Mercurial > hg > CbC > CbC_llvm
comparison lib/Target/ARM/ARMBaseInstrInfo.h @ 0:95c75e76d11b LLVM3.4
LLVM 3.4
author | Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp> |
---|---|
date | Thu, 12 Dec 2013 13:56:28 +0900 |
parents | |
children | e4204d083e25 |
comparison
equal
deleted
inserted
replaced
-1:000000000000 | 0:95c75e76d11b |
---|---|
1 //===-- ARMBaseInstrInfo.h - ARM Base 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 Base ARM implementation of the TargetInstrInfo class. | |
11 // | |
12 //===----------------------------------------------------------------------===// | |
13 | |
14 #ifndef ARMBASEINSTRUCTIONINFO_H | |
15 #define ARMBASEINSTRUCTIONINFO_H | |
16 | |
17 #include "ARM.h" | |
18 #include "llvm/ADT/DenseMap.h" | |
19 #include "llvm/ADT/SmallSet.h" | |
20 #include "llvm/CodeGen/MachineInstrBuilder.h" | |
21 #include "llvm/Target/TargetInstrInfo.h" | |
22 | |
23 #define GET_INSTRINFO_HEADER | |
24 #include "ARMGenInstrInfo.inc" | |
25 | |
26 namespace llvm { | |
27 class ARMSubtarget; | |
28 class ARMBaseRegisterInfo; | |
29 | |
30 class ARMBaseInstrInfo : public ARMGenInstrInfo { | |
31 const ARMSubtarget &Subtarget; | |
32 | |
33 protected: | |
34 // Can be only subclassed. | |
35 explicit ARMBaseInstrInfo(const ARMSubtarget &STI); | |
36 | |
37 public: | |
38 // Return whether the target has an explicit NOP encoding. | |
39 bool hasNOP() const; | |
40 | |
41 // Return the non-pre/post incrementing version of 'Opc'. Return 0 | |
42 // if there is not such an opcode. | |
43 virtual unsigned getUnindexedOpcode(unsigned Opc) const =0; | |
44 | |
45 virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI, | |
46 MachineBasicBlock::iterator &MBBI, | |
47 LiveVariables *LV) const; | |
48 | |
49 virtual const ARMBaseRegisterInfo &getRegisterInfo() const = 0; | |
50 const ARMSubtarget &getSubtarget() const { return Subtarget; } | |
51 | |
52 ScheduleHazardRecognizer * | |
53 CreateTargetHazardRecognizer(const TargetMachine *TM, | |
54 const ScheduleDAG *DAG) const; | |
55 | |
56 ScheduleHazardRecognizer * | |
57 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, | |
58 const ScheduleDAG *DAG) const; | |
59 | |
60 // Branch analysis. | |
61 virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, | |
62 MachineBasicBlock *&FBB, | |
63 SmallVectorImpl<MachineOperand> &Cond, | |
64 bool AllowModify = false) const; | |
65 virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const; | |
66 virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, | |
67 MachineBasicBlock *FBB, | |
68 const SmallVectorImpl<MachineOperand> &Cond, | |
69 DebugLoc DL) const; | |
70 | |
71 virtual | |
72 bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const; | |
73 | |
74 // Predication support. | |
75 bool isPredicated(const MachineInstr *MI) const; | |
76 | |
77 ARMCC::CondCodes getPredicate(const MachineInstr *MI) const { | |
78 int PIdx = MI->findFirstPredOperandIdx(); | |
79 return PIdx != -1 ? (ARMCC::CondCodes)MI->getOperand(PIdx).getImm() | |
80 : ARMCC::AL; | |
81 } | |
82 | |
83 virtual | |
84 bool PredicateInstruction(MachineInstr *MI, | |
85 const SmallVectorImpl<MachineOperand> &Pred) const; | |
86 | |
87 virtual | |
88 bool SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1, | |
89 const SmallVectorImpl<MachineOperand> &Pred2) const; | |
90 | |
91 virtual bool DefinesPredicate(MachineInstr *MI, | |
92 std::vector<MachineOperand> &Pred) const; | |
93 | |
94 virtual bool isPredicable(MachineInstr *MI) const; | |
95 | |
96 /// GetInstSize - Returns the size of the specified MachineInstr. | |
97 /// | |
98 virtual unsigned GetInstSizeInBytes(const MachineInstr* MI) const; | |
99 | |
100 virtual unsigned isLoadFromStackSlot(const MachineInstr *MI, | |
101 int &FrameIndex) const; | |
102 virtual unsigned isStoreToStackSlot(const MachineInstr *MI, | |
103 int &FrameIndex) const; | |
104 virtual unsigned isLoadFromStackSlotPostFE(const MachineInstr *MI, | |
105 int &FrameIndex) const; | |
106 virtual unsigned isStoreToStackSlotPostFE(const MachineInstr *MI, | |
107 int &FrameIndex) const; | |
108 | |
109 virtual void copyPhysReg(MachineBasicBlock &MBB, | |
110 MachineBasicBlock::iterator I, DebugLoc DL, | |
111 unsigned DestReg, unsigned SrcReg, | |
112 bool KillSrc) const; | |
113 | |
114 virtual void storeRegToStackSlot(MachineBasicBlock &MBB, | |
115 MachineBasicBlock::iterator MBBI, | |
116 unsigned SrcReg, bool isKill, int FrameIndex, | |
117 const TargetRegisterClass *RC, | |
118 const TargetRegisterInfo *TRI) const; | |
119 | |
120 virtual void loadRegFromStackSlot(MachineBasicBlock &MBB, | |
121 MachineBasicBlock::iterator MBBI, | |
122 unsigned DestReg, int FrameIndex, | |
123 const TargetRegisterClass *RC, | |
124 const TargetRegisterInfo *TRI) const; | |
125 | |
126 virtual bool expandPostRAPseudo(MachineBasicBlock::iterator MI) const; | |
127 | |
128 virtual void reMaterialize(MachineBasicBlock &MBB, | |
129 MachineBasicBlock::iterator MI, | |
130 unsigned DestReg, unsigned SubIdx, | |
131 const MachineInstr *Orig, | |
132 const TargetRegisterInfo &TRI) const; | |
133 | |
134 MachineInstr *duplicate(MachineInstr *Orig, MachineFunction &MF) const; | |
135 | |
136 MachineInstr *commuteInstruction(MachineInstr*, bool=false) const; | |
137 | |
138 const MachineInstrBuilder &AddDReg(MachineInstrBuilder &MIB, unsigned Reg, | |
139 unsigned SubIdx, unsigned State, | |
140 const TargetRegisterInfo *TRI) const; | |
141 | |
142 virtual bool produceSameValue(const MachineInstr *MI0, | |
143 const MachineInstr *MI1, | |
144 const MachineRegisterInfo *MRI) const; | |
145 | |
146 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to | |
147 /// determine if two loads are loading from the same base address. It should | |
148 /// only return true if the base pointers are the same and the only | |
149 /// differences between the two addresses is the offset. It also returns the | |
150 /// offsets by reference. | |
151 virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, | |
152 int64_t &Offset1, int64_t &Offset2)const; | |
153 | |
154 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to | |
155 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads | |
156 /// should be scheduled togther. On some targets if two loads are loading from | |
157 /// addresses in the same cache line, it's better if they are scheduled | |
158 /// together. This function takes two integers that represent the load offsets | |
159 /// from the common base address. It returns true if it decides it's desirable | |
160 /// to schedule the two loads together. "NumLoads" is the number of loads that | |
161 /// have already been scheduled after Load1. | |
162 virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2, | |
163 int64_t Offset1, int64_t Offset2, | |
164 unsigned NumLoads) const; | |
165 | |
166 virtual bool isSchedulingBoundary(const MachineInstr *MI, | |
167 const MachineBasicBlock *MBB, | |
168 const MachineFunction &MF) const; | |
169 | |
170 virtual bool isProfitableToIfCvt(MachineBasicBlock &MBB, | |
171 unsigned NumCycles, unsigned ExtraPredCycles, | |
172 const BranchProbability &Probability) const; | |
173 | |
174 virtual bool isProfitableToIfCvt(MachineBasicBlock &TMBB, | |
175 unsigned NumT, unsigned ExtraT, | |
176 MachineBasicBlock &FMBB, | |
177 unsigned NumF, unsigned ExtraF, | |
178 const BranchProbability &Probability) const; | |
179 | |
180 virtual bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB, | |
181 unsigned NumCycles, | |
182 const BranchProbability | |
183 &Probability) const { | |
184 return NumCycles == 1; | |
185 } | |
186 | |
187 virtual bool isProfitableToUnpredicate(MachineBasicBlock &TMBB, | |
188 MachineBasicBlock &FMBB) const; | |
189 | |
190 /// analyzeCompare - For a comparison instruction, return the source registers | |
191 /// in SrcReg and SrcReg2 if having two register operands, and the value it | |
192 /// compares against in CmpValue. Return true if the comparison instruction | |
193 /// can be analyzed. | |
194 virtual bool analyzeCompare(const MachineInstr *MI, unsigned &SrcReg, | |
195 unsigned &SrcReg2, int &CmpMask, | |
196 int &CmpValue) const; | |
197 | |
198 /// optimizeCompareInstr - Convert the instruction to set the zero flag so | |
199 /// that we can remove a "comparison with zero"; Remove a redundant CMP | |
200 /// instruction if the flags can be updated in the same way by an earlier | |
201 /// instruction such as SUB. | |
202 virtual bool optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, | |
203 unsigned SrcReg2, int CmpMask, int CmpValue, | |
204 const MachineRegisterInfo *MRI) const; | |
205 | |
206 virtual bool analyzeSelect(const MachineInstr *MI, | |
207 SmallVectorImpl<MachineOperand> &Cond, | |
208 unsigned &TrueOp, unsigned &FalseOp, | |
209 bool &Optimizable) const; | |
210 | |
211 virtual MachineInstr *optimizeSelect(MachineInstr *MI, bool) const; | |
212 | |
213 /// FoldImmediate - 'Reg' is known to be defined by a move immediate | |
214 /// instruction, try to fold the immediate into the use instruction. | |
215 virtual bool FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI, | |
216 unsigned Reg, MachineRegisterInfo *MRI) const; | |
217 | |
218 virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData, | |
219 const MachineInstr *MI) const; | |
220 | |
221 virtual | |
222 int getOperandLatency(const InstrItineraryData *ItinData, | |
223 const MachineInstr *DefMI, unsigned DefIdx, | |
224 const MachineInstr *UseMI, unsigned UseIdx) const; | |
225 virtual | |
226 int getOperandLatency(const InstrItineraryData *ItinData, | |
227 SDNode *DefNode, unsigned DefIdx, | |
228 SDNode *UseNode, unsigned UseIdx) const; | |
229 | |
230 /// VFP/NEON execution domains. | |
231 std::pair<uint16_t, uint16_t> | |
232 getExecutionDomain(const MachineInstr *MI) const; | |
233 void setExecutionDomain(MachineInstr *MI, unsigned Domain) const; | |
234 | |
235 unsigned getPartialRegUpdateClearance(const MachineInstr*, unsigned, | |
236 const TargetRegisterInfo*) const; | |
237 void breakPartialRegDependency(MachineBasicBlock::iterator, unsigned, | |
238 const TargetRegisterInfo *TRI) const; | |
239 /// Get the number of addresses by LDM or VLDM or zero for unknown. | |
240 unsigned getNumLDMAddresses(const MachineInstr *MI) const; | |
241 | |
242 private: | |
243 unsigned getInstBundleLength(const MachineInstr *MI) const; | |
244 | |
245 int getVLDMDefCycle(const InstrItineraryData *ItinData, | |
246 const MCInstrDesc &DefMCID, | |
247 unsigned DefClass, | |
248 unsigned DefIdx, unsigned DefAlign) const; | |
249 int getLDMDefCycle(const InstrItineraryData *ItinData, | |
250 const MCInstrDesc &DefMCID, | |
251 unsigned DefClass, | |
252 unsigned DefIdx, unsigned DefAlign) const; | |
253 int getVSTMUseCycle(const InstrItineraryData *ItinData, | |
254 const MCInstrDesc &UseMCID, | |
255 unsigned UseClass, | |
256 unsigned UseIdx, unsigned UseAlign) const; | |
257 int getSTMUseCycle(const InstrItineraryData *ItinData, | |
258 const MCInstrDesc &UseMCID, | |
259 unsigned UseClass, | |
260 unsigned UseIdx, unsigned UseAlign) const; | |
261 int getOperandLatency(const InstrItineraryData *ItinData, | |
262 const MCInstrDesc &DefMCID, | |
263 unsigned DefIdx, unsigned DefAlign, | |
264 const MCInstrDesc &UseMCID, | |
265 unsigned UseIdx, unsigned UseAlign) const; | |
266 | |
267 unsigned getPredicationCost(const MachineInstr *MI) const; | |
268 | |
269 unsigned getInstrLatency(const InstrItineraryData *ItinData, | |
270 const MachineInstr *MI, | |
271 unsigned *PredCost = 0) const; | |
272 | |
273 int getInstrLatency(const InstrItineraryData *ItinData, | |
274 SDNode *Node) const; | |
275 | |
276 bool hasHighOperandLatency(const InstrItineraryData *ItinData, | |
277 const MachineRegisterInfo *MRI, | |
278 const MachineInstr *DefMI, unsigned DefIdx, | |
279 const MachineInstr *UseMI, unsigned UseIdx) const; | |
280 bool hasLowDefLatency(const InstrItineraryData *ItinData, | |
281 const MachineInstr *DefMI, unsigned DefIdx) const; | |
282 | |
283 /// verifyInstruction - Perform target specific instruction verification. | |
284 bool verifyInstruction(const MachineInstr *MI, StringRef &ErrInfo) const; | |
285 | |
286 private: | |
287 /// Modeling special VFP / NEON fp MLA / MLS hazards. | |
288 | |
289 /// MLxEntryMap - Map fp MLA / MLS to the corresponding entry in the internal | |
290 /// MLx table. | |
291 DenseMap<unsigned, unsigned> MLxEntryMap; | |
292 | |
293 /// MLxHazardOpcodes - Set of add / sub and multiply opcodes that would cause | |
294 /// stalls when scheduled together with fp MLA / MLS opcodes. | |
295 SmallSet<unsigned, 16> MLxHazardOpcodes; | |
296 | |
297 public: | |
298 /// isFpMLxInstruction - Return true if the specified opcode is a fp MLA / MLS | |
299 /// instruction. | |
300 bool isFpMLxInstruction(unsigned Opcode) const { | |
301 return MLxEntryMap.count(Opcode); | |
302 } | |
303 | |
304 /// isFpMLxInstruction - This version also returns the multiply opcode and the | |
305 /// addition / subtraction opcode to expand to. Return true for 'HasLane' for | |
306 /// the MLX instructions with an extra lane operand. | |
307 bool isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc, | |
308 unsigned &AddSubOpc, bool &NegAcc, | |
309 bool &HasLane) const; | |
310 | |
311 /// canCauseFpMLxStall - Return true if an instruction of the specified opcode | |
312 /// will cause stalls when scheduled after (within 4-cycle window) a fp | |
313 /// MLA / MLS instruction. | |
314 bool canCauseFpMLxStall(unsigned Opcode) const { | |
315 return MLxHazardOpcodes.count(Opcode); | |
316 } | |
317 | |
318 /// Returns true if the instruction has a shift by immediate that can be | |
319 /// executed in one cycle less. | |
320 bool isSwiftFastImmShift(const MachineInstr *MI) const; | |
321 }; | |
322 | |
323 static inline | |
324 const MachineInstrBuilder &AddDefaultPred(const MachineInstrBuilder &MIB) { | |
325 return MIB.addImm((int64_t)ARMCC::AL).addReg(0); | |
326 } | |
327 | |
328 static inline | |
329 const MachineInstrBuilder &AddDefaultCC(const MachineInstrBuilder &MIB) { | |
330 return MIB.addReg(0); | |
331 } | |
332 | |
333 static inline | |
334 const MachineInstrBuilder &AddDefaultT1CC(const MachineInstrBuilder &MIB, | |
335 bool isDead = false) { | |
336 return MIB.addReg(ARM::CPSR, getDefRegState(true) | getDeadRegState(isDead)); | |
337 } | |
338 | |
339 static inline | |
340 const MachineInstrBuilder &AddNoT1CC(const MachineInstrBuilder &MIB) { | |
341 return MIB.addReg(0); | |
342 } | |
343 | |
344 static inline | |
345 bool isUncondBranchOpcode(int Opc) { | |
346 return Opc == ARM::B || Opc == ARM::tB || Opc == ARM::t2B; | |
347 } | |
348 | |
349 static inline | |
350 bool isCondBranchOpcode(int Opc) { | |
351 return Opc == ARM::Bcc || Opc == ARM::tBcc || Opc == ARM::t2Bcc; | |
352 } | |
353 | |
354 static inline | |
355 bool isJumpTableBranchOpcode(int Opc) { | |
356 return Opc == ARM::BR_JTr || Opc == ARM::BR_JTm || Opc == ARM::BR_JTadd || | |
357 Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT; | |
358 } | |
359 | |
360 static inline | |
361 bool isIndirectBranchOpcode(int Opc) { | |
362 return Opc == ARM::BX || Opc == ARM::MOVPCRX || Opc == ARM::tBRIND; | |
363 } | |
364 | |
365 static inline bool isPopOpcode(int Opc) { | |
366 return Opc == ARM::tPOP_RET || Opc == ARM::LDMIA_RET || | |
367 Opc == ARM::t2LDMIA_RET || Opc == ARM::tPOP || Opc == ARM::LDMIA_UPD || | |
368 Opc == ARM::t2LDMIA_UPD || Opc == ARM::VLDMDIA_UPD; | |
369 } | |
370 | |
371 static inline bool isPushOpcode(int Opc) { | |
372 return Opc == ARM::tPUSH || Opc == ARM::t2STMDB_UPD || | |
373 Opc == ARM::STMDB_UPD || Opc == ARM::VSTMDDB_UPD; | |
374 } | |
375 | |
376 /// getInstrPredicate - If instruction is predicated, returns its predicate | |
377 /// condition, otherwise returns AL. It also returns the condition code | |
378 /// register by reference. | |
379 ARMCC::CondCodes getInstrPredicate(const MachineInstr *MI, unsigned &PredReg); | |
380 | |
381 int getMatchingCondBranchOpcode(int Opc); | |
382 | |
383 /// Determine if MI can be folded into an ARM MOVCC instruction, and return the | |
384 /// opcode of the SSA instruction representing the conditional MI. | |
385 unsigned canFoldARMInstrIntoMOVCC(unsigned Reg, | |
386 MachineInstr *&MI, | |
387 const MachineRegisterInfo &MRI); | |
388 | |
389 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether | |
390 /// the instruction is encoded with an 'S' bit is determined by the optional | |
391 /// CPSR def operand. | |
392 unsigned convertAddSubFlagsOpcode(unsigned OldOpc); | |
393 | |
394 /// emitARMRegPlusImmediate / emitT2RegPlusImmediate - Emits a series of | |
395 /// instructions to materializea destreg = basereg + immediate in ARM / Thumb2 | |
396 /// code. | |
397 void emitARMRegPlusImmediate(MachineBasicBlock &MBB, | |
398 MachineBasicBlock::iterator &MBBI, DebugLoc dl, | |
399 unsigned DestReg, unsigned BaseReg, int NumBytes, | |
400 ARMCC::CondCodes Pred, unsigned PredReg, | |
401 const ARMBaseInstrInfo &TII, unsigned MIFlags = 0); | |
402 | |
403 void emitT2RegPlusImmediate(MachineBasicBlock &MBB, | |
404 MachineBasicBlock::iterator &MBBI, DebugLoc dl, | |
405 unsigned DestReg, unsigned BaseReg, int NumBytes, | |
406 ARMCC::CondCodes Pred, unsigned PredReg, | |
407 const ARMBaseInstrInfo &TII, unsigned MIFlags = 0); | |
408 void emitThumbRegPlusImmediate(MachineBasicBlock &MBB, | |
409 MachineBasicBlock::iterator &MBBI, DebugLoc dl, | |
410 unsigned DestReg, unsigned BaseReg, | |
411 int NumBytes, const TargetInstrInfo &TII, | |
412 const ARMBaseRegisterInfo& MRI, | |
413 unsigned MIFlags = 0); | |
414 | |
415 /// Tries to add registers to the reglist of a given base-updating | |
416 /// push/pop instruction to adjust the stack by an additional | |
417 /// NumBytes. This can save a few bytes per function in code-size, but | |
418 /// obviously generates more memory traffic. As such, it only takes | |
419 /// effect in functions being optimised for size. | |
420 bool tryFoldSPUpdateIntoPushPop(MachineFunction &MF, MachineInstr *MI, | |
421 unsigned NumBytes); | |
422 | |
423 /// rewriteARMFrameIndex / rewriteT2FrameIndex - | |
424 /// Rewrite MI to access 'Offset' bytes from the FP. Return false if the | |
425 /// offset could not be handled directly in MI, and return the left-over | |
426 /// portion by reference. | |
427 bool rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx, | |
428 unsigned FrameReg, int &Offset, | |
429 const ARMBaseInstrInfo &TII); | |
430 | |
431 bool rewriteT2FrameIndex(MachineInstr &MI, unsigned FrameRegIdx, | |
432 unsigned FrameReg, int &Offset, | |
433 const ARMBaseInstrInfo &TII); | |
434 | |
435 } // End llvm namespace | |
436 | |
437 #endif |