ARMISelLowering.h revision 4d3f3294535a3b622c715f2d9675d4f3e86c3378
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 ARMISELLOWERING_H 16#define ARMISELLOWERING_H 17 18#include "ARMSubtarget.h" 19#include "llvm/Target/TargetLowering.h" 20#include "llvm/Target/TargetRegisterInfo.h" 21#include "llvm/CodeGen/FastISel.h" 22#include "llvm/CodeGen/SelectionDAG.h" 23#include "llvm/CodeGen/CallingConvLower.h" 24#include <vector> 25 26namespace llvm { 27 class ARMConstantPoolValue; 28 29 namespace ARMISD { 30 // ARM Specific DAG Nodes 31 enum NodeType { 32 // Start the numbering where the builtin ops and target ops leave off. 33 FIRST_NUMBER = ISD::BUILTIN_OP_END, 34 35 Wrapper, // Wrapper - A wrapper node for TargetConstantPool, 36 // TargetExternalSymbol, and TargetGlobalAddress. 37 WrapperDYN, // WrapperDYN - A wrapper node for TargetGlobalAddress in 38 // DYN mode. 39 WrapperPIC, // WrapperPIC - A wrapper node for TargetGlobalAddress in 40 // PIC mode. 41 WrapperJT, // WrapperJT - A wrapper node for TargetJumpTable 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 52 PIC_ADD, // Add with a PC operand and a PIC label. 53 54 CMP, // ARM compare instructions. 55 CMPZ, // ARM compare that sets only Z flag. 56 CMPFP, // ARM VFP compare instruction, sets FPSCR. 57 CMPFPw0, // ARM VFP compare against zero instruction, sets FPSCR. 58 FMSTAT, // ARM fmstat instruction. 59 CMOV, // ARM conditional move instructions. 60 61 BCC_i64, 62 63 RBIT, // ARM bitreverse instruction 64 65 FTOSI, // FP to sint within a FP register. 66 FTOUI, // FP to uint within a FP register. 67 SITOF, // sint to FP within a FP register. 68 UITOF, // uint to FP within a FP register. 69 70 SRL_FLAG, // V,Flag = srl_flag X -> srl X, 1 + save carry out. 71 SRA_FLAG, // V,Flag = sra_flag X -> sra X, 1 + save carry out. 72 RRX, // V = RRX X, Flag -> srl X, 1 + shift in carry flag. 73 74 ADDC, // Add with carry 75 ADDE, // Add using carry 76 SUBC, // Sub with carry 77 SUBE, // Sub using carry 78 79 VMOVRRD, // double to two gprs. 80 VMOVDRR, // Two gprs to double. 81 82 EH_SJLJ_SETJMP, // SjLj exception handling setjmp. 83 EH_SJLJ_LONGJMP, // SjLj exception handling longjmp. 84 EH_SJLJ_DISPATCHSETUP, // SjLj exception handling dispatch setup. 85 86 TC_RETURN, // Tail call return pseudo. 87 88 THREAD_POINTER, 89 90 DYN_ALLOC, // Dynamic allocation on the stack. 91 92 MEMBARRIER, // Memory barrier (DMB) 93 MEMBARRIER_MCR, // Memory barrier (MCR) 94 95 PRELOAD, // Preload 96 97 VCEQ, // Vector compare equal. 98 VCEQZ, // Vector compare equal to zero. 99 VCGE, // Vector compare greater than or equal. 100 VCGEZ, // Vector compare greater than or equal to zero. 101 VCLEZ, // Vector compare less than or equal to zero. 102 VCGEU, // Vector compare unsigned greater than or equal. 103 VCGT, // Vector compare greater than. 104 VCGTZ, // Vector compare greater than zero. 105 VCLTZ, // Vector compare less than zero. 106 VCGTU, // Vector compare unsigned greater than. 107 VTST, // Vector test bits. 108 109 // Vector shift by immediate: 110 VSHL, // ...left 111 VSHRs, // ...right (signed) 112 VSHRu, // ...right (unsigned) 113 VSHLLs, // ...left long (signed) 114 VSHLLu, // ...left long (unsigned) 115 VSHLLi, // ...left long (with maximum shift count) 116 VSHRN, // ...right narrow 117 118 // Vector rounding shift by immediate: 119 VRSHRs, // ...right (signed) 120 VRSHRu, // ...right (unsigned) 121 VRSHRN, // ...right narrow 122 123 // Vector saturating shift by immediate: 124 VQSHLs, // ...left (signed) 125 VQSHLu, // ...left (unsigned) 126 VQSHLsu, // ...left (signed to unsigned) 127 VQSHRNs, // ...right narrow (signed) 128 VQSHRNu, // ...right narrow (unsigned) 129 VQSHRNsu, // ...right narrow (signed to unsigned) 130 131 // Vector saturating rounding shift by immediate: 132 VQRSHRNs, // ...right narrow (signed) 133 VQRSHRNu, // ...right narrow (unsigned) 134 VQRSHRNsu, // ...right narrow (signed to unsigned) 135 136 // Vector shift and insert: 137 VSLI, // ...left 138 VSRI, // ...right 139 140 // Vector get lane (VMOV scalar to ARM core register) 141 // (These are used for 8- and 16-bit element types only.) 142 VGETLANEu, // zero-extend vector extract element 143 VGETLANEs, // sign-extend vector extract element 144 145 // Vector move immediate and move negated immediate: 146 VMOVIMM, 147 VMVNIMM, 148 149 // Vector duplicate: 150 VDUP, 151 VDUPLANE, 152 153 // Vector shuffles: 154 VEXT, // extract 155 VREV64, // reverse elements within 64-bit doublewords 156 VREV32, // reverse elements within 32-bit words 157 VREV16, // reverse elements within 16-bit halfwords 158 VZIP, // zip (interleave) 159 VUZP, // unzip (deinterleave) 160 VTRN, // transpose 161 VTBL1, // 1-register shuffle with mask 162 VTBL2, // 2-register shuffle with mask 163 164 // Vector multiply long: 165 VMULLs, // ...signed 166 VMULLu, // ...unsigned 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 // Floating-point max and min: 176 FMAX, 177 FMIN, 178 179 // Bit-field insert 180 BFI, 181 182 // Vector OR with immediate 183 VORRIMM, 184 // Vector AND with NOT of immediate 185 VBICIMM, 186 187 // Vector bitwise select 188 VBSL, 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 // 64-bit atomic ops (value split into two registers) 217 ATOMADD64_DAG, 218 ATOMSUB64_DAG, 219 ATOMOR64_DAG, 220 ATOMXOR64_DAG, 221 ATOMAND64_DAG, 222 ATOMNAND64_DAG, 223 ATOMSWAP64_DAG, 224 ATOMCMPXCHG64_DAG 225 }; 226 } 227 228 /// Define some predicates that are used for node matching. 229 namespace ARM { 230 /// getVFPf32Imm / getVFPf64Imm - If the given fp immediate can be 231 /// materialized with a VMOV.f32 / VMOV.f64 (i.e. fconsts / fconstd) 232 /// instruction, returns its 8-bit integer representation. Otherwise, 233 /// returns -1. 234 int getVFPf32Imm(const APFloat &FPImm); 235 int getVFPf64Imm(const APFloat &FPImm); 236 bool isBitFieldInvertedMask(unsigned v); 237 } 238 239 //===--------------------------------------------------------------------===// 240 // ARMTargetLowering - ARM Implementation of the TargetLowering interface 241 242 class ARMTargetLowering : public TargetLowering { 243 public: 244 explicit ARMTargetLowering(TargetMachine &TM); 245 246 virtual unsigned getJumpTableEncoding(void) const; 247 248 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const; 249 250 /// ReplaceNodeResults - Replace the results of node with an illegal result 251 /// type with new values built out of custom code. 252 /// 253 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results, 254 SelectionDAG &DAG) const; 255 256 virtual const char *getTargetNodeName(unsigned Opcode) const; 257 258 virtual MachineBasicBlock * 259 EmitInstrWithCustomInserter(MachineInstr *MI, 260 MachineBasicBlock *MBB) const; 261 262 virtual void 263 AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const; 264 265 SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const; 266 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const; 267 268 bool isDesirableToTransformToIntegerOp(unsigned Opc, EVT VT) const; 269 270 /// allowsUnalignedMemoryAccesses - Returns true if the target allows 271 /// unaligned memory accesses. of the specified type. 272 /// FIXME: Add getOptimalMemOpType to implement memcpy with NEON? 273 virtual bool allowsUnalignedMemoryAccesses(EVT VT) const; 274 275 /// isLegalAddressingMode - Return true if the addressing mode represented 276 /// by AM is legal for this target, for a load/store of the specified type. 277 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty)const; 278 bool isLegalT2ScaledAddressingMode(const AddrMode &AM, EVT VT) const; 279 280 /// isLegalICmpImmediate - Return true if the specified immediate is legal 281 /// icmp immediate, that is the target has icmp instructions which can 282 /// compare a register against the immediate without having to materialize 283 /// the immediate into a register. 284 virtual bool isLegalICmpImmediate(int64_t Imm) const; 285 286 /// isLegalAddImmediate - Return true if the specified immediate is legal 287 /// add immediate, that is the target has add instructions which can 288 /// add a register and the immediate without having to materialize 289 /// the immediate into a register. 290 virtual bool isLegalAddImmediate(int64_t Imm) const; 291 292 /// getPreIndexedAddressParts - returns true by value, base pointer and 293 /// offset pointer and addressing mode by reference if the node's address 294 /// can be legally represented as pre-indexed load / store address. 295 virtual bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, 296 SDValue &Offset, 297 ISD::MemIndexedMode &AM, 298 SelectionDAG &DAG) const; 299 300 /// getPostIndexedAddressParts - returns true by value, base pointer and 301 /// offset pointer and addressing mode by reference if this node can be 302 /// combined with a load / store to form a post-indexed load / store. 303 virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, 304 SDValue &Base, SDValue &Offset, 305 ISD::MemIndexedMode &AM, 306 SelectionDAG &DAG) const; 307 308 virtual void computeMaskedBitsForTargetNode(const SDValue Op, 309 const APInt &Mask, 310 APInt &KnownZero, 311 APInt &KnownOne, 312 const SelectionDAG &DAG, 313 unsigned Depth) const; 314 315 316 virtual bool ExpandInlineAsm(CallInst *CI) const; 317 318 ConstraintType getConstraintType(const std::string &Constraint) const; 319 320 /// Examine constraint string and operand type and determine a weight value. 321 /// The operand object must already have been set up with the operand type. 322 ConstraintWeight getSingleConstraintMatchWeight( 323 AsmOperandInfo &info, const char *constraint) const; 324 325 std::pair<unsigned, const TargetRegisterClass*> 326 getRegForInlineAsmConstraint(const std::string &Constraint, 327 EVT VT) const; 328 329 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 330 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is 331 /// true it means one of the asm constraint of the inline asm instruction 332 /// being processed is 'm'. 333 virtual void LowerAsmOperandForConstraint(SDValue Op, 334 std::string &Constraint, 335 std::vector<SDValue> &Ops, 336 SelectionDAG &DAG) const; 337 338 const ARMSubtarget* getSubtarget() const { 339 return Subtarget; 340 } 341 342 /// getRegClassFor - Return the register class that should be used for the 343 /// specified value type. 344 virtual TargetRegisterClass *getRegClassFor(EVT VT) const; 345 346 /// getMaximalGlobalOffset - Returns the maximal possible offset which can 347 /// be used for loads / stores from the global. 348 virtual unsigned getMaximalGlobalOffset() const; 349 350 /// createFastISel - This method returns a target specific FastISel object, 351 /// or null if the target does not support "fast" ISel. 352 virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo) const; 353 354 Sched::Preference getSchedulingPreference(SDNode *N) const; 355 356 bool isShuffleMaskLegal(const SmallVectorImpl<int> &M, EVT VT) const; 357 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const; 358 359 /// isFPImmLegal - Returns true if the target can instruction select the 360 /// specified FP immediate natively. If false, the legalizer will 361 /// materialize the FP immediate as a load from a constant pool. 362 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const; 363 364 virtual bool getTgtMemIntrinsic(IntrinsicInfo &Info, 365 const CallInst &I, 366 unsigned Intrinsic) const; 367 protected: 368 std::pair<const TargetRegisterClass*, uint8_t> 369 findRepresentativeClass(EVT VT) const; 370 371 private: 372 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can 373 /// make the right decision when generating code for different targets. 374 const ARMSubtarget *Subtarget; 375 376 const TargetRegisterInfo *RegInfo; 377 378 const InstrItineraryData *Itins; 379 380 /// ARMPCLabelIndex - Keep track of the number of ARM PC labels created. 381 /// 382 unsigned ARMPCLabelIndex; 383 384 void addTypeForNEON(EVT VT, EVT PromotedLdStVT, EVT PromotedBitwiseVT); 385 void addDRTypeForNEON(EVT VT); 386 void addQRTypeForNEON(EVT VT); 387 388 typedef SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPassVector; 389 void PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG, 390 SDValue Chain, SDValue &Arg, 391 RegsToPassVector &RegsToPass, 392 CCValAssign &VA, CCValAssign &NextVA, 393 SDValue &StackPtr, 394 SmallVector<SDValue, 8> &MemOpChains, 395 ISD::ArgFlagsTy Flags) const; 396 SDValue GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, 397 SDValue &Root, SelectionDAG &DAG, 398 DebugLoc dl) const; 399 400 CCAssignFn *CCAssignFnForNode(CallingConv::ID CC, bool Return, 401 bool isVarArg) const; 402 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg, 403 DebugLoc dl, SelectionDAG &DAG, 404 const CCValAssign &VA, 405 ISD::ArgFlagsTy Flags) const; 406 SDValue LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const; 407 SDValue LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const; 408 SDValue LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG) const; 409 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG, 410 const ARMSubtarget *Subtarget) const; 411 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const; 412 SDValue LowerGlobalAddressDarwin(SDValue Op, SelectionDAG &DAG) const; 413 SDValue LowerGlobalAddressELF(SDValue Op, SelectionDAG &DAG) const; 414 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const; 415 SDValue LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, 416 SelectionDAG &DAG) const; 417 SDValue LowerToTLSExecModels(GlobalAddressSDNode *GA, 418 SelectionDAG &DAG) const; 419 SDValue LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) const; 420 SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const; 421 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const; 422 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const; 423 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const; 424 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const; 425 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; 426 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; 427 SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const; 428 SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const; 429 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const; 430 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG, 431 const ARMSubtarget *ST) const; 432 433 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const; 434 435 SDValue LowerCallResult(SDValue Chain, SDValue InFlag, 436 CallingConv::ID CallConv, bool isVarArg, 437 const SmallVectorImpl<ISD::InputArg> &Ins, 438 DebugLoc dl, SelectionDAG &DAG, 439 SmallVectorImpl<SDValue> &InVals) const; 440 441 virtual SDValue 442 LowerFormalArguments(SDValue Chain, 443 CallingConv::ID CallConv, bool isVarArg, 444 const SmallVectorImpl<ISD::InputArg> &Ins, 445 DebugLoc dl, SelectionDAG &DAG, 446 SmallVectorImpl<SDValue> &InVals) const; 447 448 void VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG, 449 DebugLoc dl, SDValue &Chain, unsigned ArgOffset) 450 const; 451 452 void computeRegArea(CCState &CCInfo, MachineFunction &MF, 453 unsigned &VARegSize, unsigned &VARegSaveSize) const; 454 455 virtual SDValue 456 LowerCall(SDValue Chain, SDValue Callee, 457 CallingConv::ID CallConv, bool isVarArg, 458 bool &isTailCall, 459 const SmallVectorImpl<ISD::OutputArg> &Outs, 460 const SmallVectorImpl<SDValue> &OutVals, 461 const SmallVectorImpl<ISD::InputArg> &Ins, 462 DebugLoc dl, SelectionDAG &DAG, 463 SmallVectorImpl<SDValue> &InVals) const; 464 465 /// HandleByVal - Target-specific cleanup for ByVal support. 466 virtual void HandleByVal(CCState *, unsigned &) const; 467 468 /// IsEligibleForTailCallOptimization - Check whether the call is eligible 469 /// for tail call optimization. Targets which want to do tail call 470 /// optimization should implement this function. 471 bool IsEligibleForTailCallOptimization(SDValue Callee, 472 CallingConv::ID CalleeCC, 473 bool isVarArg, 474 bool isCalleeStructRet, 475 bool isCallerStructRet, 476 const SmallVectorImpl<ISD::OutputArg> &Outs, 477 const SmallVectorImpl<SDValue> &OutVals, 478 const SmallVectorImpl<ISD::InputArg> &Ins, 479 SelectionDAG& DAG) const; 480 virtual SDValue 481 LowerReturn(SDValue Chain, 482 CallingConv::ID CallConv, bool isVarArg, 483 const SmallVectorImpl<ISD::OutputArg> &Outs, 484 const SmallVectorImpl<SDValue> &OutVals, 485 DebugLoc dl, SelectionDAG &DAG) const; 486 487 virtual bool isUsedByReturnOnly(SDNode *N) const; 488 489 virtual bool mayBeEmittedAsTailCall(CallInst *CI) const; 490 491 SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, 492 SDValue &ARMcc, SelectionDAG &DAG, DebugLoc dl) const; 493 SDValue getVFPCmp(SDValue LHS, SDValue RHS, 494 SelectionDAG &DAG, DebugLoc dl) const; 495 SDValue duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const; 496 497 SDValue OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const; 498 499 MachineBasicBlock *EmitAtomicCmpSwap(MachineInstr *MI, 500 MachineBasicBlock *BB, 501 unsigned Size) const; 502 MachineBasicBlock *EmitAtomicBinary(MachineInstr *MI, 503 MachineBasicBlock *BB, 504 unsigned Size, 505 unsigned BinOpcode) const; 506 MachineBasicBlock *EmitAtomicBinary64(MachineInstr *MI, 507 MachineBasicBlock *BB, 508 unsigned Op1, 509 unsigned Op2, 510 bool NeedsCarry = false, 511 bool IsCmpxchg = false) const; 512 MachineBasicBlock * EmitAtomicBinaryMinMax(MachineInstr *MI, 513 MachineBasicBlock *BB, 514 unsigned Size, 515 bool signExtend, 516 ARMCC::CondCodes Cond) const; 517 518 bool RemapAddSubWithFlags(MachineInstr *MI, MachineBasicBlock *BB) const; 519 }; 520 521 enum NEONModImmType { 522 VMOVModImm, 523 VMVNModImm, 524 OtherModImm 525 }; 526 527 528 namespace ARM { 529 FastISel *createFastISel(FunctionLoweringInfo &funcInfo); 530 } 531} 532 533#endif // ARMISELLOWERING_H 534