ARMISelLowering.h revision 23ff7cff52702a8bff904d8ab4c9ca67cc19d6ca
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 "ARMSubtarget.h"
19#include "llvm/Target/TargetLowering.h"
20#include "llvm/CodeGen/SelectionDAG.h"
21#include "llvm/CodeGen/CallingConvLower.h"
22#include <vector>
23
24namespace llvm {
25  class ARMConstantPoolValue;
26
27  namespace ARMISD {
28    // ARM Specific DAG Nodes
29    enum NodeType {
30      // Start the numbering where the builtin ops and target ops leave off.
31      FIRST_NUMBER = ISD::BUILTIN_OP_END,
32
33      Wrapper,      // Wrapper - A wrapper node for TargetConstantPool,
34                    // TargetExternalSymbol, and TargetGlobalAddress.
35      WrapperJT,    // WrapperJT - A wrapper node for TargetJumpTable
36
37      CALL,         // Function call.
38      CALL_PRED,    // Function call that's predicable.
39      CALL_NOLINK,  // Function call with branch not branch-and-link.
40      tCALL,        // Thumb function call.
41      BRCOND,       // Conditional branch.
42      BR_JT,        // Jumptable branch.
43      BR2_JT,       // Jumptable branch (2 level - jumptable entry is a jump).
44      RET_FLAG,     // Return with a flag operand.
45
46      PIC_ADD,      // Add with a PC operand and a PIC label.
47
48      CMP,          // ARM compare instructions.
49      CMPZ,         // ARM compare that sets only Z flag.
50      CMPFP,        // ARM VFP compare instruction, sets FPSCR.
51      CMPFPw0,      // ARM VFP compare against zero instruction, sets FPSCR.
52      FMSTAT,       // ARM fmstat instruction.
53      CMOV,         // ARM conditional move instructions.
54      CNEG,         // ARM conditional negate instructions.
55
56      RBIT,         // ARM bitreverse instruction
57
58      FTOSI,        // FP to sint within a FP register.
59      FTOUI,        // FP to uint within a FP register.
60      SITOF,        // sint to FP within a FP register.
61      UITOF,        // uint to FP within a FP register.
62
63      SRL_FLAG,     // V,Flag = srl_flag X -> srl X, 1 + save carry out.
64      SRA_FLAG,     // V,Flag = sra_flag X -> sra X, 1 + save carry out.
65      RRX,          // V = RRX X, Flag     -> srl X, 1 + shift in carry flag.
66
67      VMOVRRD,      // double to two gprs.
68      VMOVDRR,      // Two gprs to double.
69
70      EH_SJLJ_SETJMP,    // SjLj exception handling setjmp.
71      EH_SJLJ_LONGJMP,   // SjLj exception handling longjmp.
72
73      THREAD_POINTER,
74
75      DYN_ALLOC,    // Dynamic allocation on the stack.
76
77      MEMBARRIER,   // Memory barrier
78      SYNCBARRIER,  // Memory sync barrier
79
80      VCEQ,         // Vector compare equal.
81      VCGE,         // Vector compare greater than or equal.
82      VCGEU,        // Vector compare unsigned greater than or equal.
83      VCGT,         // Vector compare greater than.
84      VCGTU,        // Vector compare unsigned greater than.
85      VTST,         // Vector test bits.
86
87      // Vector shift by immediate:
88      VSHL,         // ...left
89      VSHRs,        // ...right (signed)
90      VSHRu,        // ...right (unsigned)
91      VSHLLs,       // ...left long (signed)
92      VSHLLu,       // ...left long (unsigned)
93      VSHLLi,       // ...left long (with maximum shift count)
94      VSHRN,        // ...right narrow
95
96      // Vector rounding shift by immediate:
97      VRSHRs,       // ...right (signed)
98      VRSHRu,       // ...right (unsigned)
99      VRSHRN,       // ...right narrow
100
101      // Vector saturating shift by immediate:
102      VQSHLs,       // ...left (signed)
103      VQSHLu,       // ...left (unsigned)
104      VQSHLsu,      // ...left (signed to unsigned)
105      VQSHRNs,      // ...right narrow (signed)
106      VQSHRNu,      // ...right narrow (unsigned)
107      VQSHRNsu,     // ...right narrow (signed to unsigned)
108
109      // Vector saturating rounding shift by immediate:
110      VQRSHRNs,     // ...right narrow (signed)
111      VQRSHRNu,     // ...right narrow (unsigned)
112      VQRSHRNsu,    // ...right narrow (signed to unsigned)
113
114      // Vector shift and insert:
115      VSLI,         // ...left
116      VSRI,         // ...right
117
118      // Vector get lane (VMOV scalar to ARM core register)
119      // (These are used for 8- and 16-bit element types only.)
120      VGETLANEu,    // zero-extend vector extract element
121      VGETLANEs,    // sign-extend vector extract element
122
123      // Vector duplicate:
124      VDUP,
125      VDUPLANE,
126
127      // Vector shuffles:
128      VEXT,         // extract
129      VREV64,       // reverse elements within 64-bit doublewords
130      VREV32,       // reverse elements within 32-bit words
131      VREV16,       // reverse elements within 16-bit halfwords
132      VZIP,         // zip (interleave)
133      VUZP,         // unzip (deinterleave)
134      VTRN,         // transpose
135
136      // Floating-point max and min:
137      FMAX,
138      FMIN
139    };
140  }
141
142  /// Define some predicates that are used for node matching.
143  namespace ARM {
144    /// getVMOVImm - If this is a build_vector of constants which can be
145    /// formed by using a VMOV instruction of the specified element size,
146    /// return the constant being splatted.  The ByteSize field indicates the
147    /// number of bytes of each element [1248].
148    SDValue getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
149
150    /// getVFPf32Imm / getVFPf64Imm - If the given fp immediate can be
151    /// materialized with a VMOV.f32 / VMOV.f64 (i.e. fconsts / fconstd)
152    /// instruction, returns its 8-bit integer representation. Otherwise,
153    /// returns -1.
154    int getVFPf32Imm(const APFloat &FPImm);
155    int getVFPf64Imm(const APFloat &FPImm);
156  }
157
158  //===--------------------------------------------------------------------===//
159  //  ARMTargetLowering - ARM Implementation of the TargetLowering interface
160
161  class ARMTargetLowering : public TargetLowering {
162  public:
163    explicit ARMTargetLowering(TargetMachine &TM);
164
165    virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
166
167    /// ReplaceNodeResults - Replace the results of node with an illegal result
168    /// type with new values built out of custom code.
169    ///
170    virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
171                                    SelectionDAG &DAG) const;
172
173    virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
174
175    virtual const char *getTargetNodeName(unsigned Opcode) const;
176
177    virtual MachineBasicBlock *
178      EmitInstrWithCustomInserter(MachineInstr *MI,
179                                  MachineBasicBlock *MBB) const;
180
181    /// allowsUnalignedMemoryAccesses - Returns true if the target allows
182    /// unaligned memory accesses. of the specified type.
183    /// FIXME: Add getOptimalMemOpType to implement memcpy with NEON?
184    virtual bool allowsUnalignedMemoryAccesses(EVT VT) const;
185
186    /// isLegalAddressingMode - Return true if the addressing mode represented
187    /// by AM is legal for this target, for a load/store of the specified type.
188    virtual bool isLegalAddressingMode(const AddrMode &AM, const Type *Ty)const;
189    bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const;
190
191    /// isLegalICmpImmediate - Return true if the specified immediate is legal
192    /// icmp immediate, that is the target has icmp instructions which can compare
193    /// a register against the immediate without having to materialize the
194    /// immediate into a register.
195    virtual bool isLegalICmpImmediate(int64_t Imm) const;
196
197    /// getPreIndexedAddressParts - returns true by value, base pointer and
198    /// offset pointer and addressing mode by reference if the node's address
199    /// can be legally represented as pre-indexed load / store address.
200    virtual bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
201                                           SDValue &Offset,
202                                           ISD::MemIndexedMode &AM,
203                                           SelectionDAG &DAG) const;
204
205    /// getPostIndexedAddressParts - returns true by value, base pointer and
206    /// offset pointer and addressing mode by reference if this node can be
207    /// combined with a load / store to form a post-indexed load / store.
208    virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op,
209                                            SDValue &Base, SDValue &Offset,
210                                            ISD::MemIndexedMode &AM,
211                                            SelectionDAG &DAG) const;
212
213    virtual void computeMaskedBitsForTargetNode(const SDValue Op,
214                                                const APInt &Mask,
215                                                APInt &KnownZero,
216                                                APInt &KnownOne,
217                                                const SelectionDAG &DAG,
218                                                unsigned Depth) const;
219
220
221    ConstraintType getConstraintType(const std::string &Constraint) const;
222    std::pair<unsigned, const TargetRegisterClass*>
223      getRegForInlineAsmConstraint(const std::string &Constraint,
224                                   EVT VT) const;
225    std::vector<unsigned>
226    getRegClassForInlineAsmConstraint(const std::string &Constraint,
227                                      EVT VT) const;
228
229    /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
230    /// vector.  If it is invalid, don't add anything to Ops. If hasMemory is
231    /// true it means one of the asm constraint of the inline asm instruction
232    /// being processed is 'm'.
233    virtual void LowerAsmOperandForConstraint(SDValue Op,
234                                              char ConstraintLetter,
235                                              bool hasMemory,
236                                              std::vector<SDValue> &Ops,
237                                              SelectionDAG &DAG) const;
238
239    const ARMSubtarget* getSubtarget() const {
240      return Subtarget;
241    }
242
243    /// getRegClassFor - Return the register class that should be used for the
244    /// specified value type.
245    virtual TargetRegisterClass *getRegClassFor(EVT VT) const;
246
247    /// getFunctionAlignment - Return the Log2 alignment of this function.
248    virtual unsigned getFunctionAlignment(const Function *F) const;
249
250    Sched::Preference getSchedulingPreference(SDNode *N) const;
251
252    bool isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const;
253    bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const;
254
255    /// isFPImmLegal - Returns true if the target can instruction select the
256    /// specified FP immediate natively. If false, the legalizer will
257    /// materialize the FP immediate as a load from a constant pool.
258    virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
259
260  private:
261    /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
262    /// make the right decision when generating code for different targets.
263    const ARMSubtarget *Subtarget;
264
265    /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created.
266    ///
267    unsigned ARMPCLabelIndex;
268
269    void addTypeForNEON(EVT VT, EVT PromotedLdStVT, EVT PromotedBitwiseVT);
270    void addDRTypeForNEON(EVT VT);
271    void addQRTypeForNEON(EVT VT);
272
273    typedef SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPassVector;
274    void PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
275                          SDValue Chain, SDValue &Arg,
276                          RegsToPassVector &RegsToPass,
277                          CCValAssign &VA, CCValAssign &NextVA,
278                          SDValue &StackPtr,
279                          SmallVector<SDValue, 8> &MemOpChains,
280                          ISD::ArgFlagsTy Flags) const;
281    SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
282                                 SDValue &Root, SelectionDAG &DAG,
283                                 DebugLoc dl) const;
284
285    CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return, bool isVarArg) const;
286    SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
287                             DebugLoc dl, SelectionDAG &DAG,
288                             const CCValAssign &VA,
289                             ISD::ArgFlagsTy Flags) const;
290    SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
291    SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
292    SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
293                                    const ARMSubtarget *Subtarget) const;
294    SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
295    SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const;
296    SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const;
297    SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
298    SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
299                                            SelectionDAG &DAG) const;
300    SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA,
301                                   SelectionDAG &DAG) const;
302    SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const;
303    SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
304    SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
305    SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
306    SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
307    SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
308    SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
309    SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
310    SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
311
312    SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
313                            CallingConv::ID CallConv, bool isVarArg,
314                            const SmallVectorImpl<ISD::InputArg> &Ins,
315                            DebugLoc dl, SelectionDAG &DAG,
316                            SmallVectorImpl<SDValue> &InVals) const;
317
318    virtual SDValue
319      LowerFormalArguments(SDValue Chain,
320                           CallingConv::ID CallConv, bool isVarArg,
321                           const SmallVectorImpl<ISD::InputArg> &Ins,
322                           DebugLoc dl, SelectionDAG &DAG,
323                           SmallVectorImpl<SDValue> &InVals) const;
324
325    virtual SDValue
326      LowerCall(SDValue Chain, SDValue Callee,
327                CallingConv::ID CallConv, bool isVarArg,
328                bool &isTailCall,
329                const SmallVectorImpl<ISD::OutputArg> &Outs,
330                const SmallVectorImpl<ISD::InputArg> &Ins,
331                DebugLoc dl, SelectionDAG &DAG,
332                SmallVectorImpl<SDValue> &InVals) const;
333
334    virtual SDValue
335      LowerReturn(SDValue Chain,
336                  CallingConv::ID CallConv, bool isVarArg,
337                  const SmallVectorImpl<ISD::OutputArg> &Outs,
338                  DebugLoc dl, SelectionDAG &DAG) const;
339
340    SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
341                      SDValue &ARMCC, SelectionDAG &DAG, DebugLoc dl) const;
342
343    MachineBasicBlock *EmitAtomicCmpSwap(MachineInstr *MI,
344                                         MachineBasicBlock *BB,
345                                         unsigned Size) const;
346    MachineBasicBlock *EmitAtomicBinary(MachineInstr *MI,
347                                        MachineBasicBlock *BB,
348                                        unsigned Size,
349                                        unsigned BinOpcode) const;
350
351  };
352}
353
354#endif  // ARMISELLOWERING_H
355