ARMISelDAGToDAG.cpp revision 263508
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/CodeGen/MachineFrameInfo.h" 20#include "llvm/CodeGen/MachineFunction.h" 21#include "llvm/CodeGen/MachineInstrBuilder.h" 22#include "llvm/CodeGen/MachineRegisterInfo.h" 23#include "llvm/CodeGen/SelectionDAG.h" 24#include "llvm/CodeGen/SelectionDAGISel.h" 25#include "llvm/IR/CallingConv.h" 26#include "llvm/IR/Constants.h" 27#include "llvm/IR/DerivedTypes.h" 28#include "llvm/IR/Function.h" 29#include "llvm/IR/Intrinsics.h" 30#include "llvm/IR/LLVMContext.h" 31#include "llvm/Support/CommandLine.h" 32#include "llvm/Support/Compiler.h" 33#include "llvm/Support/Debug.h" 34#include "llvm/Support/ErrorHandling.h" 35#include "llvm/Support/raw_ostream.h" 36#include "llvm/Target/TargetLowering.h" 37#include "llvm/Target/TargetOptions.h" 38 39using namespace llvm; 40 41static cl::opt<bool> 42DisableShifterOp("disable-shifter-op", cl::Hidden, 43 cl::desc("Disable isel of shifter-op"), 44 cl::init(false)); 45 46static cl::opt<bool> 47CheckVMLxHazard("check-vmlx-hazard", cl::Hidden, 48 cl::desc("Check fp vmla / vmls hazard at isel time"), 49 cl::init(true)); 50 51//===--------------------------------------------------------------------===// 52/// ARMDAGToDAGISel - ARM specific code to select ARM machine 53/// instructions for SelectionDAG operations. 54/// 55namespace { 56 57enum AddrMode2Type { 58 AM2_BASE, // Simple AM2 (+-imm12) 59 AM2_SHOP // Shifter-op AM2 60}; 61 62class ARMDAGToDAGISel : public SelectionDAGISel { 63 ARMBaseTargetMachine &TM; 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 Subtarget(&TM.getSubtarget<ARMSubtarget>()) { 74 } 75 76 virtual const char *getPassName() const { 77 return "ARM Instruction Selection"; 78 } 79 80 virtual void PreprocessISelDAG(); 81 82 /// getI32Imm - Return a target constant of type i32 with the specified 83 /// value. 84 inline SDValue getI32Imm(unsigned Imm) { 85 return CurDAG->getTargetConstant(Imm, MVT::i32); 86 } 87 88 SDNode *Select(SDNode *N); 89 90 91 bool hasNoVMLxHazardUse(SDNode *N) const; 92 bool isShifterOpProfitable(const SDValue &Shift, 93 ARM_AM::ShiftOpc ShOpcVal, unsigned ShAmt); 94 bool SelectRegShifterOperand(SDValue N, SDValue &A, 95 SDValue &B, SDValue &C, 96 bool CheckProfitability = true); 97 bool SelectImmShifterOperand(SDValue N, SDValue &A, 98 SDValue &B, bool CheckProfitability = true); 99 bool SelectShiftRegShifterOperand(SDValue N, SDValue &A, 100 SDValue &B, SDValue &C) { 101 // Don't apply the profitability check 102 return SelectRegShifterOperand(N, A, B, C, false); 103 } 104 bool SelectShiftImmShifterOperand(SDValue N, SDValue &A, 105 SDValue &B) { 106 // Don't apply the profitability check 107 return SelectImmShifterOperand(N, A, B, false); 108 } 109 110 bool SelectAddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm); 111 bool SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset, SDValue &Opc); 112 113 AddrMode2Type SelectAddrMode2Worker(SDValue N, SDValue &Base, 114 SDValue &Offset, SDValue &Opc); 115 bool SelectAddrMode2Base(SDValue N, SDValue &Base, SDValue &Offset, 116 SDValue &Opc) { 117 return SelectAddrMode2Worker(N, Base, Offset, Opc) == AM2_BASE; 118 } 119 120 bool SelectAddrMode2ShOp(SDValue N, SDValue &Base, SDValue &Offset, 121 SDValue &Opc) { 122 return SelectAddrMode2Worker(N, Base, Offset, Opc) == AM2_SHOP; 123 } 124 125 bool SelectAddrMode2(SDValue N, SDValue &Base, SDValue &Offset, 126 SDValue &Opc) { 127 SelectAddrMode2Worker(N, Base, Offset, Opc); 128// return SelectAddrMode2ShOp(N, Base, Offset, Opc); 129 // This always matches one way or another. 130 return true; 131 } 132 133 bool SelectCMOVPred(SDValue N, SDValue &Pred, SDValue &Reg) { 134 const ConstantSDNode *CN = cast<ConstantSDNode>(N); 135 Pred = CurDAG->getTargetConstant(CN->getZExtValue(), MVT::i32); 136 Reg = CurDAG->getRegister(ARM::CPSR, MVT::i32); 137 return true; 138 } 139 140 bool SelectAddrMode2OffsetReg(SDNode *Op, SDValue N, 141 SDValue &Offset, SDValue &Opc); 142 bool SelectAddrMode2OffsetImm(SDNode *Op, SDValue N, 143 SDValue &Offset, SDValue &Opc); 144 bool SelectAddrMode2OffsetImmPre(SDNode *Op, SDValue N, 145 SDValue &Offset, SDValue &Opc); 146 bool SelectAddrOffsetNone(SDValue N, SDValue &Base); 147 bool SelectAddrMode3(SDValue N, SDValue &Base, 148 SDValue &Offset, SDValue &Opc); 149 bool SelectAddrMode3Offset(SDNode *Op, SDValue N, 150 SDValue &Offset, SDValue &Opc); 151 bool SelectAddrMode5(SDValue N, SDValue &Base, 152 SDValue &Offset); 153 bool SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,SDValue &Align); 154 bool SelectAddrMode6Offset(SDNode *Op, SDValue N, SDValue &Offset); 155 156 bool SelectAddrModePC(SDValue N, SDValue &Offset, SDValue &Label); 157 158 // Thumb Addressing Modes: 159 bool SelectThumbAddrModeRR(SDValue N, SDValue &Base, SDValue &Offset); 160 bool SelectThumbAddrModeRI(SDValue N, SDValue &Base, SDValue &Offset, 161 unsigned Scale); 162 bool SelectThumbAddrModeRI5S1(SDValue N, SDValue &Base, SDValue &Offset); 163 bool SelectThumbAddrModeRI5S2(SDValue N, SDValue &Base, SDValue &Offset); 164 bool SelectThumbAddrModeRI5S4(SDValue N, SDValue &Base, SDValue &Offset); 165 bool SelectThumbAddrModeImm5S(SDValue N, unsigned Scale, SDValue &Base, 166 SDValue &OffImm); 167 bool SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base, 168 SDValue &OffImm); 169 bool SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base, 170 SDValue &OffImm); 171 bool SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base, 172 SDValue &OffImm); 173 bool SelectThumbAddrModeSP(SDValue N, SDValue &Base, SDValue &OffImm); 174 175 // Thumb 2 Addressing Modes: 176 bool SelectT2ShifterOperandReg(SDValue N, 177 SDValue &BaseReg, SDValue &Opc); 178 bool SelectT2AddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm); 179 bool SelectT2AddrModeImm8(SDValue N, SDValue &Base, 180 SDValue &OffImm); 181 bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N, 182 SDValue &OffImm); 183 bool SelectT2AddrModeSoReg(SDValue N, SDValue &Base, 184 SDValue &OffReg, SDValue &ShImm); 185 bool SelectT2AddrModeExclusive(SDValue N, SDValue &Base, SDValue &OffImm); 186 187 inline bool is_so_imm(unsigned Imm) const { 188 return ARM_AM::getSOImmVal(Imm) != -1; 189 } 190 191 inline bool is_so_imm_not(unsigned Imm) const { 192 return ARM_AM::getSOImmVal(~Imm) != -1; 193 } 194 195 inline bool is_t2_so_imm(unsigned Imm) const { 196 return ARM_AM::getT2SOImmVal(Imm) != -1; 197 } 198 199 inline bool is_t2_so_imm_not(unsigned Imm) const { 200 return ARM_AM::getT2SOImmVal(~Imm) != -1; 201 } 202 203 // Include the pieces autogenerated from the target description. 204#include "ARMGenDAGISel.inc" 205 206private: 207 /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for 208 /// ARM. 209 SDNode *SelectARMIndexedLoad(SDNode *N); 210 SDNode *SelectT2IndexedLoad(SDNode *N); 211 212 /// SelectVLD - Select NEON load intrinsics. NumVecs should be 213 /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for 214 /// loads of D registers and even subregs and odd subregs of Q registers. 215 /// For NumVecs <= 2, QOpcodes1 is not used. 216 SDNode *SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs, 217 const uint16_t *DOpcodes, 218 const uint16_t *QOpcodes0, const uint16_t *QOpcodes1); 219 220 /// SelectVST - Select NEON store intrinsics. NumVecs should 221 /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for 222 /// stores of D registers and even subregs and odd subregs of Q registers. 223 /// For NumVecs <= 2, QOpcodes1 is not used. 224 SDNode *SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs, 225 const uint16_t *DOpcodes, 226 const uint16_t *QOpcodes0, const uint16_t *QOpcodes1); 227 228 /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should 229 /// be 2, 3 or 4. The opcode arrays specify the instructions used for 230 /// load/store of D registers and Q registers. 231 SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, 232 bool isUpdating, unsigned NumVecs, 233 const uint16_t *DOpcodes, const uint16_t *QOpcodes); 234 235 /// SelectVLDDup - Select NEON load-duplicate intrinsics. NumVecs 236 /// should be 2, 3 or 4. The opcode array specifies the instructions used 237 /// for loading D registers. (Q registers are not supported.) 238 SDNode *SelectVLDDup(SDNode *N, bool isUpdating, unsigned NumVecs, 239 const uint16_t *Opcodes); 240 241 /// SelectVTBL - Select NEON VTBL and VTBX intrinsics. NumVecs should be 2, 242 /// 3 or 4. These are custom-selected so that a REG_SEQUENCE can be 243 /// generated to force the table registers to be consecutive. 244 SDNode *SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs, unsigned Opc); 245 246 /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM. 247 SDNode *SelectV6T2BitfieldExtractOp(SDNode *N, bool isSigned); 248 249 // Select special operations if node forms integer ABS pattern 250 SDNode *SelectABSOp(SDNode *N); 251 252 SDNode *SelectInlineAsm(SDNode *N); 253 254 SDNode *SelectConcatVector(SDNode *N); 255 256 SDNode *SelectAtomic(SDNode *N, unsigned Op8, unsigned Op16, unsigned Op32, unsigned Op64); 257 258 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for 259 /// inline asm expressions. 260 virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op, 261 char ConstraintCode, 262 std::vector<SDValue> &OutOps); 263 264 // Form pairs of consecutive R, S, D, or Q registers. 265 SDNode *createGPRPairNode(EVT VT, SDValue V0, SDValue V1); 266 SDNode *createSRegPairNode(EVT VT, SDValue V0, SDValue V1); 267 SDNode *createDRegPairNode(EVT VT, SDValue V0, SDValue V1); 268 SDNode *createQRegPairNode(EVT VT, SDValue V0, SDValue V1); 269 270 // Form sequences of 4 consecutive S, D, or Q registers. 271 SDNode *createQuadSRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3); 272 SDNode *createQuadDRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3); 273 SDNode *createQuadQRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3); 274 275 // Get the alignment operand for a NEON VLD or VST instruction. 276 SDValue GetVLDSTAlign(SDValue Align, unsigned NumVecs, bool is64BitVector); 277}; 278} 279 280/// isInt32Immediate - This method tests to see if the node is a 32-bit constant 281/// operand. If so Imm will receive the 32-bit value. 282static bool isInt32Immediate(SDNode *N, unsigned &Imm) { 283 if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) { 284 Imm = cast<ConstantSDNode>(N)->getZExtValue(); 285 return true; 286 } 287 return false; 288} 289 290// isInt32Immediate - This method tests to see if a constant operand. 291// If so Imm will receive the 32 bit value. 292static bool isInt32Immediate(SDValue N, unsigned &Imm) { 293 return isInt32Immediate(N.getNode(), Imm); 294} 295 296// isOpcWithIntImmediate - This method tests to see if the node is a specific 297// opcode and that it has a immediate integer right operand. 298// If so Imm will receive the 32 bit value. 299static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) { 300 return N->getOpcode() == Opc && 301 isInt32Immediate(N->getOperand(1).getNode(), Imm); 302} 303 304/// \brief Check whether a particular node is a constant value representable as 305/// (N * Scale) where (N in [\p RangeMin, \p RangeMax). 306/// 307/// \param ScaledConstant [out] - On success, the pre-scaled constant value. 308static bool isScaledConstantInRange(SDValue Node, int Scale, 309 int RangeMin, int RangeMax, 310 int &ScaledConstant) { 311 assert(Scale > 0 && "Invalid scale!"); 312 313 // Check that this is a constant. 314 const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Node); 315 if (!C) 316 return false; 317 318 ScaledConstant = (int) C->getZExtValue(); 319 if ((ScaledConstant % Scale) != 0) 320 return false; 321 322 ScaledConstant /= Scale; 323 return ScaledConstant >= RangeMin && ScaledConstant < RangeMax; 324} 325 326void ARMDAGToDAGISel::PreprocessISelDAG() { 327 if (!Subtarget->hasV6T2Ops()) 328 return; 329 330 bool isThumb2 = Subtarget->isThumb(); 331 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(), 332 E = CurDAG->allnodes_end(); I != E; ) { 333 SDNode *N = I++; // Preincrement iterator to avoid invalidation issues. 334 335 if (N->getOpcode() != ISD::ADD) 336 continue; 337 338 // Look for (add X1, (and (srl X2, c1), c2)) where c2 is constant with 339 // leading zeros, followed by consecutive set bits, followed by 1 or 2 340 // trailing zeros, e.g. 1020. 341 // Transform the expression to 342 // (add X1, (shl (and (srl X2, c1), (c2>>tz)), tz)) where tz is the number 343 // of trailing zeros of c2. The left shift would be folded as an shifter 344 // operand of 'add' and the 'and' and 'srl' would become a bits extraction 345 // node (UBFX). 346 347 SDValue N0 = N->getOperand(0); 348 SDValue N1 = N->getOperand(1); 349 unsigned And_imm = 0; 350 if (!isOpcWithIntImmediate(N1.getNode(), ISD::AND, And_imm)) { 351 if (isOpcWithIntImmediate(N0.getNode(), ISD::AND, And_imm)) 352 std::swap(N0, N1); 353 } 354 if (!And_imm) 355 continue; 356 357 // Check if the AND mask is an immediate of the form: 000.....1111111100 358 unsigned TZ = countTrailingZeros(And_imm); 359 if (TZ != 1 && TZ != 2) 360 // Be conservative here. Shifter operands aren't always free. e.g. On 361 // Swift, left shifter operand of 1 / 2 for free but others are not. 362 // e.g. 363 // ubfx r3, r1, #16, #8 364 // ldr.w r3, [r0, r3, lsl #2] 365 // vs. 366 // mov.w r9, #1020 367 // and.w r2, r9, r1, lsr #14 368 // ldr r2, [r0, r2] 369 continue; 370 And_imm >>= TZ; 371 if (And_imm & (And_imm + 1)) 372 continue; 373 374 // Look for (and (srl X, c1), c2). 375 SDValue Srl = N1.getOperand(0); 376 unsigned Srl_imm = 0; 377 if (!isOpcWithIntImmediate(Srl.getNode(), ISD::SRL, Srl_imm) || 378 (Srl_imm <= 2)) 379 continue; 380 381 // Make sure first operand is not a shifter operand which would prevent 382 // folding of the left shift. 383 SDValue CPTmp0; 384 SDValue CPTmp1; 385 SDValue CPTmp2; 386 if (isThumb2) { 387 if (SelectT2ShifterOperandReg(N0, CPTmp0, CPTmp1)) 388 continue; 389 } else { 390 if (SelectImmShifterOperand(N0, CPTmp0, CPTmp1) || 391 SelectRegShifterOperand(N0, CPTmp0, CPTmp1, CPTmp2)) 392 continue; 393 } 394 395 // Now make the transformation. 396 Srl = CurDAG->getNode(ISD::SRL, SDLoc(Srl), MVT::i32, 397 Srl.getOperand(0), 398 CurDAG->getConstant(Srl_imm+TZ, MVT::i32)); 399 N1 = CurDAG->getNode(ISD::AND, SDLoc(N1), MVT::i32, 400 Srl, CurDAG->getConstant(And_imm, MVT::i32)); 401 N1 = CurDAG->getNode(ISD::SHL, SDLoc(N1), MVT::i32, 402 N1, CurDAG->getConstant(TZ, MVT::i32)); 403 CurDAG->UpdateNodeOperands(N, N0, N1); 404 } 405} 406 407/// hasNoVMLxHazardUse - Return true if it's desirable to select a FP MLA / MLS 408/// node. VFP / NEON fp VMLA / VMLS instructions have special RAW hazards (at 409/// least on current ARM implementations) which should be avoidded. 410bool ARMDAGToDAGISel::hasNoVMLxHazardUse(SDNode *N) const { 411 if (OptLevel == CodeGenOpt::None) 412 return true; 413 414 if (!CheckVMLxHazard) 415 return true; 416 417 if (!Subtarget->isCortexA8() && !Subtarget->isCortexA9() && 418 !Subtarget->isSwift()) 419 return true; 420 421 if (!N->hasOneUse()) 422 return false; 423 424 SDNode *Use = *N->use_begin(); 425 if (Use->getOpcode() == ISD::CopyToReg) 426 return true; 427 if (Use->isMachineOpcode()) { 428 const ARMBaseInstrInfo *TII = 429 static_cast<const ARMBaseInstrInfo*>(TM.getInstrInfo()); 430 431 const MCInstrDesc &MCID = TII->get(Use->getMachineOpcode()); 432 if (MCID.mayStore()) 433 return true; 434 unsigned Opcode = MCID.getOpcode(); 435 if (Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD) 436 return true; 437 // vmlx feeding into another vmlx. We actually want to unfold 438 // the use later in the MLxExpansion pass. e.g. 439 // vmla 440 // vmla (stall 8 cycles) 441 // 442 // vmul (5 cycles) 443 // vadd (5 cycles) 444 // vmla 445 // This adds up to about 18 - 19 cycles. 446 // 447 // vmla 448 // vmul (stall 4 cycles) 449 // vadd adds up to about 14 cycles. 450 return TII->isFpMLxInstruction(Opcode); 451 } 452 453 return false; 454} 455 456bool ARMDAGToDAGISel::isShifterOpProfitable(const SDValue &Shift, 457 ARM_AM::ShiftOpc ShOpcVal, 458 unsigned ShAmt) { 459 if (!Subtarget->isLikeA9() && !Subtarget->isSwift()) 460 return true; 461 if (Shift.hasOneUse()) 462 return true; 463 // R << 2 is free. 464 return ShOpcVal == ARM_AM::lsl && 465 (ShAmt == 2 || (Subtarget->isSwift() && ShAmt == 1)); 466} 467 468bool ARMDAGToDAGISel::SelectImmShifterOperand(SDValue N, 469 SDValue &BaseReg, 470 SDValue &Opc, 471 bool CheckProfitability) { 472 if (DisableShifterOp) 473 return false; 474 475 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode()); 476 477 // Don't match base register only case. That is matched to a separate 478 // lower complexity pattern with explicit register operand. 479 if (ShOpcVal == ARM_AM::no_shift) return false; 480 481 BaseReg = N.getOperand(0); 482 unsigned ShImmVal = 0; 483 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1)); 484 if (!RHS) return false; 485 ShImmVal = RHS->getZExtValue() & 31; 486 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal), 487 MVT::i32); 488 return true; 489} 490 491bool ARMDAGToDAGISel::SelectRegShifterOperand(SDValue N, 492 SDValue &BaseReg, 493 SDValue &ShReg, 494 SDValue &Opc, 495 bool CheckProfitability) { 496 if (DisableShifterOp) 497 return false; 498 499 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode()); 500 501 // Don't match base register only case. That is matched to a separate 502 // lower complexity pattern with explicit register operand. 503 if (ShOpcVal == ARM_AM::no_shift) return false; 504 505 BaseReg = N.getOperand(0); 506 unsigned ShImmVal = 0; 507 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1)); 508 if (RHS) return false; 509 510 ShReg = N.getOperand(1); 511 if (CheckProfitability && !isShifterOpProfitable(N, ShOpcVal, ShImmVal)) 512 return false; 513 Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal), 514 MVT::i32); 515 return true; 516} 517 518 519bool ARMDAGToDAGISel::SelectAddrModeImm12(SDValue N, 520 SDValue &Base, 521 SDValue &OffImm) { 522 // Match simple R + imm12 operands. 523 524 // Base only. 525 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && 526 !CurDAG->isBaseWithConstantOffset(N)) { 527 if (N.getOpcode() == ISD::FrameIndex) { 528 // Match frame index. 529 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 530 Base = CurDAG->getTargetFrameIndex(FI, 531 getTargetLowering()->getPointerTy()); 532 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 533 return true; 534 } 535 536 if (N.getOpcode() == ARMISD::Wrapper && 537 !(Subtarget->useMovt() && 538 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) { 539 Base = N.getOperand(0); 540 } else 541 Base = N; 542 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 543 return true; 544 } 545 546 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 547 int RHSC = (int)RHS->getZExtValue(); 548 if (N.getOpcode() == ISD::SUB) 549 RHSC = -RHSC; 550 551 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned) 552 Base = N.getOperand(0); 553 if (Base.getOpcode() == ISD::FrameIndex) { 554 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 555 Base = CurDAG->getTargetFrameIndex(FI, 556 getTargetLowering()->getPointerTy()); 557 } 558 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); 559 return true; 560 } 561 } 562 563 // Base only. 564 Base = N; 565 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 566 return true; 567} 568 569 570 571bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset, 572 SDValue &Opc) { 573 if (N.getOpcode() == ISD::MUL && 574 ((!Subtarget->isLikeA9() && !Subtarget->isSwift()) || N.hasOneUse())) { 575 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 576 // X * [3,5,9] -> X + X * [2,4,8] etc. 577 int RHSC = (int)RHS->getZExtValue(); 578 if (RHSC & 1) { 579 RHSC = RHSC & ~1; 580 ARM_AM::AddrOpc AddSub = ARM_AM::add; 581 if (RHSC < 0) { 582 AddSub = ARM_AM::sub; 583 RHSC = - RHSC; 584 } 585 if (isPowerOf2_32(RHSC)) { 586 unsigned ShAmt = Log2_32(RHSC); 587 Base = Offset = N.getOperand(0); 588 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, 589 ARM_AM::lsl), 590 MVT::i32); 591 return true; 592 } 593 } 594 } 595 } 596 597 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && 598 // ISD::OR that is equivalent to an ISD::ADD. 599 !CurDAG->isBaseWithConstantOffset(N)) 600 return false; 601 602 // Leave simple R +/- imm12 operands for LDRi12 603 if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::OR) { 604 int RHSC; 605 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1, 606 -0x1000+1, 0x1000, RHSC)) // 12 bits. 607 return false; 608 } 609 610 // Otherwise this is R +/- [possibly shifted] R. 611 ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::SUB ? ARM_AM::sub:ARM_AM::add; 612 ARM_AM::ShiftOpc ShOpcVal = 613 ARM_AM::getShiftOpcForNode(N.getOperand(1).getOpcode()); 614 unsigned ShAmt = 0; 615 616 Base = N.getOperand(0); 617 Offset = N.getOperand(1); 618 619 if (ShOpcVal != ARM_AM::no_shift) { 620 // Check to see if the RHS of the shift is a constant, if not, we can't fold 621 // it. 622 if (ConstantSDNode *Sh = 623 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) { 624 ShAmt = Sh->getZExtValue(); 625 if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt)) 626 Offset = N.getOperand(1).getOperand(0); 627 else { 628 ShAmt = 0; 629 ShOpcVal = ARM_AM::no_shift; 630 } 631 } else { 632 ShOpcVal = ARM_AM::no_shift; 633 } 634 } 635 636 // Try matching (R shl C) + (R). 637 if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift && 638 !(Subtarget->isLikeA9() || Subtarget->isSwift() || 639 N.getOperand(0).hasOneUse())) { 640 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0).getOpcode()); 641 if (ShOpcVal != ARM_AM::no_shift) { 642 // Check to see if the RHS of the shift is a constant, if not, we can't 643 // fold it. 644 if (ConstantSDNode *Sh = 645 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) { 646 ShAmt = Sh->getZExtValue(); 647 if (isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt)) { 648 Offset = N.getOperand(0).getOperand(0); 649 Base = N.getOperand(1); 650 } else { 651 ShAmt = 0; 652 ShOpcVal = ARM_AM::no_shift; 653 } 654 } else { 655 ShOpcVal = ARM_AM::no_shift; 656 } 657 } 658 } 659 660 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal), 661 MVT::i32); 662 return true; 663} 664 665 666//----- 667 668AddrMode2Type ARMDAGToDAGISel::SelectAddrMode2Worker(SDValue N, 669 SDValue &Base, 670 SDValue &Offset, 671 SDValue &Opc) { 672 if (N.getOpcode() == ISD::MUL && 673 (!(Subtarget->isLikeA9() || Subtarget->isSwift()) || N.hasOneUse())) { 674 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 675 // X * [3,5,9] -> X + X * [2,4,8] etc. 676 int RHSC = (int)RHS->getZExtValue(); 677 if (RHSC & 1) { 678 RHSC = RHSC & ~1; 679 ARM_AM::AddrOpc AddSub = ARM_AM::add; 680 if (RHSC < 0) { 681 AddSub = ARM_AM::sub; 682 RHSC = - RHSC; 683 } 684 if (isPowerOf2_32(RHSC)) { 685 unsigned ShAmt = Log2_32(RHSC); 686 Base = Offset = N.getOperand(0); 687 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, 688 ARM_AM::lsl), 689 MVT::i32); 690 return AM2_SHOP; 691 } 692 } 693 } 694 } 695 696 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && 697 // ISD::OR that is equivalent to an ADD. 698 !CurDAG->isBaseWithConstantOffset(N)) { 699 Base = N; 700 if (N.getOpcode() == ISD::FrameIndex) { 701 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 702 Base = CurDAG->getTargetFrameIndex(FI, 703 getTargetLowering()->getPointerTy()); 704 } else if (N.getOpcode() == ARMISD::Wrapper && 705 !(Subtarget->useMovt() && 706 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) { 707 Base = N.getOperand(0); 708 } 709 Offset = CurDAG->getRegister(0, MVT::i32); 710 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0, 711 ARM_AM::no_shift), 712 MVT::i32); 713 return AM2_BASE; 714 } 715 716 // Match simple R +/- imm12 operands. 717 if (N.getOpcode() != ISD::SUB) { 718 int RHSC; 719 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1, 720 -0x1000+1, 0x1000, RHSC)) { // 12 bits. 721 Base = N.getOperand(0); 722 if (Base.getOpcode() == ISD::FrameIndex) { 723 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 724 Base = CurDAG->getTargetFrameIndex(FI, 725 getTargetLowering()->getPointerTy()); 726 } 727 Offset = CurDAG->getRegister(0, MVT::i32); 728 729 ARM_AM::AddrOpc AddSub = ARM_AM::add; 730 if (RHSC < 0) { 731 AddSub = ARM_AM::sub; 732 RHSC = - RHSC; 733 } 734 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC, 735 ARM_AM::no_shift), 736 MVT::i32); 737 return AM2_BASE; 738 } 739 } 740 741 if ((Subtarget->isLikeA9() || Subtarget->isSwift()) && !N.hasOneUse()) { 742 // Compute R +/- (R << N) and reuse it. 743 Base = N; 744 Offset = CurDAG->getRegister(0, MVT::i32); 745 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0, 746 ARM_AM::no_shift), 747 MVT::i32); 748 return AM2_BASE; 749 } 750 751 // Otherwise this is R +/- [possibly shifted] R. 752 ARM_AM::AddrOpc AddSub = N.getOpcode() != ISD::SUB ? ARM_AM::add:ARM_AM::sub; 753 ARM_AM::ShiftOpc ShOpcVal = 754 ARM_AM::getShiftOpcForNode(N.getOperand(1).getOpcode()); 755 unsigned ShAmt = 0; 756 757 Base = N.getOperand(0); 758 Offset = N.getOperand(1); 759 760 if (ShOpcVal != ARM_AM::no_shift) { 761 // Check to see if the RHS of the shift is a constant, if not, we can't fold 762 // it. 763 if (ConstantSDNode *Sh = 764 dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) { 765 ShAmt = Sh->getZExtValue(); 766 if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt)) 767 Offset = N.getOperand(1).getOperand(0); 768 else { 769 ShAmt = 0; 770 ShOpcVal = ARM_AM::no_shift; 771 } 772 } else { 773 ShOpcVal = ARM_AM::no_shift; 774 } 775 } 776 777 // Try matching (R shl C) + (R). 778 if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift && 779 !(Subtarget->isLikeA9() || Subtarget->isSwift() || 780 N.getOperand(0).hasOneUse())) { 781 ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0).getOpcode()); 782 if (ShOpcVal != ARM_AM::no_shift) { 783 // Check to see if the RHS of the shift is a constant, if not, we can't 784 // fold it. 785 if (ConstantSDNode *Sh = 786 dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) { 787 ShAmt = Sh->getZExtValue(); 788 if (isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt)) { 789 Offset = N.getOperand(0).getOperand(0); 790 Base = N.getOperand(1); 791 } else { 792 ShAmt = 0; 793 ShOpcVal = ARM_AM::no_shift; 794 } 795 } else { 796 ShOpcVal = ARM_AM::no_shift; 797 } 798 } 799 } 800 801 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal), 802 MVT::i32); 803 return AM2_SHOP; 804} 805 806bool ARMDAGToDAGISel::SelectAddrMode2OffsetReg(SDNode *Op, SDValue N, 807 SDValue &Offset, SDValue &Opc) { 808 unsigned Opcode = Op->getOpcode(); 809 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) 810 ? cast<LoadSDNode>(Op)->getAddressingMode() 811 : cast<StoreSDNode>(Op)->getAddressingMode(); 812 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC) 813 ? ARM_AM::add : ARM_AM::sub; 814 int Val; 815 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val)) 816 return false; 817 818 Offset = N; 819 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode()); 820 unsigned ShAmt = 0; 821 if (ShOpcVal != ARM_AM::no_shift) { 822 // Check to see if the RHS of the shift is a constant, if not, we can't fold 823 // it. 824 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 825 ShAmt = Sh->getZExtValue(); 826 if (isShifterOpProfitable(N, ShOpcVal, ShAmt)) 827 Offset = N.getOperand(0); 828 else { 829 ShAmt = 0; 830 ShOpcVal = ARM_AM::no_shift; 831 } 832 } else { 833 ShOpcVal = ARM_AM::no_shift; 834 } 835 } 836 837 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal), 838 MVT::i32); 839 return true; 840} 841 842bool ARMDAGToDAGISel::SelectAddrMode2OffsetImmPre(SDNode *Op, SDValue N, 843 SDValue &Offset, SDValue &Opc) { 844 unsigned Opcode = Op->getOpcode(); 845 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) 846 ? cast<LoadSDNode>(Op)->getAddressingMode() 847 : cast<StoreSDNode>(Op)->getAddressingMode(); 848 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC) 849 ? ARM_AM::add : ARM_AM::sub; 850 int Val; 851 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val)) { // 12 bits. 852 if (AddSub == ARM_AM::sub) Val *= -1; 853 Offset = CurDAG->getRegister(0, MVT::i32); 854 Opc = CurDAG->getTargetConstant(Val, MVT::i32); 855 return true; 856 } 857 858 return false; 859} 860 861 862bool ARMDAGToDAGISel::SelectAddrMode2OffsetImm(SDNode *Op, SDValue N, 863 SDValue &Offset, SDValue &Opc) { 864 unsigned Opcode = Op->getOpcode(); 865 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) 866 ? cast<LoadSDNode>(Op)->getAddressingMode() 867 : cast<StoreSDNode>(Op)->getAddressingMode(); 868 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC) 869 ? ARM_AM::add : ARM_AM::sub; 870 int Val; 871 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x1000, Val)) { // 12 bits. 872 Offset = CurDAG->getRegister(0, MVT::i32); 873 Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val, 874 ARM_AM::no_shift), 875 MVT::i32); 876 return true; 877 } 878 879 return false; 880} 881 882bool ARMDAGToDAGISel::SelectAddrOffsetNone(SDValue N, SDValue &Base) { 883 Base = N; 884 return true; 885} 886 887bool ARMDAGToDAGISel::SelectAddrMode3(SDValue N, 888 SDValue &Base, SDValue &Offset, 889 SDValue &Opc) { 890 if (N.getOpcode() == ISD::SUB) { 891 // X - C is canonicalize to X + -C, no need to handle it here. 892 Base = N.getOperand(0); 893 Offset = N.getOperand(1); 894 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32); 895 return true; 896 } 897 898 if (!CurDAG->isBaseWithConstantOffset(N)) { 899 Base = N; 900 if (N.getOpcode() == ISD::FrameIndex) { 901 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 902 Base = CurDAG->getTargetFrameIndex(FI, 903 getTargetLowering()->getPointerTy()); 904 } 905 Offset = CurDAG->getRegister(0, MVT::i32); 906 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32); 907 return true; 908 } 909 910 // If the RHS is +/- imm8, fold into addr mode. 911 int RHSC; 912 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1, 913 -256 + 1, 256, RHSC)) { // 8 bits. 914 Base = N.getOperand(0); 915 if (Base.getOpcode() == ISD::FrameIndex) { 916 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 917 Base = CurDAG->getTargetFrameIndex(FI, 918 getTargetLowering()->getPointerTy()); 919 } 920 Offset = CurDAG->getRegister(0, MVT::i32); 921 922 ARM_AM::AddrOpc AddSub = ARM_AM::add; 923 if (RHSC < 0) { 924 AddSub = ARM_AM::sub; 925 RHSC = -RHSC; 926 } 927 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32); 928 return true; 929 } 930 931 Base = N.getOperand(0); 932 Offset = N.getOperand(1); 933 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32); 934 return true; 935} 936 937bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N, 938 SDValue &Offset, SDValue &Opc) { 939 unsigned Opcode = Op->getOpcode(); 940 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) 941 ? cast<LoadSDNode>(Op)->getAddressingMode() 942 : cast<StoreSDNode>(Op)->getAddressingMode(); 943 ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC) 944 ? ARM_AM::add : ARM_AM::sub; 945 int Val; 946 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 256, Val)) { // 12 bits. 947 Offset = CurDAG->getRegister(0, MVT::i32); 948 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32); 949 return true; 950 } 951 952 Offset = N; 953 Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32); 954 return true; 955} 956 957bool ARMDAGToDAGISel::SelectAddrMode5(SDValue N, 958 SDValue &Base, SDValue &Offset) { 959 if (!CurDAG->isBaseWithConstantOffset(N)) { 960 Base = N; 961 if (N.getOpcode() == ISD::FrameIndex) { 962 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 963 Base = CurDAG->getTargetFrameIndex(FI, 964 getTargetLowering()->getPointerTy()); 965 } else if (N.getOpcode() == ARMISD::Wrapper && 966 !(Subtarget->useMovt() && 967 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) { 968 Base = N.getOperand(0); 969 } 970 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0), 971 MVT::i32); 972 return true; 973 } 974 975 // If the RHS is +/- imm8, fold into addr mode. 976 int RHSC; 977 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/4, 978 -256 + 1, 256, RHSC)) { 979 Base = N.getOperand(0); 980 if (Base.getOpcode() == ISD::FrameIndex) { 981 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 982 Base = CurDAG->getTargetFrameIndex(FI, 983 getTargetLowering()->getPointerTy()); 984 } 985 986 ARM_AM::AddrOpc AddSub = ARM_AM::add; 987 if (RHSC < 0) { 988 AddSub = ARM_AM::sub; 989 RHSC = -RHSC; 990 } 991 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC), 992 MVT::i32); 993 return true; 994 } 995 996 Base = N; 997 Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0), 998 MVT::i32); 999 return true; 1000} 1001 1002bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr, 1003 SDValue &Align) { 1004 Addr = N; 1005 1006 unsigned Alignment = 0; 1007 if (LSBaseSDNode *LSN = dyn_cast<LSBaseSDNode>(Parent)) { 1008 // This case occurs only for VLD1-lane/dup and VST1-lane instructions. 1009 // The maximum alignment is equal to the memory size being referenced. 1010 unsigned LSNAlign = LSN->getAlignment(); 1011 unsigned MemSize = LSN->getMemoryVT().getSizeInBits() / 8; 1012 if (LSNAlign >= MemSize && MemSize > 1) 1013 Alignment = MemSize; 1014 } else { 1015 // All other uses of addrmode6 are for intrinsics. For now just record 1016 // the raw alignment value; it will be refined later based on the legal 1017 // alignment operands for the intrinsic. 1018 Alignment = cast<MemIntrinsicSDNode>(Parent)->getAlignment(); 1019 } 1020 1021 Align = CurDAG->getTargetConstant(Alignment, MVT::i32); 1022 return true; 1023} 1024 1025bool ARMDAGToDAGISel::SelectAddrMode6Offset(SDNode *Op, SDValue N, 1026 SDValue &Offset) { 1027 LSBaseSDNode *LdSt = cast<LSBaseSDNode>(Op); 1028 ISD::MemIndexedMode AM = LdSt->getAddressingMode(); 1029 if (AM != ISD::POST_INC) 1030 return false; 1031 Offset = N; 1032 if (ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N)) { 1033 if (NC->getZExtValue() * 8 == LdSt->getMemoryVT().getSizeInBits()) 1034 Offset = CurDAG->getRegister(0, MVT::i32); 1035 } 1036 return true; 1037} 1038 1039bool ARMDAGToDAGISel::SelectAddrModePC(SDValue N, 1040 SDValue &Offset, SDValue &Label) { 1041 if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) { 1042 Offset = N.getOperand(0); 1043 SDValue N1 = N.getOperand(1); 1044 Label = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(), 1045 MVT::i32); 1046 return true; 1047 } 1048 1049 return false; 1050} 1051 1052 1053//===----------------------------------------------------------------------===// 1054// Thumb Addressing Modes 1055//===----------------------------------------------------------------------===// 1056 1057bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue N, 1058 SDValue &Base, SDValue &Offset){ 1059 if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(N)) { 1060 ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N); 1061 if (!NC || !NC->isNullValue()) 1062 return false; 1063 1064 Base = Offset = N; 1065 return true; 1066 } 1067 1068 Base = N.getOperand(0); 1069 Offset = N.getOperand(1); 1070 return true; 1071} 1072 1073bool 1074ARMDAGToDAGISel::SelectThumbAddrModeRI(SDValue N, SDValue &Base, 1075 SDValue &Offset, unsigned Scale) { 1076 if (Scale == 4) { 1077 SDValue TmpBase, TmpOffImm; 1078 if (SelectThumbAddrModeSP(N, TmpBase, TmpOffImm)) 1079 return false; // We want to select tLDRspi / tSTRspi instead. 1080 1081 if (N.getOpcode() == ARMISD::Wrapper && 1082 N.getOperand(0).getOpcode() == ISD::TargetConstantPool) 1083 return false; // We want to select tLDRpci instead. 1084 } 1085 1086 if (!CurDAG->isBaseWithConstantOffset(N)) 1087 return false; 1088 1089 // Thumb does not have [sp, r] address mode. 1090 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0)); 1091 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1)); 1092 if ((LHSR && LHSR->getReg() == ARM::SP) || 1093 (RHSR && RHSR->getReg() == ARM::SP)) 1094 return false; 1095 1096 // FIXME: Why do we explicitly check for a match here and then return false? 1097 // Presumably to allow something else to match, but shouldn't this be 1098 // documented? 1099 int RHSC; 1100 if (isScaledConstantInRange(N.getOperand(1), Scale, 0, 32, RHSC)) 1101 return false; 1102 1103 Base = N.getOperand(0); 1104 Offset = N.getOperand(1); 1105 return true; 1106} 1107 1108bool 1109ARMDAGToDAGISel::SelectThumbAddrModeRI5S1(SDValue N, 1110 SDValue &Base, 1111 SDValue &Offset) { 1112 return SelectThumbAddrModeRI(N, Base, Offset, 1); 1113} 1114 1115bool 1116ARMDAGToDAGISel::SelectThumbAddrModeRI5S2(SDValue N, 1117 SDValue &Base, 1118 SDValue &Offset) { 1119 return SelectThumbAddrModeRI(N, Base, Offset, 2); 1120} 1121 1122bool 1123ARMDAGToDAGISel::SelectThumbAddrModeRI5S4(SDValue N, 1124 SDValue &Base, 1125 SDValue &Offset) { 1126 return SelectThumbAddrModeRI(N, Base, Offset, 4); 1127} 1128 1129bool 1130ARMDAGToDAGISel::SelectThumbAddrModeImm5S(SDValue N, unsigned Scale, 1131 SDValue &Base, SDValue &OffImm) { 1132 if (Scale == 4) { 1133 SDValue TmpBase, TmpOffImm; 1134 if (SelectThumbAddrModeSP(N, TmpBase, TmpOffImm)) 1135 return false; // We want to select tLDRspi / tSTRspi instead. 1136 1137 if (N.getOpcode() == ARMISD::Wrapper && 1138 N.getOperand(0).getOpcode() == ISD::TargetConstantPool) 1139 return false; // We want to select tLDRpci instead. 1140 } 1141 1142 if (!CurDAG->isBaseWithConstantOffset(N)) { 1143 if (N.getOpcode() == ARMISD::Wrapper && 1144 !(Subtarget->useMovt() && 1145 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) { 1146 Base = N.getOperand(0); 1147 } else { 1148 Base = N; 1149 } 1150 1151 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1152 return true; 1153 } 1154 1155 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0)); 1156 RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1)); 1157 if ((LHSR && LHSR->getReg() == ARM::SP) || 1158 (RHSR && RHSR->getReg() == ARM::SP)) { 1159 ConstantSDNode *LHS = dyn_cast<ConstantSDNode>(N.getOperand(0)); 1160 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1)); 1161 unsigned LHSC = LHS ? LHS->getZExtValue() : 0; 1162 unsigned RHSC = RHS ? RHS->getZExtValue() : 0; 1163 1164 // Thumb does not have [sp, #imm5] address mode for non-zero imm5. 1165 if (LHSC != 0 || RHSC != 0) return false; 1166 1167 Base = N; 1168 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1169 return true; 1170 } 1171 1172 // If the RHS is + imm5 * scale, fold into addr mode. 1173 int RHSC; 1174 if (isScaledConstantInRange(N.getOperand(1), Scale, 0, 32, RHSC)) { 1175 Base = N.getOperand(0); 1176 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); 1177 return true; 1178 } 1179 1180 Base = N.getOperand(0); 1181 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1182 return true; 1183} 1184 1185bool 1186ARMDAGToDAGISel::SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base, 1187 SDValue &OffImm) { 1188 return SelectThumbAddrModeImm5S(N, 4, Base, OffImm); 1189} 1190 1191bool 1192ARMDAGToDAGISel::SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base, 1193 SDValue &OffImm) { 1194 return SelectThumbAddrModeImm5S(N, 2, Base, OffImm); 1195} 1196 1197bool 1198ARMDAGToDAGISel::SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base, 1199 SDValue &OffImm) { 1200 return SelectThumbAddrModeImm5S(N, 1, Base, OffImm); 1201} 1202 1203bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDValue N, 1204 SDValue &Base, SDValue &OffImm) { 1205 if (N.getOpcode() == ISD::FrameIndex) { 1206 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 1207 Base = CurDAG->getTargetFrameIndex(FI, 1208 getTargetLowering()->getPointerTy()); 1209 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1210 return true; 1211 } 1212 1213 if (!CurDAG->isBaseWithConstantOffset(N)) 1214 return false; 1215 1216 RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0)); 1217 if (N.getOperand(0).getOpcode() == ISD::FrameIndex || 1218 (LHSR && LHSR->getReg() == ARM::SP)) { 1219 // If the RHS is + imm8 * scale, fold into addr mode. 1220 int RHSC; 1221 if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/4, 0, 256, RHSC)) { 1222 Base = N.getOperand(0); 1223 if (Base.getOpcode() == ISD::FrameIndex) { 1224 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 1225 Base = CurDAG->getTargetFrameIndex(FI, 1226 getTargetLowering()->getPointerTy()); 1227 } 1228 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); 1229 return true; 1230 } 1231 } 1232 1233 return false; 1234} 1235 1236 1237//===----------------------------------------------------------------------===// 1238// Thumb 2 Addressing Modes 1239//===----------------------------------------------------------------------===// 1240 1241 1242bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDValue N, SDValue &BaseReg, 1243 SDValue &Opc) { 1244 if (DisableShifterOp) 1245 return false; 1246 1247 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOpcode()); 1248 1249 // Don't match base register only case. That is matched to a separate 1250 // lower complexity pattern with explicit register operand. 1251 if (ShOpcVal == ARM_AM::no_shift) return false; 1252 1253 BaseReg = N.getOperand(0); 1254 unsigned ShImmVal = 0; 1255 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 1256 ShImmVal = RHS->getZExtValue() & 31; 1257 Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal)); 1258 return true; 1259 } 1260 1261 return false; 1262} 1263 1264bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue N, 1265 SDValue &Base, SDValue &OffImm) { 1266 // Match simple R + imm12 operands. 1267 1268 // Base only. 1269 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && 1270 !CurDAG->isBaseWithConstantOffset(N)) { 1271 if (N.getOpcode() == ISD::FrameIndex) { 1272 // Match frame index. 1273 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 1274 Base = CurDAG->getTargetFrameIndex(FI, 1275 getTargetLowering()->getPointerTy()); 1276 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1277 return true; 1278 } 1279 1280 if (N.getOpcode() == ARMISD::Wrapper && 1281 !(Subtarget->useMovt() && 1282 N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) { 1283 Base = N.getOperand(0); 1284 if (Base.getOpcode() == ISD::TargetConstantPool) 1285 return false; // We want to select t2LDRpci instead. 1286 } else 1287 Base = N; 1288 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1289 return true; 1290 } 1291 1292 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 1293 if (SelectT2AddrModeImm8(N, Base, OffImm)) 1294 // Let t2LDRi8 handle (R - imm8). 1295 return false; 1296 1297 int RHSC = (int)RHS->getZExtValue(); 1298 if (N.getOpcode() == ISD::SUB) 1299 RHSC = -RHSC; 1300 1301 if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned) 1302 Base = N.getOperand(0); 1303 if (Base.getOpcode() == ISD::FrameIndex) { 1304 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 1305 Base = CurDAG->getTargetFrameIndex(FI, 1306 getTargetLowering()->getPointerTy()); 1307 } 1308 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); 1309 return true; 1310 } 1311 } 1312 1313 // Base only. 1314 Base = N; 1315 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1316 return true; 1317} 1318 1319bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue N, 1320 SDValue &Base, SDValue &OffImm) { 1321 // Match simple R - imm8 operands. 1322 if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && 1323 !CurDAG->isBaseWithConstantOffset(N)) 1324 return false; 1325 1326 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 1327 int RHSC = (int)RHS->getSExtValue(); 1328 if (N.getOpcode() == ISD::SUB) 1329 RHSC = -RHSC; 1330 1331 if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative) 1332 Base = N.getOperand(0); 1333 if (Base.getOpcode() == ISD::FrameIndex) { 1334 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 1335 Base = CurDAG->getTargetFrameIndex(FI, 1336 getTargetLowering()->getPointerTy()); 1337 } 1338 OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32); 1339 return true; 1340 } 1341 } 1342 1343 return false; 1344} 1345 1346bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N, 1347 SDValue &OffImm){ 1348 unsigned Opcode = Op->getOpcode(); 1349 ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) 1350 ? cast<LoadSDNode>(Op)->getAddressingMode() 1351 : cast<StoreSDNode>(Op)->getAddressingMode(); 1352 int RHSC; 1353 if (isScaledConstantInRange(N, /*Scale=*/1, 0, 0x100, RHSC)) { // 8 bits. 1354 OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC)) 1355 ? CurDAG->getTargetConstant(RHSC, MVT::i32) 1356 : CurDAG->getTargetConstant(-RHSC, MVT::i32); 1357 return true; 1358 } 1359 1360 return false; 1361} 1362 1363bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue N, 1364 SDValue &Base, 1365 SDValue &OffReg, SDValue &ShImm) { 1366 // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12. 1367 if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(N)) 1368 return false; 1369 1370 // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8. 1371 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) { 1372 int RHSC = (int)RHS->getZExtValue(); 1373 if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned) 1374 return false; 1375 else if (RHSC < 0 && RHSC >= -255) // 8 bits 1376 return false; 1377 } 1378 1379 // Look for (R + R) or (R + (R << [1,2,3])). 1380 unsigned ShAmt = 0; 1381 Base = N.getOperand(0); 1382 OffReg = N.getOperand(1); 1383 1384 // Swap if it is ((R << c) + R). 1385 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg.getOpcode()); 1386 if (ShOpcVal != ARM_AM::lsl) { 1387 ShOpcVal = ARM_AM::getShiftOpcForNode(Base.getOpcode()); 1388 if (ShOpcVal == ARM_AM::lsl) 1389 std::swap(Base, OffReg); 1390 } 1391 1392 if (ShOpcVal == ARM_AM::lsl) { 1393 // Check to see if the RHS of the shift is a constant, if not, we can't fold 1394 // it. 1395 if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) { 1396 ShAmt = Sh->getZExtValue(); 1397 if (ShAmt < 4 && isShifterOpProfitable(OffReg, ShOpcVal, ShAmt)) 1398 OffReg = OffReg.getOperand(0); 1399 else { 1400 ShAmt = 0; 1401 ShOpcVal = ARM_AM::no_shift; 1402 } 1403 } else { 1404 ShOpcVal = ARM_AM::no_shift; 1405 } 1406 } 1407 1408 ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32); 1409 1410 return true; 1411} 1412 1413bool ARMDAGToDAGISel::SelectT2AddrModeExclusive(SDValue N, SDValue &Base, 1414 SDValue &OffImm) { 1415 // This *must* succeed since it's used for the irreplacable ldrex and strex 1416 // instructions. 1417 Base = N; 1418 OffImm = CurDAG->getTargetConstant(0, MVT::i32); 1419 1420 if (N.getOpcode() != ISD::ADD || !CurDAG->isBaseWithConstantOffset(N)) 1421 return true; 1422 1423 ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1)); 1424 if (!RHS) 1425 return true; 1426 1427 uint32_t RHSC = (int)RHS->getZExtValue(); 1428 if (RHSC > 1020 || RHSC % 4 != 0) 1429 return true; 1430 1431 Base = N.getOperand(0); 1432 if (Base.getOpcode() == ISD::FrameIndex) { 1433 int FI = cast<FrameIndexSDNode>(Base)->getIndex(); 1434 Base = CurDAG->getTargetFrameIndex(FI, getTargetLowering()->getPointerTy()); 1435 } 1436 1437 OffImm = CurDAG->getTargetConstant(RHSC / 4, MVT::i32); 1438 return true; 1439} 1440 1441//===--------------------------------------------------------------------===// 1442 1443/// getAL - Returns a ARMCC::AL immediate node. 1444static inline SDValue getAL(SelectionDAG *CurDAG) { 1445 return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32); 1446} 1447 1448SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDNode *N) { 1449 LoadSDNode *LD = cast<LoadSDNode>(N); 1450 ISD::MemIndexedMode AM = LD->getAddressingMode(); 1451 if (AM == ISD::UNINDEXED) 1452 return NULL; 1453 1454 EVT LoadedVT = LD->getMemoryVT(); 1455 SDValue Offset, AMOpc; 1456 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC); 1457 unsigned Opcode = 0; 1458 bool Match = false; 1459 if (LoadedVT == MVT::i32 && isPre && 1460 SelectAddrMode2OffsetImmPre(N, LD->getOffset(), Offset, AMOpc)) { 1461 Opcode = ARM::LDR_PRE_IMM; 1462 Match = true; 1463 } else if (LoadedVT == MVT::i32 && !isPre && 1464 SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) { 1465 Opcode = ARM::LDR_POST_IMM; 1466 Match = true; 1467 } else if (LoadedVT == MVT::i32 && 1468 SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) { 1469 Opcode = isPre ? ARM::LDR_PRE_REG : ARM::LDR_POST_REG; 1470 Match = true; 1471 1472 } else if (LoadedVT == MVT::i16 && 1473 SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) { 1474 Match = true; 1475 Opcode = (LD->getExtensionType() == ISD::SEXTLOAD) 1476 ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST) 1477 : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST); 1478 } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) { 1479 if (LD->getExtensionType() == ISD::SEXTLOAD) { 1480 if (SelectAddrMode3Offset(N, LD->getOffset(), Offset, AMOpc)) { 1481 Match = true; 1482 Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST; 1483 } 1484 } else { 1485 if (isPre && 1486 SelectAddrMode2OffsetImmPre(N, LD->getOffset(), Offset, AMOpc)) { 1487 Match = true; 1488 Opcode = ARM::LDRB_PRE_IMM; 1489 } else if (!isPre && 1490 SelectAddrMode2OffsetImm(N, LD->getOffset(), Offset, AMOpc)) { 1491 Match = true; 1492 Opcode = ARM::LDRB_POST_IMM; 1493 } else if (SelectAddrMode2OffsetReg(N, LD->getOffset(), Offset, AMOpc)) { 1494 Match = true; 1495 Opcode = isPre ? ARM::LDRB_PRE_REG : ARM::LDRB_POST_REG; 1496 } 1497 } 1498 } 1499 1500 if (Match) { 1501 if (Opcode == ARM::LDR_PRE_IMM || Opcode == ARM::LDRB_PRE_IMM) { 1502 SDValue Chain = LD->getChain(); 1503 SDValue Base = LD->getBasePtr(); 1504 SDValue Ops[]= { Base, AMOpc, getAL(CurDAG), 1505 CurDAG->getRegister(0, MVT::i32), Chain }; 1506 return CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, 1507 MVT::i32, MVT::Other, Ops); 1508 } else { 1509 SDValue Chain = LD->getChain(); 1510 SDValue Base = LD->getBasePtr(); 1511 SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG), 1512 CurDAG->getRegister(0, MVT::i32), Chain }; 1513 return CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, 1514 MVT::i32, MVT::Other, Ops); 1515 } 1516 } 1517 1518 return NULL; 1519} 1520 1521SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDNode *N) { 1522 LoadSDNode *LD = cast<LoadSDNode>(N); 1523 ISD::MemIndexedMode AM = LD->getAddressingMode(); 1524 if (AM == ISD::UNINDEXED) 1525 return NULL; 1526 1527 EVT LoadedVT = LD->getMemoryVT(); 1528 bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD; 1529 SDValue Offset; 1530 bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC); 1531 unsigned Opcode = 0; 1532 bool Match = false; 1533 if (SelectT2AddrModeImm8Offset(N, LD->getOffset(), Offset)) { 1534 switch (LoadedVT.getSimpleVT().SimpleTy) { 1535 case MVT::i32: 1536 Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST; 1537 break; 1538 case MVT::i16: 1539 if (isSExtLd) 1540 Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST; 1541 else 1542 Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST; 1543 break; 1544 case MVT::i8: 1545 case MVT::i1: 1546 if (isSExtLd) 1547 Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST; 1548 else 1549 Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST; 1550 break; 1551 default: 1552 return NULL; 1553 } 1554 Match = true; 1555 } 1556 1557 if (Match) { 1558 SDValue Chain = LD->getChain(); 1559 SDValue Base = LD->getBasePtr(); 1560 SDValue Ops[]= { Base, Offset, getAL(CurDAG), 1561 CurDAG->getRegister(0, MVT::i32), Chain }; 1562 return CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, MVT::i32, 1563 MVT::Other, Ops); 1564 } 1565 1566 return NULL; 1567} 1568 1569/// \brief Form a GPRPair pseudo register from a pair of GPR regs. 1570SDNode *ARMDAGToDAGISel::createGPRPairNode(EVT VT, SDValue V0, SDValue V1) { 1571 SDLoc dl(V0.getNode()); 1572 SDValue RegClass = 1573 CurDAG->getTargetConstant(ARM::GPRPairRegClassID, MVT::i32); 1574 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::gsub_0, MVT::i32); 1575 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::gsub_1, MVT::i32); 1576 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 }; 1577 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops); 1578} 1579 1580/// \brief Form a D register from a pair of S registers. 1581SDNode *ARMDAGToDAGISel::createSRegPairNode(EVT VT, SDValue V0, SDValue V1) { 1582 SDLoc dl(V0.getNode()); 1583 SDValue RegClass = 1584 CurDAG->getTargetConstant(ARM::DPR_VFP2RegClassID, MVT::i32); 1585 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32); 1586 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32); 1587 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 }; 1588 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops); 1589} 1590 1591/// \brief Form a quad register from a pair of D registers. 1592SDNode *ARMDAGToDAGISel::createDRegPairNode(EVT VT, SDValue V0, SDValue V1) { 1593 SDLoc dl(V0.getNode()); 1594 SDValue RegClass = CurDAG->getTargetConstant(ARM::QPRRegClassID, MVT::i32); 1595 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32); 1596 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32); 1597 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 }; 1598 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops); 1599} 1600 1601/// \brief Form 4 consecutive D registers from a pair of Q registers. 1602SDNode *ARMDAGToDAGISel::createQRegPairNode(EVT VT, SDValue V0, SDValue V1) { 1603 SDLoc dl(V0.getNode()); 1604 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, MVT::i32); 1605 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32); 1606 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32); 1607 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 }; 1608 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops); 1609} 1610 1611/// \brief Form 4 consecutive S registers. 1612SDNode *ARMDAGToDAGISel::createQuadSRegsNode(EVT VT, SDValue V0, SDValue V1, 1613 SDValue V2, SDValue V3) { 1614 SDLoc dl(V0.getNode()); 1615 SDValue RegClass = 1616 CurDAG->getTargetConstant(ARM::QPR_VFP2RegClassID, MVT::i32); 1617 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32); 1618 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32); 1619 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32); 1620 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32); 1621 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1, 1622 V2, SubReg2, V3, SubReg3 }; 1623 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops); 1624} 1625 1626/// \brief Form 4 consecutive D registers. 1627SDNode *ARMDAGToDAGISel::createQuadDRegsNode(EVT VT, SDValue V0, SDValue V1, 1628 SDValue V2, SDValue V3) { 1629 SDLoc dl(V0.getNode()); 1630 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQPRRegClassID, MVT::i32); 1631 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::dsub_0, MVT::i32); 1632 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::dsub_1, MVT::i32); 1633 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::dsub_2, MVT::i32); 1634 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::dsub_3, MVT::i32); 1635 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1, 1636 V2, SubReg2, V3, SubReg3 }; 1637 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops); 1638} 1639 1640/// \brief Form 4 consecutive Q registers. 1641SDNode *ARMDAGToDAGISel::createQuadQRegsNode(EVT VT, SDValue V0, SDValue V1, 1642 SDValue V2, SDValue V3) { 1643 SDLoc dl(V0.getNode()); 1644 SDValue RegClass = CurDAG->getTargetConstant(ARM::QQQQPRRegClassID, MVT::i32); 1645 SDValue SubReg0 = CurDAG->getTargetConstant(ARM::qsub_0, MVT::i32); 1646 SDValue SubReg1 = CurDAG->getTargetConstant(ARM::qsub_1, MVT::i32); 1647 SDValue SubReg2 = CurDAG->getTargetConstant(ARM::qsub_2, MVT::i32); 1648 SDValue SubReg3 = CurDAG->getTargetConstant(ARM::qsub_3, MVT::i32); 1649 const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1, 1650 V2, SubReg2, V3, SubReg3 }; 1651 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops); 1652} 1653 1654/// GetVLDSTAlign - Get the alignment (in bytes) for the alignment operand 1655/// of a NEON VLD or VST instruction. The supported values depend on the 1656/// number of registers being loaded. 1657SDValue ARMDAGToDAGISel::GetVLDSTAlign(SDValue Align, unsigned NumVecs, 1658 bool is64BitVector) { 1659 unsigned NumRegs = NumVecs; 1660 if (!is64BitVector && NumVecs < 3) 1661 NumRegs *= 2; 1662 1663 unsigned Alignment = cast<ConstantSDNode>(Align)->getZExtValue(); 1664 if (Alignment >= 32 && NumRegs == 4) 1665 Alignment = 32; 1666 else if (Alignment >= 16 && (NumRegs == 2 || NumRegs == 4)) 1667 Alignment = 16; 1668 else if (Alignment >= 8) 1669 Alignment = 8; 1670 else 1671 Alignment = 0; 1672 1673 return CurDAG->getTargetConstant(Alignment, MVT::i32); 1674} 1675 1676// Get the register stride update opcode of a VLD/VST instruction that 1677// is otherwise equivalent to the given fixed stride updating instruction. 1678static unsigned getVLDSTRegisterUpdateOpcode(unsigned Opc) { 1679 switch (Opc) { 1680 default: break; 1681 case ARM::VLD1d8wb_fixed: return ARM::VLD1d8wb_register; 1682 case ARM::VLD1d16wb_fixed: return ARM::VLD1d16wb_register; 1683 case ARM::VLD1d32wb_fixed: return ARM::VLD1d32wb_register; 1684 case ARM::VLD1d64wb_fixed: return ARM::VLD1d64wb_register; 1685 case ARM::VLD1q8wb_fixed: return ARM::VLD1q8wb_register; 1686 case ARM::VLD1q16wb_fixed: return ARM::VLD1q16wb_register; 1687 case ARM::VLD1q32wb_fixed: return ARM::VLD1q32wb_register; 1688 case ARM::VLD1q64wb_fixed: return ARM::VLD1q64wb_register; 1689 1690 case ARM::VST1d8wb_fixed: return ARM::VST1d8wb_register; 1691 case ARM::VST1d16wb_fixed: return ARM::VST1d16wb_register; 1692 case ARM::VST1d32wb_fixed: return ARM::VST1d32wb_register; 1693 case ARM::VST1d64wb_fixed: return ARM::VST1d64wb_register; 1694 case ARM::VST1q8wb_fixed: return ARM::VST1q8wb_register; 1695 case ARM::VST1q16wb_fixed: return ARM::VST1q16wb_register; 1696 case ARM::VST1q32wb_fixed: return ARM::VST1q32wb_register; 1697 case ARM::VST1q64wb_fixed: return ARM::VST1q64wb_register; 1698 case ARM::VST1d64TPseudoWB_fixed: return ARM::VST1d64TPseudoWB_register; 1699 case ARM::VST1d64QPseudoWB_fixed: return ARM::VST1d64QPseudoWB_register; 1700 1701 case ARM::VLD2d8wb_fixed: return ARM::VLD2d8wb_register; 1702 case ARM::VLD2d16wb_fixed: return ARM::VLD2d16wb_register; 1703 case ARM::VLD2d32wb_fixed: return ARM::VLD2d32wb_register; 1704 case ARM::VLD2q8PseudoWB_fixed: return ARM::VLD2q8PseudoWB_register; 1705 case ARM::VLD2q16PseudoWB_fixed: return ARM::VLD2q16PseudoWB_register; 1706 case ARM::VLD2q32PseudoWB_fixed: return ARM::VLD2q32PseudoWB_register; 1707 1708 case ARM::VST2d8wb_fixed: return ARM::VST2d8wb_register; 1709 case ARM::VST2d16wb_fixed: return ARM::VST2d16wb_register; 1710 case ARM::VST2d32wb_fixed: return ARM::VST2d32wb_register; 1711 case ARM::VST2q8PseudoWB_fixed: return ARM::VST2q8PseudoWB_register; 1712 case ARM::VST2q16PseudoWB_fixed: return ARM::VST2q16PseudoWB_register; 1713 case ARM::VST2q32PseudoWB_fixed: return ARM::VST2q32PseudoWB_register; 1714 1715 case ARM::VLD2DUPd8wb_fixed: return ARM::VLD2DUPd8wb_register; 1716 case ARM::VLD2DUPd16wb_fixed: return ARM::VLD2DUPd16wb_register; 1717 case ARM::VLD2DUPd32wb_fixed: return ARM::VLD2DUPd32wb_register; 1718 } 1719 return Opc; // If not one we handle, return it unchanged. 1720} 1721 1722SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs, 1723 const uint16_t *DOpcodes, 1724 const uint16_t *QOpcodes0, 1725 const uint16_t *QOpcodes1) { 1726 assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range"); 1727 SDLoc dl(N); 1728 1729 SDValue MemAddr, Align; 1730 unsigned AddrOpIdx = isUpdating ? 1 : 2; 1731 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align)) 1732 return NULL; 1733 1734 SDValue Chain = N->getOperand(0); 1735 EVT VT = N->getValueType(0); 1736 bool is64BitVector = VT.is64BitVector(); 1737 Align = GetVLDSTAlign(Align, NumVecs, is64BitVector); 1738 1739 unsigned OpcodeIndex; 1740 switch (VT.getSimpleVT().SimpleTy) { 1741 default: llvm_unreachable("unhandled vld type"); 1742 // Double-register operations: 1743 case MVT::v8i8: OpcodeIndex = 0; break; 1744 case MVT::v4i16: OpcodeIndex = 1; break; 1745 case MVT::v2f32: 1746 case MVT::v2i32: OpcodeIndex = 2; break; 1747 case MVT::v1i64: OpcodeIndex = 3; break; 1748 // Quad-register operations: 1749 case MVT::v16i8: OpcodeIndex = 0; break; 1750 case MVT::v8i16: OpcodeIndex = 1; break; 1751 case MVT::v4f32: 1752 case MVT::v4i32: OpcodeIndex = 2; break; 1753 case MVT::v2i64: OpcodeIndex = 3; 1754 assert(NumVecs == 1 && "v2i64 type only supported for VLD1"); 1755 break; 1756 } 1757 1758 EVT ResTy; 1759 if (NumVecs == 1) 1760 ResTy = VT; 1761 else { 1762 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs; 1763 if (!is64BitVector) 1764 ResTyElts *= 2; 1765 ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, ResTyElts); 1766 } 1767 std::vector<EVT> ResTys; 1768 ResTys.push_back(ResTy); 1769 if (isUpdating) 1770 ResTys.push_back(MVT::i32); 1771 ResTys.push_back(MVT::Other); 1772 1773 SDValue Pred = getAL(CurDAG); 1774 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 1775 SDNode *VLd; 1776 SmallVector<SDValue, 7> Ops; 1777 1778 // Double registers and VLD1/VLD2 quad registers are directly supported. 1779 if (is64BitVector || NumVecs <= 2) { 1780 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] : 1781 QOpcodes0[OpcodeIndex]); 1782 Ops.push_back(MemAddr); 1783 Ops.push_back(Align); 1784 if (isUpdating) { 1785 SDValue Inc = N->getOperand(AddrOpIdx + 1); 1786 // FIXME: VLD1/VLD2 fixed increment doesn't need Reg0. Remove the reg0 1787 // case entirely when the rest are updated to that form, too. 1788 if ((NumVecs == 1 || NumVecs == 2) && !isa<ConstantSDNode>(Inc.getNode())) 1789 Opc = getVLDSTRegisterUpdateOpcode(Opc); 1790 // We use a VLD1 for v1i64 even if the pseudo says vld2/3/4, so 1791 // check for that explicitly too. Horribly hacky, but temporary. 1792 if ((NumVecs != 1 && NumVecs != 2 && Opc != ARM::VLD1q64wb_fixed) || 1793 !isa<ConstantSDNode>(Inc.getNode())) 1794 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc); 1795 } 1796 Ops.push_back(Pred); 1797 Ops.push_back(Reg0); 1798 Ops.push_back(Chain); 1799 VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops); 1800 1801 } else { 1802 // Otherwise, quad registers are loaded with two separate instructions, 1803 // where one loads the even registers and the other loads the odd registers. 1804 EVT AddrTy = MemAddr.getValueType(); 1805 1806 // Load the even subregs. This is always an updating load, so that it 1807 // provides the address to the second load for the odd subregs. 1808 SDValue ImplDef = 1809 SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, ResTy), 0); 1810 const SDValue OpsA[] = { MemAddr, Align, Reg0, ImplDef, Pred, Reg0, Chain }; 1811 SDNode *VLdA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl, 1812 ResTy, AddrTy, MVT::Other, OpsA); 1813 Chain = SDValue(VLdA, 2); 1814 1815 // Load the odd subregs. 1816 Ops.push_back(SDValue(VLdA, 1)); 1817 Ops.push_back(Align); 1818 if (isUpdating) { 1819 SDValue Inc = N->getOperand(AddrOpIdx + 1); 1820 assert(isa<ConstantSDNode>(Inc.getNode()) && 1821 "only constant post-increment update allowed for VLD3/4"); 1822 (void)Inc; 1823 Ops.push_back(Reg0); 1824 } 1825 Ops.push_back(SDValue(VLdA, 0)); 1826 Ops.push_back(Pred); 1827 Ops.push_back(Reg0); 1828 Ops.push_back(Chain); 1829 VLd = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys, Ops); 1830 } 1831 1832 // Transfer memoperands. 1833 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 1834 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 1835 cast<MachineSDNode>(VLd)->setMemRefs(MemOp, MemOp + 1); 1836 1837 if (NumVecs == 1) 1838 return VLd; 1839 1840 // Extract out the subregisters. 1841 SDValue SuperReg = SDValue(VLd, 0); 1842 assert(ARM::dsub_7 == ARM::dsub_0+7 && 1843 ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering"); 1844 unsigned Sub0 = (is64BitVector ? ARM::dsub_0 : ARM::qsub_0); 1845 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) 1846 ReplaceUses(SDValue(N, Vec), 1847 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg)); 1848 ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, 1)); 1849 if (isUpdating) 1850 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLd, 2)); 1851 return NULL; 1852} 1853 1854SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs, 1855 const uint16_t *DOpcodes, 1856 const uint16_t *QOpcodes0, 1857 const uint16_t *QOpcodes1) { 1858 assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range"); 1859 SDLoc dl(N); 1860 1861 SDValue MemAddr, Align; 1862 unsigned AddrOpIdx = isUpdating ? 1 : 2; 1863 unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1) 1864 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align)) 1865 return NULL; 1866 1867 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 1868 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 1869 1870 SDValue Chain = N->getOperand(0); 1871 EVT VT = N->getOperand(Vec0Idx).getValueType(); 1872 bool is64BitVector = VT.is64BitVector(); 1873 Align = GetVLDSTAlign(Align, NumVecs, is64BitVector); 1874 1875 unsigned OpcodeIndex; 1876 switch (VT.getSimpleVT().SimpleTy) { 1877 default: llvm_unreachable("unhandled vst type"); 1878 // Double-register operations: 1879 case MVT::v8i8: OpcodeIndex = 0; break; 1880 case MVT::v4i16: OpcodeIndex = 1; break; 1881 case MVT::v2f32: 1882 case MVT::v2i32: OpcodeIndex = 2; break; 1883 case MVT::v1i64: OpcodeIndex = 3; break; 1884 // Quad-register operations: 1885 case MVT::v16i8: OpcodeIndex = 0; break; 1886 case MVT::v8i16: OpcodeIndex = 1; break; 1887 case MVT::v4f32: 1888 case MVT::v4i32: OpcodeIndex = 2; break; 1889 case MVT::v2i64: OpcodeIndex = 3; 1890 assert(NumVecs == 1 && "v2i64 type only supported for VST1"); 1891 break; 1892 } 1893 1894 std::vector<EVT> ResTys; 1895 if (isUpdating) 1896 ResTys.push_back(MVT::i32); 1897 ResTys.push_back(MVT::Other); 1898 1899 SDValue Pred = getAL(CurDAG); 1900 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 1901 SmallVector<SDValue, 7> Ops; 1902 1903 // Double registers and VST1/VST2 quad registers are directly supported. 1904 if (is64BitVector || NumVecs <= 2) { 1905 SDValue SrcReg; 1906 if (NumVecs == 1) { 1907 SrcReg = N->getOperand(Vec0Idx); 1908 } else if (is64BitVector) { 1909 // Form a REG_SEQUENCE to force register allocation. 1910 SDValue V0 = N->getOperand(Vec0Idx + 0); 1911 SDValue V1 = N->getOperand(Vec0Idx + 1); 1912 if (NumVecs == 2) 1913 SrcReg = SDValue(createDRegPairNode(MVT::v2i64, V0, V1), 0); 1914 else { 1915 SDValue V2 = N->getOperand(Vec0Idx + 2); 1916 // If it's a vst3, form a quad D-register and leave the last part as 1917 // an undef. 1918 SDValue V3 = (NumVecs == 3) 1919 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,dl,VT), 0) 1920 : N->getOperand(Vec0Idx + 3); 1921 SrcReg = SDValue(createQuadDRegsNode(MVT::v4i64, V0, V1, V2, V3), 0); 1922 } 1923 } else { 1924 // Form a QQ register. 1925 SDValue Q0 = N->getOperand(Vec0Idx); 1926 SDValue Q1 = N->getOperand(Vec0Idx + 1); 1927 SrcReg = SDValue(createQRegPairNode(MVT::v4i64, Q0, Q1), 0); 1928 } 1929 1930 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] : 1931 QOpcodes0[OpcodeIndex]); 1932 Ops.push_back(MemAddr); 1933 Ops.push_back(Align); 1934 if (isUpdating) { 1935 SDValue Inc = N->getOperand(AddrOpIdx + 1); 1936 // FIXME: VST1/VST2 fixed increment doesn't need Reg0. Remove the reg0 1937 // case entirely when the rest are updated to that form, too. 1938 if (NumVecs <= 2 && !isa<ConstantSDNode>(Inc.getNode())) 1939 Opc = getVLDSTRegisterUpdateOpcode(Opc); 1940 // We use a VST1 for v1i64 even if the pseudo says vld2/3/4, so 1941 // check for that explicitly too. Horribly hacky, but temporary. 1942 if ((NumVecs > 2 && Opc != ARM::VST1q64wb_fixed) || 1943 !isa<ConstantSDNode>(Inc.getNode())) 1944 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc); 1945 } 1946 Ops.push_back(SrcReg); 1947 Ops.push_back(Pred); 1948 Ops.push_back(Reg0); 1949 Ops.push_back(Chain); 1950 SDNode *VSt = CurDAG->getMachineNode(Opc, dl, ResTys, Ops); 1951 1952 // Transfer memoperands. 1953 cast<MachineSDNode>(VSt)->setMemRefs(MemOp, MemOp + 1); 1954 1955 return VSt; 1956 } 1957 1958 // Otherwise, quad registers are stored with two separate instructions, 1959 // where one stores the even registers and the other stores the odd registers. 1960 1961 // Form the QQQQ REG_SEQUENCE. 1962 SDValue V0 = N->getOperand(Vec0Idx + 0); 1963 SDValue V1 = N->getOperand(Vec0Idx + 1); 1964 SDValue V2 = N->getOperand(Vec0Idx + 2); 1965 SDValue V3 = (NumVecs == 3) 1966 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0) 1967 : N->getOperand(Vec0Idx + 3); 1968 SDValue RegSeq = SDValue(createQuadQRegsNode(MVT::v8i64, V0, V1, V2, V3), 0); 1969 1970 // Store the even D registers. This is always an updating store, so that it 1971 // provides the address to the second store for the odd subregs. 1972 const SDValue OpsA[] = { MemAddr, Align, Reg0, RegSeq, Pred, Reg0, Chain }; 1973 SDNode *VStA = CurDAG->getMachineNode(QOpcodes0[OpcodeIndex], dl, 1974 MemAddr.getValueType(), 1975 MVT::Other, OpsA); 1976 cast<MachineSDNode>(VStA)->setMemRefs(MemOp, MemOp + 1); 1977 Chain = SDValue(VStA, 1); 1978 1979 // Store the odd D registers. 1980 Ops.push_back(SDValue(VStA, 0)); 1981 Ops.push_back(Align); 1982 if (isUpdating) { 1983 SDValue Inc = N->getOperand(AddrOpIdx + 1); 1984 assert(isa<ConstantSDNode>(Inc.getNode()) && 1985 "only constant post-increment update allowed for VST3/4"); 1986 (void)Inc; 1987 Ops.push_back(Reg0); 1988 } 1989 Ops.push_back(RegSeq); 1990 Ops.push_back(Pred); 1991 Ops.push_back(Reg0); 1992 Ops.push_back(Chain); 1993 SDNode *VStB = CurDAG->getMachineNode(QOpcodes1[OpcodeIndex], dl, ResTys, 1994 Ops); 1995 cast<MachineSDNode>(VStB)->setMemRefs(MemOp, MemOp + 1); 1996 return VStB; 1997} 1998 1999SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad, 2000 bool isUpdating, unsigned NumVecs, 2001 const uint16_t *DOpcodes, 2002 const uint16_t *QOpcodes) { 2003 assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range"); 2004 SDLoc dl(N); 2005 2006 SDValue MemAddr, Align; 2007 unsigned AddrOpIdx = isUpdating ? 1 : 2; 2008 unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1) 2009 if (!SelectAddrMode6(N, N->getOperand(AddrOpIdx), MemAddr, Align)) 2010 return NULL; 2011 2012 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 2013 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 2014 2015 SDValue Chain = N->getOperand(0); 2016 unsigned Lane = 2017 cast<ConstantSDNode>(N->getOperand(Vec0Idx + NumVecs))->getZExtValue(); 2018 EVT VT = N->getOperand(Vec0Idx).getValueType(); 2019 bool is64BitVector = VT.is64BitVector(); 2020 2021 unsigned Alignment = 0; 2022 if (NumVecs != 3) { 2023 Alignment = cast<ConstantSDNode>(Align)->getZExtValue(); 2024 unsigned NumBytes = NumVecs * VT.getVectorElementType().getSizeInBits()/8; 2025 if (Alignment > NumBytes) 2026 Alignment = NumBytes; 2027 if (Alignment < 8 && Alignment < NumBytes) 2028 Alignment = 0; 2029 // Alignment must be a power of two; make sure of that. 2030 Alignment = (Alignment & -Alignment); 2031 if (Alignment == 1) 2032 Alignment = 0; 2033 } 2034 Align = CurDAG->getTargetConstant(Alignment, MVT::i32); 2035 2036 unsigned OpcodeIndex; 2037 switch (VT.getSimpleVT().SimpleTy) { 2038 default: llvm_unreachable("unhandled vld/vst lane type"); 2039 // Double-register operations: 2040 case MVT::v8i8: OpcodeIndex = 0; break; 2041 case MVT::v4i16: OpcodeIndex = 1; break; 2042 case MVT::v2f32: 2043 case MVT::v2i32: OpcodeIndex = 2; break; 2044 // Quad-register operations: 2045 case MVT::v8i16: OpcodeIndex = 0; break; 2046 case MVT::v4f32: 2047 case MVT::v4i32: OpcodeIndex = 1; break; 2048 } 2049 2050 std::vector<EVT> ResTys; 2051 if (IsLoad) { 2052 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs; 2053 if (!is64BitVector) 2054 ResTyElts *= 2; 2055 ResTys.push_back(EVT::getVectorVT(*CurDAG->getContext(), 2056 MVT::i64, ResTyElts)); 2057 } 2058 if (isUpdating) 2059 ResTys.push_back(MVT::i32); 2060 ResTys.push_back(MVT::Other); 2061 2062 SDValue Pred = getAL(CurDAG); 2063 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 2064 2065 SmallVector<SDValue, 8> Ops; 2066 Ops.push_back(MemAddr); 2067 Ops.push_back(Align); 2068 if (isUpdating) { 2069 SDValue Inc = N->getOperand(AddrOpIdx + 1); 2070 Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc); 2071 } 2072 2073 SDValue SuperReg; 2074 SDValue V0 = N->getOperand(Vec0Idx + 0); 2075 SDValue V1 = N->getOperand(Vec0Idx + 1); 2076 if (NumVecs == 2) { 2077 if (is64BitVector) 2078 SuperReg = SDValue(createDRegPairNode(MVT::v2i64, V0, V1), 0); 2079 else 2080 SuperReg = SDValue(createQRegPairNode(MVT::v4i64, V0, V1), 0); 2081 } else { 2082 SDValue V2 = N->getOperand(Vec0Idx + 2); 2083 SDValue V3 = (NumVecs == 3) 2084 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0) 2085 : N->getOperand(Vec0Idx + 3); 2086 if (is64BitVector) 2087 SuperReg = SDValue(createQuadDRegsNode(MVT::v4i64, V0, V1, V2, V3), 0); 2088 else 2089 SuperReg = SDValue(createQuadQRegsNode(MVT::v8i64, V0, V1, V2, V3), 0); 2090 } 2091 Ops.push_back(SuperReg); 2092 Ops.push_back(getI32Imm(Lane)); 2093 Ops.push_back(Pred); 2094 Ops.push_back(Reg0); 2095 Ops.push_back(Chain); 2096 2097 unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] : 2098 QOpcodes[OpcodeIndex]); 2099 SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops); 2100 cast<MachineSDNode>(VLdLn)->setMemRefs(MemOp, MemOp + 1); 2101 if (!IsLoad) 2102 return VLdLn; 2103 2104 // Extract the subregisters. 2105 SuperReg = SDValue(VLdLn, 0); 2106 assert(ARM::dsub_7 == ARM::dsub_0+7 && 2107 ARM::qsub_3 == ARM::qsub_0+3 && "Unexpected subreg numbering"); 2108 unsigned Sub0 = is64BitVector ? ARM::dsub_0 : ARM::qsub_0; 2109 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) 2110 ReplaceUses(SDValue(N, Vec), 2111 CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg)); 2112 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, 1)); 2113 if (isUpdating) 2114 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdLn, 2)); 2115 return NULL; 2116} 2117 2118SDNode *ARMDAGToDAGISel::SelectVLDDup(SDNode *N, bool isUpdating, 2119 unsigned NumVecs, 2120 const uint16_t *Opcodes) { 2121 assert(NumVecs >=2 && NumVecs <= 4 && "VLDDup NumVecs out-of-range"); 2122 SDLoc dl(N); 2123 2124 SDValue MemAddr, Align; 2125 if (!SelectAddrMode6(N, N->getOperand(1), MemAddr, Align)) 2126 return NULL; 2127 2128 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 2129 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 2130 2131 SDValue Chain = N->getOperand(0); 2132 EVT VT = N->getValueType(0); 2133 2134 unsigned Alignment = 0; 2135 if (NumVecs != 3) { 2136 Alignment = cast<ConstantSDNode>(Align)->getZExtValue(); 2137 unsigned NumBytes = NumVecs * VT.getVectorElementType().getSizeInBits()/8; 2138 if (Alignment > NumBytes) 2139 Alignment = NumBytes; 2140 if (Alignment < 8 && Alignment < NumBytes) 2141 Alignment = 0; 2142 // Alignment must be a power of two; make sure of that. 2143 Alignment = (Alignment & -Alignment); 2144 if (Alignment == 1) 2145 Alignment = 0; 2146 } 2147 Align = CurDAG->getTargetConstant(Alignment, MVT::i32); 2148 2149 unsigned OpcodeIndex; 2150 switch (VT.getSimpleVT().SimpleTy) { 2151 default: llvm_unreachable("unhandled vld-dup type"); 2152 case MVT::v8i8: OpcodeIndex = 0; break; 2153 case MVT::v4i16: OpcodeIndex = 1; break; 2154 case MVT::v2f32: 2155 case MVT::v2i32: OpcodeIndex = 2; break; 2156 } 2157 2158 SDValue Pred = getAL(CurDAG); 2159 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 2160 SDValue SuperReg; 2161 unsigned Opc = Opcodes[OpcodeIndex]; 2162 SmallVector<SDValue, 6> Ops; 2163 Ops.push_back(MemAddr); 2164 Ops.push_back(Align); 2165 if (isUpdating) { 2166 // fixed-stride update instructions don't have an explicit writeback 2167 // operand. It's implicit in the opcode itself. 2168 SDValue Inc = N->getOperand(2); 2169 if (!isa<ConstantSDNode>(Inc.getNode())) 2170 Ops.push_back(Inc); 2171 // FIXME: VLD3 and VLD4 haven't been updated to that form yet. 2172 else if (NumVecs > 2) 2173 Ops.push_back(Reg0); 2174 } 2175 Ops.push_back(Pred); 2176 Ops.push_back(Reg0); 2177 Ops.push_back(Chain); 2178 2179 unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs; 2180 std::vector<EVT> ResTys; 2181 ResTys.push_back(EVT::getVectorVT(*CurDAG->getContext(), MVT::i64,ResTyElts)); 2182 if (isUpdating) 2183 ResTys.push_back(MVT::i32); 2184 ResTys.push_back(MVT::Other); 2185 SDNode *VLdDup = CurDAG->getMachineNode(Opc, dl, ResTys, Ops); 2186 cast<MachineSDNode>(VLdDup)->setMemRefs(MemOp, MemOp + 1); 2187 SuperReg = SDValue(VLdDup, 0); 2188 2189 // Extract the subregisters. 2190 assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering"); 2191 unsigned SubIdx = ARM::dsub_0; 2192 for (unsigned Vec = 0; Vec < NumVecs; ++Vec) 2193 ReplaceUses(SDValue(N, Vec), 2194 CurDAG->getTargetExtractSubreg(SubIdx+Vec, dl, VT, SuperReg)); 2195 ReplaceUses(SDValue(N, NumVecs), SDValue(VLdDup, 1)); 2196 if (isUpdating) 2197 ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdDup, 2)); 2198 return NULL; 2199} 2200 2201SDNode *ARMDAGToDAGISel::SelectVTBL(SDNode *N, bool IsExt, unsigned NumVecs, 2202 unsigned Opc) { 2203 assert(NumVecs >= 2 && NumVecs <= 4 && "VTBL NumVecs out-of-range"); 2204 SDLoc dl(N); 2205 EVT VT = N->getValueType(0); 2206 unsigned FirstTblReg = IsExt ? 2 : 1; 2207 2208 // Form a REG_SEQUENCE to force register allocation. 2209 SDValue RegSeq; 2210 SDValue V0 = N->getOperand(FirstTblReg + 0); 2211 SDValue V1 = N->getOperand(FirstTblReg + 1); 2212 if (NumVecs == 2) 2213 RegSeq = SDValue(createDRegPairNode(MVT::v16i8, V0, V1), 0); 2214 else { 2215 SDValue V2 = N->getOperand(FirstTblReg + 2); 2216 // If it's a vtbl3, form a quad D-register and leave the last part as 2217 // an undef. 2218 SDValue V3 = (NumVecs == 3) 2219 ? SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0) 2220 : N->getOperand(FirstTblReg + 3); 2221 RegSeq = SDValue(createQuadDRegsNode(MVT::v4i64, V0, V1, V2, V3), 0); 2222 } 2223 2224 SmallVector<SDValue, 6> Ops; 2225 if (IsExt) 2226 Ops.push_back(N->getOperand(1)); 2227 Ops.push_back(RegSeq); 2228 Ops.push_back(N->getOperand(FirstTblReg + NumVecs)); 2229 Ops.push_back(getAL(CurDAG)); // predicate 2230 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // predicate register 2231 return CurDAG->getMachineNode(Opc, dl, VT, Ops); 2232} 2233 2234SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDNode *N, 2235 bool isSigned) { 2236 if (!Subtarget->hasV6T2Ops()) 2237 return NULL; 2238 2239 unsigned Opc = isSigned 2240 ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX) 2241 : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX); 2242 2243 // For unsigned extracts, check for a shift right and mask 2244 unsigned And_imm = 0; 2245 if (N->getOpcode() == ISD::AND) { 2246 if (isOpcWithIntImmediate(N, ISD::AND, And_imm)) { 2247 2248 // The immediate is a mask of the low bits iff imm & (imm+1) == 0 2249 if (And_imm & (And_imm + 1)) 2250 return NULL; 2251 2252 unsigned Srl_imm = 0; 2253 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SRL, 2254 Srl_imm)) { 2255 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!"); 2256 2257 // Note: The width operand is encoded as width-1. 2258 unsigned Width = CountTrailingOnes_32(And_imm) - 1; 2259 unsigned LSB = Srl_imm; 2260 2261 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 2262 2263 if ((LSB + Width + 1) == N->getValueType(0).getSizeInBits()) { 2264 // It's cheaper to use a right shift to extract the top bits. 2265 if (Subtarget->isThumb()) { 2266 Opc = isSigned ? ARM::t2ASRri : ARM::t2LSRri; 2267 SDValue Ops[] = { N->getOperand(0).getOperand(0), 2268 CurDAG->getTargetConstant(LSB, MVT::i32), 2269 getAL(CurDAG), Reg0, Reg0 }; 2270 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5); 2271 } 2272 2273 // ARM models shift instructions as MOVsi with shifter operand. 2274 ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(ISD::SRL); 2275 SDValue ShOpc = 2276 CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, LSB), 2277 MVT::i32); 2278 SDValue Ops[] = { N->getOperand(0).getOperand(0), ShOpc, 2279 getAL(CurDAG), Reg0, Reg0 }; 2280 return CurDAG->SelectNodeTo(N, ARM::MOVsi, MVT::i32, Ops, 5); 2281 } 2282 2283 SDValue Ops[] = { N->getOperand(0).getOperand(0), 2284 CurDAG->getTargetConstant(LSB, MVT::i32), 2285 CurDAG->getTargetConstant(Width, MVT::i32), 2286 getAL(CurDAG), Reg0 }; 2287 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5); 2288 } 2289 } 2290 return NULL; 2291 } 2292 2293 // Otherwise, we're looking for a shift of a shift 2294 unsigned Shl_imm = 0; 2295 if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::SHL, Shl_imm)) { 2296 assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!"); 2297 unsigned Srl_imm = 0; 2298 if (isInt32Immediate(N->getOperand(1), Srl_imm)) { 2299 assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!"); 2300 // Note: The width operand is encoded as width-1. 2301 unsigned Width = 32 - Srl_imm - 1; 2302 int LSB = Srl_imm - Shl_imm; 2303 if (LSB < 0) 2304 return NULL; 2305 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 2306 SDValue Ops[] = { N->getOperand(0).getOperand(0), 2307 CurDAG->getTargetConstant(LSB, MVT::i32), 2308 CurDAG->getTargetConstant(Width, MVT::i32), 2309 getAL(CurDAG), Reg0 }; 2310 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5); 2311 } 2312 } 2313 return NULL; 2314} 2315 2316/// Target-specific DAG combining for ISD::XOR. 2317/// Target-independent combining lowers SELECT_CC nodes of the form 2318/// select_cc setg[ge] X, 0, X, -X 2319/// select_cc setgt X, -1, X, -X 2320/// select_cc setl[te] X, 0, -X, X 2321/// select_cc setlt X, 1, -X, X 2322/// which represent Integer ABS into: 2323/// Y = sra (X, size(X)-1); xor (add (X, Y), Y) 2324/// ARM instruction selection detects the latter and matches it to 2325/// ARM::ABS or ARM::t2ABS machine node. 2326SDNode *ARMDAGToDAGISel::SelectABSOp(SDNode *N){ 2327 SDValue XORSrc0 = N->getOperand(0); 2328 SDValue XORSrc1 = N->getOperand(1); 2329 EVT VT = N->getValueType(0); 2330 2331 if (Subtarget->isThumb1Only()) 2332 return NULL; 2333 2334 if (XORSrc0.getOpcode() != ISD::ADD || XORSrc1.getOpcode() != ISD::SRA) 2335 return NULL; 2336 2337 SDValue ADDSrc0 = XORSrc0.getOperand(0); 2338 SDValue ADDSrc1 = XORSrc0.getOperand(1); 2339 SDValue SRASrc0 = XORSrc1.getOperand(0); 2340 SDValue SRASrc1 = XORSrc1.getOperand(1); 2341 ConstantSDNode *SRAConstant = dyn_cast<ConstantSDNode>(SRASrc1); 2342 EVT XType = SRASrc0.getValueType(); 2343 unsigned Size = XType.getSizeInBits() - 1; 2344 2345 if (ADDSrc1 == XORSrc1 && ADDSrc0 == SRASrc0 && 2346 XType.isInteger() && SRAConstant != NULL && 2347 Size == SRAConstant->getZExtValue()) { 2348 unsigned Opcode = Subtarget->isThumb2() ? ARM::t2ABS : ARM::ABS; 2349 return CurDAG->SelectNodeTo(N, Opcode, VT, ADDSrc0); 2350 } 2351 2352 return NULL; 2353} 2354 2355SDNode *ARMDAGToDAGISel::SelectConcatVector(SDNode *N) { 2356 // The only time a CONCAT_VECTORS operation can have legal types is when 2357 // two 64-bit vectors are concatenated to a 128-bit vector. 2358 EVT VT = N->getValueType(0); 2359 if (!VT.is128BitVector() || N->getNumOperands() != 2) 2360 llvm_unreachable("unexpected CONCAT_VECTORS"); 2361 return createDRegPairNode(VT, N->getOperand(0), N->getOperand(1)); 2362} 2363 2364SDNode *ARMDAGToDAGISel::SelectAtomic(SDNode *Node, unsigned Op8, 2365 unsigned Op16,unsigned Op32, 2366 unsigned Op64) { 2367 // Mostly direct translation to the given operations, except that we preserve 2368 // the AtomicOrdering for use later on. 2369 AtomicSDNode *AN = cast<AtomicSDNode>(Node); 2370 EVT VT = AN->getMemoryVT(); 2371 2372 unsigned Op; 2373 SDVTList VTs = CurDAG->getVTList(AN->getValueType(0), MVT::Other); 2374 if (VT == MVT::i8) 2375 Op = Op8; 2376 else if (VT == MVT::i16) 2377 Op = Op16; 2378 else if (VT == MVT::i32) 2379 Op = Op32; 2380 else if (VT == MVT::i64) { 2381 Op = Op64; 2382 VTs = CurDAG->getVTList(MVT::i32, MVT::i32, MVT::Other); 2383 } else 2384 llvm_unreachable("Unexpected atomic operation"); 2385 2386 SmallVector<SDValue, 6> Ops; 2387 for (unsigned i = 1; i < AN->getNumOperands(); ++i) 2388 Ops.push_back(AN->getOperand(i)); 2389 2390 Ops.push_back(CurDAG->getTargetConstant(AN->getOrdering(), MVT::i32)); 2391 Ops.push_back(AN->getOperand(0)); // Chain moves to the end 2392 2393 return CurDAG->SelectNodeTo(Node, Op, VTs, &Ops[0], Ops.size()); 2394} 2395 2396SDNode *ARMDAGToDAGISel::Select(SDNode *N) { 2397 SDLoc dl(N); 2398 2399 if (N->isMachineOpcode()) { 2400 N->setNodeId(-1); 2401 return NULL; // Already selected. 2402 } 2403 2404 switch (N->getOpcode()) { 2405 default: break; 2406 case ISD::INLINEASM: { 2407 SDNode *ResNode = SelectInlineAsm(N); 2408 if (ResNode) 2409 return ResNode; 2410 break; 2411 } 2412 case ISD::XOR: { 2413 // Select special operations if XOR node forms integer ABS pattern 2414 SDNode *ResNode = SelectABSOp(N); 2415 if (ResNode) 2416 return ResNode; 2417 // Other cases are autogenerated. 2418 break; 2419 } 2420 case ISD::Constant: { 2421 unsigned Val = cast<ConstantSDNode>(N)->getZExtValue(); 2422 bool UseCP = true; 2423 if (Subtarget->hasThumb2()) 2424 // Thumb2-aware targets have the MOVT instruction, so all immediates can 2425 // be done with MOV + MOVT, at worst. 2426 UseCP = 0; 2427 else { 2428 if (Subtarget->isThumb()) { 2429 UseCP = (Val > 255 && // MOV 2430 ~Val > 255 && // MOV + MVN 2431 !ARM_AM::isThumbImmShiftedVal(Val)); // MOV + LSL 2432 } else 2433 UseCP = (ARM_AM::getSOImmVal(Val) == -1 && // MOV 2434 ARM_AM::getSOImmVal(~Val) == -1 && // MVN 2435 !ARM_AM::isSOImmTwoPartVal(Val)); // two instrs. 2436 } 2437 2438 if (UseCP) { 2439 SDValue CPIdx = 2440 CurDAG->getTargetConstantPool(ConstantInt::get( 2441 Type::getInt32Ty(*CurDAG->getContext()), Val), 2442 getTargetLowering()->getPointerTy()); 2443 2444 SDNode *ResNode; 2445 if (Subtarget->isThumb1Only()) { 2446 SDValue Pred = getAL(CurDAG); 2447 SDValue PredReg = CurDAG->getRegister(0, MVT::i32); 2448 SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() }; 2449 ResNode = CurDAG->getMachineNode(ARM::tLDRpci, dl, MVT::i32, MVT::Other, 2450 Ops); 2451 } else { 2452 SDValue Ops[] = { 2453 CPIdx, 2454 CurDAG->getTargetConstant(0, MVT::i32), 2455 getAL(CurDAG), 2456 CurDAG->getRegister(0, MVT::i32), 2457 CurDAG->getEntryNode() 2458 }; 2459 ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other, 2460 Ops); 2461 } 2462 ReplaceUses(SDValue(N, 0), SDValue(ResNode, 0)); 2463 return NULL; 2464 } 2465 2466 // Other cases are autogenerated. 2467 break; 2468 } 2469 case ISD::FrameIndex: { 2470 // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm. 2471 int FI = cast<FrameIndexSDNode>(N)->getIndex(); 2472 SDValue TFI = CurDAG->getTargetFrameIndex(FI, 2473 getTargetLowering()->getPointerTy()); 2474 if (Subtarget->isThumb1Only()) { 2475 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32), 2476 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) }; 2477 return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, Ops, 4); 2478 } else { 2479 unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ? 2480 ARM::t2ADDri : ARM::ADDri); 2481 SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32), 2482 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), 2483 CurDAG->getRegister(0, MVT::i32) }; 2484 return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5); 2485 } 2486 } 2487 case ISD::SRL: 2488 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false)) 2489 return I; 2490 break; 2491 case ISD::SRA: 2492 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, true)) 2493 return I; 2494 break; 2495 case ISD::MUL: 2496 if (Subtarget->isThumb1Only()) 2497 break; 2498 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) { 2499 unsigned RHSV = C->getZExtValue(); 2500 if (!RHSV) break; 2501 if (isPowerOf2_32(RHSV-1)) { // 2^n+1? 2502 unsigned ShImm = Log2_32(RHSV-1); 2503 if (ShImm >= 32) 2504 break; 2505 SDValue V = N->getOperand(0); 2506 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm); 2507 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32); 2508 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 2509 if (Subtarget->isThumb()) { 2510 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 }; 2511 return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6); 2512 } else { 2513 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 }; 2514 return CurDAG->SelectNodeTo(N, ARM::ADDrsi, MVT::i32, Ops, 7); 2515 } 2516 } 2517 if (isPowerOf2_32(RHSV+1)) { // 2^n-1? 2518 unsigned ShImm = Log2_32(RHSV+1); 2519 if (ShImm >= 32) 2520 break; 2521 SDValue V = N->getOperand(0); 2522 ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm); 2523 SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32); 2524 SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); 2525 if (Subtarget->isThumb()) { 2526 SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 }; 2527 return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 6); 2528 } else { 2529 SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 }; 2530 return CurDAG->SelectNodeTo(N, ARM::RSBrsi, MVT::i32, Ops, 7); 2531 } 2532 } 2533 } 2534 break; 2535 case ISD::AND: { 2536 // Check for unsigned bitfield extract 2537 if (SDNode *I = SelectV6T2BitfieldExtractOp(N, false)) 2538 return I; 2539 2540 // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits 2541 // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits 2542 // are entirely contributed by c2 and lower 16-bits are entirely contributed 2543 // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)). 2544 // Select it to: "movt x, ((c1 & 0xffff) >> 16) 2545 EVT VT = N->getValueType(0); 2546 if (VT != MVT::i32) 2547 break; 2548 unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2()) 2549 ? ARM::t2MOVTi16 2550 : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0); 2551 if (!Opc) 2552 break; 2553 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); 2554 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2555 if (!N1C) 2556 break; 2557 if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) { 2558 SDValue N2 = N0.getOperand(1); 2559 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2); 2560 if (!N2C) 2561 break; 2562 unsigned N1CVal = N1C->getZExtValue(); 2563 unsigned N2CVal = N2C->getZExtValue(); 2564 if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) && 2565 (N1CVal & 0xffffU) == 0xffffU && 2566 (N2CVal & 0xffffU) == 0x0U) { 2567 SDValue Imm16 = CurDAG->getTargetConstant((N2CVal & 0xFFFF0000U) >> 16, 2568 MVT::i32); 2569 SDValue Ops[] = { N0.getOperand(0), Imm16, 2570 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) }; 2571 return CurDAG->getMachineNode(Opc, dl, VT, Ops); 2572 } 2573 } 2574 break; 2575 } 2576 case ARMISD::VMOVRRD: 2577 return CurDAG->getMachineNode(ARM::VMOVRRD, dl, MVT::i32, MVT::i32, 2578 N->getOperand(0), getAL(CurDAG), 2579 CurDAG->getRegister(0, MVT::i32)); 2580 case ISD::UMUL_LOHI: { 2581 if (Subtarget->isThumb1Only()) 2582 break; 2583 if (Subtarget->isThumb()) { 2584 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), 2585 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) }; 2586 return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops); 2587 } else { 2588 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), 2589 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), 2590 CurDAG->getRegister(0, MVT::i32) }; 2591 return CurDAG->getMachineNode(Subtarget->hasV6Ops() ? 2592 ARM::UMULL : ARM::UMULLv5, 2593 dl, MVT::i32, MVT::i32, Ops); 2594 } 2595 } 2596 case ISD::SMUL_LOHI: { 2597 if (Subtarget->isThumb1Only()) 2598 break; 2599 if (Subtarget->isThumb()) { 2600 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), 2601 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) }; 2602 return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops); 2603 } else { 2604 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), 2605 getAL(CurDAG), CurDAG->getRegister(0, MVT::i32), 2606 CurDAG->getRegister(0, MVT::i32) }; 2607 return CurDAG->getMachineNode(Subtarget->hasV6Ops() ? 2608 ARM::SMULL : ARM::SMULLv5, 2609 dl, MVT::i32, MVT::i32, Ops); 2610 } 2611 } 2612 case ARMISD::UMLAL:{ 2613 if (Subtarget->isThumb()) { 2614 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2), 2615 N->getOperand(3), getAL(CurDAG), 2616 CurDAG->getRegister(0, MVT::i32)}; 2617 return CurDAG->getMachineNode(ARM::t2UMLAL, dl, MVT::i32, MVT::i32, Ops); 2618 }else{ 2619 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2), 2620 N->getOperand(3), getAL(CurDAG), 2621 CurDAG->getRegister(0, MVT::i32), 2622 CurDAG->getRegister(0, MVT::i32) }; 2623 return CurDAG->getMachineNode(Subtarget->hasV6Ops() ? 2624 ARM::UMLAL : ARM::UMLALv5, 2625 dl, MVT::i32, MVT::i32, Ops); 2626 } 2627 } 2628 case ARMISD::SMLAL:{ 2629 if (Subtarget->isThumb()) { 2630 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2), 2631 N->getOperand(3), getAL(CurDAG), 2632 CurDAG->getRegister(0, MVT::i32)}; 2633 return CurDAG->getMachineNode(ARM::t2SMLAL, dl, MVT::i32, MVT::i32, Ops); 2634 }else{ 2635 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2), 2636 N->getOperand(3), getAL(CurDAG), 2637 CurDAG->getRegister(0, MVT::i32), 2638 CurDAG->getRegister(0, MVT::i32) }; 2639 return CurDAG->getMachineNode(Subtarget->hasV6Ops() ? 2640 ARM::SMLAL : ARM::SMLALv5, 2641 dl, MVT::i32, MVT::i32, Ops); 2642 } 2643 } 2644 case ISD::LOAD: { 2645 SDNode *ResNode = 0; 2646 if (Subtarget->isThumb() && Subtarget->hasThumb2()) 2647 ResNode = SelectT2IndexedLoad(N); 2648 else 2649 ResNode = SelectARMIndexedLoad(N); 2650 if (ResNode) 2651 return ResNode; 2652 // Other cases are autogenerated. 2653 break; 2654 } 2655 case ARMISD::BRCOND: { 2656 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc) 2657 // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc) 2658 // Pattern complexity = 6 cost = 1 size = 0 2659 2660 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc) 2661 // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc) 2662 // Pattern complexity = 6 cost = 1 size = 0 2663 2664 // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc) 2665 // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc) 2666 // Pattern complexity = 6 cost = 1 size = 0 2667 2668 unsigned Opc = Subtarget->isThumb() ? 2669 ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc; 2670 SDValue Chain = N->getOperand(0); 2671 SDValue N1 = N->getOperand(1); 2672 SDValue N2 = N->getOperand(2); 2673 SDValue N3 = N->getOperand(3); 2674 SDValue InFlag = N->getOperand(4); 2675 assert(N1.getOpcode() == ISD::BasicBlock); 2676 assert(N2.getOpcode() == ISD::Constant); 2677 assert(N3.getOpcode() == ISD::Register); 2678 2679 SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned) 2680 cast<ConstantSDNode>(N2)->getZExtValue()), 2681 MVT::i32); 2682 SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag }; 2683 SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other, 2684 MVT::Glue, Ops); 2685 Chain = SDValue(ResNode, 0); 2686 if (N->getNumValues() == 2) { 2687 InFlag = SDValue(ResNode, 1); 2688 ReplaceUses(SDValue(N, 1), InFlag); 2689 } 2690 ReplaceUses(SDValue(N, 0), 2691 SDValue(Chain.getNode(), Chain.getResNo())); 2692 return NULL; 2693 } 2694 case ARMISD::VZIP: { 2695 unsigned Opc = 0; 2696 EVT VT = N->getValueType(0); 2697 switch (VT.getSimpleVT().SimpleTy) { 2698 default: return NULL; 2699 case MVT::v8i8: Opc = ARM::VZIPd8; break; 2700 case MVT::v4i16: Opc = ARM::VZIPd16; break; 2701 case MVT::v2f32: 2702 // vzip.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm. 2703 case MVT::v2i32: Opc = ARM::VTRNd32; break; 2704 case MVT::v16i8: Opc = ARM::VZIPq8; break; 2705 case MVT::v8i16: Opc = ARM::VZIPq16; break; 2706 case MVT::v4f32: 2707 case MVT::v4i32: Opc = ARM::VZIPq32; break; 2708 } 2709 SDValue Pred = getAL(CurDAG); 2710 SDValue PredReg = CurDAG->getRegister(0, MVT::i32); 2711 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg }; 2712 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops); 2713 } 2714 case ARMISD::VUZP: { 2715 unsigned Opc = 0; 2716 EVT VT = N->getValueType(0); 2717 switch (VT.getSimpleVT().SimpleTy) { 2718 default: return NULL; 2719 case MVT::v8i8: Opc = ARM::VUZPd8; break; 2720 case MVT::v4i16: Opc = ARM::VUZPd16; break; 2721 case MVT::v2f32: 2722 // vuzp.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm. 2723 case MVT::v2i32: Opc = ARM::VTRNd32; break; 2724 case MVT::v16i8: Opc = ARM::VUZPq8; break; 2725 case MVT::v8i16: Opc = ARM::VUZPq16; break; 2726 case MVT::v4f32: 2727 case MVT::v4i32: Opc = ARM::VUZPq32; break; 2728 } 2729 SDValue Pred = getAL(CurDAG); 2730 SDValue PredReg = CurDAG->getRegister(0, MVT::i32); 2731 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg }; 2732 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops); 2733 } 2734 case ARMISD::VTRN: { 2735 unsigned Opc = 0; 2736 EVT VT = N->getValueType(0); 2737 switch (VT.getSimpleVT().SimpleTy) { 2738 default: return NULL; 2739 case MVT::v8i8: Opc = ARM::VTRNd8; break; 2740 case MVT::v4i16: Opc = ARM::VTRNd16; break; 2741 case MVT::v2f32: 2742 case MVT::v2i32: Opc = ARM::VTRNd32; break; 2743 case MVT::v16i8: Opc = ARM::VTRNq8; break; 2744 case MVT::v8i16: Opc = ARM::VTRNq16; break; 2745 case MVT::v4f32: 2746 case MVT::v4i32: Opc = ARM::VTRNq32; break; 2747 } 2748 SDValue Pred = getAL(CurDAG); 2749 SDValue PredReg = CurDAG->getRegister(0, MVT::i32); 2750 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg }; 2751 return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops); 2752 } 2753 case ARMISD::BUILD_VECTOR: { 2754 EVT VecVT = N->getValueType(0); 2755 EVT EltVT = VecVT.getVectorElementType(); 2756 unsigned NumElts = VecVT.getVectorNumElements(); 2757 if (EltVT == MVT::f64) { 2758 assert(NumElts == 2 && "unexpected type for BUILD_VECTOR"); 2759 return createDRegPairNode(VecVT, N->getOperand(0), N->getOperand(1)); 2760 } 2761 assert(EltVT == MVT::f32 && "unexpected type for BUILD_VECTOR"); 2762 if (NumElts == 2) 2763 return createSRegPairNode(VecVT, N->getOperand(0), N->getOperand(1)); 2764 assert(NumElts == 4 && "unexpected type for BUILD_VECTOR"); 2765 return createQuadSRegsNode(VecVT, N->getOperand(0), N->getOperand(1), 2766 N->getOperand(2), N->getOperand(3)); 2767 } 2768 2769 case ARMISD::VLD2DUP: { 2770 static const uint16_t Opcodes[] = { ARM::VLD2DUPd8, ARM::VLD2DUPd16, 2771 ARM::VLD2DUPd32 }; 2772 return SelectVLDDup(N, false, 2, Opcodes); 2773 } 2774 2775 case ARMISD::VLD3DUP: { 2776 static const uint16_t Opcodes[] = { ARM::VLD3DUPd8Pseudo, 2777 ARM::VLD3DUPd16Pseudo, 2778 ARM::VLD3DUPd32Pseudo }; 2779 return SelectVLDDup(N, false, 3, Opcodes); 2780 } 2781 2782 case ARMISD::VLD4DUP: { 2783 static const uint16_t Opcodes[] = { ARM::VLD4DUPd8Pseudo, 2784 ARM::VLD4DUPd16Pseudo, 2785 ARM::VLD4DUPd32Pseudo }; 2786 return SelectVLDDup(N, false, 4, Opcodes); 2787 } 2788 2789 case ARMISD::VLD2DUP_UPD: { 2790 static const uint16_t Opcodes[] = { ARM::VLD2DUPd8wb_fixed, 2791 ARM::VLD2DUPd16wb_fixed, 2792 ARM::VLD2DUPd32wb_fixed }; 2793 return SelectVLDDup(N, true, 2, Opcodes); 2794 } 2795 2796 case ARMISD::VLD3DUP_UPD: { 2797 static const uint16_t Opcodes[] = { ARM::VLD3DUPd8Pseudo_UPD, 2798 ARM::VLD3DUPd16Pseudo_UPD, 2799 ARM::VLD3DUPd32Pseudo_UPD }; 2800 return SelectVLDDup(N, true, 3, Opcodes); 2801 } 2802 2803 case ARMISD::VLD4DUP_UPD: { 2804 static const uint16_t Opcodes[] = { ARM::VLD4DUPd8Pseudo_UPD, 2805 ARM::VLD4DUPd16Pseudo_UPD, 2806 ARM::VLD4DUPd32Pseudo_UPD }; 2807 return SelectVLDDup(N, true, 4, Opcodes); 2808 } 2809 2810 case ARMISD::VLD1_UPD: { 2811 static const uint16_t DOpcodes[] = { ARM::VLD1d8wb_fixed, 2812 ARM::VLD1d16wb_fixed, 2813 ARM::VLD1d32wb_fixed, 2814 ARM::VLD1d64wb_fixed }; 2815 static const uint16_t QOpcodes[] = { ARM::VLD1q8wb_fixed, 2816 ARM::VLD1q16wb_fixed, 2817 ARM::VLD1q32wb_fixed, 2818 ARM::VLD1q64wb_fixed }; 2819 return SelectVLD(N, true, 1, DOpcodes, QOpcodes, 0); 2820 } 2821 2822 case ARMISD::VLD2_UPD: { 2823 static const uint16_t DOpcodes[] = { ARM::VLD2d8wb_fixed, 2824 ARM::VLD2d16wb_fixed, 2825 ARM::VLD2d32wb_fixed, 2826 ARM::VLD1q64wb_fixed}; 2827 static const uint16_t QOpcodes[] = { ARM::VLD2q8PseudoWB_fixed, 2828 ARM::VLD2q16PseudoWB_fixed, 2829 ARM::VLD2q32PseudoWB_fixed }; 2830 return SelectVLD(N, true, 2, DOpcodes, QOpcodes, 0); 2831 } 2832 2833 case ARMISD::VLD3_UPD: { 2834 static const uint16_t DOpcodes[] = { ARM::VLD3d8Pseudo_UPD, 2835 ARM::VLD3d16Pseudo_UPD, 2836 ARM::VLD3d32Pseudo_UPD, 2837 ARM::VLD1q64wb_fixed}; 2838 static const uint16_t QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD, 2839 ARM::VLD3q16Pseudo_UPD, 2840 ARM::VLD3q32Pseudo_UPD }; 2841 static const uint16_t QOpcodes1[] = { ARM::VLD3q8oddPseudo_UPD, 2842 ARM::VLD3q16oddPseudo_UPD, 2843 ARM::VLD3q32oddPseudo_UPD }; 2844 return SelectVLD(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1); 2845 } 2846 2847 case ARMISD::VLD4_UPD: { 2848 static const uint16_t DOpcodes[] = { ARM::VLD4d8Pseudo_UPD, 2849 ARM::VLD4d16Pseudo_UPD, 2850 ARM::VLD4d32Pseudo_UPD, 2851 ARM::VLD1q64wb_fixed}; 2852 static const uint16_t QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD, 2853 ARM::VLD4q16Pseudo_UPD, 2854 ARM::VLD4q32Pseudo_UPD }; 2855 static const uint16_t QOpcodes1[] = { ARM::VLD4q8oddPseudo_UPD, 2856 ARM::VLD4q16oddPseudo_UPD, 2857 ARM::VLD4q32oddPseudo_UPD }; 2858 return SelectVLD(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1); 2859 } 2860 2861 case ARMISD::VLD2LN_UPD: { 2862 static const uint16_t DOpcodes[] = { ARM::VLD2LNd8Pseudo_UPD, 2863 ARM::VLD2LNd16Pseudo_UPD, 2864 ARM::VLD2LNd32Pseudo_UPD }; 2865 static const uint16_t QOpcodes[] = { ARM::VLD2LNq16Pseudo_UPD, 2866 ARM::VLD2LNq32Pseudo_UPD }; 2867 return SelectVLDSTLane(N, true, true, 2, DOpcodes, QOpcodes); 2868 } 2869 2870 case ARMISD::VLD3LN_UPD: { 2871 static const uint16_t DOpcodes[] = { ARM::VLD3LNd8Pseudo_UPD, 2872 ARM::VLD3LNd16Pseudo_UPD, 2873 ARM::VLD3LNd32Pseudo_UPD }; 2874 static const uint16_t QOpcodes[] = { ARM::VLD3LNq16Pseudo_UPD, 2875 ARM::VLD3LNq32Pseudo_UPD }; 2876 return SelectVLDSTLane(N, true, true, 3, DOpcodes, QOpcodes); 2877 } 2878 2879 case ARMISD::VLD4LN_UPD: { 2880 static const uint16_t DOpcodes[] = { ARM::VLD4LNd8Pseudo_UPD, 2881 ARM::VLD4LNd16Pseudo_UPD, 2882 ARM::VLD4LNd32Pseudo_UPD }; 2883 static const uint16_t QOpcodes[] = { ARM::VLD4LNq16Pseudo_UPD, 2884 ARM::VLD4LNq32Pseudo_UPD }; 2885 return SelectVLDSTLane(N, true, true, 4, DOpcodes, QOpcodes); 2886 } 2887 2888 case ARMISD::VST1_UPD: { 2889 static const uint16_t DOpcodes[] = { ARM::VST1d8wb_fixed, 2890 ARM::VST1d16wb_fixed, 2891 ARM::VST1d32wb_fixed, 2892 ARM::VST1d64wb_fixed }; 2893 static const uint16_t QOpcodes[] = { ARM::VST1q8wb_fixed, 2894 ARM::VST1q16wb_fixed, 2895 ARM::VST1q32wb_fixed, 2896 ARM::VST1q64wb_fixed }; 2897 return SelectVST(N, true, 1, DOpcodes, QOpcodes, 0); 2898 } 2899 2900 case ARMISD::VST2_UPD: { 2901 static const uint16_t DOpcodes[] = { ARM::VST2d8wb_fixed, 2902 ARM::VST2d16wb_fixed, 2903 ARM::VST2d32wb_fixed, 2904 ARM::VST1q64wb_fixed}; 2905 static const uint16_t QOpcodes[] = { ARM::VST2q8PseudoWB_fixed, 2906 ARM::VST2q16PseudoWB_fixed, 2907 ARM::VST2q32PseudoWB_fixed }; 2908 return SelectVST(N, true, 2, DOpcodes, QOpcodes, 0); 2909 } 2910 2911 case ARMISD::VST3_UPD: { 2912 static const uint16_t DOpcodes[] = { ARM::VST3d8Pseudo_UPD, 2913 ARM::VST3d16Pseudo_UPD, 2914 ARM::VST3d32Pseudo_UPD, 2915 ARM::VST1d64TPseudoWB_fixed}; 2916 static const uint16_t QOpcodes0[] = { ARM::VST3q8Pseudo_UPD, 2917 ARM::VST3q16Pseudo_UPD, 2918 ARM::VST3q32Pseudo_UPD }; 2919 static const uint16_t QOpcodes1[] = { ARM::VST3q8oddPseudo_UPD, 2920 ARM::VST3q16oddPseudo_UPD, 2921 ARM::VST3q32oddPseudo_UPD }; 2922 return SelectVST(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1); 2923 } 2924 2925 case ARMISD::VST4_UPD: { 2926 static const uint16_t DOpcodes[] = { ARM::VST4d8Pseudo_UPD, 2927 ARM::VST4d16Pseudo_UPD, 2928 ARM::VST4d32Pseudo_UPD, 2929 ARM::VST1d64QPseudoWB_fixed}; 2930 static const uint16_t QOpcodes0[] = { ARM::VST4q8Pseudo_UPD, 2931 ARM::VST4q16Pseudo_UPD, 2932 ARM::VST4q32Pseudo_UPD }; 2933 static const uint16_t QOpcodes1[] = { ARM::VST4q8oddPseudo_UPD, 2934 ARM::VST4q16oddPseudo_UPD, 2935 ARM::VST4q32oddPseudo_UPD }; 2936 return SelectVST(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1); 2937 } 2938 2939 case ARMISD::VST2LN_UPD: { 2940 static const uint16_t DOpcodes[] = { ARM::VST2LNd8Pseudo_UPD, 2941 ARM::VST2LNd16Pseudo_UPD, 2942 ARM::VST2LNd32Pseudo_UPD }; 2943 static const uint16_t QOpcodes[] = { ARM::VST2LNq16Pseudo_UPD, 2944 ARM::VST2LNq32Pseudo_UPD }; 2945 return SelectVLDSTLane(N, false, true, 2, DOpcodes, QOpcodes); 2946 } 2947 2948 case ARMISD::VST3LN_UPD: { 2949 static const uint16_t DOpcodes[] = { ARM::VST3LNd8Pseudo_UPD, 2950 ARM::VST3LNd16Pseudo_UPD, 2951 ARM::VST3LNd32Pseudo_UPD }; 2952 static const uint16_t QOpcodes[] = { ARM::VST3LNq16Pseudo_UPD, 2953 ARM::VST3LNq32Pseudo_UPD }; 2954 return SelectVLDSTLane(N, false, true, 3, DOpcodes, QOpcodes); 2955 } 2956 2957 case ARMISD::VST4LN_UPD: { 2958 static const uint16_t DOpcodes[] = { ARM::VST4LNd8Pseudo_UPD, 2959 ARM::VST4LNd16Pseudo_UPD, 2960 ARM::VST4LNd32Pseudo_UPD }; 2961 static const uint16_t QOpcodes[] = { ARM::VST4LNq16Pseudo_UPD, 2962 ARM::VST4LNq32Pseudo_UPD }; 2963 return SelectVLDSTLane(N, false, true, 4, DOpcodes, QOpcodes); 2964 } 2965 2966 case ISD::INTRINSIC_VOID: 2967 case ISD::INTRINSIC_W_CHAIN: { 2968 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 2969 switch (IntNo) { 2970 default: 2971 break; 2972 2973 case Intrinsic::arm_ldrexd: { 2974 SDValue MemAddr = N->getOperand(2); 2975 SDLoc dl(N); 2976 SDValue Chain = N->getOperand(0); 2977 2978 bool isThumb = Subtarget->isThumb() && Subtarget->hasThumb2(); 2979 unsigned NewOpc = isThumb ? ARM::t2LDREXD :ARM::LDREXD; 2980 2981 // arm_ldrexd returns a i64 value in {i32, i32} 2982 std::vector<EVT> ResTys; 2983 if (isThumb) { 2984 ResTys.push_back(MVT::i32); 2985 ResTys.push_back(MVT::i32); 2986 } else 2987 ResTys.push_back(MVT::Untyped); 2988 ResTys.push_back(MVT::Other); 2989 2990 // Place arguments in the right order. 2991 SmallVector<SDValue, 7> Ops; 2992 Ops.push_back(MemAddr); 2993 Ops.push_back(getAL(CurDAG)); 2994 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); 2995 Ops.push_back(Chain); 2996 SDNode *Ld = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops); 2997 // Transfer memoperands. 2998 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 2999 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 3000 cast<MachineSDNode>(Ld)->setMemRefs(MemOp, MemOp + 1); 3001 3002 // Remap uses. 3003 SDValue OutChain = isThumb ? SDValue(Ld, 2) : SDValue(Ld, 1); 3004 if (!SDValue(N, 0).use_empty()) { 3005 SDValue Result; 3006 if (isThumb) 3007 Result = SDValue(Ld, 0); 3008 else { 3009 SDValue SubRegIdx = CurDAG->getTargetConstant(ARM::gsub_0, MVT::i32); 3010 SDNode *ResNode = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 3011 dl, MVT::i32, SDValue(Ld, 0), SubRegIdx); 3012 Result = SDValue(ResNode,0); 3013 } 3014 ReplaceUses(SDValue(N, 0), Result); 3015 } 3016 if (!SDValue(N, 1).use_empty()) { 3017 SDValue Result; 3018 if (isThumb) 3019 Result = SDValue(Ld, 1); 3020 else { 3021 SDValue SubRegIdx = CurDAG->getTargetConstant(ARM::gsub_1, MVT::i32); 3022 SDNode *ResNode = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG, 3023 dl, MVT::i32, SDValue(Ld, 0), SubRegIdx); 3024 Result = SDValue(ResNode,0); 3025 } 3026 ReplaceUses(SDValue(N, 1), Result); 3027 } 3028 ReplaceUses(SDValue(N, 2), OutChain); 3029 return NULL; 3030 } 3031 3032 case Intrinsic::arm_strexd: { 3033 SDLoc dl(N); 3034 SDValue Chain = N->getOperand(0); 3035 SDValue Val0 = N->getOperand(2); 3036 SDValue Val1 = N->getOperand(3); 3037 SDValue MemAddr = N->getOperand(4); 3038 3039 // Store exclusive double return a i32 value which is the return status 3040 // of the issued store. 3041 EVT ResTys[] = { MVT::i32, MVT::Other }; 3042 3043 bool isThumb = Subtarget->isThumb() && Subtarget->hasThumb2(); 3044 // Place arguments in the right order. 3045 SmallVector<SDValue, 7> Ops; 3046 if (isThumb) { 3047 Ops.push_back(Val0); 3048 Ops.push_back(Val1); 3049 } else 3050 // arm_strexd uses GPRPair. 3051 Ops.push_back(SDValue(createGPRPairNode(MVT::Untyped, Val0, Val1), 0)); 3052 Ops.push_back(MemAddr); 3053 Ops.push_back(getAL(CurDAG)); 3054 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); 3055 Ops.push_back(Chain); 3056 3057 unsigned NewOpc = isThumb ? ARM::t2STREXD : ARM::STREXD; 3058 3059 SDNode *St = CurDAG->getMachineNode(NewOpc, dl, ResTys, Ops); 3060 // Transfer memoperands. 3061 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); 3062 MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand(); 3063 cast<MachineSDNode>(St)->setMemRefs(MemOp, MemOp + 1); 3064 3065 return St; 3066 } 3067 3068 case Intrinsic::arm_neon_vld1: { 3069 static const uint16_t DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16, 3070 ARM::VLD1d32, ARM::VLD1d64 }; 3071 static const uint16_t QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16, 3072 ARM::VLD1q32, ARM::VLD1q64}; 3073 return SelectVLD(N, false, 1, DOpcodes, QOpcodes, 0); 3074 } 3075 3076 case Intrinsic::arm_neon_vld2: { 3077 static const uint16_t DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16, 3078 ARM::VLD2d32, ARM::VLD1q64 }; 3079 static const uint16_t QOpcodes[] = { ARM::VLD2q8Pseudo, ARM::VLD2q16Pseudo, 3080 ARM::VLD2q32Pseudo }; 3081 return SelectVLD(N, false, 2, DOpcodes, QOpcodes, 0); 3082 } 3083 3084 case Intrinsic::arm_neon_vld3: { 3085 static const uint16_t DOpcodes[] = { ARM::VLD3d8Pseudo, 3086 ARM::VLD3d16Pseudo, 3087 ARM::VLD3d32Pseudo, 3088 ARM::VLD1d64TPseudo }; 3089 static const uint16_t QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD, 3090 ARM::VLD3q16Pseudo_UPD, 3091 ARM::VLD3q32Pseudo_UPD }; 3092 static const uint16_t QOpcodes1[] = { ARM::VLD3q8oddPseudo, 3093 ARM::VLD3q16oddPseudo, 3094 ARM::VLD3q32oddPseudo }; 3095 return SelectVLD(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1); 3096 } 3097 3098 case Intrinsic::arm_neon_vld4: { 3099 static const uint16_t DOpcodes[] = { ARM::VLD4d8Pseudo, 3100 ARM::VLD4d16Pseudo, 3101 ARM::VLD4d32Pseudo, 3102 ARM::VLD1d64QPseudo }; 3103 static const uint16_t QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD, 3104 ARM::VLD4q16Pseudo_UPD, 3105 ARM::VLD4q32Pseudo_UPD }; 3106 static const uint16_t QOpcodes1[] = { ARM::VLD4q8oddPseudo, 3107 ARM::VLD4q16oddPseudo, 3108 ARM::VLD4q32oddPseudo }; 3109 return SelectVLD(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1); 3110 } 3111 3112 case Intrinsic::arm_neon_vld2lane: { 3113 static const uint16_t DOpcodes[] = { ARM::VLD2LNd8Pseudo, 3114 ARM::VLD2LNd16Pseudo, 3115 ARM::VLD2LNd32Pseudo }; 3116 static const uint16_t QOpcodes[] = { ARM::VLD2LNq16Pseudo, 3117 ARM::VLD2LNq32Pseudo }; 3118 return SelectVLDSTLane(N, true, false, 2, DOpcodes, QOpcodes); 3119 } 3120 3121 case Intrinsic::arm_neon_vld3lane: { 3122 static const uint16_t DOpcodes[] = { ARM::VLD3LNd8Pseudo, 3123 ARM::VLD3LNd16Pseudo, 3124 ARM::VLD3LNd32Pseudo }; 3125 static const uint16_t QOpcodes[] = { ARM::VLD3LNq16Pseudo, 3126 ARM::VLD3LNq32Pseudo }; 3127 return SelectVLDSTLane(N, true, false, 3, DOpcodes, QOpcodes); 3128 } 3129 3130 case Intrinsic::arm_neon_vld4lane: { 3131 static const uint16_t DOpcodes[] = { ARM::VLD4LNd8Pseudo, 3132 ARM::VLD4LNd16Pseudo, 3133 ARM::VLD4LNd32Pseudo }; 3134 static const uint16_t QOpcodes[] = { ARM::VLD4LNq16Pseudo, 3135 ARM::VLD4LNq32Pseudo }; 3136 return SelectVLDSTLane(N, true, false, 4, DOpcodes, QOpcodes); 3137 } 3138 3139 case Intrinsic::arm_neon_vst1: { 3140 static const uint16_t DOpcodes[] = { ARM::VST1d8, ARM::VST1d16, 3141 ARM::VST1d32, ARM::VST1d64 }; 3142 static const uint16_t QOpcodes[] = { ARM::VST1q8, ARM::VST1q16, 3143 ARM::VST1q32, ARM::VST1q64 }; 3144 return SelectVST(N, false, 1, DOpcodes, QOpcodes, 0); 3145 } 3146 3147 case Intrinsic::arm_neon_vst2: { 3148 static const uint16_t DOpcodes[] = { ARM::VST2d8, ARM::VST2d16, 3149 ARM::VST2d32, ARM::VST1q64 }; 3150 static uint16_t QOpcodes[] = { ARM::VST2q8Pseudo, ARM::VST2q16Pseudo, 3151 ARM::VST2q32Pseudo }; 3152 return SelectVST(N, false, 2, DOpcodes, QOpcodes, 0); 3153 } 3154 3155 case Intrinsic::arm_neon_vst3: { 3156 static const uint16_t DOpcodes[] = { ARM::VST3d8Pseudo, 3157 ARM::VST3d16Pseudo, 3158 ARM::VST3d32Pseudo, 3159 ARM::VST1d64TPseudo }; 3160 static const uint16_t QOpcodes0[] = { ARM::VST3q8Pseudo_UPD, 3161 ARM::VST3q16Pseudo_UPD, 3162 ARM::VST3q32Pseudo_UPD }; 3163 static const uint16_t QOpcodes1[] = { ARM::VST3q8oddPseudo, 3164 ARM::VST3q16oddPseudo, 3165 ARM::VST3q32oddPseudo }; 3166 return SelectVST(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1); 3167 } 3168 3169 case Intrinsic::arm_neon_vst4: { 3170 static const uint16_t DOpcodes[] = { ARM::VST4d8Pseudo, 3171 ARM::VST4d16Pseudo, 3172 ARM::VST4d32Pseudo, 3173 ARM::VST1d64QPseudo }; 3174 static const uint16_t QOpcodes0[] = { ARM::VST4q8Pseudo_UPD, 3175 ARM::VST4q16Pseudo_UPD, 3176 ARM::VST4q32Pseudo_UPD }; 3177 static const uint16_t QOpcodes1[] = { ARM::VST4q8oddPseudo, 3178 ARM::VST4q16oddPseudo, 3179 ARM::VST4q32oddPseudo }; 3180 return SelectVST(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1); 3181 } 3182 3183 case Intrinsic::arm_neon_vst2lane: { 3184 static const uint16_t DOpcodes[] = { ARM::VST2LNd8Pseudo, 3185 ARM::VST2LNd16Pseudo, 3186 ARM::VST2LNd32Pseudo }; 3187 static const uint16_t QOpcodes[] = { ARM::VST2LNq16Pseudo, 3188 ARM::VST2LNq32Pseudo }; 3189 return SelectVLDSTLane(N, false, false, 2, DOpcodes, QOpcodes); 3190 } 3191 3192 case Intrinsic::arm_neon_vst3lane: { 3193 static const uint16_t DOpcodes[] = { ARM::VST3LNd8Pseudo, 3194 ARM::VST3LNd16Pseudo, 3195 ARM::VST3LNd32Pseudo }; 3196 static const uint16_t QOpcodes[] = { ARM::VST3LNq16Pseudo, 3197 ARM::VST3LNq32Pseudo }; 3198 return SelectVLDSTLane(N, false, false, 3, DOpcodes, QOpcodes); 3199 } 3200 3201 case Intrinsic::arm_neon_vst4lane: { 3202 static const uint16_t DOpcodes[] = { ARM::VST4LNd8Pseudo, 3203 ARM::VST4LNd16Pseudo, 3204 ARM::VST4LNd32Pseudo }; 3205 static const uint16_t QOpcodes[] = { ARM::VST4LNq16Pseudo, 3206 ARM::VST4LNq32Pseudo }; 3207 return SelectVLDSTLane(N, false, false, 4, DOpcodes, QOpcodes); 3208 } 3209 } 3210 break; 3211 } 3212 3213 case ISD::INTRINSIC_WO_CHAIN: { 3214 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); 3215 switch (IntNo) { 3216 default: 3217 break; 3218 3219 case Intrinsic::arm_neon_vtbl2: 3220 return SelectVTBL(N, false, 2, ARM::VTBL2); 3221 case Intrinsic::arm_neon_vtbl3: 3222 return SelectVTBL(N, false, 3, ARM::VTBL3Pseudo); 3223 case Intrinsic::arm_neon_vtbl4: 3224 return SelectVTBL(N, false, 4, ARM::VTBL4Pseudo); 3225 3226 case Intrinsic::arm_neon_vtbx2: 3227 return SelectVTBL(N, true, 2, ARM::VTBX2); 3228 case Intrinsic::arm_neon_vtbx3: 3229 return SelectVTBL(N, true, 3, ARM::VTBX3Pseudo); 3230 case Intrinsic::arm_neon_vtbx4: 3231 return SelectVTBL(N, true, 4, ARM::VTBX4Pseudo); 3232 } 3233 break; 3234 } 3235 3236 case ARMISD::VTBL1: { 3237 SDLoc dl(N); 3238 EVT VT = N->getValueType(0); 3239 SmallVector<SDValue, 6> Ops; 3240 3241 Ops.push_back(N->getOperand(0)); 3242 Ops.push_back(N->getOperand(1)); 3243 Ops.push_back(getAL(CurDAG)); // Predicate 3244 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // Predicate Register 3245 return CurDAG->getMachineNode(ARM::VTBL1, dl, VT, Ops); 3246 } 3247 case ARMISD::VTBL2: { 3248 SDLoc dl(N); 3249 EVT VT = N->getValueType(0); 3250 3251 // Form a REG_SEQUENCE to force register allocation. 3252 SDValue V0 = N->getOperand(0); 3253 SDValue V1 = N->getOperand(1); 3254 SDValue RegSeq = SDValue(createDRegPairNode(MVT::v16i8, V0, V1), 0); 3255 3256 SmallVector<SDValue, 6> Ops; 3257 Ops.push_back(RegSeq); 3258 Ops.push_back(N->getOperand(2)); 3259 Ops.push_back(getAL(CurDAG)); // Predicate 3260 Ops.push_back(CurDAG->getRegister(0, MVT::i32)); // Predicate Register 3261 return CurDAG->getMachineNode(ARM::VTBL2, dl, VT, Ops); 3262 } 3263 3264 case ISD::CONCAT_VECTORS: 3265 return SelectConcatVector(N); 3266 3267 case ISD::ATOMIC_LOAD: 3268 if (cast<AtomicSDNode>(N)->getMemoryVT() == MVT::i64) 3269 return SelectAtomic(N, 0, 0, 0, ARM::ATOMIC_LOAD_I64); 3270 else 3271 break; 3272 3273 case ISD::ATOMIC_STORE: 3274 if (cast<AtomicSDNode>(N)->getMemoryVT() == MVT::i64) 3275 return SelectAtomic(N, 0, 0, 0, ARM::ATOMIC_STORE_I64); 3276 else 3277 break; 3278 3279 case ISD::ATOMIC_LOAD_ADD: 3280 return SelectAtomic(N, 3281 ARM::ATOMIC_LOAD_ADD_I8, 3282 ARM::ATOMIC_LOAD_ADD_I16, 3283 ARM::ATOMIC_LOAD_ADD_I32, 3284 ARM::ATOMIC_LOAD_ADD_I64); 3285 case ISD::ATOMIC_LOAD_SUB: 3286 return SelectAtomic(N, 3287 ARM::ATOMIC_LOAD_SUB_I8, 3288 ARM::ATOMIC_LOAD_SUB_I16, 3289 ARM::ATOMIC_LOAD_SUB_I32, 3290 ARM::ATOMIC_LOAD_SUB_I64); 3291 case ISD::ATOMIC_LOAD_AND: 3292 return SelectAtomic(N, 3293 ARM::ATOMIC_LOAD_AND_I8, 3294 ARM::ATOMIC_LOAD_AND_I16, 3295 ARM::ATOMIC_LOAD_AND_I32, 3296 ARM::ATOMIC_LOAD_AND_I64); 3297 case ISD::ATOMIC_LOAD_OR: 3298 return SelectAtomic(N, 3299 ARM::ATOMIC_LOAD_OR_I8, 3300 ARM::ATOMIC_LOAD_OR_I16, 3301 ARM::ATOMIC_LOAD_OR_I32, 3302 ARM::ATOMIC_LOAD_OR_I64); 3303 case ISD::ATOMIC_LOAD_XOR: 3304 return SelectAtomic(N, 3305 ARM::ATOMIC_LOAD_XOR_I8, 3306 ARM::ATOMIC_LOAD_XOR_I16, 3307 ARM::ATOMIC_LOAD_XOR_I32, 3308 ARM::ATOMIC_LOAD_XOR_I64); 3309 case ISD::ATOMIC_LOAD_NAND: 3310 return SelectAtomic(N, 3311 ARM::ATOMIC_LOAD_NAND_I8, 3312 ARM::ATOMIC_LOAD_NAND_I16, 3313 ARM::ATOMIC_LOAD_NAND_I32, 3314 ARM::ATOMIC_LOAD_NAND_I64); 3315 case ISD::ATOMIC_LOAD_MIN: 3316 return SelectAtomic(N, 3317 ARM::ATOMIC_LOAD_MIN_I8, 3318 ARM::ATOMIC_LOAD_MIN_I16, 3319 ARM::ATOMIC_LOAD_MIN_I32, 3320 ARM::ATOMIC_LOAD_MIN_I64); 3321 case ISD::ATOMIC_LOAD_MAX: 3322 return SelectAtomic(N, 3323 ARM::ATOMIC_LOAD_MAX_I8, 3324 ARM::ATOMIC_LOAD_MAX_I16, 3325 ARM::ATOMIC_LOAD_MAX_I32, 3326 ARM::ATOMIC_LOAD_MAX_I64); 3327 case ISD::ATOMIC_LOAD_UMIN: 3328 return SelectAtomic(N, 3329 ARM::ATOMIC_LOAD_UMIN_I8, 3330 ARM::ATOMIC_LOAD_UMIN_I16, 3331 ARM::ATOMIC_LOAD_UMIN_I32, 3332 ARM::ATOMIC_LOAD_UMIN_I64); 3333 case ISD::ATOMIC_LOAD_UMAX: 3334 return SelectAtomic(N, 3335 ARM::ATOMIC_LOAD_UMAX_I8, 3336 ARM::ATOMIC_LOAD_UMAX_I16, 3337 ARM::ATOMIC_LOAD_UMAX_I32, 3338 ARM::ATOMIC_LOAD_UMAX_I64); 3339 case ISD::ATOMIC_SWAP: 3340 return SelectAtomic(N, 3341 ARM::ATOMIC_SWAP_I8, 3342 ARM::ATOMIC_SWAP_I16, 3343 ARM::ATOMIC_SWAP_I32, 3344 ARM::ATOMIC_SWAP_I64); 3345 case ISD::ATOMIC_CMP_SWAP: 3346 return SelectAtomic(N, 3347 ARM::ATOMIC_CMP_SWAP_I8, 3348 ARM::ATOMIC_CMP_SWAP_I16, 3349 ARM::ATOMIC_CMP_SWAP_I32, 3350 ARM::ATOMIC_CMP_SWAP_I64); 3351 } 3352 3353 return SelectCode(N); 3354} 3355 3356SDNode *ARMDAGToDAGISel::SelectInlineAsm(SDNode *N){ 3357 std::vector<SDValue> AsmNodeOperands; 3358 unsigned Flag, Kind; 3359 bool Changed = false; 3360 unsigned NumOps = N->getNumOperands(); 3361 3362 // Normally, i64 data is bounded to two arbitrary GRPs for "%r" constraint. 3363 // However, some instrstions (e.g. ldrexd/strexd in ARM mode) require 3364 // (even/even+1) GPRs and use %n and %Hn to refer to the individual regs 3365 // respectively. Since there is no constraint to explicitly specify a 3366 // reg pair, we use GPRPair reg class for "%r" for 64-bit data. For Thumb, 3367 // the 64-bit data may be referred by H, Q, R modifiers, so we still pack 3368 // them into a GPRPair. 3369 3370 SDLoc dl(N); 3371 SDValue Glue = N->getGluedNode() ? N->getOperand(NumOps-1) : SDValue(0,0); 3372 3373 SmallVector<bool, 8> OpChanged; 3374 // Glue node will be appended late. 3375 for(unsigned i = 0, e = N->getGluedNode() ? NumOps - 1 : NumOps; i < e; ++i) { 3376 SDValue op = N->getOperand(i); 3377 AsmNodeOperands.push_back(op); 3378 3379 if (i < InlineAsm::Op_FirstOperand) 3380 continue; 3381 3382 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(i))) { 3383 Flag = C->getZExtValue(); 3384 Kind = InlineAsm::getKind(Flag); 3385 } 3386 else 3387 continue; 3388 3389 // Immediate operands to inline asm in the SelectionDAG are modeled with 3390 // two operands. The first is a constant of value InlineAsm::Kind_Imm, and 3391 // the second is a constant with the value of the immediate. If we get here 3392 // and we have a Kind_Imm, skip the next operand, and continue. 3393 if (Kind == InlineAsm::Kind_Imm) { 3394 SDValue op = N->getOperand(++i); 3395 AsmNodeOperands.push_back(op); 3396 continue; 3397 } 3398 3399 unsigned NumRegs = InlineAsm::getNumOperandRegisters(Flag); 3400 if (NumRegs) 3401 OpChanged.push_back(false); 3402 3403 unsigned DefIdx = 0; 3404 bool IsTiedToChangedOp = false; 3405 // If it's a use that is tied with a previous def, it has no 3406 // reg class constraint. 3407 if (Changed && InlineAsm::isUseOperandTiedToDef(Flag, DefIdx)) 3408 IsTiedToChangedOp = OpChanged[DefIdx]; 3409 3410 if (Kind != InlineAsm::Kind_RegUse && Kind != InlineAsm::Kind_RegDef 3411 && Kind != InlineAsm::Kind_RegDefEarlyClobber) 3412 continue; 3413 3414 unsigned RC; 3415 bool HasRC = InlineAsm::hasRegClassConstraint(Flag, RC); 3416 if ((!IsTiedToChangedOp && (!HasRC || RC != ARM::GPRRegClassID)) 3417 || NumRegs != 2) 3418 continue; 3419 3420 assert((i+2 < NumOps) && "Invalid number of operands in inline asm"); 3421 SDValue V0 = N->getOperand(i+1); 3422 SDValue V1 = N->getOperand(i+2); 3423 unsigned Reg0 = cast<RegisterSDNode>(V0)->getReg(); 3424 unsigned Reg1 = cast<RegisterSDNode>(V1)->getReg(); 3425 SDValue PairedReg; 3426 MachineRegisterInfo &MRI = MF->getRegInfo(); 3427 3428 if (Kind == InlineAsm::Kind_RegDef || 3429 Kind == InlineAsm::Kind_RegDefEarlyClobber) { 3430 // Replace the two GPRs with 1 GPRPair and copy values from GPRPair to 3431 // the original GPRs. 3432 3433 unsigned GPVR = MRI.createVirtualRegister(&ARM::GPRPairRegClass); 3434 PairedReg = CurDAG->getRegister(GPVR, MVT::Untyped); 3435 SDValue Chain = SDValue(N,0); 3436 3437 SDNode *GU = N->getGluedUser(); 3438 SDValue RegCopy = CurDAG->getCopyFromReg(Chain, dl, GPVR, MVT::Untyped, 3439 Chain.getValue(1)); 3440 3441 // Extract values from a GPRPair reg and copy to the original GPR reg. 3442 SDValue Sub0 = CurDAG->getTargetExtractSubreg(ARM::gsub_0, dl, MVT::i32, 3443 RegCopy); 3444 SDValue Sub1 = CurDAG->getTargetExtractSubreg(ARM::gsub_1, dl, MVT::i32, 3445 RegCopy); 3446 SDValue T0 = CurDAG->getCopyToReg(Sub0, dl, Reg0, Sub0, 3447 RegCopy.getValue(1)); 3448 SDValue T1 = CurDAG->getCopyToReg(Sub1, dl, Reg1, Sub1, T0.getValue(1)); 3449 3450 // Update the original glue user. 3451 std::vector<SDValue> Ops(GU->op_begin(), GU->op_end()-1); 3452 Ops.push_back(T1.getValue(1)); 3453 CurDAG->UpdateNodeOperands(GU, &Ops[0], Ops.size()); 3454 GU = T1.getNode(); 3455 } 3456 else { 3457 // For Kind == InlineAsm::Kind_RegUse, we first copy two GPRs into a 3458 // GPRPair and then pass the GPRPair to the inline asm. 3459 SDValue Chain = AsmNodeOperands[InlineAsm::Op_InputChain]; 3460 3461 // As REG_SEQ doesn't take RegisterSDNode, we copy them first. 3462 SDValue T0 = CurDAG->getCopyFromReg(Chain, dl, Reg0, MVT::i32, 3463 Chain.getValue(1)); 3464 SDValue T1 = CurDAG->getCopyFromReg(Chain, dl, Reg1, MVT::i32, 3465 T0.getValue(1)); 3466 SDValue Pair = SDValue(createGPRPairNode(MVT::Untyped, T0, T1), 0); 3467 3468 // Copy REG_SEQ into a GPRPair-typed VR and replace the original two 3469 // i32 VRs of inline asm with it. 3470 unsigned GPVR = MRI.createVirtualRegister(&ARM::GPRPairRegClass); 3471 PairedReg = CurDAG->getRegister(GPVR, MVT::Untyped); 3472 Chain = CurDAG->getCopyToReg(T1, dl, GPVR, Pair, T1.getValue(1)); 3473 3474 AsmNodeOperands[InlineAsm::Op_InputChain] = Chain; 3475 Glue = Chain.getValue(1); 3476 } 3477 3478 Changed = true; 3479 3480 if(PairedReg.getNode()) { 3481 OpChanged[OpChanged.size() -1 ] = true; 3482 Flag = InlineAsm::getFlagWord(Kind, 1 /* RegNum*/); 3483 if (IsTiedToChangedOp) 3484 Flag = InlineAsm::getFlagWordForMatchingOp(Flag, DefIdx); 3485 else 3486 Flag = InlineAsm::getFlagWordForRegClass(Flag, ARM::GPRPairRegClassID); 3487 // Replace the current flag. 3488 AsmNodeOperands[AsmNodeOperands.size() -1] = CurDAG->getTargetConstant( 3489 Flag, MVT::i32); 3490 // Add the new register node and skip the original two GPRs. 3491 AsmNodeOperands.push_back(PairedReg); 3492 // Skip the next two GPRs. 3493 i += 2; 3494 } 3495 } 3496 3497 if (Glue.getNode()) 3498 AsmNodeOperands.push_back(Glue); 3499 if (!Changed) 3500 return NULL; 3501 3502 SDValue New = CurDAG->getNode(ISD::INLINEASM, SDLoc(N), 3503 CurDAG->getVTList(MVT::Other, MVT::Glue), &AsmNodeOperands[0], 3504 AsmNodeOperands.size()); 3505 New->setNodeId(-1); 3506 return New.getNode(); 3507} 3508 3509 3510bool ARMDAGToDAGISel:: 3511SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode, 3512 std::vector<SDValue> &OutOps) { 3513 assert(ConstraintCode == 'm' && "unexpected asm memory constraint"); 3514 // Require the address to be in a register. That is safe for all ARM 3515 // variants and it is hard to do anything much smarter without knowing 3516 // how the operand is used. 3517 OutOps.push_back(Op); 3518 return false; 3519} 3520 3521/// createARMISelDag - This pass converts a legalized DAG into a 3522/// ARM-specific DAG, ready for instruction scheduling. 3523/// 3524FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM, 3525 CodeGenOpt::Level OptLevel) { 3526 return new ARMDAGToDAGISel(TM, OptLevel); 3527} 3528