ARMISelDAGToDAG.cpp revision 5b81584f7403ffdb9cc6babaaeb0411c080e0f81
1//===-- ARMISelDAGToDAG.cpp - A dag to dag inst selector for ARM ----------===// 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 an instruction selector for the ARM target. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "arm-isel" 15#include "ARM.h" 16#include "ARMBaseInstrInfo.h" 17#include "ARMTargetMachine.h" 18#include "MCTargetDesc/ARMAddressingModes.h" 19#include "llvm/CallingConv.h" 20#include "llvm/Constants.h" 21#include "llvm/DerivedTypes.h" 22#include "llvm/Function.h" 23#include "llvm/Intrinsics.h" 24#include "llvm/LLVMContext.h" 25#include "llvm/CodeGen/MachineFrameInfo.h" 26#include "llvm/CodeGen/MachineFunction.h" 27#include "llvm/CodeGen/MachineInstrBuilder.h" 28#include "llvm/CodeGen/SelectionDAG.h" 29#include "llvm/CodeGen/SelectionDAGISel.h" 30#include "llvm/Target/TargetLowering.h" 31#include "llvm/Target/TargetOptions.h" 32#include "llvm/Support/CommandLine.h" 33#include "llvm/Support/Compiler.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/Support/ErrorHandling.h" 36#include "llvm/Support/raw_ostream.h" 37 38using namespace llvm; 39 40static cl::opt<bool> 41DisableShifterOp("disable-shifter-op", cl::Hidden, 42 cl::desc("Disable isel of shifter-op"), 43 cl::init(false)); 44 45static cl::opt<bool> 46CheckVMLxHazard("check-vmlx-hazard", cl::Hidden, 47 cl::desc("Check fp vmla / vmls hazard at isel time"), 48 cl::init(true)); 49 50//===--------------------------------------------------------------------===// 51/// ARMDAGToDAGISel - ARM specific code to select ARM machine 52/// instructions for SelectionDAG operations. 53/// 54namespace { 55 56enum AddrMode2Type { 57 AM2_BASE, // Simple AM2 (+-imm12) 58 AM2_SHOP // Shifter-op AM2 59}; 60 61class ARMDAGToDAGISel : public SelectionDAGISel { 62 ARMBaseTargetMachine &TM; 63 const ARMBaseInstrInfo *TII; 64 65 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can 66 /// make the right decision when generating code for different targets. 67 const ARMSubtarget *Subtarget; 68 69public: 70 explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm, 71 CodeGenOpt::Level OptLevel) 72 : SelectionDAGISel(tm, OptLevel), TM(tm), 73 TII(static_cast<const ARMBaseInstrInfo*>(TM.getInstrInfo())), 74 Subtarget(&TM.getSubtarget<ARMSubtarget>()) { 75 } 76 77 virtual const char *getPassName() const { 78 return "ARM Instruction Selection"; 79 } 80 81 /// getI32Imm - Return a target constant of type i32 with the specified 82 /// value. 83 inline SDValue getI32Imm(unsigned Imm) { 84 return CurDAG->getTargetConstant(Imm, MVT::i32); 85 } 86 87 SDNode *Select(SDNode *N); 88 89 90 bool hasNoVMLxHazardUse(SDNode *N) const; 91 bool isShifterOpProfitable(const SDValue &Shift, 92 ARM_AM::ShiftOpc ShOpcVal, unsigned ShAmt); 93 bool SelectRegShifterOperand(SDValue N, SDValue &A, 94 SDValue &B, SDValue &C, 95 bool CheckProfitability = true); 96 bool SelectImmShifterOperand(SDValue N, SDValue &A, 97 SDValue &B, bool CheckProfitability = true); 98 bool SelectShiftRegShifterOperand(SDValue N, SDValue &A, 99 SDValue &B, SDValue &C) { 100 // Don't apply the profitability check 101 return SelectRegShifterOperand(N, A, B, C, false); 102 } 103 bool SelectShiftImmShifterOperand(SDValue N, SDValue &A, 104 SDValue &B) { 105 // Don't apply the profitability check 106 return SelectImmShifterOperand(N, A, B, false); 107 } 108 109 bool SelectAddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm); 110 bool SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset, SDValue &Opc); 111 112 AddrMode2Type SelectAddrMode2Worker(SDValue N, SDValue &Base, 113 SDValue &Offset, SDValue &Opc); 114 bool SelectAddrMode2Base(SDValue N, SDValue &Base, SDValue &Offset, 115 SDValue &Opc) { 116 return SelectAddrMode2Worker(N, Base, Offset, Opc) == AM2_BASE; 117 } 118 119 bool SelectAddrMode2ShOp(SDValue N, SDValue &Base, SDValue &Offset, 120 SDValue &Opc) { 121 return SelectAddrMode2Worker(N, Base, Offset, Opc) == AM2_SHOP; 122 } 123 124 bool SelectAddrMode2(SDValue N, SDValue &Base, SDValue &Offset, 125 SDValue &Opc) { 126 SelectAddrMode2Worker(N, Base, Offset, Opc); 127// return SelectAddrMode2ShOp(N, Base, Offset, Opc); 128 // This always matches one way or another. 129 return true; 130 } 131 132 bool SelectAddrMode2OffsetReg(SDNode *Op, SDValue N, 133 SDValue &Offset, SDValue &Opc); 134 bool SelectAddrMode2OffsetImm(SDNode *Op, SDValue N, 135 SDValue &Offset, SDValue &Opc); 136 bool SelectAddrOffsetNone(SDValue N, SDValue &Base); 137 bool SelectAddrMode3(SDValue N, SDValue &Base, 138 SDValue &Offset, SDValue &Opc); 139 bool SelectAddrMode3Offset(SDNode *Op, SDValue N, 140 SDValue &Offset, SDValue &Opc); 141 bool SelectAddrMode5(SDValue N, SDValue &Base, 142 SDValue &Offset); 143 bool SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,SDValue &Align); 144 bool SelectAddrMode6Offset(SDNode *Op, SDValue N, SDValue &Offset); 145 146 bool SelectAddrModePC(SDValue N, SDValue &Offset, SDValue &Label); 147 148 // Thumb Addressing Modes: 149 bool SelectThumbAddrModeRR(SDValue N, SDValue &Base, SDValue &Offset); 150 bool SelectThumbAddrModeRI(SDValue N, SDValue &Base, SDValue &Offset, 151 unsigned Scale); 152 bool SelectThumbAddrModeRI5S1(SDValue N, SDValue &Base, SDValue &Offset); 153 bool SelectThumbAddrModeRI5S2(SDValue N, SDValue &Base, SDValue &Offset); 154 bool SelectThumbAddrModeRI5S4(SDValue N, SDValue &Base, SDValue &Offset); 155 bool SelectThumbAddrModeImm5S(SDValue N, unsigned Scale, SDValue &Base, 156 SDValue &OffImm); 157 bool SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base, 158 SDValue &OffImm); 159 bool SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base, 160 SDValue &OffImm); 161 bool SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base, 162 SDValue &OffImm); 163 bool SelectThumbAddrModeSP(SDValue N, SDValue &Base, SDValue &OffImm); 164 165 // Thumb 2 Addressing Modes: 166 bool SelectT2ShifterOperandReg(SDValue N, 167 SDValue &BaseReg, SDValue &Opc); 168 bool SelectT2AddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm); 169 bool SelectT2AddrModeImm8(SDValue N, SDValue &Base, 170 SDValue &OffImm); 171 bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N, 172 SDValue &OffImm); 173 bool SelectT2AddrModeSoReg(SDValue N, SDValue &Base, 174 SDValue &OffReg, SDValue &ShImm); 175 176 inline bool is_so_imm(unsigned Imm) const { 177 return ARM_AM::getSOImmVal(Imm) != -1; 178 } 179 180 inline bool is_so_imm_not(unsigned Imm) const { 181 return ARM_AM::getSOImmVal(~Imm) != -1; 182 } 183 184 inline bool is_t2_so_imm(unsigned Imm) const { 185 return ARM_AM::getT2SOImmVal(Imm) != -1; 186 } 187 188 inline bool is_t2_so_imm_not(unsigned Imm) const { 189 return ARM_AM::getT2SOImmVal(~Imm) != -1; 190 } 191 192 // Include the pieces autogenerated from the target description. 193#include "ARMGenDAGISel.inc" 194 195private: 196 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for 197 /// ARM. 198 SDNode *SelectARMIndexedLoad(SDNode *N); 199 SDNode *SelectT2IndexedLoad(SDNode *N); 200 201 /// SelectVLD - Select NEON load intrinsics. NumVecs should be 202 /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for 203 /// loads of D registers and even subregs and odd subregs of Q registers. 204 /// For NumVecs <= 2, QOpcodes1 is not used. 205 SDNode *SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs, 206 unsigned *DOpcodes, 207 unsigned *QOpcodes0, unsigned *QOpcodes1); 208 209 /// SelectVST - Select NEON store intrinsics. NumVecs should 210 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for 211 /// stores of D registers and even subregs and odd subregs of Q registers. 212 /// For NumVecs <= 2, QOpcodes1 is not used. 213 SDNode *SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs, 214 unsigned *DOpcodes, 215 unsigned *QOpcodes0, unsigned *QOpcodes1); 216 217 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should 218 /// be 2, 3 or 4. The opcode arrays specify the instructions used for 219 /// load/store of D registers and Q registers. 220 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, 221 bool isUpdating, unsigned NumVecs, 222 unsigned *DOpcodes, unsigned *QOpcodes); 223 224 /// SelectVLDDup - Select NEON load-duplicate intrinsics. NumVecs 225 /// should be 2, 3 or 4. The opcode array specifies the instructions used 226 /// for loading D registers. (Q registers are not supported.) 227 SDNode *SelectVLDDup(SDNode *N, bool isUpdating, unsigned NumVecs, 228 unsigned *Opcodes); 229 230 /// SelectVTBL - Select NEON VTBL and VTBX intrinsics. NumVecs should be 2, 231 /// 3 or 4. These are custom-selected so that a REG_SEQUENCE can be 232 /// generated to force the table registers to be consecutive. 233 SDNode *SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs, unsigned Opc); 234 235 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM. 236 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned); 237 238 /// SelectCMOVOp - Select CMOV instructions for ARM. 239 SDNode *SelectCMOVOp(SDNode *N); 240 SDNode *SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal, 241 ARMCC::CondCodes CCVal, SDValue CCR, 242 SDValue InFlag); 243 SDNode *SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal, 244 ARMCC::CondCodes CCVal, SDValue CCR, 245 SDValue InFlag); 246 SDNode *SelectT2CMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal, 247 ARMCC::CondCodes CCVal, SDValue CCR, 248 SDValue InFlag); 249 SDNode *SelectARMCMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal, 250 ARMCC::CondCodes CCVal, SDValue CCR, 251 SDValue InFlag); 252 253 SDNode *SelectConcatVector(SDNode *N); 254 255 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for 256 /// inline asm expressions. 257 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op, 258 char ConstraintCode, 259 std::vector<SDValue> &OutOps); 260 261 // Form pairs of consecutive S, D, or Q registers. 262 SDNode *PairSRegs(EVT VT, SDValue V0, SDValue V1); 263 SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1); 264 SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1); 265 266 // Form sequences of 4 consecutive S, D, or Q registers. 267 SDNode *QuadSRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3); 268 SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3); 269 SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3); 270 271 // Get the alignment operand for a NEON VLD or VST instruction. 272 SDValue GetVLDSTAlign(SDValue Align, unsigned NumVecs, bool is64BitVector); 273}; 274} 275 276/// isInt32Immediate - This method tests to see if the node is a 32-bit constant 277/// operand. If so Imm will receive the 32-bit value. 278static bool isInt32Immediate(SDNode *N, unsigned &Imm) { 279 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) { 280 Imm = cast<ConstantSDNode>(N)->getZExtValue(); 281 return true; 282 } 283 return false; 284} 285 286// isInt32Immediate - This method tests to see if a constant operand. 287// If so Imm will receive the 32 bit value. 288static bool isInt32Immediate(SDValue N, unsigned &Imm) { 289 return isInt32Immediate(N.getNode(), Imm); 290} 291 292// isOpcWithIntImmediate - This method tests to see if the node is a specific 293// opcode and that it has a immediate integer right operand. 294// If so Imm will receive the 32 bit value. 295static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) { 296 return N->getOpcode() == Opc && 297 isInt32Immediate(N->getOperand(1).getNode(), Imm); 298} 299 300/// \brief Check whether a particular node is a constant value representable as 301/// (N * Scale) where (N in [\arg RangeMin, \arg RangeMax). 302/// 303/// \param ScaledConstant [out] - On success, the pre-scaled constant value. 304static bool isScaledConstantInRange(SDValue Node, unsigned Scale, 305 int RangeMin, int RangeMax, 306 int &ScaledConstant) { 307 assert(Scale && "Invalid scale!"); 308 309 // Check that this is a constant. 310 const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Node); 311 if (!C) 312 return false; 313 314 ScaledConstant = (int) C->getZExtValue(); 315 if ((ScaledConstant % Scale) != 0) 316 return false; 317 318 ScaledConstant /= Scale; 319 return ScaledConstant >= RangeMin && ScaledConstant < RangeMax; 320} 321 322/// hasNoVMLxHazardUse - Return true if it's desirable to select a FP MLA / MLS 323/// node. VFP / NEON fp VMLA / VMLS instructions have special RAW hazards (at 324/// least on current ARM implementations) which should be avoidded. 325bool ARMDAGToDAGISel::hasNoVMLxHazardUse(SDNode *N) const { 326 if (OptLevel == CodeGenOpt::None) 327 return true; 328 329 if (!CheckVMLxHazard) 330 return true; 331 332 if (!Subtarget->isCortexA8() && !Subtarget->isCortexA9()) 333 return true; 334 335 if (!N->hasOneUse()) 336 return false; 337 338 SDNode *Use = *N->use_begin(); 339 if (Use->getOpcode() == ISD::CopyToReg) 340 return true; 341 if (Use->isMachineOpcode()) { 342 const MCInstrDesc &MCID = TII->get(Use->getMachineOpcode()); 343 if (MCID.mayStore()) 344 return true; 345 unsigned Opcode = MCID.getOpcode(); 346 if (Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD) 347 return true; 348 // vmlx feeding into another vmlx. We actually want to unfold 349 // the use later in the MLxExpansion pass. e.g. 350 // vmla 351 // vmla (stall 8 cycles) 352 // 353 // vmul (5 cycles) 354 // vadd (5 cycles) 355 // vmla 356 // This adds up to about 18 - 19 cycles. 357 // 358 // vmla 359 // vmul (stall 4 cycles) 360 // vadd adds up to about 14 cycles. 361 return TII->isFpMLxInstruction(Opcode); 362 } 363 364 return false; 365} 366 367bool ARMDAGToDAGISel::isShifterOpProfitable(const SDValue &Shift, 368 ARM_AM::ShiftOpc ShOpcVal, 369 unsigned ShAmt) { 370 if (!Subtarget->isCortexA9()) 371 return true; 372 if (Shift.hasOneUse()) 373 return true; 374 // R << 2 is free. 375 return ShOpcVal == ARM_AM::lsl && ShAmt == 2; 376} 377 378bool ARMDAGToDAGISel::SelectImmShifterOperand(SDValue N, 379 SDValue &BaseReg, 380 SDValue &Opc, 381 bool CheckProfitability) { 382 if (DisableShifterOp) 383 return false; 384 385 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode()); 386 387 // Don't match base register only case. That is matched to a separate 388 // lower complexity pattern with explicit register operand. 389 if (ShOpcVal == ARM_AM::no_shift) return false; 390 391 BaseReg = N.getOperand(0); 392 unsigned ShImmVal = 0; 393 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1)); 394 if (!RHS) return false; 395 ShImmVal = RHS->getZExtValue() & 31; 396 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal), 397 MVT::i32); 398 return true; 399} 400 401bool ARMDAGToDAGISel::SelectRegShifterOperand(SDValue N, 402 SDValue &BaseReg, 403 SDValue &ShReg, 404 SDValue &Opc, 405 bool CheckProfitability) { 406 if (DisableShifterOp) 407 return false; 408 409 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode()); 410 411 // Don't match base register only case. That is matched to a separate 412 // lower complexity pattern with explicit register operand. 413 if (ShOpcVal == ARM_AM::no_shift) return false; 414 415 BaseReg = N.getOperand(0); 416 unsigned ShImmVal = 0; 417 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1)); 418 if (RHS) return false; 419 420 ShReg = N.getOperand(1); 421 if (CheckProfitability && !isShifterOpProfitable(N, ShOpcVal, ShImmVal)) 422 return false; 423 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal), 424 MVT::i32); 425 return true; 426} 427 428 429bool ARMDAGToDAGISel::SelectAddrModeImm12(SDValue N, 430 SDValue &Base, 431 SDValue &OffImm) { 432 // Match simple R + imm12 operands. 433 434 // Base only. 435 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && 436 !CurDAG->isBaseWithConstantOffset(N)) { 437 if (N.getOpcode() == ISD::FrameIndex) { 438 // Match frame index. 439 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 440 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 441 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 442 return true; 443 } 444 445 if (N.getOpcode() == ARMISD::Wrapper && 446 !(Subtarget->useMovt() && 447 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) { 448 Base = N.getOperand(0); 449 } else 450 Base = N; 451 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 452 return true; 453 } 454 455 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 456 int RHSC = (int)RHS->getZExtValue(); 457 if (N.getOpcode() == ISD::SUB) 458 RHSC = -RHSC; 459 460 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned) 461 Base = N.getOperand(0); 462 if (Base.getOpcode() == ISD::FrameIndex) { 463 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 464 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 465 } 466 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); 467 return true; 468 } 469 } 470 471 // Base only. 472 Base = N; 473 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 474 return true; 475} 476 477 478 479bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset, 480 SDValue &Opc) { 481 if (N.getOpcode() == ISD::MUL && 482 (!Subtarget->isCortexA9() || N.hasOneUse())) { 483 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 484 // X * [3,5,9] -> X + X * [2,4,8] etc. 485 int RHSC = (int)RHS->getZExtValue(); 486 if (RHSC & 1) { 487 RHSC = RHSC & ~1; 488 ARM_AM::AddrOpc AddSub = ARM_AM::add; 489 if (RHSC < 0) { 490 AddSub = ARM_AM::sub; 491 RHSC = - RHSC; 492 } 493 if (isPowerOf2_32(RHSC)) { 494 unsigned ShAmt = Log2_32(RHSC); 495 Base = Offset = N.getOperand(0); 496 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, 497 ARM_AM::lsl), 498 MVT::i32); 499 return true; 500 } 501 } 502 } 503 } 504 505 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && 506 // ISD::OR that is equivalent to an ISD::ADD. 507 !CurDAG->isBaseWithConstantOffset(N)) 508 return false; 509 510 // Leave simple R +/- imm12 operands for LDRi12 511 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::OR) { 512 int RHSC; 513 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1, 514 -0x1000+1, 0x1000, RHSC)) // 12 bits. 515 return false; 516 } 517 518 if (Subtarget->isCortexA9() && !N.hasOneUse()) 519 // Compute R +/- (R << N) and reuse it. 520 return false; 521 522 // Otherwise this is R +/- [possibly shifted] R. 523 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::SUB ? ARM_AM::sub:ARM_AM::add; 524 ARM_AM::ShiftOpc ShOpcVal = 525 ARM_AM::getShiftOpcForNode(N.getOperand(1).getOpcode()); 526 unsigned ShAmt = 0; 527 528 Base = N.getOperand(0); 529 Offset = N.getOperand(1); 530 531 if (ShOpcVal != ARM_AM::no_shift) { 532 // Check to see if the RHS of the shift is a constant, if not, we can't fold 533 // it. 534 if (ConstantSDNode *Sh = 535 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) { 536 ShAmt = Sh->getZExtValue(); 537 if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt)) 538 Offset = N.getOperand(1).getOperand(0); 539 else { 540 ShAmt = 0; 541 ShOpcVal = ARM_AM::no_shift; 542 } 543 } else { 544 ShOpcVal = ARM_AM::no_shift; 545 } 546 } 547 548 // Try matching (R shl C) + (R). 549 if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift && 550 !(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) { 551 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0).getOpcode()); 552 if (ShOpcVal != ARM_AM::no_shift) { 553 // Check to see if the RHS of the shift is a constant, if not, we can't 554 // fold it. 555 if (ConstantSDNode *Sh = 556 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) { 557 ShAmt = Sh->getZExtValue(); 558 if (!Subtarget->isCortexA9() || 559 (N.hasOneUse() && 560 isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt))) { 561 Offset = N.getOperand(0).getOperand(0); 562 Base = N.getOperand(1); 563 } else { 564 ShAmt = 0; 565 ShOpcVal = ARM_AM::no_shift; 566 } 567 } else { 568 ShOpcVal = ARM_AM::no_shift; 569 } 570 } 571 } 572 573 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal), 574 MVT::i32); 575 return true; 576} 577 578 579 580 581//----- 582 583AddrMode2Type ARMDAGToDAGISel::SelectAddrMode2Worker(SDValue N, 584 SDValue &Base, 585 SDValue &Offset, 586 SDValue &Opc) { 587 if (N.getOpcode() == ISD::MUL && 588 (!Subtarget->isCortexA9() || N.hasOneUse())) { 589 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 590 // X * [3,5,9] -> X + X * [2,4,8] etc. 591 int RHSC = (int)RHS->getZExtValue(); 592 if (RHSC & 1) { 593 RHSC = RHSC & ~1; 594 ARM_AM::AddrOpc AddSub = ARM_AM::add; 595 if (RHSC < 0) { 596 AddSub = ARM_AM::sub; 597 RHSC = - RHSC; 598 } 599 if (isPowerOf2_32(RHSC)) { 600 unsigned ShAmt = Log2_32(RHSC); 601 Base = Offset = N.getOperand(0); 602 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, 603 ARM_AM::lsl), 604 MVT::i32); 605 return AM2_SHOP; 606 } 607 } 608 } 609 } 610 611 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && 612 // ISD::OR that is equivalent to an ADD. 613 !CurDAG->isBaseWithConstantOffset(N)) { 614 Base = N; 615 if (N.getOpcode() == ISD::FrameIndex) { 616 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 617 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 618 } else if (N.getOpcode() == ARMISD::Wrapper && 619 !(Subtarget->useMovt() && 620 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) { 621 Base = N.getOperand(0); 622 } 623 Offset = CurDAG->getRegister(0, MVT::i32); 624 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0, 625 ARM_AM::no_shift), 626 MVT::i32); 627 return AM2_BASE; 628 } 629 630 // Match simple R +/- imm12 operands. 631 if (N.getOpcode() != ISD::SUB) { 632 int RHSC; 633 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1, 634 -0x1000+1, 0x1000, RHSC)) { // 12 bits. 635 Base = N.getOperand(0); 636 if (Base.getOpcode() == ISD::FrameIndex) { 637 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 638 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 639 } 640 Offset = CurDAG->getRegister(0, MVT::i32); 641 642 ARM_AM::AddrOpc AddSub = ARM_AM::add; 643 if (RHSC < 0) { 644 AddSub = ARM_AM::sub; 645 RHSC = - RHSC; 646 } 647 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC, 648 ARM_AM::no_shift), 649 MVT::i32); 650 return AM2_BASE; 651 } 652 } 653 654 if (Subtarget->isCortexA9() && !N.hasOneUse()) { 655 // Compute R +/- (R << N) and reuse it. 656 Base = N; 657 Offset = CurDAG->getRegister(0, MVT::i32); 658 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0, 659 ARM_AM::no_shift), 660 MVT::i32); 661 return AM2_BASE; 662 } 663 664 // Otherwise this is R +/- [possibly shifted] R. 665 ARM_AM::AddrOpc AddSub = N.getOpcode() != ISD::SUB ? ARM_AM::add:ARM_AM::sub; 666 ARM_AM::ShiftOpc ShOpcVal = 667 ARM_AM::getShiftOpcForNode(N.getOperand(1).getOpcode()); 668 unsigned ShAmt = 0; 669 670 Base = N.getOperand(0); 671 Offset = N.getOperand(1); 672 673 if (ShOpcVal != ARM_AM::no_shift) { 674 // Check to see if the RHS of the shift is a constant, if not, we can't fold 675 // it. 676 if (ConstantSDNode *Sh = 677 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) { 678 ShAmt = Sh->getZExtValue(); 679 if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt)) 680 Offset = N.getOperand(1).getOperand(0); 681 else { 682 ShAmt = 0; 683 ShOpcVal = ARM_AM::no_shift; 684 } 685 } else { 686 ShOpcVal = ARM_AM::no_shift; 687 } 688 } 689 690 // Try matching (R shl C) + (R). 691 if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift && 692 !(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) { 693 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0).getOpcode()); 694 if (ShOpcVal != ARM_AM::no_shift) { 695 // Check to see if the RHS of the shift is a constant, if not, we can't 696 // fold it. 697 if (ConstantSDNode *Sh = 698 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) { 699 ShAmt = Sh->getZExtValue(); 700 if (!Subtarget->isCortexA9() || 701 (N.hasOneUse() && 702 isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt))) { 703 Offset = N.getOperand(0).getOperand(0); 704 Base = N.getOperand(1); 705 } else { 706 ShAmt = 0; 707 ShOpcVal = ARM_AM::no_shift; 708 } 709 } else { 710 ShOpcVal = ARM_AM::no_shift; 711 } 712 } 713 } 714 715 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal), 716 MVT::i32); 717 return AM2_SHOP; 718} 719 720bool ARMDAGToDAGISel::SelectAddrMode2OffsetReg(SDNode *Op, SDValue N, 721 SDValue &Offset, SDValue &Opc) { 722 unsigned Opcode = Op->getOpcode(); 723 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) 724 ? cast<LoadSDNode>(Op)->getAddressingMode() 725 : cast<StoreSDNode>(Op)->getAddressingMode(); 726 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC) 727 ? ARM_AM::add : ARM_AM::sub; 728 int Val; 729 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val)) 730 return false; 731 732 Offset = N; 733 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode()); 734 unsigned ShAmt = 0; 735 if (ShOpcVal != ARM_AM::no_shift) { 736 // Check to see if the RHS of the shift is a constant, if not, we can't fold 737 // it. 738 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 739 ShAmt = Sh->getZExtValue(); 740 if (isShifterOpProfitable(N, ShOpcVal, ShAmt)) 741 Offset = N.getOperand(0); 742 else { 743 ShAmt = 0; 744 ShOpcVal = ARM_AM::no_shift; 745 } 746 } else { 747 ShOpcVal = ARM_AM::no_shift; 748 } 749 } 750 751 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal), 752 MVT::i32); 753 return true; 754} 755 756bool ARMDAGToDAGISel::SelectAddrMode2OffsetImm(SDNode *Op, SDValue N, 757 SDValue &Offset, SDValue &Opc) { 758 unsigned Opcode = Op->getOpcode(); 759 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) 760 ? cast<LoadSDNode>(Op)->getAddressingMode() 761 : cast<StoreSDNode>(Op)->getAddressingMode(); 762 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC) 763 ? ARM_AM::add : ARM_AM::sub; 764 int Val; 765 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val)) { // 12 bits. 766 Offset = CurDAG->getRegister(0, MVT::i32); 767 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val, 768 ARM_AM::no_shift), 769 MVT::i32); 770 return true; 771 } 772 773 return false; 774} 775 776bool ARMDAGToDAGISel::SelectAddrOffsetNone(SDValue N, SDValue &Base) { 777 Base = N; 778 return true; 779} 780 781bool ARMDAGToDAGISel::SelectAddrMode3(SDValue N, 782 SDValue &Base, SDValue &Offset, 783 SDValue &Opc) { 784 if (N.getOpcode() == ISD::SUB) { 785 // X - C is canonicalize to X + -C, no need to handle it here. 786 Base = N.getOperand(0); 787 Offset = N.getOperand(1); 788 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32); 789 return true; 790 } 791 792 if (!CurDAG->isBaseWithConstantOffset(N)) { 793 Base = N; 794 if (N.getOpcode() == ISD::FrameIndex) { 795 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 796 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 797 } 798 Offset = CurDAG->getRegister(0, MVT::i32); 799 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32); 800 return true; 801 } 802 803 // If the RHS is +/- imm8, fold into addr mode. 804 int RHSC; 805 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1, 806 -256 + 1, 256, RHSC)) { // 8 bits. 807 Base = N.getOperand(0); 808 if (Base.getOpcode() == ISD::FrameIndex) { 809 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 810 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 811 } 812 Offset = CurDAG->getRegister(0, MVT::i32); 813 814 ARM_AM::AddrOpc AddSub = ARM_AM::add; 815 if (RHSC < 0) { 816 AddSub = ARM_AM::sub; 817 RHSC = -RHSC; 818 } 819 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32); 820 return true; 821 } 822 823 Base = N.getOperand(0); 824 Offset = N.getOperand(1); 825 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32); 826 return true; 827} 828 829bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N, 830 SDValue &Offset, SDValue &Opc) { 831 unsigned Opcode = Op->getOpcode(); 832 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) 833 ? cast<LoadSDNode>(Op)->getAddressingMode() 834 : cast<StoreSDNode>(Op)->getAddressingMode(); 835 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC) 836 ? ARM_AM::add : ARM_AM::sub; 837 int Val; 838 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 256, Val)) { // 12 bits. 839 Offset = CurDAG->getRegister(0, MVT::i32); 840 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32); 841 return true; 842 } 843 844 Offset = N; 845 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32); 846 return true; 847} 848 849bool ARMDAGToDAGISel::SelectAddrMode5(SDValue N, 850 SDValue &Base, SDValue &Offset) { 851 if (!CurDAG->isBaseWithConstantOffset(N)) { 852 Base = N; 853 if (N.getOpcode() == ISD::FrameIndex) { 854 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 855 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 856 } else if (N.getOpcode() == ARMISD::Wrapper && 857 !(Subtarget->useMovt() && 858 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) { 859 Base = N.getOperand(0); 860 } 861 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0), 862 MVT::i32); 863 return true; 864 } 865 866 // If the RHS is +/- imm8, fold into addr mode. 867 int RHSC; 868 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/4, 869 -256 + 1, 256, RHSC)) { 870 Base = N.getOperand(0); 871 if (Base.getOpcode() == ISD::FrameIndex) { 872 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 873 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 874 } 875 876 ARM_AM::AddrOpc AddSub = ARM_AM::add; 877 if (RHSC < 0) { 878 AddSub = ARM_AM::sub; 879 RHSC = -RHSC; 880 } 881 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC), 882 MVT::i32); 883 return true; 884 } 885 886 Base = N; 887 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0), 888 MVT::i32); 889 return true; 890} 891 892bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr, 893 SDValue &Align) { 894 Addr = N; 895 896 unsigned Alignment = 0; 897 if (LSBaseSDNode *LSN = dyn_cast<LSBaseSDNode>(Parent)) { 898 // This case occurs only for VLD1-lane/dup and VST1-lane instructions. 899 // The maximum alignment is equal to the memory size being referenced. 900 unsigned LSNAlign = LSN->getAlignment(); 901 unsigned MemSize = LSN->getMemoryVT().getSizeInBits() / 8; 902 if (LSNAlign > MemSize && MemSize > 1) 903 Alignment = MemSize; 904 } else { 905 // All other uses of addrmode6 are for intrinsics. For now just record 906 // the raw alignment value; it will be refined later based on the legal 907 // alignment operands for the intrinsic. 908 Alignment = cast<MemIntrinsicSDNode>(Parent)->getAlignment(); 909 } 910 911 Align = CurDAG->getTargetConstant(Alignment, MVT::i32); 912 return true; 913} 914 915bool ARMDAGToDAGISel::SelectAddrMode6Offset(SDNode *Op, SDValue N, 916 SDValue &Offset) { 917 LSBaseSDNode *LdSt = cast<LSBaseSDNode>(Op); 918 ISD::MemIndexedMode AM = LdSt->getAddressingMode(); 919 if (AM != ISD::POST_INC) 920 return false; 921 Offset = N; 922 if (ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N)) { 923 if (NC->getZExtValue() * 8 == LdSt->getMemoryVT().getSizeInBits()) 924 Offset = CurDAG->getRegister(0, MVT::i32); 925 } 926 return true; 927} 928 929bool ARMDAGToDAGISel::SelectAddrModePC(SDValue N, 930 SDValue &Offset, SDValue &Label) { 931 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) { 932 Offset = N.getOperand(0); 933 SDValue N1 = N.getOperand(1); 934 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(), 935 MVT::i32); 936 return true; 937 } 938 939 return false; 940} 941 942 943//===----------------------------------------------------------------------===// 944// Thumb Addressing Modes 945//===----------------------------------------------------------------------===// 946 947bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue N, 948 SDValue &Base, SDValue &Offset){ 949 if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(N)) { 950 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N); 951 if (!NC || !NC->isNullValue()) 952 return false; 953 954 Base = Offset = N; 955 return true; 956 } 957 958 Base = N.getOperand(0); 959 Offset = N.getOperand(1); 960 return true; 961} 962 963bool 964ARMDAGToDAGISel::SelectThumbAddrModeRI(SDValue N, SDValue &Base, 965 SDValue &Offset, unsigned Scale) { 966 if (Scale == 4) { 967 SDValue TmpBase, TmpOffImm; 968 if (SelectThumbAddrModeSP(N, TmpBase, TmpOffImm)) 969 return false; // We want to select tLDRspi / tSTRspi instead. 970 971 if (N.getOpcode() == ARMISD::Wrapper && 972 N.getOperand(0).getOpcode() == ISD::TargetConstantPool) 973 return false; // We want to select tLDRpci instead. 974 } 975 976 if (!CurDAG->isBaseWithConstantOffset(N)) 977 return false; 978 979 // Thumb does not have [sp, r] address mode. 980 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0)); 981 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1)); 982 if ((LHSR && LHSR->getReg() == ARM::SP) || 983 (RHSR && RHSR->getReg() == ARM::SP)) 984 return false; 985 986 // FIXME: Why do we explicitly check for a match here and then return false? 987 // Presumably to allow something else to match, but shouldn't this be 988 // documented? 989 int RHSC; 990 if (isScaledConstantInRange(N.getOperand(1), Scale, 0, 32, RHSC)) 991 return false; 992 993 Base = N.getOperand(0); 994 Offset = N.getOperand(1); 995 return true; 996} 997 998bool 999ARMDAGToDAGISel::SelectThumbAddrModeRI5S1(SDValue N, 1000 SDValue &Base, 1001 SDValue &Offset) { 1002 return SelectThumbAddrModeRI(N, Base, Offset, 1); 1003} 1004 1005bool 1006ARMDAGToDAGISel::SelectThumbAddrModeRI5S2(SDValue N, 1007 SDValue &Base, 1008 SDValue &Offset) { 1009 return SelectThumbAddrModeRI(N, Base, Offset, 2); 1010} 1011 1012bool 1013ARMDAGToDAGISel::SelectThumbAddrModeRI5S4(SDValue N, 1014 SDValue &Base, 1015 SDValue &Offset) { 1016 return SelectThumbAddrModeRI(N, Base, Offset, 4); 1017} 1018 1019bool 1020ARMDAGToDAGISel::SelectThumbAddrModeImm5S(SDValue N, unsigned Scale, 1021 SDValue &Base, SDValue &OffImm) { 1022 if (Scale == 4) { 1023 SDValue TmpBase, TmpOffImm; 1024 if (SelectThumbAddrModeSP(N, TmpBase, TmpOffImm)) 1025 return false; // We want to select tLDRspi / tSTRspi instead. 1026 1027 if (N.getOpcode() == ARMISD::Wrapper && 1028 N.getOperand(0).getOpcode() == ISD::TargetConstantPool) 1029 return false; // We want to select tLDRpci instead. 1030 } 1031 1032 if (!CurDAG->isBaseWithConstantOffset(N)) { 1033 if (N.getOpcode() == ARMISD::Wrapper && 1034 !(Subtarget->useMovt() && 1035 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) { 1036 Base = N.getOperand(0); 1037 } else { 1038 Base = N; 1039 } 1040 1041 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1042 return true; 1043 } 1044 1045 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0)); 1046 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1)); 1047 if ((LHSR && LHSR->getReg() == ARM::SP) || 1048 (RHSR && RHSR->getReg() == ARM::SP)) { 1049 ConstantSDNode *LHS = dyn_cast<ConstantSDNode>(N.getOperand(0)); 1050 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1)); 1051 unsigned LHSC = LHS ? LHS->getZExtValue() : 0; 1052 unsigned RHSC = RHS ? RHS->getZExtValue() : 0; 1053 1054 // Thumb does not have [sp, #imm5] address mode for non-zero imm5. 1055 if (LHSC != 0 || RHSC != 0) return false; 1056 1057 Base = N; 1058 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1059 return true; 1060 } 1061 1062 // If the RHS is + imm5 * scale, fold into addr mode. 1063 int RHSC; 1064 if (isScaledConstantInRange(N.getOperand(1), Scale, 0, 32, RHSC)) { 1065 Base = N.getOperand(0); 1066 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); 1067 return true; 1068 } 1069 1070 Base = N.getOperand(0); 1071 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1072 return true; 1073} 1074 1075bool 1076ARMDAGToDAGISel::SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base, 1077 SDValue &OffImm) { 1078 return SelectThumbAddrModeImm5S(N, 4, Base, OffImm); 1079} 1080 1081bool 1082ARMDAGToDAGISel::SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base, 1083 SDValue &OffImm) { 1084 return SelectThumbAddrModeImm5S(N, 2, Base, OffImm); 1085} 1086 1087bool 1088ARMDAGToDAGISel::SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base, 1089 SDValue &OffImm) { 1090 return SelectThumbAddrModeImm5S(N, 1, Base, OffImm); 1091} 1092 1093bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDValue N, 1094 SDValue &Base, SDValue &OffImm) { 1095 if (N.getOpcode() == ISD::FrameIndex) { 1096 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 1097 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 1098 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1099 return true; 1100 } 1101 1102 if (!CurDAG->isBaseWithConstantOffset(N)) 1103 return false; 1104 1105 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0)); 1106 if (N.getOperand(0).getOpcode() == ISD::FrameIndex || 1107 (LHSR && LHSR->getReg() == ARM::SP)) { 1108 // If the RHS is + imm8 * scale, fold into addr mode. 1109 int RHSC; 1110 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/4, 0, 256, RHSC)) { 1111 Base = N.getOperand(0); 1112 if (Base.getOpcode() == ISD::FrameIndex) { 1113 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 1114 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 1115 } 1116 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); 1117 return true; 1118 } 1119 } 1120 1121 return false; 1122} 1123 1124 1125//===----------------------------------------------------------------------===// 1126// Thumb 2 Addressing Modes 1127//===----------------------------------------------------------------------===// 1128 1129 1130bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDValue N, SDValue &BaseReg, 1131 SDValue &Opc) { 1132 if (DisableShifterOp) 1133 return false; 1134 1135 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode()); 1136 1137 // Don't match base register only case. That is matched to a separate 1138 // lower complexity pattern with explicit register operand. 1139 if (ShOpcVal == ARM_AM::no_shift) return false; 1140 1141 BaseReg = N.getOperand(0); 1142 unsigned ShImmVal = 0; 1143 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 1144 ShImmVal = RHS->getZExtValue() & 31; 1145 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal)); 1146 return true; 1147 } 1148 1149 return false; 1150} 1151 1152bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue N, 1153 SDValue &Base, SDValue &OffImm) { 1154 // Match simple R + imm12 operands. 1155 1156 // Base only. 1157 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && 1158 !CurDAG->isBaseWithConstantOffset(N)) { 1159 if (N.getOpcode() == ISD::FrameIndex) { 1160 // Match frame index. 1161 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 1162 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 1163 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1164 return true; 1165 } 1166 1167 if (N.getOpcode() == ARMISD::Wrapper && 1168 !(Subtarget->useMovt() && 1169 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) { 1170 Base = N.getOperand(0); 1171 if (Base.getOpcode() == ISD::TargetConstantPool) 1172 return false; // We want to select t2LDRpci instead. 1173 } else 1174 Base = N; 1175 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1176 return true; 1177 } 1178 1179 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 1180 if (SelectT2AddrModeImm8(N, Base, OffImm)) 1181 // Let t2LDRi8 handle (R - imm8). 1182 return false; 1183 1184 int RHSC = (int)RHS->getZExtValue(); 1185 if (N.getOpcode() == ISD::SUB) 1186 RHSC = -RHSC; 1187 1188 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned) 1189 Base = N.getOperand(0); 1190 if (Base.getOpcode() == ISD::FrameIndex) { 1191 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 1192 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 1193 } 1194 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); 1195 return true; 1196 } 1197 } 1198 1199 // Base only. 1200 Base = N; 1201 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1202 return true; 1203} 1204 1205bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue N, 1206 SDValue &Base, SDValue &OffImm) { 1207 // Match simple R - imm8 operands. 1208 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && 1209 !CurDAG->isBaseWithConstantOffset(N)) 1210 return false; 1211 1212 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 1213 int RHSC = (int)RHS->getSExtValue(); 1214 if (N.getOpcode() == ISD::SUB) 1215 RHSC = -RHSC; 1216 1217 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative) 1218 Base = N.getOperand(0); 1219 if (Base.getOpcode() == ISD::FrameIndex) { 1220 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 1221 Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 1222 } 1223 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); 1224 return true; 1225 } 1226 } 1227 1228 return false; 1229} 1230 1231bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N, 1232 SDValue &OffImm){ 1233 unsigned Opcode = Op->getOpcode(); 1234 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) 1235 ? cast<LoadSDNode>(Op)->getAddressingMode() 1236 : cast<StoreSDNode>(Op)->getAddressingMode(); 1237 int RHSC; 1238 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x100, RHSC)) { // 8 bits. 1239 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC)) 1240 ? CurDAG->getTargetConstant(RHSC, MVT::i32) 1241 : CurDAG->getTargetConstant(-RHSC, MVT::i32); 1242 return true; 1243 } 1244 1245 return false; 1246} 1247 1248bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue N, 1249 SDValue &Base, 1250 SDValue &OffReg, SDValue &ShImm) { 1251 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12. 1252 if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(N)) 1253 return false; 1254 1255 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8. 1256 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 1257 int RHSC = (int)RHS->getZExtValue(); 1258 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned) 1259 return false; 1260 else if (RHSC < 0 && RHSC >= -255) // 8 bits 1261 return false; 1262 } 1263 1264 if (Subtarget->isCortexA9() && !N.hasOneUse()) { 1265 // Compute R + (R << [1,2,3]) and reuse it. 1266 Base = N; 1267 return false; 1268 } 1269 1270 // Look for (R + R) or (R + (R << [1,2,3])). 1271 unsigned ShAmt = 0; 1272 Base = N.getOperand(0); 1273 OffReg = N.getOperand(1); 1274 1275 // Swap if it is ((R << c) + R). 1276 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg.getOpcode()); 1277 if (ShOpcVal != ARM_AM::lsl) { 1278 ShOpcVal = ARM_AM::getShiftOpcForNode(Base.getOpcode()); 1279 if (ShOpcVal == ARM_AM::lsl) 1280 std::swap(Base, OffReg); 1281 } 1282 1283 if (ShOpcVal == ARM_AM::lsl) { 1284 // Check to see if the RHS of the shift is a constant, if not, we can't fold 1285 // it. 1286 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) { 1287 ShAmt = Sh->getZExtValue(); 1288 if (ShAmt < 4 && isShifterOpProfitable(OffReg, ShOpcVal, ShAmt)) 1289 OffReg = OffReg.getOperand(0); 1290 else { 1291 ShAmt = 0; 1292 ShOpcVal = ARM_AM::no_shift; 1293 } 1294 } else { 1295 ShOpcVal = ARM_AM::no_shift; 1296 } 1297 } 1298 1299 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32); 1300 1301 return true; 1302} 1303 1304//===--------------------------------------------------------------------===// 1305 1306/// getAL - Returns a ARMCC::AL immediate node. 1307static inline SDValue getAL(SelectionDAG *CurDAG) { 1308 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32); 1309} 1310 1311SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) { 1312 LoadSDNode *LD = cast<LoadSDNode>(N); 1313 ISD::MemIndexedMode AM = LD->getAddressingMode(); 1314 if (AM == ISD::UNINDEXED) 1315 return NULL; 1316 1317 EVT LoadedVT = LD->getMemoryVT(); 1318 SDValue Offset, AMOpc; 1319 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC); 1320 unsigned Opcode = 0; 1321 bool Match = false; 1322 if (LoadedVT == MVT::i32 && 1323 SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) { 1324 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST_IMM; 1325 Match = true; 1326 } else if (LoadedVT == MVT::i32 && 1327 SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) { 1328 Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST_REG; 1329 Match = true; 1330 1331 } else if (LoadedVT == MVT::i16 && 1332 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) { 1333 Match = true; 1334 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD) 1335 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST) 1336 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST); 1337 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) { 1338 if (LD->getExtensionType() == ISD::SEXTLOAD) { 1339 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) { 1340 Match = true; 1341 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST; 1342 } 1343 } else { 1344 if (SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) { 1345 Match = true; 1346 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST_IMM; 1347 } else if (SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) { 1348 Match = true; 1349 Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST_REG; 1350 } 1351 } 1352 } 1353 1354 if (Match) { 1355 SDValue Chain = LD->getChain(); 1356 SDValue Base = LD->getBasePtr(); 1357 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG), 1358 CurDAG->getRegister(0, MVT::i32), Chain }; 1359 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32, 1360 MVT::Other, Ops, 6); 1361 } 1362 1363 return NULL; 1364} 1365 1366SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) { 1367 LoadSDNode *LD = cast<LoadSDNode>(N); 1368 ISD::MemIndexedMode AM = LD->getAddressingMode(); 1369 if (AM == ISD::UNINDEXED) 1370 return NULL; 1371 1372 EVT LoadedVT = LD->getMemoryVT(); 1373 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD; 1374 SDValue Offset; 1375 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC); 1376 unsigned Opcode = 0; 1377 bool Match = false; 1378 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) { 1379 switch (LoadedVT.getSimpleVT().SimpleTy) { 1380 case MVT::i32: 1381 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST; 1382 break; 1383 case MVT::i16: 1384 if (isSExtLd) 1385 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST; 1386 else 1387 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST; 1388 break; 1389 case MVT::i8: 1390 case MVT::i1: 1391 if (isSExtLd) 1392 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST; 1393 else 1394 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST; 1395 break; 1396 default: 1397 return NULL; 1398 } 1399 Match = true; 1400 } 1401 1402 if (Match) { 1403 SDValue Chain = LD->getChain(); 1404 SDValue Base = LD->getBasePtr(); 1405 SDValue Ops[]= { Base, Offset, getAL(CurDAG), 1406 CurDAG->getRegister(0, MVT::i32), Chain }; 1407 return CurDAG->getMachineNode(Opcode, N->getDebugLoc(), MVT::i32, MVT::i32, 1408 MVT::Other, Ops, 5); 1409 } 1410 1411 return NULL; 1412} 1413 1414/// PairSRegs - Form a D register from a pair of S registers. 1415/// 1416SDNode *ARMDAGToDAGISel::PairSRegs(EVT VT, SDValue V0, SDValue V1) { 1417 DebugLoc dl = V0.getNode()->getDebugLoc(); 1418 SDValue RegClass = 1419 CurDAG->getTargetConstant(ARM::DPR_VFP2RegClassID, MVT::i32); 1420 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32); 1421 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32); 1422 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 }; 1423 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 5); 1424} 1425 1426/// PairDRegs - Form a quad register from a pair of D registers. 1427/// 1428SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) { 1429 DebugLoc dl = V0.getNode()->getDebugLoc(); 1430 SDValue RegClass = CurDAG->getTargetConstant(ARM::QPRRegClassID, MVT::i32); 1431 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32); 1432 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32); 1433 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 }; 1434 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 5); 1435} 1436 1437/// PairQRegs - Form 4 consecutive D registers from a pair of Q registers. 1438/// 1439SDNode *ARMDAGToDAGISel::PairQRegs(EVT VT, SDValue V0, SDValue V1) { 1440 DebugLoc dl = V0.getNode()->getDebugLoc(); 1441 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, MVT::i32); 1442 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32); 1443 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32); 1444 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 }; 1445 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 5); 1446} 1447 1448/// QuadSRegs - Form 4 consecutive S registers. 1449/// 1450SDNode *ARMDAGToDAGISel::QuadSRegs(EVT VT, SDValue V0, SDValue V1, 1451 SDValue V2, SDValue V3) { 1452 DebugLoc dl = V0.getNode()->getDebugLoc(); 1453 SDValue RegClass = 1454 CurDAG->getTargetConstant(ARM::QPR_VFP2RegClassID, MVT::i32); 1455 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32); 1456 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32); 1457 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32); 1458 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32); 1459 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1, 1460 V2, SubReg2, V3, SubReg3 }; 1461 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 9); 1462} 1463 1464/// QuadDRegs - Form 4 consecutive D registers. 1465/// 1466SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1, 1467 SDValue V2, SDValue V3) { 1468 DebugLoc dl = V0.getNode()->getDebugLoc(); 1469 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, MVT::i32); 1470 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32); 1471 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32); 1472 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32); 1473 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32); 1474 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1, 1475 V2, SubReg2, V3, SubReg3 }; 1476 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 9); 1477} 1478 1479/// QuadQRegs - Form 4 consecutive Q registers. 1480/// 1481SDNode *ARMDAGToDAGISel::QuadQRegs(EVT VT, SDValue V0, SDValue V1, 1482 SDValue V2, SDValue V3) { 1483 DebugLoc dl = V0.getNode()->getDebugLoc(); 1484 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQQQPRRegClassID, MVT::i32); 1485 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32); 1486 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32); 1487 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, MVT::i32); 1488 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, MVT::i32); 1489 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1, 1490 V2, SubReg2, V3, SubReg3 }; 1491 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 9); 1492} 1493 1494/// GetVLDSTAlign - Get the alignment (in bytes) for the alignment operand 1495/// of a NEON VLD or VST instruction. The supported values depend on the 1496/// number of registers being loaded. 1497SDValue ARMDAGToDAGISel::GetVLDSTAlign(SDValue Align, unsigned NumVecs, 1498 bool is64BitVector) { 1499 unsigned NumRegs = NumVecs; 1500 if (!is64BitVector && NumVecs < 3) 1501 NumRegs *= 2; 1502 1503 unsigned Alignment = cast<ConstantSDNode>(Align)->getZExtValue(); 1504 if (Alignment >= 32 && NumRegs == 4) 1505 Alignment = 32; 1506 else if (Alignment >= 16 && (NumRegs == 2 || NumRegs == 4)) 1507 Alignment = 16; 1508 else if (Alignment >= 8) 1509 Alignment = 8; 1510 else 1511 Alignment = 0; 1512 1513 return CurDAG->getTargetConstant(Alignment, MVT::i32); 1514} 1515 1516SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs, 1517 unsigned *DOpcodes, unsigned *QOpcodes0, 1518 unsigned *QOpcodes1) { 1519 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range"); 1520 DebugLoc dl = N->getDebugLoc(); 1521 1522 SDValue MemAddr, Align; 1523 unsigned AddrOpIdx = isUpdating ? 1 : 2; 1524 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align)) 1525 return NULL; 1526 1527 SDValue Chain = N->getOperand(0); 1528 EVT VT = N->getValueType(0); 1529 bool is64BitVector = VT.is64BitVector(); 1530 Align = GetVLDSTAlign(Align, NumVecs, is64BitVector); 1531 1532 unsigned OpcodeIndex; 1533 switch (VT.getSimpleVT().SimpleTy) { 1534 default: llvm_unreachable("unhandled vld type"); 1535 // Double-register operations: 1536 case MVT::v8i8: OpcodeIndex = 0; break; 1537 case MVT::v4i16: OpcodeIndex = 1; break; 1538 case MVT::v2f32: 1539 case MVT::v2i32: OpcodeIndex = 2; break; 1540 case MVT::v1i64: OpcodeIndex = 3; break; 1541 // Quad-register operations: 1542 case MVT::v16i8: OpcodeIndex = 0; break; 1543 case MVT::v8i16: OpcodeIndex = 1; break; 1544 case MVT::v4f32: 1545 case MVT::v4i32: OpcodeIndex = 2; break; 1546 case MVT::v2i64: OpcodeIndex = 3; 1547 assert(NumVecs == 1 && "v2i64 type only supported for VLD1"); 1548 break; 1549 } 1550 1551 EVT ResTy; 1552 if (NumVecs == 1) 1553 ResTy = VT; 1554 else { 1555 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs; 1556 if (!is64BitVector) 1557 ResTyElts *= 2; 1558 ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, ResTyElts); 1559 } 1560 std::vector<EVT> ResTys; 1561 ResTys.push_back(ResTy); 1562 if (isUpdating) 1563 ResTys.push_back(MVT::i32); 1564 ResTys.push_back(MVT::Other); 1565 1566 SDValue Pred = getAL(CurDAG); 1567 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 1568 SDNode *VLd; 1569 SmallVector<SDValue, 7> Ops; 1570 1571 // Double registers and VLD1/VLD2 quad registers are directly supported. 1572 if (is64BitVector || NumVecs <= 2) { 1573 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] : 1574 QOpcodes0[OpcodeIndex]); 1575 Ops.push_back(MemAddr); 1576 Ops.push_back(Align); 1577 if (isUpdating) { 1578 SDValue Inc = N->getOperand(AddrOpIdx + 1); 1579 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc); 1580 } 1581 Ops.push_back(Pred); 1582 Ops.push_back(Reg0); 1583 Ops.push_back(Chain); 1584 VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), Ops.size()); 1585 1586 } else { 1587 // Otherwise, quad registers are loaded with two separate instructions, 1588 // where one loads the even registers and the other loads the odd registers. 1589 EVT AddrTy = MemAddr.getValueType(); 1590 1591 // Load the even subregs. This is always an updating load, so that it 1592 // provides the address to the second load for the odd subregs. 1593 SDValue ImplDef = 1594 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, ResTy), 0); 1595 const SDValue OpsA[] = { MemAddr, Align, Reg0, ImplDef, Pred, Reg0, Chain }; 1596 SDNode *VLdA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl, 1597 ResTy, AddrTy, MVT::Other, OpsA, 7); 1598 Chain = SDValue(VLdA, 2); 1599 1600 // Load the odd subregs. 1601 Ops.push_back(SDValue(VLdA, 1)); 1602 Ops.push_back(Align); 1603 if (isUpdating) { 1604 SDValue Inc = N->getOperand(AddrOpIdx + 1); 1605 assert(isa<ConstantSDNode>(Inc.getNode()) && 1606 "only constant post-increment update allowed for VLD3/4"); 1607 (void)Inc; 1608 Ops.push_back(Reg0); 1609 } 1610 Ops.push_back(SDValue(VLdA, 0)); 1611 Ops.push_back(Pred); 1612 Ops.push_back(Reg0); 1613 Ops.push_back(Chain); 1614 VLd = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys, 1615 Ops.data(), Ops.size()); 1616 } 1617 1618 // Transfer memoperands. 1619 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 1620 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 1621 cast<MachineSDNode>(VLd)->setMemRefs(MemOp, MemOp + 1); 1622 1623 if (NumVecs == 1) 1624 return VLd; 1625 1626 // Extract out the subregisters. 1627 SDValue SuperReg = SDValue(VLd, 0); 1628 assert(ARM::dsub_7 == ARM::dsub_0+7 && 1629 ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering"); 1630 unsigned Sub0 = (is64BitVector ? ARM::dsub_0 : ARM::qsub_0); 1631 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) 1632 ReplaceUses(SDValue(N, Vec), 1633 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg)); 1634 ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, 1)); 1635 if (isUpdating) 1636 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLd, 2)); 1637 return NULL; 1638} 1639 1640SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs, 1641 unsigned *DOpcodes, unsigned *QOpcodes0, 1642 unsigned *QOpcodes1) { 1643 assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range"); 1644 DebugLoc dl = N->getDebugLoc(); 1645 1646 SDValue MemAddr, Align; 1647 unsigned AddrOpIdx = isUpdating ? 1 : 2; 1648 unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1) 1649 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align)) 1650 return NULL; 1651 1652 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 1653 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 1654 1655 SDValue Chain = N->getOperand(0); 1656 EVT VT = N->getOperand(Vec0Idx).getValueType(); 1657 bool is64BitVector = VT.is64BitVector(); 1658 Align = GetVLDSTAlign(Align, NumVecs, is64BitVector); 1659 1660 unsigned OpcodeIndex; 1661 switch (VT.getSimpleVT().SimpleTy) { 1662 default: llvm_unreachable("unhandled vst type"); 1663 // Double-register operations: 1664 case MVT::v8i8: OpcodeIndex = 0; break; 1665 case MVT::v4i16: OpcodeIndex = 1; break; 1666 case MVT::v2f32: 1667 case MVT::v2i32: OpcodeIndex = 2; break; 1668 case MVT::v1i64: OpcodeIndex = 3; break; 1669 // Quad-register operations: 1670 case MVT::v16i8: OpcodeIndex = 0; break; 1671 case MVT::v8i16: OpcodeIndex = 1; break; 1672 case MVT::v4f32: 1673 case MVT::v4i32: OpcodeIndex = 2; break; 1674 case MVT::v2i64: OpcodeIndex = 3; 1675 assert(NumVecs == 1 && "v2i64 type only supported for VST1"); 1676 break; 1677 } 1678 1679 std::vector<EVT> ResTys; 1680 if (isUpdating) 1681 ResTys.push_back(MVT::i32); 1682 ResTys.push_back(MVT::Other); 1683 1684 SDValue Pred = getAL(CurDAG); 1685 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 1686 SmallVector<SDValue, 7> Ops; 1687 1688 // Double registers and VST1/VST2 quad registers are directly supported. 1689 if (is64BitVector || NumVecs <= 2) { 1690 SDValue SrcReg; 1691 if (NumVecs == 1) { 1692 SrcReg = N->getOperand(Vec0Idx); 1693 } else if (is64BitVector) { 1694 // Form a REG_SEQUENCE to force register allocation. 1695 SDValue V0 = N->getOperand(Vec0Idx + 0); 1696 SDValue V1 = N->getOperand(Vec0Idx + 1); 1697 if (NumVecs == 2) 1698 SrcReg = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0); 1699 else { 1700 SDValue V2 = N->getOperand(Vec0Idx + 2); 1701 // If it's a vst3, form a quad D-register and leave the last part as 1702 // an undef. 1703 SDValue V3 = (NumVecs == 3) 1704 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0) 1705 : N->getOperand(Vec0Idx + 3); 1706 SrcReg = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0); 1707 } 1708 } else { 1709 // Form a QQ register. 1710 SDValue Q0 = N->getOperand(Vec0Idx); 1711 SDValue Q1 = N->getOperand(Vec0Idx + 1); 1712 SrcReg = SDValue(PairQRegs(MVT::v4i64, Q0, Q1), 0); 1713 } 1714 1715 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] : 1716 QOpcodes0[OpcodeIndex]); 1717 Ops.push_back(MemAddr); 1718 Ops.push_back(Align); 1719 if (isUpdating) { 1720 SDValue Inc = N->getOperand(AddrOpIdx + 1); 1721 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc); 1722 } 1723 Ops.push_back(SrcReg); 1724 Ops.push_back(Pred); 1725 Ops.push_back(Reg0); 1726 Ops.push_back(Chain); 1727 SDNode *VSt = 1728 CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), Ops.size()); 1729 1730 // Transfer memoperands. 1731 cast<MachineSDNode>(VSt)->setMemRefs(MemOp, MemOp + 1); 1732 1733 return VSt; 1734 } 1735 1736 // Otherwise, quad registers are stored with two separate instructions, 1737 // where one stores the even registers and the other stores the odd registers. 1738 1739 // Form the QQQQ REG_SEQUENCE. 1740 SDValue V0 = N->getOperand(Vec0Idx + 0); 1741 SDValue V1 = N->getOperand(Vec0Idx + 1); 1742 SDValue V2 = N->getOperand(Vec0Idx + 2); 1743 SDValue V3 = (NumVecs == 3) 1744 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0) 1745 : N->getOperand(Vec0Idx + 3); 1746 SDValue RegSeq = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0); 1747 1748 // Store the even D registers. This is always an updating store, so that it 1749 // provides the address to the second store for the odd subregs. 1750 const SDValue OpsA[] = { MemAddr, Align, Reg0, RegSeq, Pred, Reg0, Chain }; 1751 SDNode *VStA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl, 1752 MemAddr.getValueType(), 1753 MVT::Other, OpsA, 7); 1754 cast<MachineSDNode>(VStA)->setMemRefs(MemOp, MemOp + 1); 1755 Chain = SDValue(VStA, 1); 1756 1757 // Store the odd D registers. 1758 Ops.push_back(SDValue(VStA, 0)); 1759 Ops.push_back(Align); 1760 if (isUpdating) { 1761 SDValue Inc = N->getOperand(AddrOpIdx + 1); 1762 assert(isa<ConstantSDNode>(Inc.getNode()) && 1763 "only constant post-increment update allowed for VST3/4"); 1764 (void)Inc; 1765 Ops.push_back(Reg0); 1766 } 1767 Ops.push_back(RegSeq); 1768 Ops.push_back(Pred); 1769 Ops.push_back(Reg0); 1770 Ops.push_back(Chain); 1771 SDNode *VStB = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys, 1772 Ops.data(), Ops.size()); 1773 cast<MachineSDNode>(VStB)->setMemRefs(MemOp, MemOp + 1); 1774 return VStB; 1775} 1776 1777SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad, 1778 bool isUpdating, unsigned NumVecs, 1779 unsigned *DOpcodes, 1780 unsigned *QOpcodes) { 1781 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range"); 1782 DebugLoc dl = N->getDebugLoc(); 1783 1784 SDValue MemAddr, Align; 1785 unsigned AddrOpIdx = isUpdating ? 1 : 2; 1786 unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1) 1787 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align)) 1788 return NULL; 1789 1790 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 1791 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 1792 1793 SDValue Chain = N->getOperand(0); 1794 unsigned Lane = 1795 cast<ConstantSDNode>(N->getOperand(Vec0Idx + NumVecs))->getZExtValue(); 1796 EVT VT = N->getOperand(Vec0Idx).getValueType(); 1797 bool is64BitVector = VT.is64BitVector(); 1798 1799 unsigned Alignment = 0; 1800 if (NumVecs != 3) { 1801 Alignment = cast<ConstantSDNode>(Align)->getZExtValue(); 1802 unsigned NumBytes = NumVecs * VT.getVectorElementType().getSizeInBits()/8; 1803 if (Alignment > NumBytes) 1804 Alignment = NumBytes; 1805 if (Alignment < 8 && Alignment < NumBytes) 1806 Alignment = 0; 1807 // Alignment must be a power of two; make sure of that. 1808 Alignment = (Alignment & -Alignment); 1809 if (Alignment == 1) 1810 Alignment = 0; 1811 } 1812 Align = CurDAG->getTargetConstant(Alignment, MVT::i32); 1813 1814 unsigned OpcodeIndex; 1815 switch (VT.getSimpleVT().SimpleTy) { 1816 default: llvm_unreachable("unhandled vld/vst lane type"); 1817 // Double-register operations: 1818 case MVT::v8i8: OpcodeIndex = 0; break; 1819 case MVT::v4i16: OpcodeIndex = 1; break; 1820 case MVT::v2f32: 1821 case MVT::v2i32: OpcodeIndex = 2; break; 1822 // Quad-register operations: 1823 case MVT::v8i16: OpcodeIndex = 0; break; 1824 case MVT::v4f32: 1825 case MVT::v4i32: OpcodeIndex = 1; break; 1826 } 1827 1828 std::vector<EVT> ResTys; 1829 if (IsLoad) { 1830 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs; 1831 if (!is64BitVector) 1832 ResTyElts *= 2; 1833 ResTys.push_back(EVT::getVectorVT(*CurDAG->getContext(), 1834 MVT::i64, ResTyElts)); 1835 } 1836 if (isUpdating) 1837 ResTys.push_back(MVT::i32); 1838 ResTys.push_back(MVT::Other); 1839 1840 SDValue Pred = getAL(CurDAG); 1841 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 1842 1843 SmallVector<SDValue, 8> Ops; 1844 Ops.push_back(MemAddr); 1845 Ops.push_back(Align); 1846 if (isUpdating) { 1847 SDValue Inc = N->getOperand(AddrOpIdx + 1); 1848 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc); 1849 } 1850 1851 SDValue SuperReg; 1852 SDValue V0 = N->getOperand(Vec0Idx + 0); 1853 SDValue V1 = N->getOperand(Vec0Idx + 1); 1854 if (NumVecs == 2) { 1855 if (is64BitVector) 1856 SuperReg = SDValue(PairDRegs(MVT::v2i64, V0, V1), 0); 1857 else 1858 SuperReg = SDValue(PairQRegs(MVT::v4i64, V0, V1), 0); 1859 } else { 1860 SDValue V2 = N->getOperand(Vec0Idx + 2); 1861 SDValue V3 = (NumVecs == 3) 1862 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0) 1863 : N->getOperand(Vec0Idx + 3); 1864 if (is64BitVector) 1865 SuperReg = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0); 1866 else 1867 SuperReg = SDValue(QuadQRegs(MVT::v8i64, V0, V1, V2, V3), 0); 1868 } 1869 Ops.push_back(SuperReg); 1870 Ops.push_back(getI32Imm(Lane)); 1871 Ops.push_back(Pred); 1872 Ops.push_back(Reg0); 1873 Ops.push_back(Chain); 1874 1875 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] : 1876 QOpcodes[OpcodeIndex]); 1877 SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, 1878 Ops.data(), Ops.size()); 1879 cast<MachineSDNode>(VLdLn)->setMemRefs(MemOp, MemOp + 1); 1880 if (!IsLoad) 1881 return VLdLn; 1882 1883 // Extract the subregisters. 1884 SuperReg = SDValue(VLdLn, 0); 1885 assert(ARM::dsub_7 == ARM::dsub_0+7 && 1886 ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering"); 1887 unsigned Sub0 = is64BitVector ? ARM::dsub_0 : ARM::qsub_0; 1888 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) 1889 ReplaceUses(SDValue(N, Vec), 1890 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg)); 1891 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, 1)); 1892 if (isUpdating) 1893 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdLn, 2)); 1894 return NULL; 1895} 1896 1897SDNode *ARMDAGToDAGISel::SelectVLDDup(SDNode *N, bool isUpdating, 1898 unsigned NumVecs, unsigned *Opcodes) { 1899 assert(NumVecs >=2 && NumVecs <= 4 && "VLDDup NumVecs out-of-range"); 1900 DebugLoc dl = N->getDebugLoc(); 1901 1902 SDValue MemAddr, Align; 1903 if (!SelectAddrMode6(N, N->getOperand(1), MemAddr, Align)) 1904 return NULL; 1905 1906 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 1907 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 1908 1909 SDValue Chain = N->getOperand(0); 1910 EVT VT = N->getValueType(0); 1911 1912 unsigned Alignment = 0; 1913 if (NumVecs != 3) { 1914 Alignment = cast<ConstantSDNode>(Align)->getZExtValue(); 1915 unsigned NumBytes = NumVecs * VT.getVectorElementType().getSizeInBits()/8; 1916 if (Alignment > NumBytes) 1917 Alignment = NumBytes; 1918 if (Alignment < 8 && Alignment < NumBytes) 1919 Alignment = 0; 1920 // Alignment must be a power of two; make sure of that. 1921 Alignment = (Alignment & -Alignment); 1922 if (Alignment == 1) 1923 Alignment = 0; 1924 } 1925 Align = CurDAG->getTargetConstant(Alignment, MVT::i32); 1926 1927 unsigned OpcodeIndex; 1928 switch (VT.getSimpleVT().SimpleTy) { 1929 default: llvm_unreachable("unhandled vld-dup type"); 1930 case MVT::v8i8: OpcodeIndex = 0; break; 1931 case MVT::v4i16: OpcodeIndex = 1; break; 1932 case MVT::v2f32: 1933 case MVT::v2i32: OpcodeIndex = 2; break; 1934 } 1935 1936 SDValue Pred = getAL(CurDAG); 1937 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 1938 SDValue SuperReg; 1939 unsigned Opc = Opcodes[OpcodeIndex]; 1940 SmallVector<SDValue, 6> Ops; 1941 Ops.push_back(MemAddr); 1942 Ops.push_back(Align); 1943 if (isUpdating) { 1944 SDValue Inc = N->getOperand(2); 1945 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc); 1946 } 1947 Ops.push_back(Pred); 1948 Ops.push_back(Reg0); 1949 Ops.push_back(Chain); 1950 1951 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs; 1952 std::vector<EVT> ResTys; 1953 ResTys.push_back(EVT::getVectorVT(*CurDAG->getContext(), MVT::i64,ResTyElts)); 1954 if (isUpdating) 1955 ResTys.push_back(MVT::i32); 1956 ResTys.push_back(MVT::Other); 1957 SDNode *VLdDup = 1958 CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), Ops.size()); 1959 cast<MachineSDNode>(VLdDup)->setMemRefs(MemOp, MemOp + 1); 1960 SuperReg = SDValue(VLdDup, 0); 1961 1962 // Extract the subregisters. 1963 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering"); 1964 unsigned SubIdx = ARM::dsub_0; 1965 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) 1966 ReplaceUses(SDValue(N, Vec), 1967 CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, SuperReg)); 1968 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdDup, 1)); 1969 if (isUpdating) 1970 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdDup, 2)); 1971 return NULL; 1972} 1973 1974SDNode *ARMDAGToDAGISel::SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs, 1975 unsigned Opc) { 1976 assert(NumVecs >= 2 && NumVecs <= 4 && "VTBL NumVecs out-of-range"); 1977 DebugLoc dl = N->getDebugLoc(); 1978 EVT VT = N->getValueType(0); 1979 unsigned FirstTblReg = IsExt ? 2 : 1; 1980 1981 // Form a REG_SEQUENCE to force register allocation. 1982 SDValue RegSeq; 1983 SDValue V0 = N->getOperand(FirstTblReg + 0); 1984 SDValue V1 = N->getOperand(FirstTblReg + 1); 1985 if (NumVecs == 2) 1986 RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0); 1987 else { 1988 SDValue V2 = N->getOperand(FirstTblReg + 2); 1989 // If it's a vtbl3, form a quad D-register and leave the last part as 1990 // an undef. 1991 SDValue V3 = (NumVecs == 3) 1992 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0) 1993 : N->getOperand(FirstTblReg + 3); 1994 RegSeq = SDValue(QuadDRegs(MVT::v4i64, V0, V1, V2, V3), 0); 1995 } 1996 1997 SmallVector<SDValue, 6> Ops; 1998 if (IsExt) 1999 Ops.push_back(N->getOperand(1)); 2000 Ops.push_back(RegSeq); 2001 Ops.push_back(N->getOperand(FirstTblReg + NumVecs)); 2002 Ops.push_back(getAL(CurDAG)); // predicate 2003 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // predicate register 2004 return CurDAG->getMachineNode(Opc, dl, VT, Ops.data(), Ops.size()); 2005} 2006 2007SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N, 2008 bool isSigned) { 2009 if (!Subtarget->hasV6T2Ops()) 2010 return NULL; 2011 2012 unsigned Opc = isSigned ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX) 2013 : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX); 2014 2015 2016 // For unsigned extracts, check for a shift right and mask 2017 unsigned And_imm = 0; 2018 if (N->getOpcode() == ISD::AND) { 2019 if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) { 2020 2021 // The immediate is a mask of the low bits iff imm & (imm+1) == 0 2022 if (And_imm & (And_imm + 1)) 2023 return NULL; 2024 2025 unsigned Srl_imm = 0; 2026 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL, 2027 Srl_imm)) { 2028 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!"); 2029 2030 // Note: The width operand is encoded as width-1. 2031 unsigned Width = CountTrailingOnes_32(And_imm) - 1; 2032 unsigned LSB = Srl_imm; 2033 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 2034 SDValue Ops[] = { N->getOperand(0).getOperand(0), 2035 CurDAG->getTargetConstant(LSB, MVT::i32), 2036 CurDAG->getTargetConstant(Width, MVT::i32), 2037 getAL(CurDAG), Reg0 }; 2038 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5); 2039 } 2040 } 2041 return NULL; 2042 } 2043 2044 // Otherwise, we're looking for a shift of a shift 2045 unsigned Shl_imm = 0; 2046 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) { 2047 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!"); 2048 unsigned Srl_imm = 0; 2049 if (isInt32Immediate(N->getOperand(1), Srl_imm)) { 2050 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!"); 2051 // Note: The width operand is encoded as width-1. 2052 unsigned Width = 32 - Srl_imm - 1; 2053 int LSB = Srl_imm - Shl_imm; 2054 if (LSB < 0) 2055 return NULL; 2056 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 2057 SDValue Ops[] = { N->getOperand(0).getOperand(0), 2058 CurDAG->getTargetConstant(LSB, MVT::i32), 2059 CurDAG->getTargetConstant(Width, MVT::i32), 2060 getAL(CurDAG), Reg0 }; 2061 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5); 2062 } 2063 } 2064 return NULL; 2065} 2066 2067SDNode *ARMDAGToDAGISel:: 2068SelectT2CMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal, 2069 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) { 2070 SDValue CPTmp0; 2071 SDValue CPTmp1; 2072 if (SelectT2ShifterOperandReg(TrueVal, CPTmp0, CPTmp1)) { 2073 unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue(); 2074 unsigned SOShOp = ARM_AM::getSORegShOp(SOVal); 2075 unsigned Opc = 0; 2076 switch (SOShOp) { 2077 case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break; 2078 case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break; 2079 case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break; 2080 case ARM_AM::ror: Opc = ARM::t2MOVCCror; break; 2081 default: 2082 llvm_unreachable("Unknown so_reg opcode!"); 2083 break; 2084 } 2085 SDValue SOShImm = 2086 CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32); 2087 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32); 2088 SDValue Ops[] = { FalseVal, CPTmp0, SOShImm, CC, CCR, InFlag }; 2089 return CurDAG->SelectNodeTo(N, Opc, MVT::i32,Ops, 6); 2090 } 2091 return 0; 2092} 2093 2094SDNode *ARMDAGToDAGISel:: 2095SelectARMCMOVShiftOp(SDNode *N, SDValue FalseVal, SDValue TrueVal, 2096 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) { 2097 SDValue CPTmp0; 2098 SDValue CPTmp1; 2099 SDValue CPTmp2; 2100 if (SelectImmShifterOperand(TrueVal, CPTmp0, CPTmp2)) { 2101 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32); 2102 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp2, CC, CCR, InFlag }; 2103 return CurDAG->SelectNodeTo(N, ARM::MOVCCsi, MVT::i32, Ops, 6); 2104 } 2105 2106 if (SelectRegShifterOperand(TrueVal, CPTmp0, CPTmp1, CPTmp2)) { 2107 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32); 2108 SDValue Ops[] = { FalseVal, CPTmp0, CPTmp1, CPTmp2, CC, CCR, InFlag }; 2109 return CurDAG->SelectNodeTo(N, ARM::MOVCCsr, MVT::i32, Ops, 7); 2110 } 2111 return 0; 2112} 2113 2114SDNode *ARMDAGToDAGISel:: 2115SelectT2CMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal, 2116 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) { 2117 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal); 2118 if (!T) 2119 return 0; 2120 2121 unsigned Opc = 0; 2122 unsigned TrueImm = T->getZExtValue(); 2123 if (is_t2_so_imm(TrueImm)) { 2124 Opc = ARM::t2MOVCCi; 2125 } else if (TrueImm <= 0xffff) { 2126 Opc = ARM::t2MOVCCi16; 2127 } else if (is_t2_so_imm_not(TrueImm)) { 2128 TrueImm = ~TrueImm; 2129 Opc = ARM::t2MVNCCi; 2130 } else if (TrueVal.getNode()->hasOneUse() && Subtarget->hasV6T2Ops()) { 2131 // Large immediate. 2132 Opc = ARM::t2MOVCCi32imm; 2133 } 2134 2135 if (Opc) { 2136 SDValue True = CurDAG->getTargetConstant(TrueImm, MVT::i32); 2137 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32); 2138 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag }; 2139 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5); 2140 } 2141 2142 return 0; 2143} 2144 2145SDNode *ARMDAGToDAGISel:: 2146SelectARMCMOVImmOp(SDNode *N, SDValue FalseVal, SDValue TrueVal, 2147 ARMCC::CondCodes CCVal, SDValue CCR, SDValue InFlag) { 2148 ConstantSDNode *T = dyn_cast<ConstantSDNode>(TrueVal); 2149 if (!T) 2150 return 0; 2151 2152 unsigned Opc = 0; 2153 unsigned TrueImm = T->getZExtValue(); 2154 bool isSoImm = is_so_imm(TrueImm); 2155 if (isSoImm) { 2156 Opc = ARM::MOVCCi; 2157 } else if (Subtarget->hasV6T2Ops() && TrueImm <= 0xffff) { 2158 Opc = ARM::MOVCCi16; 2159 } else if (is_so_imm_not(TrueImm)) { 2160 TrueImm = ~TrueImm; 2161 Opc = ARM::MVNCCi; 2162 } else if (TrueVal.getNode()->hasOneUse() && 2163 (Subtarget->hasV6T2Ops() || ARM_AM::isSOImmTwoPartVal(TrueImm))) { 2164 // Large immediate. 2165 Opc = ARM::MOVCCi32imm; 2166 } 2167 2168 if (Opc) { 2169 SDValue True = CurDAG->getTargetConstant(TrueImm, MVT::i32); 2170 SDValue CC = CurDAG->getTargetConstant(CCVal, MVT::i32); 2171 SDValue Ops[] = { FalseVal, True, CC, CCR, InFlag }; 2172 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5); 2173 } 2174 2175 return 0; 2176} 2177 2178SDNode *ARMDAGToDAGISel::SelectCMOVOp(SDNode *N) { 2179 EVT VT = N->getValueType(0); 2180 SDValue FalseVal = N->getOperand(0); 2181 SDValue TrueVal = N->getOperand(1); 2182 SDValue CC = N->getOperand(2); 2183 SDValue CCR = N->getOperand(3); 2184 SDValue InFlag = N->getOperand(4); 2185 assert(CC.getOpcode() == ISD::Constant); 2186 assert(CCR.getOpcode() == ISD::Register); 2187 ARMCC::CondCodes CCVal = 2188 (ARMCC::CondCodes)cast<ConstantSDNode>(CC)->getZExtValue(); 2189 2190 if (!Subtarget->isThumb1Only() && VT == MVT::i32) { 2191 // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc) 2192 // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc) 2193 // Pattern complexity = 18 cost = 1 size = 0 2194 SDValue CPTmp0; 2195 SDValue CPTmp1; 2196 SDValue CPTmp2; 2197 if (Subtarget->isThumb()) { 2198 SDNode *Res = SelectT2CMOVShiftOp(N, FalseVal, TrueVal, 2199 CCVal, CCR, InFlag); 2200 if (!Res) 2201 Res = SelectT2CMOVShiftOp(N, TrueVal, FalseVal, 2202 ARMCC::getOppositeCondition(CCVal), CCR, InFlag); 2203 if (Res) 2204 return Res; 2205 } else { 2206 SDNode *Res = SelectARMCMOVShiftOp(N, FalseVal, TrueVal, 2207 CCVal, CCR, InFlag); 2208 if (!Res) 2209 Res = SelectARMCMOVShiftOp(N, TrueVal, FalseVal, 2210 ARMCC::getOppositeCondition(CCVal), CCR, InFlag); 2211 if (Res) 2212 return Res; 2213 } 2214 2215 // Pattern: (ARMcmov:i32 GPR:i32:$false, 2216 // (imm:i32)<<P:Pred_so_imm>>:$true, 2217 // (imm:i32):$cc) 2218 // Emits: (MOVCCi:i32 GPR:i32:$false, 2219 // (so_imm:i32 (imm:i32):$true), (imm:i32):$cc) 2220 // Pattern complexity = 10 cost = 1 size = 0 2221 if (Subtarget->isThumb()) { 2222 SDNode *Res = SelectT2CMOVImmOp(N, FalseVal, TrueVal, 2223 CCVal, CCR, InFlag); 2224 if (!Res) 2225 Res = SelectT2CMOVImmOp(N, TrueVal, FalseVal, 2226 ARMCC::getOppositeCondition(CCVal), CCR, InFlag); 2227 if (Res) 2228 return Res; 2229 } else { 2230 SDNode *Res = SelectARMCMOVImmOp(N, FalseVal, TrueVal, 2231 CCVal, CCR, InFlag); 2232 if (!Res) 2233 Res = SelectARMCMOVImmOp(N, TrueVal, FalseVal, 2234 ARMCC::getOppositeCondition(CCVal), CCR, InFlag); 2235 if (Res) 2236 return Res; 2237 } 2238 } 2239 2240 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc) 2241 // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc) 2242 // Pattern complexity = 6 cost = 1 size = 0 2243 // 2244 // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc) 2245 // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc) 2246 // Pattern complexity = 6 cost = 11 size = 0 2247 // 2248 // Also VMOVScc and VMOVDcc. 2249 SDValue Tmp2 = CurDAG->getTargetConstant(CCVal, MVT::i32); 2250 SDValue Ops[] = { FalseVal, TrueVal, Tmp2, CCR, InFlag }; 2251 unsigned Opc = 0; 2252 switch (VT.getSimpleVT().SimpleTy) { 2253 default: assert(false && "Illegal conditional move type!"); 2254 break; 2255 case MVT::i32: 2256 Opc = Subtarget->isThumb() 2257 ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo) 2258 : ARM::MOVCCr; 2259 break; 2260 case MVT::f32: 2261 Opc = ARM::VMOVScc; 2262 break; 2263 case MVT::f64: 2264 Opc = ARM::VMOVDcc; 2265 break; 2266 } 2267 return CurDAG->SelectNodeTo(N, Opc, VT, Ops, 5); 2268} 2269 2270SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) { 2271 // The only time a CONCAT_VECTORS operation can have legal types is when 2272 // two 64-bit vectors are concatenated to a 128-bit vector. 2273 EVT VT = N->getValueType(0); 2274 if (!VT.is128BitVector() || N->getNumOperands() != 2) 2275 llvm_unreachable("unexpected CONCAT_VECTORS"); 2276 return PairDRegs(VT, N->getOperand(0), N->getOperand(1)); 2277} 2278 2279SDNode *ARMDAGToDAGISel::Select(SDNode *N) { 2280 DebugLoc dl = N->getDebugLoc(); 2281 2282 if (N->isMachineOpcode()) 2283 return NULL; // Already selected. 2284 2285 switch (N->getOpcode()) { 2286 default: break; 2287 case ISD::Constant: { 2288 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue(); 2289 bool UseCP = true; 2290 if (Subtarget->hasThumb2()) 2291 // Thumb2-aware targets have the MOVT instruction, so all immediates can 2292 // be done with MOV + MOVT, at worst. 2293 UseCP = 0; 2294 else { 2295 if (Subtarget->isThumb()) { 2296 UseCP = (Val > 255 && // MOV 2297 ~Val > 255 && // MOV + MVN 2298 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL 2299 } else 2300 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV 2301 ARM_AM::getSOImmVal(~Val) == -1 && // MVN 2302 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs. 2303 } 2304 2305 if (UseCP) { 2306 SDValue CPIdx = 2307 CurDAG->getTargetConstantPool(ConstantInt::get( 2308 Type::getInt32Ty(*CurDAG->getContext()), Val), 2309 TLI.getPointerTy()); 2310 2311 SDNode *ResNode; 2312 if (Subtarget->isThumb1Only()) { 2313 SDValue Pred = getAL(CurDAG); 2314 SDValue PredReg = CurDAG->getRegister(0, MVT::i32); 2315 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() }; 2316 ResNode = CurDAG->getMachineNode(ARM::tLDRpci, dl, MVT::i32, MVT::Other, 2317 Ops, 4); 2318 } else { 2319 SDValue Ops[] = { 2320 CPIdx, 2321 CurDAG->getTargetConstant(0, MVT::i32), 2322 getAL(CurDAG), 2323 CurDAG->getRegister(0, MVT::i32), 2324 CurDAG->getEntryNode() 2325 }; 2326 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other, 2327 Ops, 5); 2328 } 2329 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0)); 2330 return NULL; 2331 } 2332 2333 // Other cases are autogenerated. 2334 break; 2335 } 2336 case ISD::FrameIndex: { 2337 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm. 2338 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 2339 SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy()); 2340 if (Subtarget->isThumb1Only()) { 2341 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32), 2342 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) }; 2343 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, Ops, 4); 2344 } else { 2345 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ? 2346 ARM::t2ADDri : ARM::ADDri); 2347 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32), 2348 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), 2349 CurDAG->getRegister(0, MVT::i32) }; 2350 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5); 2351 } 2352 } 2353 case ISD::SRL: 2354 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false)) 2355 return I; 2356 break; 2357 case ISD::SRA: 2358 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true)) 2359 return I; 2360 break; 2361 case ISD::MUL: 2362 if (Subtarget->isThumb1Only()) 2363 break; 2364 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) { 2365 unsigned RHSV = C->getZExtValue(); 2366 if (!RHSV) break; 2367 if (isPowerOf2_32(RHSV-1)) { // 2^n+1? 2368 unsigned ShImm = Log2_32(RHSV-1); 2369 if (ShImm >= 32) 2370 break; 2371 SDValue V = N->getOperand(0); 2372 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm); 2373 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32); 2374 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 2375 if (Subtarget->isThumb()) { 2376 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 }; 2377 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6); 2378 } else { 2379 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 }; 2380 return CurDAG->SelectNodeTo(N, ARM::ADDrsi, MVT::i32, Ops, 7); 2381 } 2382 } 2383 if (isPowerOf2_32(RHSV+1)) { // 2^n-1? 2384 unsigned ShImm = Log2_32(RHSV+1); 2385 if (ShImm >= 32) 2386 break; 2387 SDValue V = N->getOperand(0); 2388 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm); 2389 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32); 2390 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 2391 if (Subtarget->isThumb()) { 2392 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 }; 2393 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 6); 2394 } else { 2395 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 }; 2396 return CurDAG->SelectNodeTo(N, ARM::RSBrsi, MVT::i32, Ops, 7); 2397 } 2398 } 2399 } 2400 break; 2401 case ISD::AND: { 2402 // Check for unsigned bitfield extract 2403 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false)) 2404 return I; 2405 2406 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits 2407 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits 2408 // are entirely contributed by c2 and lower 16-bits are entirely contributed 2409 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)). 2410 // Select it to: "movt x, ((c1 & 0xffff) >> 16) 2411 EVT VT = N->getValueType(0); 2412 if (VT != MVT::i32) 2413 break; 2414 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2()) 2415 ? ARM::t2MOVTi16 2416 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0); 2417 if (!Opc) 2418 break; 2419 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); 2420 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2421 if (!N1C) 2422 break; 2423 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) { 2424 SDValue N2 = N0.getOperand(1); 2425 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2); 2426 if (!N2C) 2427 break; 2428 unsigned N1CVal = N1C->getZExtValue(); 2429 unsigned N2CVal = N2C->getZExtValue(); 2430 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) && 2431 (N1CVal & 0xffffU) == 0xffffU && 2432 (N2CVal & 0xffffU) == 0x0U) { 2433 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16, 2434 MVT::i32); 2435 SDValue Ops[] = { N0.getOperand(0), Imm16, 2436 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) }; 2437 return CurDAG->getMachineNode(Opc, dl, VT, Ops, 4); 2438 } 2439 } 2440 break; 2441 } 2442 case ARMISD::VMOVRRD: 2443 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32, 2444 N->getOperand(0), getAL(CurDAG), 2445 CurDAG->getRegister(0, MVT::i32)); 2446 case ISD::UMUL_LOHI: { 2447 if (Subtarget->isThumb1Only()) 2448 break; 2449 if (Subtarget->isThumb()) { 2450 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), 2451 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), 2452 CurDAG->getRegister(0, MVT::i32) }; 2453 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32,Ops,4); 2454 } else { 2455 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), 2456 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), 2457 CurDAG->getRegister(0, MVT::i32) }; 2458 return CurDAG->getMachineNode(Subtarget->hasV6Ops() ? 2459 ARM::UMULL : ARM::UMULLv5, 2460 dl, MVT::i32, MVT::i32, Ops, 5); 2461 } 2462 } 2463 case ISD::SMUL_LOHI: { 2464 if (Subtarget->isThumb1Only()) 2465 break; 2466 if (Subtarget->isThumb()) { 2467 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), 2468 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) }; 2469 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32,Ops,4); 2470 } else { 2471 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), 2472 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), 2473 CurDAG->getRegister(0, MVT::i32) }; 2474 return CurDAG->getMachineNode(Subtarget->hasV6Ops() ? 2475 ARM::SMULL : ARM::SMULLv5, 2476 dl, MVT::i32, MVT::i32, Ops, 5); 2477 } 2478 } 2479 case ISD::LOAD: { 2480 SDNode *ResNode = 0; 2481 if (Subtarget->isThumb() && Subtarget->hasThumb2()) 2482 ResNode = SelectT2IndexedLoad(N); 2483 else 2484 ResNode = SelectARMIndexedLoad(N); 2485 if (ResNode) 2486 return ResNode; 2487 // Other cases are autogenerated. 2488 break; 2489 } 2490 case ARMISD::BRCOND: { 2491 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc) 2492 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc) 2493 // Pattern complexity = 6 cost = 1 size = 0 2494 2495 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc) 2496 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc) 2497 // Pattern complexity = 6 cost = 1 size = 0 2498 2499 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc) 2500 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc) 2501 // Pattern complexity = 6 cost = 1 size = 0 2502 2503 unsigned Opc = Subtarget->isThumb() ? 2504 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc; 2505 SDValue Chain = N->getOperand(0); 2506 SDValue N1 = N->getOperand(1); 2507 SDValue N2 = N->getOperand(2); 2508 SDValue N3 = N->getOperand(3); 2509 SDValue InFlag = N->getOperand(4); 2510 assert(N1.getOpcode() == ISD::BasicBlock); 2511 assert(N2.getOpcode() == ISD::Constant); 2512 assert(N3.getOpcode() == ISD::Register); 2513 2514 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned) 2515 cast<ConstantSDNode>(N2)->getZExtValue()), 2516 MVT::i32); 2517 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag }; 2518 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other, 2519 MVT::Glue, Ops, 5); 2520 Chain = SDValue(ResNode, 0); 2521 if (N->getNumValues() == 2) { 2522 InFlag = SDValue(ResNode, 1); 2523 ReplaceUses(SDValue(N, 1), InFlag); 2524 } 2525 ReplaceUses(SDValue(N, 0), 2526 SDValue(Chain.getNode(), Chain.getResNo())); 2527 return NULL; 2528 } 2529 case ARMISD::CMOV: 2530 return SelectCMOVOp(N); 2531 case ARMISD::VZIP: { 2532 unsigned Opc = 0; 2533 EVT VT = N->getValueType(0); 2534 switch (VT.getSimpleVT().SimpleTy) { 2535 default: return NULL; 2536 case MVT::v8i8: Opc = ARM::VZIPd8; break; 2537 case MVT::v4i16: Opc = ARM::VZIPd16; break; 2538 case MVT::v2f32: 2539 case MVT::v2i32: Opc = ARM::VZIPd32; break; 2540 case MVT::v16i8: Opc = ARM::VZIPq8; break; 2541 case MVT::v8i16: Opc = ARM::VZIPq16; break; 2542 case MVT::v4f32: 2543 case MVT::v4i32: Opc = ARM::VZIPq32; break; 2544 } 2545 SDValue Pred = getAL(CurDAG); 2546 SDValue PredReg = CurDAG->getRegister(0, MVT::i32); 2547 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg }; 2548 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4); 2549 } 2550 case ARMISD::VUZP: { 2551 unsigned Opc = 0; 2552 EVT VT = N->getValueType(0); 2553 switch (VT.getSimpleVT().SimpleTy) { 2554 default: return NULL; 2555 case MVT::v8i8: Opc = ARM::VUZPd8; break; 2556 case MVT::v4i16: Opc = ARM::VUZPd16; break; 2557 case MVT::v2f32: 2558 case MVT::v2i32: Opc = ARM::VUZPd32; break; 2559 case MVT::v16i8: Opc = ARM::VUZPq8; break; 2560 case MVT::v8i16: Opc = ARM::VUZPq16; break; 2561 case MVT::v4f32: 2562 case MVT::v4i32: Opc = ARM::VUZPq32; break; 2563 } 2564 SDValue Pred = getAL(CurDAG); 2565 SDValue PredReg = CurDAG->getRegister(0, MVT::i32); 2566 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg }; 2567 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4); 2568 } 2569 case ARMISD::VTRN: { 2570 unsigned Opc = 0; 2571 EVT VT = N->getValueType(0); 2572 switch (VT.getSimpleVT().SimpleTy) { 2573 default: return NULL; 2574 case MVT::v8i8: Opc = ARM::VTRNd8; break; 2575 case MVT::v4i16: Opc = ARM::VTRNd16; break; 2576 case MVT::v2f32: 2577 case MVT::v2i32: Opc = ARM::VTRNd32; break; 2578 case MVT::v16i8: Opc = ARM::VTRNq8; break; 2579 case MVT::v8i16: Opc = ARM::VTRNq16; break; 2580 case MVT::v4f32: 2581 case MVT::v4i32: Opc = ARM::VTRNq32; break; 2582 } 2583 SDValue Pred = getAL(CurDAG); 2584 SDValue PredReg = CurDAG->getRegister(0, MVT::i32); 2585 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg }; 2586 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4); 2587 } 2588 case ARMISD::BUILD_VECTOR: { 2589 EVT VecVT = N->getValueType(0); 2590 EVT EltVT = VecVT.getVectorElementType(); 2591 unsigned NumElts = VecVT.getVectorNumElements(); 2592 if (EltVT == MVT::f64) { 2593 assert(NumElts == 2 && "unexpected type for BUILD_VECTOR"); 2594 return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1)); 2595 } 2596 assert(EltVT == MVT::f32 && "unexpected type for BUILD_VECTOR"); 2597 if (NumElts == 2) 2598 return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1)); 2599 assert(NumElts == 4 && "unexpected type for BUILD_VECTOR"); 2600 return QuadSRegs(VecVT, N->getOperand(0), N->getOperand(1), 2601 N->getOperand(2), N->getOperand(3)); 2602 } 2603 2604 case ARMISD::VLD2DUP: { 2605 unsigned Opcodes[] = { ARM::VLD2DUPd8Pseudo, ARM::VLD2DUPd16Pseudo, 2606 ARM::VLD2DUPd32Pseudo }; 2607 return SelectVLDDup(N, false, 2, Opcodes); 2608 } 2609 2610 case ARMISD::VLD3DUP: { 2611 unsigned Opcodes[] = { ARM::VLD3DUPd8Pseudo, ARM::VLD3DUPd16Pseudo, 2612 ARM::VLD3DUPd32Pseudo }; 2613 return SelectVLDDup(N, false, 3, Opcodes); 2614 } 2615 2616 case ARMISD::VLD4DUP: { 2617 unsigned Opcodes[] = { ARM::VLD4DUPd8Pseudo, ARM::VLD4DUPd16Pseudo, 2618 ARM::VLD4DUPd32Pseudo }; 2619 return SelectVLDDup(N, false, 4, Opcodes); 2620 } 2621 2622 case ARMISD::VLD2DUP_UPD: { 2623 unsigned Opcodes[] = { ARM::VLD2DUPd8Pseudo_UPD, ARM::VLD2DUPd16Pseudo_UPD, 2624 ARM::VLD2DUPd32Pseudo_UPD }; 2625 return SelectVLDDup(N, true, 2, Opcodes); 2626 } 2627 2628 case ARMISD::VLD3DUP_UPD: { 2629 unsigned Opcodes[] = { ARM::VLD3DUPd8Pseudo_UPD, ARM::VLD3DUPd16Pseudo_UPD, 2630 ARM::VLD3DUPd32Pseudo_UPD }; 2631 return SelectVLDDup(N, true, 3, Opcodes); 2632 } 2633 2634 case ARMISD::VLD4DUP_UPD: { 2635 unsigned Opcodes[] = { ARM::VLD4DUPd8Pseudo_UPD, ARM::VLD4DUPd16Pseudo_UPD, 2636 ARM::VLD4DUPd32Pseudo_UPD }; 2637 return SelectVLDDup(N, true, 4, Opcodes); 2638 } 2639 2640 case ARMISD::VLD1_UPD: { 2641 unsigned DOpcodes[] = { ARM::VLD1d8_UPD, ARM::VLD1d16_UPD, 2642 ARM::VLD1d32_UPD, ARM::VLD1d64_UPD }; 2643 unsigned QOpcodes[] = { ARM::VLD1q8Pseudo_UPD, ARM::VLD1q16Pseudo_UPD, 2644 ARM::VLD1q32Pseudo_UPD, ARM::VLD1q64Pseudo_UPD }; 2645 return SelectVLD(N, true, 1, DOpcodes, QOpcodes, 0); 2646 } 2647 2648 case ARMISD::VLD2_UPD: { 2649 unsigned DOpcodes[] = { ARM::VLD2d8Pseudo_UPD, ARM::VLD2d16Pseudo_UPD, 2650 ARM::VLD2d32Pseudo_UPD, ARM::VLD1q64Pseudo_UPD }; 2651 unsigned QOpcodes[] = { ARM::VLD2q8Pseudo_UPD, ARM::VLD2q16Pseudo_UPD, 2652 ARM::VLD2q32Pseudo_UPD }; 2653 return SelectVLD(N, true, 2, DOpcodes, QOpcodes, 0); 2654 } 2655 2656 case ARMISD::VLD3_UPD: { 2657 unsigned DOpcodes[] = { ARM::VLD3d8Pseudo_UPD, ARM::VLD3d16Pseudo_UPD, 2658 ARM::VLD3d32Pseudo_UPD, ARM::VLD1d64TPseudo_UPD }; 2659 unsigned QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD, 2660 ARM::VLD3q16Pseudo_UPD, 2661 ARM::VLD3q32Pseudo_UPD }; 2662 unsigned QOpcodes1[] = { ARM::VLD3q8oddPseudo_UPD, 2663 ARM::VLD3q16oddPseudo_UPD, 2664 ARM::VLD3q32oddPseudo_UPD }; 2665 return SelectVLD(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1); 2666 } 2667 2668 case ARMISD::VLD4_UPD: { 2669 unsigned DOpcodes[] = { ARM::VLD4d8Pseudo_UPD, ARM::VLD4d16Pseudo_UPD, 2670 ARM::VLD4d32Pseudo_UPD, ARM::VLD1d64QPseudo_UPD }; 2671 unsigned QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD, 2672 ARM::VLD4q16Pseudo_UPD, 2673 ARM::VLD4q32Pseudo_UPD }; 2674 unsigned QOpcodes1[] = { ARM::VLD4q8oddPseudo_UPD, 2675 ARM::VLD4q16oddPseudo_UPD, 2676 ARM::VLD4q32oddPseudo_UPD }; 2677 return SelectVLD(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1); 2678 } 2679 2680 case ARMISD::VLD2LN_UPD: { 2681 unsigned DOpcodes[] = { ARM::VLD2LNd8Pseudo_UPD, ARM::VLD2LNd16Pseudo_UPD, 2682 ARM::VLD2LNd32Pseudo_UPD }; 2683 unsigned QOpcodes[] = { ARM::VLD2LNq16Pseudo_UPD, 2684 ARM::VLD2LNq32Pseudo_UPD }; 2685 return SelectVLDSTLane(N, true, true, 2, DOpcodes, QOpcodes); 2686 } 2687 2688 case ARMISD::VLD3LN_UPD: { 2689 unsigned DOpcodes[] = { ARM::VLD3LNd8Pseudo_UPD, ARM::VLD3LNd16Pseudo_UPD, 2690 ARM::VLD3LNd32Pseudo_UPD }; 2691 unsigned QOpcodes[] = { ARM::VLD3LNq16Pseudo_UPD, 2692 ARM::VLD3LNq32Pseudo_UPD }; 2693 return SelectVLDSTLane(N, true, true, 3, DOpcodes, QOpcodes); 2694 } 2695 2696 case ARMISD::VLD4LN_UPD: { 2697 unsigned DOpcodes[] = { ARM::VLD4LNd8Pseudo_UPD, ARM::VLD4LNd16Pseudo_UPD, 2698 ARM::VLD4LNd32Pseudo_UPD }; 2699 unsigned QOpcodes[] = { ARM::VLD4LNq16Pseudo_UPD, 2700 ARM::VLD4LNq32Pseudo_UPD }; 2701 return SelectVLDSTLane(N, true, true, 4, DOpcodes, QOpcodes); 2702 } 2703 2704 case ARMISD::VST1_UPD: { 2705 unsigned DOpcodes[] = { ARM::VST1d8_UPD, ARM::VST1d16_UPD, 2706 ARM::VST1d32_UPD, ARM::VST1d64_UPD }; 2707 unsigned QOpcodes[] = { ARM::VST1q8Pseudo_UPD, ARM::VST1q16Pseudo_UPD, 2708 ARM::VST1q32Pseudo_UPD, ARM::VST1q64Pseudo_UPD }; 2709 return SelectVST(N, true, 1, DOpcodes, QOpcodes, 0); 2710 } 2711 2712 case ARMISD::VST2_UPD: { 2713 unsigned DOpcodes[] = { ARM::VST2d8Pseudo_UPD, ARM::VST2d16Pseudo_UPD, 2714 ARM::VST2d32Pseudo_UPD, ARM::VST1q64Pseudo_UPD }; 2715 unsigned QOpcodes[] = { ARM::VST2q8Pseudo_UPD, ARM::VST2q16Pseudo_UPD, 2716 ARM::VST2q32Pseudo_UPD }; 2717 return SelectVST(N, true, 2, DOpcodes, QOpcodes, 0); 2718 } 2719 2720 case ARMISD::VST3_UPD: { 2721 unsigned DOpcodes[] = { ARM::VST3d8Pseudo_UPD, ARM::VST3d16Pseudo_UPD, 2722 ARM::VST3d32Pseudo_UPD, ARM::VST1d64TPseudo_UPD }; 2723 unsigned QOpcodes0[] = { ARM::VST3q8Pseudo_UPD, 2724 ARM::VST3q16Pseudo_UPD, 2725 ARM::VST3q32Pseudo_UPD }; 2726 unsigned QOpcodes1[] = { ARM::VST3q8oddPseudo_UPD, 2727 ARM::VST3q16oddPseudo_UPD, 2728 ARM::VST3q32oddPseudo_UPD }; 2729 return SelectVST(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1); 2730 } 2731 2732 case ARMISD::VST4_UPD: { 2733 unsigned DOpcodes[] = { ARM::VST4d8Pseudo_UPD, ARM::VST4d16Pseudo_UPD, 2734 ARM::VST4d32Pseudo_UPD, ARM::VST1d64QPseudo_UPD }; 2735 unsigned QOpcodes0[] = { ARM::VST4q8Pseudo_UPD, 2736 ARM::VST4q16Pseudo_UPD, 2737 ARM::VST4q32Pseudo_UPD }; 2738 unsigned QOpcodes1[] = { ARM::VST4q8oddPseudo_UPD, 2739 ARM::VST4q16oddPseudo_UPD, 2740 ARM::VST4q32oddPseudo_UPD }; 2741 return SelectVST(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1); 2742 } 2743 2744 case ARMISD::VST2LN_UPD: { 2745 unsigned DOpcodes[] = { ARM::VST2LNd8Pseudo_UPD, ARM::VST2LNd16Pseudo_UPD, 2746 ARM::VST2LNd32Pseudo_UPD }; 2747 unsigned QOpcodes[] = { ARM::VST2LNq16Pseudo_UPD, 2748 ARM::VST2LNq32Pseudo_UPD }; 2749 return SelectVLDSTLane(N, false, true, 2, DOpcodes, QOpcodes); 2750 } 2751 2752 case ARMISD::VST3LN_UPD: { 2753 unsigned DOpcodes[] = { ARM::VST3LNd8Pseudo_UPD, ARM::VST3LNd16Pseudo_UPD, 2754 ARM::VST3LNd32Pseudo_UPD }; 2755 unsigned QOpcodes[] = { ARM::VST3LNq16Pseudo_UPD, 2756 ARM::VST3LNq32Pseudo_UPD }; 2757 return SelectVLDSTLane(N, false, true, 3, DOpcodes, QOpcodes); 2758 } 2759 2760 case ARMISD::VST4LN_UPD: { 2761 unsigned DOpcodes[] = { ARM::VST4LNd8Pseudo_UPD, ARM::VST4LNd16Pseudo_UPD, 2762 ARM::VST4LNd32Pseudo_UPD }; 2763 unsigned QOpcodes[] = { ARM::VST4LNq16Pseudo_UPD, 2764 ARM::VST4LNq32Pseudo_UPD }; 2765 return SelectVLDSTLane(N, false, true, 4, DOpcodes, QOpcodes); 2766 } 2767 2768 case ISD::INTRINSIC_VOID: 2769 case ISD::INTRINSIC_W_CHAIN: { 2770 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 2771 switch (IntNo) { 2772 default: 2773 break; 2774 2775 case Intrinsic::arm_ldrexd: { 2776 SDValue MemAddr = N->getOperand(2); 2777 DebugLoc dl = N->getDebugLoc(); 2778 SDValue Chain = N->getOperand(0); 2779 2780 unsigned NewOpc = ARM::LDREXD; 2781 if (Subtarget->isThumb() && Subtarget->hasThumb2()) 2782 NewOpc = ARM::t2LDREXD; 2783 2784 // arm_ldrexd returns a i64 value in {i32, i32} 2785 std::vector<EVT> ResTys; 2786 ResTys.push_back(MVT::i32); 2787 ResTys.push_back(MVT::i32); 2788 ResTys.push_back(MVT::Other); 2789 2790 // place arguments in the right order 2791 SmallVector<SDValue, 7> Ops; 2792 Ops.push_back(MemAddr); 2793 Ops.push_back(getAL(CurDAG)); 2794 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); 2795 Ops.push_back(Chain); 2796 SDNode *Ld = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops.data(), 2797 Ops.size()); 2798 // Transfer memoperands. 2799 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 2800 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 2801 cast<MachineSDNode>(Ld)->setMemRefs(MemOp, MemOp + 1); 2802 2803 // Until there's support for specifing explicit register constraints 2804 // like the use of even/odd register pair, hardcode ldrexd to always 2805 // use the pair [R0, R1] to hold the load result. 2806 Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ARM::R0, 2807 SDValue(Ld, 0), SDValue(0,0)); 2808 Chain = CurDAG->getCopyToReg(Chain, dl, ARM::R1, 2809 SDValue(Ld, 1), Chain.getValue(1)); 2810 2811 // Remap uses. 2812 SDValue Glue = Chain.getValue(1); 2813 if (!SDValue(N, 0).use_empty()) { 2814 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, 2815 ARM::R0, MVT::i32, Glue); 2816 Glue = Result.getValue(2); 2817 ReplaceUses(SDValue(N, 0), Result); 2818 } 2819 if (!SDValue(N, 1).use_empty()) { 2820 SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, 2821 ARM::R1, MVT::i32, Glue); 2822 Glue = Result.getValue(2); 2823 ReplaceUses(SDValue(N, 1), Result); 2824 } 2825 2826 ReplaceUses(SDValue(N, 2), SDValue(Ld, 2)); 2827 return NULL; 2828 } 2829 2830 case Intrinsic::arm_strexd: { 2831 DebugLoc dl = N->getDebugLoc(); 2832 SDValue Chain = N->getOperand(0); 2833 SDValue Val0 = N->getOperand(2); 2834 SDValue Val1 = N->getOperand(3); 2835 SDValue MemAddr = N->getOperand(4); 2836 2837 // Until there's support for specifing explicit register constraints 2838 // like the use of even/odd register pair, hardcode strexd to always 2839 // use the pair [R2, R3] to hold the i64 (i32, i32) value to be stored. 2840 Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, ARM::R2, Val0, 2841 SDValue(0, 0)); 2842 Chain = CurDAG->getCopyToReg(Chain, dl, ARM::R3, Val1, Chain.getValue(1)); 2843 2844 SDValue Glue = Chain.getValue(1); 2845 Val0 = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, 2846 ARM::R2, MVT::i32, Glue); 2847 Glue = Val0.getValue(1); 2848 Val1 = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl, 2849 ARM::R3, MVT::i32, Glue); 2850 2851 // Store exclusive double return a i32 value which is the return status 2852 // of the issued store. 2853 std::vector<EVT> ResTys; 2854 ResTys.push_back(MVT::i32); 2855 ResTys.push_back(MVT::Other); 2856 2857 // place arguments in the right order 2858 SmallVector<SDValue, 7> Ops; 2859 Ops.push_back(Val0); 2860 Ops.push_back(Val1); 2861 Ops.push_back(MemAddr); 2862 Ops.push_back(getAL(CurDAG)); 2863 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); 2864 Ops.push_back(Chain); 2865 2866 unsigned NewOpc = ARM::STREXD; 2867 if (Subtarget->isThumb() && Subtarget->hasThumb2()) 2868 NewOpc = ARM::t2STREXD; 2869 2870 SDNode *St = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops.data(), 2871 Ops.size()); 2872 // Transfer memoperands. 2873 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 2874 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 2875 cast<MachineSDNode>(St)->setMemRefs(MemOp, MemOp + 1); 2876 2877 return St; 2878 } 2879 2880 case Intrinsic::arm_neon_vld1: { 2881 unsigned DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16, 2882 ARM::VLD1d32, ARM::VLD1d64 }; 2883 unsigned QOpcodes[] = { ARM::VLD1q8Pseudo, ARM::VLD1q16Pseudo, 2884 ARM::VLD1q32Pseudo, ARM::VLD1q64Pseudo }; 2885 return SelectVLD(N, false, 1, DOpcodes, QOpcodes, 0); 2886 } 2887 2888 case Intrinsic::arm_neon_vld2: { 2889 unsigned DOpcodes[] = { ARM::VLD2d8Pseudo, ARM::VLD2d16Pseudo, 2890 ARM::VLD2d32Pseudo, ARM::VLD1q64Pseudo }; 2891 unsigned QOpcodes[] = { ARM::VLD2q8Pseudo, ARM::VLD2q16Pseudo, 2892 ARM::VLD2q32Pseudo }; 2893 return SelectVLD(N, false, 2, DOpcodes, QOpcodes, 0); 2894 } 2895 2896 case Intrinsic::arm_neon_vld3: { 2897 unsigned DOpcodes[] = { ARM::VLD3d8Pseudo, ARM::VLD3d16Pseudo, 2898 ARM::VLD3d32Pseudo, ARM::VLD1d64TPseudo }; 2899 unsigned QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD, 2900 ARM::VLD3q16Pseudo_UPD, 2901 ARM::VLD3q32Pseudo_UPD }; 2902 unsigned QOpcodes1[] = { ARM::VLD3q8oddPseudo, 2903 ARM::VLD3q16oddPseudo, 2904 ARM::VLD3q32oddPseudo }; 2905 return SelectVLD(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1); 2906 } 2907 2908 case Intrinsic::arm_neon_vld4: { 2909 unsigned DOpcodes[] = { ARM::VLD4d8Pseudo, ARM::VLD4d16Pseudo, 2910 ARM::VLD4d32Pseudo, ARM::VLD1d64QPseudo }; 2911 unsigned QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD, 2912 ARM::VLD4q16Pseudo_UPD, 2913 ARM::VLD4q32Pseudo_UPD }; 2914 unsigned QOpcodes1[] = { ARM::VLD4q8oddPseudo, 2915 ARM::VLD4q16oddPseudo, 2916 ARM::VLD4q32oddPseudo }; 2917 return SelectVLD(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1); 2918 } 2919 2920 case Intrinsic::arm_neon_vld2lane: { 2921 unsigned DOpcodes[] = { ARM::VLD2LNd8Pseudo, ARM::VLD2LNd16Pseudo, 2922 ARM::VLD2LNd32Pseudo }; 2923 unsigned QOpcodes[] = { ARM::VLD2LNq16Pseudo, ARM::VLD2LNq32Pseudo }; 2924 return SelectVLDSTLane(N, true, false, 2, DOpcodes, QOpcodes); 2925 } 2926 2927 case Intrinsic::arm_neon_vld3lane: { 2928 unsigned DOpcodes[] = { ARM::VLD3LNd8Pseudo, ARM::VLD3LNd16Pseudo, 2929 ARM::VLD3LNd32Pseudo }; 2930 unsigned QOpcodes[] = { ARM::VLD3LNq16Pseudo, ARM::VLD3LNq32Pseudo }; 2931 return SelectVLDSTLane(N, true, false, 3, DOpcodes, QOpcodes); 2932 } 2933 2934 case Intrinsic::arm_neon_vld4lane: { 2935 unsigned DOpcodes[] = { ARM::VLD4LNd8Pseudo, ARM::VLD4LNd16Pseudo, 2936 ARM::VLD4LNd32Pseudo }; 2937 unsigned QOpcodes[] = { ARM::VLD4LNq16Pseudo, ARM::VLD4LNq32Pseudo }; 2938 return SelectVLDSTLane(N, true, false, 4, DOpcodes, QOpcodes); 2939 } 2940 2941 case Intrinsic::arm_neon_vst1: { 2942 unsigned DOpcodes[] = { ARM::VST1d8, ARM::VST1d16, 2943 ARM::VST1d32, ARM::VST1d64 }; 2944 unsigned QOpcodes[] = { ARM::VST1q8Pseudo, ARM::VST1q16Pseudo, 2945 ARM::VST1q32Pseudo, ARM::VST1q64Pseudo }; 2946 return SelectVST(N, false, 1, DOpcodes, QOpcodes, 0); 2947 } 2948 2949 case Intrinsic::arm_neon_vst2: { 2950 unsigned DOpcodes[] = { ARM::VST2d8Pseudo, ARM::VST2d16Pseudo, 2951 ARM::VST2d32Pseudo, ARM::VST1q64Pseudo }; 2952 unsigned QOpcodes[] = { ARM::VST2q8Pseudo, ARM::VST2q16Pseudo, 2953 ARM::VST2q32Pseudo }; 2954 return SelectVST(N, false, 2, DOpcodes, QOpcodes, 0); 2955 } 2956 2957 case Intrinsic::arm_neon_vst3: { 2958 unsigned DOpcodes[] = { ARM::VST3d8Pseudo, ARM::VST3d16Pseudo, 2959 ARM::VST3d32Pseudo, ARM::VST1d64TPseudo }; 2960 unsigned QOpcodes0[] = { ARM::VST3q8Pseudo_UPD, 2961 ARM::VST3q16Pseudo_UPD, 2962 ARM::VST3q32Pseudo_UPD }; 2963 unsigned QOpcodes1[] = { ARM::VST3q8oddPseudo, 2964 ARM::VST3q16oddPseudo, 2965 ARM::VST3q32oddPseudo }; 2966 return SelectVST(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1); 2967 } 2968 2969 case Intrinsic::arm_neon_vst4: { 2970 unsigned DOpcodes[] = { ARM::VST4d8Pseudo, ARM::VST4d16Pseudo, 2971 ARM::VST4d32Pseudo, ARM::VST1d64QPseudo }; 2972 unsigned QOpcodes0[] = { ARM::VST4q8Pseudo_UPD, 2973 ARM::VST4q16Pseudo_UPD, 2974 ARM::VST4q32Pseudo_UPD }; 2975 unsigned QOpcodes1[] = { ARM::VST4q8oddPseudo, 2976 ARM::VST4q16oddPseudo, 2977 ARM::VST4q32oddPseudo }; 2978 return SelectVST(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1); 2979 } 2980 2981 case Intrinsic::arm_neon_vst2lane: { 2982 unsigned DOpcodes[] = { ARM::VST2LNd8Pseudo, ARM::VST2LNd16Pseudo, 2983 ARM::VST2LNd32Pseudo }; 2984 unsigned QOpcodes[] = { ARM::VST2LNq16Pseudo, ARM::VST2LNq32Pseudo }; 2985 return SelectVLDSTLane(N, false, false, 2, DOpcodes, QOpcodes); 2986 } 2987 2988 case Intrinsic::arm_neon_vst3lane: { 2989 unsigned DOpcodes[] = { ARM::VST3LNd8Pseudo, ARM::VST3LNd16Pseudo, 2990 ARM::VST3LNd32Pseudo }; 2991 unsigned QOpcodes[] = { ARM::VST3LNq16Pseudo, ARM::VST3LNq32Pseudo }; 2992 return SelectVLDSTLane(N, false, false, 3, DOpcodes, QOpcodes); 2993 } 2994 2995 case Intrinsic::arm_neon_vst4lane: { 2996 unsigned DOpcodes[] = { ARM::VST4LNd8Pseudo, ARM::VST4LNd16Pseudo, 2997 ARM::VST4LNd32Pseudo }; 2998 unsigned QOpcodes[] = { ARM::VST4LNq16Pseudo, ARM::VST4LNq32Pseudo }; 2999 return SelectVLDSTLane(N, false, false, 4, DOpcodes, QOpcodes); 3000 } 3001 } 3002 break; 3003 } 3004 3005 case ISD::INTRINSIC_WO_CHAIN: { 3006 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); 3007 switch (IntNo) { 3008 default: 3009 break; 3010 3011 case Intrinsic::arm_neon_vtbl2: 3012 return SelectVTBL(N, false, 2, ARM::VTBL2Pseudo); 3013 case Intrinsic::arm_neon_vtbl3: 3014 return SelectVTBL(N, false, 3, ARM::VTBL3Pseudo); 3015 case Intrinsic::arm_neon_vtbl4: 3016 return SelectVTBL(N, false, 4, ARM::VTBL4Pseudo); 3017 3018 case Intrinsic::arm_neon_vtbx2: 3019 return SelectVTBL(N, true, 2, ARM::VTBX2Pseudo); 3020 case Intrinsic::arm_neon_vtbx3: 3021 return SelectVTBL(N, true, 3, ARM::VTBX3Pseudo); 3022 case Intrinsic::arm_neon_vtbx4: 3023 return SelectVTBL(N, true, 4, ARM::VTBX4Pseudo); 3024 } 3025 break; 3026 } 3027 3028 case ARMISD::VTBL1: { 3029 DebugLoc dl = N->getDebugLoc(); 3030 EVT VT = N->getValueType(0); 3031 SmallVector<SDValue, 6> Ops; 3032 3033 Ops.push_back(N->getOperand(0)); 3034 Ops.push_back(N->getOperand(1)); 3035 Ops.push_back(getAL(CurDAG)); // Predicate 3036 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // Predicate Register 3037 return CurDAG->getMachineNode(ARM::VTBL1, dl, VT, Ops.data(), Ops.size()); 3038 } 3039 case ARMISD::VTBL2: { 3040 DebugLoc dl = N->getDebugLoc(); 3041 EVT VT = N->getValueType(0); 3042 3043 // Form a REG_SEQUENCE to force register allocation. 3044 SDValue V0 = N->getOperand(0); 3045 SDValue V1 = N->getOperand(1); 3046 SDValue RegSeq = SDValue(PairDRegs(MVT::v16i8, V0, V1), 0); 3047 3048 SmallVector<SDValue, 6> Ops; 3049 Ops.push_back(RegSeq); 3050 Ops.push_back(N->getOperand(2)); 3051 Ops.push_back(getAL(CurDAG)); // Predicate 3052 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // Predicate Register 3053 return CurDAG->getMachineNode(ARM::VTBL2Pseudo, dl, VT, 3054 Ops.data(), Ops.size()); 3055 } 3056 3057 case ISD::CONCAT_VECTORS: 3058 return SelectConcatVector(N); 3059 } 3060 3061 return SelectCode(N); 3062} 3063 3064bool ARMDAGToDAGISel:: 3065SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode, 3066 std::vector<SDValue> &OutOps) { 3067 assert(ConstraintCode == 'm' && "unexpected asm memory constraint"); 3068 // Require the address to be in a register. That is safe for all ARM 3069 // variants and it is hard to do anything much smarter without knowing 3070 // how the operand is used. 3071 OutOps.push_back(Op); 3072 return false; 3073} 3074 3075/// createARMISelDag - This pass converts a legalized DAG into a 3076/// ARM-specific DAG, ready for instruction scheduling. 3077/// 3078FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM, 3079 CodeGenOpt::Level OptLevel) { 3080 return new ARMDAGToDAGISel(TM, OptLevel); 3081} 3082