1 //===- HexagonInstrInfo.cpp - Hexagon Instruction Information -------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains the Hexagon implementation of the TargetInstrInfo class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "HexagonInstrInfo.h" 15 #include "Hexagon.h" 16 #include "HexagonFrameLowering.h" 17 #include "HexagonHazardRecognizer.h" 18 #include "HexagonRegisterInfo.h" 19 #include "HexagonSubtarget.h" 20 #include "llvm/ADT/ArrayRef.h" 21 #include "llvm/ADT/SmallPtrSet.h" 22 #include "llvm/ADT/SmallVector.h" 23 #include "llvm/ADT/StringRef.h" 24 #include "llvm/CodeGen/DFAPacketizer.h" 25 #include "llvm/CodeGen/LivePhysRegs.h" 26 #include "llvm/CodeGen/MachineBasicBlock.h" 27 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 28 #include "llvm/CodeGen/MachineFrameInfo.h" 29 #include "llvm/CodeGen/MachineFunction.h" 30 #include "llvm/CodeGen/MachineInstr.h" 31 #include "llvm/CodeGen/MachineInstrBuilder.h" 32 #include "llvm/CodeGen/MachineInstrBundle.h" 33 #include "llvm/CodeGen/MachineLoopInfo.h" 34 #include "llvm/CodeGen/MachineMemOperand.h" 35 #include "llvm/CodeGen/MachineOperand.h" 36 #include "llvm/CodeGen/MachineRegisterInfo.h" 37 #include "llvm/CodeGen/MachineValueType.h" 38 #include "llvm/CodeGen/ScheduleDAG.h" 39 #include "llvm/IR/DebugLoc.h" 40 #include "llvm/MC/MCAsmInfo.h" 41 #include "llvm/MC/MCInstrDesc.h" 42 #include "llvm/MC/MCInstrItineraries.h" 43 #include "llvm/MC/MCRegisterInfo.h" 44 #include "llvm/Support/BranchProbability.h" 45 #include "llvm/Support/CommandLine.h" 46 #include "llvm/Support/Debug.h" 47 #include "llvm/Support/ErrorHandling.h" 48 #include "llvm/Support/MathExtras.h" 49 #include "llvm/Support/raw_ostream.h" 50 #include "llvm/Target/TargetInstrInfo.h" 51 #include "llvm/Target/TargetMachine.h" 52 #include "llvm/Target/TargetOpcodes.h" 53 #include "llvm/Target/TargetRegisterInfo.h" 54 #include "llvm/Target/TargetSubtargetInfo.h" 55 #include <cassert> 56 #include <cctype> 57 #include <cstdint> 58 #include <cstring> 59 #include <iterator> 60 #include <string> 61 #include <utility> 62 63 using namespace llvm; 64 65 #define DEBUG_TYPE "hexagon-instrinfo" 66 67 #define GET_INSTRINFO_CTOR_DTOR 68 #define GET_INSTRMAP_INFO 69 #include "HexagonDepTimingClasses.h" 70 #include "HexagonGenDFAPacketizer.inc" 71 #include "HexagonGenInstrInfo.inc" 72 73 cl::opt<bool> ScheduleInlineAsm("hexagon-sched-inline-asm", cl::Hidden, 74 cl::init(false), cl::desc("Do not consider inline-asm a scheduling/" 75 "packetization boundary.")); 76 77 static cl::opt<bool> EnableBranchPrediction("hexagon-enable-branch-prediction", 78 cl::Hidden, cl::init(true), cl::desc("Enable branch prediction")); 79 80 static cl::opt<bool> DisableNVSchedule("disable-hexagon-nv-schedule", 81 cl::Hidden, cl::ZeroOrMore, cl::init(false), 82 cl::desc("Disable schedule adjustment for new value stores.")); 83 84 static cl::opt<bool> EnableTimingClassLatency( 85 "enable-timing-class-latency", cl::Hidden, cl::init(false), 86 cl::desc("Enable timing class latency")); 87 88 static cl::opt<bool> EnableALUForwarding( 89 "enable-alu-forwarding", cl::Hidden, cl::init(true), 90 cl::desc("Enable vec alu forwarding")); 91 92 static cl::opt<bool> EnableACCForwarding( 93 "enable-acc-forwarding", cl::Hidden, cl::init(true), 94 cl::desc("Enable vec acc forwarding")); 95 96 static cl::opt<bool> BranchRelaxAsmLarge("branch-relax-asm-large", 97 cl::init(true), cl::Hidden, cl::ZeroOrMore, cl::desc("branch relax asm")); 98 99 static cl::opt<bool> UseDFAHazardRec("dfa-hazard-rec", 100 cl::init(true), cl::Hidden, cl::ZeroOrMore, 101 cl::desc("Use the DFA based hazard recognizer.")); 102 103 /// Constants for Hexagon instructions. 104 const int Hexagon_MEMW_OFFSET_MAX = 4095; 105 const int Hexagon_MEMW_OFFSET_MIN = -4096; 106 const int Hexagon_MEMD_OFFSET_MAX = 8191; 107 const int Hexagon_MEMD_OFFSET_MIN = -8192; 108 const int Hexagon_MEMH_OFFSET_MAX = 2047; 109 const int Hexagon_MEMH_OFFSET_MIN = -2048; 110 const int Hexagon_MEMB_OFFSET_MAX = 1023; 111 const int Hexagon_MEMB_OFFSET_MIN = -1024; 112 const int Hexagon_ADDI_OFFSET_MAX = 32767; 113 const int Hexagon_ADDI_OFFSET_MIN = -32768; 114 115 // Pin the vtable to this file. 116 void HexagonInstrInfo::anchor() {} 117 118 HexagonInstrInfo::HexagonInstrInfo(HexagonSubtarget &ST) 119 : HexagonGenInstrInfo(Hexagon::ADJCALLSTACKDOWN, Hexagon::ADJCALLSTACKUP), 120 Subtarget(ST) {} 121 122 static bool isIntRegForSubInst(unsigned Reg) { 123 return (Reg >= Hexagon::R0 && Reg <= Hexagon::R7) || 124 (Reg >= Hexagon::R16 && Reg <= Hexagon::R23); 125 } 126 127 static bool isDblRegForSubInst(unsigned Reg, const HexagonRegisterInfo &HRI) { 128 return isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::isub_lo)) && 129 isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::isub_hi)); 130 } 131 132 /// Calculate number of instructions excluding the debug instructions. 133 static unsigned nonDbgMICount(MachineBasicBlock::const_instr_iterator MIB, 134 MachineBasicBlock::const_instr_iterator MIE) { 135 unsigned Count = 0; 136 for (; MIB != MIE; ++MIB) { 137 if (!MIB->isDebugValue()) 138 ++Count; 139 } 140 return Count; 141 } 142 143 /// Find the hardware loop instruction used to set-up the specified loop. 144 /// On Hexagon, we have two instructions used to set-up the hardware loop 145 /// (LOOP0, LOOP1) with corresponding endloop (ENDLOOP0, ENDLOOP1) instructions 146 /// to indicate the end of a loop. 147 static MachineInstr *findLoopInstr(MachineBasicBlock *BB, unsigned EndLoopOp, 148 MachineBasicBlock *TargetBB, 149 SmallPtrSet<MachineBasicBlock *, 8> &Visited) { 150 unsigned LOOPi; 151 unsigned LOOPr; 152 if (EndLoopOp == Hexagon::ENDLOOP0) { 153 LOOPi = Hexagon::J2_loop0i; 154 LOOPr = Hexagon::J2_loop0r; 155 } else { // EndLoopOp == Hexagon::EndLOOP1 156 LOOPi = Hexagon::J2_loop1i; 157 LOOPr = Hexagon::J2_loop1r; 158 } 159 160 // The loop set-up instruction will be in a predecessor block 161 for (MachineBasicBlock *PB : BB->predecessors()) { 162 // If this has been visited, already skip it. 163 if (!Visited.insert(PB).second) 164 continue; 165 if (PB == BB) 166 continue; 167 for (auto I = PB->instr_rbegin(), E = PB->instr_rend(); I != E; ++I) { 168 unsigned Opc = I->getOpcode(); 169 if (Opc == LOOPi || Opc == LOOPr) 170 return &*I; 171 // We've reached a different loop, which means the loop01 has been 172 // removed. 173 if (Opc == EndLoopOp && I->getOperand(0).getMBB() != TargetBB) 174 return nullptr; 175 } 176 // Check the predecessors for the LOOP instruction. 177 if (MachineInstr *Loop = findLoopInstr(PB, EndLoopOp, TargetBB, Visited)) 178 return Loop; 179 } 180 return nullptr; 181 } 182 183 /// Gather register def/uses from MI. 184 /// This treats possible (predicated) defs as actually happening ones 185 /// (conservatively). 186 static inline void parseOperands(const MachineInstr &MI, 187 SmallVector<unsigned, 4> &Defs, SmallVector<unsigned, 8> &Uses) { 188 Defs.clear(); 189 Uses.clear(); 190 191 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 192 const MachineOperand &MO = MI.getOperand(i); 193 194 if (!MO.isReg()) 195 continue; 196 197 unsigned Reg = MO.getReg(); 198 if (!Reg) 199 continue; 200 201 if (MO.isUse()) 202 Uses.push_back(MO.getReg()); 203 204 if (MO.isDef()) 205 Defs.push_back(MO.getReg()); 206 } 207 } 208 209 // Position dependent, so check twice for swap. 210 static bool isDuplexPairMatch(unsigned Ga, unsigned Gb) { 211 switch (Ga) { 212 case HexagonII::HSIG_None: 213 default: 214 return false; 215 case HexagonII::HSIG_L1: 216 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_A); 217 case HexagonII::HSIG_L2: 218 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 || 219 Gb == HexagonII::HSIG_A); 220 case HexagonII::HSIG_S1: 221 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 || 222 Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_A); 223 case HexagonII::HSIG_S2: 224 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 || 225 Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_S2 || 226 Gb == HexagonII::HSIG_A); 227 case HexagonII::HSIG_A: 228 return (Gb == HexagonII::HSIG_A); 229 case HexagonII::HSIG_Compound: 230 return (Gb == HexagonII::HSIG_Compound); 231 } 232 return false; 233 } 234 235 /// isLoadFromStackSlot - If the specified machine instruction is a direct 236 /// load from a stack slot, return the virtual or physical register number of 237 /// the destination along with the FrameIndex of the loaded stack slot. If 238 /// not, return 0. This predicate must return 0 if the instruction has 239 /// any side effects other than loading from the stack slot. 240 unsigned HexagonInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 241 int &FrameIndex) const { 242 switch (MI.getOpcode()) { 243 default: 244 break; 245 case Hexagon::L2_loadri_io: 246 case Hexagon::L2_loadrd_io: 247 case Hexagon::V6_vL32b_ai: 248 case Hexagon::V6_vL32b_nt_ai: 249 case Hexagon::V6_vL32Ub_ai: 250 case Hexagon::LDriw_pred: 251 case Hexagon::LDriw_mod: 252 case Hexagon::PS_vloadrq_ai: 253 case Hexagon::PS_vloadrw_ai: 254 case Hexagon::PS_vloadrw_nt_ai: { 255 const MachineOperand OpFI = MI.getOperand(1); 256 if (!OpFI.isFI()) 257 return 0; 258 const MachineOperand OpOff = MI.getOperand(2); 259 if (!OpOff.isImm() || OpOff.getImm() != 0) 260 return 0; 261 FrameIndex = OpFI.getIndex(); 262 return MI.getOperand(0).getReg(); 263 } 264 265 case Hexagon::L2_ploadrit_io: 266 case Hexagon::L2_ploadrif_io: 267 case Hexagon::L2_ploadrdt_io: 268 case Hexagon::L2_ploadrdf_io: { 269 const MachineOperand OpFI = MI.getOperand(2); 270 if (!OpFI.isFI()) 271 return 0; 272 const MachineOperand OpOff = MI.getOperand(3); 273 if (!OpOff.isImm() || OpOff.getImm() != 0) 274 return 0; 275 FrameIndex = OpFI.getIndex(); 276 return MI.getOperand(0).getReg(); 277 } 278 } 279 280 return 0; 281 } 282 283 /// isStoreToStackSlot - If the specified machine instruction is a direct 284 /// store to a stack slot, return the virtual or physical register number of 285 /// the source reg along with the FrameIndex of the loaded stack slot. If 286 /// not, return 0. This predicate must return 0 if the instruction has 287 /// any side effects other than storing to the stack slot. 288 unsigned HexagonInstrInfo::isStoreToStackSlot(const MachineInstr &MI, 289 int &FrameIndex) const { 290 switch (MI.getOpcode()) { 291 default: 292 break; 293 case Hexagon::S2_storerb_io: 294 case Hexagon::S2_storerh_io: 295 case Hexagon::S2_storeri_io: 296 case Hexagon::S2_storerd_io: 297 case Hexagon::V6_vS32b_ai: 298 case Hexagon::V6_vS32Ub_ai: 299 case Hexagon::STriw_pred: 300 case Hexagon::STriw_mod: 301 case Hexagon::PS_vstorerq_ai: 302 case Hexagon::PS_vstorerw_ai: { 303 const MachineOperand &OpFI = MI.getOperand(0); 304 if (!OpFI.isFI()) 305 return 0; 306 const MachineOperand &OpOff = MI.getOperand(1); 307 if (!OpOff.isImm() || OpOff.getImm() != 0) 308 return 0; 309 FrameIndex = OpFI.getIndex(); 310 return MI.getOperand(2).getReg(); 311 } 312 313 case Hexagon::S2_pstorerbt_io: 314 case Hexagon::S2_pstorerbf_io: 315 case Hexagon::S2_pstorerht_io: 316 case Hexagon::S2_pstorerhf_io: 317 case Hexagon::S2_pstorerit_io: 318 case Hexagon::S2_pstorerif_io: 319 case Hexagon::S2_pstorerdt_io: 320 case Hexagon::S2_pstorerdf_io: { 321 const MachineOperand &OpFI = MI.getOperand(1); 322 if (!OpFI.isFI()) 323 return 0; 324 const MachineOperand &OpOff = MI.getOperand(2); 325 if (!OpOff.isImm() || OpOff.getImm() != 0) 326 return 0; 327 FrameIndex = OpFI.getIndex(); 328 return MI.getOperand(3).getReg(); 329 } 330 } 331 332 return 0; 333 } 334 335 /// This function can analyze one/two way branching only and should (mostly) be 336 /// called by target independent side. 337 /// First entry is always the opcode of the branching instruction, except when 338 /// the Cond vector is supposed to be empty, e.g., when AnalyzeBranch fails, a 339 /// BB with only unconditional jump. Subsequent entries depend upon the opcode, 340 /// e.g. Jump_c p will have 341 /// Cond[0] = Jump_c 342 /// Cond[1] = p 343 /// HW-loop ENDLOOP: 344 /// Cond[0] = ENDLOOP 345 /// Cond[1] = MBB 346 /// New value jump: 347 /// Cond[0] = Hexagon::CMPEQri_f_Jumpnv_t_V4 -- specific opcode 348 /// Cond[1] = R 349 /// Cond[2] = Imm 350 bool HexagonInstrInfo::analyzeBranch(MachineBasicBlock &MBB, 351 MachineBasicBlock *&TBB, 352 MachineBasicBlock *&FBB, 353 SmallVectorImpl<MachineOperand> &Cond, 354 bool AllowModify) const { 355 TBB = nullptr; 356 FBB = nullptr; 357 Cond.clear(); 358 359 // If the block has no terminators, it just falls into the block after it. 360 MachineBasicBlock::instr_iterator I = MBB.instr_end(); 361 if (I == MBB.instr_begin()) 362 return false; 363 364 // A basic block may looks like this: 365 // 366 // [ insn 367 // EH_LABEL 368 // insn 369 // insn 370 // insn 371 // EH_LABEL 372 // insn ] 373 // 374 // It has two succs but does not have a terminator 375 // Don't know how to handle it. 376 do { 377 --I; 378 if (I->isEHLabel()) 379 // Don't analyze EH branches. 380 return true; 381 } while (I != MBB.instr_begin()); 382 383 I = MBB.instr_end(); 384 --I; 385 386 while (I->isDebugValue()) { 387 if (I == MBB.instr_begin()) 388 return false; 389 --I; 390 } 391 392 bool JumpToBlock = I->getOpcode() == Hexagon::J2_jump && 393 I->getOperand(0).isMBB(); 394 // Delete the J2_jump if it's equivalent to a fall-through. 395 if (AllowModify && JumpToBlock && 396 MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) { 397 DEBUG(dbgs() << "\nErasing the jump to successor block\n";); 398 I->eraseFromParent(); 399 I = MBB.instr_end(); 400 if (I == MBB.instr_begin()) 401 return false; 402 --I; 403 } 404 if (!isUnpredicatedTerminator(*I)) 405 return false; 406 407 // Get the last instruction in the block. 408 MachineInstr *LastInst = &*I; 409 MachineInstr *SecondLastInst = nullptr; 410 // Find one more terminator if present. 411 while (true) { 412 if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(*I)) { 413 if (!SecondLastInst) 414 SecondLastInst = &*I; 415 else 416 // This is a third branch. 417 return true; 418 } 419 if (I == MBB.instr_begin()) 420 break; 421 --I; 422 } 423 424 int LastOpcode = LastInst->getOpcode(); 425 int SecLastOpcode = SecondLastInst ? SecondLastInst->getOpcode() : 0; 426 // If the branch target is not a basic block, it could be a tail call. 427 // (It is, if the target is a function.) 428 if (LastOpcode == Hexagon::J2_jump && !LastInst->getOperand(0).isMBB()) 429 return true; 430 if (SecLastOpcode == Hexagon::J2_jump && 431 !SecondLastInst->getOperand(0).isMBB()) 432 return true; 433 434 bool LastOpcodeHasJMP_c = PredOpcodeHasJMP_c(LastOpcode); 435 bool LastOpcodeHasNVJump = isNewValueJump(*LastInst); 436 437 if (LastOpcodeHasJMP_c && !LastInst->getOperand(1).isMBB()) 438 return true; 439 440 // If there is only one terminator instruction, process it. 441 if (LastInst && !SecondLastInst) { 442 if (LastOpcode == Hexagon::J2_jump) { 443 TBB = LastInst->getOperand(0).getMBB(); 444 return false; 445 } 446 if (isEndLoopN(LastOpcode)) { 447 TBB = LastInst->getOperand(0).getMBB(); 448 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode())); 449 Cond.push_back(LastInst->getOperand(0)); 450 return false; 451 } 452 if (LastOpcodeHasJMP_c) { 453 TBB = LastInst->getOperand(1).getMBB(); 454 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode())); 455 Cond.push_back(LastInst->getOperand(0)); 456 return false; 457 } 458 // Only supporting rr/ri versions of new-value jumps. 459 if (LastOpcodeHasNVJump && (LastInst->getNumExplicitOperands() == 3)) { 460 TBB = LastInst->getOperand(2).getMBB(); 461 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode())); 462 Cond.push_back(LastInst->getOperand(0)); 463 Cond.push_back(LastInst->getOperand(1)); 464 return false; 465 } 466 DEBUG(dbgs() << "\nCant analyze BB#" << MBB.getNumber() 467 << " with one jump\n";); 468 // Otherwise, don't know what this is. 469 return true; 470 } 471 472 bool SecLastOpcodeHasJMP_c = PredOpcodeHasJMP_c(SecLastOpcode); 473 bool SecLastOpcodeHasNVJump = isNewValueJump(*SecondLastInst); 474 if (SecLastOpcodeHasJMP_c && (LastOpcode == Hexagon::J2_jump)) { 475 if (!SecondLastInst->getOperand(1).isMBB()) 476 return true; 477 TBB = SecondLastInst->getOperand(1).getMBB(); 478 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode())); 479 Cond.push_back(SecondLastInst->getOperand(0)); 480 FBB = LastInst->getOperand(0).getMBB(); 481 return false; 482 } 483 484 // Only supporting rr/ri versions of new-value jumps. 485 if (SecLastOpcodeHasNVJump && 486 (SecondLastInst->getNumExplicitOperands() == 3) && 487 (LastOpcode == Hexagon::J2_jump)) { 488 TBB = SecondLastInst->getOperand(2).getMBB(); 489 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode())); 490 Cond.push_back(SecondLastInst->getOperand(0)); 491 Cond.push_back(SecondLastInst->getOperand(1)); 492 FBB = LastInst->getOperand(0).getMBB(); 493 return false; 494 } 495 496 // If the block ends with two Hexagon:JMPs, handle it. The second one is not 497 // executed, so remove it. 498 if (SecLastOpcode == Hexagon::J2_jump && LastOpcode == Hexagon::J2_jump) { 499 TBB = SecondLastInst->getOperand(0).getMBB(); 500 I = LastInst->getIterator(); 501 if (AllowModify) 502 I->eraseFromParent(); 503 return false; 504 } 505 506 // If the block ends with an ENDLOOP, and J2_jump, handle it. 507 if (isEndLoopN(SecLastOpcode) && LastOpcode == Hexagon::J2_jump) { 508 TBB = SecondLastInst->getOperand(0).getMBB(); 509 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode())); 510 Cond.push_back(SecondLastInst->getOperand(0)); 511 FBB = LastInst->getOperand(0).getMBB(); 512 return false; 513 } 514 DEBUG(dbgs() << "\nCant analyze BB#" << MBB.getNumber() 515 << " with two jumps";); 516 // Otherwise, can't handle this. 517 return true; 518 } 519 520 unsigned HexagonInstrInfo::removeBranch(MachineBasicBlock &MBB, 521 int *BytesRemoved) const { 522 assert(!BytesRemoved && "code size not handled"); 523 524 DEBUG(dbgs() << "\nRemoving branches out of BB#" << MBB.getNumber()); 525 MachineBasicBlock::iterator I = MBB.end(); 526 unsigned Count = 0; 527 while (I != MBB.begin()) { 528 --I; 529 if (I->isDebugValue()) 530 continue; 531 // Only removing branches from end of MBB. 532 if (!I->isBranch()) 533 return Count; 534 if (Count && (I->getOpcode() == Hexagon::J2_jump)) 535 llvm_unreachable("Malformed basic block: unconditional branch not last"); 536 MBB.erase(&MBB.back()); 537 I = MBB.end(); 538 ++Count; 539 } 540 return Count; 541 } 542 543 unsigned HexagonInstrInfo::insertBranch(MachineBasicBlock &MBB, 544 MachineBasicBlock *TBB, 545 MachineBasicBlock *FBB, 546 ArrayRef<MachineOperand> Cond, 547 const DebugLoc &DL, 548 int *BytesAdded) const { 549 unsigned BOpc = Hexagon::J2_jump; 550 unsigned BccOpc = Hexagon::J2_jumpt; 551 assert(validateBranchCond(Cond) && "Invalid branching condition"); 552 assert(TBB && "insertBranch must not be told to insert a fallthrough"); 553 assert(!BytesAdded && "code size not handled"); 554 555 // Check if reverseBranchCondition has asked to reverse this branch 556 // If we want to reverse the branch an odd number of times, we want 557 // J2_jumpf. 558 if (!Cond.empty() && Cond[0].isImm()) 559 BccOpc = Cond[0].getImm(); 560 561 if (!FBB) { 562 if (Cond.empty()) { 563 // Due to a bug in TailMerging/CFG Optimization, we need to add a 564 // special case handling of a predicated jump followed by an 565 // unconditional jump. If not, Tail Merging and CFG Optimization go 566 // into an infinite loop. 567 MachineBasicBlock *NewTBB, *NewFBB; 568 SmallVector<MachineOperand, 4> Cond; 569 auto Term = MBB.getFirstTerminator(); 570 if (Term != MBB.end() && isPredicated(*Term) && 571 !analyzeBranch(MBB, NewTBB, NewFBB, Cond, false) && 572 MachineFunction::iterator(NewTBB) == ++MBB.getIterator()) { 573 reverseBranchCondition(Cond); 574 removeBranch(MBB); 575 return insertBranch(MBB, TBB, nullptr, Cond, DL); 576 } 577 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB); 578 } else if (isEndLoopN(Cond[0].getImm())) { 579 int EndLoopOp = Cond[0].getImm(); 580 assert(Cond[1].isMBB()); 581 // Since we're adding an ENDLOOP, there better be a LOOP instruction. 582 // Check for it, and change the BB target if needed. 583 SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs; 584 MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, Cond[1].getMBB(), 585 VisitedBBs); 586 assert(Loop != nullptr && "Inserting an ENDLOOP without a LOOP"); 587 Loop->getOperand(0).setMBB(TBB); 588 // Add the ENDLOOP after the finding the LOOP0. 589 BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB); 590 } else if (isNewValueJump(Cond[0].getImm())) { 591 assert((Cond.size() == 3) && "Only supporting rr/ri version of nvjump"); 592 // New value jump 593 // (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset) 594 // (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset) 595 unsigned Flags1 = getUndefRegState(Cond[1].isUndef()); 596 DEBUG(dbgs() << "\nInserting NVJump for BB#" << MBB.getNumber();); 597 if (Cond[2].isReg()) { 598 unsigned Flags2 = getUndefRegState(Cond[2].isUndef()); 599 BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1). 600 addReg(Cond[2].getReg(), Flags2).addMBB(TBB); 601 } else if(Cond[2].isImm()) { 602 BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1). 603 addImm(Cond[2].getImm()).addMBB(TBB); 604 } else 605 llvm_unreachable("Invalid condition for branching"); 606 } else { 607 assert((Cond.size() == 2) && "Malformed cond vector"); 608 const MachineOperand &RO = Cond[1]; 609 unsigned Flags = getUndefRegState(RO.isUndef()); 610 BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB); 611 } 612 return 1; 613 } 614 assert((!Cond.empty()) && 615 "Cond. cannot be empty when multiple branchings are required"); 616 assert((!isNewValueJump(Cond[0].getImm())) && 617 "NV-jump cannot be inserted with another branch"); 618 // Special case for hardware loops. The condition is a basic block. 619 if (isEndLoopN(Cond[0].getImm())) { 620 int EndLoopOp = Cond[0].getImm(); 621 assert(Cond[1].isMBB()); 622 // Since we're adding an ENDLOOP, there better be a LOOP instruction. 623 // Check for it, and change the BB target if needed. 624 SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs; 625 MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, Cond[1].getMBB(), 626 VisitedBBs); 627 assert(Loop != nullptr && "Inserting an ENDLOOP without a LOOP"); 628 Loop->getOperand(0).setMBB(TBB); 629 // Add the ENDLOOP after the finding the LOOP0. 630 BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB); 631 } else { 632 const MachineOperand &RO = Cond[1]; 633 unsigned Flags = getUndefRegState(RO.isUndef()); 634 BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB); 635 } 636 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB); 637 638 return 2; 639 } 640 641 /// Analyze the loop code to find the loop induction variable and compare used 642 /// to compute the number of iterations. Currently, we analyze loop that are 643 /// controlled using hardware loops. In this case, the induction variable 644 /// instruction is null. For all other cases, this function returns true, which 645 /// means we're unable to analyze it. 646 bool HexagonInstrInfo::analyzeLoop(MachineLoop &L, 647 MachineInstr *&IndVarInst, 648 MachineInstr *&CmpInst) const { 649 650 MachineBasicBlock *LoopEnd = L.getBottomBlock(); 651 MachineBasicBlock::iterator I = LoopEnd->getFirstTerminator(); 652 // We really "analyze" only hardware loops right now. 653 if (I != LoopEnd->end() && isEndLoopN(I->getOpcode())) { 654 IndVarInst = nullptr; 655 CmpInst = &*I; 656 return false; 657 } 658 return true; 659 } 660 661 /// Generate code to reduce the loop iteration by one and check if the loop is 662 /// finished. Return the value/register of the new loop count. this function 663 /// assumes the nth iteration is peeled first. 664 unsigned HexagonInstrInfo::reduceLoopCount(MachineBasicBlock &MBB, 665 MachineInstr *IndVar, MachineInstr &Cmp, 666 SmallVectorImpl<MachineOperand> &Cond, 667 SmallVectorImpl<MachineInstr *> &PrevInsts, 668 unsigned Iter, unsigned MaxIter) const { 669 // We expect a hardware loop currently. This means that IndVar is set 670 // to null, and the compare is the ENDLOOP instruction. 671 assert((!IndVar) && isEndLoopN(Cmp.getOpcode()) 672 && "Expecting a hardware loop"); 673 MachineFunction *MF = MBB.getParent(); 674 DebugLoc DL = Cmp.getDebugLoc(); 675 SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs; 676 MachineInstr *Loop = findLoopInstr(&MBB, Cmp.getOpcode(), 677 Cmp.getOperand(0).getMBB(), VisitedBBs); 678 if (!Loop) 679 return 0; 680 // If the loop trip count is a compile-time value, then just change the 681 // value. 682 if (Loop->getOpcode() == Hexagon::J2_loop0i || 683 Loop->getOpcode() == Hexagon::J2_loop1i) { 684 int64_t Offset = Loop->getOperand(1).getImm(); 685 if (Offset <= 1) 686 Loop->eraseFromParent(); 687 else 688 Loop->getOperand(1).setImm(Offset - 1); 689 return Offset - 1; 690 } 691 // The loop trip count is a run-time value. We generate code to subtract 692 // one from the trip count, and update the loop instruction. 693 assert(Loop->getOpcode() == Hexagon::J2_loop0r && "Unexpected instruction"); 694 unsigned LoopCount = Loop->getOperand(1).getReg(); 695 // Check if we're done with the loop. 696 unsigned LoopEnd = createVR(MF, MVT::i1); 697 MachineInstr *NewCmp = BuildMI(&MBB, DL, get(Hexagon::C2_cmpgtui), LoopEnd). 698 addReg(LoopCount).addImm(1); 699 unsigned NewLoopCount = createVR(MF, MVT::i32); 700 MachineInstr *NewAdd = BuildMI(&MBB, DL, get(Hexagon::A2_addi), NewLoopCount). 701 addReg(LoopCount).addImm(-1); 702 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 703 // Update the previously generated instructions with the new loop counter. 704 for (SmallVectorImpl<MachineInstr *>::iterator I = PrevInsts.begin(), 705 E = PrevInsts.end(); I != E; ++I) 706 (*I)->substituteRegister(LoopCount, NewLoopCount, 0, HRI); 707 PrevInsts.clear(); 708 PrevInsts.push_back(NewCmp); 709 PrevInsts.push_back(NewAdd); 710 // Insert the new loop instruction if this is the last time the loop is 711 // decremented. 712 if (Iter == MaxIter) 713 BuildMI(&MBB, DL, get(Hexagon::J2_loop0r)). 714 addMBB(Loop->getOperand(0).getMBB()).addReg(NewLoopCount); 715 // Delete the old loop instruction. 716 if (Iter == 0) 717 Loop->eraseFromParent(); 718 Cond.push_back(MachineOperand::CreateImm(Hexagon::J2_jumpf)); 719 Cond.push_back(NewCmp->getOperand(0)); 720 return NewLoopCount; 721 } 722 723 bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB, 724 unsigned NumCycles, unsigned ExtraPredCycles, 725 BranchProbability Probability) const { 726 return nonDbgBBSize(&MBB) <= 3; 727 } 728 729 bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB, 730 unsigned NumTCycles, unsigned ExtraTCycles, MachineBasicBlock &FMBB, 731 unsigned NumFCycles, unsigned ExtraFCycles, BranchProbability Probability) 732 const { 733 return nonDbgBBSize(&TMBB) <= 3 && nonDbgBBSize(&FMBB) <= 3; 734 } 735 736 bool HexagonInstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB, 737 unsigned NumInstrs, BranchProbability Probability) const { 738 return NumInstrs <= 4; 739 } 740 741 void HexagonInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 742 MachineBasicBlock::iterator I, 743 const DebugLoc &DL, unsigned DestReg, 744 unsigned SrcReg, bool KillSrc) const { 745 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 746 unsigned KillFlag = getKillRegState(KillSrc); 747 748 if (Hexagon::IntRegsRegClass.contains(SrcReg, DestReg)) { 749 BuildMI(MBB, I, DL, get(Hexagon::A2_tfr), DestReg) 750 .addReg(SrcReg, KillFlag); 751 return; 752 } 753 if (Hexagon::DoubleRegsRegClass.contains(SrcReg, DestReg)) { 754 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrp), DestReg) 755 .addReg(SrcReg, KillFlag); 756 return; 757 } 758 if (Hexagon::PredRegsRegClass.contains(SrcReg, DestReg)) { 759 // Map Pd = Ps to Pd = or(Ps, Ps). 760 BuildMI(MBB, I, DL, get(Hexagon::C2_or), DestReg) 761 .addReg(SrcReg).addReg(SrcReg, KillFlag); 762 return; 763 } 764 if (Hexagon::CtrRegsRegClass.contains(DestReg) && 765 Hexagon::IntRegsRegClass.contains(SrcReg)) { 766 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg) 767 .addReg(SrcReg, KillFlag); 768 return; 769 } 770 if (Hexagon::IntRegsRegClass.contains(DestReg) && 771 Hexagon::CtrRegsRegClass.contains(SrcReg)) { 772 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrcrr), DestReg) 773 .addReg(SrcReg, KillFlag); 774 return; 775 } 776 if (Hexagon::ModRegsRegClass.contains(DestReg) && 777 Hexagon::IntRegsRegClass.contains(SrcReg)) { 778 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg) 779 .addReg(SrcReg, KillFlag); 780 return; 781 } 782 if (Hexagon::PredRegsRegClass.contains(SrcReg) && 783 Hexagon::IntRegsRegClass.contains(DestReg)) { 784 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg) 785 .addReg(SrcReg, KillFlag); 786 return; 787 } 788 if (Hexagon::IntRegsRegClass.contains(SrcReg) && 789 Hexagon::PredRegsRegClass.contains(DestReg)) { 790 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrrp), DestReg) 791 .addReg(SrcReg, KillFlag); 792 return; 793 } 794 if (Hexagon::PredRegsRegClass.contains(SrcReg) && 795 Hexagon::IntRegsRegClass.contains(DestReg)) { 796 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg) 797 .addReg(SrcReg, KillFlag); 798 return; 799 } 800 if (Hexagon::HvxVRRegClass.contains(SrcReg, DestReg)) { 801 BuildMI(MBB, I, DL, get(Hexagon::V6_vassign), DestReg). 802 addReg(SrcReg, KillFlag); 803 return; 804 } 805 if (Hexagon::HvxWRRegClass.contains(SrcReg, DestReg)) { 806 unsigned LoSrc = HRI.getSubReg(SrcReg, Hexagon::vsub_lo); 807 unsigned HiSrc = HRI.getSubReg(SrcReg, Hexagon::vsub_hi); 808 BuildMI(MBB, I, DL, get(Hexagon::V6_vcombine), DestReg) 809 .addReg(HiSrc, KillFlag) 810 .addReg(LoSrc, KillFlag); 811 return; 812 } 813 if (Hexagon::HvxQRRegClass.contains(SrcReg, DestReg)) { 814 BuildMI(MBB, I, DL, get(Hexagon::V6_pred_and), DestReg) 815 .addReg(SrcReg) 816 .addReg(SrcReg, KillFlag); 817 return; 818 } 819 if (Hexagon::HvxQRRegClass.contains(SrcReg) && 820 Hexagon::HvxVRRegClass.contains(DestReg)) { 821 llvm_unreachable("Unimplemented pred to vec"); 822 return; 823 } 824 if (Hexagon::HvxQRRegClass.contains(DestReg) && 825 Hexagon::HvxVRRegClass.contains(SrcReg)) { 826 llvm_unreachable("Unimplemented vec to pred"); 827 return; 828 } 829 830 #ifndef NDEBUG 831 // Show the invalid registers to ease debugging. 832 dbgs() << "Invalid registers for copy in BB#" << MBB.getNumber() 833 << ": " << PrintReg(DestReg, &HRI) 834 << " = " << PrintReg(SrcReg, &HRI) << '\n'; 835 #endif 836 llvm_unreachable("Unimplemented"); 837 } 838 839 void HexagonInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 840 MachineBasicBlock::iterator I, unsigned SrcReg, bool isKill, int FI, 841 const TargetRegisterClass *RC, const TargetRegisterInfo *TRI) const { 842 DebugLoc DL = MBB.findDebugLoc(I); 843 MachineFunction &MF = *MBB.getParent(); 844 MachineFrameInfo &MFI = MF.getFrameInfo(); 845 unsigned SlotAlign = MFI.getObjectAlignment(FI); 846 unsigned RegAlign = TRI->getSpillAlignment(*RC); 847 unsigned KillFlag = getKillRegState(isKill); 848 bool HasAlloca = MFI.hasVarSizedObjects(); 849 const HexagonFrameLowering &HFI = *Subtarget.getFrameLowering(); 850 851 MachineMemOperand *MMO = MF.getMachineMemOperand( 852 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore, 853 MFI.getObjectSize(FI), SlotAlign); 854 855 if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) { 856 BuildMI(MBB, I, DL, get(Hexagon::S2_storeri_io)) 857 .addFrameIndex(FI).addImm(0) 858 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 859 } else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) { 860 BuildMI(MBB, I, DL, get(Hexagon::S2_storerd_io)) 861 .addFrameIndex(FI).addImm(0) 862 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 863 } else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) { 864 BuildMI(MBB, I, DL, get(Hexagon::STriw_pred)) 865 .addFrameIndex(FI).addImm(0) 866 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 867 } else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) { 868 BuildMI(MBB, I, DL, get(Hexagon::STriw_mod)) 869 .addFrameIndex(FI).addImm(0) 870 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 871 } else if (Hexagon::HvxQRRegClass.hasSubClassEq(RC)) { 872 BuildMI(MBB, I, DL, get(Hexagon::PS_vstorerq_ai)) 873 .addFrameIndex(FI).addImm(0) 874 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 875 } else if (Hexagon::HvxVRRegClass.hasSubClassEq(RC)) { 876 // If there are variable-sized objects, spills will not be aligned. 877 if (HasAlloca) 878 SlotAlign = HFI.getStackAlignment(); 879 unsigned Opc = SlotAlign < RegAlign ? Hexagon::V6_vS32Ub_ai 880 : Hexagon::V6_vS32b_ai; 881 MachineMemOperand *MMOA = MF.getMachineMemOperand( 882 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore, 883 MFI.getObjectSize(FI), SlotAlign); 884 BuildMI(MBB, I, DL, get(Opc)) 885 .addFrameIndex(FI).addImm(0) 886 .addReg(SrcReg, KillFlag).addMemOperand(MMOA); 887 } else if (Hexagon::HvxWRRegClass.hasSubClassEq(RC)) { 888 // If there are variable-sized objects, spills will not be aligned. 889 if (HasAlloca) 890 SlotAlign = HFI.getStackAlignment(); 891 unsigned Opc = SlotAlign < RegAlign ? Hexagon::PS_vstorerwu_ai 892 : Hexagon::PS_vstorerw_ai; 893 MachineMemOperand *MMOA = MF.getMachineMemOperand( 894 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore, 895 MFI.getObjectSize(FI), SlotAlign); 896 BuildMI(MBB, I, DL, get(Opc)) 897 .addFrameIndex(FI).addImm(0) 898 .addReg(SrcReg, KillFlag).addMemOperand(MMOA); 899 } else { 900 llvm_unreachable("Unimplemented"); 901 } 902 } 903 904 void HexagonInstrInfo::loadRegFromStackSlot( 905 MachineBasicBlock &MBB, MachineBasicBlock::iterator I, unsigned DestReg, 906 int FI, const TargetRegisterClass *RC, 907 const TargetRegisterInfo *TRI) const { 908 DebugLoc DL = MBB.findDebugLoc(I); 909 MachineFunction &MF = *MBB.getParent(); 910 MachineFrameInfo &MFI = MF.getFrameInfo(); 911 unsigned SlotAlign = MFI.getObjectAlignment(FI); 912 unsigned RegAlign = TRI->getSpillAlignment(*RC); 913 bool HasAlloca = MFI.hasVarSizedObjects(); 914 const HexagonFrameLowering &HFI = *Subtarget.getFrameLowering(); 915 916 MachineMemOperand *MMO = MF.getMachineMemOperand( 917 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad, 918 MFI.getObjectSize(FI), SlotAlign); 919 920 if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) { 921 BuildMI(MBB, I, DL, get(Hexagon::L2_loadri_io), DestReg) 922 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 923 } else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) { 924 BuildMI(MBB, I, DL, get(Hexagon::L2_loadrd_io), DestReg) 925 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 926 } else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) { 927 BuildMI(MBB, I, DL, get(Hexagon::LDriw_pred), DestReg) 928 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 929 } else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) { 930 BuildMI(MBB, I, DL, get(Hexagon::LDriw_mod), DestReg) 931 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 932 } else if (Hexagon::HvxQRRegClass.hasSubClassEq(RC)) { 933 BuildMI(MBB, I, DL, get(Hexagon::PS_vloadrq_ai), DestReg) 934 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 935 } else if (Hexagon::HvxVRRegClass.hasSubClassEq(RC)) { 936 // If there are variable-sized objects, spills will not be aligned. 937 if (HasAlloca) 938 SlotAlign = HFI.getStackAlignment(); 939 unsigned Opc = SlotAlign < RegAlign ? Hexagon::V6_vL32Ub_ai 940 : Hexagon::V6_vL32b_ai; 941 MachineMemOperand *MMOA = MF.getMachineMemOperand( 942 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad, 943 MFI.getObjectSize(FI), SlotAlign); 944 BuildMI(MBB, I, DL, get(Opc), DestReg) 945 .addFrameIndex(FI).addImm(0).addMemOperand(MMOA); 946 } else if (Hexagon::HvxWRRegClass.hasSubClassEq(RC)) { 947 // If there are variable-sized objects, spills will not be aligned. 948 if (HasAlloca) 949 SlotAlign = HFI.getStackAlignment(); 950 unsigned Opc = SlotAlign < RegAlign ? Hexagon::PS_vloadrwu_ai 951 : Hexagon::PS_vloadrw_ai; 952 MachineMemOperand *MMOA = MF.getMachineMemOperand( 953 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad, 954 MFI.getObjectSize(FI), SlotAlign); 955 BuildMI(MBB, I, DL, get(Opc), DestReg) 956 .addFrameIndex(FI).addImm(0).addMemOperand(MMOA); 957 } else { 958 llvm_unreachable("Can't store this register to stack slot"); 959 } 960 } 961 962 static void getLiveRegsAt(LivePhysRegs &Regs, const MachineInstr &MI) { 963 const MachineBasicBlock &B = *MI.getParent(); 964 Regs.addLiveOuts(B); 965 auto E = ++MachineBasicBlock::const_iterator(MI.getIterator()).getReverse(); 966 for (auto I = B.rbegin(); I != E; ++I) 967 Regs.stepBackward(*I); 968 } 969 970 /// expandPostRAPseudo - This function is called for all pseudo instructions 971 /// that remain after register allocation. Many pseudo instructions are 972 /// created to help register allocation. This is the place to convert them 973 /// into real instructions. The target can edit MI in place, or it can insert 974 /// new instructions and erase MI. The function should return true if 975 /// anything was changed. 976 bool HexagonInstrInfo::expandPostRAPseudo(MachineInstr &MI) const { 977 MachineBasicBlock &MBB = *MI.getParent(); 978 MachineFunction &MF = *MBB.getParent(); 979 MachineRegisterInfo &MRI = MF.getRegInfo(); 980 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 981 DebugLoc DL = MI.getDebugLoc(); 982 unsigned Opc = MI.getOpcode(); 983 984 switch (Opc) { 985 case TargetOpcode::COPY: { 986 MachineOperand &MD = MI.getOperand(0); 987 MachineOperand &MS = MI.getOperand(1); 988 MachineBasicBlock::iterator MBBI = MI.getIterator(); 989 if (MD.getReg() != MS.getReg() && !MS.isUndef()) { 990 copyPhysReg(MBB, MI, DL, MD.getReg(), MS.getReg(), MS.isKill()); 991 std::prev(MBBI)->copyImplicitOps(*MBB.getParent(), MI); 992 } 993 MBB.erase(MBBI); 994 return true; 995 } 996 case Hexagon::PS_aligna: 997 BuildMI(MBB, MI, DL, get(Hexagon::A2_andir), MI.getOperand(0).getReg()) 998 .addReg(HRI.getFrameRegister()) 999 .addImm(-MI.getOperand(1).getImm()); 1000 MBB.erase(MI); 1001 return true; 1002 case Hexagon::V6_vassignp: { 1003 unsigned SrcReg = MI.getOperand(1).getReg(); 1004 unsigned DstReg = MI.getOperand(0).getReg(); 1005 unsigned Kill = getKillRegState(MI.getOperand(1).isKill()); 1006 BuildMI(MBB, MI, DL, get(Hexagon::V6_vcombine), DstReg) 1007 .addReg(HRI.getSubReg(SrcReg, Hexagon::vsub_hi), Kill) 1008 .addReg(HRI.getSubReg(SrcReg, Hexagon::vsub_lo), Kill); 1009 MBB.erase(MI); 1010 return true; 1011 } 1012 case Hexagon::V6_lo: { 1013 unsigned SrcReg = MI.getOperand(1).getReg(); 1014 unsigned DstReg = MI.getOperand(0).getReg(); 1015 unsigned SrcSubLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo); 1016 copyPhysReg(MBB, MI, DL, DstReg, SrcSubLo, MI.getOperand(1).isKill()); 1017 MBB.erase(MI); 1018 MRI.clearKillFlags(SrcSubLo); 1019 return true; 1020 } 1021 case Hexagon::V6_hi: { 1022 unsigned SrcReg = MI.getOperand(1).getReg(); 1023 unsigned DstReg = MI.getOperand(0).getReg(); 1024 unsigned SrcSubHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi); 1025 copyPhysReg(MBB, MI, DL, DstReg, SrcSubHi, MI.getOperand(1).isKill()); 1026 MBB.erase(MI); 1027 MRI.clearKillFlags(SrcSubHi); 1028 return true; 1029 } 1030 case Hexagon::PS_vstorerw_ai: 1031 case Hexagon::PS_vstorerwu_ai: { 1032 bool Aligned = Opc == Hexagon::PS_vstorerw_ai; 1033 unsigned SrcReg = MI.getOperand(2).getReg(); 1034 unsigned SrcSubHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi); 1035 unsigned SrcSubLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo); 1036 unsigned NewOpc = Aligned ? Hexagon::V6_vS32b_ai : Hexagon::V6_vS32Ub_ai; 1037 unsigned Offset = HRI.getSpillSize(Hexagon::HvxVRRegClass); 1038 1039 MachineInstr *MI1New = 1040 BuildMI(MBB, MI, DL, get(NewOpc)) 1041 .add(MI.getOperand(0)) 1042 .addImm(MI.getOperand(1).getImm()) 1043 .addReg(SrcSubLo) 1044 .setMemRefs(MI.memoperands_begin(), MI.memoperands_end()); 1045 MI1New->getOperand(0).setIsKill(false); 1046 BuildMI(MBB, MI, DL, get(NewOpc)) 1047 .add(MI.getOperand(0)) 1048 // The Vectors are indexed in multiples of vector size. 1049 .addImm(MI.getOperand(1).getImm() + Offset) 1050 .addReg(SrcSubHi) 1051 .setMemRefs(MI.memoperands_begin(), MI.memoperands_end()); 1052 MBB.erase(MI); 1053 return true; 1054 } 1055 case Hexagon::PS_vloadrw_ai: 1056 case Hexagon::PS_vloadrwu_ai: { 1057 bool Aligned = Opc == Hexagon::PS_vloadrw_ai; 1058 unsigned DstReg = MI.getOperand(0).getReg(); 1059 unsigned NewOpc = Aligned ? Hexagon::V6_vL32b_ai : Hexagon::V6_vL32Ub_ai; 1060 unsigned Offset = HRI.getSpillSize(Hexagon::HvxVRRegClass); 1061 1062 MachineInstr *MI1New = BuildMI(MBB, MI, DL, get(NewOpc), 1063 HRI.getSubReg(DstReg, Hexagon::vsub_lo)) 1064 .add(MI.getOperand(1)) 1065 .addImm(MI.getOperand(2).getImm()) 1066 .setMemRefs(MI.memoperands_begin(), MI.memoperands_end()); 1067 MI1New->getOperand(1).setIsKill(false); 1068 BuildMI(MBB, MI, DL, get(NewOpc), HRI.getSubReg(DstReg, Hexagon::vsub_hi)) 1069 .add(MI.getOperand(1)) 1070 // The Vectors are indexed in multiples of vector size. 1071 .addImm(MI.getOperand(2).getImm() + Offset) 1072 .setMemRefs(MI.memoperands_begin(), MI.memoperands_end()); 1073 MBB.erase(MI); 1074 return true; 1075 } 1076 case Hexagon::PS_true: { 1077 unsigned Reg = MI.getOperand(0).getReg(); 1078 BuildMI(MBB, MI, DL, get(Hexagon::C2_orn), Reg) 1079 .addReg(Reg, RegState::Undef) 1080 .addReg(Reg, RegState::Undef); 1081 MBB.erase(MI); 1082 return true; 1083 } 1084 case Hexagon::PS_false: { 1085 unsigned Reg = MI.getOperand(0).getReg(); 1086 BuildMI(MBB, MI, DL, get(Hexagon::C2_andn), Reg) 1087 .addReg(Reg, RegState::Undef) 1088 .addReg(Reg, RegState::Undef); 1089 MBB.erase(MI); 1090 return true; 1091 } 1092 case Hexagon::PS_vmulw: { 1093 // Expand a 64-bit vector multiply into 2 32-bit scalar multiplies. 1094 unsigned DstReg = MI.getOperand(0).getReg(); 1095 unsigned Src1Reg = MI.getOperand(1).getReg(); 1096 unsigned Src2Reg = MI.getOperand(2).getReg(); 1097 unsigned Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::isub_hi); 1098 unsigned Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::isub_lo); 1099 unsigned Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::isub_hi); 1100 unsigned Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::isub_lo); 1101 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_mpyi), 1102 HRI.getSubReg(DstReg, Hexagon::isub_hi)) 1103 .addReg(Src1SubHi) 1104 .addReg(Src2SubHi); 1105 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_mpyi), 1106 HRI.getSubReg(DstReg, Hexagon::isub_lo)) 1107 .addReg(Src1SubLo) 1108 .addReg(Src2SubLo); 1109 MBB.erase(MI); 1110 MRI.clearKillFlags(Src1SubHi); 1111 MRI.clearKillFlags(Src1SubLo); 1112 MRI.clearKillFlags(Src2SubHi); 1113 MRI.clearKillFlags(Src2SubLo); 1114 return true; 1115 } 1116 case Hexagon::PS_vmulw_acc: { 1117 // Expand 64-bit vector multiply with addition into 2 scalar multiplies. 1118 unsigned DstReg = MI.getOperand(0).getReg(); 1119 unsigned Src1Reg = MI.getOperand(1).getReg(); 1120 unsigned Src2Reg = MI.getOperand(2).getReg(); 1121 unsigned Src3Reg = MI.getOperand(3).getReg(); 1122 unsigned Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::isub_hi); 1123 unsigned Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::isub_lo); 1124 unsigned Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::isub_hi); 1125 unsigned Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::isub_lo); 1126 unsigned Src3SubHi = HRI.getSubReg(Src3Reg, Hexagon::isub_hi); 1127 unsigned Src3SubLo = HRI.getSubReg(Src3Reg, Hexagon::isub_lo); 1128 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_maci), 1129 HRI.getSubReg(DstReg, Hexagon::isub_hi)) 1130 .addReg(Src1SubHi) 1131 .addReg(Src2SubHi) 1132 .addReg(Src3SubHi); 1133 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_maci), 1134 HRI.getSubReg(DstReg, Hexagon::isub_lo)) 1135 .addReg(Src1SubLo) 1136 .addReg(Src2SubLo) 1137 .addReg(Src3SubLo); 1138 MBB.erase(MI); 1139 MRI.clearKillFlags(Src1SubHi); 1140 MRI.clearKillFlags(Src1SubLo); 1141 MRI.clearKillFlags(Src2SubHi); 1142 MRI.clearKillFlags(Src2SubLo); 1143 MRI.clearKillFlags(Src3SubHi); 1144 MRI.clearKillFlags(Src3SubLo); 1145 return true; 1146 } 1147 case Hexagon::PS_pselect: { 1148 const MachineOperand &Op0 = MI.getOperand(0); 1149 const MachineOperand &Op1 = MI.getOperand(1); 1150 const MachineOperand &Op2 = MI.getOperand(2); 1151 const MachineOperand &Op3 = MI.getOperand(3); 1152 unsigned Rd = Op0.getReg(); 1153 unsigned Pu = Op1.getReg(); 1154 unsigned Rs = Op2.getReg(); 1155 unsigned Rt = Op3.getReg(); 1156 DebugLoc DL = MI.getDebugLoc(); 1157 unsigned K1 = getKillRegState(Op1.isKill()); 1158 unsigned K2 = getKillRegState(Op2.isKill()); 1159 unsigned K3 = getKillRegState(Op3.isKill()); 1160 if (Rd != Rs) 1161 BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpt), Rd) 1162 .addReg(Pu, (Rd == Rt) ? K1 : 0) 1163 .addReg(Rs, K2); 1164 if (Rd != Rt) 1165 BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpf), Rd) 1166 .addReg(Pu, K1) 1167 .addReg(Rt, K3); 1168 MBB.erase(MI); 1169 return true; 1170 } 1171 case Hexagon::PS_vselect: { 1172 const MachineOperand &Op0 = MI.getOperand(0); 1173 const MachineOperand &Op1 = MI.getOperand(1); 1174 const MachineOperand &Op2 = MI.getOperand(2); 1175 const MachineOperand &Op3 = MI.getOperand(3); 1176 LivePhysRegs LiveAtMI(HRI); 1177 getLiveRegsAt(LiveAtMI, MI); 1178 bool IsDestLive = !LiveAtMI.available(MRI, Op0.getReg()); 1179 unsigned PReg = Op1.getReg(); 1180 assert(Op1.getSubReg() == 0); 1181 unsigned PState = getRegState(Op1); 1182 1183 if (Op0.getReg() != Op2.getReg()) { 1184 unsigned S = Op0.getReg() != Op3.getReg() ? PState & ~RegState::Kill 1185 : PState; 1186 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vcmov)) 1187 .add(Op0) 1188 .addReg(PReg, S) 1189 .add(Op2); 1190 if (IsDestLive) 1191 T.addReg(Op0.getReg(), RegState::Implicit); 1192 IsDestLive = true; 1193 } 1194 if (Op0.getReg() != Op3.getReg()) { 1195 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vncmov)) 1196 .add(Op0) 1197 .addReg(PReg, PState) 1198 .add(Op3); 1199 if (IsDestLive) 1200 T.addReg(Op0.getReg(), RegState::Implicit); 1201 } 1202 MBB.erase(MI); 1203 return true; 1204 } 1205 case Hexagon::PS_wselect: { 1206 MachineOperand &Op0 = MI.getOperand(0); 1207 MachineOperand &Op1 = MI.getOperand(1); 1208 MachineOperand &Op2 = MI.getOperand(2); 1209 MachineOperand &Op3 = MI.getOperand(3); 1210 LivePhysRegs LiveAtMI(HRI); 1211 getLiveRegsAt(LiveAtMI, MI); 1212 bool IsDestLive = !LiveAtMI.available(MRI, Op0.getReg()); 1213 unsigned PReg = Op1.getReg(); 1214 assert(Op1.getSubReg() == 0); 1215 unsigned PState = getRegState(Op1); 1216 1217 if (Op0.getReg() != Op2.getReg()) { 1218 unsigned S = Op0.getReg() != Op3.getReg() ? PState & ~RegState::Kill 1219 : PState; 1220 unsigned SrcLo = HRI.getSubReg(Op2.getReg(), Hexagon::vsub_lo); 1221 unsigned SrcHi = HRI.getSubReg(Op2.getReg(), Hexagon::vsub_hi); 1222 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vccombine)) 1223 .add(Op0) 1224 .addReg(PReg, S) 1225 .add(Op1) 1226 .addReg(SrcHi) 1227 .addReg(SrcLo); 1228 if (IsDestLive) 1229 T.addReg(Op0.getReg(), RegState::Implicit); 1230 IsDestLive = true; 1231 } 1232 if (Op0.getReg() != Op3.getReg()) { 1233 unsigned SrcLo = HRI.getSubReg(Op3.getReg(), Hexagon::vsub_lo); 1234 unsigned SrcHi = HRI.getSubReg(Op3.getReg(), Hexagon::vsub_hi); 1235 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vnccombine)) 1236 .add(Op0) 1237 .addReg(PReg, PState) 1238 .addReg(SrcHi) 1239 .addReg(SrcLo); 1240 if (IsDestLive) 1241 T.addReg(Op0.getReg(), RegState::Implicit); 1242 } 1243 MBB.erase(MI); 1244 return true; 1245 } 1246 case Hexagon::PS_tailcall_i: 1247 MI.setDesc(get(Hexagon::J2_jump)); 1248 return true; 1249 case Hexagon::PS_tailcall_r: 1250 case Hexagon::PS_jmpret: 1251 MI.setDesc(get(Hexagon::J2_jumpr)); 1252 return true; 1253 case Hexagon::PS_jmprett: 1254 MI.setDesc(get(Hexagon::J2_jumprt)); 1255 return true; 1256 case Hexagon::PS_jmpretf: 1257 MI.setDesc(get(Hexagon::J2_jumprf)); 1258 return true; 1259 case Hexagon::PS_jmprettnewpt: 1260 MI.setDesc(get(Hexagon::J2_jumprtnewpt)); 1261 return true; 1262 case Hexagon::PS_jmpretfnewpt: 1263 MI.setDesc(get(Hexagon::J2_jumprfnewpt)); 1264 return true; 1265 case Hexagon::PS_jmprettnew: 1266 MI.setDesc(get(Hexagon::J2_jumprtnew)); 1267 return true; 1268 case Hexagon::PS_jmpretfnew: 1269 MI.setDesc(get(Hexagon::J2_jumprfnew)); 1270 return true; 1271 } 1272 1273 return false; 1274 } 1275 1276 // We indicate that we want to reverse the branch by 1277 // inserting the reversed branching opcode. 1278 bool HexagonInstrInfo::reverseBranchCondition( 1279 SmallVectorImpl<MachineOperand> &Cond) const { 1280 if (Cond.empty()) 1281 return true; 1282 assert(Cond[0].isImm() && "First entry in the cond vector not imm-val"); 1283 unsigned opcode = Cond[0].getImm(); 1284 //unsigned temp; 1285 assert(get(opcode).isBranch() && "Should be a branching condition."); 1286 if (isEndLoopN(opcode)) 1287 return true; 1288 unsigned NewOpcode = getInvertedPredicatedOpcode(opcode); 1289 Cond[0].setImm(NewOpcode); 1290 return false; 1291 } 1292 1293 void HexagonInstrInfo::insertNoop(MachineBasicBlock &MBB, 1294 MachineBasicBlock::iterator MI) const { 1295 DebugLoc DL; 1296 BuildMI(MBB, MI, DL, get(Hexagon::A2_nop)); 1297 } 1298 1299 bool HexagonInstrInfo::isPostIncrement(const MachineInstr &MI) const { 1300 return getAddrMode(MI) == HexagonII::PostInc; 1301 } 1302 1303 // Returns true if an instruction is predicated irrespective of the predicate 1304 // sense. For example, all of the following will return true. 1305 // if (p0) R1 = add(R2, R3) 1306 // if (!p0) R1 = add(R2, R3) 1307 // if (p0.new) R1 = add(R2, R3) 1308 // if (!p0.new) R1 = add(R2, R3) 1309 // Note: New-value stores are not included here as in the current 1310 // implementation, we don't need to check their predicate sense. 1311 bool HexagonInstrInfo::isPredicated(const MachineInstr &MI) const { 1312 const uint64_t F = MI.getDesc().TSFlags; 1313 return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask; 1314 } 1315 1316 bool HexagonInstrInfo::PredicateInstruction( 1317 MachineInstr &MI, ArrayRef<MachineOperand> Cond) const { 1318 if (Cond.empty() || isNewValueJump(Cond[0].getImm()) || 1319 isEndLoopN(Cond[0].getImm())) { 1320 DEBUG(dbgs() << "\nCannot predicate:"; MI.dump();); 1321 return false; 1322 } 1323 int Opc = MI.getOpcode(); 1324 assert (isPredicable(MI) && "Expected predicable instruction"); 1325 bool invertJump = predOpcodeHasNot(Cond); 1326 1327 // We have to predicate MI "in place", i.e. after this function returns, 1328 // MI will need to be transformed into a predicated form. To avoid com- 1329 // plicated manipulations with the operands (handling tied operands, 1330 // etc.), build a new temporary instruction, then overwrite MI with it. 1331 1332 MachineBasicBlock &B = *MI.getParent(); 1333 DebugLoc DL = MI.getDebugLoc(); 1334 unsigned PredOpc = getCondOpcode(Opc, invertJump); 1335 MachineInstrBuilder T = BuildMI(B, MI, DL, get(PredOpc)); 1336 unsigned NOp = 0, NumOps = MI.getNumOperands(); 1337 while (NOp < NumOps) { 1338 MachineOperand &Op = MI.getOperand(NOp); 1339 if (!Op.isReg() || !Op.isDef() || Op.isImplicit()) 1340 break; 1341 T.add(Op); 1342 NOp++; 1343 } 1344 1345 unsigned PredReg, PredRegPos, PredRegFlags; 1346 bool GotPredReg = getPredReg(Cond, PredReg, PredRegPos, PredRegFlags); 1347 (void)GotPredReg; 1348 assert(GotPredReg); 1349 T.addReg(PredReg, PredRegFlags); 1350 while (NOp < NumOps) 1351 T.add(MI.getOperand(NOp++)); 1352 1353 MI.setDesc(get(PredOpc)); 1354 while (unsigned n = MI.getNumOperands()) 1355 MI.RemoveOperand(n-1); 1356 for (unsigned i = 0, n = T->getNumOperands(); i < n; ++i) 1357 MI.addOperand(T->getOperand(i)); 1358 1359 MachineBasicBlock::instr_iterator TI = T->getIterator(); 1360 B.erase(TI); 1361 1362 MachineRegisterInfo &MRI = B.getParent()->getRegInfo(); 1363 MRI.clearKillFlags(PredReg); 1364 return true; 1365 } 1366 1367 bool HexagonInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1, 1368 ArrayRef<MachineOperand> Pred2) const { 1369 // TODO: Fix this 1370 return false; 1371 } 1372 1373 bool HexagonInstrInfo::DefinesPredicate(MachineInstr &MI, 1374 std::vector<MachineOperand> &Pred) const { 1375 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 1376 1377 for (unsigned oper = 0; oper < MI.getNumOperands(); ++oper) { 1378 MachineOperand MO = MI.getOperand(oper); 1379 if (MO.isReg()) { 1380 if (!MO.isDef()) 1381 continue; 1382 const TargetRegisterClass* RC = HRI.getMinimalPhysRegClass(MO.getReg()); 1383 if (RC == &Hexagon::PredRegsRegClass) { 1384 Pred.push_back(MO); 1385 return true; 1386 } 1387 continue; 1388 } else if (MO.isRegMask()) { 1389 for (unsigned PR : Hexagon::PredRegsRegClass) { 1390 if (!MI.modifiesRegister(PR, &HRI)) 1391 continue; 1392 Pred.push_back(MO); 1393 return true; 1394 } 1395 } 1396 } 1397 return false; 1398 } 1399 1400 bool HexagonInstrInfo::isPredicable(const MachineInstr &MI) const { 1401 if (!MI.getDesc().isPredicable()) 1402 return false; 1403 1404 if (MI.isCall() || isTailCall(MI)) { 1405 if (!Subtarget.usePredicatedCalls()) 1406 return false; 1407 } 1408 return true; 1409 } 1410 1411 bool HexagonInstrInfo::isSchedulingBoundary(const MachineInstr &MI, 1412 const MachineBasicBlock *MBB, 1413 const MachineFunction &MF) const { 1414 // Debug info is never a scheduling boundary. It's necessary to be explicit 1415 // due to the special treatment of IT instructions below, otherwise a 1416 // dbg_value followed by an IT will result in the IT instruction being 1417 // considered a scheduling hazard, which is wrong. It should be the actual 1418 // instruction preceding the dbg_value instruction(s), just like it is 1419 // when debug info is not present. 1420 if (MI.isDebugValue()) 1421 return false; 1422 1423 // Throwing call is a boundary. 1424 if (MI.isCall()) { 1425 // Don't mess around with no return calls. 1426 if (doesNotReturn(MI)) 1427 return true; 1428 // If any of the block's successors is a landing pad, this could be a 1429 // throwing call. 1430 for (auto I : MBB->successors()) 1431 if (I->isEHPad()) 1432 return true; 1433 } 1434 1435 // Terminators and labels can't be scheduled around. 1436 if (MI.getDesc().isTerminator() || MI.isPosition()) 1437 return true; 1438 1439 if (MI.isInlineAsm() && !ScheduleInlineAsm) 1440 return true; 1441 1442 return false; 1443 } 1444 1445 /// Measure the specified inline asm to determine an approximation of its 1446 /// length. 1447 /// Comments (which run till the next SeparatorString or newline) do not 1448 /// count as an instruction. 1449 /// Any other non-whitespace text is considered an instruction, with 1450 /// multiple instructions separated by SeparatorString or newlines. 1451 /// Variable-length instructions are not handled here; this function 1452 /// may be overloaded in the target code to do that. 1453 /// Hexagon counts the number of ##'s and adjust for that many 1454 /// constant exenders. 1455 unsigned HexagonInstrInfo::getInlineAsmLength(const char *Str, 1456 const MCAsmInfo &MAI) const { 1457 StringRef AStr(Str); 1458 // Count the number of instructions in the asm. 1459 bool atInsnStart = true; 1460 unsigned Length = 0; 1461 for (; *Str; ++Str) { 1462 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(), 1463 strlen(MAI.getSeparatorString())) == 0) 1464 atInsnStart = true; 1465 if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) { 1466 Length += MAI.getMaxInstLength(); 1467 atInsnStart = false; 1468 } 1469 if (atInsnStart && strncmp(Str, MAI.getCommentString().data(), 1470 MAI.getCommentString().size()) == 0) 1471 atInsnStart = false; 1472 } 1473 1474 // Add to size number of constant extenders seen * 4. 1475 StringRef Occ("##"); 1476 Length += AStr.count(Occ)*4; 1477 return Length; 1478 } 1479 1480 ScheduleHazardRecognizer* 1481 HexagonInstrInfo::CreateTargetPostRAHazardRecognizer( 1482 const InstrItineraryData *II, const ScheduleDAG *DAG) const { 1483 if (UseDFAHazardRec) 1484 return new HexagonHazardRecognizer(II, this, Subtarget); 1485 return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG); 1486 } 1487 1488 /// \brief For a comparison instruction, return the source registers in 1489 /// \p SrcReg and \p SrcReg2 if having two register operands, and the value it 1490 /// compares against in CmpValue. Return true if the comparison instruction 1491 /// can be analyzed. 1492 bool HexagonInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg, 1493 unsigned &SrcReg2, int &Mask, 1494 int &Value) const { 1495 unsigned Opc = MI.getOpcode(); 1496 1497 // Set mask and the first source register. 1498 switch (Opc) { 1499 case Hexagon::C2_cmpeq: 1500 case Hexagon::C2_cmpeqp: 1501 case Hexagon::C2_cmpgt: 1502 case Hexagon::C2_cmpgtp: 1503 case Hexagon::C2_cmpgtu: 1504 case Hexagon::C2_cmpgtup: 1505 case Hexagon::C4_cmpneq: 1506 case Hexagon::C4_cmplte: 1507 case Hexagon::C4_cmplteu: 1508 case Hexagon::C2_cmpeqi: 1509 case Hexagon::C2_cmpgti: 1510 case Hexagon::C2_cmpgtui: 1511 case Hexagon::C4_cmpneqi: 1512 case Hexagon::C4_cmplteui: 1513 case Hexagon::C4_cmpltei: 1514 SrcReg = MI.getOperand(1).getReg(); 1515 Mask = ~0; 1516 break; 1517 case Hexagon::A4_cmpbeq: 1518 case Hexagon::A4_cmpbgt: 1519 case Hexagon::A4_cmpbgtu: 1520 case Hexagon::A4_cmpbeqi: 1521 case Hexagon::A4_cmpbgti: 1522 case Hexagon::A4_cmpbgtui: 1523 SrcReg = MI.getOperand(1).getReg(); 1524 Mask = 0xFF; 1525 break; 1526 case Hexagon::A4_cmpheq: 1527 case Hexagon::A4_cmphgt: 1528 case Hexagon::A4_cmphgtu: 1529 case Hexagon::A4_cmpheqi: 1530 case Hexagon::A4_cmphgti: 1531 case Hexagon::A4_cmphgtui: 1532 SrcReg = MI.getOperand(1).getReg(); 1533 Mask = 0xFFFF; 1534 break; 1535 } 1536 1537 // Set the value/second source register. 1538 switch (Opc) { 1539 case Hexagon::C2_cmpeq: 1540 case Hexagon::C2_cmpeqp: 1541 case Hexagon::C2_cmpgt: 1542 case Hexagon::C2_cmpgtp: 1543 case Hexagon::C2_cmpgtu: 1544 case Hexagon::C2_cmpgtup: 1545 case Hexagon::A4_cmpbeq: 1546 case Hexagon::A4_cmpbgt: 1547 case Hexagon::A4_cmpbgtu: 1548 case Hexagon::A4_cmpheq: 1549 case Hexagon::A4_cmphgt: 1550 case Hexagon::A4_cmphgtu: 1551 case Hexagon::C4_cmpneq: 1552 case Hexagon::C4_cmplte: 1553 case Hexagon::C4_cmplteu: 1554 SrcReg2 = MI.getOperand(2).getReg(); 1555 return true; 1556 1557 case Hexagon::C2_cmpeqi: 1558 case Hexagon::C2_cmpgtui: 1559 case Hexagon::C2_cmpgti: 1560 case Hexagon::C4_cmpneqi: 1561 case Hexagon::C4_cmplteui: 1562 case Hexagon::C4_cmpltei: 1563 case Hexagon::A4_cmpbeqi: 1564 case Hexagon::A4_cmpbgti: 1565 case Hexagon::A4_cmpbgtui: 1566 case Hexagon::A4_cmpheqi: 1567 case Hexagon::A4_cmphgti: 1568 case Hexagon::A4_cmphgtui: 1569 SrcReg2 = 0; 1570 Value = MI.getOperand(2).getImm(); 1571 return true; 1572 } 1573 1574 return false; 1575 } 1576 1577 unsigned HexagonInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 1578 const MachineInstr &MI, 1579 unsigned *PredCost) const { 1580 return getInstrTimingClassLatency(ItinData, MI); 1581 } 1582 1583 DFAPacketizer *HexagonInstrInfo::CreateTargetScheduleState( 1584 const TargetSubtargetInfo &STI) const { 1585 const InstrItineraryData *II = STI.getInstrItineraryData(); 1586 return static_cast<const HexagonSubtarget&>(STI).createDFAPacketizer(II); 1587 } 1588 1589 // Inspired by this pair: 1590 // %R13<def> = L2_loadri_io %R29, 136; mem:LD4[FixedStack0] 1591 // S2_storeri_io %R29, 132, %R1<kill>; flags: mem:ST4[FixedStack1] 1592 // Currently AA considers the addresses in these instructions to be aliasing. 1593 bool HexagonInstrInfo::areMemAccessesTriviallyDisjoint( 1594 MachineInstr &MIa, MachineInstr &MIb, AliasAnalysis *AA) const { 1595 if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects() || 1596 MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef()) 1597 return false; 1598 1599 // Instructions that are pure loads, not loads and stores like memops are not 1600 // dependent. 1601 if (MIa.mayLoad() && !isMemOp(MIa) && MIb.mayLoad() && !isMemOp(MIb)) 1602 return true; 1603 1604 // Get the base register in MIa. 1605 unsigned BasePosA, OffsetPosA; 1606 if (!getBaseAndOffsetPosition(MIa, BasePosA, OffsetPosA)) 1607 return false; 1608 const MachineOperand &BaseA = MIa.getOperand(BasePosA); 1609 unsigned BaseRegA = BaseA.getReg(); 1610 unsigned BaseSubA = BaseA.getSubReg(); 1611 1612 // Get the base register in MIb. 1613 unsigned BasePosB, OffsetPosB; 1614 if (!getBaseAndOffsetPosition(MIb, BasePosB, OffsetPosB)) 1615 return false; 1616 const MachineOperand &BaseB = MIb.getOperand(BasePosB); 1617 unsigned BaseRegB = BaseB.getReg(); 1618 unsigned BaseSubB = BaseB.getSubReg(); 1619 1620 if (BaseRegA != BaseRegB || BaseSubA != BaseSubB) 1621 return false; 1622 1623 // Get the access sizes. 1624 unsigned SizeA = getMemAccessSize(MIa); 1625 unsigned SizeB = getMemAccessSize(MIb); 1626 1627 // Get the offsets. Handle immediates only for now. 1628 const MachineOperand &OffA = MIa.getOperand(OffsetPosA); 1629 const MachineOperand &OffB = MIb.getOperand(OffsetPosB); 1630 if (!MIa.getOperand(OffsetPosA).isImm() || 1631 !MIb.getOperand(OffsetPosB).isImm()) 1632 return false; 1633 int OffsetA = isPostIncrement(MIa) ? 0 : OffA.getImm(); 1634 int OffsetB = isPostIncrement(MIb) ? 0 : OffB.getImm(); 1635 1636 // This is a mem access with the same base register and known offsets from it. 1637 // Reason about it. 1638 if (OffsetA > OffsetB) { 1639 uint64_t OffDiff = (uint64_t)((int64_t)OffsetA - (int64_t)OffsetB); 1640 return SizeB <= OffDiff; 1641 } 1642 if (OffsetA < OffsetB) { 1643 uint64_t OffDiff = (uint64_t)((int64_t)OffsetB - (int64_t)OffsetA); 1644 return SizeA <= OffDiff; 1645 } 1646 1647 return false; 1648 } 1649 1650 /// If the instruction is an increment of a constant value, return the amount. 1651 bool HexagonInstrInfo::getIncrementValue(const MachineInstr &MI, 1652 int &Value) const { 1653 if (isPostIncrement(MI)) { 1654 unsigned AccessSize; 1655 return getBaseAndOffset(MI, Value, AccessSize); 1656 } 1657 if (MI.getOpcode() == Hexagon::A2_addi) { 1658 Value = MI.getOperand(2).getImm(); 1659 return true; 1660 } 1661 1662 return false; 1663 } 1664 1665 std::pair<unsigned, unsigned> 1666 HexagonInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const { 1667 return std::make_pair(TF & ~HexagonII::MO_Bitmasks, 1668 TF & HexagonII::MO_Bitmasks); 1669 } 1670 1671 ArrayRef<std::pair<unsigned, const char*>> 1672 HexagonInstrInfo::getSerializableDirectMachineOperandTargetFlags() const { 1673 using namespace HexagonII; 1674 1675 static const std::pair<unsigned, const char*> Flags[] = { 1676 {MO_PCREL, "hexagon-pcrel"}, 1677 {MO_GOT, "hexagon-got"}, 1678 {MO_LO16, "hexagon-lo16"}, 1679 {MO_HI16, "hexagon-hi16"}, 1680 {MO_GPREL, "hexagon-gprel"}, 1681 {MO_GDGOT, "hexagon-gdgot"}, 1682 {MO_GDPLT, "hexagon-gdplt"}, 1683 {MO_IE, "hexagon-ie"}, 1684 {MO_IEGOT, "hexagon-iegot"}, 1685 {MO_TPREL, "hexagon-tprel"} 1686 }; 1687 return makeArrayRef(Flags); 1688 } 1689 1690 ArrayRef<std::pair<unsigned, const char*>> 1691 HexagonInstrInfo::getSerializableBitmaskMachineOperandTargetFlags() const { 1692 using namespace HexagonII; 1693 1694 static const std::pair<unsigned, const char*> Flags[] = { 1695 {HMOTF_ConstExtended, "hexagon-ext"} 1696 }; 1697 return makeArrayRef(Flags); 1698 } 1699 1700 unsigned HexagonInstrInfo::createVR(MachineFunction *MF, MVT VT) const { 1701 MachineRegisterInfo &MRI = MF->getRegInfo(); 1702 const TargetRegisterClass *TRC; 1703 if (VT == MVT::i1) { 1704 TRC = &Hexagon::PredRegsRegClass; 1705 } else if (VT == MVT::i32 || VT == MVT::f32) { 1706 TRC = &Hexagon::IntRegsRegClass; 1707 } else if (VT == MVT::i64 || VT == MVT::f64) { 1708 TRC = &Hexagon::DoubleRegsRegClass; 1709 } else { 1710 llvm_unreachable("Cannot handle this register class"); 1711 } 1712 1713 unsigned NewReg = MRI.createVirtualRegister(TRC); 1714 return NewReg; 1715 } 1716 1717 bool HexagonInstrInfo::isAbsoluteSet(const MachineInstr &MI) const { 1718 return (getAddrMode(MI) == HexagonII::AbsoluteSet); 1719 } 1720 1721 bool HexagonInstrInfo::isAccumulator(const MachineInstr &MI) const { 1722 const uint64_t F = MI.getDesc().TSFlags; 1723 return((F >> HexagonII::AccumulatorPos) & HexagonII::AccumulatorMask); 1724 } 1725 1726 bool HexagonInstrInfo::isComplex(const MachineInstr &MI) const { 1727 return !isTC1(MI) && !isTC2Early(MI) && !MI.getDesc().mayLoad() && 1728 !MI.getDesc().mayStore() && 1729 MI.getDesc().getOpcode() != Hexagon::S2_allocframe && 1730 MI.getDesc().getOpcode() != Hexagon::L2_deallocframe && 1731 !isMemOp(MI) && !MI.isBranch() && !MI.isReturn() && !MI.isCall(); 1732 } 1733 1734 // Return true if the instruction is a compund branch instruction. 1735 bool HexagonInstrInfo::isCompoundBranchInstr(const MachineInstr &MI) const { 1736 return getType(MI) == HexagonII::TypeCJ && MI.isBranch(); 1737 } 1738 1739 // TODO: In order to have isExtendable for fpimm/f32Ext, we need to handle 1740 // isFPImm and later getFPImm as well. 1741 bool HexagonInstrInfo::isConstExtended(const MachineInstr &MI) const { 1742 const uint64_t F = MI.getDesc().TSFlags; 1743 unsigned isExtended = (F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask; 1744 if (isExtended) // Instruction must be extended. 1745 return true; 1746 1747 unsigned isExtendable = 1748 (F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask; 1749 if (!isExtendable) 1750 return false; 1751 1752 if (MI.isCall()) 1753 return false; 1754 1755 short ExtOpNum = getCExtOpNum(MI); 1756 const MachineOperand &MO = MI.getOperand(ExtOpNum); 1757 // Use MO operand flags to determine if MO 1758 // has the HMOTF_ConstExtended flag set. 1759 if (MO.getTargetFlags() & HexagonII::HMOTF_ConstExtended) 1760 return true; 1761 // If this is a Machine BB address we are talking about, and it is 1762 // not marked as extended, say so. 1763 if (MO.isMBB()) 1764 return false; 1765 1766 // We could be using an instruction with an extendable immediate and shoehorn 1767 // a global address into it. If it is a global address it will be constant 1768 // extended. We do this for COMBINE. 1769 if (MO.isGlobal() || MO.isSymbol() || MO.isBlockAddress() || 1770 MO.isJTI() || MO.isCPI() || MO.isFPImm()) 1771 return true; 1772 1773 // If the extendable operand is not 'Immediate' type, the instruction should 1774 // have 'isExtended' flag set. 1775 assert(MO.isImm() && "Extendable operand must be Immediate type"); 1776 1777 int MinValue = getMinValue(MI); 1778 int MaxValue = getMaxValue(MI); 1779 int ImmValue = MO.getImm(); 1780 1781 return (ImmValue < MinValue || ImmValue > MaxValue); 1782 } 1783 1784 bool HexagonInstrInfo::isDeallocRet(const MachineInstr &MI) const { 1785 switch (MI.getOpcode()) { 1786 case Hexagon::L4_return: 1787 case Hexagon::L4_return_t: 1788 case Hexagon::L4_return_f: 1789 case Hexagon::L4_return_tnew_pnt: 1790 case Hexagon::L4_return_fnew_pnt: 1791 case Hexagon::L4_return_tnew_pt: 1792 case Hexagon::L4_return_fnew_pt: 1793 return true; 1794 } 1795 return false; 1796 } 1797 1798 // Return true when ConsMI uses a register defined by ProdMI. 1799 bool HexagonInstrInfo::isDependent(const MachineInstr &ProdMI, 1800 const MachineInstr &ConsMI) const { 1801 if (!ProdMI.getDesc().getNumDefs()) 1802 return false; 1803 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 1804 1805 SmallVector<unsigned, 4> DefsA; 1806 SmallVector<unsigned, 4> DefsB; 1807 SmallVector<unsigned, 8> UsesA; 1808 SmallVector<unsigned, 8> UsesB; 1809 1810 parseOperands(ProdMI, DefsA, UsesA); 1811 parseOperands(ConsMI, DefsB, UsesB); 1812 1813 for (auto &RegA : DefsA) 1814 for (auto &RegB : UsesB) { 1815 // True data dependency. 1816 if (RegA == RegB) 1817 return true; 1818 1819 if (TargetRegisterInfo::isPhysicalRegister(RegA)) 1820 for (MCSubRegIterator SubRegs(RegA, &HRI); SubRegs.isValid(); ++SubRegs) 1821 if (RegB == *SubRegs) 1822 return true; 1823 1824 if (TargetRegisterInfo::isPhysicalRegister(RegB)) 1825 for (MCSubRegIterator SubRegs(RegB, &HRI); SubRegs.isValid(); ++SubRegs) 1826 if (RegA == *SubRegs) 1827 return true; 1828 } 1829 1830 return false; 1831 } 1832 1833 // Returns true if the instruction is alread a .cur. 1834 bool HexagonInstrInfo::isDotCurInst(const MachineInstr &MI) const { 1835 switch (MI.getOpcode()) { 1836 case Hexagon::V6_vL32b_cur_pi: 1837 case Hexagon::V6_vL32b_cur_ai: 1838 return true; 1839 } 1840 return false; 1841 } 1842 1843 // Returns true, if any one of the operands is a dot new 1844 // insn, whether it is predicated dot new or register dot new. 1845 bool HexagonInstrInfo::isDotNewInst(const MachineInstr &MI) const { 1846 if (isNewValueInst(MI) || (isPredicated(MI) && isPredicatedNew(MI))) 1847 return true; 1848 1849 return false; 1850 } 1851 1852 /// Symmetrical. See if these two instructions are fit for duplex pair. 1853 bool HexagonInstrInfo::isDuplexPair(const MachineInstr &MIa, 1854 const MachineInstr &MIb) const { 1855 HexagonII::SubInstructionGroup MIaG = getDuplexCandidateGroup(MIa); 1856 HexagonII::SubInstructionGroup MIbG = getDuplexCandidateGroup(MIb); 1857 return (isDuplexPairMatch(MIaG, MIbG) || isDuplexPairMatch(MIbG, MIaG)); 1858 } 1859 1860 bool HexagonInstrInfo::isEarlySourceInstr(const MachineInstr &MI) const { 1861 if (MI.mayLoad() || MI.mayStore() || MI.isCompare()) 1862 return true; 1863 1864 // Multiply 1865 unsigned SchedClass = MI.getDesc().getSchedClass(); 1866 return is_TC4x(SchedClass) || is_TC3x(SchedClass); 1867 } 1868 1869 bool HexagonInstrInfo::isEndLoopN(unsigned Opcode) const { 1870 return (Opcode == Hexagon::ENDLOOP0 || 1871 Opcode == Hexagon::ENDLOOP1); 1872 } 1873 1874 bool HexagonInstrInfo::isExpr(unsigned OpType) const { 1875 switch(OpType) { 1876 case MachineOperand::MO_MachineBasicBlock: 1877 case MachineOperand::MO_GlobalAddress: 1878 case MachineOperand::MO_ExternalSymbol: 1879 case MachineOperand::MO_JumpTableIndex: 1880 case MachineOperand::MO_ConstantPoolIndex: 1881 case MachineOperand::MO_BlockAddress: 1882 return true; 1883 default: 1884 return false; 1885 } 1886 } 1887 1888 bool HexagonInstrInfo::isExtendable(const MachineInstr &MI) const { 1889 const MCInstrDesc &MID = MI.getDesc(); 1890 const uint64_t F = MID.TSFlags; 1891 if ((F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask) 1892 return true; 1893 1894 // TODO: This is largely obsolete now. Will need to be removed 1895 // in consecutive patches. 1896 switch (MI.getOpcode()) { 1897 // PS_fi and PS_fia remain special cases. 1898 case Hexagon::PS_fi: 1899 case Hexagon::PS_fia: 1900 return true; 1901 default: 1902 return false; 1903 } 1904 return false; 1905 } 1906 1907 // This returns true in two cases: 1908 // - The OP code itself indicates that this is an extended instruction. 1909 // - One of MOs has been marked with HMOTF_ConstExtended flag. 1910 bool HexagonInstrInfo::isExtended(const MachineInstr &MI) const { 1911 // First check if this is permanently extended op code. 1912 const uint64_t F = MI.getDesc().TSFlags; 1913 if ((F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask) 1914 return true; 1915 // Use MO operand flags to determine if one of MI's operands 1916 // has HMOTF_ConstExtended flag set. 1917 for (const MachineOperand &MO : MI.operands()) 1918 if (MO.getTargetFlags() & HexagonII::HMOTF_ConstExtended) 1919 return true; 1920 return false; 1921 } 1922 1923 bool HexagonInstrInfo::isFloat(const MachineInstr &MI) const { 1924 unsigned Opcode = MI.getOpcode(); 1925 const uint64_t F = get(Opcode).TSFlags; 1926 return (F >> HexagonII::FPPos) & HexagonII::FPMask; 1927 } 1928 1929 // No V60 HVX VMEM with A_INDIRECT. 1930 bool HexagonInstrInfo::isHVXMemWithAIndirect(const MachineInstr &I, 1931 const MachineInstr &J) const { 1932 if (!isHVXVec(I)) 1933 return false; 1934 if (!I.mayLoad() && !I.mayStore()) 1935 return false; 1936 return J.isIndirectBranch() || isIndirectCall(J) || isIndirectL4Return(J); 1937 } 1938 1939 bool HexagonInstrInfo::isIndirectCall(const MachineInstr &MI) const { 1940 switch (MI.getOpcode()) { 1941 case Hexagon::J2_callr: 1942 case Hexagon::J2_callrf: 1943 case Hexagon::J2_callrt: 1944 case Hexagon::PS_call_nr: 1945 return true; 1946 } 1947 return false; 1948 } 1949 1950 bool HexagonInstrInfo::isIndirectL4Return(const MachineInstr &MI) const { 1951 switch (MI.getOpcode()) { 1952 case Hexagon::L4_return: 1953 case Hexagon::L4_return_t: 1954 case Hexagon::L4_return_f: 1955 case Hexagon::L4_return_fnew_pnt: 1956 case Hexagon::L4_return_fnew_pt: 1957 case Hexagon::L4_return_tnew_pnt: 1958 case Hexagon::L4_return_tnew_pt: 1959 return true; 1960 } 1961 return false; 1962 } 1963 1964 bool HexagonInstrInfo::isJumpR(const MachineInstr &MI) const { 1965 switch (MI.getOpcode()) { 1966 case Hexagon::J2_jumpr: 1967 case Hexagon::J2_jumprt: 1968 case Hexagon::J2_jumprf: 1969 case Hexagon::J2_jumprtnewpt: 1970 case Hexagon::J2_jumprfnewpt: 1971 case Hexagon::J2_jumprtnew: 1972 case Hexagon::J2_jumprfnew: 1973 return true; 1974 } 1975 return false; 1976 } 1977 1978 // Return true if a given MI can accommodate given offset. 1979 // Use abs estimate as oppose to the exact number. 1980 // TODO: This will need to be changed to use MC level 1981 // definition of instruction extendable field size. 1982 bool HexagonInstrInfo::isJumpWithinBranchRange(const MachineInstr &MI, 1983 unsigned offset) const { 1984 // This selection of jump instructions matches to that what 1985 // analyzeBranch can parse, plus NVJ. 1986 if (isNewValueJump(MI)) // r9:2 1987 return isInt<11>(offset); 1988 1989 switch (MI.getOpcode()) { 1990 // Still missing Jump to address condition on register value. 1991 default: 1992 return false; 1993 case Hexagon::J2_jump: // bits<24> dst; // r22:2 1994 case Hexagon::J2_call: 1995 case Hexagon::PS_call_nr: 1996 return isInt<24>(offset); 1997 case Hexagon::J2_jumpt: //bits<17> dst; // r15:2 1998 case Hexagon::J2_jumpf: 1999 case Hexagon::J2_jumptnew: 2000 case Hexagon::J2_jumptnewpt: 2001 case Hexagon::J2_jumpfnew: 2002 case Hexagon::J2_jumpfnewpt: 2003 case Hexagon::J2_callt: 2004 case Hexagon::J2_callf: 2005 return isInt<17>(offset); 2006 case Hexagon::J2_loop0i: 2007 case Hexagon::J2_loop0iext: 2008 case Hexagon::J2_loop0r: 2009 case Hexagon::J2_loop0rext: 2010 case Hexagon::J2_loop1i: 2011 case Hexagon::J2_loop1iext: 2012 case Hexagon::J2_loop1r: 2013 case Hexagon::J2_loop1rext: 2014 return isInt<9>(offset); 2015 // TODO: Add all the compound branches here. Can we do this in Relation model? 2016 case Hexagon::J4_cmpeqi_tp0_jump_nt: 2017 case Hexagon::J4_cmpeqi_tp1_jump_nt: 2018 return isInt<11>(offset); 2019 } 2020 } 2021 2022 bool HexagonInstrInfo::isLateInstrFeedsEarlyInstr(const MachineInstr &LRMI, 2023 const MachineInstr &ESMI) const { 2024 bool isLate = isLateResultInstr(LRMI); 2025 bool isEarly = isEarlySourceInstr(ESMI); 2026 2027 DEBUG(dbgs() << "V60" << (isLate ? "-LR " : " -- ")); 2028 DEBUG(LRMI.dump()); 2029 DEBUG(dbgs() << "V60" << (isEarly ? "-ES " : " -- ")); 2030 DEBUG(ESMI.dump()); 2031 2032 if (isLate && isEarly) { 2033 DEBUG(dbgs() << "++Is Late Result feeding Early Source\n"); 2034 return true; 2035 } 2036 2037 return false; 2038 } 2039 2040 bool HexagonInstrInfo::isLateResultInstr(const MachineInstr &MI) const { 2041 switch (MI.getOpcode()) { 2042 case TargetOpcode::EXTRACT_SUBREG: 2043 case TargetOpcode::INSERT_SUBREG: 2044 case TargetOpcode::SUBREG_TO_REG: 2045 case TargetOpcode::REG_SEQUENCE: 2046 case TargetOpcode::IMPLICIT_DEF: 2047 case TargetOpcode::COPY: 2048 case TargetOpcode::INLINEASM: 2049 case TargetOpcode::PHI: 2050 return false; 2051 default: 2052 break; 2053 } 2054 2055 unsigned SchedClass = MI.getDesc().getSchedClass(); 2056 return !is_TC1(SchedClass); 2057 } 2058 2059 bool HexagonInstrInfo::isLateSourceInstr(const MachineInstr &MI) const { 2060 // Instructions with iclass A_CVI_VX and attribute A_CVI_LATE uses a multiply 2061 // resource, but all operands can be received late like an ALU instruction. 2062 return getType(MI) == HexagonII::TypeCVI_VX_LATE; 2063 } 2064 2065 bool HexagonInstrInfo::isLoopN(const MachineInstr &MI) const { 2066 unsigned Opcode = MI.getOpcode(); 2067 return Opcode == Hexagon::J2_loop0i || 2068 Opcode == Hexagon::J2_loop0r || 2069 Opcode == Hexagon::J2_loop0iext || 2070 Opcode == Hexagon::J2_loop0rext || 2071 Opcode == Hexagon::J2_loop1i || 2072 Opcode == Hexagon::J2_loop1r || 2073 Opcode == Hexagon::J2_loop1iext || 2074 Opcode == Hexagon::J2_loop1rext; 2075 } 2076 2077 bool HexagonInstrInfo::isMemOp(const MachineInstr &MI) const { 2078 switch (MI.getOpcode()) { 2079 default: return false; 2080 case Hexagon::L4_iadd_memopw_io: 2081 case Hexagon::L4_isub_memopw_io: 2082 case Hexagon::L4_add_memopw_io: 2083 case Hexagon::L4_sub_memopw_io: 2084 case Hexagon::L4_and_memopw_io: 2085 case Hexagon::L4_or_memopw_io: 2086 case Hexagon::L4_iadd_memoph_io: 2087 case Hexagon::L4_isub_memoph_io: 2088 case Hexagon::L4_add_memoph_io: 2089 case Hexagon::L4_sub_memoph_io: 2090 case Hexagon::L4_and_memoph_io: 2091 case Hexagon::L4_or_memoph_io: 2092 case Hexagon::L4_iadd_memopb_io: 2093 case Hexagon::L4_isub_memopb_io: 2094 case Hexagon::L4_add_memopb_io: 2095 case Hexagon::L4_sub_memopb_io: 2096 case Hexagon::L4_and_memopb_io: 2097 case Hexagon::L4_or_memopb_io: 2098 case Hexagon::L4_ior_memopb_io: 2099 case Hexagon::L4_ior_memoph_io: 2100 case Hexagon::L4_ior_memopw_io: 2101 case Hexagon::L4_iand_memopb_io: 2102 case Hexagon::L4_iand_memoph_io: 2103 case Hexagon::L4_iand_memopw_io: 2104 return true; 2105 } 2106 return false; 2107 } 2108 2109 bool HexagonInstrInfo::isNewValue(const MachineInstr &MI) const { 2110 const uint64_t F = MI.getDesc().TSFlags; 2111 return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask; 2112 } 2113 2114 bool HexagonInstrInfo::isNewValue(unsigned Opcode) const { 2115 const uint64_t F = get(Opcode).TSFlags; 2116 return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask; 2117 } 2118 2119 bool HexagonInstrInfo::isNewValueInst(const MachineInstr &MI) const { 2120 return isNewValueJump(MI) || isNewValueStore(MI); 2121 } 2122 2123 bool HexagonInstrInfo::isNewValueJump(const MachineInstr &MI) const { 2124 return isNewValue(MI) && MI.isBranch(); 2125 } 2126 2127 bool HexagonInstrInfo::isNewValueJump(unsigned Opcode) const { 2128 return isNewValue(Opcode) && get(Opcode).isBranch() && isPredicated(Opcode); 2129 } 2130 2131 bool HexagonInstrInfo::isNewValueStore(const MachineInstr &MI) const { 2132 const uint64_t F = MI.getDesc().TSFlags; 2133 return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask; 2134 } 2135 2136 bool HexagonInstrInfo::isNewValueStore(unsigned Opcode) const { 2137 const uint64_t F = get(Opcode).TSFlags; 2138 return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask; 2139 } 2140 2141 // Returns true if a particular operand is extendable for an instruction. 2142 bool HexagonInstrInfo::isOperandExtended(const MachineInstr &MI, 2143 unsigned OperandNum) const { 2144 const uint64_t F = MI.getDesc().TSFlags; 2145 return ((F >> HexagonII::ExtendableOpPos) & HexagonII::ExtendableOpMask) 2146 == OperandNum; 2147 } 2148 2149 bool HexagonInstrInfo::isPredicatedNew(const MachineInstr &MI) const { 2150 const uint64_t F = MI.getDesc().TSFlags; 2151 assert(isPredicated(MI)); 2152 return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask; 2153 } 2154 2155 bool HexagonInstrInfo::isPredicatedNew(unsigned Opcode) const { 2156 const uint64_t F = get(Opcode).TSFlags; 2157 assert(isPredicated(Opcode)); 2158 return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask; 2159 } 2160 2161 bool HexagonInstrInfo::isPredicatedTrue(const MachineInstr &MI) const { 2162 const uint64_t F = MI.getDesc().TSFlags; 2163 return !((F >> HexagonII::PredicatedFalsePos) & 2164 HexagonII::PredicatedFalseMask); 2165 } 2166 2167 bool HexagonInstrInfo::isPredicatedTrue(unsigned Opcode) const { 2168 const uint64_t F = get(Opcode).TSFlags; 2169 // Make sure that the instruction is predicated. 2170 assert((F>> HexagonII::PredicatedPos) & HexagonII::PredicatedMask); 2171 return !((F >> HexagonII::PredicatedFalsePos) & 2172 HexagonII::PredicatedFalseMask); 2173 } 2174 2175 bool HexagonInstrInfo::isPredicated(unsigned Opcode) const { 2176 const uint64_t F = get(Opcode).TSFlags; 2177 return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask; 2178 } 2179 2180 bool HexagonInstrInfo::isPredicateLate(unsigned Opcode) const { 2181 const uint64_t F = get(Opcode).TSFlags; 2182 return ~(F >> HexagonII::PredicateLatePos) & HexagonII::PredicateLateMask; 2183 } 2184 2185 bool HexagonInstrInfo::isPredictedTaken(unsigned Opcode) const { 2186 const uint64_t F = get(Opcode).TSFlags; 2187 assert(get(Opcode).isBranch() && 2188 (isPredicatedNew(Opcode) || isNewValue(Opcode))); 2189 return (F >> HexagonII::TakenPos) & HexagonII::TakenMask; 2190 } 2191 2192 bool HexagonInstrInfo::isSaveCalleeSavedRegsCall(const MachineInstr &MI) const { 2193 return MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4 || 2194 MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT || 2195 MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_PIC || 2196 MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT_PIC; 2197 } 2198 2199 bool HexagonInstrInfo::isSignExtendingLoad(const MachineInstr &MI) const { 2200 switch (MI.getOpcode()) { 2201 // Byte 2202 case Hexagon::L2_loadrb_io: 2203 case Hexagon::L4_loadrb_ur: 2204 case Hexagon::L4_loadrb_ap: 2205 case Hexagon::L2_loadrb_pr: 2206 case Hexagon::L2_loadrb_pbr: 2207 case Hexagon::L2_loadrb_pi: 2208 case Hexagon::L2_loadrb_pci: 2209 case Hexagon::L2_loadrb_pcr: 2210 case Hexagon::L2_loadbsw2_io: 2211 case Hexagon::L4_loadbsw2_ur: 2212 case Hexagon::L4_loadbsw2_ap: 2213 case Hexagon::L2_loadbsw2_pr: 2214 case Hexagon::L2_loadbsw2_pbr: 2215 case Hexagon::L2_loadbsw2_pi: 2216 case Hexagon::L2_loadbsw2_pci: 2217 case Hexagon::L2_loadbsw2_pcr: 2218 case Hexagon::L2_loadbsw4_io: 2219 case Hexagon::L4_loadbsw4_ur: 2220 case Hexagon::L4_loadbsw4_ap: 2221 case Hexagon::L2_loadbsw4_pr: 2222 case Hexagon::L2_loadbsw4_pbr: 2223 case Hexagon::L2_loadbsw4_pi: 2224 case Hexagon::L2_loadbsw4_pci: 2225 case Hexagon::L2_loadbsw4_pcr: 2226 case Hexagon::L4_loadrb_rr: 2227 case Hexagon::L2_ploadrbt_io: 2228 case Hexagon::L2_ploadrbt_pi: 2229 case Hexagon::L2_ploadrbf_io: 2230 case Hexagon::L2_ploadrbf_pi: 2231 case Hexagon::L2_ploadrbtnew_io: 2232 case Hexagon::L2_ploadrbfnew_io: 2233 case Hexagon::L4_ploadrbt_rr: 2234 case Hexagon::L4_ploadrbf_rr: 2235 case Hexagon::L4_ploadrbtnew_rr: 2236 case Hexagon::L4_ploadrbfnew_rr: 2237 case Hexagon::L2_ploadrbtnew_pi: 2238 case Hexagon::L2_ploadrbfnew_pi: 2239 case Hexagon::L4_ploadrbt_abs: 2240 case Hexagon::L4_ploadrbf_abs: 2241 case Hexagon::L4_ploadrbtnew_abs: 2242 case Hexagon::L4_ploadrbfnew_abs: 2243 case Hexagon::L2_loadrbgp: 2244 // Half 2245 case Hexagon::L2_loadrh_io: 2246 case Hexagon::L4_loadrh_ur: 2247 case Hexagon::L4_loadrh_ap: 2248 case Hexagon::L2_loadrh_pr: 2249 case Hexagon::L2_loadrh_pbr: 2250 case Hexagon::L2_loadrh_pi: 2251 case Hexagon::L2_loadrh_pci: 2252 case Hexagon::L2_loadrh_pcr: 2253 case Hexagon::L4_loadrh_rr: 2254 case Hexagon::L2_ploadrht_io: 2255 case Hexagon::L2_ploadrht_pi: 2256 case Hexagon::L2_ploadrhf_io: 2257 case Hexagon::L2_ploadrhf_pi: 2258 case Hexagon::L2_ploadrhtnew_io: 2259 case Hexagon::L2_ploadrhfnew_io: 2260 case Hexagon::L4_ploadrht_rr: 2261 case Hexagon::L4_ploadrhf_rr: 2262 case Hexagon::L4_ploadrhtnew_rr: 2263 case Hexagon::L4_ploadrhfnew_rr: 2264 case Hexagon::L2_ploadrhtnew_pi: 2265 case Hexagon::L2_ploadrhfnew_pi: 2266 case Hexagon::L4_ploadrht_abs: 2267 case Hexagon::L4_ploadrhf_abs: 2268 case Hexagon::L4_ploadrhtnew_abs: 2269 case Hexagon::L4_ploadrhfnew_abs: 2270 case Hexagon::L2_loadrhgp: 2271 return true; 2272 default: 2273 return false; 2274 } 2275 } 2276 2277 bool HexagonInstrInfo::isSolo(const MachineInstr &MI) const { 2278 const uint64_t F = MI.getDesc().TSFlags; 2279 return (F >> HexagonII::SoloPos) & HexagonII::SoloMask; 2280 } 2281 2282 bool HexagonInstrInfo::isSpillPredRegOp(const MachineInstr &MI) const { 2283 switch (MI.getOpcode()) { 2284 case Hexagon::STriw_pred: 2285 case Hexagon::LDriw_pred: 2286 return true; 2287 default: 2288 return false; 2289 } 2290 } 2291 2292 bool HexagonInstrInfo::isTailCall(const MachineInstr &MI) const { 2293 if (!MI.isBranch()) 2294 return false; 2295 2296 for (auto &Op : MI.operands()) 2297 if (Op.isGlobal() || Op.isSymbol()) 2298 return true; 2299 return false; 2300 } 2301 2302 // Returns true when SU has a timing class TC1. 2303 bool HexagonInstrInfo::isTC1(const MachineInstr &MI) const { 2304 unsigned SchedClass = MI.getDesc().getSchedClass(); 2305 return is_TC1(SchedClass); 2306 } 2307 2308 bool HexagonInstrInfo::isTC2(const MachineInstr &MI) const { 2309 unsigned SchedClass = MI.getDesc().getSchedClass(); 2310 return is_TC2(SchedClass); 2311 } 2312 2313 bool HexagonInstrInfo::isTC2Early(const MachineInstr &MI) const { 2314 unsigned SchedClass = MI.getDesc().getSchedClass(); 2315 return is_TC2early(SchedClass); 2316 } 2317 2318 bool HexagonInstrInfo::isTC4x(const MachineInstr &MI) const { 2319 unsigned SchedClass = MI.getDesc().getSchedClass(); 2320 return is_TC4x(SchedClass); 2321 } 2322 2323 // Schedule this ASAP. 2324 bool HexagonInstrInfo::isToBeScheduledASAP(const MachineInstr &MI1, 2325 const MachineInstr &MI2) const { 2326 if (mayBeCurLoad(MI1)) { 2327 // if (result of SU is used in Next) return true; 2328 unsigned DstReg = MI1.getOperand(0).getReg(); 2329 int N = MI2.getNumOperands(); 2330 for (int I = 0; I < N; I++) 2331 if (MI2.getOperand(I).isReg() && DstReg == MI2.getOperand(I).getReg()) 2332 return true; 2333 } 2334 if (mayBeNewStore(MI2)) 2335 if (MI2.getOpcode() == Hexagon::V6_vS32b_pi) 2336 if (MI1.getOperand(0).isReg() && MI2.getOperand(3).isReg() && 2337 MI1.getOperand(0).getReg() == MI2.getOperand(3).getReg()) 2338 return true; 2339 return false; 2340 } 2341 2342 bool HexagonInstrInfo::isHVXVec(const MachineInstr &MI) const { 2343 const uint64_t V = getType(MI); 2344 return HexagonII::TypeCVI_FIRST <= V && V <= HexagonII::TypeCVI_LAST; 2345 } 2346 2347 // Check if the Offset is a valid auto-inc imm by Load/Store Type. 2348 bool HexagonInstrInfo::isValidAutoIncImm(const EVT VT, int Offset) const { 2349 int Size = VT.getSizeInBits() / 8; 2350 if (Offset % Size != 0) 2351 return false; 2352 int Count = Offset / Size; 2353 2354 switch (VT.getSimpleVT().SimpleTy) { 2355 // For scalars the auto-inc is s4 2356 case MVT::i8: 2357 case MVT::i16: 2358 case MVT::i32: 2359 case MVT::i64: 2360 return isInt<4>(Count); 2361 // For HVX vectors the auto-inc is s3 2362 case MVT::v64i8: 2363 case MVT::v32i16: 2364 case MVT::v16i32: 2365 case MVT::v8i64: 2366 case MVT::v128i8: 2367 case MVT::v64i16: 2368 case MVT::v32i32: 2369 case MVT::v16i64: 2370 return isInt<3>(Count); 2371 default: 2372 break; 2373 } 2374 2375 llvm_unreachable("Not an valid type!"); 2376 } 2377 2378 bool HexagonInstrInfo::isValidOffset(unsigned Opcode, int Offset, 2379 const TargetRegisterInfo *TRI, bool Extend) const { 2380 // This function is to check whether the "Offset" is in the correct range of 2381 // the given "Opcode". If "Offset" is not in the correct range, "A2_addi" is 2382 // inserted to calculate the final address. Due to this reason, the function 2383 // assumes that the "Offset" has correct alignment. 2384 // We used to assert if the offset was not properly aligned, however, 2385 // there are cases where a misaligned pointer recast can cause this 2386 // problem, and we need to allow for it. The front end warns of such 2387 // misaligns with respect to load size. 2388 switch (Opcode) { 2389 case Hexagon::PS_vstorerq_ai: 2390 case Hexagon::PS_vstorerw_ai: 2391 case Hexagon::PS_vstorerw_nt_ai: 2392 case Hexagon::PS_vloadrq_ai: 2393 case Hexagon::PS_vloadrw_ai: 2394 case Hexagon::PS_vloadrw_nt_ai: 2395 case Hexagon::V6_vL32b_ai: 2396 case Hexagon::V6_vS32b_ai: 2397 case Hexagon::V6_vL32b_nt_ai: 2398 case Hexagon::V6_vS32b_nt_ai: 2399 case Hexagon::V6_vL32Ub_ai: 2400 case Hexagon::V6_vS32Ub_ai: { 2401 unsigned VectorSize = TRI->getSpillSize(Hexagon::HvxVRRegClass); 2402 assert(isPowerOf2_32(VectorSize)); 2403 if (Offset & (VectorSize-1)) 2404 return false; 2405 return isInt<4>(Offset >> Log2_32(VectorSize)); 2406 } 2407 2408 case Hexagon::J2_loop0i: 2409 case Hexagon::J2_loop1i: 2410 return isUInt<10>(Offset); 2411 2412 case Hexagon::S4_storeirb_io: 2413 case Hexagon::S4_storeirbt_io: 2414 case Hexagon::S4_storeirbf_io: 2415 return isUInt<6>(Offset); 2416 2417 case Hexagon::S4_storeirh_io: 2418 case Hexagon::S4_storeirht_io: 2419 case Hexagon::S4_storeirhf_io: 2420 return isShiftedUInt<6,1>(Offset); 2421 2422 case Hexagon::S4_storeiri_io: 2423 case Hexagon::S4_storeirit_io: 2424 case Hexagon::S4_storeirif_io: 2425 return isShiftedUInt<6,2>(Offset); 2426 } 2427 2428 if (Extend) 2429 return true; 2430 2431 switch (Opcode) { 2432 case Hexagon::L2_loadri_io: 2433 case Hexagon::S2_storeri_io: 2434 return (Offset >= Hexagon_MEMW_OFFSET_MIN) && 2435 (Offset <= Hexagon_MEMW_OFFSET_MAX); 2436 2437 case Hexagon::L2_loadrd_io: 2438 case Hexagon::S2_storerd_io: 2439 return (Offset >= Hexagon_MEMD_OFFSET_MIN) && 2440 (Offset <= Hexagon_MEMD_OFFSET_MAX); 2441 2442 case Hexagon::L2_loadrh_io: 2443 case Hexagon::L2_loadruh_io: 2444 case Hexagon::S2_storerh_io: 2445 case Hexagon::S2_storerf_io: 2446 return (Offset >= Hexagon_MEMH_OFFSET_MIN) && 2447 (Offset <= Hexagon_MEMH_OFFSET_MAX); 2448 2449 case Hexagon::L2_loadrb_io: 2450 case Hexagon::L2_loadrub_io: 2451 case Hexagon::S2_storerb_io: 2452 return (Offset >= Hexagon_MEMB_OFFSET_MIN) && 2453 (Offset <= Hexagon_MEMB_OFFSET_MAX); 2454 2455 case Hexagon::A2_addi: 2456 return (Offset >= Hexagon_ADDI_OFFSET_MIN) && 2457 (Offset <= Hexagon_ADDI_OFFSET_MAX); 2458 2459 case Hexagon::L4_iadd_memopw_io: 2460 case Hexagon::L4_isub_memopw_io: 2461 case Hexagon::L4_add_memopw_io: 2462 case Hexagon::L4_sub_memopw_io: 2463 case Hexagon::L4_and_memopw_io: 2464 case Hexagon::L4_or_memopw_io: 2465 return (0 <= Offset && Offset <= 255); 2466 2467 case Hexagon::L4_iadd_memoph_io: 2468 case Hexagon::L4_isub_memoph_io: 2469 case Hexagon::L4_add_memoph_io: 2470 case Hexagon::L4_sub_memoph_io: 2471 case Hexagon::L4_and_memoph_io: 2472 case Hexagon::L4_or_memoph_io: 2473 return (0 <= Offset && Offset <= 127); 2474 2475 case Hexagon::L4_iadd_memopb_io: 2476 case Hexagon::L4_isub_memopb_io: 2477 case Hexagon::L4_add_memopb_io: 2478 case Hexagon::L4_sub_memopb_io: 2479 case Hexagon::L4_and_memopb_io: 2480 case Hexagon::L4_or_memopb_io: 2481 return (0 <= Offset && Offset <= 63); 2482 2483 // LDriw_xxx and STriw_xxx are pseudo operations, so it has to take offset of 2484 // any size. Later pass knows how to handle it. 2485 case Hexagon::STriw_pred: 2486 case Hexagon::LDriw_pred: 2487 case Hexagon::STriw_mod: 2488 case Hexagon::LDriw_mod: 2489 return true; 2490 2491 case Hexagon::PS_fi: 2492 case Hexagon::PS_fia: 2493 case Hexagon::INLINEASM: 2494 return true; 2495 2496 case Hexagon::L2_ploadrbt_io: 2497 case Hexagon::L2_ploadrbf_io: 2498 case Hexagon::L2_ploadrubt_io: 2499 case Hexagon::L2_ploadrubf_io: 2500 case Hexagon::S2_pstorerbt_io: 2501 case Hexagon::S2_pstorerbf_io: 2502 return isUInt<6>(Offset); 2503 2504 case Hexagon::L2_ploadrht_io: 2505 case Hexagon::L2_ploadrhf_io: 2506 case Hexagon::L2_ploadruht_io: 2507 case Hexagon::L2_ploadruhf_io: 2508 case Hexagon::S2_pstorerht_io: 2509 case Hexagon::S2_pstorerhf_io: 2510 return isShiftedUInt<6,1>(Offset); 2511 2512 case Hexagon::L2_ploadrit_io: 2513 case Hexagon::L2_ploadrif_io: 2514 case Hexagon::S2_pstorerit_io: 2515 case Hexagon::S2_pstorerif_io: 2516 return isShiftedUInt<6,2>(Offset); 2517 2518 case Hexagon::L2_ploadrdt_io: 2519 case Hexagon::L2_ploadrdf_io: 2520 case Hexagon::S2_pstorerdt_io: 2521 case Hexagon::S2_pstorerdf_io: 2522 return isShiftedUInt<6,3>(Offset); 2523 } // switch 2524 2525 llvm_unreachable("No offset range is defined for this opcode. " 2526 "Please define it in the above switch statement!"); 2527 } 2528 2529 bool HexagonInstrInfo::isVecAcc(const MachineInstr &MI) const { 2530 return isHVXVec(MI) && isAccumulator(MI); 2531 } 2532 2533 bool HexagonInstrInfo::isVecALU(const MachineInstr &MI) const { 2534 const uint64_t F = get(MI.getOpcode()).TSFlags; 2535 const uint64_t V = ((F >> HexagonII::TypePos) & HexagonII::TypeMask); 2536 return 2537 V == HexagonII::TypeCVI_VA || 2538 V == HexagonII::TypeCVI_VA_DV; 2539 } 2540 2541 bool HexagonInstrInfo::isVecUsableNextPacket(const MachineInstr &ProdMI, 2542 const MachineInstr &ConsMI) const { 2543 if (EnableACCForwarding && isVecAcc(ProdMI) && isVecAcc(ConsMI)) 2544 return true; 2545 2546 if (EnableALUForwarding && (isVecALU(ConsMI) || isLateSourceInstr(ConsMI))) 2547 return true; 2548 2549 if (mayBeNewStore(ConsMI)) 2550 return true; 2551 2552 return false; 2553 } 2554 2555 bool HexagonInstrInfo::isZeroExtendingLoad(const MachineInstr &MI) const { 2556 switch (MI.getOpcode()) { 2557 // Byte 2558 case Hexagon::L2_loadrub_io: 2559 case Hexagon::L4_loadrub_ur: 2560 case Hexagon::L4_loadrub_ap: 2561 case Hexagon::L2_loadrub_pr: 2562 case Hexagon::L2_loadrub_pbr: 2563 case Hexagon::L2_loadrub_pi: 2564 case Hexagon::L2_loadrub_pci: 2565 case Hexagon::L2_loadrub_pcr: 2566 case Hexagon::L2_loadbzw2_io: 2567 case Hexagon::L4_loadbzw2_ur: 2568 case Hexagon::L4_loadbzw2_ap: 2569 case Hexagon::L2_loadbzw2_pr: 2570 case Hexagon::L2_loadbzw2_pbr: 2571 case Hexagon::L2_loadbzw2_pi: 2572 case Hexagon::L2_loadbzw2_pci: 2573 case Hexagon::L2_loadbzw2_pcr: 2574 case Hexagon::L2_loadbzw4_io: 2575 case Hexagon::L4_loadbzw4_ur: 2576 case Hexagon::L4_loadbzw4_ap: 2577 case Hexagon::L2_loadbzw4_pr: 2578 case Hexagon::L2_loadbzw4_pbr: 2579 case Hexagon::L2_loadbzw4_pi: 2580 case Hexagon::L2_loadbzw4_pci: 2581 case Hexagon::L2_loadbzw4_pcr: 2582 case Hexagon::L4_loadrub_rr: 2583 case Hexagon::L2_ploadrubt_io: 2584 case Hexagon::L2_ploadrubt_pi: 2585 case Hexagon::L2_ploadrubf_io: 2586 case Hexagon::L2_ploadrubf_pi: 2587 case Hexagon::L2_ploadrubtnew_io: 2588 case Hexagon::L2_ploadrubfnew_io: 2589 case Hexagon::L4_ploadrubt_rr: 2590 case Hexagon::L4_ploadrubf_rr: 2591 case Hexagon::L4_ploadrubtnew_rr: 2592 case Hexagon::L4_ploadrubfnew_rr: 2593 case Hexagon::L2_ploadrubtnew_pi: 2594 case Hexagon::L2_ploadrubfnew_pi: 2595 case Hexagon::L4_ploadrubt_abs: 2596 case Hexagon::L4_ploadrubf_abs: 2597 case Hexagon::L4_ploadrubtnew_abs: 2598 case Hexagon::L4_ploadrubfnew_abs: 2599 case Hexagon::L2_loadrubgp: 2600 // Half 2601 case Hexagon::L2_loadruh_io: 2602 case Hexagon::L4_loadruh_ur: 2603 case Hexagon::L4_loadruh_ap: 2604 case Hexagon::L2_loadruh_pr: 2605 case Hexagon::L2_loadruh_pbr: 2606 case Hexagon::L2_loadruh_pi: 2607 case Hexagon::L2_loadruh_pci: 2608 case Hexagon::L2_loadruh_pcr: 2609 case Hexagon::L4_loadruh_rr: 2610 case Hexagon::L2_ploadruht_io: 2611 case Hexagon::L2_ploadruht_pi: 2612 case Hexagon::L2_ploadruhf_io: 2613 case Hexagon::L2_ploadruhf_pi: 2614 case Hexagon::L2_ploadruhtnew_io: 2615 case Hexagon::L2_ploadruhfnew_io: 2616 case Hexagon::L4_ploadruht_rr: 2617 case Hexagon::L4_ploadruhf_rr: 2618 case Hexagon::L4_ploadruhtnew_rr: 2619 case Hexagon::L4_ploadruhfnew_rr: 2620 case Hexagon::L2_ploadruhtnew_pi: 2621 case Hexagon::L2_ploadruhfnew_pi: 2622 case Hexagon::L4_ploadruht_abs: 2623 case Hexagon::L4_ploadruhf_abs: 2624 case Hexagon::L4_ploadruhtnew_abs: 2625 case Hexagon::L4_ploadruhfnew_abs: 2626 case Hexagon::L2_loadruhgp: 2627 return true; 2628 default: 2629 return false; 2630 } 2631 } 2632 2633 // Add latency to instruction. 2634 bool HexagonInstrInfo::addLatencyToSchedule(const MachineInstr &MI1, 2635 const MachineInstr &MI2) const { 2636 if (isHVXVec(MI1) && isHVXVec(MI2)) 2637 if (!isVecUsableNextPacket(MI1, MI2)) 2638 return true; 2639 return false; 2640 } 2641 2642 /// \brief Get the base register and byte offset of a load/store instr. 2643 bool HexagonInstrInfo::getMemOpBaseRegImmOfs(MachineInstr &LdSt, 2644 unsigned &BaseReg, int64_t &Offset, const TargetRegisterInfo *TRI) 2645 const { 2646 unsigned AccessSize = 0; 2647 int OffsetVal = 0; 2648 BaseReg = getBaseAndOffset(LdSt, OffsetVal, AccessSize); 2649 Offset = OffsetVal; 2650 return BaseReg != 0; 2651 } 2652 2653 /// \brief Can these instructions execute at the same time in a bundle. 2654 bool HexagonInstrInfo::canExecuteInBundle(const MachineInstr &First, 2655 const MachineInstr &Second) const { 2656 if (Second.mayStore() && First.getOpcode() == Hexagon::S2_allocframe) { 2657 const MachineOperand &Op = Second.getOperand(0); 2658 if (Op.isReg() && Op.isUse() && Op.getReg() == Hexagon::R29) 2659 return true; 2660 } 2661 if (DisableNVSchedule) 2662 return false; 2663 if (mayBeNewStore(Second)) { 2664 // Make sure the definition of the first instruction is the value being 2665 // stored. 2666 const MachineOperand &Stored = 2667 Second.getOperand(Second.getNumOperands() - 1); 2668 if (!Stored.isReg()) 2669 return false; 2670 for (unsigned i = 0, e = First.getNumOperands(); i < e; ++i) { 2671 const MachineOperand &Op = First.getOperand(i); 2672 if (Op.isReg() && Op.isDef() && Op.getReg() == Stored.getReg()) 2673 return true; 2674 } 2675 } 2676 return false; 2677 } 2678 2679 bool HexagonInstrInfo::doesNotReturn(const MachineInstr &CallMI) const { 2680 unsigned Opc = CallMI.getOpcode(); 2681 return Opc == Hexagon::PS_call_nr || Opc == Hexagon::PS_callr_nr; 2682 } 2683 2684 bool HexagonInstrInfo::hasEHLabel(const MachineBasicBlock *B) const { 2685 for (auto &I : *B) 2686 if (I.isEHLabel()) 2687 return true; 2688 return false; 2689 } 2690 2691 // Returns true if an instruction can be converted into a non-extended 2692 // equivalent instruction. 2693 bool HexagonInstrInfo::hasNonExtEquivalent(const MachineInstr &MI) const { 2694 short NonExtOpcode; 2695 // Check if the instruction has a register form that uses register in place 2696 // of the extended operand, if so return that as the non-extended form. 2697 if (Hexagon::getRegForm(MI.getOpcode()) >= 0) 2698 return true; 2699 2700 if (MI.getDesc().mayLoad() || MI.getDesc().mayStore()) { 2701 // Check addressing mode and retrieve non-ext equivalent instruction. 2702 2703 switch (getAddrMode(MI)) { 2704 case HexagonII::Absolute: 2705 // Load/store with absolute addressing mode can be converted into 2706 // base+offset mode. 2707 NonExtOpcode = Hexagon::changeAddrMode_abs_io(MI.getOpcode()); 2708 break; 2709 case HexagonII::BaseImmOffset: 2710 // Load/store with base+offset addressing mode can be converted into 2711 // base+register offset addressing mode. However left shift operand should 2712 // be set to 0. 2713 NonExtOpcode = Hexagon::changeAddrMode_io_rr(MI.getOpcode()); 2714 break; 2715 case HexagonII::BaseLongOffset: 2716 NonExtOpcode = Hexagon::changeAddrMode_ur_rr(MI.getOpcode()); 2717 break; 2718 default: 2719 return false; 2720 } 2721 if (NonExtOpcode < 0) 2722 return false; 2723 return true; 2724 } 2725 return false; 2726 } 2727 2728 bool HexagonInstrInfo::hasPseudoInstrPair(const MachineInstr &MI) const { 2729 return Hexagon::getRealHWInstr(MI.getOpcode(), 2730 Hexagon::InstrType_Pseudo) >= 0; 2731 } 2732 2733 bool HexagonInstrInfo::hasUncondBranch(const MachineBasicBlock *B) 2734 const { 2735 MachineBasicBlock::const_iterator I = B->getFirstTerminator(), E = B->end(); 2736 while (I != E) { 2737 if (I->isBarrier()) 2738 return true; 2739 ++I; 2740 } 2741 return false; 2742 } 2743 2744 // Returns true, if a LD insn can be promoted to a cur load. 2745 bool HexagonInstrInfo::mayBeCurLoad(const MachineInstr &MI) const { 2746 const uint64_t F = MI.getDesc().TSFlags; 2747 return ((F >> HexagonII::mayCVLoadPos) & HexagonII::mayCVLoadMask) && 2748 Subtarget.hasV60TOps(); 2749 } 2750 2751 // Returns true, if a ST insn can be promoted to a new-value store. 2752 bool HexagonInstrInfo::mayBeNewStore(const MachineInstr &MI) const { 2753 const uint64_t F = MI.getDesc().TSFlags; 2754 return (F >> HexagonII::mayNVStorePos) & HexagonII::mayNVStoreMask; 2755 } 2756 2757 bool HexagonInstrInfo::producesStall(const MachineInstr &ProdMI, 2758 const MachineInstr &ConsMI) const { 2759 // There is no stall when ProdMI is not a V60 vector. 2760 if (!isHVXVec(ProdMI)) 2761 return false; 2762 2763 // There is no stall when ProdMI and ConsMI are not dependent. 2764 if (!isDependent(ProdMI, ConsMI)) 2765 return false; 2766 2767 // When Forward Scheduling is enabled, there is no stall if ProdMI and ConsMI 2768 // are scheduled in consecutive packets. 2769 if (isVecUsableNextPacket(ProdMI, ConsMI)) 2770 return false; 2771 2772 return true; 2773 } 2774 2775 bool HexagonInstrInfo::producesStall(const MachineInstr &MI, 2776 MachineBasicBlock::const_instr_iterator BII) const { 2777 // There is no stall when I is not a V60 vector. 2778 if (!isHVXVec(MI)) 2779 return false; 2780 2781 MachineBasicBlock::const_instr_iterator MII = BII; 2782 MachineBasicBlock::const_instr_iterator MIE = MII->getParent()->instr_end(); 2783 2784 if (!(*MII).isBundle()) { 2785 const MachineInstr &J = *MII; 2786 return producesStall(J, MI); 2787 } 2788 2789 for (++MII; MII != MIE && MII->isInsideBundle(); ++MII) { 2790 const MachineInstr &J = *MII; 2791 if (producesStall(J, MI)) 2792 return true; 2793 } 2794 return false; 2795 } 2796 2797 bool HexagonInstrInfo::predCanBeUsedAsDotNew(const MachineInstr &MI, 2798 unsigned PredReg) const { 2799 for (const MachineOperand &MO : MI.operands()) { 2800 // Predicate register must be explicitly defined. 2801 if (MO.isRegMask() && MO.clobbersPhysReg(PredReg)) 2802 return false; 2803 if (MO.isReg() && MO.isDef() && MO.isImplicit() && (MO.getReg() == PredReg)) 2804 return false; 2805 } 2806 2807 // Hexagon Programmer's Reference says that decbin, memw_locked, and 2808 // memd_locked cannot be used as .new as well, 2809 // but we don't seem to have these instructions defined. 2810 return MI.getOpcode() != Hexagon::A4_tlbmatch; 2811 } 2812 2813 bool HexagonInstrInfo::PredOpcodeHasJMP_c(unsigned Opcode) const { 2814 return Opcode == Hexagon::J2_jumpt || 2815 Opcode == Hexagon::J2_jumptpt || 2816 Opcode == Hexagon::J2_jumpf || 2817 Opcode == Hexagon::J2_jumpfpt || 2818 Opcode == Hexagon::J2_jumptnew || 2819 Opcode == Hexagon::J2_jumpfnew || 2820 Opcode == Hexagon::J2_jumptnewpt || 2821 Opcode == Hexagon::J2_jumpfnewpt; 2822 } 2823 2824 bool HexagonInstrInfo::predOpcodeHasNot(ArrayRef<MachineOperand> Cond) const { 2825 if (Cond.empty() || !isPredicated(Cond[0].getImm())) 2826 return false; 2827 return !isPredicatedTrue(Cond[0].getImm()); 2828 } 2829 2830 unsigned HexagonInstrInfo::getAddrMode(const MachineInstr &MI) const { 2831 const uint64_t F = MI.getDesc().TSFlags; 2832 return (F >> HexagonII::AddrModePos) & HexagonII::AddrModeMask; 2833 } 2834 2835 // Returns the base register in a memory access (load/store). The offset is 2836 // returned in Offset and the access size is returned in AccessSize. 2837 // If the base register has a subregister or the offset field does not contain 2838 // an immediate value, return 0. 2839 unsigned HexagonInstrInfo::getBaseAndOffset(const MachineInstr &MI, 2840 int &Offset, unsigned &AccessSize) const { 2841 // Return if it is not a base+offset type instruction or a MemOp. 2842 if (getAddrMode(MI) != HexagonII::BaseImmOffset && 2843 getAddrMode(MI) != HexagonII::BaseLongOffset && 2844 !isMemOp(MI) && !isPostIncrement(MI)) 2845 return 0; 2846 2847 AccessSize = getMemAccessSize(MI); 2848 2849 unsigned BasePos = 0, OffsetPos = 0; 2850 if (!getBaseAndOffsetPosition(MI, BasePos, OffsetPos)) 2851 return 0; 2852 2853 // Post increment updates its EA after the mem access, 2854 // so we need to treat its offset as zero. 2855 if (isPostIncrement(MI)) { 2856 Offset = 0; 2857 } else { 2858 const MachineOperand &OffsetOp = MI.getOperand(OffsetPos); 2859 if (!OffsetOp.isImm()) 2860 return 0; 2861 Offset = OffsetOp.getImm(); 2862 } 2863 2864 const MachineOperand &BaseOp = MI.getOperand(BasePos); 2865 if (BaseOp.getSubReg() != 0) 2866 return 0; 2867 return BaseOp.getReg(); 2868 } 2869 2870 /// Return the position of the base and offset operands for this instruction. 2871 bool HexagonInstrInfo::getBaseAndOffsetPosition(const MachineInstr &MI, 2872 unsigned &BasePos, unsigned &OffsetPos) const { 2873 // Deal with memops first. 2874 if (isMemOp(MI)) { 2875 BasePos = 0; 2876 OffsetPos = 1; 2877 } else if (MI.mayStore()) { 2878 BasePos = 0; 2879 OffsetPos = 1; 2880 } else if (MI.mayLoad()) { 2881 BasePos = 1; 2882 OffsetPos = 2; 2883 } else 2884 return false; 2885 2886 if (isPredicated(MI)) { 2887 BasePos++; 2888 OffsetPos++; 2889 } 2890 if (isPostIncrement(MI)) { 2891 BasePos++; 2892 OffsetPos++; 2893 } 2894 2895 if (!MI.getOperand(BasePos).isReg() || !MI.getOperand(OffsetPos).isImm()) 2896 return false; 2897 2898 return true; 2899 } 2900 2901 // Inserts branching instructions in reverse order of their occurrence. 2902 // e.g. jump_t t1 (i1) 2903 // jump t2 (i2) 2904 // Jumpers = {i2, i1} 2905 SmallVector<MachineInstr*, 2> HexagonInstrInfo::getBranchingInstrs( 2906 MachineBasicBlock& MBB) const { 2907 SmallVector<MachineInstr*, 2> Jumpers; 2908 // If the block has no terminators, it just falls into the block after it. 2909 MachineBasicBlock::instr_iterator I = MBB.instr_end(); 2910 if (I == MBB.instr_begin()) 2911 return Jumpers; 2912 2913 // A basic block may looks like this: 2914 // 2915 // [ insn 2916 // EH_LABEL 2917 // insn 2918 // insn 2919 // insn 2920 // EH_LABEL 2921 // insn ] 2922 // 2923 // It has two succs but does not have a terminator 2924 // Don't know how to handle it. 2925 do { 2926 --I; 2927 if (I->isEHLabel()) 2928 return Jumpers; 2929 } while (I != MBB.instr_begin()); 2930 2931 I = MBB.instr_end(); 2932 --I; 2933 2934 while (I->isDebugValue()) { 2935 if (I == MBB.instr_begin()) 2936 return Jumpers; 2937 --I; 2938 } 2939 if (!isUnpredicatedTerminator(*I)) 2940 return Jumpers; 2941 2942 // Get the last instruction in the block. 2943 MachineInstr *LastInst = &*I; 2944 Jumpers.push_back(LastInst); 2945 MachineInstr *SecondLastInst = nullptr; 2946 // Find one more terminator if present. 2947 do { 2948 if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(*I)) { 2949 if (!SecondLastInst) { 2950 SecondLastInst = &*I; 2951 Jumpers.push_back(SecondLastInst); 2952 } else // This is a third branch. 2953 return Jumpers; 2954 } 2955 if (I == MBB.instr_begin()) 2956 break; 2957 --I; 2958 } while (true); 2959 return Jumpers; 2960 } 2961 2962 // Returns Operand Index for the constant extended instruction. 2963 unsigned HexagonInstrInfo::getCExtOpNum(const MachineInstr &MI) const { 2964 const uint64_t F = MI.getDesc().TSFlags; 2965 return (F >> HexagonII::ExtendableOpPos) & HexagonII::ExtendableOpMask; 2966 } 2967 2968 // See if instruction could potentially be a duplex candidate. 2969 // If so, return its group. Zero otherwise. 2970 HexagonII::CompoundGroup HexagonInstrInfo::getCompoundCandidateGroup( 2971 const MachineInstr &MI) const { 2972 unsigned DstReg, SrcReg, Src1Reg, Src2Reg; 2973 2974 switch (MI.getOpcode()) { 2975 default: 2976 return HexagonII::HCG_None; 2977 // 2978 // Compound pairs. 2979 // "p0=cmp.eq(Rs16,Rt16); if (p0.new) jump:nt #r9:2" 2980 // "Rd16=#U6 ; jump #r9:2" 2981 // "Rd16=Rs16 ; jump #r9:2" 2982 // 2983 case Hexagon::C2_cmpeq: 2984 case Hexagon::C2_cmpgt: 2985 case Hexagon::C2_cmpgtu: 2986 DstReg = MI.getOperand(0).getReg(); 2987 Src1Reg = MI.getOperand(1).getReg(); 2988 Src2Reg = MI.getOperand(2).getReg(); 2989 if (Hexagon::PredRegsRegClass.contains(DstReg) && 2990 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) && 2991 isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg)) 2992 return HexagonII::HCG_A; 2993 break; 2994 case Hexagon::C2_cmpeqi: 2995 case Hexagon::C2_cmpgti: 2996 case Hexagon::C2_cmpgtui: 2997 // P0 = cmp.eq(Rs,#u2) 2998 DstReg = MI.getOperand(0).getReg(); 2999 SrcReg = MI.getOperand(1).getReg(); 3000 if (Hexagon::PredRegsRegClass.contains(DstReg) && 3001 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) && 3002 isIntRegForSubInst(SrcReg) && MI.getOperand(2).isImm() && 3003 ((isUInt<5>(MI.getOperand(2).getImm())) || 3004 (MI.getOperand(2).getImm() == -1))) 3005 return HexagonII::HCG_A; 3006 break; 3007 case Hexagon::A2_tfr: 3008 // Rd = Rs 3009 DstReg = MI.getOperand(0).getReg(); 3010 SrcReg = MI.getOperand(1).getReg(); 3011 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg)) 3012 return HexagonII::HCG_A; 3013 break; 3014 case Hexagon::A2_tfrsi: 3015 // Rd = #u6 3016 // Do not test for #u6 size since the const is getting extended 3017 // regardless and compound could be formed. 3018 DstReg = MI.getOperand(0).getReg(); 3019 if (isIntRegForSubInst(DstReg)) 3020 return HexagonII::HCG_A; 3021 break; 3022 case Hexagon::S2_tstbit_i: 3023 DstReg = MI.getOperand(0).getReg(); 3024 Src1Reg = MI.getOperand(1).getReg(); 3025 if (Hexagon::PredRegsRegClass.contains(DstReg) && 3026 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) && 3027 MI.getOperand(2).isImm() && 3028 isIntRegForSubInst(Src1Reg) && (MI.getOperand(2).getImm() == 0)) 3029 return HexagonII::HCG_A; 3030 break; 3031 // The fact that .new form is used pretty much guarantees 3032 // that predicate register will match. Nevertheless, 3033 // there could be some false positives without additional 3034 // checking. 3035 case Hexagon::J2_jumptnew: 3036 case Hexagon::J2_jumpfnew: 3037 case Hexagon::J2_jumptnewpt: 3038 case Hexagon::J2_jumpfnewpt: 3039 Src1Reg = MI.getOperand(0).getReg(); 3040 if (Hexagon::PredRegsRegClass.contains(Src1Reg) && 3041 (Hexagon::P0 == Src1Reg || Hexagon::P1 == Src1Reg)) 3042 return HexagonII::HCG_B; 3043 break; 3044 // Transfer and jump: 3045 // Rd=#U6 ; jump #r9:2 3046 // Rd=Rs ; jump #r9:2 3047 // Do not test for jump range here. 3048 case Hexagon::J2_jump: 3049 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4: 3050 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4_PIC: 3051 return HexagonII::HCG_C; 3052 } 3053 3054 return HexagonII::HCG_None; 3055 } 3056 3057 // Returns -1 when there is no opcode found. 3058 unsigned HexagonInstrInfo::getCompoundOpcode(const MachineInstr &GA, 3059 const MachineInstr &GB) const { 3060 assert(getCompoundCandidateGroup(GA) == HexagonII::HCG_A); 3061 assert(getCompoundCandidateGroup(GB) == HexagonII::HCG_B); 3062 if ((GA.getOpcode() != Hexagon::C2_cmpeqi) || 3063 (GB.getOpcode() != Hexagon::J2_jumptnew)) 3064 return -1; 3065 unsigned DestReg = GA.getOperand(0).getReg(); 3066 if (!GB.readsRegister(DestReg)) 3067 return -1; 3068 if (DestReg == Hexagon::P0) 3069 return Hexagon::J4_cmpeqi_tp0_jump_nt; 3070 if (DestReg == Hexagon::P1) 3071 return Hexagon::J4_cmpeqi_tp1_jump_nt; 3072 return -1; 3073 } 3074 3075 int HexagonInstrInfo::getCondOpcode(int Opc, bool invertPredicate) const { 3076 enum Hexagon::PredSense inPredSense; 3077 inPredSense = invertPredicate ? Hexagon::PredSense_false : 3078 Hexagon::PredSense_true; 3079 int CondOpcode = Hexagon::getPredOpcode(Opc, inPredSense); 3080 if (CondOpcode >= 0) // Valid Conditional opcode/instruction 3081 return CondOpcode; 3082 3083 llvm_unreachable("Unexpected predicable instruction"); 3084 } 3085 3086 // Return the cur value instruction for a given store. 3087 int HexagonInstrInfo::getDotCurOp(const MachineInstr &MI) const { 3088 switch (MI.getOpcode()) { 3089 default: llvm_unreachable("Unknown .cur type"); 3090 case Hexagon::V6_vL32b_pi: 3091 return Hexagon::V6_vL32b_cur_pi; 3092 case Hexagon::V6_vL32b_ai: 3093 return Hexagon::V6_vL32b_cur_ai; 3094 case Hexagon::V6_vL32b_nt_pi: 3095 return Hexagon::V6_vL32b_nt_cur_pi; 3096 case Hexagon::V6_vL32b_nt_ai: 3097 return Hexagon::V6_vL32b_nt_cur_ai; 3098 } 3099 return 0; 3100 } 3101 3102 // Return the regular version of the .cur instruction. 3103 int HexagonInstrInfo::getNonDotCurOp(const MachineInstr &MI) const { 3104 switch (MI.getOpcode()) { 3105 default: llvm_unreachable("Unknown .cur type"); 3106 case Hexagon::V6_vL32b_cur_pi: 3107 return Hexagon::V6_vL32b_pi; 3108 case Hexagon::V6_vL32b_cur_ai: 3109 return Hexagon::V6_vL32b_ai; 3110 case Hexagon::V6_vL32b_nt_cur_pi: 3111 return Hexagon::V6_vL32b_nt_pi; 3112 case Hexagon::V6_vL32b_nt_cur_ai: 3113 return Hexagon::V6_vL32b_nt_ai; 3114 } 3115 return 0; 3116 } 3117 3118 // The diagram below shows the steps involved in the conversion of a predicated 3119 // store instruction to its .new predicated new-value form. 3120 // 3121 // Note: It doesn't include conditional new-value stores as they can't be 3122 // converted to .new predicate. 3123 // 3124 // p.new NV store [ if(p0.new)memw(R0+#0)=R2.new ] 3125 // ^ ^ 3126 // / \ (not OK. it will cause new-value store to be 3127 // / X conditional on p0.new while R2 producer is 3128 // / \ on p0) 3129 // / \. 3130 // p.new store p.old NV store 3131 // [if(p0.new)memw(R0+#0)=R2] [if(p0)memw(R0+#0)=R2.new] 3132 // ^ ^ 3133 // \ / 3134 // \ / 3135 // \ / 3136 // p.old store 3137 // [if (p0)memw(R0+#0)=R2] 3138 // 3139 // The following set of instructions further explains the scenario where 3140 // conditional new-value store becomes invalid when promoted to .new predicate 3141 // form. 3142 // 3143 // { 1) if (p0) r0 = add(r1, r2) 3144 // 2) p0 = cmp.eq(r3, #0) } 3145 // 3146 // 3) if (p0) memb(r1+#0) = r0 --> this instruction can't be grouped with 3147 // the first two instructions because in instr 1, r0 is conditional on old value 3148 // of p0 but its use in instr 3 is conditional on p0 modified by instr 2 which 3149 // is not valid for new-value stores. 3150 // Predicated new value stores (i.e. if (p0) memw(..)=r0.new) are excluded 3151 // from the "Conditional Store" list. Because a predicated new value store 3152 // would NOT be promoted to a double dot new store. See diagram below: 3153 // This function returns yes for those stores that are predicated but not 3154 // yet promoted to predicate dot new instructions. 3155 // 3156 // +---------------------+ 3157 // /-----| if (p0) memw(..)=r0 |---------\~ 3158 // || +---------------------+ || 3159 // promote || /\ /\ || promote 3160 // || /||\ /||\ || 3161 // \||/ demote || \||/ 3162 // \/ || || \/ 3163 // +-------------------------+ || +-------------------------+ 3164 // | if (p0.new) memw(..)=r0 | || | if (p0) memw(..)=r0.new | 3165 // +-------------------------+ || +-------------------------+ 3166 // || || || 3167 // || demote \||/ 3168 // promote || \/ NOT possible 3169 // || || /\~ 3170 // \||/ || /||\~ 3171 // \/ || || 3172 // +-----------------------------+ 3173 // | if (p0.new) memw(..)=r0.new | 3174 // +-----------------------------+ 3175 // Double Dot New Store 3176 // 3177 // Returns the most basic instruction for the .new predicated instructions and 3178 // new-value stores. 3179 // For example, all of the following instructions will be converted back to the 3180 // same instruction: 3181 // 1) if (p0.new) memw(R0+#0) = R1.new ---> 3182 // 2) if (p0) memw(R0+#0)= R1.new -------> if (p0) memw(R0+#0) = R1 3183 // 3) if (p0.new) memw(R0+#0) = R1 ---> 3184 // 3185 // To understand the translation of instruction 1 to its original form, consider 3186 // a packet with 3 instructions. 3187 // { p0 = cmp.eq(R0,R1) 3188 // if (p0.new) R2 = add(R3, R4) 3189 // R5 = add (R3, R1) 3190 // } 3191 // if (p0) memw(R5+#0) = R2 <--- trying to include it in the previous packet 3192 // 3193 // This instruction can be part of the previous packet only if both p0 and R2 3194 // are promoted to .new values. This promotion happens in steps, first 3195 // predicate register is promoted to .new and in the next iteration R2 is 3196 // promoted. Therefore, in case of dependence check failure (due to R5) during 3197 // next iteration, it should be converted back to its most basic form. 3198 3199 // Return the new value instruction for a given store. 3200 int HexagonInstrInfo::getDotNewOp(const MachineInstr &MI) const { 3201 int NVOpcode = Hexagon::getNewValueOpcode(MI.getOpcode()); 3202 if (NVOpcode >= 0) // Valid new-value store instruction. 3203 return NVOpcode; 3204 3205 switch (MI.getOpcode()) { 3206 default: 3207 report_fatal_error(std::string("Unknown .new type: ") + 3208 std::to_string(MI.getOpcode())); 3209 case Hexagon::S4_storerb_ur: 3210 return Hexagon::S4_storerbnew_ur; 3211 3212 case Hexagon::S2_storerb_pci: 3213 return Hexagon::S2_storerb_pci; 3214 3215 case Hexagon::S2_storeri_pci: 3216 return Hexagon::S2_storeri_pci; 3217 3218 case Hexagon::S2_storerh_pci: 3219 return Hexagon::S2_storerh_pci; 3220 3221 case Hexagon::S2_storerd_pci: 3222 return Hexagon::S2_storerd_pci; 3223 3224 case Hexagon::S2_storerf_pci: 3225 return Hexagon::S2_storerf_pci; 3226 3227 case Hexagon::V6_vS32b_ai: 3228 return Hexagon::V6_vS32b_new_ai; 3229 3230 case Hexagon::V6_vS32b_pi: 3231 return Hexagon::V6_vS32b_new_pi; 3232 } 3233 return 0; 3234 } 3235 3236 // Returns the opcode to use when converting MI, which is a conditional jump, 3237 // into a conditional instruction which uses the .new value of the predicate. 3238 // We also use branch probabilities to add a hint to the jump. 3239 // If MBPI is null, all edges will be treated as equally likely for the 3240 // purposes of establishing a predication hint. 3241 int HexagonInstrInfo::getDotNewPredJumpOp(const MachineInstr &MI, 3242 const MachineBranchProbabilityInfo *MBPI) const { 3243 // We assume that block can have at most two successors. 3244 const MachineBasicBlock *Src = MI.getParent(); 3245 const MachineOperand &BrTarget = MI.getOperand(1); 3246 bool Taken = false; 3247 const BranchProbability OneHalf(1, 2); 3248 3249 auto getEdgeProbability = [MBPI] (const MachineBasicBlock *Src, 3250 const MachineBasicBlock *Dst) { 3251 if (MBPI) 3252 return MBPI->getEdgeProbability(Src, Dst); 3253 return BranchProbability(1, Src->succ_size()); 3254 }; 3255 3256 if (BrTarget.isMBB()) { 3257 const MachineBasicBlock *Dst = BrTarget.getMBB(); 3258 Taken = getEdgeProbability(Src, Dst) >= OneHalf; 3259 } else { 3260 // The branch target is not a basic block (most likely a function). 3261 // Since BPI only gives probabilities for targets that are basic blocks, 3262 // try to identify another target of this branch (potentially a fall- 3263 // -through) and check the probability of that target. 3264 // 3265 // The only handled branch combinations are: 3266 // - one conditional branch, 3267 // - one conditional branch followed by one unconditional branch. 3268 // Otherwise, assume not-taken. 3269 assert(MI.isConditionalBranch()); 3270 const MachineBasicBlock &B = *MI.getParent(); 3271 bool SawCond = false, Bad = false; 3272 for (const MachineInstr &I : B) { 3273 if (!I.isBranch()) 3274 continue; 3275 if (I.isConditionalBranch()) { 3276 SawCond = true; 3277 if (&I != &MI) { 3278 Bad = true; 3279 break; 3280 } 3281 } 3282 if (I.isUnconditionalBranch() && !SawCond) { 3283 Bad = true; 3284 break; 3285 } 3286 } 3287 if (!Bad) { 3288 MachineBasicBlock::const_instr_iterator It(MI); 3289 MachineBasicBlock::const_instr_iterator NextIt = std::next(It); 3290 if (NextIt == B.instr_end()) { 3291 // If this branch is the last, look for the fall-through block. 3292 for (const MachineBasicBlock *SB : B.successors()) { 3293 if (!B.isLayoutSuccessor(SB)) 3294 continue; 3295 Taken = getEdgeProbability(Src, SB) < OneHalf; 3296 break; 3297 } 3298 } else { 3299 assert(NextIt->isUnconditionalBranch()); 3300 // Find the first MBB operand and assume it's the target. 3301 const MachineBasicBlock *BT = nullptr; 3302 for (const MachineOperand &Op : NextIt->operands()) { 3303 if (!Op.isMBB()) 3304 continue; 3305 BT = Op.getMBB(); 3306 break; 3307 } 3308 Taken = BT && getEdgeProbability(Src, BT) < OneHalf; 3309 } 3310 } // if (!Bad) 3311 } 3312 3313 // The Taken flag should be set to something reasonable by this point. 3314 3315 switch (MI.getOpcode()) { 3316 case Hexagon::J2_jumpt: 3317 return Taken ? Hexagon::J2_jumptnewpt : Hexagon::J2_jumptnew; 3318 case Hexagon::J2_jumpf: 3319 return Taken ? Hexagon::J2_jumpfnewpt : Hexagon::J2_jumpfnew; 3320 3321 default: 3322 llvm_unreachable("Unexpected jump instruction."); 3323 } 3324 } 3325 3326 // Return .new predicate version for an instruction. 3327 int HexagonInstrInfo::getDotNewPredOp(const MachineInstr &MI, 3328 const MachineBranchProbabilityInfo *MBPI) const { 3329 switch (MI.getOpcode()) { 3330 // Condtional Jumps 3331 case Hexagon::J2_jumpt: 3332 case Hexagon::J2_jumpf: 3333 return getDotNewPredJumpOp(MI, MBPI); 3334 } 3335 3336 int NewOpcode = Hexagon::getPredNewOpcode(MI.getOpcode()); 3337 if (NewOpcode >= 0) 3338 return NewOpcode; 3339 return 0; 3340 } 3341 3342 int HexagonInstrInfo::getDotOldOp(const MachineInstr &MI) const { 3343 int NewOp = MI.getOpcode(); 3344 if (isPredicated(NewOp) && isPredicatedNew(NewOp)) { // Get predicate old form 3345 NewOp = Hexagon::getPredOldOpcode(NewOp); 3346 // All Hexagon architectures have prediction bits on dot-new branches, 3347 // but only Hexagon V60+ has prediction bits on dot-old ones. Make sure 3348 // to pick the right opcode when converting back to dot-old. 3349 if (!Subtarget.getFeatureBits()[Hexagon::ArchV60]) { 3350 switch (NewOp) { 3351 case Hexagon::J2_jumptpt: 3352 NewOp = Hexagon::J2_jumpt; 3353 break; 3354 case Hexagon::J2_jumpfpt: 3355 NewOp = Hexagon::J2_jumpf; 3356 break; 3357 case Hexagon::J2_jumprtpt: 3358 NewOp = Hexagon::J2_jumprt; 3359 break; 3360 case Hexagon::J2_jumprfpt: 3361 NewOp = Hexagon::J2_jumprf; 3362 break; 3363 } 3364 } 3365 assert(NewOp >= 0 && 3366 "Couldn't change predicate new instruction to its old form."); 3367 } 3368 3369 if (isNewValueStore(NewOp)) { // Convert into non-new-value format 3370 NewOp = Hexagon::getNonNVStore(NewOp); 3371 assert(NewOp >= 0 && "Couldn't change new-value store to its old form."); 3372 } 3373 3374 if (Subtarget.hasV60TOps()) 3375 return NewOp; 3376 3377 // Subtargets prior to V60 didn't support 'taken' forms of predicated jumps. 3378 switch (NewOp) { 3379 case Hexagon::J2_jumpfpt: 3380 return Hexagon::J2_jumpf; 3381 case Hexagon::J2_jumptpt: 3382 return Hexagon::J2_jumpt; 3383 case Hexagon::J2_jumprfpt: 3384 return Hexagon::J2_jumprf; 3385 case Hexagon::J2_jumprtpt: 3386 return Hexagon::J2_jumprt; 3387 } 3388 return NewOp; 3389 } 3390 3391 // See if instruction could potentially be a duplex candidate. 3392 // If so, return its group. Zero otherwise. 3393 HexagonII::SubInstructionGroup HexagonInstrInfo::getDuplexCandidateGroup( 3394 const MachineInstr &MI) const { 3395 unsigned DstReg, SrcReg, Src1Reg, Src2Reg; 3396 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 3397 3398 switch (MI.getOpcode()) { 3399 default: 3400 return HexagonII::HSIG_None; 3401 // 3402 // Group L1: 3403 // 3404 // Rd = memw(Rs+#u4:2) 3405 // Rd = memub(Rs+#u4:0) 3406 case Hexagon::L2_loadri_io: 3407 DstReg = MI.getOperand(0).getReg(); 3408 SrcReg = MI.getOperand(1).getReg(); 3409 // Special case this one from Group L2. 3410 // Rd = memw(r29+#u5:2) 3411 if (isIntRegForSubInst(DstReg)) { 3412 if (Hexagon::IntRegsRegClass.contains(SrcReg) && 3413 HRI.getStackRegister() == SrcReg && 3414 MI.getOperand(2).isImm() && 3415 isShiftedUInt<5,2>(MI.getOperand(2).getImm())) 3416 return HexagonII::HSIG_L2; 3417 // Rd = memw(Rs+#u4:2) 3418 if (isIntRegForSubInst(SrcReg) && 3419 (MI.getOperand(2).isImm() && 3420 isShiftedUInt<4,2>(MI.getOperand(2).getImm()))) 3421 return HexagonII::HSIG_L1; 3422 } 3423 break; 3424 case Hexagon::L2_loadrub_io: 3425 // Rd = memub(Rs+#u4:0) 3426 DstReg = MI.getOperand(0).getReg(); 3427 SrcReg = MI.getOperand(1).getReg(); 3428 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 3429 MI.getOperand(2).isImm() && isUInt<4>(MI.getOperand(2).getImm())) 3430 return HexagonII::HSIG_L1; 3431 break; 3432 // 3433 // Group L2: 3434 // 3435 // Rd = memh/memuh(Rs+#u3:1) 3436 // Rd = memb(Rs+#u3:0) 3437 // Rd = memw(r29+#u5:2) - Handled above. 3438 // Rdd = memd(r29+#u5:3) 3439 // deallocframe 3440 // [if ([!]p0[.new])] dealloc_return 3441 // [if ([!]p0[.new])] jumpr r31 3442 case Hexagon::L2_loadrh_io: 3443 case Hexagon::L2_loadruh_io: 3444 // Rd = memh/memuh(Rs+#u3:1) 3445 DstReg = MI.getOperand(0).getReg(); 3446 SrcReg = MI.getOperand(1).getReg(); 3447 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 3448 MI.getOperand(2).isImm() && 3449 isShiftedUInt<3,1>(MI.getOperand(2).getImm())) 3450 return HexagonII::HSIG_L2; 3451 break; 3452 case Hexagon::L2_loadrb_io: 3453 // Rd = memb(Rs+#u3:0) 3454 DstReg = MI.getOperand(0).getReg(); 3455 SrcReg = MI.getOperand(1).getReg(); 3456 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 3457 MI.getOperand(2).isImm() && 3458 isUInt<3>(MI.getOperand(2).getImm())) 3459 return HexagonII::HSIG_L2; 3460 break; 3461 case Hexagon::L2_loadrd_io: 3462 // Rdd = memd(r29+#u5:3) 3463 DstReg = MI.getOperand(0).getReg(); 3464 SrcReg = MI.getOperand(1).getReg(); 3465 if (isDblRegForSubInst(DstReg, HRI) && 3466 Hexagon::IntRegsRegClass.contains(SrcReg) && 3467 HRI.getStackRegister() == SrcReg && 3468 MI.getOperand(2).isImm() && 3469 isShiftedUInt<5,3>(MI.getOperand(2).getImm())) 3470 return HexagonII::HSIG_L2; 3471 break; 3472 // dealloc_return is not documented in Hexagon Manual, but marked 3473 // with A_SUBINSN attribute in iset_v4classic.py. 3474 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4: 3475 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4_PIC: 3476 case Hexagon::L4_return: 3477 case Hexagon::L2_deallocframe: 3478 return HexagonII::HSIG_L2; 3479 case Hexagon::EH_RETURN_JMPR: 3480 case Hexagon::PS_jmpret: 3481 // jumpr r31 3482 // Actual form JMPR %PC<imp-def>, %R31<imp-use>, %R0<imp-use,internal>. 3483 DstReg = MI.getOperand(0).getReg(); 3484 if (Hexagon::IntRegsRegClass.contains(DstReg) && (Hexagon::R31 == DstReg)) 3485 return HexagonII::HSIG_L2; 3486 break; 3487 case Hexagon::PS_jmprett: 3488 case Hexagon::PS_jmpretf: 3489 case Hexagon::PS_jmprettnewpt: 3490 case Hexagon::PS_jmpretfnewpt: 3491 case Hexagon::PS_jmprettnew: 3492 case Hexagon::PS_jmpretfnew: 3493 DstReg = MI.getOperand(1).getReg(); 3494 SrcReg = MI.getOperand(0).getReg(); 3495 // [if ([!]p0[.new])] jumpr r31 3496 if ((Hexagon::PredRegsRegClass.contains(SrcReg) && 3497 (Hexagon::P0 == SrcReg)) && 3498 (Hexagon::IntRegsRegClass.contains(DstReg) && (Hexagon::R31 == DstReg))) 3499 return HexagonII::HSIG_L2; 3500 break; 3501 case Hexagon::L4_return_t: 3502 case Hexagon::L4_return_f: 3503 case Hexagon::L4_return_tnew_pnt: 3504 case Hexagon::L4_return_fnew_pnt: 3505 case Hexagon::L4_return_tnew_pt: 3506 case Hexagon::L4_return_fnew_pt: 3507 // [if ([!]p0[.new])] dealloc_return 3508 SrcReg = MI.getOperand(0).getReg(); 3509 if (Hexagon::PredRegsRegClass.contains(SrcReg) && (Hexagon::P0 == SrcReg)) 3510 return HexagonII::HSIG_L2; 3511 break; 3512 // 3513 // Group S1: 3514 // 3515 // memw(Rs+#u4:2) = Rt 3516 // memb(Rs+#u4:0) = Rt 3517 case Hexagon::S2_storeri_io: 3518 // Special case this one from Group S2. 3519 // memw(r29+#u5:2) = Rt 3520 Src1Reg = MI.getOperand(0).getReg(); 3521 Src2Reg = MI.getOperand(2).getReg(); 3522 if (Hexagon::IntRegsRegClass.contains(Src1Reg) && 3523 isIntRegForSubInst(Src2Reg) && 3524 HRI.getStackRegister() == Src1Reg && MI.getOperand(1).isImm() && 3525 isShiftedUInt<5,2>(MI.getOperand(1).getImm())) 3526 return HexagonII::HSIG_S2; 3527 // memw(Rs+#u4:2) = Rt 3528 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) && 3529 MI.getOperand(1).isImm() && 3530 isShiftedUInt<4,2>(MI.getOperand(1).getImm())) 3531 return HexagonII::HSIG_S1; 3532 break; 3533 case Hexagon::S2_storerb_io: 3534 // memb(Rs+#u4:0) = Rt 3535 Src1Reg = MI.getOperand(0).getReg(); 3536 Src2Reg = MI.getOperand(2).getReg(); 3537 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) && 3538 MI.getOperand(1).isImm() && isUInt<4>(MI.getOperand(1).getImm())) 3539 return HexagonII::HSIG_S1; 3540 break; 3541 // 3542 // Group S2: 3543 // 3544 // memh(Rs+#u3:1) = Rt 3545 // memw(r29+#u5:2) = Rt 3546 // memd(r29+#s6:3) = Rtt 3547 // memw(Rs+#u4:2) = #U1 3548 // memb(Rs+#u4) = #U1 3549 // allocframe(#u5:3) 3550 case Hexagon::S2_storerh_io: 3551 // memh(Rs+#u3:1) = Rt 3552 Src1Reg = MI.getOperand(0).getReg(); 3553 Src2Reg = MI.getOperand(2).getReg(); 3554 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) && 3555 MI.getOperand(1).isImm() && 3556 isShiftedUInt<3,1>(MI.getOperand(1).getImm())) 3557 return HexagonII::HSIG_S1; 3558 break; 3559 case Hexagon::S2_storerd_io: 3560 // memd(r29+#s6:3) = Rtt 3561 Src1Reg = MI.getOperand(0).getReg(); 3562 Src2Reg = MI.getOperand(2).getReg(); 3563 if (isDblRegForSubInst(Src2Reg, HRI) && 3564 Hexagon::IntRegsRegClass.contains(Src1Reg) && 3565 HRI.getStackRegister() == Src1Reg && MI.getOperand(1).isImm() && 3566 isShiftedInt<6,3>(MI.getOperand(1).getImm())) 3567 return HexagonII::HSIG_S2; 3568 break; 3569 case Hexagon::S4_storeiri_io: 3570 // memw(Rs+#u4:2) = #U1 3571 Src1Reg = MI.getOperand(0).getReg(); 3572 if (isIntRegForSubInst(Src1Reg) && MI.getOperand(1).isImm() && 3573 isShiftedUInt<4,2>(MI.getOperand(1).getImm()) && 3574 MI.getOperand(2).isImm() && isUInt<1>(MI.getOperand(2).getImm())) 3575 return HexagonII::HSIG_S2; 3576 break; 3577 case Hexagon::S4_storeirb_io: 3578 // memb(Rs+#u4) = #U1 3579 Src1Reg = MI.getOperand(0).getReg(); 3580 if (isIntRegForSubInst(Src1Reg) && 3581 MI.getOperand(1).isImm() && isUInt<4>(MI.getOperand(1).getImm()) && 3582 MI.getOperand(2).isImm() && isUInt<1>(MI.getOperand(2).getImm())) 3583 return HexagonII::HSIG_S2; 3584 break; 3585 case Hexagon::S2_allocframe: 3586 if (MI.getOperand(0).isImm() && 3587 isShiftedUInt<5,3>(MI.getOperand(0).getImm())) 3588 return HexagonII::HSIG_S1; 3589 break; 3590 // 3591 // Group A: 3592 // 3593 // Rx = add(Rx,#s7) 3594 // Rd = Rs 3595 // Rd = #u6 3596 // Rd = #-1 3597 // if ([!]P0[.new]) Rd = #0 3598 // Rd = add(r29,#u6:2) 3599 // Rx = add(Rx,Rs) 3600 // P0 = cmp.eq(Rs,#u2) 3601 // Rdd = combine(#0,Rs) 3602 // Rdd = combine(Rs,#0) 3603 // Rdd = combine(#u2,#U2) 3604 // Rd = add(Rs,#1) 3605 // Rd = add(Rs,#-1) 3606 // Rd = sxth/sxtb/zxtb/zxth(Rs) 3607 // Rd = and(Rs,#1) 3608 case Hexagon::A2_addi: 3609 DstReg = MI.getOperand(0).getReg(); 3610 SrcReg = MI.getOperand(1).getReg(); 3611 if (isIntRegForSubInst(DstReg)) { 3612 // Rd = add(r29,#u6:2) 3613 if (Hexagon::IntRegsRegClass.contains(SrcReg) && 3614 HRI.getStackRegister() == SrcReg && MI.getOperand(2).isImm() && 3615 isShiftedUInt<6,2>(MI.getOperand(2).getImm())) 3616 return HexagonII::HSIG_A; 3617 // Rx = add(Rx,#s7) 3618 if ((DstReg == SrcReg) && MI.getOperand(2).isImm() && 3619 isInt<7>(MI.getOperand(2).getImm())) 3620 return HexagonII::HSIG_A; 3621 // Rd = add(Rs,#1) 3622 // Rd = add(Rs,#-1) 3623 if (isIntRegForSubInst(SrcReg) && MI.getOperand(2).isImm() && 3624 ((MI.getOperand(2).getImm() == 1) || 3625 (MI.getOperand(2).getImm() == -1))) 3626 return HexagonII::HSIG_A; 3627 } 3628 break; 3629 case Hexagon::A2_add: 3630 // Rx = add(Rx,Rs) 3631 DstReg = MI.getOperand(0).getReg(); 3632 Src1Reg = MI.getOperand(1).getReg(); 3633 Src2Reg = MI.getOperand(2).getReg(); 3634 if (isIntRegForSubInst(DstReg) && (DstReg == Src1Reg) && 3635 isIntRegForSubInst(Src2Reg)) 3636 return HexagonII::HSIG_A; 3637 break; 3638 case Hexagon::A2_andir: 3639 // Same as zxtb. 3640 // Rd16=and(Rs16,#255) 3641 // Rd16=and(Rs16,#1) 3642 DstReg = MI.getOperand(0).getReg(); 3643 SrcReg = MI.getOperand(1).getReg(); 3644 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 3645 MI.getOperand(2).isImm() && 3646 ((MI.getOperand(2).getImm() == 1) || 3647 (MI.getOperand(2).getImm() == 255))) 3648 return HexagonII::HSIG_A; 3649 break; 3650 case Hexagon::A2_tfr: 3651 // Rd = Rs 3652 DstReg = MI.getOperand(0).getReg(); 3653 SrcReg = MI.getOperand(1).getReg(); 3654 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg)) 3655 return HexagonII::HSIG_A; 3656 break; 3657 case Hexagon::A2_tfrsi: 3658 // Rd = #u6 3659 // Do not test for #u6 size since the const is getting extended 3660 // regardless and compound could be formed. 3661 // Rd = #-1 3662 DstReg = MI.getOperand(0).getReg(); 3663 if (isIntRegForSubInst(DstReg)) 3664 return HexagonII::HSIG_A; 3665 break; 3666 case Hexagon::C2_cmoveit: 3667 case Hexagon::C2_cmovenewit: 3668 case Hexagon::C2_cmoveif: 3669 case Hexagon::C2_cmovenewif: 3670 // if ([!]P0[.new]) Rd = #0 3671 // Actual form: 3672 // %R16<def> = C2_cmovenewit %P0<internal>, 0, %R16<imp-use,undef>; 3673 DstReg = MI.getOperand(0).getReg(); 3674 SrcReg = MI.getOperand(1).getReg(); 3675 if (isIntRegForSubInst(DstReg) && 3676 Hexagon::PredRegsRegClass.contains(SrcReg) && Hexagon::P0 == SrcReg && 3677 MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0) 3678 return HexagonII::HSIG_A; 3679 break; 3680 case Hexagon::C2_cmpeqi: 3681 // P0 = cmp.eq(Rs,#u2) 3682 DstReg = MI.getOperand(0).getReg(); 3683 SrcReg = MI.getOperand(1).getReg(); 3684 if (Hexagon::PredRegsRegClass.contains(DstReg) && 3685 Hexagon::P0 == DstReg && isIntRegForSubInst(SrcReg) && 3686 MI.getOperand(2).isImm() && isUInt<2>(MI.getOperand(2).getImm())) 3687 return HexagonII::HSIG_A; 3688 break; 3689 case Hexagon::A2_combineii: 3690 case Hexagon::A4_combineii: 3691 // Rdd = combine(#u2,#U2) 3692 DstReg = MI.getOperand(0).getReg(); 3693 if (isDblRegForSubInst(DstReg, HRI) && 3694 ((MI.getOperand(1).isImm() && isUInt<2>(MI.getOperand(1).getImm())) || 3695 (MI.getOperand(1).isGlobal() && 3696 isUInt<2>(MI.getOperand(1).getOffset()))) && 3697 ((MI.getOperand(2).isImm() && isUInt<2>(MI.getOperand(2).getImm())) || 3698 (MI.getOperand(2).isGlobal() && 3699 isUInt<2>(MI.getOperand(2).getOffset())))) 3700 return HexagonII::HSIG_A; 3701 break; 3702 case Hexagon::A4_combineri: 3703 // Rdd = combine(Rs,#0) 3704 DstReg = MI.getOperand(0).getReg(); 3705 SrcReg = MI.getOperand(1).getReg(); 3706 if (isDblRegForSubInst(DstReg, HRI) && isIntRegForSubInst(SrcReg) && 3707 ((MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0) || 3708 (MI.getOperand(2).isGlobal() && MI.getOperand(2).getOffset() == 0))) 3709 return HexagonII::HSIG_A; 3710 break; 3711 case Hexagon::A4_combineir: 3712 // Rdd = combine(#0,Rs) 3713 DstReg = MI.getOperand(0).getReg(); 3714 SrcReg = MI.getOperand(2).getReg(); 3715 if (isDblRegForSubInst(DstReg, HRI) && isIntRegForSubInst(SrcReg) && 3716 ((MI.getOperand(1).isImm() && MI.getOperand(1).getImm() == 0) || 3717 (MI.getOperand(1).isGlobal() && MI.getOperand(1).getOffset() == 0))) 3718 return HexagonII::HSIG_A; 3719 break; 3720 case Hexagon::A2_sxtb: 3721 case Hexagon::A2_sxth: 3722 case Hexagon::A2_zxtb: 3723 case Hexagon::A2_zxth: 3724 // Rd = sxth/sxtb/zxtb/zxth(Rs) 3725 DstReg = MI.getOperand(0).getReg(); 3726 SrcReg = MI.getOperand(1).getReg(); 3727 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg)) 3728 return HexagonII::HSIG_A; 3729 break; 3730 } 3731 3732 return HexagonII::HSIG_None; 3733 } 3734 3735 short HexagonInstrInfo::getEquivalentHWInstr(const MachineInstr &MI) const { 3736 return Hexagon::getRealHWInstr(MI.getOpcode(), Hexagon::InstrType_Real); 3737 } 3738 3739 unsigned HexagonInstrInfo::getInstrTimingClassLatency( 3740 const InstrItineraryData *ItinData, const MachineInstr &MI) const { 3741 // Default to one cycle for no itinerary. However, an "empty" itinerary may 3742 // still have a MinLatency property, which getStageLatency checks. 3743 if (!ItinData) 3744 return getInstrLatency(ItinData, MI); 3745 3746 if (MI.isTransient()) 3747 return 0; 3748 return ItinData->getStageLatency(MI.getDesc().getSchedClass()); 3749 } 3750 3751 /// getOperandLatency - Compute and return the use operand latency of a given 3752 /// pair of def and use. 3753 /// In most cases, the static scheduling itinerary was enough to determine the 3754 /// operand latency. But it may not be possible for instructions with variable 3755 /// number of defs / uses. 3756 /// 3757 /// This is a raw interface to the itinerary that may be directly overriden by 3758 /// a target. Use computeOperandLatency to get the best estimate of latency. 3759 int HexagonInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 3760 const MachineInstr &DefMI, 3761 unsigned DefIdx, 3762 const MachineInstr &UseMI, 3763 unsigned UseIdx) const { 3764 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 3765 3766 // Get DefIdx and UseIdx for super registers. 3767 MachineOperand DefMO = DefMI.getOperand(DefIdx); 3768 3769 if (HRI.isPhysicalRegister(DefMO.getReg())) { 3770 if (DefMO.isImplicit()) { 3771 for (MCSuperRegIterator SR(DefMO.getReg(), &HRI); SR.isValid(); ++SR) { 3772 int Idx = DefMI.findRegisterDefOperandIdx(*SR, false, false, &HRI); 3773 if (Idx != -1) { 3774 DefIdx = Idx; 3775 break; 3776 } 3777 } 3778 } 3779 3780 MachineOperand UseMO = UseMI.getOperand(UseIdx); 3781 if (UseMO.isImplicit()) { 3782 for (MCSuperRegIterator SR(UseMO.getReg(), &HRI); SR.isValid(); ++SR) { 3783 int Idx = UseMI.findRegisterUseOperandIdx(*SR, false, &HRI); 3784 if (Idx != -1) { 3785 UseIdx = Idx; 3786 break; 3787 } 3788 } 3789 } 3790 } 3791 3792 return TargetInstrInfo::getOperandLatency(ItinData, DefMI, DefIdx, 3793 UseMI, UseIdx); 3794 } 3795 3796 // inverts the predication logic. 3797 // p -> NotP 3798 // NotP -> P 3799 bool HexagonInstrInfo::getInvertedPredSense( 3800 SmallVectorImpl<MachineOperand> &Cond) const { 3801 if (Cond.empty()) 3802 return false; 3803 unsigned Opc = getInvertedPredicatedOpcode(Cond[0].getImm()); 3804 Cond[0].setImm(Opc); 3805 return true; 3806 } 3807 3808 unsigned HexagonInstrInfo::getInvertedPredicatedOpcode(const int Opc) const { 3809 int InvPredOpcode; 3810 InvPredOpcode = isPredicatedTrue(Opc) ? Hexagon::getFalsePredOpcode(Opc) 3811 : Hexagon::getTruePredOpcode(Opc); 3812 if (InvPredOpcode >= 0) // Valid instruction with the inverted predicate. 3813 return InvPredOpcode; 3814 3815 llvm_unreachable("Unexpected predicated instruction"); 3816 } 3817 3818 // Returns the max value that doesn't need to be extended. 3819 int HexagonInstrInfo::getMaxValue(const MachineInstr &MI) const { 3820 const uint64_t F = MI.getDesc().TSFlags; 3821 unsigned isSigned = (F >> HexagonII::ExtentSignedPos) 3822 & HexagonII::ExtentSignedMask; 3823 unsigned bits = (F >> HexagonII::ExtentBitsPos) 3824 & HexagonII::ExtentBitsMask; 3825 3826 if (isSigned) // if value is signed 3827 return ~(-1U << (bits - 1)); 3828 else 3829 return ~(-1U << bits); 3830 } 3831 3832 unsigned HexagonInstrInfo::getMemAccessSize(const MachineInstr &MI) const { 3833 using namespace HexagonII; 3834 3835 const uint64_t F = MI.getDesc().TSFlags; 3836 unsigned S = (F >> MemAccessSizePos) & MemAccesSizeMask; 3837 unsigned Size = getMemAccessSizeInBytes(MemAccessSize(S)); 3838 if (Size != 0) 3839 return Size; 3840 3841 // Handle vector access sizes. 3842 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 3843 switch (S) { 3844 case HexagonII::HVXVectorAccess: 3845 return HRI.getSpillSize(Hexagon::HvxVRRegClass); 3846 default: 3847 llvm_unreachable("Unexpected instruction"); 3848 } 3849 } 3850 3851 // Returns the min value that doesn't need to be extended. 3852 int HexagonInstrInfo::getMinValue(const MachineInstr &MI) const { 3853 const uint64_t F = MI.getDesc().TSFlags; 3854 unsigned isSigned = (F >> HexagonII::ExtentSignedPos) 3855 & HexagonII::ExtentSignedMask; 3856 unsigned bits = (F >> HexagonII::ExtentBitsPos) 3857 & HexagonII::ExtentBitsMask; 3858 3859 if (isSigned) // if value is signed 3860 return -1U << (bits - 1); 3861 else 3862 return 0; 3863 } 3864 3865 // Returns opcode of the non-extended equivalent instruction. 3866 short HexagonInstrInfo::getNonExtOpcode(const MachineInstr &MI) const { 3867 // Check if the instruction has a register form that uses register in place 3868 // of the extended operand, if so return that as the non-extended form. 3869 short NonExtOpcode = Hexagon::getRegForm(MI.getOpcode()); 3870 if (NonExtOpcode >= 0) 3871 return NonExtOpcode; 3872 3873 if (MI.getDesc().mayLoad() || MI.getDesc().mayStore()) { 3874 // Check addressing mode and retrieve non-ext equivalent instruction. 3875 switch (getAddrMode(MI)) { 3876 case HexagonII::Absolute: 3877 return Hexagon::changeAddrMode_abs_io(MI.getOpcode()); 3878 case HexagonII::BaseImmOffset: 3879 return Hexagon::changeAddrMode_io_rr(MI.getOpcode()); 3880 case HexagonII::BaseLongOffset: 3881 return Hexagon::changeAddrMode_ur_rr(MI.getOpcode()); 3882 3883 default: 3884 return -1; 3885 } 3886 } 3887 return -1; 3888 } 3889 3890 bool HexagonInstrInfo::getPredReg(ArrayRef<MachineOperand> Cond, 3891 unsigned &PredReg, unsigned &PredRegPos, unsigned &PredRegFlags) const { 3892 if (Cond.empty()) 3893 return false; 3894 assert(Cond.size() == 2); 3895 if (isNewValueJump(Cond[0].getImm()) || Cond[1].isMBB()) { 3896 DEBUG(dbgs() << "No predregs for new-value jumps/endloop"); 3897 return false; 3898 } 3899 PredReg = Cond[1].getReg(); 3900 PredRegPos = 1; 3901 // See IfConversion.cpp why we add RegState::Implicit | RegState::Undef 3902 PredRegFlags = 0; 3903 if (Cond[1].isImplicit()) 3904 PredRegFlags = RegState::Implicit; 3905 if (Cond[1].isUndef()) 3906 PredRegFlags |= RegState::Undef; 3907 return true; 3908 } 3909 3910 short HexagonInstrInfo::getPseudoInstrPair(const MachineInstr &MI) const { 3911 return Hexagon::getRealHWInstr(MI.getOpcode(), Hexagon::InstrType_Pseudo); 3912 } 3913 3914 short HexagonInstrInfo::getRegForm(const MachineInstr &MI) const { 3915 return Hexagon::getRegForm(MI.getOpcode()); 3916 } 3917 3918 // Return the number of bytes required to encode the instruction. 3919 // Hexagon instructions are fixed length, 4 bytes, unless they 3920 // use a constant extender, which requires another 4 bytes. 3921 // For debug instructions and prolog labels, return 0. 3922 unsigned HexagonInstrInfo::getSize(const MachineInstr &MI) const { 3923 if (MI.isDebugValue() || MI.isPosition()) 3924 return 0; 3925 3926 unsigned Size = MI.getDesc().getSize(); 3927 if (!Size) 3928 // Assume the default insn size in case it cannot be determined 3929 // for whatever reason. 3930 Size = HEXAGON_INSTR_SIZE; 3931 3932 if (isConstExtended(MI) || isExtended(MI)) 3933 Size += HEXAGON_INSTR_SIZE; 3934 3935 // Try and compute number of instructions in asm. 3936 if (BranchRelaxAsmLarge && MI.getOpcode() == Hexagon::INLINEASM) { 3937 const MachineBasicBlock &MBB = *MI.getParent(); 3938 const MachineFunction *MF = MBB.getParent(); 3939 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo(); 3940 3941 // Count the number of register definitions to find the asm string. 3942 unsigned NumDefs = 0; 3943 for (; MI.getOperand(NumDefs).isReg() && MI.getOperand(NumDefs).isDef(); 3944 ++NumDefs) 3945 assert(NumDefs != MI.getNumOperands()-2 && "No asm string?"); 3946 3947 assert(MI.getOperand(NumDefs).isSymbol() && "No asm string?"); 3948 // Disassemble the AsmStr and approximate number of instructions. 3949 const char *AsmStr = MI.getOperand(NumDefs).getSymbolName(); 3950 Size = getInlineAsmLength(AsmStr, *MAI); 3951 } 3952 3953 return Size; 3954 } 3955 3956 uint64_t HexagonInstrInfo::getType(const MachineInstr &MI) const { 3957 const uint64_t F = MI.getDesc().TSFlags; 3958 return (F >> HexagonII::TypePos) & HexagonII::TypeMask; 3959 } 3960 3961 unsigned HexagonInstrInfo::getUnits(const MachineInstr &MI) const { 3962 const InstrItineraryData &II = *Subtarget.getInstrItineraryData(); 3963 const InstrStage &IS = *II.beginStage(MI.getDesc().getSchedClass()); 3964 3965 return IS.getUnits(); 3966 } 3967 3968 // Calculate size of the basic block without debug instructions. 3969 unsigned HexagonInstrInfo::nonDbgBBSize(const MachineBasicBlock *BB) const { 3970 return nonDbgMICount(BB->instr_begin(), BB->instr_end()); 3971 } 3972 3973 unsigned HexagonInstrInfo::nonDbgBundleSize( 3974 MachineBasicBlock::const_iterator BundleHead) const { 3975 assert(BundleHead->isBundle() && "Not a bundle header"); 3976 auto MII = BundleHead.getInstrIterator(); 3977 // Skip the bundle header. 3978 return nonDbgMICount(++MII, getBundleEnd(BundleHead.getInstrIterator())); 3979 } 3980 3981 /// immediateExtend - Changes the instruction in place to one using an immediate 3982 /// extender. 3983 void HexagonInstrInfo::immediateExtend(MachineInstr &MI) const { 3984 assert((isExtendable(MI)||isConstExtended(MI)) && 3985 "Instruction must be extendable"); 3986 // Find which operand is extendable. 3987 short ExtOpNum = getCExtOpNum(MI); 3988 MachineOperand &MO = MI.getOperand(ExtOpNum); 3989 // This needs to be something we understand. 3990 assert((MO.isMBB() || MO.isImm()) && 3991 "Branch with unknown extendable field type"); 3992 // Mark given operand as extended. 3993 MO.addTargetFlag(HexagonII::HMOTF_ConstExtended); 3994 } 3995 3996 bool HexagonInstrInfo::invertAndChangeJumpTarget( 3997 MachineInstr &MI, MachineBasicBlock *NewTarget) const { 3998 DEBUG(dbgs() << "\n[invertAndChangeJumpTarget] to BB#" 3999 << NewTarget->getNumber(); MI.dump();); 4000 assert(MI.isBranch()); 4001 unsigned NewOpcode = getInvertedPredicatedOpcode(MI.getOpcode()); 4002 int TargetPos = MI.getNumOperands() - 1; 4003 // In general branch target is the last operand, 4004 // but some implicit defs added at the end might change it. 4005 while ((TargetPos > -1) && !MI.getOperand(TargetPos).isMBB()) 4006 --TargetPos; 4007 assert((TargetPos >= 0) && MI.getOperand(TargetPos).isMBB()); 4008 MI.getOperand(TargetPos).setMBB(NewTarget); 4009 if (EnableBranchPrediction && isPredicatedNew(MI)) { 4010 NewOpcode = reversePrediction(NewOpcode); 4011 } 4012 MI.setDesc(get(NewOpcode)); 4013 return true; 4014 } 4015 4016 void HexagonInstrInfo::genAllInsnTimingClasses(MachineFunction &MF) const { 4017 /* +++ The code below is used to generate complete set of Hexagon Insn +++ */ 4018 MachineFunction::iterator A = MF.begin(); 4019 MachineBasicBlock &B = *A; 4020 MachineBasicBlock::iterator I = B.begin(); 4021 DebugLoc DL = I->getDebugLoc(); 4022 MachineInstr *NewMI; 4023 4024 for (unsigned insn = TargetOpcode::GENERIC_OP_END+1; 4025 insn < Hexagon::INSTRUCTION_LIST_END; ++insn) { 4026 NewMI = BuildMI(B, I, DL, get(insn)); 4027 DEBUG(dbgs() << "\n" << getName(NewMI->getOpcode()) << 4028 " Class: " << NewMI->getDesc().getSchedClass()); 4029 NewMI->eraseFromParent(); 4030 } 4031 /* --- The code above is used to generate complete set of Hexagon Insn --- */ 4032 } 4033 4034 // inverts the predication logic. 4035 // p -> NotP 4036 // NotP -> P 4037 bool HexagonInstrInfo::reversePredSense(MachineInstr &MI) const { 4038 DEBUG(dbgs() << "\nTrying to reverse pred. sense of:"; MI.dump()); 4039 MI.setDesc(get(getInvertedPredicatedOpcode(MI.getOpcode()))); 4040 return true; 4041 } 4042 4043 // Reverse the branch prediction. 4044 unsigned HexagonInstrInfo::reversePrediction(unsigned Opcode) const { 4045 int PredRevOpcode = -1; 4046 if (isPredictedTaken(Opcode)) 4047 PredRevOpcode = Hexagon::notTakenBranchPrediction(Opcode); 4048 else 4049 PredRevOpcode = Hexagon::takenBranchPrediction(Opcode); 4050 assert(PredRevOpcode > 0); 4051 return PredRevOpcode; 4052 } 4053 4054 // TODO: Add more rigorous validation. 4055 bool HexagonInstrInfo::validateBranchCond(const ArrayRef<MachineOperand> &Cond) 4056 const { 4057 return Cond.empty() || (Cond[0].isImm() && (Cond.size() != 1)); 4058 } 4059 4060 // Addressing mode relations. 4061 short HexagonInstrInfo::changeAddrMode_abs_io(short Opc) const { 4062 return Opc >= 0 ? Hexagon::changeAddrMode_abs_io(Opc) : Opc; 4063 } 4064 4065 short HexagonInstrInfo::changeAddrMode_io_abs(short Opc) const { 4066 return Opc >= 0 ? Hexagon::changeAddrMode_io_abs(Opc) : Opc; 4067 } 4068 4069 short HexagonInstrInfo::changeAddrMode_io_rr(short Opc) const { 4070 return Opc >= 0 ? Hexagon::changeAddrMode_io_rr(Opc) : Opc; 4071 } 4072 4073 short HexagonInstrInfo::changeAddrMode_rr_io(short Opc) const { 4074 return Opc >= 0 ? Hexagon::changeAddrMode_rr_io(Opc) : Opc; 4075 } 4076 4077 short HexagonInstrInfo::changeAddrMode_rr_ur(short Opc) const { 4078 return Opc >= 0 ? Hexagon::changeAddrMode_rr_ur(Opc) : Opc; 4079 } 4080 4081 short HexagonInstrInfo::changeAddrMode_ur_rr(short Opc) const { 4082 return Opc >= 0 ? Hexagon::changeAddrMode_ur_rr(Opc) : Opc; 4083 } 4084