Use BranchProbability instead of floating points in IfConverter.
[llvm/stm8.git] / lib / Target / ARM / ARMBaseInstrInfo.h
blob1d8799fd6684aff401dbef4604a027f982090925
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.
12 //===----------------------------------------------------------------------===//
14 #ifndef ARMBASEINSTRUCTIONINFO_H
15 #define ARMBASEINSTRUCTIONINFO_H
17 #include "ARM.h"
18 #include "llvm/CodeGen/MachineInstrBuilder.h"
19 #include "llvm/Target/TargetInstrInfo.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/SmallSet.h"
23 #define GET_INSTRINFO_HEADER
24 #include "ARMGenInstrInfo.inc"
26 namespace llvm {
27 class ARMSubtarget;
28 class ARMBaseRegisterInfo;
30 /// ARMII - This namespace holds all of the target specific flags that
31 /// instruction info tracks.
32 ///
33 namespace ARMII {
34 enum {
35 //===------------------------------------------------------------------===//
36 // Instruction Flags.
38 //===------------------------------------------------------------------===//
39 // This four-bit field describes the addressing mode used.
40 AddrModeMask = 0x1f, // The AddrMode enums are declared in ARMBaseInfo.h
42 // Size* - Flags to keep track of the size of an instruction.
43 SizeShift = 5,
44 SizeMask = 7 << SizeShift,
45 SizeSpecial = 1, // 0 byte pseudo or special case.
46 Size8Bytes = 2,
47 Size4Bytes = 3,
48 Size2Bytes = 4,
50 // IndexMode - Unindex, pre-indexed, or post-indexed are valid for load
51 // and store ops only. Generic "updating" flag is used for ld/st multiple.
52 // The index mode enums are declared in ARMBaseInfo.h
53 IndexModeShift = 8,
54 IndexModeMask = 3 << IndexModeShift,
56 //===------------------------------------------------------------------===//
57 // Instruction encoding formats.
59 FormShift = 10,
60 FormMask = 0x3f << FormShift,
62 // Pseudo instructions
63 Pseudo = 0 << FormShift,
65 // Multiply instructions
66 MulFrm = 1 << FormShift,
68 // Branch instructions
69 BrFrm = 2 << FormShift,
70 BrMiscFrm = 3 << FormShift,
72 // Data Processing instructions
73 DPFrm = 4 << FormShift,
74 DPSoRegFrm = 5 << FormShift,
76 // Load and Store
77 LdFrm = 6 << FormShift,
78 StFrm = 7 << FormShift,
79 LdMiscFrm = 8 << FormShift,
80 StMiscFrm = 9 << FormShift,
81 LdStMulFrm = 10 << FormShift,
83 LdStExFrm = 11 << FormShift,
85 // Miscellaneous arithmetic instructions
86 ArithMiscFrm = 12 << FormShift,
87 SatFrm = 13 << FormShift,
89 // Extend instructions
90 ExtFrm = 14 << FormShift,
92 // VFP formats
93 VFPUnaryFrm = 15 << FormShift,
94 VFPBinaryFrm = 16 << FormShift,
95 VFPConv1Frm = 17 << FormShift,
96 VFPConv2Frm = 18 << FormShift,
97 VFPConv3Frm = 19 << FormShift,
98 VFPConv4Frm = 20 << FormShift,
99 VFPConv5Frm = 21 << FormShift,
100 VFPLdStFrm = 22 << FormShift,
101 VFPLdStMulFrm = 23 << FormShift,
102 VFPMiscFrm = 24 << FormShift,
104 // Thumb format
105 ThumbFrm = 25 << FormShift,
107 // Miscelleaneous format
108 MiscFrm = 26 << FormShift,
110 // NEON formats
111 NGetLnFrm = 27 << FormShift,
112 NSetLnFrm = 28 << FormShift,
113 NDupFrm = 29 << FormShift,
114 NLdStFrm = 30 << FormShift,
115 N1RegModImmFrm= 31 << FormShift,
116 N2RegFrm = 32 << FormShift,
117 NVCVTFrm = 33 << FormShift,
118 NVDupLnFrm = 34 << FormShift,
119 N2RegVShLFrm = 35 << FormShift,
120 N2RegVShRFrm = 36 << FormShift,
121 N3RegFrm = 37 << FormShift,
122 N3RegVShFrm = 38 << FormShift,
123 NVExtFrm = 39 << FormShift,
124 NVMulSLFrm = 40 << FormShift,
125 NVTBLFrm = 41 << FormShift,
127 //===------------------------------------------------------------------===//
128 // Misc flags.
130 // UnaryDP - Indicates this is a unary data processing instruction, i.e.
131 // it doesn't have a Rn operand.
132 UnaryDP = 1 << 16,
134 // Xform16Bit - Indicates this Thumb2 instruction may be transformed into
135 // a 16-bit Thumb instruction if certain conditions are met.
136 Xform16Bit = 1 << 17,
138 //===------------------------------------------------------------------===//
139 // Code domain.
140 DomainShift = 18,
141 DomainMask = 7 << DomainShift,
142 DomainGeneral = 0 << DomainShift,
143 DomainVFP = 1 << DomainShift,
144 DomainNEON = 2 << DomainShift,
145 DomainNEONA8 = 4 << DomainShift,
147 //===------------------------------------------------------------------===//
148 // Field shifts - such shifts are used to set field while generating
149 // machine instructions.
151 // FIXME: This list will need adjusting/fixing as the MC code emitter
152 // takes shape and the ARMCodeEmitter.cpp bits go away.
153 ShiftTypeShift = 4,
155 M_BitShift = 5,
156 ShiftImmShift = 5,
157 ShiftShift = 7,
158 N_BitShift = 7,
159 ImmHiShift = 8,
160 SoRotImmShift = 8,
161 RegRsShift = 8,
162 ExtRotImmShift = 10,
163 RegRdLoShift = 12,
164 RegRdShift = 12,
165 RegRdHiShift = 16,
166 RegRnShift = 16,
167 S_BitShift = 20,
168 W_BitShift = 21,
169 AM3_I_BitShift = 22,
170 D_BitShift = 22,
171 U_BitShift = 23,
172 P_BitShift = 24,
173 I_BitShift = 25,
174 CondShift = 28
178 class ARMBaseInstrInfo : public ARMGenInstrInfo {
179 const ARMSubtarget &Subtarget;
181 protected:
182 // Can be only subclassed.
183 explicit ARMBaseInstrInfo(const ARMSubtarget &STI);
185 public:
186 // Return the non-pre/post incrementing version of 'Opc'. Return 0
187 // if there is not such an opcode.
188 virtual unsigned getUnindexedOpcode(unsigned Opc) const =0;
190 virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI,
191 MachineBasicBlock::iterator &MBBI,
192 LiveVariables *LV) const;
194 virtual const ARMBaseRegisterInfo &getRegisterInfo() const =0;
195 const ARMSubtarget &getSubtarget() const { return Subtarget; }
197 ScheduleHazardRecognizer *
198 CreateTargetHazardRecognizer(const TargetMachine *TM,
199 const ScheduleDAG *DAG) const;
201 ScheduleHazardRecognizer *
202 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
203 const ScheduleDAG *DAG) const;
205 // Branch analysis.
206 virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
207 MachineBasicBlock *&FBB,
208 SmallVectorImpl<MachineOperand> &Cond,
209 bool AllowModify = false) const;
210 virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const;
211 virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
212 MachineBasicBlock *FBB,
213 const SmallVectorImpl<MachineOperand> &Cond,
214 DebugLoc DL) const;
216 virtual
217 bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const;
219 // Predication support.
220 bool isPredicated(const MachineInstr *MI) const {
221 int PIdx = MI->findFirstPredOperandIdx();
222 return PIdx != -1 && MI->getOperand(PIdx).getImm() != ARMCC::AL;
225 ARMCC::CondCodes getPredicate(const MachineInstr *MI) const {
226 int PIdx = MI->findFirstPredOperandIdx();
227 return PIdx != -1 ? (ARMCC::CondCodes)MI->getOperand(PIdx).getImm()
228 : ARMCC::AL;
231 virtual
232 bool PredicateInstruction(MachineInstr *MI,
233 const SmallVectorImpl<MachineOperand> &Pred) const;
235 virtual
236 bool SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
237 const SmallVectorImpl<MachineOperand> &Pred2) const;
239 virtual bool DefinesPredicate(MachineInstr *MI,
240 std::vector<MachineOperand> &Pred) const;
242 virtual bool isPredicable(MachineInstr *MI) const;
244 /// GetInstSize - Returns the size of the specified MachineInstr.
246 virtual unsigned GetInstSizeInBytes(const MachineInstr* MI) const;
248 virtual unsigned isLoadFromStackSlot(const MachineInstr *MI,
249 int &FrameIndex) const;
250 virtual unsigned isStoreToStackSlot(const MachineInstr *MI,
251 int &FrameIndex) const;
253 virtual void copyPhysReg(MachineBasicBlock &MBB,
254 MachineBasicBlock::iterator I, DebugLoc DL,
255 unsigned DestReg, unsigned SrcReg,
256 bool KillSrc) const;
258 virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
259 MachineBasicBlock::iterator MBBI,
260 unsigned SrcReg, bool isKill, int FrameIndex,
261 const TargetRegisterClass *RC,
262 const TargetRegisterInfo *TRI) const;
264 virtual void loadRegFromStackSlot(MachineBasicBlock &MBB,
265 MachineBasicBlock::iterator MBBI,
266 unsigned DestReg, int FrameIndex,
267 const TargetRegisterClass *RC,
268 const TargetRegisterInfo *TRI) const;
270 virtual MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF,
271 int FrameIx,
272 uint64_t Offset,
273 const MDNode *MDPtr,
274 DebugLoc DL) const;
276 virtual void reMaterialize(MachineBasicBlock &MBB,
277 MachineBasicBlock::iterator MI,
278 unsigned DestReg, unsigned SubIdx,
279 const MachineInstr *Orig,
280 const TargetRegisterInfo &TRI) const;
282 MachineInstr *duplicate(MachineInstr *Orig, MachineFunction &MF) const;
284 virtual bool produceSameValue(const MachineInstr *MI0,
285 const MachineInstr *MI1,
286 const MachineRegisterInfo *MRI) const;
288 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
289 /// determine if two loads are loading from the same base address. It should
290 /// only return true if the base pointers are the same and the only
291 /// differences between the two addresses is the offset. It also returns the
292 /// offsets by reference.
293 virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
294 int64_t &Offset1, int64_t &Offset2)const;
296 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
297 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads
298 /// should be scheduled togther. On some targets if two loads are loading from
299 /// addresses in the same cache line, it's better if they are scheduled
300 /// together. This function takes two integers that represent the load offsets
301 /// from the common base address. It returns true if it decides it's desirable
302 /// to schedule the two loads together. "NumLoads" is the number of loads that
303 /// have already been scheduled after Load1.
304 virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
305 int64_t Offset1, int64_t Offset2,
306 unsigned NumLoads) const;
308 virtual bool isSchedulingBoundary(const MachineInstr *MI,
309 const MachineBasicBlock *MBB,
310 const MachineFunction &MF) const;
312 virtual bool isProfitableToIfCvt(MachineBasicBlock &MBB,
313 unsigned NumCycles, unsigned ExtraPredCycles,
314 const BranchProbability &Probability) const;
316 virtual bool isProfitableToIfCvt(MachineBasicBlock &TMBB,
317 unsigned NumT, unsigned ExtraT,
318 MachineBasicBlock &FMBB,
319 unsigned NumF, unsigned ExtraF,
320 const BranchProbability &Probability) const;
322 virtual bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
323 unsigned NumCycles,
324 const BranchProbability
325 &Probability) const {
326 return NumCycles == 1;
329 /// AnalyzeCompare - For a comparison instruction, return the source register
330 /// in SrcReg and the value it compares against in CmpValue. Return true if
331 /// the comparison instruction can be analyzed.
332 virtual bool AnalyzeCompare(const MachineInstr *MI, unsigned &SrcReg,
333 int &CmpMask, int &CmpValue) const;
335 /// OptimizeCompareInstr - Convert the instruction to set the zero flag so
336 /// that we can remove a "comparison with zero".
337 virtual bool OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg,
338 int CmpMask, int CmpValue,
339 const MachineRegisterInfo *MRI) const;
341 /// FoldImmediate - 'Reg' is known to be defined by a move immediate
342 /// instruction, try to fold the immediate into the use instruction.
343 virtual bool FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI,
344 unsigned Reg, MachineRegisterInfo *MRI) const;
346 virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData,
347 const MachineInstr *MI) const;
349 virtual
350 int getOperandLatency(const InstrItineraryData *ItinData,
351 const MachineInstr *DefMI, unsigned DefIdx,
352 const MachineInstr *UseMI, unsigned UseIdx) const;
353 virtual
354 int getOperandLatency(const InstrItineraryData *ItinData,
355 SDNode *DefNode, unsigned DefIdx,
356 SDNode *UseNode, unsigned UseIdx) const;
357 private:
358 int getVLDMDefCycle(const InstrItineraryData *ItinData,
359 const MCInstrDesc &DefMCID,
360 unsigned DefClass,
361 unsigned DefIdx, unsigned DefAlign) const;
362 int getLDMDefCycle(const InstrItineraryData *ItinData,
363 const MCInstrDesc &DefMCID,
364 unsigned DefClass,
365 unsigned DefIdx, unsigned DefAlign) const;
366 int getVSTMUseCycle(const InstrItineraryData *ItinData,
367 const MCInstrDesc &UseMCID,
368 unsigned UseClass,
369 unsigned UseIdx, unsigned UseAlign) const;
370 int getSTMUseCycle(const InstrItineraryData *ItinData,
371 const MCInstrDesc &UseMCID,
372 unsigned UseClass,
373 unsigned UseIdx, unsigned UseAlign) const;
374 int getOperandLatency(const InstrItineraryData *ItinData,
375 const MCInstrDesc &DefMCID,
376 unsigned DefIdx, unsigned DefAlign,
377 const MCInstrDesc &UseMCID,
378 unsigned UseIdx, unsigned UseAlign) const;
380 int getInstrLatency(const InstrItineraryData *ItinData,
381 const MachineInstr *MI, unsigned *PredCost = 0) const;
383 int getInstrLatency(const InstrItineraryData *ItinData,
384 SDNode *Node) const;
386 bool hasHighOperandLatency(const InstrItineraryData *ItinData,
387 const MachineRegisterInfo *MRI,
388 const MachineInstr *DefMI, unsigned DefIdx,
389 const MachineInstr *UseMI, unsigned UseIdx) const;
390 bool hasLowDefLatency(const InstrItineraryData *ItinData,
391 const MachineInstr *DefMI, unsigned DefIdx) const;
393 private:
394 /// Modeling special VFP / NEON fp MLA / MLS hazards.
396 /// MLxEntryMap - Map fp MLA / MLS to the corresponding entry in the internal
397 /// MLx table.
398 DenseMap<unsigned, unsigned> MLxEntryMap;
400 /// MLxHazardOpcodes - Set of add / sub and multiply opcodes that would cause
401 /// stalls when scheduled together with fp MLA / MLS opcodes.
402 SmallSet<unsigned, 16> MLxHazardOpcodes;
404 public:
405 /// isFpMLxInstruction - Return true if the specified opcode is a fp MLA / MLS
406 /// instruction.
407 bool isFpMLxInstruction(unsigned Opcode) const {
408 return MLxEntryMap.count(Opcode);
411 /// isFpMLxInstruction - This version also returns the multiply opcode and the
412 /// addition / subtraction opcode to expand to. Return true for 'HasLane' for
413 /// the MLX instructions with an extra lane operand.
414 bool isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
415 unsigned &AddSubOpc, bool &NegAcc,
416 bool &HasLane) const;
418 /// canCauseFpMLxStall - Return true if an instruction of the specified opcode
419 /// will cause stalls when scheduled after (within 4-cycle window) a fp
420 /// MLA / MLS instruction.
421 bool canCauseFpMLxStall(unsigned Opcode) const {
422 return MLxHazardOpcodes.count(Opcode);
426 static inline
427 const MachineInstrBuilder &AddDefaultPred(const MachineInstrBuilder &MIB) {
428 return MIB.addImm((int64_t)ARMCC::AL).addReg(0);
431 static inline
432 const MachineInstrBuilder &AddDefaultCC(const MachineInstrBuilder &MIB) {
433 return MIB.addReg(0);
436 static inline
437 const MachineInstrBuilder &AddDefaultT1CC(const MachineInstrBuilder &MIB,
438 bool isDead = false) {
439 return MIB.addReg(ARM::CPSR, getDefRegState(true) | getDeadRegState(isDead));
442 static inline
443 const MachineInstrBuilder &AddNoT1CC(const MachineInstrBuilder &MIB) {
444 return MIB.addReg(0);
447 static inline
448 bool isUncondBranchOpcode(int Opc) {
449 return Opc == ARM::B || Opc == ARM::tB || Opc == ARM::t2B;
452 static inline
453 bool isCondBranchOpcode(int Opc) {
454 return Opc == ARM::Bcc || Opc == ARM::tBcc || Opc == ARM::t2Bcc;
457 static inline
458 bool isJumpTableBranchOpcode(int Opc) {
459 return Opc == ARM::BR_JTr || Opc == ARM::BR_JTm || Opc == ARM::BR_JTadd ||
460 Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT;
463 static inline
464 bool isIndirectBranchOpcode(int Opc) {
465 return Opc == ARM::BX || Opc == ARM::MOVPCRX || Opc == ARM::tBRIND;
468 /// getInstrPredicate - If instruction is predicated, returns its predicate
469 /// condition, otherwise returns AL. It also returns the condition code
470 /// register by reference.
471 ARMCC::CondCodes getInstrPredicate(const MachineInstr *MI, unsigned &PredReg);
473 int getMatchingCondBranchOpcode(int Opc);
475 /// emitARMRegPlusImmediate / emitT2RegPlusImmediate - Emits a series of
476 /// instructions to materializea destreg = basereg + immediate in ARM / Thumb2
477 /// code.
478 void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
479 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
480 unsigned DestReg, unsigned BaseReg, int NumBytes,
481 ARMCC::CondCodes Pred, unsigned PredReg,
482 const ARMBaseInstrInfo &TII, unsigned MIFlags = 0);
484 void emitT2RegPlusImmediate(MachineBasicBlock &MBB,
485 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
486 unsigned DestReg, unsigned BaseReg, int NumBytes,
487 ARMCC::CondCodes Pred, unsigned PredReg,
488 const ARMBaseInstrInfo &TII, unsigned MIFlags = 0);
489 void emitThumbRegPlusImmediate(MachineBasicBlock &MBB,
490 MachineBasicBlock::iterator &MBBI, DebugLoc dl,
491 unsigned DestReg, unsigned BaseReg,
492 int NumBytes, const TargetInstrInfo &TII,
493 const ARMBaseRegisterInfo& MRI,
494 unsigned MIFlags = 0);
497 /// rewriteARMFrameIndex / rewriteT2FrameIndex -
498 /// Rewrite MI to access 'Offset' bytes from the FP. Return false if the
499 /// offset could not be handled directly in MI, and return the left-over
500 /// portion by reference.
501 bool rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
502 unsigned FrameReg, int &Offset,
503 const ARMBaseInstrInfo &TII);
505 bool rewriteT2FrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
506 unsigned FrameReg, int &Offset,
507 const ARMBaseInstrInfo &TII);
509 } // End llvm namespace
511 #endif