X86InstrInfo.h revision e5609f37323b105c7720d5d423a9203d1e869c29
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  // Turn CMov opcode into condition code.
65  CondCode getCondFromCMovOpc(unsigned Opc);
66
67  /// GetOppositeBranchCondition - Return the inverse of the specified cond,
68  /// e.g. turning COND_E to COND_NE.
69  CondCode GetOppositeBranchCondition(X86::CondCode CC);
70}  // end namespace X86;
71
72
73/// isGlobalStubReference - Return true if the specified TargetFlag operand is
74/// a reference to a stub for a global, not the global itself.
75inline static bool isGlobalStubReference(unsigned char TargetFlag) {
76  switch (TargetFlag) {
77  case X86II::MO_DLLIMPORT: // dllimport stub.
78  case X86II::MO_GOTPCREL:  // rip-relative GOT reference.
79  case X86II::MO_GOT:       // normal GOT reference.
80  case X86II::MO_DARWIN_NONLAZY_PIC_BASE:        // Normal $non_lazy_ptr ref.
81  case X86II::MO_DARWIN_NONLAZY:                 // Normal $non_lazy_ptr ref.
82  case X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE: // Hidden $non_lazy_ptr ref.
83    return true;
84  default:
85    return false;
86  }
87}
88
89/// isGlobalRelativeToPICBase - Return true if the specified global value
90/// reference is relative to a 32-bit PIC base (X86ISD::GlobalBaseReg).  If this
91/// is true, the addressing mode has the PIC base register added in (e.g. EBX).
92inline static bool isGlobalRelativeToPICBase(unsigned char TargetFlag) {
93  switch (TargetFlag) {
94  case X86II::MO_GOTOFF:                         // isPICStyleGOT: local global.
95  case X86II::MO_GOT:                            // isPICStyleGOT: other global.
96  case X86II::MO_PIC_BASE_OFFSET:                // Darwin local global.
97  case X86II::MO_DARWIN_NONLAZY_PIC_BASE:        // Darwin/32 external global.
98  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  /// Given an operand within a MachineInstr, insert preceding code to put it
196  /// into the right format for a particular kind of LEA instruction. This may
197  /// involve using an appropriate super-register instead (with an implicit use
198  /// of the original) or creating a new virtual register and inserting COPY
199  /// instructions to get the data into the right class.
200  ///
201  /// Reference parameters are set to indicate how caller should add this
202  /// operand to the LEA instruction.
203  bool classifyLEAReg(MachineInstr *MI, const MachineOperand &Src,
204                      unsigned LEAOpcode, bool AllowSP,
205                      unsigned &NewSrc, bool &isKill,
206                      bool &isUndef, MachineOperand &ImplicitOp) const;
207
208  /// convertToThreeAddress - This method must be implemented by targets that
209  /// set the M_CONVERTIBLE_TO_3_ADDR flag.  When this flag is set, the target
210  /// may be able to convert a two-address instruction into a true
211  /// three-address instruction on demand.  This allows the X86 target (for
212  /// example) to convert ADD and SHL instructions into LEA instructions if they
213  /// would require register copies due to two-addressness.
214  ///
215  /// This method returns a null pointer if the transformation cannot be
216  /// performed, otherwise it returns the new instruction.
217  ///
218  virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI,
219                                              MachineBasicBlock::iterator &MBBI,
220                                              LiveVariables *LV) const;
221
222  /// commuteInstruction - We have a few instructions that must be hacked on to
223  /// commute them.
224  ///
225  virtual MachineInstr *commuteInstruction(MachineInstr *MI, bool NewMI) const;
226
227  // Branch analysis.
228  virtual bool isUnpredicatedTerminator(const MachineInstr* MI) const;
229  virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
230                             MachineBasicBlock *&FBB,
231                             SmallVectorImpl<MachineOperand> &Cond,
232                             bool AllowModify) const;
233  virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const;
234  virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
235                                MachineBasicBlock *FBB,
236                                const SmallVectorImpl<MachineOperand> &Cond,
237                                DebugLoc DL) const;
238  virtual bool canInsertSelect(const MachineBasicBlock&,
239                               const SmallVectorImpl<MachineOperand> &Cond,
240                               unsigned, unsigned, int&, int&, int&) const;
241  virtual void insertSelect(MachineBasicBlock &MBB,
242                            MachineBasicBlock::iterator MI, DebugLoc DL,
243                            unsigned DstReg,
244                            const SmallVectorImpl<MachineOperand> &Cond,
245                            unsigned TrueReg, unsigned FalseReg) const;
246  virtual void copyPhysReg(MachineBasicBlock &MBB,
247                           MachineBasicBlock::iterator MI, DebugLoc DL,
248                           unsigned DestReg, unsigned SrcReg,
249                           bool KillSrc) const;
250  virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
251                                   MachineBasicBlock::iterator MI,
252                                   unsigned SrcReg, bool isKill, int FrameIndex,
253                                   const TargetRegisterClass *RC,
254                                   const TargetRegisterInfo *TRI) const;
255
256  virtual void storeRegToAddr(MachineFunction &MF, unsigned SrcReg, bool isKill,
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 void loadRegFromStackSlot(MachineBasicBlock &MBB,
264                                    MachineBasicBlock::iterator MI,
265                                    unsigned DestReg, int FrameIndex,
266                                    const TargetRegisterClass *RC,
267                                    const TargetRegisterInfo *TRI) const;
268
269  virtual void loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
270                               SmallVectorImpl<MachineOperand> &Addr,
271                               const TargetRegisterClass *RC,
272                               MachineInstr::mmo_iterator MMOBegin,
273                               MachineInstr::mmo_iterator MMOEnd,
274                               SmallVectorImpl<MachineInstr*> &NewMIs) const;
275
276  virtual bool expandPostRAPseudo(MachineBasicBlock::iterator MI) const;
277
278  virtual
279  MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF,
280                                         int FrameIx, uint64_t Offset,
281                                         const MDNode *MDPtr,
282                                         DebugLoc DL) const;
283
284  /// foldMemoryOperand - If this target supports it, fold a load or store of
285  /// the specified stack slot into the specified machine instruction for the
286  /// specified operand(s).  If this is possible, the target should perform the
287  /// folding and return true, otherwise it should return false.  If it folds
288  /// the instruction, it is likely that the MachineInstruction the iterator
289  /// references has been changed.
290  virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
291                                              MachineInstr* MI,
292                                           const SmallVectorImpl<unsigned> &Ops,
293                                              int FrameIndex) const;
294
295  /// foldMemoryOperand - Same as the previous version except it allows folding
296  /// of any load and store from / to any address, not just from a specific
297  /// stack slot.
298  virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
299                                              MachineInstr* MI,
300                                           const SmallVectorImpl<unsigned> &Ops,
301                                              MachineInstr* LoadMI) const;
302
303  /// canFoldMemoryOperand - Returns true if the specified load / store is
304  /// folding is possible.
305  virtual bool canFoldMemoryOperand(const MachineInstr*,
306                                    const SmallVectorImpl<unsigned> &) const;
307
308  /// unfoldMemoryOperand - Separate a single instruction which folded a load or
309  /// a store or a load and a store into two or more instruction. If this is
310  /// possible, returns true as well as the new instructions by reference.
311  virtual bool unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
312                           unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
313                           SmallVectorImpl<MachineInstr*> &NewMIs) const;
314
315  virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
316                           SmallVectorImpl<SDNode*> &NewNodes) const;
317
318  /// getOpcodeAfterMemoryUnfold - Returns the opcode of the would be new
319  /// instruction after load / store are unfolded from an instruction of the
320  /// specified opcode. It returns zero if the specified unfolding is not
321  /// possible. If LoadRegIndex is non-null, it is filled in with the operand
322  /// index of the operand which will hold the register holding the loaded
323  /// value.
324  virtual unsigned getOpcodeAfterMemoryUnfold(unsigned Opc,
325                                      bool UnfoldLoad, bool UnfoldStore,
326                                      unsigned *LoadRegIndex = 0) const;
327
328  /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler
329  /// to determine if two loads are loading from the same base address. It
330  /// should only return true if the base pointers are the same and the
331  /// only differences between the two addresses are the offset. It also returns
332  /// the offsets by reference.
333  virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
334                                       int64_t &Offset1, int64_t &Offset2) const;
335
336  /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
337  /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
338  /// be scheduled togther. On some targets if two loads are loading from
339  /// addresses in the same cache line, it's better if they are scheduled
340  /// together. This function takes two integers that represent the load offsets
341  /// from the common base address. It returns true if it decides it's desirable
342  /// to schedule the two loads together. "NumLoads" is the number of loads that
343  /// have already been scheduled after Load1.
344  virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
345                                       int64_t Offset1, int64_t Offset2,
346                                       unsigned NumLoads) const;
347
348  virtual void getNoopForMachoTarget(MCInst &NopInst) const;
349
350  virtual
351  bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const;
352
353  /// isSafeToMoveRegClassDefs - Return true if it's safe to move a machine
354  /// instruction that defines the specified register class.
355  bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const;
356
357  static bool isX86_64ExtendedReg(const MachineOperand &MO) {
358    if (!MO.isReg()) return false;
359    return X86II::isX86_64ExtendedReg(MO.getReg());
360  }
361
362  /// getGlobalBaseReg - Return a virtual register initialized with the
363  /// the global base register value. Output instructions required to
364  /// initialize the register in the function entry block, if necessary.
365  ///
366  unsigned getGlobalBaseReg(MachineFunction *MF) const;
367
368  std::pair<uint16_t, uint16_t>
369  getExecutionDomain(const MachineInstr *MI) const;
370
371  void setExecutionDomain(MachineInstr *MI, unsigned Domain) const;
372
373  unsigned getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
374                                        const TargetRegisterInfo *TRI) const;
375  void breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
376                                 const TargetRegisterInfo *TRI) const;
377
378  MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
379                                      MachineInstr* MI,
380                                      unsigned OpNum,
381                                      const SmallVectorImpl<MachineOperand> &MOs,
382                                      unsigned Size, unsigned Alignment) const;
383
384  bool isHighLatencyDef(int opc) const;
385
386  bool hasHighOperandLatency(const InstrItineraryData *ItinData,
387                             const MachineRegisterInfo *MRI,
388                             const MachineInstr *DefMI, unsigned DefIdx,
389                             const MachineInstr *UseMI, unsigned UseIdx) const;
390
391  /// analyzeCompare - For a comparison instruction, return the source registers
392  /// in SrcReg and SrcReg2 if having two register operands, and the value it
393  /// compares against in CmpValue. Return true if the comparison instruction
394  /// can be analyzed.
395  virtual bool analyzeCompare(const MachineInstr *MI, unsigned &SrcReg,
396                              unsigned &SrcReg2,
397                              int &CmpMask, int &CmpValue) const;
398
399  /// optimizeCompareInstr - Check if there exists an earlier instruction that
400  /// operates on the same source operands and sets flags in the same way as
401  /// Compare; remove Compare if possible.
402  virtual bool optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg,
403                                    unsigned SrcReg2, int CmpMask, int CmpValue,
404                                    const MachineRegisterInfo *MRI) const;
405
406  /// optimizeLoadInstr - Try to remove the load by folding it to a register
407  /// operand at the use. We fold the load instructions if and only if the
408  /// def and use are in the same BB. We only look at one load and see
409  /// whether it can be folded into MI. FoldAsLoadDefReg is the virtual register
410  /// defined by the load we are trying to fold. DefMI returns the machine
411  /// instruction that defines FoldAsLoadDefReg, and the function returns
412  /// the machine instruction generated due to folding.
413  virtual MachineInstr* optimizeLoadInstr(MachineInstr *MI,
414                        const MachineRegisterInfo *MRI,
415                        unsigned &FoldAsLoadDefReg,
416                        MachineInstr *&DefMI) const;
417
418private:
419  MachineInstr * convertToThreeAddressWithLEA(unsigned MIOpc,
420                                              MachineFunction::iterator &MFI,
421                                              MachineBasicBlock::iterator &MBBI,
422                                              LiveVariables *LV) const;
423
424  /// isFrameOperand - Return true and the FrameIndex if the specified
425  /// operand and follow operands form a reference to the stack frame.
426  bool isFrameOperand(const MachineInstr *MI, unsigned int Op,
427                      int &FrameIndex) const;
428};
429
430} // End llvm namespace
431
432#endif
433