1 //===-- ARMISelLowering.h - ARM DAG Lowering Interface ----------*- 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 defines the interfaces that ARM uses to lower LLVM code into a 11 // selection DAG. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #ifndef LLVM_LIB_TARGET_ARM_ARMISELLOWERING_H 16 #define LLVM_LIB_TARGET_ARM_ARMISELLOWERING_H 17 18 #include "MCTargetDesc/ARMBaseInfo.h" 19 #include "llvm/CodeGen/CallingConvLower.h" 20 #include "llvm/CodeGen/SelectionDAG.h" 21 #include "llvm/Target/TargetLowering.h" 22 #include <vector> 23 24 namespace llvm { 25 class ARMConstantPoolValue; 26 class ARMSubtarget; 27 28 namespace ARMISD { 29 // ARM Specific DAG Nodes 30 enum NodeType : unsigned { 31 // Start the numbering where the builtin ops and target ops leave off. 32 FIRST_NUMBER = ISD::BUILTIN_OP_END, 33 34 Wrapper, // Wrapper - A wrapper node for TargetConstantPool, 35 // TargetExternalSymbol, and TargetGlobalAddress. 36 WrapperPIC, // WrapperPIC - A wrapper node for TargetGlobalAddress in 37 // PIC mode. 38 WrapperJT, // WrapperJT - A wrapper node for TargetJumpTable 39 40 // Add pseudo op to model memcpy for struct byval. 41 COPY_STRUCT_BYVAL, 42 43 CALL, // Function call. 44 CALL_PRED, // Function call that's predicable. 45 CALL_NOLINK, // Function call with branch not branch-and-link. 46 tCALL, // Thumb function call. 47 BRCOND, // Conditional branch. 48 BR_JT, // Jumptable branch. 49 BR2_JT, // Jumptable branch (2 level - jumptable entry is a jump). 50 RET_FLAG, // Return with a flag operand. 51 INTRET_FLAG, // Interrupt return with an LR-offset and a flag operand. 52 53 PIC_ADD, // Add with a PC operand and a PIC label. 54 55 CMP, // ARM compare instructions. 56 CMN, // ARM CMN instructions. 57 CMPZ, // ARM compare that sets only Z flag. 58 CMPFP, // ARM VFP compare instruction, sets FPSCR. 59 CMPFPw0, // ARM VFP compare against zero instruction, sets FPSCR. 60 FMSTAT, // ARM fmstat instruction. 61 62 CMOV, // ARM conditional move instructions. 63 64 BCC_i64, 65 66 SRL_FLAG, // V,Flag = srl_flag X -> srl X, 1 + save carry out. 67 SRA_FLAG, // V,Flag = sra_flag X -> sra X, 1 + save carry out. 68 RRX, // V = RRX X, Flag -> srl X, 1 + shift in carry flag. 69 70 ADDC, // Add with carry 71 ADDE, // Add using carry 72 SUBC, // Sub with carry 73 SUBE, // Sub using carry 74 75 VMOVRRD, // double to two gprs. 76 VMOVDRR, // Two gprs to double. 77 78 EH_SJLJ_SETJMP, // SjLj exception handling setjmp. 79 EH_SJLJ_LONGJMP, // SjLj exception handling longjmp. 80 EH_SJLJ_SETUP_DISPATCH, // SjLj exception handling setup_dispatch. 81 82 TC_RETURN, // Tail call return pseudo. 83 84 THREAD_POINTER, 85 86 DYN_ALLOC, // Dynamic allocation on the stack. 87 88 MEMBARRIER_MCR, // Memory barrier (MCR) 89 90 PRELOAD, // Preload 91 92 WIN__CHKSTK, // Windows' __chkstk call to do stack probing. 93 WIN__DBZCHK, // Windows' divide by zero check 94 95 VCEQ, // Vector compare equal. 96 VCEQZ, // Vector compare equal to zero. 97 VCGE, // Vector compare greater than or equal. 98 VCGEZ, // Vector compare greater than or equal to zero. 99 VCLEZ, // Vector compare less than or equal to zero. 100 VCGEU, // Vector compare unsigned greater than or equal. 101 VCGT, // Vector compare greater than. 102 VCGTZ, // Vector compare greater than zero. 103 VCLTZ, // Vector compare less than zero. 104 VCGTU, // Vector compare unsigned greater than. 105 VTST, // Vector test bits. 106 107 // Vector shift by immediate: 108 VSHL, // ...left 109 VSHRs, // ...right (signed) 110 VSHRu, // ...right (unsigned) 111 112 // Vector rounding shift by immediate: 113 VRSHRs, // ...right (signed) 114 VRSHRu, // ...right (unsigned) 115 VRSHRN, // ...right narrow 116 117 // Vector saturating shift by immediate: 118 VQSHLs, // ...left (signed) 119 VQSHLu, // ...left (unsigned) 120 VQSHLsu, // ...left (signed to unsigned) 121 VQSHRNs, // ...right narrow (signed) 122 VQSHRNu, // ...right narrow (unsigned) 123 VQSHRNsu, // ...right narrow (signed to unsigned) 124 125 // Vector saturating rounding shift by immediate: 126 VQRSHRNs, // ...right narrow (signed) 127 VQRSHRNu, // ...right narrow (unsigned) 128 VQRSHRNsu, // ...right narrow (signed to unsigned) 129 130 // Vector shift and insert: 131 VSLI, // ...left 132 VSRI, // ...right 133 134 // Vector get lane (VMOV scalar to ARM core register) 135 // (These are used for 8- and 16-bit element types only.) 136 VGETLANEu, // zero-extend vector extract element 137 VGETLANEs, // sign-extend vector extract element 138 139 // Vector move immediate and move negated immediate: 140 VMOVIMM, 141 VMVNIMM, 142 143 // Vector move f32 immediate: 144 VMOVFPIMM, 145 146 // Vector duplicate: 147 VDUP, 148 VDUPLANE, 149 150 // Vector shuffles: 151 VEXT, // extract 152 VREV64, // reverse elements within 64-bit doublewords 153 VREV32, // reverse elements within 32-bit words 154 VREV16, // reverse elements within 16-bit halfwords 155 VZIP, // zip (interleave) 156 VUZP, // unzip (deinterleave) 157 VTRN, // transpose 158 VTBL1, // 1-register shuffle with mask 159 VTBL2, // 2-register shuffle with mask 160 161 // Vector multiply long: 162 VMULLs, // ...signed 163 VMULLu, // ...unsigned 164 165 UMLAL, // 64bit Unsigned Accumulate Multiply 166 SMLAL, // 64bit Signed Accumulate Multiply 167 168 // Operands of the standard BUILD_VECTOR node are not legalized, which 169 // is fine if BUILD_VECTORs are always lowered to shuffles or other 170 // operations, but for ARM some BUILD_VECTORs are legal as-is and their 171 // operands need to be legalized. Define an ARM-specific version of 172 // BUILD_VECTOR for this purpose. 173 BUILD_VECTOR, 174 175 // Bit-field insert 176 BFI, 177 178 // Vector OR with immediate 179 VORRIMM, 180 // Vector AND with NOT of immediate 181 VBICIMM, 182 183 // Vector bitwise select 184 VBSL, 185 186 // Pseudo-instruction representing a memory copy using ldm/stm 187 // instructions. 188 MEMCPY, 189 190 // Vector load N-element structure to all lanes: 191 VLD2DUP = ISD::FIRST_TARGET_MEMORY_OPCODE, 192 VLD3DUP, 193 VLD4DUP, 194 195 // NEON loads with post-increment base updates: 196 VLD1_UPD, 197 VLD2_UPD, 198 VLD3_UPD, 199 VLD4_UPD, 200 VLD2LN_UPD, 201 VLD3LN_UPD, 202 VLD4LN_UPD, 203 VLD2DUP_UPD, 204 VLD3DUP_UPD, 205 VLD4DUP_UPD, 206 207 // NEON stores with post-increment base updates: 208 VST1_UPD, 209 VST2_UPD, 210 VST3_UPD, 211 VST4_UPD, 212 VST2LN_UPD, 213 VST3LN_UPD, 214 VST4LN_UPD 215 }; 216 } 217 218 /// Define some predicates that are used for node matching. 219 namespace ARM { 220 bool isBitFieldInvertedMask(unsigned v); 221 } 222 223 //===--------------------------------------------------------------------===// 224 // ARMTargetLowering - ARM Implementation of the TargetLowering interface 225 226 class ARMTargetLowering : public TargetLowering { 227 public: 228 explicit ARMTargetLowering(const TargetMachine &TM, 229 const ARMSubtarget &STI); 230 231 unsigned getJumpTableEncoding() const override; 232 bool useSoftFloat() const override; 233 234 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override; 235 236 /// ReplaceNodeResults - Replace the results of node with an illegal result 237 /// type with new values built out of custom code. 238 /// 239 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results, 240 SelectionDAG &DAG) const override; 241 242 const char *getTargetNodeName(unsigned Opcode) const override; 243 isSelectSupported(SelectSupportKind Kind)244 bool isSelectSupported(SelectSupportKind Kind) const override { 245 // ARM does not support scalar condition selects on vectors. 246 return (Kind != ScalarCondVectorVal); 247 } 248 249 /// getSetCCResultType - Return the value type to use for ISD::SETCC. 250 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context, 251 EVT VT) const override; 252 253 MachineBasicBlock * 254 EmitInstrWithCustomInserter(MachineInstr *MI, 255 MachineBasicBlock *MBB) const override; 256 257 void AdjustInstrPostInstrSelection(MachineInstr *MI, 258 SDNode *Node) const override; 259 260 SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const; 261 SDValue PerformCMOVToBFICombine(SDNode *N, SelectionDAG &DAG) const; 262 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override; 263 264 bool isDesirableToTransformToIntegerOp(unsigned Opc, EVT VT) const override; 265 266 /// allowsMisalignedMemoryAccesses - Returns true if the target allows 267 /// unaligned memory accesses of the specified type. Returns whether it 268 /// is "fast" by reference in the second argument. 269 bool allowsMisalignedMemoryAccesses(EVT VT, unsigned AddrSpace, 270 unsigned Align, 271 bool *Fast) const override; 272 273 EVT getOptimalMemOpType(uint64_t Size, 274 unsigned DstAlign, unsigned SrcAlign, 275 bool IsMemset, bool ZeroMemset, 276 bool MemcpyStrSrc, 277 MachineFunction &MF) const override; 278 279 using TargetLowering::isZExtFree; 280 bool isZExtFree(SDValue Val, EVT VT2) const override; 281 282 bool isVectorLoadExtDesirable(SDValue ExtVal) const override; 283 284 bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const override; 285 286 287 /// isLegalAddressingMode - Return true if the addressing mode represented 288 /// by AM is legal for this target, for a load/store of the specified type. 289 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, 290 Type *Ty, unsigned AS) const override; 291 bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const; 292 293 /// isLegalICmpImmediate - Return true if the specified immediate is legal 294 /// icmp immediate, that is the target has icmp instructions which can 295 /// compare a register against the immediate without having to materialize 296 /// the immediate into a register. 297 bool isLegalICmpImmediate(int64_t Imm) const override; 298 299 /// isLegalAddImmediate - Return true if the specified immediate is legal 300 /// add immediate, that is the target has add instructions which can 301 /// add a register and the immediate without having to materialize 302 /// the immediate into a register. 303 bool isLegalAddImmediate(int64_t Imm) const override; 304 305 /// getPreIndexedAddressParts - returns true by value, base pointer and 306 /// offset pointer and addressing mode by reference if the node's address 307 /// can be legally represented as pre-indexed load / store address. 308 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, SDValue &Offset, 309 ISD::MemIndexedMode &AM, 310 SelectionDAG &DAG) const override; 311 312 /// getPostIndexedAddressParts - returns true by value, base pointer and 313 /// offset pointer and addressing mode by reference if this node can be 314 /// combined with a load / store to form a post-indexed load / store. 315 bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, SDValue &Base, 316 SDValue &Offset, ISD::MemIndexedMode &AM, 317 SelectionDAG &DAG) const override; 318 319 void computeKnownBitsForTargetNode(const SDValue Op, APInt &KnownZero, 320 APInt &KnownOne, 321 const SelectionDAG &DAG, 322 unsigned Depth) const override; 323 324 325 bool ExpandInlineAsm(CallInst *CI) const override; 326 327 ConstraintType getConstraintType(StringRef Constraint) const override; 328 329 /// Examine constraint string and operand type and determine a weight value. 330 /// The operand object must already have been set up with the operand type. 331 ConstraintWeight getSingleConstraintMatchWeight( 332 AsmOperandInfo &info, const char *constraint) const override; 333 334 std::pair<unsigned, const TargetRegisterClass *> 335 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, 336 StringRef Constraint, MVT VT) const override; 337 338 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 339 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is 340 /// true it means one of the asm constraint of the inline asm instruction 341 /// being processed is 'm'. 342 void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint, 343 std::vector<SDValue> &Ops, 344 SelectionDAG &DAG) const override; 345 346 unsigned getInlineAsmMemConstraint(StringRef ConstraintCode)347 getInlineAsmMemConstraint(StringRef ConstraintCode) const override { 348 if (ConstraintCode == "Q") 349 return InlineAsm::Constraint_Q; 350 else if (ConstraintCode == "o") 351 return InlineAsm::Constraint_o; 352 else if (ConstraintCode.size() == 2) { 353 if (ConstraintCode[0] == 'U') { 354 switch(ConstraintCode[1]) { 355 default: 356 break; 357 case 'm': 358 return InlineAsm::Constraint_Um; 359 case 'n': 360 return InlineAsm::Constraint_Un; 361 case 'q': 362 return InlineAsm::Constraint_Uq; 363 case 's': 364 return InlineAsm::Constraint_Us; 365 case 't': 366 return InlineAsm::Constraint_Ut; 367 case 'v': 368 return InlineAsm::Constraint_Uv; 369 case 'y': 370 return InlineAsm::Constraint_Uy; 371 } 372 } 373 } 374 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode); 375 } 376 getSubtarget()377 const ARMSubtarget* getSubtarget() const { 378 return Subtarget; 379 } 380 381 /// getRegClassFor - Return the register class that should be used for the 382 /// specified value type. 383 const TargetRegisterClass *getRegClassFor(MVT VT) const override; 384 385 /// Returns true if a cast between SrcAS and DestAS is a noop. isNoopAddrSpaceCast(unsigned SrcAS,unsigned DestAS)386 bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override { 387 // Addrspacecasts are always noops. 388 return true; 389 } 390 391 bool shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize, 392 unsigned &PrefAlign) const override; 393 394 /// createFastISel - This method returns a target specific FastISel object, 395 /// or null if the target does not support "fast" ISel. 396 FastISel *createFastISel(FunctionLoweringInfo &funcInfo, 397 const TargetLibraryInfo *libInfo) const override; 398 399 Sched::Preference getSchedulingPreference(SDNode *N) const override; 400 401 bool 402 isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const override; 403 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override; 404 405 /// isFPImmLegal - Returns true if the target can instruction select the 406 /// specified FP immediate natively. If false, the legalizer will 407 /// materialize the FP immediate as a load from a constant pool. 408 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override; 409 410 bool getTgtMemIntrinsic(IntrinsicInfo &Info, 411 const CallInst &I, 412 unsigned Intrinsic) const override; 413 414 /// \brief Returns true if it is beneficial to convert a load of a constant 415 /// to just the constant itself. 416 bool shouldConvertConstantLoadToIntImm(const APInt &Imm, 417 Type *Ty) const override; 418 419 /// \brief Returns true if an argument of type Ty needs to be passed in a 420 /// contiguous block of registers in calling convention CallConv. 421 bool functionArgumentNeedsConsecutiveRegisters( 422 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const override; 423 424 /// If a physical register, this returns the register that receives the 425 /// exception address on entry to an EH pad. 426 unsigned 427 getExceptionPointerRegister(const Constant *PersonalityFn) const override; 428 429 /// If a physical register, this returns the register that receives the 430 /// exception typeid on entry to a landing pad. 431 unsigned 432 getExceptionSelectorRegister(const Constant *PersonalityFn) const override; 433 434 Instruction *makeDMB(IRBuilder<> &Builder, ARM_MB::MemBOpt Domain) const; 435 Value *emitLoadLinked(IRBuilder<> &Builder, Value *Addr, 436 AtomicOrdering Ord) const override; 437 Value *emitStoreConditional(IRBuilder<> &Builder, Value *Val, 438 Value *Addr, AtomicOrdering Ord) const override; 439 440 void emitAtomicCmpXchgNoStoreLLBalance(IRBuilder<> &Builder) const override; 441 442 Instruction* emitLeadingFence(IRBuilder<> &Builder, AtomicOrdering Ord, 443 bool IsStore, bool IsLoad) const override; 444 Instruction* emitTrailingFence(IRBuilder<> &Builder, AtomicOrdering Ord, 445 bool IsStore, bool IsLoad) const override; 446 getMaxSupportedInterleaveFactor()447 unsigned getMaxSupportedInterleaveFactor() const override { return 4; } 448 449 bool lowerInterleavedLoad(LoadInst *LI, 450 ArrayRef<ShuffleVectorInst *> Shuffles, 451 ArrayRef<unsigned> Indices, 452 unsigned Factor) const override; 453 bool lowerInterleavedStore(StoreInst *SI, ShuffleVectorInst *SVI, 454 unsigned Factor) const override; 455 456 TargetLoweringBase::AtomicExpansionKind 457 shouldExpandAtomicLoadInIR(LoadInst *LI) const override; 458 bool shouldExpandAtomicStoreInIR(StoreInst *SI) const override; 459 TargetLoweringBase::AtomicExpansionKind 460 shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const override; 461 bool shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override; 462 463 bool useLoadStackGuardNode() const override; 464 465 bool canCombineStoreAndExtract(Type *VectorTy, Value *Idx, 466 unsigned &Cost) const override; 467 468 bool isCheapToSpeculateCttz() const override; 469 bool isCheapToSpeculateCtlz() const override; 470 471 protected: 472 std::pair<const TargetRegisterClass *, uint8_t> 473 findRepresentativeClass(const TargetRegisterInfo *TRI, 474 MVT VT) const override; 475 476 private: 477 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can 478 /// make the right decision when generating code for different targets. 479 const ARMSubtarget *Subtarget; 480 481 const TargetRegisterInfo *RegInfo; 482 483 const InstrItineraryData *Itins; 484 485 /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created. 486 /// 487 unsigned ARMPCLabelIndex; 488 489 void addTypeForNEON(MVT VT, MVT PromotedLdStVT, MVT PromotedBitwiseVT); 490 void addDRTypeForNEON(MVT VT); 491 void addQRTypeForNEON(MVT VT); 492 std::pair<SDValue, SDValue> getARMXALUOOp(SDValue Op, SelectionDAG &DAG, SDValue &ARMcc) const; 493 494 typedef SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPassVector; 495 void PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG, 496 SDValue Chain, SDValue &Arg, 497 RegsToPassVector &RegsToPass, 498 CCValAssign &VA, CCValAssign &NextVA, 499 SDValue &StackPtr, 500 SmallVectorImpl<SDValue> &MemOpChains, 501 ISD::ArgFlagsTy Flags) const; 502 SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, 503 SDValue &Root, SelectionDAG &DAG, 504 SDLoc dl) const; 505 506 CallingConv::ID getEffectiveCallingConv(CallingConv::ID CC, 507 bool isVarArg) const; 508 CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return, 509 bool isVarArg) const; 510 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg, 511 SDLoc dl, SelectionDAG &DAG, 512 const CCValAssign &VA, 513 ISD::ArgFlagsTy Flags) const; 514 SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const; 515 SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const; 516 SDValue LowerEH_SJLJ_SETUP_DISPATCH(SDValue Op, SelectionDAG &DAG) const; 517 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG, 518 const ARMSubtarget *Subtarget) const; 519 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const; 520 SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const; 521 SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const; 522 SDValue LowerGlobalAddressWindows(SDValue Op, SelectionDAG &DAG) const; 523 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const; 524 SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, 525 SelectionDAG &DAG) const; 526 SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA, 527 SelectionDAG &DAG, 528 TLSModel::Model model) const; 529 SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const; 530 SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) const; 531 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const; 532 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const; 533 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const; 534 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const; 535 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; 536 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; 537 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const; 538 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const; 539 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const; 540 SDValue LowerConstantFP(SDValue Op, SelectionDAG &DAG, 541 const ARMSubtarget *ST) const; 542 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG, 543 const ARMSubtarget *ST) const; 544 SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const; 545 SDValue LowerDivRem(SDValue Op, SelectionDAG &DAG) const; 546 SDValue LowerDIV_Windows(SDValue Op, SelectionDAG &DAG, bool Signed) const; 547 void ExpandDIV_Windows(SDValue Op, SelectionDAG &DAG, bool Signed, 548 SmallVectorImpl<SDValue> &Results) const; 549 SDValue LowerWindowsDIVLibCall(SDValue Op, SelectionDAG &DAG, bool Signed, 550 SDValue &Chain) const; 551 SDValue LowerREM(SDNode *N, SelectionDAG &DAG) const; 552 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const; 553 SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const; 554 SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const; 555 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const; 556 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; 557 558 unsigned getRegisterByName(const char* RegName, EVT VT, 559 SelectionDAG &DAG) const override; 560 561 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster 562 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be 563 /// expanded to FMAs when this method returns true, otherwise fmuladd is 564 /// expanded to fmul + fadd. 565 /// 566 /// ARM supports both fused and unfused multiply-add operations; we already 567 /// lower a pair of fmul and fadd to the latter so it's not clear that there 568 /// would be a gain or that the gain would be worthwhile enough to risk 569 /// correctness bugs. isFMAFasterThanFMulAndFAdd(EVT VT)570 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override { return false; } 571 572 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const; 573 574 SDValue LowerCallResult(SDValue Chain, SDValue InFlag, 575 CallingConv::ID CallConv, bool isVarArg, 576 const SmallVectorImpl<ISD::InputArg> &Ins, 577 SDLoc dl, SelectionDAG &DAG, 578 SmallVectorImpl<SDValue> &InVals, 579 bool isThisReturn, SDValue ThisVal) const; 580 581 SDValue 582 LowerFormalArguments(SDValue Chain, 583 CallingConv::ID CallConv, bool isVarArg, 584 const SmallVectorImpl<ISD::InputArg> &Ins, 585 SDLoc dl, SelectionDAG &DAG, 586 SmallVectorImpl<SDValue> &InVals) const override; 587 588 int StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG, 589 SDLoc dl, SDValue &Chain, 590 const Value *OrigArg, 591 unsigned InRegsParamRecordIdx, 592 int ArgOffset, 593 unsigned ArgSize) const; 594 595 void VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG, 596 SDLoc dl, SDValue &Chain, 597 unsigned ArgOffset, 598 unsigned TotalArgRegsSaveSize, 599 bool ForceMutable = false) const; 600 601 SDValue 602 LowerCall(TargetLowering::CallLoweringInfo &CLI, 603 SmallVectorImpl<SDValue> &InVals) const override; 604 605 /// HandleByVal - Target-specific cleanup for ByVal support. 606 void HandleByVal(CCState *, unsigned &, unsigned) const override; 607 608 /// IsEligibleForTailCallOptimization - Check whether the call is eligible 609 /// for tail call optimization. Targets which want to do tail call 610 /// optimization should implement this function. 611 bool IsEligibleForTailCallOptimization(SDValue Callee, 612 CallingConv::ID CalleeCC, 613 bool isVarArg, 614 bool isCalleeStructRet, 615 bool isCallerStructRet, 616 const SmallVectorImpl<ISD::OutputArg> &Outs, 617 const SmallVectorImpl<SDValue> &OutVals, 618 const SmallVectorImpl<ISD::InputArg> &Ins, 619 SelectionDAG& DAG) const; 620 621 bool CanLowerReturn(CallingConv::ID CallConv, 622 MachineFunction &MF, bool isVarArg, 623 const SmallVectorImpl<ISD::OutputArg> &Outs, 624 LLVMContext &Context) const override; 625 626 SDValue 627 LowerReturn(SDValue Chain, 628 CallingConv::ID CallConv, bool isVarArg, 629 const SmallVectorImpl<ISD::OutputArg> &Outs, 630 const SmallVectorImpl<SDValue> &OutVals, 631 SDLoc dl, SelectionDAG &DAG) const override; 632 633 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override; 634 635 bool mayBeEmittedAsTailCall(CallInst *CI) const override; 636 637 SDValue getCMOV(SDLoc dl, EVT VT, SDValue FalseVal, SDValue TrueVal, 638 SDValue ARMcc, SDValue CCR, SDValue Cmp, 639 SelectionDAG &DAG) const; 640 SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, 641 SDValue &ARMcc, SelectionDAG &DAG, SDLoc dl) const; 642 SDValue getVFPCmp(SDValue LHS, SDValue RHS, 643 SelectionDAG &DAG, SDLoc dl) const; 644 SDValue duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const; 645 646 SDValue OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const; 647 648 void SetupEntryBlockForSjLj(MachineInstr *MI, 649 MachineBasicBlock *MBB, 650 MachineBasicBlock *DispatchBB, int FI) const; 651 652 void EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const; 653 654 bool RemapAddSubWithFlags(MachineInstr *MI, MachineBasicBlock *BB) const; 655 656 MachineBasicBlock *EmitStructByval(MachineInstr *MI, 657 MachineBasicBlock *MBB) const; 658 659 MachineBasicBlock *EmitLowered__chkstk(MachineInstr *MI, 660 MachineBasicBlock *MBB) const; 661 MachineBasicBlock *EmitLowered__dbzchk(MachineInstr *MI, 662 MachineBasicBlock *MBB) const; 663 }; 664 665 enum NEONModImmType { 666 VMOVModImm, 667 VMVNModImm, 668 OtherModImm 669 }; 670 671 namespace ARM { 672 FastISel *createFastISel(FunctionLoweringInfo &funcInfo, 673 const TargetLibraryInfo *libInfo); 674 } 675 } 676 677 #endif // ARMISELLOWERING_H 678