ARMISelLowering.h revision 31928dfc03d92322f9f2fb1c4a7878024d3cc9d1
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      INTRET_FLAG,  // Interrupt return with an LR-offset and a flag operand.
56
57      PIC_ADD,      // Add with a PC operand and a PIC label.
58
59      CMP,          // ARM compare instructions.
60      CMN,          // ARM CMN instructions.
61      CMPZ,         // ARM compare that sets only Z flag.
62      CMPFP,        // ARM VFP compare instruction, sets FPSCR.
63      CMPFPw0,      // ARM VFP compare against zero instruction, sets FPSCR.
64      FMSTAT,       // ARM fmstat instruction.
65
66      CMOV,         // ARM conditional move instructions.
67
68      BCC_i64,
69
70      RBIT,         // ARM bitreverse instruction
71
72      FTOSI,        // FP to sint within a FP register.
73      FTOUI,        // FP to uint within a FP register.
74      SITOF,        // sint to FP within a FP register.
75      UITOF,        // uint to FP within a FP register.
76
77      SRL_FLAG,     // V,Flag = srl_flag X -> srl X, 1 + save carry out.
78      SRA_FLAG,     // V,Flag = sra_flag X -> sra X, 1 + save carry out.
79      RRX,          // V = RRX X, Flag     -> srl X, 1 + shift in carry flag.
80
81      ADDC,         // Add with carry
82      ADDE,         // Add using carry
83      SUBC,         // Sub with carry
84      SUBE,         // Sub using carry
85
86      VMOVRRD,      // double to two gprs.
87      VMOVDRR,      // Two gprs to double.
88
89      EH_SJLJ_SETJMP,         // SjLj exception handling setjmp.
90      EH_SJLJ_LONGJMP,        // SjLj exception handling longjmp.
91
92      TC_RETURN,    // Tail call return pseudo.
93
94      THREAD_POINTER,
95
96      DYN_ALLOC,    // Dynamic allocation on the stack.
97
98      MEMBARRIER_MCR, // Memory barrier (MCR)
99
100      PRELOAD,      // Preload
101
102      VCEQ,         // Vector compare equal.
103      VCEQZ,        // Vector compare equal to zero.
104      VCGE,         // Vector compare greater than or equal.
105      VCGEZ,        // Vector compare greater than or equal to zero.
106      VCLEZ,        // Vector compare less than or equal to zero.
107      VCGEU,        // Vector compare unsigned greater than or equal.
108      VCGT,         // Vector compare greater than.
109      VCGTZ,        // Vector compare greater than zero.
110      VCLTZ,        // Vector compare less than zero.
111      VCGTU,        // Vector compare unsigned greater than.
112      VTST,         // Vector test bits.
113
114      // Vector shift by immediate:
115      VSHL,         // ...left
116      VSHRs,        // ...right (signed)
117      VSHRu,        // ...right (unsigned)
118      VSHLLs,       // ...left long (signed)
119      VSHLLu,       // ...left long (unsigned)
120      VSHLLi,       // ...left long (with maximum shift count)
121      VSHRN,        // ...right narrow
122
123      // Vector rounding shift by immediate:
124      VRSHRs,       // ...right (signed)
125      VRSHRu,       // ...right (unsigned)
126      VRSHRN,       // ...right narrow
127
128      // Vector saturating shift by immediate:
129      VQSHLs,       // ...left (signed)
130      VQSHLu,       // ...left (unsigned)
131      VQSHLsu,      // ...left (signed to unsigned)
132      VQSHRNs,      // ...right narrow (signed)
133      VQSHRNu,      // ...right narrow (unsigned)
134      VQSHRNsu,     // ...right narrow (signed to unsigned)
135
136      // Vector saturating rounding shift by immediate:
137      VQRSHRNs,     // ...right narrow (signed)
138      VQRSHRNu,     // ...right narrow (unsigned)
139      VQRSHRNsu,    // ...right narrow (signed to unsigned)
140
141      // Vector shift and insert:
142      VSLI,         // ...left
143      VSRI,         // ...right
144
145      // Vector get lane (VMOV scalar to ARM core register)
146      // (These are used for 8- and 16-bit element types only.)
147      VGETLANEu,    // zero-extend vector extract element
148      VGETLANEs,    // sign-extend vector extract element
149
150      // Vector move immediate and move negated immediate:
151      VMOVIMM,
152      VMVNIMM,
153
154      // Vector move f32 immediate:
155      VMOVFPIMM,
156
157      // Vector duplicate:
158      VDUP,
159      VDUPLANE,
160
161      // Vector shuffles:
162      VEXT,         // extract
163      VREV64,       // reverse elements within 64-bit doublewords
164      VREV32,       // reverse elements within 32-bit words
165      VREV16,       // reverse elements within 16-bit halfwords
166      VZIP,         // zip (interleave)
167      VUZP,         // unzip (deinterleave)
168      VTRN,         // transpose
169      VTBL1,        // 1-register shuffle with mask
170      VTBL2,        // 2-register shuffle with mask
171
172      // Vector multiply long:
173      VMULLs,       // ...signed
174      VMULLu,       // ...unsigned
175
176      UMLAL,        // 64bit Unsigned Accumulate Multiply
177      SMLAL,        // 64bit Signed Accumulate Multiply
178
179      // Operands of the standard BUILD_VECTOR node are not legalized, which
180      // is fine if BUILD_VECTORs are always lowered to shuffles or other
181      // operations, but for ARM some BUILD_VECTORs are legal as-is and their
182      // operands need to be legalized.  Define an ARM-specific version of
183      // BUILD_VECTOR for this purpose.
184      BUILD_VECTOR,
185
186      // Floating-point max and min:
187      FMAX,
188      FMIN,
189      VMAXNM,
190      VMINNM,
191
192      // Bit-field insert
193      BFI,
194
195      // Vector OR with immediate
196      VORRIMM,
197      // Vector AND with NOT of immediate
198      VBICIMM,
199
200      // Vector bitwise select
201      VBSL,
202
203      // Vector load N-element structure to all lanes:
204      VLD2DUP = ISD::FIRST_TARGET_MEMORY_OPCODE,
205      VLD3DUP,
206      VLD4DUP,
207
208      // NEON loads with post-increment base updates:
209      VLD1_UPD,
210      VLD2_UPD,
211      VLD3_UPD,
212      VLD4_UPD,
213      VLD2LN_UPD,
214      VLD3LN_UPD,
215      VLD4LN_UPD,
216      VLD2DUP_UPD,
217      VLD3DUP_UPD,
218      VLD4DUP_UPD,
219
220      // NEON stores with post-increment base updates:
221      VST1_UPD,
222      VST2_UPD,
223      VST3_UPD,
224      VST4_UPD,
225      VST2LN_UPD,
226      VST3LN_UPD,
227      VST4LN_UPD
228    };
229  }
230
231  /// Define some predicates that are used for node matching.
232  namespace ARM {
233    bool isBitFieldInvertedMask(unsigned v);
234  }
235
236  //===--------------------------------------------------------------------===//
237  //  ARMTargetLowering - ARM Implementation of the TargetLowering interface
238
239  class ARMTargetLowering : public TargetLowering {
240  public:
241    explicit ARMTargetLowering(TargetMachine &TM);
242
243    virtual unsigned getJumpTableEncoding() const;
244
245    virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
246
247    /// ReplaceNodeResults - Replace the results of node with an illegal result
248    /// type with new values built out of custom code.
249    ///
250    virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
251                                    SelectionDAG &DAG) const;
252
253    virtual const char *getTargetNodeName(unsigned Opcode) const;
254
255    virtual bool isSelectSupported(SelectSupportKind Kind) const {
256      // ARM does not support scalar condition selects on vectors.
257      return (Kind != ScalarCondVectorVal);
258    }
259
260    /// getSetCCResultType - Return the value type to use for ISD::SETCC.
261    virtual EVT getSetCCResultType(LLVMContext &Context, EVT VT) const;
262
263    virtual MachineBasicBlock *
264      EmitInstrWithCustomInserter(MachineInstr *MI,
265                                  MachineBasicBlock *MBB) const;
266
267    virtual void
268    AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const;
269
270    SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const;
271    virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
272
273    bool isDesirableToTransformToIntegerOp(unsigned Opc, EVT VT) const;
274
275    /// allowsUnalignedMemoryAccesses - Returns true if the target allows
276    /// unaligned memory accesses of the specified type. Returns whether it
277    /// is "fast" by reference in the second argument.
278    virtual bool allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const;
279
280    virtual EVT getOptimalMemOpType(uint64_t Size,
281                                    unsigned DstAlign, unsigned SrcAlign,
282                                    bool IsMemset, bool ZeroMemset,
283                                    bool MemcpyStrSrc,
284                                    MachineFunction &MF) const;
285
286    using TargetLowering::isZExtFree;
287    virtual bool isZExtFree(SDValue Val, EVT VT2) const;
288
289    virtual bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const;
290
291
292    /// isLegalAddressingMode - Return true if the addressing mode represented
293    /// by AM is legal for this target, for a load/store of the specified type.
294    virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const;
295    bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const;
296
297    /// isLegalICmpImmediate - Return true if the specified immediate is legal
298    /// icmp immediate, that is the target has icmp instructions which can
299    /// compare a register against the immediate without having to materialize
300    /// the immediate into a register.
301    virtual bool isLegalICmpImmediate(int64_t Imm) const;
302
303    /// isLegalAddImmediate - Return true if the specified immediate is legal
304    /// add immediate, that is the target has add instructions which can
305    /// add a register and the immediate without having to materialize
306    /// the immediate into a register.
307    virtual bool isLegalAddImmediate(int64_t Imm) const;
308
309    /// getPreIndexedAddressParts - returns true by value, base pointer and
310    /// offset pointer and addressing mode by reference if the node's address
311    /// can be legally represented as pre-indexed load / store address.
312    virtual bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
313                                           SDValue &Offset,
314                                           ISD::MemIndexedMode &AM,
315                                           SelectionDAG &DAG) const;
316
317    /// getPostIndexedAddressParts - returns true by value, base pointer and
318    /// offset pointer and addressing mode by reference if this node can be
319    /// combined with a load / store to form a post-indexed load / store.
320    virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op,
321                                            SDValue &Base, SDValue &Offset,
322                                            ISD::MemIndexedMode &AM,
323                                            SelectionDAG &DAG) const;
324
325    virtual void computeMaskedBitsForTargetNode(const SDValue Op,
326                                                APInt &KnownZero,
327                                                APInt &KnownOne,
328                                                const SelectionDAG &DAG,
329                                                unsigned Depth) const;
330
331
332    virtual bool ExpandInlineAsm(CallInst *CI) const;
333
334    ConstraintType getConstraintType(const std::string &Constraint) const;
335
336    /// Examine constraint string and operand type and determine a weight value.
337    /// The operand object must already have been set up with the operand type.
338    ConstraintWeight getSingleConstraintMatchWeight(
339      AsmOperandInfo &info, const char *constraint) const;
340
341    std::pair<unsigned, const TargetRegisterClass*>
342      getRegForInlineAsmConstraint(const std::string &Constraint,
343                                   MVT VT) const;
344
345    /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
346    /// vector.  If it is invalid, don't add anything to Ops. If hasMemory is
347    /// true it means one of the asm constraint of the inline asm instruction
348    /// being processed is 'm'.
349    virtual void LowerAsmOperandForConstraint(SDValue Op,
350                                              std::string &Constraint,
351                                              std::vector<SDValue> &Ops,
352                                              SelectionDAG &DAG) const;
353
354    const ARMSubtarget* getSubtarget() const {
355      return Subtarget;
356    }
357
358    /// getRegClassFor - Return the register class that should be used for the
359    /// specified value type.
360    virtual const TargetRegisterClass *getRegClassFor(MVT VT) const;
361
362    /// getMaximalGlobalOffset - Returns the maximal possible offset which can
363    /// be used for loads / stores from the global.
364    virtual unsigned getMaximalGlobalOffset() const;
365
366    /// Returns true if a cast between SrcAS and DestAS is a noop.
367    virtual bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const {
368      // Addrspacecasts are always noops.
369      return true;
370    }
371
372    /// createFastISel - This method returns a target specific FastISel object,
373    /// or null if the target does not support "fast" ISel.
374    virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
375                                     const TargetLibraryInfo *libInfo) const;
376
377    Sched::Preference getSchedulingPreference(SDNode *N) const;
378
379    bool isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const;
380    bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const;
381
382    /// isFPImmLegal - Returns true if the target can instruction select the
383    /// specified FP immediate natively. If false, the legalizer will
384    /// materialize the FP immediate as a load from a constant pool.
385    virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
386
387    virtual bool getTgtMemIntrinsic(IntrinsicInfo &Info,
388                                    const CallInst &I,
389                                    unsigned Intrinsic) const;
390  protected:
391    std::pair<const TargetRegisterClass*, uint8_t>
392    findRepresentativeClass(MVT VT) const;
393
394  private:
395    /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
396    /// make the right decision when generating code for different targets.
397    const ARMSubtarget *Subtarget;
398
399    const TargetRegisterInfo *RegInfo;
400
401    const InstrItineraryData *Itins;
402
403    /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created.
404    ///
405    unsigned ARMPCLabelIndex;
406
407    void addTypeForNEON(MVT VT, MVT PromotedLdStVT, MVT PromotedBitwiseVT);
408    void addDRTypeForNEON(MVT VT);
409    void addQRTypeForNEON(MVT VT);
410
411    typedef SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPassVector;
412    void PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
413                          SDValue Chain, SDValue &Arg,
414                          RegsToPassVector &RegsToPass,
415                          CCValAssign &VA, CCValAssign &NextVA,
416                          SDValue &StackPtr,
417                          SmallVectorImpl<SDValue> &MemOpChains,
418                          ISD::ArgFlagsTy Flags) const;
419    SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
420                                 SDValue &Root, SelectionDAG &DAG,
421                                 SDLoc dl) const;
422
423    CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return,
424                                  bool isVarArg) const;
425    SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
426                             SDLoc dl, SelectionDAG &DAG,
427                             const CCValAssign &VA,
428                             ISD::ArgFlagsTy Flags) const;
429    SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
430    SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
431    SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
432                                    const ARMSubtarget *Subtarget) const;
433    SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
434    SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const;
435    SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const;
436    SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
437    SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
438                                            SelectionDAG &DAG) const;
439    SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA,
440                                 SelectionDAG &DAG,
441                                 TLSModel::Model model) const;
442    SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const;
443    SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
444    SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
445    SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
446    SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
447    SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
448    SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
449    SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
450    SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
451    SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
452    SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
453    SDValue LowerConstantFP(SDValue Op, SelectionDAG &DAG,
454                            const ARMSubtarget *ST) const;
455    SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
456                              const ARMSubtarget *ST) const;
457    SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const;
458    SDValue LowerDivRem(SDValue Op, SelectionDAG &DAG) const;
459
460    /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
461    /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
462    /// expanded to FMAs when this method returns true, otherwise fmuladd is
463    /// expanded to fmul + fadd.
464    ///
465    /// ARM supports both fused and unfused multiply-add operations; we already
466    /// lower a pair of fmul and fadd to the latter so it's not clear that there
467    /// would be a gain or that the gain would be worthwhile enough to risk
468    /// correctness bugs.
469    virtual bool isFMAFasterThanFMulAndFAdd(EVT VT) const { return false; }
470
471    SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const;
472
473    SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
474                            CallingConv::ID CallConv, bool isVarArg,
475                            const SmallVectorImpl<ISD::InputArg> &Ins,
476                            SDLoc dl, SelectionDAG &DAG,
477                            SmallVectorImpl<SDValue> &InVals,
478                            bool isThisReturn, SDValue ThisVal) const;
479
480    virtual SDValue
481      LowerFormalArguments(SDValue Chain,
482                           CallingConv::ID CallConv, bool isVarArg,
483                           const SmallVectorImpl<ISD::InputArg> &Ins,
484                           SDLoc dl, SelectionDAG &DAG,
485                           SmallVectorImpl<SDValue> &InVals) const;
486
487    int StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
488                       SDLoc dl, SDValue &Chain,
489                       const Value *OrigArg,
490                       unsigned InRegsParamRecordIdx,
491                       unsigned OffsetFromOrigArg,
492                       unsigned ArgOffset,
493                       unsigned ArgSize,
494                       bool ForceMutable) const;
495
496    void VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
497                              SDLoc dl, SDValue &Chain,
498                              unsigned ArgOffset,
499                              bool ForceMutable = false) const;
500
501    void computeRegArea(CCState &CCInfo, MachineFunction &MF,
502                        unsigned InRegsParamRecordIdx,
503                        unsigned ArgSize,
504                        unsigned &ArgRegsSize,
505                        unsigned &ArgRegsSaveSize) const;
506
507    virtual SDValue
508      LowerCall(TargetLowering::CallLoweringInfo &CLI,
509                SmallVectorImpl<SDValue> &InVals) const;
510
511    /// HandleByVal - Target-specific cleanup for ByVal support.
512    virtual void HandleByVal(CCState *, unsigned &, unsigned) const;
513
514    /// IsEligibleForTailCallOptimization - Check whether the call is eligible
515    /// for tail call optimization. Targets which want to do tail call
516    /// optimization should implement this function.
517    bool IsEligibleForTailCallOptimization(SDValue Callee,
518                                           CallingConv::ID CalleeCC,
519                                           bool isVarArg,
520                                           bool isCalleeStructRet,
521                                           bool isCallerStructRet,
522                                    const SmallVectorImpl<ISD::OutputArg> &Outs,
523                                    const SmallVectorImpl<SDValue> &OutVals,
524                                    const SmallVectorImpl<ISD::InputArg> &Ins,
525                                           SelectionDAG& DAG) const;
526
527    virtual bool CanLowerReturn(CallingConv::ID CallConv,
528                                MachineFunction &MF, bool isVarArg,
529                                const SmallVectorImpl<ISD::OutputArg> &Outs,
530                                LLVMContext &Context) const;
531
532    virtual SDValue
533      LowerReturn(SDValue Chain,
534                  CallingConv::ID CallConv, bool isVarArg,
535                  const SmallVectorImpl<ISD::OutputArg> &Outs,
536                  const SmallVectorImpl<SDValue> &OutVals,
537                  SDLoc dl, SelectionDAG &DAG) const;
538
539    virtual bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const;
540
541    virtual bool mayBeEmittedAsTailCall(CallInst *CI) const;
542
543    SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
544                      SDValue &ARMcc, SelectionDAG &DAG, SDLoc dl) const;
545    SDValue getVFPCmp(SDValue LHS, SDValue RHS,
546                      SelectionDAG &DAG, SDLoc dl) const;
547    SDValue duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const;
548
549    SDValue OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const;
550
551    MachineBasicBlock *EmitAtomicCmpSwap(MachineInstr *MI,
552                                         MachineBasicBlock *BB,
553                                         unsigned Size) const;
554    MachineBasicBlock *EmitAtomicBinary(MachineInstr *MI,
555                                        MachineBasicBlock *BB,
556                                        unsigned Size,
557                                        unsigned BinOpcode) const;
558    MachineBasicBlock *EmitAtomicBinary64(MachineInstr *MI,
559                                          MachineBasicBlock *BB,
560                                          unsigned Op1,
561                                          unsigned Op2,
562                                          bool NeedsCarry = false,
563                                          bool IsCmpxchg = false,
564                                          bool IsMinMax = false,
565                                          ARMCC::CondCodes CC = ARMCC::AL) const;
566    MachineBasicBlock * EmitAtomicBinaryMinMax(MachineInstr *MI,
567                                               MachineBasicBlock *BB,
568                                               unsigned Size,
569                                               bool signExtend,
570                                               ARMCC::CondCodes Cond) const;
571    MachineBasicBlock *EmitAtomicLoad64(MachineInstr *MI,
572                                        MachineBasicBlock *BB) const;
573
574    void SetupEntryBlockForSjLj(MachineInstr *MI,
575                                MachineBasicBlock *MBB,
576                                MachineBasicBlock *DispatchBB, int FI) const;
577
578    MachineBasicBlock *EmitSjLjDispatchBlock(MachineInstr *MI,
579                                             MachineBasicBlock *MBB) const;
580
581    bool RemapAddSubWithFlags(MachineInstr *MI, MachineBasicBlock *BB) const;
582
583    MachineBasicBlock *EmitStructByval(MachineInstr *MI,
584                                       MachineBasicBlock *MBB) const;
585  };
586
587  enum NEONModImmType {
588    VMOVModImm,
589    VMVNModImm,
590    OtherModImm
591  };
592
593
594  namespace ARM {
595    FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
596                             const TargetLibraryInfo *libInfo);
597  }
598}
599
600#endif  // ARMISELLOWERING_H
601