1255365Sdes//===-- X86InstrInfo.h - X86 Instruction Information ------------*- C++ -*-===//
2255365Sdes//
3255365Sdes//                     The LLVM Compiler Infrastructure
4255365Sdes//
5255365Sdes// This file is distributed under the University of Illinois Open Source
6255365Sdes// License. See LICENSE.TXT for details.
7255365Sdes//
8255365Sdes//===----------------------------------------------------------------------===//
9255365Sdes//
10255365Sdes// This file contains the X86 implementation of the TargetInstrInfo class.
11255365Sdes//
12255365Sdes//===----------------------------------------------------------------------===//
13255365Sdes
14255365Sdes#ifndef X86INSTRUCTIONINFO_H
15255365Sdes#define X86INSTRUCTIONINFO_H
16255365Sdes
17255365Sdes#include "X86.h"
18255365Sdes#include "X86RegisterInfo.h"
19255365Sdes#include "llvm/ADT/DenseMap.h"
20255365Sdes#include "llvm/Target/TargetInstrInfo.h"
21255365Sdes
22255365Sdes#define GET_INSTRINFO_HEADER
23255365Sdes#include "X86GenInstrInfo.inc"
24255365Sdes
25255365Sdesnamespace llvm {
26255365Sdes  class X86RegisterInfo;
27255365Sdes  class X86TargetMachine;
28255365Sdes
29255365Sdesnamespace X86 {
30255365Sdes  // X86 specific condition code. These correspond to X86_*_COND in
31255365Sdes  // X86InstrInfo.td. They must be kept in synch.
32255365Sdes  enum CondCode {
33255365Sdes    COND_A  = 0,
34255365Sdes    COND_AE = 1,
35255365Sdes    COND_B  = 2,
36255365Sdes    COND_BE = 3,
37255365Sdes    COND_E  = 4,
38255365Sdes    COND_G  = 5,
39255365Sdes    COND_GE = 6,
40255365Sdes    COND_L  = 7,
41255365Sdes    COND_LE = 8,
42255365Sdes    COND_NE = 9,
43255365Sdes    COND_NO = 10,
44255365Sdes    COND_NP = 11,
45255365Sdes    COND_NS = 12,
46255365Sdes    COND_O  = 13,
47255365Sdes    COND_P  = 14,
48255365Sdes    COND_S  = 15,
49255365Sdes
50255365Sdes    // Artificial condition codes. These are used by AnalyzeBranch
51255365Sdes    // to indicate a block terminated with two conditional branches to
52255365Sdes    // the same location. This occurs in code using FCMP_OEQ or FCMP_UNE,
53255365Sdes    // which can't be represented on x86 with a single condition. These
54255365Sdes    // are never used in MachineInstrs.
55255365Sdes    COND_NE_OR_P,
56255365Sdes    COND_NP_OR_E,
57255365Sdes
58255365Sdes    COND_INVALID
59255365Sdes  };
60255365Sdes
61255365Sdes  // Turn condition code into conditional branch opcode.
62255365Sdes  unsigned GetCondBranchFromCond(CondCode CC);
63255365Sdes
64255365Sdes  // Turn CMov opcode into condition code.
65255365Sdes  CondCode getCondFromCMovOpc(unsigned Opc);
66255365Sdes
67255365Sdes  /// GetOppositeBranchCondition - Return the inverse of the specified cond,
68255365Sdes  /// e.g. turning COND_E to COND_NE.
69255365Sdes  CondCode GetOppositeBranchCondition(X86::CondCode CC);
70255365Sdes}  // end namespace X86;
71255365Sdes
72255365Sdes
73255365Sdes/// isGlobalStubReference - Return true if the specified TargetFlag operand is
74255365Sdes/// a reference to a stub for a global, not the global itself.
75255365Sdesinline static bool isGlobalStubReference(unsigned char TargetFlag) {
76255365Sdes  switch (TargetFlag) {
77255365Sdes  case X86II::MO_DLLIMPORT: // dllimport stub.
78255365Sdes  case X86II::MO_GOTPCREL:  // rip-relative GOT reference.
79255365Sdes  case X86II::MO_GOT:       // normal GOT reference.
80255365Sdes  case X86II::MO_DARWIN_NONLAZY_PIC_BASE:        // Normal $non_lazy_ptr ref.
81255365Sdes  case X86II::MO_DARWIN_NONLAZY:                 // Normal $non_lazy_ptr ref.
82255365Sdes  case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE: // Hidden $non_lazy_ptr ref.
83255365Sdes    return true;
84255365Sdes  default:
85255365Sdes    return false;
86255365Sdes  }
87255365Sdes}
88255365Sdes
89255365Sdes/// isGlobalRelativeToPICBase - Return true if the specified global value
90255365Sdes/// reference is relative to a 32-bit PIC base (X86ISD::GlobalBaseReg).  If this
91255365Sdes/// is true, the addressing mode has the PIC base register added in (e.g. EBX).
92255365Sdesinline static bool isGlobalRelativeToPICBase(unsigned char TargetFlag) {
93255365Sdes  switch (TargetFlag) {
94255365Sdes  case X86II::MO_GOTOFF:                         // isPICStyleGOT: local global.
95255365Sdes  case X86II::MO_GOT:                            // isPICStyleGOT: other global.
96255365Sdes  case X86II::MO_PIC_BASE_OFFSET:                // Darwin local global.
97255365Sdes  case X86II::MO_DARWIN_NONLAZY_PIC_BASE:        // Darwin/32 external global.
98255365Sdes  case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE: // Darwin/32 hidden global.
99  case X86II::MO_TLVP:                           // ??? Pretty sure..
100    return true;
101  default:
102    return false;
103  }
104}
105
106inline static bool isScale(const MachineOperand &MO) {
107  return MO.isImm() &&
108    (MO.getImm() == 1 || MO.getImm() == 2 ||
109     MO.getImm() == 4 || MO.getImm() == 8);
110}
111
112inline static bool isLeaMem(const MachineInstr *MI, unsigned Op) {
113  if (MI->getOperand(Op).isFI()) return true;
114  return Op+4 <= MI->getNumOperands() &&
115    MI->getOperand(Op  ).isReg() && isScale(MI->getOperand(Op+1)) &&
116    MI->getOperand(Op+2).isReg() &&
117    (MI->getOperand(Op+3).isImm() ||
118     MI->getOperand(Op+3).isGlobal() ||
119     MI->getOperand(Op+3).isCPI() ||
120     MI->getOperand(Op+3).isJTI());
121}
122
123inline static bool isMem(const MachineInstr *MI, unsigned Op) {
124  if (MI->getOperand(Op).isFI()) return true;
125  return Op+5 <= MI->getNumOperands() &&
126    MI->getOperand(Op+4).isReg() &&
127    isLeaMem(MI, Op);
128}
129
130class X86InstrInfo : public X86GenInstrInfo {
131  X86TargetMachine &TM;
132  const X86RegisterInfo RI;
133
134  /// RegOp2MemOpTable3Addr, RegOp2MemOpTable0, RegOp2MemOpTable1,
135  /// RegOp2MemOpTable2, RegOp2MemOpTable3 - Load / store folding opcode maps.
136  ///
137  typedef DenseMap<unsigned,
138                   std::pair<unsigned, unsigned> > RegOp2MemOpTableType;
139  RegOp2MemOpTableType RegOp2MemOpTable2Addr;
140  RegOp2MemOpTableType RegOp2MemOpTable0;
141  RegOp2MemOpTableType RegOp2MemOpTable1;
142  RegOp2MemOpTableType RegOp2MemOpTable2;
143  RegOp2MemOpTableType RegOp2MemOpTable3;
144
145  /// MemOp2RegOpTable - Load / store unfolding opcode map.
146  ///
147  typedef DenseMap<unsigned,
148                   std::pair<unsigned, unsigned> > MemOp2RegOpTableType;
149  MemOp2RegOpTableType MemOp2RegOpTable;
150
151  static void AddTableEntry(RegOp2MemOpTableType &R2MTable,
152                            MemOp2RegOpTableType &M2RTable,
153                            unsigned RegOp, unsigned MemOp, unsigned Flags);
154
155public:
156  explicit X86InstrInfo(X86TargetMachine &tm);
157
158  /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info.  As
159  /// such, whenever a client has an instance of instruction info, it should
160  /// always be able to get register info as well (through this method).
161  ///
162  virtual const X86RegisterInfo &getRegisterInfo() const { return RI; }
163
164  /// isCoalescableExtInstr - Return true if the instruction is a "coalescable"
165  /// extension instruction. That is, it's like a copy where it's legal for the
166  /// source to overlap the destination. e.g. X86::MOVSX64rr32. If this returns
167  /// true, then it's expected the pre-extension value is available as a subreg
168  /// of the result register. This also returns the sub-register index in
169  /// SubIdx.
170  virtual bool isCoalescableExtInstr(const MachineInstr &MI,
171                                     unsigned &SrcReg, unsigned &DstReg,
172                                     unsigned &SubIdx) const;
173
174  unsigned isLoadFromStackSlot(const MachineInstr *MI, int &FrameIndex) const;
175  /// isLoadFromStackSlotPostFE - Check for post-frame ptr elimination
176  /// stack locations as well.  This uses a heuristic so it isn't
177  /// reliable for correctness.
178  unsigned isLoadFromStackSlotPostFE(const MachineInstr *MI,
179                                     int &FrameIndex) const;
180
181  unsigned isStoreToStackSlot(const MachineInstr *MI, int &FrameIndex) const;
182  /// isStoreToStackSlotPostFE - Check for post-frame ptr elimination
183  /// stack locations as well.  This uses a heuristic so it isn't
184  /// reliable for correctness.
185  unsigned isStoreToStackSlotPostFE(const MachineInstr *MI,
186                                    int &FrameIndex) const;
187
188  bool isReallyTriviallyReMaterializable(const MachineInstr *MI,
189                                         AliasAnalysis *AA) const;
190  void reMaterialize(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
191                     unsigned DestReg, unsigned SubIdx,
192                     const MachineInstr *Orig,
193                     const TargetRegisterInfo &TRI) const;
194
195  /// convertToThreeAddress - This method must be implemented by targets that
196  /// set the M_CONVERTIBLE_TO_3_ADDR flag.  When this flag is set, the target
197  /// may be able to convert a two-address instruction into a true
198  /// three-address instruction on demand.  This allows the X86 target (for
199  /// example) to convert ADD and SHL instructions into LEA instructions if they
200  /// would require register copies due to two-addressness.
201  ///
202  /// This method returns a null pointer if the transformation cannot be
203  /// performed, otherwise it returns the new instruction.
204  ///
205  virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI,
206                                              MachineBasicBlock::iterator &MBBI,
207                                              LiveVariables *LV) const;
208
209  /// commuteInstruction - We have a few instructions that must be hacked on to
210  /// commute them.
211  ///
212  virtual MachineInstr *commuteInstruction(MachineInstr *MI, bool NewMI) const;
213
214  // Branch analysis.
215  virtual bool isUnpredicatedTerminator(const MachineInstr* MI) const;
216  virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
217                             MachineBasicBlock *&FBB,
218                             SmallVectorImpl<MachineOperand> &Cond,
219                             bool AllowModify) const;
220  virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const;
221  virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
222                                MachineBasicBlock *FBB,
223                                const SmallVectorImpl<MachineOperand> &Cond,
224                                DebugLoc DL) const;
225  virtual bool canInsertSelect(const MachineBasicBlock&,
226                               const SmallVectorImpl<MachineOperand> &Cond,
227                               unsigned, unsigned, int&, int&, int&) const;
228  virtual void insertSelect(MachineBasicBlock &MBB,
229                            MachineBasicBlock::iterator MI, DebugLoc DL,
230                            unsigned DstReg,
231                            const SmallVectorImpl<MachineOperand> &Cond,
232                            unsigned TrueReg, unsigned FalseReg) const;
233  virtual void copyPhysReg(MachineBasicBlock &MBB,
234                           MachineBasicBlock::iterator MI, DebugLoc DL,
235                           unsigned DestReg, unsigned SrcReg,
236                           bool KillSrc) const;
237  virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
238                                   MachineBasicBlock::iterator MI,
239                                   unsigned SrcReg, bool isKill, int FrameIndex,
240                                   const TargetRegisterClass *RC,
241                                   const TargetRegisterInfo *TRI) const;
242
243  virtual void storeRegToAddr(MachineFunction &MF, unsigned SrcReg, bool isKill,
244                              SmallVectorImpl<MachineOperand> &Addr,
245                              const TargetRegisterClass *RC,
246                              MachineInstr::mmo_iterator MMOBegin,
247                              MachineInstr::mmo_iterator MMOEnd,
248                              SmallVectorImpl<MachineInstr*> &NewMIs) const;
249
250  virtual void loadRegFromStackSlot(MachineBasicBlock &MBB,
251                                    MachineBasicBlock::iterator MI,
252                                    unsigned DestReg, int FrameIndex,
253                                    const TargetRegisterClass *RC,
254                                    const TargetRegisterInfo *TRI) const;
255
256  virtual void loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
257                               SmallVectorImpl<MachineOperand> &Addr,
258                               const TargetRegisterClass *RC,
259                               MachineInstr::mmo_iterator MMOBegin,
260                               MachineInstr::mmo_iterator MMOEnd,
261                               SmallVectorImpl<MachineInstr*> &NewMIs) const;
262
263  virtual bool expandPostRAPseudo(MachineBasicBlock::iterator MI) const;
264
265  virtual
266  MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF,
267                                         int FrameIx, uint64_t Offset,
268                                         const MDNode *MDPtr,
269                                         DebugLoc DL) const;
270
271  /// foldMemoryOperand - If this target supports it, fold a load or store of
272  /// the specified stack slot into the specified machine instruction for the
273  /// specified operand(s).  If this is possible, the target should perform the
274  /// folding and return true, otherwise it should return false.  If it folds
275  /// the instruction, it is likely that the MachineInstruction the iterator
276  /// references has been changed.
277  virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
278                                              MachineInstr* MI,
279                                           const SmallVectorImpl<unsigned> &Ops,
280                                              int FrameIndex) const;
281
282  /// foldMemoryOperand - Same as the previous version except it allows folding
283  /// of any load and store from / to any address, not just from a specific
284  /// stack slot.
285  virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
286                                              MachineInstr* MI,
287                                           const SmallVectorImpl<unsigned> &Ops,
288                                              MachineInstr* LoadMI) const;
289
290  /// canFoldMemoryOperand - Returns true if the specified load / store is
291  /// folding is possible.
292  virtual bool canFoldMemoryOperand(const MachineInstr*,
293                                    const SmallVectorImpl<unsigned> &) const;
294
295  /// unfoldMemoryOperand - Separate a single instruction which folded a load or
296  /// a store or a load and a store into two or more instruction. If this is
297  /// possible, returns true as well as the new instructions by reference.
298  virtual bool unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
299                           unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
300                           SmallVectorImpl<MachineInstr*> &NewMIs) const;
301
302  virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
303                           SmallVectorImpl<SDNode*> &NewNodes) const;
304
305  /// getOpcodeAfterMemoryUnfold - Returns the opcode of the would be new
306  /// instruction after load / store are unfolded from an instruction of the
307  /// specified opcode. It returns zero if the specified unfolding is not
308  /// possible. If LoadRegIndex is non-null, it is filled in with the operand
309  /// index of the operand which will hold the register holding the loaded
310  /// value.
311  virtual unsigned getOpcodeAfterMemoryUnfold(unsigned Opc,
312                                      bool UnfoldLoad, bool UnfoldStore,
313                                      unsigned *LoadRegIndex = 0) const;
314
315  /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler
316  /// to determine if two loads are loading from the same base address. It
317  /// should only return true if the base pointers are the same and the
318  /// only differences between the two addresses are the offset. It also returns
319  /// the offsets by reference.
320  virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
321                                       int64_t &Offset1, int64_t &Offset2) const;
322
323  /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
324  /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
325  /// be scheduled togther. On some targets if two loads are loading from
326  /// addresses in the same cache line, it's better if they are scheduled
327  /// together. This function takes two integers that represent the load offsets
328  /// from the common base address. It returns true if it decides it's desirable
329  /// to schedule the two loads together. "NumLoads" is the number of loads that
330  /// have already been scheduled after Load1.
331  virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
332                                       int64_t Offset1, int64_t Offset2,
333                                       unsigned NumLoads) const;
334
335  virtual void getNoopForMachoTarget(MCInst &NopInst) const;
336
337  virtual
338  bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const;
339
340  /// isSafeToMoveRegClassDefs - Return true if it's safe to move a machine
341  /// instruction that defines the specified register class.
342  bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const;
343
344  static bool isX86_64ExtendedReg(const MachineOperand &MO) {
345    if (!MO.isReg()) return false;
346    return X86II::isX86_64ExtendedReg(MO.getReg());
347  }
348
349  /// getGlobalBaseReg - Return a virtual register initialized with the
350  /// the global base register value. Output instructions required to
351  /// initialize the register in the function entry block, if necessary.
352  ///
353  unsigned getGlobalBaseReg(MachineFunction *MF) const;
354
355  std::pair<uint16_t, uint16_t>
356  getExecutionDomain(const MachineInstr *MI) const;
357
358  void setExecutionDomain(MachineInstr *MI, unsigned Domain) const;
359
360  unsigned getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
361                                        const TargetRegisterInfo *TRI) const;
362  void breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
363                                 const TargetRegisterInfo *TRI) const;
364
365  MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
366                                      MachineInstr* MI,
367                                      unsigned OpNum,
368                                      const SmallVectorImpl<MachineOperand> &MOs,
369                                      unsigned Size, unsigned Alignment) const;
370
371  bool isHighLatencyDef(int opc) const;
372
373  bool hasHighOperandLatency(const InstrItineraryData *ItinData,
374                             const MachineRegisterInfo *MRI,
375                             const MachineInstr *DefMI, unsigned DefIdx,
376                             const MachineInstr *UseMI, unsigned UseIdx) const;
377
378  /// analyzeCompare - For a comparison instruction, return the source registers
379  /// in SrcReg and SrcReg2 if having two register operands, and the value it
380  /// compares against in CmpValue. Return true if the comparison instruction
381  /// can be analyzed.
382  virtual bool analyzeCompare(const MachineInstr *MI, unsigned &SrcReg,
383                              unsigned &SrcReg2,
384                              int &CmpMask, int &CmpValue) const;
385
386  /// optimizeCompareInstr - Check if there exists an earlier instruction that
387  /// operates on the same source operands and sets flags in the same way as
388  /// Compare; remove Compare if possible.
389  virtual bool optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg,
390                                    unsigned SrcReg2, int CmpMask, int CmpValue,
391                                    const MachineRegisterInfo *MRI) const;
392
393  /// optimizeLoadInstr - Try to remove the load by folding it to a register
394  /// operand at the use. We fold the load instructions if and only if the
395  /// def and use are in the same BB. We only look at one load and see
396  /// whether it can be folded into MI. FoldAsLoadDefReg is the virtual register
397  /// defined by the load we are trying to fold. DefMI returns the machine
398  /// instruction that defines FoldAsLoadDefReg, and the function returns
399  /// the machine instruction generated due to folding.
400  virtual MachineInstr* optimizeLoadInstr(MachineInstr *MI,
401                        const MachineRegisterInfo *MRI,
402                        unsigned &FoldAsLoadDefReg,
403                        MachineInstr *&DefMI) const;
404
405private:
406  MachineInstr * convertToThreeAddressWithLEA(unsigned MIOpc,
407                                              MachineFunction::iterator &MFI,
408                                              MachineBasicBlock::iterator &MBBI,
409                                              LiveVariables *LV) const;
410
411  /// isFrameOperand - Return true and the FrameIndex if the specified
412  /// operand and follow operands form a reference to the stack frame.
413  bool isFrameOperand(const MachineInstr *MI, unsigned int Op,
414                      int &FrameIndex) const;
415};
416
417} // End llvm namespace
418
419#endif
420