQ,         // Vector compare equal.
102      VCEQZ,        // Vector compare equal to zero.
103      VCGE,         // Vector compare greater than or equal.
104      VCGEZ,        // Vector compare greater than or equal to zero.
105      VCLEZ,        // Vector compare less than or equal to zero.
106      VCGEU,        // Vector compare unsigned greater than or equal.
107      VCGT,         // Vector compare greater than.
108      VCGTZ,        // Vector compare greater than zero.
109      VCLTZ,        // Vector compare less than zero.
110      VCGTU,        // Vector compare unsigned greater than.
111      VTST,         // Vector test bits.
112
113      // Vector shift by immediate:
114      VSHL,         // ...left
115      VSHRs,        // ...right (signed)
116      VSHRu,        // ...right (unsigned)
117      VSHLLs,       // ...left long (signed)
118      VSHLLu,       // ...left long (unsigned)
119      VSHLLi,       // ...left long (with maximum shift count)
120      VSHRN,        // ...right narrow
121
122      // Vector rounding shift by immediate:
123      VRSHRs,       // ...right (signed)
124      VRSHRu,       // ...right (unsigned)
125      VRSHRN,       // ...right narrow
126
127      // Vector saturating shift by immediate:
128      VQSHLs,       // ...left (signed)
129      VQSHLu,       // ...left (unsigned)
130      VQSHLsu,      // ...left (signed to unsigned)
131      VQSHRNs,      // ...right narrow (signed)
132      VQSHRNu,      // ...right narrow (unsigned)
133      VQSHRNsu,     // ...right narrow (signed to unsigned)
134
135      // Vector saturating rounding shift by immediate:
136      VQRSHRNs,     // ...right narrow (signed)
137      VQRSHRNu,     // ...right narrow (unsigned)
138      VQRSHRNsu,    // ...right narrow (signed to unsigned)
139
140      // Vector shift and insert:
141      VSLI,         // ...left
142      VSRI,         // ...right
143
144      // Vector get lane (VMOV scalar to ARM core register)
145      // (These are used for 8- and 16-bit element types only.)
146      VGETLANEu,    // zero-extend vector extract element
147      VGETLANEs,    // sign-extend vector extract element
148
149      // Vector move immediate and move negated immediate:
150      VMOVIMM,
151      VMVNIMM,
152
153      // Vector move f32 immediate:
154      VMOVFPIMM,
155
156      // Vector duplicate:
157      VDUP,
158      VDUPLANE,
159
160      // Vector shuffles:
161      VEXT,         // extract
162      VREV64,       // reverse elements within 64-bit doublewords
163      VREV32,       // reverse elements within 32-bit words
164      VREV16,       // reverse elements within 16-bit halfwords
165      VZIP,         // zip (interleave)
166      VUZP,         // unzip (deinterleave)
167      VTRN,         // transpose
168      VTBL1,        // 1-register shuffle with mask
169      VTBL2,        // 2-register shuffle with mask
170
171      // Vector multiply long:
172      VMULLs,       // ...signed
173      VMULLu,       // ...unsigned
174
175      UMLAL,        // 64bit Unsigned Accumulate Multiply
176      SMLAL,        // 64bit Signed Accumulate Multiply
177
178      // Operands of the standard BUILD_VECTOR node are not legalized, which
179      // is fine if BUILD_VECTORs are always lowered to shuffles or other
180      // operations, but for ARM some BUILD_VECTORs are legal as-is and their
181      // operands need to be legalized.  Define an ARM-specific version of
182      // BUILD_VECTOR for this purpose.
183      BUILD_VECTOR,
184
185      // Floating-point max and min:
186      FMAX,
187      FMIN,
188      VMAXNM,
189      VMINNM,
190
191      // Bit-field insert
192      BFI,
193
194      // Vector OR with immediate
195      VORRIMM,
196      // Vector AND with NOT of immediate
197      VBICIMM,
198
199      // Vector bitwise select
200      VBSL,
201
202      // Vector load N-element structure to all lanes:
203      VLD2DUP = ISD::FIRST_TARGET_MEMORY_OPCODE,
204      VLD3DUP,
205      VLD4DUP,
206
207      // NEON loads with post-increment base updates:
208      VLD1_UPD,
209      VLD2_UPD,
210      VLD3_UPD,
211      VLD4_UPD,
212      VLD2LN_UPD,
213      VLD3LN_UPD,
214      VLD4LN_UPD,
215      VLD2DUP_UPD,
216      VLD3DUP_UPD,
217      VLD4DUP_UPD,
218
219      // NEON stores with post-increment base updates:
220      VST1_UPD,
221      VST2_UPD,
222      VST3_UPD,
223      VST4_UPD,
224      VST2LN_UPD,
225      VST3LN_UPD,
226      VST4LN_UPD
227    };
228  }
229
230  /// Define some predicates that are used for node matching.
231  namespace ARM {
232    bool isBitFieldInvertedMask(unsigned v);
233  }
234
235  //===--------------------------------------------------------------------===//
236  //  ARMTargetLowering - ARM Implementation of the TargetLowering interface
237
238  class ARMTargetLowering : public TargetLowering {
239  public:
240    explicit ARMTargetLowering(TargetMachine &TM);
241
242    virtual unsigned getJumpTableEncoding() const;
243
244    virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
245
246    /// ReplaceNodeResults - Replace the results of node with an illegal result
247    /// type with new values built out of custom code.
248    ///
249    virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
250                                    SelectionDAG &DAG) const;
251
252    virtual const char *getTargetNodeName(unsigned Opcode) const;
253
254    virtual bool isSelectSupported(SelectSupportKind Kind) const {
255      // ARM does not support scalar condition selects on vectors.
256      return (Kind != ScalarCondVectorVal);
257    }
258
259    /// getSetCCResultType - Return the value type to use for ISD::SETCC.
260    virtual EVT getSetCCResultType(LLVMContext &Context, EVT VT) const;
261
262    virtual MachineBasicBlock *
263      EmitInstrWithCustomInserter(MachineInstr *MI,
264                                  MachineBasicBlock *MBB) const;
265
266    virtual void
267    AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const;
268
269    SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const;
270    virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
271
272    bool isDesirableToTransformToIntegerOp(unsigned Opc, EVT VT) const;
273
274    /// allowsUnalignedMemoryAccesses - Returns true if the target allows
275    /// unaligned memory accesses of the specified type. Returns whether it
276    /// is "fast" by reference in the second argument.
277    virtual bool allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const;
278
279    virtual EVT getOptimalMemOpType(uint64_t Size,
280                                    unsigned DstAlign, unsigned SrcAlign,
281                                    bool IsMemset, bool ZeroMemset,
282                                    bool MemcpyStrSrc,
283                                    MachineFunction &MF) const;
284
285    using TargetLowering::isZExtFree;
286    virtual bool isZExtFree(SDValue Val, EVT VT2) const;
287
288    virtual bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const;
289
290
291    /// isLegalAddressingMode - Return true if the addressing mode represented
292    /// by AM is legal for this target, for a load/store of the specified type.
293    virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const;
294    bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const;
295
296    /// isLegalICmpImmediate - Return true if the specified immediate is legal
297    /// icmp immediate, that is the target has icmp instructions which can
298    /// compare a register against the immediate without having to materialize
299    /// the immediate into a register.
300    virtual bool isLegalICmpImmediate(int64_t Imm) const;
301
302    /// isLegalAddImmediate - Return true if the specified immediate is legal
303    /// add immediate, that is the target has add instructions which can
304    /// add a register and the immediate without having to materialize
305    /// the immediate into a register.
306    virtual bool isLegalAddImmediate(int64_t Imm) const;
307
308    /// getPreIndexedAddressParts - returns true by value, base pointer and
309    /// offset pointer and addressing mode by reference if the node's address
310    /// can be legally represented as pre-indexed load / store address.
311    virtual bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
312                                           SDValue &Offset,
313                                           ISD::MemIndexedMode &AM,
314                                           SelectionDAG &DAG) const;
315
316    /// getPostIndexedAddressParts - returns true by value, base pointer and
317    /// offset pointer and addressing mode by reference if this node can be
318    /// combined with a load / store to form a post-indexed load / store.
319    virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op,
320                                            SDValue &Base, SDValue &Offset,
321                                            ISD::MemIndexedMode &AM,
322                                            SelectionDAG &DAG) const;
323
324    virtual void computeMaskedBitsForTargetNode(const SDValue Op,
325                                                APInt &KnownZero,
326                                                APInt &KnownOne,
327                                                const SelectionDAG &DAG,
328                                                unsigned Depth) const;
329
330
331    virtual bool ExpandInlineAsm(CallInst *CI) const;
332
333    ConstraintType getConstraintType(const std::string &Constraint) const;
334
335    /// Examine constraint string and operand type and determine a weight value.
336    /// The operand object must already have been set up with the operand type.
337    ConstraintWeight getSingleConstraintMatchWeight(
338      AsmOperandInfo &info, const char *constraint) const;
339
340    std::pair<unsigned, const TargetRegisterClass*>
341      getRegForInlineAsmConstraint(const std::string &Constraint,
342                                   MVT VT) const;
343
344    /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
345    /// vector.  If it is invalid, don't add anything to Ops. If hasMemory is
346    /// true it means one of the asm constraint of the inline asm instruction
347    /// being processed is 'm'.
348    virtual void LowerAsmOperandForConstraint(SDValue Op,
349                                              std::string &Constraint,
350                                              std::vector<SDValue> &Ops,
351                                              SelectionDAG &DAG) const;
352
353    const ARMSubtarget* getSubtarget() const {
354      return Subtarget;
355    }
356
357    /// getRegClassFor - Return the register class that should be used for the
358    /// specified value type.
359    virtual const TargetRegisterClass *getRegClassFor(MVT VT) const;
360
361    /// getMaximalGlobalOffset - Returns the maximal possible offset which can
362    /// be used for loads / stores from the global.
363    virtual unsigned getMaximalGlobalOffset() const;
364
365    /// createFastISel - This method returns a target specific FastISel object,
366    /// or null if the target does not support "fast" ISel.
367    virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
368                                     const TargetLibraryInfo *libInfo) const;
369
370    Sched::Preference getSchedulingPreference(SDNode *N) const;
371
372    bool isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const;
373    bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const;
374
375    /// isFPImmLegal - Returns true if the target can instruction select the
376    /// specified FP immediate natively. If false, the legalizer will
377    /// materialize the FP immediate as a load from a constant pool.
378    virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
379
380    virtual bool getTgtMemIntrinsic(IntrinsicInfo &Info,
381                                    const CallInst &I,
382                                    unsigned Intrinsic) const;
383  protected:
384    std::pair<const TargetRegisterClass*, uint8_t>
385    findRepresentativeClass(MVT VT) const;
386
387  private:
388    /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
389    /// make the right decision when generating code for different targets.
390    const ARMSubtarget *Subtarget;
391
392    const TargetRegisterInfo *RegInfo;
393
394    const InstrItineraryData *Itins;
395
396    /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created.
397    ///
398    unsigned ARMPCLabelIndex;
399
400    void addTypeForNEON(MVT VT, MVT PromotedLdStVT, MVT PromotedBitwiseVT);
401    void addDRTypeForNEON(MVT VT);
402    void addQRTypeForNEON(MVT VT);
403
404    typedef SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPassVector;
405    void PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
406                          SDValue Chain, SDValue &Arg,
407                          RegsToPassVector &RegsToPass,
408                          CCValAssign &VA, CCValAssign &NextVA,
409                          SDValue &StackPtr,
410                          SmallVectorImpl<SDValue> &MemOpChains,
411                          ISD::ArgFlagsTy Flags) const;
412    SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
413                                 SDValue &Root, SelectionDAG &DAG,
414                                 SDLoc dl) const;
415
416    CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return,
417                                  bool isVarArg) const;
418    SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
419                             SDLoc dl, SelectionDAG &DAG,
420                             const CCValAssign &VA,
421                             ISD::ArgFlagsTy Flags) const;
422    SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
423    SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
424    SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
425                                    const ARMSubtarget *Subtarget) const;
426    SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
427    SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const;
428    SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const;
429    SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
430    SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
431                                            SelectionDAG &DAG) const;
432    SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA,
433                                 SelectionDAG &DAG,
434                                 TLSModel::Model model) const;
435    SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const;
436    SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
437    SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
438    SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
439    SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
440    SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
441    SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
442    SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
443    SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
444    SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
445    SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
446    SDValue LowerConstantFP(SDValue Op, SelectionDAG &DAG,
447                            const ARMSubtarget *ST) const;
448    SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
449                              const ARMSubtarget *ST) const;
450    SDValue LowerDivRem(SDValue Op, SelectionDAG &DAG) const;
451
452    /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
453    /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
454    /// expanded to FMAs when this method returns true, otherwise fmuladd is
455    /// expanded to fmul + fadd.
456    ///
457    /// ARM supports both fused and unfused multiply-add operations; we already
458    /// lower a pair of fmul and fadd to the latter so it's not clear that there
459    /// would be a gain or that the gain would be worthwhile enough to risk
460    /// correctness bugs.
461    virtual bool isFMAFasterThanFMulAndFAdd(EVT VT) const { return false; }
462
463    SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const;
464
465    SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
466                            CallingConv::ID CallConv, bool isVarArg,
467                            const SmallVectorImpl<ISD::InputArg> &Ins,
468                            SDLoc dl, SelectionDAG &DAG,
469                            SmallVectorImpl<SDValue> &InVals,
470                            bool isThisReturn, SDValue ThisVal) const;
471
472    virtual SDValue
473      LowerFormalArguments(SDValue Chain,
474                           CallingConv::ID CallConv, bool isVarArg,
475                           const SmallVectorImpl<ISD::InputArg> &Ins,
476                           SDLoc dl, SelectionDAG &DAG,
477                           SmallVectorImpl<SDValue> &InVals) const;
478
479    int StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
480                       SDLoc dl, SDValue &Chain,
481                       const Value *OrigArg,
482                       unsigned InRegsParamRecordIdx,
483                       unsigned OffsetFromOrigArg,
484                       unsigned ArgOffset,
485                       unsigned ArgSize,
486                       bool ForceMutable) const;
487
488    void VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
489                              SDLoc dl, SDValue &Chain,
490                              unsigned ArgOffset,
491                              bool ForceMutable = false) const;
492
493    void computeRegArea(CCState &CCInfo, MachineFunction &MF,
494                        unsigned InRegsParamRecordIdx,
495                        unsigned ArgSize,
496                        unsigned &ArgRegsSize,
497                        unsigned &ArgRegsSaveSize) const;
498
499    virtual SDValue
500      LowerCall(TargetLowering::CallLoweringInfo &CLI,
501                SmallVectorImpl<SDValue> &InVals) const;
502
503    /// HandleByVal - Target-specific cleanup for ByVal support.
504    virtual void HandleByVal(CCState *, unsigned &, unsigned) const;
505
506    /// IsEligibleForTailCallOptimization - Check whether the call is eligible
507    /// for tail call optimization. Targets which want to do tail call
508    /// optimization should implement this function.
509    bool IsEligibleForTailCallOptimization(SDValue Callee,
510                                           CallingConv::ID CalleeCC,
511                                           bool isVarArg,
512                                           bool isCalleeStructRet,
513                                           bool isCallerStructRet,
514                                    const SmallVectorImpl<ISD::OutputArg> &Outs,
515                                    const SmallVectorImpl<SDValue> &OutVals,
516                                    const SmallVectorImpl<ISD::InputArg> &Ins,
517                                           SelectionDAG& DAG) const;
518
519    virtual bool CanLowerReturn(CallingConv::ID CallConv,
520                                MachineFunction &MF, bool isVarArg,
521                                const SmallVectorImpl<ISD::OutputArg> &Outs,
522                                LLVMContext &Context) const;
523
524    virtual SDValue
525      LowerReturn(SDValue Chain,
526                  CallingConv::ID CallConv, bool isVarArg,
527                  const SmallVectorImpl<ISD::OutputArg> &Outs,
528                  const SmallVectorImpl<SDValue> &OutVals,
529                  SDLoc dl, SelectionDAG &DAG) const;
530
531    virtual bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const;
532
533    virtual bool mayBeEmittedAsTailCall(CallInst *CI) const;
534
535    SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
536                      SDValue &ARMcc, SelectionDAG &DAG, SDLoc dl) const;
537    SDValue getVFPCmp(SDValue LHS, SDValue RHS,
538                      SelectionDAG &DAG, SDLoc dl) const;
539    SDValue duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const;
540
541    SDValue OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const;
542
543    MachineBasicBlock *EmitAtomicCmpSwap(MachineInstr *MI,
544                                         MachineBasicBlock *BB,
545                                         unsigned Size) const;
546    MachineBasicBlock *EmitAtomicBinary(MachineInstr *MI,
547                                        MachineBasicBlock *BB,
548                                        unsigned Size,
549                                        unsigned BinOpcode) const;
550    MachineBasicBlock *EmitAtomicBinary64(MachineInstr *MI,
551                                          MachineBasicBlock *BB,
552                                          unsigned Op1,
553                                          unsigned Op2,
554                                          bool NeedsCarry = false,
555                                          bool IsCmpxchg = false,
556                                          bool IsMinMax = false,
557                                          ARMCC::CondCodes CC = ARMCC::AL) const;
558    MachineBasicBlock * EmitAtomicBinaryMinMax(MachineInstr *MI,
559                                               MachineBasicBlock *BB,
560                                               unsigned Size,
561                                               bool signExtend,
562                                               ARMCC::CondCodes Cond) const;
563    MachineBasicBlock *EmitAtomicLoad64(MachineInstr *MI,
564                                        MachineBasicBlock *BB) const;
565
566    void SetupEntryBlockForSjLj(MachineInstr *MI,
567                                MachineBasicBlock *MBB,
568                                MachineBasicBlock *DispatchBB, int FI) const;
569
570    MachineBasicBlock *EmitSjLjDispatchBlock(MachineInstr *MI,
571                                             MachineBasicBlock *MBB) const;
572
573    bool RemapAddSubWithFlags(MachineInstr *MI, MachineBasicBlock *BB) const;
574
575    MachineBasicBlock *EmitStructByval(MachineInstr *MI,
576                                       MachineBasicBlock *MBB) const;
577  };
578
579  enum NEONModImmType {
580    VMOVModImm,
581    VMVNModImm,
582    OtherModImm
583  };
584
585
586  namespace ARM {
587    FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
588                             const TargetLibraryInfo *libInfo);
589  }
590}
591
592#endif  // ARMISELLOWERING_H
593