X86InstrInfo.h revision 177cf1e1a3685209ab805f82897902a8d2b61661
1//===-- X86InstrInfo.h - X86 Instruction Information ------------*- 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 contains the X86 implementation of the TargetInstrInfo class.
11//
12//===----------------------------------------------------------------------===//
13
14#ifndef X86INSTRUCTIONINFO_H
15#define X86INSTRUCTIONINFO_H
16
17#include "X86.h"
18#include "X86RegisterInfo.h"
19#include "llvm/ADT/DenseMap.h"
20#include "llvm/Target/TargetInstrInfo.h"
21
22#define GET_INSTRINFO_HEADER
23#include "X86GenInstrInfo.inc"
24
25namespace llvm {
26  class X86RegisterInfo;
27  class X86TargetMachine;
28
29namespace X86 {
30  // X86 specific condition code. These correspond to X86_*_COND in
31  // X86InstrInfo.td. They must be kept in synch.
32  enum CondCode {
33    COND_A  = 0,
34    COND_AE = 1,
35    COND_B  = 2,
36    COND_BE = 3,
37    COND_E  = 4,
38    COND_G  = 5,
39    COND_GE = 6,
40    COND_L  = 7,
41    COND_LE = 8,
42    COND_NE = 9,
43    COND_NO = 10,
44    COND_NP = 11,
45    COND_NS = 12,
46    COND_O  = 13,
47    COND_P  = 14,
48    COND_S  = 15,
49
50    // Artificial condition codes. These are used by AnalyzeBranch
51    // to indicate a block terminated with two conditional branches to
52    // the same location. This occurs in code using FCMP_OEQ or FCMP_UNE,
53    // which can't be represented on x86 with a single condition. These
54    // are never used in MachineInstrs.
55    COND_NE_OR_P,
56    COND_NP_OR_E,
57
58    COND_INVALID
59  };
60
61  // Turn condition code into conditional branch opcode.
62  unsigned GetCondBranchFromCond(CondCode CC);
63
64  /// GetOppositeBranchCondition - Return the inverse of the specified cond,
65  /// e.g. turning COND_E to COND_NE.
66  CondCode GetOppositeBranchCondition(X86::CondCode CC);
67}  // end namespace X86;
68
69
70/// isGlobalStubReference - Return true if the specified TargetFlag operand is
71/// a reference to a stub for a global, not the global itself.
72inline static bool isGlobalStubReference(unsigned char TargetFlag) {
73  switch (TargetFlag) {
74  case X86II::MO_DLLIMPORT: // dllimport stub.
75  case X86II::MO_GOTPCREL:  // rip-relative GOT reference.
76  case X86II::MO_GOT:       // normal GOT reference.
77  case X86II::MO_DARWIN_NONLAZY_PIC_BASE:        // Normal $non_lazy_ptr ref.
78  case X86II::MO_DARWIN_NONLAZY:                 // Normal $non_lazy_ptr ref.
79  case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE: // Hidden $non_lazy_ptr ref.
80    return true;
81  default:
82    return false;
83  }
84}
85
86/// isGlobalRelativeToPICBase - Return true if the specified global value
87/// reference is relative to a 32-bit PIC base (X86ISD::GlobalBaseReg).  If this
88/// is true, the addressing mode has the PIC base register added in (e.g. EBX).
89inline static bool isGlobalRelativeToPICBase(unsigned char TargetFlag) {
90  switch (TargetFlag) {
91  case X86II::MO_GOTOFF:                         // isPICStyleGOT: local global.
92  case X86II::MO_GOT:                            // isPICStyleGOT: other global.
93  case X86II::MO_PIC_BASE_OFFSET:                // Darwin local global.
94  case X86II::MO_DARWIN_NONLAZY_PIC_BASE:        // Darwin/32 external global.
95  case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE: // Darwin/32 hidden global.
96  case X86II::MO_TLVP:                           // ??? Pretty sure..
97    return true;
98  default:
99    return false;
100  }
101}
102
103inline static bool isScale(const MachineOperand &MO) {
104  return MO.isImm() &&
105    (MO.getImm() == 1 || MO.getImm() == 2 ||
106     MO.getImm() == 4 || MO.getImm() == 8);
107}
108
109inline static bool isLeaMem(const MachineInstr *MI, unsigned Op) {
110  if (MI->getOperand(Op).isFI()) return true;
111  return Op+4 <= MI->getNumOperands() &&
112    MI->getOperand(Op  ).isReg() && isScale(MI->getOperand(Op+1)) &&
113    MI->getOperand(Op+2).isReg() &&
114    (MI->getOperand(Op+3).isImm() ||
115     MI->getOperand(Op+3).isGlobal() ||
116     MI->getOperand(Op+3).isCPI() ||
117     MI->getOperand(Op+3).isJTI());
118}
119
120inline static bool isMem(const MachineInstr *MI, unsigned Op) {
121  if (MI->getOperand(Op).isFI()) return true;
122  return Op+5 <= MI->getNumOperands() &&
123    MI->getOperand(Op+4).isReg() &&
124    isLeaMem(MI, Op);
125}
126
127class X86InstrInfo : public X86GenInstrInfo {
128  X86TargetMachine &TM;
129  const X86RegisterInfo RI;
130
131  /// RegOp2MemOpTable3Addr, RegOp2MemOpTable2Addr,
132  /// RegOp2MemOpTable0, RegOp2MemOpTable1,
133  /// RegOp2MemOpTable2 - Load / store folding opcode maps.
134  ///
135  typedef DenseMap<unsigned,
136                   std::pair<unsigned, unsigned> > RegOp2MemOpTableType;
137  RegOp2MemOpTableType RegOp2MemOpTable2Addr;
138  RegOp2MemOpTableType RegOp2MemOpTable0;
139  RegOp2MemOpTableType RegOp2MemOpTable1;
140  RegOp2MemOpTableType RegOp2MemOpTable2;
141  RegOp2MemOpTableType RegOp2MemOpTable3;
142
143  /// MemOp2RegOpTable - Load / store unfolding opcode map.
144  ///
145  typedef DenseMap<unsigned,
146                   std::pair<unsigned, unsigned> > MemOp2RegOpTableType;
147  MemOp2RegOpTableType MemOp2RegOpTable;
148
149  void AddTableEntry(RegOp2MemOpTableType &R2MTable,
150                     MemOp2RegOpTableType &M2RTable,
151                     unsigned RegOp, unsigned MemOp, unsigned Flags);
152
153public:
154  explicit X86InstrInfo(X86TargetMachine &tm);
155
156  /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info.  As
157  /// such, whenever a client has an instance of instruction info, it should
158  /// always be able to get register info as well (through this method).
159  ///
160  virtual const X86RegisterInfo &getRegisterInfo() const { return RI; }
161
162  /// isCoalescableExtInstr - Return true if the instruction is a "coalescable"
163  /// extension instruction. That is, it's like a copy where it's legal for the
164  /// source to overlap the destination. e.g. X86::MOVSX64rr32. If this returns
165  /// true, then it's expected the pre-extension value is available as a subreg
166  /// of the result register. This also returns the sub-register index in
167  /// SubIdx.
168  virtual bool isCoalescableExtInstr(const MachineInstr &MI,
169                                     unsigned &SrcReg, unsigned &DstReg,
170                                     unsigned &SubIdx) const;
171
172  unsigned isLoadFromStackSlot(const MachineInstr *MI, int &FrameIndex) const;
173  /// isLoadFromStackSlotPostFE - Check for post-frame ptr elimination
174  /// stack locations as well.  This uses a heuristic so it isn't
175  /// reliable for correctness.
176  unsigned isLoadFromStackSlotPostFE(const MachineInstr *MI,
177                                     int &FrameIndex) const;
178
179  unsigned isStoreToStackSlot(const MachineInstr *MI, int &FrameIndex) const;
180  /// isStoreToStackSlotPostFE - Check for post-frame ptr elimination
181  /// stack locations as well.  This uses a heuristic so it isn't
182  /// reliable for correctness.
183  unsigned isStoreToStackSlotPostFE(const MachineInstr *MI,
184                                    int &FrameIndex) const;
185
186  bool isReallyTriviallyReMaterializable(const MachineInstr *MI,
187                                         AliasAnalysis *AA) const;
188  void reMaterialize(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
189                     unsigned DestReg, unsigned SubIdx,
190                     const MachineInstr *Orig,
191                     const TargetRegisterInfo &TRI) const;
192
193  /// convertToThreeAddress - This method must be implemented by targets that
194  /// set the M_CONVERTIBLE_TO_3_ADDR flag.  When this flag is set, the target
195  /// may be able to convert a two-address instruction into a true
196  /// three-address instruction on demand.  This allows the X86 target (for
197  /// example) to convert ADD and SHL instructions into LEA instructions if they
198  /// would require register copies due to two-addressness.
199  ///
200  /// This method returns a null pointer if the transformation cannot be
201  /// performed, otherwise it returns the new instruction.
202  ///
203  virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI,
204                                              MachineBasicBlock::iterator &MBBI,
205                                              LiveVariables *LV) const;
206
207  /// commuteInstruction - We have a few instructions that must be hacked on to
208  /// commute them.
209  ///
210  virtual MachineInstr *commuteInstruction(MachineInstr *MI, bool NewMI) const;
211
212  // Branch analysis.
213  virtual bool isUnpredicatedTerminator(const MachineInstr* MI) const;
214  virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
215                             MachineBasicBlock *&FBB,
216                             SmallVectorImpl<MachineOperand> &Cond,
217                             bool AllowModify) const;
218  virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const;
219  virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
220                                MachineBasicBlock *FBB,
221                                const SmallVectorImpl<MachineOperand> &Cond,
222                                DebugLoc DL) const;
223  virtual void copyPhysReg(MachineBasicBlock &MBB,
224                           MachineBasicBlock::iterator MI, DebugLoc DL,
225                           unsigned DestReg, unsigned SrcReg,
226                           bool KillSrc) const;
227  virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
228                                   MachineBasicBlock::iterator MI,
229                                   unsigned SrcReg, bool isKill, int FrameIndex,
230                                   const TargetRegisterClass *RC,
231                                   const TargetRegisterInfo *TRI) const;
232
233  virtual void storeRegToAddr(MachineFunction &MF, unsigned SrcReg, bool isKill,
234                              SmallVectorImpl<MachineOperand> &Addr,
235                              const TargetRegisterClass *RC,
236                              MachineInstr::mmo_iterator MMOBegin,
237                              MachineInstr::mmo_iterator MMOEnd,
238                              SmallVectorImpl<MachineInstr*> &NewMIs) const;
239
240  virtual void loadRegFromStackSlot(MachineBasicBlock &MBB,
241                                    MachineBasicBlock::iterator MI,
242                                    unsigned DestReg, int FrameIndex,
243                                    const TargetRegisterClass *RC,
244                                    const TargetRegisterInfo *TRI) const;
245
246  virtual void loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
247                               SmallVectorImpl<MachineOperand> &Addr,
248                               const TargetRegisterClass *RC,
249                               MachineInstr::mmo_iterator MMOBegin,
250                               MachineInstr::mmo_iterator MMOEnd,
251                               SmallVectorImpl<MachineInstr*> &NewMIs) const;
252
253  virtual bool expandPostRAPseudo(MachineBasicBlock::iterator MI) const;
254
255  virtual
256  MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF,
257                                         int FrameIx, uint64_t Offset,
258                                         const MDNode *MDPtr,
259                                         DebugLoc DL) const;
260
261  /// foldMemoryOperand - If this target supports it, fold a load or store of
262  /// the specified stack slot into the specified machine instruction for the
263  /// specified operand(s).  If this is possible, the target should perform the
264  /// folding and return true, otherwise it should return false.  If it folds
265  /// the instruction, it is likely that the MachineInstruction the iterator
266  /// references has been changed.
267  virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
268                                              MachineInstr* MI,
269                                           const SmallVectorImpl<unsigned> &Ops,
270                                              int FrameIndex) const;
271
272  /// foldMemoryOperand - Same as the previous version except it allows folding
273  /// of any load and store from / to any address, not just from a specific
274  /// stack slot.
275  virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
276                                              MachineInstr* MI,
277                                           const SmallVectorImpl<unsigned> &Ops,
278                                              MachineInstr* LoadMI) const;
279
280  /// canFoldMemoryOperand - Returns true if the specified load / store is
281  /// folding is possible.
282  virtual bool canFoldMemoryOperand(const MachineInstr*,
283                                    const SmallVectorImpl<unsigned> &) const;
284
285  /// unfoldMemoryOperand - Separate a single instruction which folded a load or
286  /// a store or a load and a store into two or more instruction. If this is
287  /// possible, returns true as well as the new instructions by reference.
288  virtual bool unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
289                           unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
290                           SmallVectorImpl<MachineInstr*> &NewMIs) const;
291
292  virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
293                           SmallVectorImpl<SDNode*> &NewNodes) const;
294
295  /// getOpcodeAfterMemoryUnfold - Returns the opcode of the would be new
296  /// instruction after load / store are unfolded from an instruction of the
297  /// specified opcode. It returns zero if the specified unfolding is not
298  /// possible. If LoadRegIndex is non-null, it is filled in with the operand
299  /// index of the operand which will hold the register holding the loaded
300  /// value.
301  virtual unsigned getOpcodeAfterMemoryUnfold(unsigned Opc,
302                                      bool UnfoldLoad, bool UnfoldStore,
303                                      unsigned *LoadRegIndex = 0) const;
304
305  /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler
306  /// to determine if two loads are loading from the same base address. It
307  /// should only return true if the base pointers are the same and the
308  /// only differences between the two addresses are the offset. It also returns
309  /// the offsets by reference.
310  virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
311                                       int64_t &Offset1, int64_t &Offset2) const;
312
313  /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
314  /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
315  /// be scheduled togther. On some targets if two loads are loading from
316  /// addresses in the same cache line, it's better if they are scheduled
317  /// together. This function takes two integers that represent the load offsets
318  /// from the common base address. It returns true if it decides it's desirable
319  /// to schedule the two loads together. "NumLoads" is the number of loads that
320  /// have already been scheduled after Load1.
321  virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
322                                       int64_t Offset1, int64_t Offset2,
323                                       unsigned NumLoads) const;
324
325  virtual void getNoopForMachoTarget(MCInst &NopInst) const;
326
327  virtual
328  bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const;
329
330  /// isSafeToMoveRegClassDefs - Return true if it's safe to move a machine
331  /// instruction that defines the specified register class.
332  bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const;
333
334  static bool isX86_64ExtendedReg(const MachineOperand &MO) {
335    if (!MO.isReg()) return false;
336    return X86II::isX86_64ExtendedReg(MO.getReg());
337  }
338
339  /// getGlobalBaseReg - Return a virtual register initialized with the
340  /// the global base register value. Output instructions required to
341  /// initialize the register in the function entry block, if necessary.
342  ///
343  unsigned getGlobalBaseReg(MachineFunction *MF) const;
344
345  std::pair<uint16_t, uint16_t>
346  getExecutionDomain(const MachineInstr *MI) const;
347
348  void setExecutionDomain(MachineInstr *MI, unsigned Domain) const;
349
350  unsigned getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
351                                        const TargetRegisterInfo *TRI) const;
352  void breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
353                                 const TargetRegisterInfo *TRI) const;
354
355  MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
356                                      MachineInstr* MI,
357                                      unsigned OpNum,
358                                      const SmallVectorImpl<MachineOperand> &MOs,
359                                      unsigned Size, unsigned Alignment) const;
360
361  bool isHighLatencyDef(int opc) const;
362
363  bool hasHighOperandLatency(const InstrItineraryData *ItinData,
364                             const MachineRegisterInfo *MRI,
365                             const MachineInstr *DefMI, unsigned DefIdx,
366                             const MachineInstr *UseMI, unsigned UseIdx) const;
367
368private:
369  MachineInstr * convertToThreeAddressWithLEA(unsigned MIOpc,
370                                              MachineFunction::iterator &MFI,
371                                              MachineBasicBlock::iterator &MBBI,
372                                              LiveVariables *LV) const;
373
374  /// isFrameOperand - Return true and the FrameIndex if the specified
375  /// operand and follow operands form a reference to the stack frame.
376  bool isFrameOperand(const MachineInstr *MI, unsigned int Op,
377                      int &FrameIndex) const;
378};
379
380} // End llvm namespace
381
382#endif
383