ARMISelLowering.h revision 088880cb192fb6dd5b1bf85af62023c5ca3da38f
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      FTOSI,        // FP to sint within a FP register.
57      FTOUI,        // FP to uint within a FP register.
58      SITOF,        // sint to FP within a FP register.
59      UITOF,        // uint to FP within a FP register.
60
61      SRL_FLAG,     // V,Flag = srl_flag X -> srl X, 1 + save carry out.
62      SRA_FLAG,     // V,Flag = sra_flag X -> sra X, 1 + save carry out.
63      RRX,          // V = RRX X, Flag     -> srl X, 1 + shift in carry flag.
64
65      FMRRD,        // double to two gprs.
66      FMDRR,        // Two gprs to double.
67
68      EH_SJLJ_SETJMP,    // SjLj exception handling setjmp.
69      EH_SJLJ_LONGJMP,   // SjLj exception handling longjmp.
70
71      THREAD_POINTER,
72
73      DYN_ALLOC,    // Dynamic allocation on the stack.
74
75      VCEQ,         // Vector compare equal.
76      VCGE,         // Vector compare greater than or equal.
77      VCGEU,        // Vector compare unsigned greater than or equal.
78      VCGT,         // Vector compare greater than.
79      VCGTU,        // Vector compare unsigned greater than.
80      VTST,         // Vector test bits.
81
82      // Vector shift by immediate:
83      VSHL,         // ...left
84      VSHRs,        // ...right (signed)
85      VSHRu,        // ...right (unsigned)
86      VSHLLs,       // ...left long (signed)
87      VSHLLu,       // ...left long (unsigned)
88      VSHLLi,       // ...left long (with maximum shift count)
89      VSHRN,        // ...right narrow
90
91      // Vector rounding shift by immediate:
92      VRSHRs,       // ...right (signed)
93      VRSHRu,       // ...right (unsigned)
94      VRSHRN,       // ...right narrow
95
96      // Vector saturating shift by immediate:
97      VQSHLs,       // ...left (signed)
98      VQSHLu,       // ...left (unsigned)
99      VQSHLsu,      // ...left (signed to unsigned)
100      VQSHRNs,      // ...right narrow (signed)
101      VQSHRNu,      // ...right narrow (unsigned)
102      VQSHRNsu,     // ...right narrow (signed to unsigned)
103
104      // Vector saturating rounding shift by immediate:
105      VQRSHRNs,     // ...right narrow (signed)
106      VQRSHRNu,     // ...right narrow (unsigned)
107      VQRSHRNsu,    // ...right narrow (signed to unsigned)
108
109      // Vector shift and insert:
110      VSLI,         // ...left
111      VSRI,         // ...right
112
113      // Vector get lane (VMOV scalar to ARM core register)
114      // (These are used for 8- and 16-bit element types only.)
115      VGETLANEu,    // zero-extend vector extract element
116      VGETLANEs,    // sign-extend vector extract element
117
118      // Vector duplicate:
119      VDUP,
120      VDUPLANE,
121
122      // Vector load/store with (de)interleaving
123      VLD2D,
124      VLD3D,
125      VLD4D,
126      VST2D,
127      VST3D,
128      VST4D,
129
130      // Vector shuffles:
131      VREV64,       // reverse elements within 64-bit doublewords
132      VREV32,       // reverse elements within 32-bit words
133      VREV16        // reverse elements within 16-bit halfwords
134    };
135  }
136
137  /// Define some predicates that are used for node matching.
138  namespace ARM {
139    /// getVMOVImm - If this is a build_vector of constants which can be
140    /// formed by using a VMOV instruction of the specified element size,
141    /// return the constant being splatted.  The ByteSize field indicates the
142    /// number of bytes of each element [1248].
143    SDValue getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
144  }
145
146  //===--------------------------------------------------------------------===//
147  //  ARMTargetLowering - ARM Implementation of the TargetLowering interface
148
149  class ARMTargetLowering : public TargetLowering {
150    int VarArgsFrameIndex;            // FrameIndex for start of varargs area.
151  public:
152    explicit ARMTargetLowering(TargetMachine &TM);
153
154    virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG);
155
156    /// ReplaceNodeResults - Replace the results of node with an illegal result
157    /// type with new values built out of custom code.
158    ///
159    virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
160                                    SelectionDAG &DAG);
161
162    virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
163
164    virtual const char *getTargetNodeName(unsigned Opcode) const;
165
166    virtual MachineBasicBlock *EmitInstrWithCustomInserter(MachineInstr *MI,
167                                                  MachineBasicBlock *MBB) const;
168
169    /// allowsUnalignedMemoryAccesses - Returns true if the target allows
170    /// unaligned memory accesses. of the specified type.
171    /// FIXME: Add getOptimalMemOpType to implement memcpy with NEON?
172    virtual bool allowsUnalignedMemoryAccesses(EVT VT) const;
173
174    /// isLegalAddressingMode - Return true if the addressing mode represented
175    /// by AM is legal for this target, for a load/store of the specified type.
176    virtual bool isLegalAddressingMode(const AddrMode &AM, const Type *Ty)const;
177    bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const;
178
179    /// getPreIndexedAddressParts - returns true by value, base pointer and
180    /// offset pointer and addressing mode by reference if the node's address
181    /// can be legally represented as pre-indexed load / store address.
182    virtual bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
183                                           SDValue &Offset,
184                                           ISD::MemIndexedMode &AM,
185                                           SelectionDAG &DAG) const;
186
187    /// getPostIndexedAddressParts - returns true by value, base pointer and
188    /// offset pointer and addressing mode by reference if this node can be
189    /// combined with a load / store to form a post-indexed load / store.
190    virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op,
191                                            SDValue &Base, SDValue &Offset,
192                                            ISD::MemIndexedMode &AM,
193                                            SelectionDAG &DAG) const;
194
195    virtual void computeMaskedBitsForTargetNode(const SDValue Op,
196                                                const APInt &Mask,
197                                                APInt &KnownZero,
198                                                APInt &KnownOne,
199                                                const SelectionDAG &DAG,
200                                                unsigned Depth) const;
201
202
203    ConstraintType getConstraintType(const std::string &Constraint) const;
204    std::pair<unsigned, const TargetRegisterClass*>
205      getRegForInlineAsmConstraint(const std::string &Constraint,
206                                   EVT VT) const;
207    std::vector<unsigned>
208    getRegClassForInlineAsmConstraint(const std::string &Constraint,
209                                      EVT VT) const;
210
211    /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
212    /// vector.  If it is invalid, don't add anything to Ops. If hasMemory is
213    /// true it means one of the asm constraint of the inline asm instruction
214    /// being processed is 'm'.
215    virtual void LowerAsmOperandForConstraint(SDValue Op,
216                                              char ConstraintLetter,
217                                              bool hasMemory,
218                                              std::vector<SDValue> &Ops,
219                                              SelectionDAG &DAG) const;
220
221    virtual const ARMSubtarget* getSubtarget() {
222      return Subtarget;
223    }
224
225    /// getFunctionAlignment - Return the Log2 alignment of this function.
226    virtual unsigned getFunctionAlignment(const Function *F) const;
227
228  private:
229    /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
230    /// make the right decision when generating code for different targets.
231    const ARMSubtarget *Subtarget;
232
233    /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created.
234    ///
235    unsigned ARMPCLabelIndex;
236
237    void addTypeForNEON(EVT VT, EVT PromotedLdStVT, EVT PromotedBitwiseVT);
238    void addDRTypeForNEON(EVT VT);
239    void addQRTypeForNEON(EVT VT);
240
241    typedef SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPassVector;
242    void PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
243                          SDValue Chain, SDValue &Arg,
244                          RegsToPassVector &RegsToPass,
245                          CCValAssign &VA, CCValAssign &NextVA,
246                          SDValue &StackPtr,
247                          SmallVector<SDValue, 8> &MemOpChains,
248                          ISD::ArgFlagsTy Flags);
249    SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
250                                 SDValue &Root, SelectionDAG &DAG, DebugLoc dl);
251
252    CCAssignFn *CCAssignFnForNode(unsigned CC, bool Return, bool isVarArg) const;
253    SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
254                             DebugLoc dl, SelectionDAG &DAG,
255                             const CCValAssign &VA,
256                             ISD::ArgFlagsTy Flags);
257    SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG);
258    SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG);
259    SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG);
260    SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG);
261    SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG);
262    SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
263                                            SelectionDAG &DAG);
264    SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA,
265                                   SelectionDAG &DAG);
266    SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG);
267    SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG);
268    SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG);
269    SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG);
270
271    SDValue EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
272                                      SDValue Chain,
273                                      SDValue Dst, SDValue Src,
274                                      SDValue Size, unsigned Align,
275                                      bool AlwaysInline,
276                                      const Value *DstSV, uint64_t DstSVOff,
277                                      const Value *SrcSV, uint64_t SrcSVOff);
278    SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
279                            unsigned CallConv, bool isVarArg,
280                            const SmallVectorImpl<ISD::InputArg> &Ins,
281                            DebugLoc dl, SelectionDAG &DAG,
282                            SmallVectorImpl<SDValue> &InVals);
283
284    virtual SDValue
285      LowerFormalArguments(SDValue Chain,
286                           unsigned CallConv, bool isVarArg,
287                           const SmallVectorImpl<ISD::InputArg> &Ins,
288                           DebugLoc dl, SelectionDAG &DAG,
289                           SmallVectorImpl<SDValue> &InVals);
290
291    virtual SDValue
292      LowerCall(SDValue Chain, SDValue Callee,
293                unsigned CallConv, bool isVarArg,
294                bool isTailCall,
295                const SmallVectorImpl<ISD::OutputArg> &Outs,
296                const SmallVectorImpl<ISD::InputArg> &Ins,
297                DebugLoc dl, SelectionDAG &DAG,
298                SmallVectorImpl<SDValue> &InVals);
299
300    virtual SDValue
301      LowerReturn(SDValue Chain,
302                  unsigned CallConv, bool isVarArg,
303                  const SmallVectorImpl<ISD::OutputArg> &Outs,
304                  DebugLoc dl, SelectionDAG &DAG);
305  };
306}
307
308#endif  // ARMISELLOWERING_H
309