ARMBaseInstrInfo.h revision b3a6817d06ab88f410012b5c5cc17fc4f07058c3
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/CodeGen/MachineInstrBuilder.h" 19#include "llvm/Target/TargetInstrInfo.h" 20 21namespace llvm { 22 class ARMSubtarget; 23 class ARMBaseRegisterInfo; 24 25/// ARMII - This namespace holds all of the target specific flags that 26/// instruction info tracks. 27/// 28namespace ARMII { 29 enum { 30 //===------------------------------------------------------------------===// 31 // Instruction Flags. 32 33 //===------------------------------------------------------------------===// 34 // This four-bit field describes the addressing mode used. 35 36 AddrModeMask = 0x1f, 37 AddrModeNone = 0, 38 AddrMode1 = 1, 39 AddrMode2 = 2, 40 AddrMode3 = 3, 41 AddrMode4 = 4, 42 AddrMode5 = 5, 43 AddrMode6 = 6, 44 AddrModeT1_1 = 7, 45 AddrModeT1_2 = 8, 46 AddrModeT1_4 = 9, 47 AddrModeT1_s = 10, // i8 * 4 for pc and sp relative data 48 AddrModeT2_i12 = 11, 49 AddrModeT2_i8 = 12, 50 AddrModeT2_so = 13, 51 AddrModeT2_pc = 14, // +/- i12 for pc relative data 52 AddrModeT2_i8s4 = 15, // i8 * 4 53 54 // Size* - Flags to keep track of the size of an instruction. 55 SizeShift = 5, 56 SizeMask = 7 << SizeShift, 57 SizeSpecial = 1, // 0 byte pseudo or special case. 58 Size8Bytes = 2, 59 Size4Bytes = 3, 60 Size2Bytes = 4, 61 62 // IndexMode - Unindex, pre-indexed, or post-indexed are valid for load 63 // and store ops only. Generic "updating" flag is used for ld/st multiple. 64 IndexModeShift = 8, 65 IndexModeMask = 3 << IndexModeShift, 66 IndexModePre = 1, 67 IndexModePost = 2, 68 IndexModeUpd = 3, 69 70 //===------------------------------------------------------------------===// 71 // Instruction encoding formats. 72 // 73 FormShift = 10, 74 FormMask = 0x3f << FormShift, 75 76 // Pseudo instructions 77 Pseudo = 0 << FormShift, 78 79 // Multiply instructions 80 MulFrm = 1 << FormShift, 81 82 // Branch instructions 83 BrFrm = 2 << FormShift, 84 BrMiscFrm = 3 << FormShift, 85 86 // Data Processing instructions 87 DPFrm = 4 << FormShift, 88 DPSoRegFrm = 5 << FormShift, 89 90 // Load and Store 91 LdFrm = 6 << FormShift, 92 StFrm = 7 << FormShift, 93 LdMiscFrm = 8 << FormShift, 94 StMiscFrm = 9 << FormShift, 95 LdStMulFrm = 10 << FormShift, 96 97 LdStExFrm = 11 << FormShift, 98 99 // Miscellaneous arithmetic instructions 100 ArithMiscFrm = 12 << FormShift, 101 SatFrm = 13 << FormShift, 102 103 // Extend instructions 104 ExtFrm = 14 << FormShift, 105 106 // VFP formats 107 VFPUnaryFrm = 15 << FormShift, 108 VFPBinaryFrm = 16 << FormShift, 109 VFPConv1Frm = 17 << FormShift, 110 VFPConv2Frm = 18 << FormShift, 111 VFPConv3Frm = 19 << FormShift, 112 VFPConv4Frm = 20 << FormShift, 113 VFPConv5Frm = 21 << FormShift, 114 VFPLdStFrm = 22 << FormShift, 115 VFPLdStMulFrm = 23 << FormShift, 116 VFPMiscFrm = 24 << FormShift, 117 118 // Thumb format 119 ThumbFrm = 25 << FormShift, 120 121 // Miscelleaneous format 122 MiscFrm = 26 << FormShift, 123 124 // NEON formats 125 NGetLnFrm = 27 << FormShift, 126 NSetLnFrm = 28 << FormShift, 127 NDupFrm = 29 << FormShift, 128 NLdStFrm = 30 << FormShift, 129 N1RegModImmFrm= 31 << FormShift, 130 N2RegFrm = 32 << FormShift, 131 NVCVTFrm = 33 << FormShift, 132 NVDupLnFrm = 34 << FormShift, 133 N2RegVShLFrm = 35 << FormShift, 134 N2RegVShRFrm = 36 << FormShift, 135 N3RegFrm = 37 << FormShift, 136 N3RegVShFrm = 38 << FormShift, 137 NVExtFrm = 39 << FormShift, 138 NVMulSLFrm = 40 << FormShift, 139 NVTBLFrm = 41 << FormShift, 140 141 //===------------------------------------------------------------------===// 142 // Misc flags. 143 144 // UnaryDP - Indicates this is a unary data processing instruction, i.e. 145 // it doesn't have a Rn operand. 146 UnaryDP = 1 << 16, 147 148 // Xform16Bit - Indicates this Thumb2 instruction may be transformed into 149 // a 16-bit Thumb instruction if certain conditions are met. 150 Xform16Bit = 1 << 17, 151 152 //===------------------------------------------------------------------===// 153 // Code domain. 154 DomainShift = 18, 155 DomainMask = 3 << DomainShift, 156 DomainGeneral = 0 << DomainShift, 157 DomainVFP = 1 << DomainShift, 158 DomainNEON = 2 << DomainShift, 159 160 //===------------------------------------------------------------------===// 161 // Field shifts - such shifts are used to set field while generating 162 // machine instructions. 163 // 164 // FIXME: This list will need adjusting/fixing as the MC code emitter 165 // takes shape and the ARMCodeEmitter.cpp bits go away. 166 ShiftTypeShift = 4, 167 168 M_BitShift = 5, 169 ShiftImmShift = 5, 170 ShiftShift = 7, 171 N_BitShift = 7, 172 ImmHiShift = 8, 173 SoRotImmShift = 8, 174 RegRsShift = 8, 175 ExtRotImmShift = 10, 176 RegRdLoShift = 12, 177 RegRdShift = 12, 178 RegRdHiShift = 16, 179 RegRnShift = 16, 180 S_BitShift = 20, 181 W_BitShift = 21, 182 AM3_I_BitShift = 22, 183 D_BitShift = 22, 184 U_BitShift = 23, 185 P_BitShift = 24, 186 I_BitShift = 25, 187 CondShift = 28 188 }; 189} 190 191class ARMBaseInstrInfo : public TargetInstrInfoImpl { 192 const ARMSubtarget &Subtarget; 193protected: 194 // Can be only subclassed. 195 explicit ARMBaseInstrInfo(const ARMSubtarget &STI); 196public: 197 // Return the non-pre/post incrementing version of 'Opc'. Return 0 198 // if there is not such an opcode. 199 virtual unsigned getUnindexedOpcode(unsigned Opc) const =0; 200 201 virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI, 202 MachineBasicBlock::iterator &MBBI, 203 LiveVariables *LV) const; 204 205 virtual const ARMBaseRegisterInfo &getRegisterInfo() const =0; 206 const ARMSubtarget &getSubtarget() const { return Subtarget; } 207 208 bool spillCalleeSavedRegisters(MachineBasicBlock &MBB, 209 MachineBasicBlock::iterator MI, 210 const std::vector<CalleeSavedInfo> &CSI, 211 const TargetRegisterInfo *TRI) const; 212 213 // Branch analysis. 214 virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, 215 MachineBasicBlock *&FBB, 216 SmallVectorImpl<MachineOperand> &Cond, 217 bool AllowModify = false) 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; 223 224 virtual 225 bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const; 226 227 // Predication support. 228 bool isPredicated(const MachineInstr *MI) const { 229 int PIdx = MI->findFirstPredOperandIdx(); 230 return PIdx != -1 && MI->getOperand(PIdx).getImm() != ARMCC::AL; 231 } 232 233 ARMCC::CondCodes getPredicate(const MachineInstr *MI) const { 234 int PIdx = MI->findFirstPredOperandIdx(); 235 return PIdx != -1 ? (ARMCC::CondCodes)MI->getOperand(PIdx).getImm() 236 : ARMCC::AL; 237 } 238 239 virtual 240 bool PredicateInstruction(MachineInstr *MI, 241 const SmallVectorImpl<MachineOperand> &Pred) const; 242 243 virtual 244 bool SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1, 245 const SmallVectorImpl<MachineOperand> &Pred2) const; 246 247 virtual bool DefinesPredicate(MachineInstr *MI, 248 std::vector<MachineOperand> &Pred) const; 249 250 virtual bool isPredicable(MachineInstr *MI) const; 251 252 /// GetInstSize - Returns the size of the specified MachineInstr. 253 /// 254 virtual unsigned GetInstSizeInBytes(const MachineInstr* MI) const; 255 256 virtual unsigned isLoadFromStackSlot(const MachineInstr *MI, 257 int &FrameIndex) const; 258 virtual unsigned isStoreToStackSlot(const MachineInstr *MI, 259 int &FrameIndex) const; 260 261 virtual void copyPhysReg(MachineBasicBlock &MBB, 262 MachineBasicBlock::iterator I, DebugLoc DL, 263 unsigned DestReg, unsigned SrcReg, 264 bool KillSrc) const; 265 266 virtual void storeRegToStackSlot(MachineBasicBlock &MBB, 267 MachineBasicBlock::iterator MBBI, 268 unsigned SrcReg, bool isKill, int FrameIndex, 269 const TargetRegisterClass *RC, 270 const TargetRegisterInfo *TRI) const; 271 272 virtual void loadRegFromStackSlot(MachineBasicBlock &MBB, 273 MachineBasicBlock::iterator MBBI, 274 unsigned DestReg, int FrameIndex, 275 const TargetRegisterClass *RC, 276 const TargetRegisterInfo *TRI) const; 277 278 virtual MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF, 279 int FrameIx, 280 uint64_t Offset, 281 const MDNode *MDPtr, 282 DebugLoc DL) const; 283 284 virtual void reMaterialize(MachineBasicBlock &MBB, 285 MachineBasicBlock::iterator MI, 286 unsigned DestReg, unsigned SubIdx, 287 const MachineInstr *Orig, 288 const TargetRegisterInfo &TRI) const; 289 290 MachineInstr *duplicate(MachineInstr *Orig, MachineFunction &MF) const; 291 292 virtual bool produceSameValue(const MachineInstr *MI0, 293 const MachineInstr *MI1) const; 294 295 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to 296 /// determine if two loads are loading from the same base address. It should 297 /// only return true if the base pointers are the same and the only 298 /// differences between the two addresses is the offset. It also returns the 299 /// offsets by reference. 300 virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, 301 int64_t &Offset1, int64_t &Offset2)const; 302 303 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to 304 /// determine (in conjuction with areLoadsFromSameBasePtr) if two loads should 305 /// be scheduled togther. On some targets if two loads are loading from 306 /// addresses in the same cache line, it's better if they are scheduled 307 /// together. This function takes two integers that represent the load offsets 308 /// from the common base address. It returns true if it decides it's desirable 309 /// to schedule the two loads together. "NumLoads" is the number of loads that 310 /// have already been scheduled after Load1. 311 virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2, 312 int64_t Offset1, int64_t Offset2, 313 unsigned NumLoads) const; 314 315 virtual bool isSchedulingBoundary(const MachineInstr *MI, 316 const MachineBasicBlock *MBB, 317 const MachineFunction &MF) const; 318 319 virtual bool isProfitableToIfCvt(MachineBasicBlock &MBB, 320 unsigned NumInstrs, 321 float Prob, float Confidence) const; 322 323 virtual bool isProfitableToIfCvt(MachineBasicBlock &TMBB,unsigned NumT, 324 MachineBasicBlock &FMBB,unsigned NumF, 325 float Probability, float Confidence) const; 326 327 virtual bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB, 328 unsigned NumInstrs, 329 float Probability, 330 float Confidence) const { 331 return NumInstrs == 1; 332 } 333 334 /// AnalyzeCompare - For a comparison instruction, return the source register 335 /// in SrcReg and the value it compares against in CmpValue. Return true if 336 /// the comparison instruction can be analyzed. 337 virtual bool AnalyzeCompare(const MachineInstr *MI, unsigned &SrcReg, 338 int &CmpMask, int &CmpValue) const; 339 340 /// OptimizeCompareInstr - Convert the instruction to set the zero flag so 341 /// that we can remove a "comparison with zero". 342 virtual bool OptimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, 343 int CmpMask, int CmpValue, 344 const MachineRegisterInfo *MRI, 345 MachineBasicBlock::iterator &MII) const; 346 347 virtual unsigned getNumMicroOps(const MachineInstr *MI, 348 const InstrItineraryData *ItinData) const; 349 350 virtual 351 int getOperandLatency(const InstrItineraryData *ItinData, 352 const MachineInstr *DefMI, unsigned DefIdx, 353 const MachineInstr *UseMI, unsigned UseIdx) const; 354 virtual 355 int getOperandLatency(const InstrItineraryData *ItinData, 356 SDNode *DefNode, unsigned DefIdx, 357 SDNode *UseNode, unsigned UseIdx) const; 358private: 359 int getVLDMDefCycle(const InstrItineraryData *ItinData, 360 const TargetInstrDesc &DefTID, 361 unsigned DefClass, 362 unsigned DefIdx, unsigned DefAlign) const; 363 int getLDMDefCycle(const InstrItineraryData *ItinData, 364 const TargetInstrDesc &DefTID, 365 unsigned DefClass, 366 unsigned DefIdx, unsigned DefAlign) const; 367 int getVSTMUseCycle(const InstrItineraryData *ItinData, 368 const TargetInstrDesc &UseTID, 369 unsigned UseClass, 370 unsigned UseIdx, unsigned UseAlign) const; 371 int getSTMUseCycle(const InstrItineraryData *ItinData, 372 const TargetInstrDesc &UseTID, 373 unsigned UseClass, 374 unsigned UseIdx, unsigned UseAlign) const; 375 int getOperandLatency(const InstrItineraryData *ItinData, 376 const TargetInstrDesc &DefTID, 377 unsigned DefIdx, unsigned DefAlign, 378 const TargetInstrDesc &UseTID, 379 unsigned UseIdx, unsigned UseAlign) const; 380 381 bool hasHighOperandLatency(const InstrItineraryData *ItinData, 382 const MachineRegisterInfo *MRI, 383 const MachineInstr *DefMI, unsigned DefIdx, 384 const MachineInstr *UseMI, unsigned UseIdx) const; 385}; 386 387static inline 388const MachineInstrBuilder &AddDefaultPred(const MachineInstrBuilder &MIB) { 389 return MIB.addImm((int64_t)ARMCC::AL).addReg(0); 390} 391 392static inline 393const MachineInstrBuilder &AddDefaultCC(const MachineInstrBuilder &MIB) { 394 return MIB.addReg(0); 395} 396 397static inline 398const MachineInstrBuilder &AddDefaultT1CC(const MachineInstrBuilder &MIB, 399 bool isDead = false) { 400 return MIB.addReg(ARM::CPSR, getDefRegState(true) | getDeadRegState(isDead)); 401} 402 403static inline 404const MachineInstrBuilder &AddNoT1CC(const MachineInstrBuilder &MIB) { 405 return MIB.addReg(0); 406} 407 408static inline 409bool isUncondBranchOpcode(int Opc) { 410 return Opc == ARM::B || Opc == ARM::tB || Opc == ARM::t2B; 411} 412 413static inline 414bool isCondBranchOpcode(int Opc) { 415 return Opc == ARM::Bcc || Opc == ARM::tBcc || Opc == ARM::t2Bcc; 416} 417 418static inline 419bool isJumpTableBranchOpcode(int Opc) { 420 return Opc == ARM::BR_JTr || Opc == ARM::BR_JTm || Opc == ARM::BR_JTadd || 421 Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT; 422} 423 424static inline 425bool isIndirectBranchOpcode(int Opc) { 426 return Opc == ARM::BRIND || Opc == ARM::MOVPCRX || Opc == ARM::tBRIND; 427} 428 429/// getInstrPredicate - If instruction is predicated, returns its predicate 430/// condition, otherwise returns AL. It also returns the condition code 431/// register by reference. 432ARMCC::CondCodes getInstrPredicate(const MachineInstr *MI, unsigned &PredReg); 433 434int getMatchingCondBranchOpcode(int Opc); 435 436/// emitARMRegPlusImmediate / emitT2RegPlusImmediate - Emits a series of 437/// instructions to materializea destreg = basereg + immediate in ARM / Thumb2 438/// code. 439void emitARMRegPlusImmediate(MachineBasicBlock &MBB, 440 MachineBasicBlock::iterator &MBBI, DebugLoc dl, 441 unsigned DestReg, unsigned BaseReg, int NumBytes, 442 ARMCC::CondCodes Pred, unsigned PredReg, 443 const ARMBaseInstrInfo &TII); 444 445void emitT2RegPlusImmediate(MachineBasicBlock &MBB, 446 MachineBasicBlock::iterator &MBBI, DebugLoc dl, 447 unsigned DestReg, unsigned BaseReg, int NumBytes, 448 ARMCC::CondCodes Pred, unsigned PredReg, 449 const ARMBaseInstrInfo &TII); 450void emitThumbRegPlusImmediate(MachineBasicBlock &MBB, 451 MachineBasicBlock::iterator &MBBI, 452 unsigned DestReg, unsigned BaseReg, 453 int NumBytes, const TargetInstrInfo &TII, 454 const ARMBaseRegisterInfo& MRI, 455 DebugLoc dl); 456 457 458/// rewriteARMFrameIndex / rewriteT2FrameIndex - 459/// Rewrite MI to access 'Offset' bytes from the FP. Return false if the 460/// offset could not be handled directly in MI, and return the left-over 461/// portion by reference. 462bool rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx, 463 unsigned FrameReg, int &Offset, 464 const ARMBaseInstrInfo &TII); 465 466bool rewriteT2FrameIndex(MachineInstr &MI, unsigned FrameRegIdx, 467 unsigned FrameReg, int &Offset, 468 const ARMBaseInstrInfo &TII); 469 470} // End llvm namespace 471 472#endif 473