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 "HexagonRegisterInfo.h" 17 #include "HexagonSubtarget.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SmallVector.h" 20 #include "llvm/CodeGen/DFAPacketizer.h" 21 #include "llvm/CodeGen/MachineFrameInfo.h" 22 #include "llvm/CodeGen/MachineInstrBuilder.h" 23 #include "llvm/CodeGen/MachineMemOperand.h" 24 #include "llvm/CodeGen/MachineRegisterInfo.h" 25 #include "llvm/CodeGen/PseudoSourceValue.h" 26 #include "llvm/MC/MCAsmInfo.h" 27 #include "llvm/Support/CommandLine.h" 28 #include "llvm/Support/Debug.h" 29 #include "llvm/Support/MathExtras.h" 30 #include "llvm/Support/raw_ostream.h" 31 #include <cctype> 32 33 using namespace llvm; 34 35 #define DEBUG_TYPE "hexagon-instrinfo" 36 37 #define GET_INSTRINFO_CTOR_DTOR 38 #define GET_INSTRMAP_INFO 39 #include "HexagonGenInstrInfo.inc" 40 #include "HexagonGenDFAPacketizer.inc" 41 42 using namespace llvm; 43 44 cl::opt<bool> ScheduleInlineAsm("hexagon-sched-inline-asm", cl::Hidden, 45 cl::init(false), cl::desc("Do not consider inline-asm a scheduling/" 46 "packetization boundary.")); 47 48 static cl::opt<bool> EnableBranchPrediction("hexagon-enable-branch-prediction", 49 cl::Hidden, cl::init(true), cl::desc("Enable branch prediction")); 50 51 static cl::opt<bool> DisableNVSchedule("disable-hexagon-nv-schedule", 52 cl::Hidden, cl::ZeroOrMore, cl::init(false), 53 cl::desc("Disable schedule adjustment for new value stores.")); 54 55 static cl::opt<bool> EnableTimingClassLatency( 56 "enable-timing-class-latency", cl::Hidden, cl::init(false), 57 cl::desc("Enable timing class latency")); 58 59 static cl::opt<bool> EnableALUForwarding( 60 "enable-alu-forwarding", cl::Hidden, cl::init(true), 61 cl::desc("Enable vec alu forwarding")); 62 63 static cl::opt<bool> EnableACCForwarding( 64 "enable-acc-forwarding", cl::Hidden, cl::init(true), 65 cl::desc("Enable vec acc forwarding")); 66 67 static cl::opt<bool> BranchRelaxAsmLarge("branch-relax-asm-large", 68 cl::init(true), cl::Hidden, cl::ZeroOrMore, cl::desc("branch relax asm")); 69 70 /// 71 /// Constants for Hexagon instructions. 72 /// 73 const int Hexagon_MEMV_OFFSET_MAX_128B = 896; // #s4: -8*128...7*128 74 const int Hexagon_MEMV_OFFSET_MIN_128B = -1024; // #s4 75 const int Hexagon_MEMV_OFFSET_MAX = 448; // #s4: -8*64...7*64 76 const int Hexagon_MEMV_OFFSET_MIN = -512; // #s4 77 const int Hexagon_MEMW_OFFSET_MAX = 4095; 78 const int Hexagon_MEMW_OFFSET_MIN = -4096; 79 const int Hexagon_MEMD_OFFSET_MAX = 8191; 80 const int Hexagon_MEMD_OFFSET_MIN = -8192; 81 const int Hexagon_MEMH_OFFSET_MAX = 2047; 82 const int Hexagon_MEMH_OFFSET_MIN = -2048; 83 const int Hexagon_MEMB_OFFSET_MAX = 1023; 84 const int Hexagon_MEMB_OFFSET_MIN = -1024; 85 const int Hexagon_ADDI_OFFSET_MAX = 32767; 86 const int Hexagon_ADDI_OFFSET_MIN = -32768; 87 const int Hexagon_MEMD_AUTOINC_MAX = 56; 88 const int Hexagon_MEMD_AUTOINC_MIN = -64; 89 const int Hexagon_MEMW_AUTOINC_MAX = 28; 90 const int Hexagon_MEMW_AUTOINC_MIN = -32; 91 const int Hexagon_MEMH_AUTOINC_MAX = 14; 92 const int Hexagon_MEMH_AUTOINC_MIN = -16; 93 const int Hexagon_MEMB_AUTOINC_MAX = 7; 94 const int Hexagon_MEMB_AUTOINC_MIN = -8; 95 const int Hexagon_MEMV_AUTOINC_MAX = 192; // #s3 96 const int Hexagon_MEMV_AUTOINC_MIN = -256; // #s3 97 const int Hexagon_MEMV_AUTOINC_MAX_128B = 384; // #s3 98 const int Hexagon_MEMV_AUTOINC_MIN_128B = -512; // #s3 99 100 // Pin the vtable to this file. 101 void HexagonInstrInfo::anchor() {} 102 103 HexagonInstrInfo::HexagonInstrInfo(HexagonSubtarget &ST) 104 : HexagonGenInstrInfo(Hexagon::ADJCALLSTACKDOWN, Hexagon::ADJCALLSTACKUP), 105 RI() {} 106 107 108 static bool isIntRegForSubInst(unsigned Reg) { 109 return (Reg >= Hexagon::R0 && Reg <= Hexagon::R7) || 110 (Reg >= Hexagon::R16 && Reg <= Hexagon::R23); 111 } 112 113 114 static bool isDblRegForSubInst(unsigned Reg, const HexagonRegisterInfo &HRI) { 115 return isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::subreg_loreg)) && 116 isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::subreg_hireg)); 117 } 118 119 120 /// Calculate number of instructions excluding the debug instructions. 121 static unsigned nonDbgMICount(MachineBasicBlock::const_instr_iterator MIB, 122 MachineBasicBlock::const_instr_iterator MIE) { 123 unsigned Count = 0; 124 for (; MIB != MIE; ++MIB) { 125 if (!MIB->isDebugValue()) 126 ++Count; 127 } 128 return Count; 129 } 130 131 132 /// Find the hardware loop instruction used to set-up the specified loop. 133 /// On Hexagon, we have two instructions used to set-up the hardware loop 134 /// (LOOP0, LOOP1) with corresponding endloop (ENDLOOP0, ENDLOOP1) instructions 135 /// to indicate the end of a loop. 136 static MachineInstr *findLoopInstr(MachineBasicBlock *BB, int EndLoopOp, 137 SmallPtrSet<MachineBasicBlock *, 8> &Visited) { 138 int LOOPi; 139 int LOOPr; 140 if (EndLoopOp == Hexagon::ENDLOOP0) { 141 LOOPi = Hexagon::J2_loop0i; 142 LOOPr = Hexagon::J2_loop0r; 143 } else { // EndLoopOp == Hexagon::EndLOOP1 144 LOOPi = Hexagon::J2_loop1i; 145 LOOPr = Hexagon::J2_loop1r; 146 } 147 148 // The loop set-up instruction will be in a predecessor block 149 for (MachineBasicBlock::pred_iterator PB = BB->pred_begin(), 150 PE = BB->pred_end(); PB != PE; ++PB) { 151 // If this has been visited, already skip it. 152 if (!Visited.insert(*PB).second) 153 continue; 154 if (*PB == BB) 155 continue; 156 for (MachineBasicBlock::reverse_instr_iterator I = (*PB)->instr_rbegin(), 157 E = (*PB)->instr_rend(); I != E; ++I) { 158 int Opc = I->getOpcode(); 159 if (Opc == LOOPi || Opc == LOOPr) 160 return &*I; 161 // We've reached a different loop, which means the loop0 has been removed. 162 if (Opc == EndLoopOp) 163 return 0; 164 } 165 // Check the predecessors for the LOOP instruction. 166 MachineInstr *loop = findLoopInstr(*PB, EndLoopOp, Visited); 167 if (loop) 168 return loop; 169 } 170 return 0; 171 } 172 173 174 /// Gather register def/uses from MI. 175 /// This treats possible (predicated) defs as actually happening ones 176 /// (conservatively). 177 static inline void parseOperands(const MachineInstr *MI, 178 SmallVector<unsigned, 4> &Defs, SmallVector<unsigned, 8> &Uses) { 179 Defs.clear(); 180 Uses.clear(); 181 182 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 183 const MachineOperand &MO = MI->getOperand(i); 184 185 if (!MO.isReg()) 186 continue; 187 188 unsigned Reg = MO.getReg(); 189 if (!Reg) 190 continue; 191 192 if (MO.isUse()) 193 Uses.push_back(MO.getReg()); 194 195 if (MO.isDef()) 196 Defs.push_back(MO.getReg()); 197 } 198 } 199 200 201 // Position dependent, so check twice for swap. 202 static bool isDuplexPairMatch(unsigned Ga, unsigned Gb) { 203 switch (Ga) { 204 case HexagonII::HSIG_None: 205 default: 206 return false; 207 case HexagonII::HSIG_L1: 208 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_A); 209 case HexagonII::HSIG_L2: 210 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 || 211 Gb == HexagonII::HSIG_A); 212 case HexagonII::HSIG_S1: 213 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 || 214 Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_A); 215 case HexagonII::HSIG_S2: 216 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 || 217 Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_S2 || 218 Gb == HexagonII::HSIG_A); 219 case HexagonII::HSIG_A: 220 return (Gb == HexagonII::HSIG_A); 221 case HexagonII::HSIG_Compound: 222 return (Gb == HexagonII::HSIG_Compound); 223 } 224 return false; 225 } 226 227 228 229 /// isLoadFromStackSlot - If the specified machine instruction is a direct 230 /// load from a stack slot, return the virtual or physical register number of 231 /// the destination along with the FrameIndex of the loaded stack slot. If 232 /// not, return 0. This predicate must return 0 if the instruction has 233 /// any side effects other than loading from the stack slot. 234 unsigned HexagonInstrInfo::isLoadFromStackSlot(const MachineInstr *MI, 235 int &FrameIndex) const { 236 switch (MI->getOpcode()) { 237 default: 238 break; 239 case Hexagon::L2_loadrb_io: 240 case Hexagon::L2_loadrub_io: 241 case Hexagon::L2_loadrh_io: 242 case Hexagon::L2_loadruh_io: 243 case Hexagon::L2_loadri_io: 244 case Hexagon::L2_loadrd_io: 245 case Hexagon::V6_vL32b_ai: 246 case Hexagon::V6_vL32b_ai_128B: 247 case Hexagon::V6_vL32Ub_ai: 248 case Hexagon::V6_vL32Ub_ai_128B: 249 case Hexagon::LDriw_pred: 250 case Hexagon::LDriw_mod: 251 case Hexagon::LDriq_pred_V6: 252 case Hexagon::LDriq_pred_vec_V6: 253 case Hexagon::LDriv_pseudo_V6: 254 case Hexagon::LDrivv_pseudo_V6: 255 case Hexagon::LDriq_pred_V6_128B: 256 case Hexagon::LDriq_pred_vec_V6_128B: 257 case Hexagon::LDriv_pseudo_V6_128B: 258 case Hexagon::LDrivv_pseudo_V6_128B: { 259 const MachineOperand OpFI = MI->getOperand(1); 260 if (!OpFI.isFI()) 261 return 0; 262 const MachineOperand OpOff = MI->getOperand(2); 263 if (!OpOff.isImm() || OpOff.getImm() != 0) 264 return 0; 265 FrameIndex = OpFI.getIndex(); 266 return MI->getOperand(0).getReg(); 267 } 268 269 case Hexagon::L2_ploadrbt_io: 270 case Hexagon::L2_ploadrbf_io: 271 case Hexagon::L2_ploadrubt_io: 272 case Hexagon::L2_ploadrubf_io: 273 case Hexagon::L2_ploadrht_io: 274 case Hexagon::L2_ploadrhf_io: 275 case Hexagon::L2_ploadruht_io: 276 case Hexagon::L2_ploadruhf_io: 277 case Hexagon::L2_ploadrit_io: 278 case Hexagon::L2_ploadrif_io: 279 case Hexagon::L2_ploadrdt_io: 280 case Hexagon::L2_ploadrdf_io: { 281 const MachineOperand OpFI = MI->getOperand(2); 282 if (!OpFI.isFI()) 283 return 0; 284 const MachineOperand OpOff = MI->getOperand(3); 285 if (!OpOff.isImm() || OpOff.getImm() != 0) 286 return 0; 287 FrameIndex = OpFI.getIndex(); 288 return MI->getOperand(0).getReg(); 289 } 290 } 291 292 return 0; 293 } 294 295 296 /// isStoreToStackSlot - If the specified machine instruction is a direct 297 /// store to a stack slot, return the virtual or physical register number of 298 /// the source reg along with the FrameIndex of the loaded stack slot. If 299 /// not, return 0. This predicate must return 0 if the instruction has 300 /// any side effects other than storing to the stack slot. 301 unsigned HexagonInstrInfo::isStoreToStackSlot(const MachineInstr *MI, 302 int &FrameIndex) const { 303 switch (MI->getOpcode()) { 304 default: 305 break; 306 case Hexagon::S2_storerb_io: 307 case Hexagon::S2_storerh_io: 308 case Hexagon::S2_storeri_io: 309 case Hexagon::S2_storerd_io: 310 case Hexagon::V6_vS32b_ai: 311 case Hexagon::V6_vS32b_ai_128B: 312 case Hexagon::V6_vS32Ub_ai: 313 case Hexagon::V6_vS32Ub_ai_128B: 314 case Hexagon::STriw_pred: 315 case Hexagon::STriw_mod: 316 case Hexagon::STriq_pred_V6: 317 case Hexagon::STriq_pred_vec_V6: 318 case Hexagon::STriv_pseudo_V6: 319 case Hexagon::STrivv_pseudo_V6: 320 case Hexagon::STriq_pred_V6_128B: 321 case Hexagon::STriq_pred_vec_V6_128B: 322 case Hexagon::STriv_pseudo_V6_128B: 323 case Hexagon::STrivv_pseudo_V6_128B: { 324 const MachineOperand &OpFI = MI->getOperand(0); 325 if (!OpFI.isFI()) 326 return 0; 327 const MachineOperand &OpOff = MI->getOperand(1); 328 if (!OpOff.isImm() || OpOff.getImm() != 0) 329 return 0; 330 FrameIndex = OpFI.getIndex(); 331 return MI->getOperand(2).getReg(); 332 } 333 334 case Hexagon::S2_pstorerbt_io: 335 case Hexagon::S2_pstorerbf_io: 336 case Hexagon::S2_pstorerht_io: 337 case Hexagon::S2_pstorerhf_io: 338 case Hexagon::S2_pstorerit_io: 339 case Hexagon::S2_pstorerif_io: 340 case Hexagon::S2_pstorerdt_io: 341 case Hexagon::S2_pstorerdf_io: { 342 const MachineOperand &OpFI = MI->getOperand(1); 343 if (!OpFI.isFI()) 344 return 0; 345 const MachineOperand &OpOff = MI->getOperand(2); 346 if (!OpOff.isImm() || OpOff.getImm() != 0) 347 return 0; 348 FrameIndex = OpFI.getIndex(); 349 return MI->getOperand(3).getReg(); 350 } 351 } 352 353 return 0; 354 } 355 356 357 /// This function can analyze one/two way branching only and should (mostly) be 358 /// called by target independent side. 359 /// First entry is always the opcode of the branching instruction, except when 360 /// the Cond vector is supposed to be empty, e.g., when AnalyzeBranch fails, a 361 /// BB with only unconditional jump. Subsequent entries depend upon the opcode, 362 /// e.g. Jump_c p will have 363 /// Cond[0] = Jump_c 364 /// Cond[1] = p 365 /// HW-loop ENDLOOP: 366 /// Cond[0] = ENDLOOP 367 /// Cond[1] = MBB 368 /// New value jump: 369 /// Cond[0] = Hexagon::CMPEQri_f_Jumpnv_t_V4 -- specific opcode 370 /// Cond[1] = R 371 /// Cond[2] = Imm 372 /// 373 bool HexagonInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, 374 MachineBasicBlock *&TBB, 375 MachineBasicBlock *&FBB, 376 SmallVectorImpl<MachineOperand> &Cond, 377 bool AllowModify) const { 378 TBB = nullptr; 379 FBB = nullptr; 380 Cond.clear(); 381 382 // If the block has no terminators, it just falls into the block after it. 383 MachineBasicBlock::instr_iterator I = MBB.instr_end(); 384 if (I == MBB.instr_begin()) 385 return false; 386 387 // A basic block may looks like this: 388 // 389 // [ insn 390 // EH_LABEL 391 // insn 392 // insn 393 // insn 394 // EH_LABEL 395 // insn ] 396 // 397 // It has two succs but does not have a terminator 398 // Don't know how to handle it. 399 do { 400 --I; 401 if (I->isEHLabel()) 402 // Don't analyze EH branches. 403 return true; 404 } while (I != MBB.instr_begin()); 405 406 I = MBB.instr_end(); 407 --I; 408 409 while (I->isDebugValue()) { 410 if (I == MBB.instr_begin()) 411 return false; 412 --I; 413 } 414 415 bool JumpToBlock = I->getOpcode() == Hexagon::J2_jump && 416 I->getOperand(0).isMBB(); 417 // Delete the J2_jump if it's equivalent to a fall-through. 418 if (AllowModify && JumpToBlock && 419 MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) { 420 DEBUG(dbgs()<< "\nErasing the jump to successor block\n";); 421 I->eraseFromParent(); 422 I = MBB.instr_end(); 423 if (I == MBB.instr_begin()) 424 return false; 425 --I; 426 } 427 if (!isUnpredicatedTerminator(*I)) 428 return false; 429 430 // Get the last instruction in the block. 431 MachineInstr *LastInst = &*I; 432 MachineInstr *SecondLastInst = nullptr; 433 // Find one more terminator if present. 434 for (;;) { 435 if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(*I)) { 436 if (!SecondLastInst) 437 SecondLastInst = &*I; 438 else 439 // This is a third branch. 440 return true; 441 } 442 if (I == MBB.instr_begin()) 443 break; 444 --I; 445 } 446 447 int LastOpcode = LastInst->getOpcode(); 448 int SecLastOpcode = SecondLastInst ? SecondLastInst->getOpcode() : 0; 449 // If the branch target is not a basic block, it could be a tail call. 450 // (It is, if the target is a function.) 451 if (LastOpcode == Hexagon::J2_jump && !LastInst->getOperand(0).isMBB()) 452 return true; 453 if (SecLastOpcode == Hexagon::J2_jump && 454 !SecondLastInst->getOperand(0).isMBB()) 455 return true; 456 457 bool LastOpcodeHasJMP_c = PredOpcodeHasJMP_c(LastOpcode); 458 bool LastOpcodeHasNVJump = isNewValueJump(LastInst); 459 460 if (LastOpcodeHasJMP_c && !LastInst->getOperand(1).isMBB()) 461 return true; 462 463 // If there is only one terminator instruction, process it. 464 if (LastInst && !SecondLastInst) { 465 if (LastOpcode == Hexagon::J2_jump) { 466 TBB = LastInst->getOperand(0).getMBB(); 467 return false; 468 } 469 if (isEndLoopN(LastOpcode)) { 470 TBB = LastInst->getOperand(0).getMBB(); 471 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode())); 472 Cond.push_back(LastInst->getOperand(0)); 473 return false; 474 } 475 if (LastOpcodeHasJMP_c) { 476 TBB = LastInst->getOperand(1).getMBB(); 477 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode())); 478 Cond.push_back(LastInst->getOperand(0)); 479 return false; 480 } 481 // Only supporting rr/ri versions of new-value jumps. 482 if (LastOpcodeHasNVJump && (LastInst->getNumExplicitOperands() == 3)) { 483 TBB = LastInst->getOperand(2).getMBB(); 484 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode())); 485 Cond.push_back(LastInst->getOperand(0)); 486 Cond.push_back(LastInst->getOperand(1)); 487 return false; 488 } 489 DEBUG(dbgs() << "\nCant analyze BB#" << MBB.getNumber() 490 << " with one jump\n";); 491 // Otherwise, don't know what this is. 492 return true; 493 } 494 495 bool SecLastOpcodeHasJMP_c = PredOpcodeHasJMP_c(SecLastOpcode); 496 bool SecLastOpcodeHasNVJump = isNewValueJump(SecondLastInst); 497 if (SecLastOpcodeHasJMP_c && (LastOpcode == Hexagon::J2_jump)) { 498 if (!SecondLastInst->getOperand(1).isMBB()) 499 return true; 500 TBB = SecondLastInst->getOperand(1).getMBB(); 501 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode())); 502 Cond.push_back(SecondLastInst->getOperand(0)); 503 FBB = LastInst->getOperand(0).getMBB(); 504 return false; 505 } 506 507 // Only supporting rr/ri versions of new-value jumps. 508 if (SecLastOpcodeHasNVJump && 509 (SecondLastInst->getNumExplicitOperands() == 3) && 510 (LastOpcode == Hexagon::J2_jump)) { 511 TBB = SecondLastInst->getOperand(2).getMBB(); 512 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode())); 513 Cond.push_back(SecondLastInst->getOperand(0)); 514 Cond.push_back(SecondLastInst->getOperand(1)); 515 FBB = LastInst->getOperand(0).getMBB(); 516 return false; 517 } 518 519 // If the block ends with two Hexagon:JMPs, handle it. The second one is not 520 // executed, so remove it. 521 if (SecLastOpcode == Hexagon::J2_jump && LastOpcode == Hexagon::J2_jump) { 522 TBB = SecondLastInst->getOperand(0).getMBB(); 523 I = LastInst->getIterator(); 524 if (AllowModify) 525 I->eraseFromParent(); 526 return false; 527 } 528 529 // If the block ends with an ENDLOOP, and J2_jump, handle it. 530 if (isEndLoopN(SecLastOpcode) && LastOpcode == Hexagon::J2_jump) { 531 TBB = SecondLastInst->getOperand(0).getMBB(); 532 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode())); 533 Cond.push_back(SecondLastInst->getOperand(0)); 534 FBB = LastInst->getOperand(0).getMBB(); 535 return false; 536 } 537 DEBUG(dbgs() << "\nCant analyze BB#" << MBB.getNumber() 538 << " with two jumps";); 539 // Otherwise, can't handle this. 540 return true; 541 } 542 543 544 unsigned HexagonInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { 545 DEBUG(dbgs() << "\nRemoving branches out of BB#" << MBB.getNumber()); 546 MachineBasicBlock::iterator I = MBB.end(); 547 unsigned Count = 0; 548 while (I != MBB.begin()) { 549 --I; 550 if (I->isDebugValue()) 551 continue; 552 // Only removing branches from end of MBB. 553 if (!I->isBranch()) 554 return Count; 555 if (Count && (I->getOpcode() == Hexagon::J2_jump)) 556 llvm_unreachable("Malformed basic block: unconditional branch not last"); 557 MBB.erase(&MBB.back()); 558 I = MBB.end(); 559 ++Count; 560 } 561 return Count; 562 } 563 564 unsigned HexagonInstrInfo::InsertBranch(MachineBasicBlock &MBB, 565 MachineBasicBlock *TBB, 566 MachineBasicBlock *FBB, 567 ArrayRef<MachineOperand> Cond, 568 const DebugLoc &DL) const { 569 unsigned BOpc = Hexagon::J2_jump; 570 unsigned BccOpc = Hexagon::J2_jumpt; 571 assert(validateBranchCond(Cond) && "Invalid branching condition"); 572 assert(TBB && "InsertBranch must not be told to insert a fallthrough"); 573 574 // Check if ReverseBranchCondition has asked to reverse this branch 575 // If we want to reverse the branch an odd number of times, we want 576 // J2_jumpf. 577 if (!Cond.empty() && Cond[0].isImm()) 578 BccOpc = Cond[0].getImm(); 579 580 if (!FBB) { 581 if (Cond.empty()) { 582 // Due to a bug in TailMerging/CFG Optimization, we need to add a 583 // special case handling of a predicated jump followed by an 584 // unconditional jump. If not, Tail Merging and CFG Optimization go 585 // into an infinite loop. 586 MachineBasicBlock *NewTBB, *NewFBB; 587 SmallVector<MachineOperand, 4> Cond; 588 MachineInstr *Term = MBB.getFirstTerminator(); 589 if (Term != MBB.end() && isPredicated(*Term) && 590 !AnalyzeBranch(MBB, NewTBB, NewFBB, Cond, false)) { 591 MachineBasicBlock *NextBB = &*++MBB.getIterator(); 592 if (NewTBB == NextBB) { 593 ReverseBranchCondition(Cond); 594 RemoveBranch(MBB); 595 return InsertBranch(MBB, TBB, nullptr, Cond, DL); 596 } 597 } 598 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB); 599 } else if (isEndLoopN(Cond[0].getImm())) { 600 int EndLoopOp = Cond[0].getImm(); 601 assert(Cond[1].isMBB()); 602 // Since we're adding an ENDLOOP, there better be a LOOP instruction. 603 // Check for it, and change the BB target if needed. 604 SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs; 605 MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, VisitedBBs); 606 assert(Loop != 0 && "Inserting an ENDLOOP without a LOOP"); 607 Loop->getOperand(0).setMBB(TBB); 608 // Add the ENDLOOP after the finding the LOOP0. 609 BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB); 610 } else if (isNewValueJump(Cond[0].getImm())) { 611 assert((Cond.size() == 3) && "Only supporting rr/ri version of nvjump"); 612 // New value jump 613 // (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset) 614 // (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset) 615 unsigned Flags1 = getUndefRegState(Cond[1].isUndef()); 616 DEBUG(dbgs() << "\nInserting NVJump for BB#" << MBB.getNumber();); 617 if (Cond[2].isReg()) { 618 unsigned Flags2 = getUndefRegState(Cond[2].isUndef()); 619 BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1). 620 addReg(Cond[2].getReg(), Flags2).addMBB(TBB); 621 } else if(Cond[2].isImm()) { 622 BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1). 623 addImm(Cond[2].getImm()).addMBB(TBB); 624 } else 625 llvm_unreachable("Invalid condition for branching"); 626 } else { 627 assert((Cond.size() == 2) && "Malformed cond vector"); 628 const MachineOperand &RO = Cond[1]; 629 unsigned Flags = getUndefRegState(RO.isUndef()); 630 BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB); 631 } 632 return 1; 633 } 634 assert((!Cond.empty()) && 635 "Cond. cannot be empty when multiple branchings are required"); 636 assert((!isNewValueJump(Cond[0].getImm())) && 637 "NV-jump cannot be inserted with another branch"); 638 // Special case for hardware loops. The condition is a basic block. 639 if (isEndLoopN(Cond[0].getImm())) { 640 int EndLoopOp = Cond[0].getImm(); 641 assert(Cond[1].isMBB()); 642 // Since we're adding an ENDLOOP, there better be a LOOP instruction. 643 // Check for it, and change the BB target if needed. 644 SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs; 645 MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, VisitedBBs); 646 assert(Loop != 0 && "Inserting an ENDLOOP without a LOOP"); 647 Loop->getOperand(0).setMBB(TBB); 648 // Add the ENDLOOP after the finding the LOOP0. 649 BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB); 650 } else { 651 const MachineOperand &RO = Cond[1]; 652 unsigned Flags = getUndefRegState(RO.isUndef()); 653 BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB); 654 } 655 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB); 656 657 return 2; 658 } 659 660 661 bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB, 662 unsigned NumCycles, unsigned ExtraPredCycles, 663 BranchProbability Probability) const { 664 return nonDbgBBSize(&MBB) <= 3; 665 } 666 667 668 bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB, 669 unsigned NumTCycles, unsigned ExtraTCycles, MachineBasicBlock &FMBB, 670 unsigned NumFCycles, unsigned ExtraFCycles, BranchProbability Probability) 671 const { 672 return nonDbgBBSize(&TMBB) <= 3 && nonDbgBBSize(&FMBB) <= 3; 673 } 674 675 676 bool HexagonInstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB, 677 unsigned NumInstrs, BranchProbability Probability) const { 678 return NumInstrs <= 4; 679 } 680 681 void HexagonInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 682 MachineBasicBlock::iterator I, 683 const DebugLoc &DL, unsigned DestReg, 684 unsigned SrcReg, bool KillSrc) const { 685 auto &HRI = getRegisterInfo(); 686 unsigned KillFlag = getKillRegState(KillSrc); 687 688 if (Hexagon::IntRegsRegClass.contains(SrcReg, DestReg)) { 689 BuildMI(MBB, I, DL, get(Hexagon::A2_tfr), DestReg) 690 .addReg(SrcReg, KillFlag); 691 return; 692 } 693 if (Hexagon::DoubleRegsRegClass.contains(SrcReg, DestReg)) { 694 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrp), DestReg) 695 .addReg(SrcReg, KillFlag); 696 return; 697 } 698 if (Hexagon::PredRegsRegClass.contains(SrcReg, DestReg)) { 699 // Map Pd = Ps to Pd = or(Ps, Ps). 700 BuildMI(MBB, I, DL, get(Hexagon::C2_or), DestReg) 701 .addReg(SrcReg).addReg(SrcReg, KillFlag); 702 return; 703 } 704 if (Hexagon::CtrRegsRegClass.contains(DestReg) && 705 Hexagon::IntRegsRegClass.contains(SrcReg)) { 706 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg) 707 .addReg(SrcReg, KillFlag); 708 return; 709 } 710 if (Hexagon::IntRegsRegClass.contains(DestReg) && 711 Hexagon::CtrRegsRegClass.contains(SrcReg)) { 712 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrcrr), DestReg) 713 .addReg(SrcReg, KillFlag); 714 return; 715 } 716 if (Hexagon::ModRegsRegClass.contains(DestReg) && 717 Hexagon::IntRegsRegClass.contains(SrcReg)) { 718 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg) 719 .addReg(SrcReg, KillFlag); 720 return; 721 } 722 if (Hexagon::PredRegsRegClass.contains(SrcReg) && 723 Hexagon::IntRegsRegClass.contains(DestReg)) { 724 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg) 725 .addReg(SrcReg, KillFlag); 726 return; 727 } 728 if (Hexagon::IntRegsRegClass.contains(SrcReg) && 729 Hexagon::PredRegsRegClass.contains(DestReg)) { 730 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrrp), DestReg) 731 .addReg(SrcReg, KillFlag); 732 return; 733 } 734 if (Hexagon::PredRegsRegClass.contains(SrcReg) && 735 Hexagon::IntRegsRegClass.contains(DestReg)) { 736 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg) 737 .addReg(SrcReg, KillFlag); 738 return; 739 } 740 if (Hexagon::VectorRegsRegClass.contains(SrcReg, DestReg)) { 741 BuildMI(MBB, I, DL, get(Hexagon::V6_vassign), DestReg). 742 addReg(SrcReg, KillFlag); 743 return; 744 } 745 if (Hexagon::VecDblRegsRegClass.contains(SrcReg, DestReg)) { 746 BuildMI(MBB, I, DL, get(Hexagon::V6_vcombine), DestReg) 747 .addReg(HRI.getSubReg(SrcReg, Hexagon::subreg_hireg), KillFlag) 748 .addReg(HRI.getSubReg(SrcReg, Hexagon::subreg_loreg), KillFlag); 749 return; 750 } 751 if (Hexagon::VecPredRegsRegClass.contains(SrcReg, DestReg)) { 752 BuildMI(MBB, I, DL, get(Hexagon::V6_pred_and), DestReg) 753 .addReg(SrcReg) 754 .addReg(SrcReg, KillFlag); 755 return; 756 } 757 if (Hexagon::VecPredRegsRegClass.contains(SrcReg) && 758 Hexagon::VectorRegsRegClass.contains(DestReg)) { 759 llvm_unreachable("Unimplemented pred to vec"); 760 return; 761 } 762 if (Hexagon::VecPredRegsRegClass.contains(DestReg) && 763 Hexagon::VectorRegsRegClass.contains(SrcReg)) { 764 llvm_unreachable("Unimplemented vec to pred"); 765 return; 766 } 767 if (Hexagon::VecPredRegs128BRegClass.contains(SrcReg, DestReg)) { 768 unsigned DstHi = HRI.getSubReg(DestReg, Hexagon::subreg_hireg); 769 BuildMI(MBB, I, DL, get(Hexagon::V6_pred_and), DstHi) 770 .addReg(HRI.getSubReg(SrcReg, Hexagon::subreg_hireg), KillFlag); 771 unsigned DstLo = HRI.getSubReg(DestReg, Hexagon::subreg_loreg); 772 BuildMI(MBB, I, DL, get(Hexagon::V6_pred_and), DstLo) 773 .addReg(HRI.getSubReg(SrcReg, Hexagon::subreg_loreg), KillFlag); 774 return; 775 } 776 777 #ifndef NDEBUG 778 // Show the invalid registers to ease debugging. 779 dbgs() << "Invalid registers for copy in BB#" << MBB.getNumber() 780 << ": " << PrintReg(DestReg, &HRI) 781 << " = " << PrintReg(SrcReg, &HRI) << '\n'; 782 #endif 783 llvm_unreachable("Unimplemented"); 784 } 785 786 787 void HexagonInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 788 MachineBasicBlock::iterator I, unsigned SrcReg, bool isKill, int FI, 789 const TargetRegisterClass *RC, const TargetRegisterInfo *TRI) const { 790 DebugLoc DL = MBB.findDebugLoc(I); 791 MachineFunction &MF = *MBB.getParent(); 792 MachineFrameInfo &MFI = *MF.getFrameInfo(); 793 unsigned Align = MFI.getObjectAlignment(FI); 794 unsigned KillFlag = getKillRegState(isKill); 795 796 MachineMemOperand *MMO = MF.getMachineMemOperand( 797 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore, 798 MFI.getObjectSize(FI), Align); 799 800 if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) { 801 BuildMI(MBB, I, DL, get(Hexagon::S2_storeri_io)) 802 .addFrameIndex(FI).addImm(0) 803 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 804 } else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) { 805 BuildMI(MBB, I, DL, get(Hexagon::S2_storerd_io)) 806 .addFrameIndex(FI).addImm(0) 807 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 808 } else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) { 809 BuildMI(MBB, I, DL, get(Hexagon::STriw_pred)) 810 .addFrameIndex(FI).addImm(0) 811 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 812 } else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) { 813 BuildMI(MBB, I, DL, get(Hexagon::STriw_mod)) 814 .addFrameIndex(FI).addImm(0) 815 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 816 } else if (Hexagon::VecPredRegs128BRegClass.hasSubClassEq(RC)) { 817 BuildMI(MBB, I, DL, get(Hexagon::STriq_pred_V6_128B)) 818 .addFrameIndex(FI).addImm(0) 819 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 820 } else if (Hexagon::VecPredRegsRegClass.hasSubClassEq(RC)) { 821 BuildMI(MBB, I, DL, get(Hexagon::STriq_pred_V6)) 822 .addFrameIndex(FI).addImm(0) 823 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 824 } else if (Hexagon::VectorRegs128BRegClass.hasSubClassEq(RC)) { 825 DEBUG(dbgs() << "++Generating 128B vector spill"); 826 BuildMI(MBB, I, DL, get(Hexagon::STriv_pseudo_V6_128B)) 827 .addFrameIndex(FI).addImm(0) 828 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 829 } else if (Hexagon::VectorRegsRegClass.hasSubClassEq(RC)) { 830 DEBUG(dbgs() << "++Generating vector spill"); 831 BuildMI(MBB, I, DL, get(Hexagon::STriv_pseudo_V6)) 832 .addFrameIndex(FI).addImm(0) 833 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 834 } else if (Hexagon::VecDblRegsRegClass.hasSubClassEq(RC)) { 835 DEBUG(dbgs() << "++Generating double vector spill"); 836 BuildMI(MBB, I, DL, get(Hexagon::STrivv_pseudo_V6)) 837 .addFrameIndex(FI).addImm(0) 838 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 839 } else if (Hexagon::VecDblRegs128BRegClass.hasSubClassEq(RC)) { 840 DEBUG(dbgs() << "++Generating 128B double vector spill"); 841 BuildMI(MBB, I, DL, get(Hexagon::STrivv_pseudo_V6_128B)) 842 .addFrameIndex(FI).addImm(0) 843 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 844 } else { 845 llvm_unreachable("Unimplemented"); 846 } 847 } 848 849 850 void HexagonInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, 851 MachineBasicBlock::iterator I, unsigned DestReg, int FI, 852 const TargetRegisterClass *RC, const TargetRegisterInfo *TRI) const { 853 DebugLoc DL = MBB.findDebugLoc(I); 854 MachineFunction &MF = *MBB.getParent(); 855 MachineFrameInfo &MFI = *MF.getFrameInfo(); 856 unsigned Align = MFI.getObjectAlignment(FI); 857 858 MachineMemOperand *MMO = MF.getMachineMemOperand( 859 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad, 860 MFI.getObjectSize(FI), Align); 861 862 if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) { 863 BuildMI(MBB, I, DL, get(Hexagon::L2_loadri_io), DestReg) 864 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 865 } else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) { 866 BuildMI(MBB, I, DL, get(Hexagon::L2_loadrd_io), DestReg) 867 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 868 } else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) { 869 BuildMI(MBB, I, DL, get(Hexagon::LDriw_pred), DestReg) 870 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 871 } else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) { 872 BuildMI(MBB, I, DL, get(Hexagon::LDriw_mod), DestReg) 873 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 874 } else if (Hexagon::VecPredRegs128BRegClass.hasSubClassEq(RC)) { 875 BuildMI(MBB, I, DL, get(Hexagon::LDriq_pred_V6_128B), DestReg) 876 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 877 } else if (Hexagon::VecPredRegsRegClass.hasSubClassEq(RC)) { 878 BuildMI(MBB, I, DL, get(Hexagon::LDriq_pred_V6), DestReg) 879 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 880 } else if (Hexagon::VecDblRegs128BRegClass.hasSubClassEq(RC)) { 881 DEBUG(dbgs() << "++Generating 128B double vector restore"); 882 BuildMI(MBB, I, DL, get(Hexagon::LDrivv_pseudo_V6_128B), DestReg) 883 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 884 } else if (Hexagon::VectorRegs128BRegClass.hasSubClassEq(RC)) { 885 DEBUG(dbgs() << "++Generating 128B vector restore"); 886 BuildMI(MBB, I, DL, get(Hexagon::LDriv_pseudo_V6_128B), DestReg) 887 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 888 } else if (Hexagon::VectorRegsRegClass.hasSubClassEq(RC)) { 889 DEBUG(dbgs() << "++Generating vector restore"); 890 BuildMI(MBB, I, DL, get(Hexagon::LDriv_pseudo_V6), DestReg) 891 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 892 } else if (Hexagon::VecDblRegsRegClass.hasSubClassEq(RC)) { 893 DEBUG(dbgs() << "++Generating double vector restore"); 894 BuildMI(MBB, I, DL, get(Hexagon::LDrivv_pseudo_V6), DestReg) 895 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 896 } else { 897 llvm_unreachable("Can't store this register to stack slot"); 898 } 899 } 900 901 902 /// expandPostRAPseudo - This function is called for all pseudo instructions 903 /// that remain after register allocation. Many pseudo instructions are 904 /// created to help register allocation. This is the place to convert them 905 /// into real instructions. The target can edit MI in place, or it can insert 906 /// new instructions and erase MI. The function should return true if 907 /// anything was changed. 908 bool HexagonInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) 909 const { 910 const HexagonRegisterInfo &HRI = getRegisterInfo(); 911 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo(); 912 MachineBasicBlock &MBB = *MI->getParent(); 913 DebugLoc DL = MI->getDebugLoc(); 914 unsigned Opc = MI->getOpcode(); 915 const unsigned VecOffset = 1; 916 bool Is128B = false; 917 918 switch (Opc) { 919 case TargetOpcode::COPY: { 920 MachineOperand &MD = MI->getOperand(0); 921 MachineOperand &MS = MI->getOperand(1); 922 if (MD.getReg() != MS.getReg() && !MS.isUndef()) { 923 copyPhysReg(MBB, MI, DL, MD.getReg(), MS.getReg(), MS.isKill()); 924 std::prev(MI)->copyImplicitOps(*MBB.getParent(), *MI); 925 } 926 MBB.erase(MI); 927 return true; 928 } 929 case Hexagon::ALIGNA: 930 BuildMI(MBB, MI, DL, get(Hexagon::A2_andir), MI->getOperand(0).getReg()) 931 .addReg(HRI.getFrameRegister()) 932 .addImm(-MI->getOperand(1).getImm()); 933 MBB.erase(MI); 934 return true; 935 case Hexagon::HEXAGON_V6_vassignp_128B: 936 case Hexagon::HEXAGON_V6_vassignp: { 937 unsigned SrcReg = MI->getOperand(1).getReg(); 938 unsigned DstReg = MI->getOperand(0).getReg(); 939 if (SrcReg != DstReg) 940 copyPhysReg(MBB, MI, DL, DstReg, SrcReg, MI->getOperand(1).isKill()); 941 MBB.erase(MI); 942 return true; 943 } 944 case Hexagon::HEXAGON_V6_lo_128B: 945 case Hexagon::HEXAGON_V6_lo: { 946 unsigned SrcReg = MI->getOperand(1).getReg(); 947 unsigned DstReg = MI->getOperand(0).getReg(); 948 unsigned SrcSubLo = HRI.getSubReg(SrcReg, Hexagon::subreg_loreg); 949 copyPhysReg(MBB, MI, DL, DstReg, SrcSubLo, MI->getOperand(1).isKill()); 950 MBB.erase(MI); 951 MRI.clearKillFlags(SrcSubLo); 952 return true; 953 } 954 case Hexagon::HEXAGON_V6_hi_128B: 955 case Hexagon::HEXAGON_V6_hi: { 956 unsigned SrcReg = MI->getOperand(1).getReg(); 957 unsigned DstReg = MI->getOperand(0).getReg(); 958 unsigned SrcSubHi = HRI.getSubReg(SrcReg, Hexagon::subreg_hireg); 959 copyPhysReg(MBB, MI, DL, DstReg, SrcSubHi, MI->getOperand(1).isKill()); 960 MBB.erase(MI); 961 MRI.clearKillFlags(SrcSubHi); 962 return true; 963 } 964 case Hexagon::STrivv_indexed_128B: 965 Is128B = true; 966 case Hexagon::STrivv_indexed: { 967 unsigned SrcReg = MI->getOperand(2).getReg(); 968 unsigned SrcSubHi = HRI.getSubReg(SrcReg, Hexagon::subreg_hireg); 969 unsigned SrcSubLo = HRI.getSubReg(SrcReg, Hexagon::subreg_loreg); 970 unsigned NewOpcd = Is128B ? Hexagon::V6_vS32b_ai_128B 971 : Hexagon::V6_vS32b_ai; 972 unsigned Offset = Is128B ? VecOffset << 7 : VecOffset << 6; 973 MachineInstr *MI1New = BuildMI(MBB, MI, DL, get(NewOpcd)) 974 .addOperand(MI->getOperand(0)) 975 .addImm(MI->getOperand(1).getImm()) 976 .addReg(SrcSubLo) 977 .setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 978 MI1New->getOperand(0).setIsKill(false); 979 BuildMI(MBB, MI, DL, get(NewOpcd)) 980 .addOperand(MI->getOperand(0)) 981 // The Vectors are indexed in multiples of vector size. 982 .addImm(MI->getOperand(1).getImm()+Offset) 983 .addReg(SrcSubHi) 984 .setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 985 MBB.erase(MI); 986 return true; 987 } 988 case Hexagon::LDrivv_pseudo_V6_128B: 989 case Hexagon::LDrivv_indexed_128B: 990 Is128B = true; 991 case Hexagon::LDrivv_pseudo_V6: 992 case Hexagon::LDrivv_indexed: { 993 unsigned NewOpcd = Is128B ? Hexagon::V6_vL32b_ai_128B 994 : Hexagon::V6_vL32b_ai; 995 unsigned DstReg = MI->getOperand(0).getReg(); 996 unsigned Offset = Is128B ? VecOffset << 7 : VecOffset << 6; 997 MachineInstr *MI1New = 998 BuildMI(MBB, MI, DL, get(NewOpcd), 999 HRI.getSubReg(DstReg, Hexagon::subreg_loreg)) 1000 .addOperand(MI->getOperand(1)) 1001 .addImm(MI->getOperand(2).getImm()); 1002 MI1New->getOperand(1).setIsKill(false); 1003 BuildMI(MBB, MI, DL, get(NewOpcd), 1004 HRI.getSubReg(DstReg, Hexagon::subreg_hireg)) 1005 .addOperand(MI->getOperand(1)) 1006 // The Vectors are indexed in multiples of vector size. 1007 .addImm(MI->getOperand(2).getImm() + Offset) 1008 .setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 1009 MBB.erase(MI); 1010 return true; 1011 } 1012 case Hexagon::LDriv_pseudo_V6_128B: 1013 Is128B = true; 1014 case Hexagon::LDriv_pseudo_V6: { 1015 unsigned DstReg = MI->getOperand(0).getReg(); 1016 unsigned NewOpc = Is128B ? Hexagon::V6_vL32b_ai_128B 1017 : Hexagon::V6_vL32b_ai; 1018 int32_t Off = MI->getOperand(2).getImm(); 1019 BuildMI(MBB, MI, DL, get(NewOpc), DstReg) 1020 .addOperand(MI->getOperand(1)) 1021 .addImm(Off) 1022 .setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 1023 MBB.erase(MI); 1024 return true; 1025 } 1026 case Hexagon::STriv_pseudo_V6_128B: 1027 Is128B = true; 1028 case Hexagon::STriv_pseudo_V6: { 1029 unsigned NewOpc = Is128B ? Hexagon::V6_vS32b_ai_128B 1030 : Hexagon::V6_vS32b_ai; 1031 int32_t Off = MI->getOperand(1).getImm(); 1032 BuildMI(MBB, MI, DL, get(NewOpc)) 1033 .addOperand(MI->getOperand(0)) 1034 .addImm(Off) 1035 .addOperand(MI->getOperand(2)) 1036 .setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 1037 MBB.erase(MI); 1038 return true; 1039 } 1040 case Hexagon::TFR_PdTrue: { 1041 unsigned Reg = MI->getOperand(0).getReg(); 1042 BuildMI(MBB, MI, DL, get(Hexagon::C2_orn), Reg) 1043 .addReg(Reg, RegState::Undef) 1044 .addReg(Reg, RegState::Undef); 1045 MBB.erase(MI); 1046 return true; 1047 } 1048 case Hexagon::TFR_PdFalse: { 1049 unsigned Reg = MI->getOperand(0).getReg(); 1050 BuildMI(MBB, MI, DL, get(Hexagon::C2_andn), Reg) 1051 .addReg(Reg, RegState::Undef) 1052 .addReg(Reg, RegState::Undef); 1053 MBB.erase(MI); 1054 return true; 1055 } 1056 case Hexagon::VMULW: { 1057 // Expand a 64-bit vector multiply into 2 32-bit scalar multiplies. 1058 unsigned DstReg = MI->getOperand(0).getReg(); 1059 unsigned Src1Reg = MI->getOperand(1).getReg(); 1060 unsigned Src2Reg = MI->getOperand(2).getReg(); 1061 unsigned Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::subreg_hireg); 1062 unsigned Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::subreg_loreg); 1063 unsigned Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::subreg_hireg); 1064 unsigned Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::subreg_loreg); 1065 BuildMI(MBB, MI, MI->getDebugLoc(), get(Hexagon::M2_mpyi), 1066 HRI.getSubReg(DstReg, Hexagon::subreg_hireg)).addReg(Src1SubHi) 1067 .addReg(Src2SubHi); 1068 BuildMI(MBB, MI, MI->getDebugLoc(), get(Hexagon::M2_mpyi), 1069 HRI.getSubReg(DstReg, Hexagon::subreg_loreg)).addReg(Src1SubLo) 1070 .addReg(Src2SubLo); 1071 MBB.erase(MI); 1072 MRI.clearKillFlags(Src1SubHi); 1073 MRI.clearKillFlags(Src1SubLo); 1074 MRI.clearKillFlags(Src2SubHi); 1075 MRI.clearKillFlags(Src2SubLo); 1076 return true; 1077 } 1078 case Hexagon::VMULW_ACC: { 1079 // Expand 64-bit vector multiply with addition into 2 scalar multiplies. 1080 unsigned DstReg = MI->getOperand(0).getReg(); 1081 unsigned Src1Reg = MI->getOperand(1).getReg(); 1082 unsigned Src2Reg = MI->getOperand(2).getReg(); 1083 unsigned Src3Reg = MI->getOperand(3).getReg(); 1084 unsigned Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::subreg_hireg); 1085 unsigned Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::subreg_loreg); 1086 unsigned Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::subreg_hireg); 1087 unsigned Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::subreg_loreg); 1088 unsigned Src3SubHi = HRI.getSubReg(Src3Reg, Hexagon::subreg_hireg); 1089 unsigned Src3SubLo = HRI.getSubReg(Src3Reg, Hexagon::subreg_loreg); 1090 BuildMI(MBB, MI, MI->getDebugLoc(), get(Hexagon::M2_maci), 1091 HRI.getSubReg(DstReg, Hexagon::subreg_hireg)).addReg(Src1SubHi) 1092 .addReg(Src2SubHi).addReg(Src3SubHi); 1093 BuildMI(MBB, MI, MI->getDebugLoc(), get(Hexagon::M2_maci), 1094 HRI.getSubReg(DstReg, Hexagon::subreg_loreg)).addReg(Src1SubLo) 1095 .addReg(Src2SubLo).addReg(Src3SubLo); 1096 MBB.erase(MI); 1097 MRI.clearKillFlags(Src1SubHi); 1098 MRI.clearKillFlags(Src1SubLo); 1099 MRI.clearKillFlags(Src2SubHi); 1100 MRI.clearKillFlags(Src2SubLo); 1101 MRI.clearKillFlags(Src3SubHi); 1102 MRI.clearKillFlags(Src3SubLo); 1103 return true; 1104 } 1105 case Hexagon::Insert4: { 1106 unsigned DstReg = MI->getOperand(0).getReg(); 1107 unsigned Src1Reg = MI->getOperand(1).getReg(); 1108 unsigned Src2Reg = MI->getOperand(2).getReg(); 1109 unsigned Src3Reg = MI->getOperand(3).getReg(); 1110 unsigned Src4Reg = MI->getOperand(4).getReg(); 1111 unsigned Src1RegIsKill = getKillRegState(MI->getOperand(1).isKill()); 1112 unsigned Src2RegIsKill = getKillRegState(MI->getOperand(2).isKill()); 1113 unsigned Src3RegIsKill = getKillRegState(MI->getOperand(3).isKill()); 1114 unsigned Src4RegIsKill = getKillRegState(MI->getOperand(4).isKill()); 1115 unsigned DstSubHi = HRI.getSubReg(DstReg, Hexagon::subreg_hireg); 1116 unsigned DstSubLo = HRI.getSubReg(DstReg, Hexagon::subreg_loreg); 1117 BuildMI(MBB, MI, MI->getDebugLoc(), get(Hexagon::S2_insert), 1118 HRI.getSubReg(DstReg, Hexagon::subreg_loreg)).addReg(DstSubLo) 1119 .addReg(Src1Reg, Src1RegIsKill).addImm(16).addImm(0); 1120 BuildMI(MBB, MI, MI->getDebugLoc(), get(Hexagon::S2_insert), 1121 HRI.getSubReg(DstReg, Hexagon::subreg_loreg)).addReg(DstSubLo) 1122 .addReg(Src2Reg, Src2RegIsKill).addImm(16).addImm(16); 1123 BuildMI(MBB, MI, MI->getDebugLoc(), get(Hexagon::S2_insert), 1124 HRI.getSubReg(DstReg, Hexagon::subreg_hireg)).addReg(DstSubHi) 1125 .addReg(Src3Reg, Src3RegIsKill).addImm(16).addImm(0); 1126 BuildMI(MBB, MI, MI->getDebugLoc(), get(Hexagon::S2_insert), 1127 HRI.getSubReg(DstReg, Hexagon::subreg_hireg)).addReg(DstSubHi) 1128 .addReg(Src4Reg, Src4RegIsKill).addImm(16).addImm(16); 1129 MBB.erase(MI); 1130 MRI.clearKillFlags(DstReg); 1131 MRI.clearKillFlags(DstSubHi); 1132 MRI.clearKillFlags(DstSubLo); 1133 return true; 1134 } 1135 case Hexagon::MUX64_rr: { 1136 const MachineOperand &Op0 = MI->getOperand(0); 1137 const MachineOperand &Op1 = MI->getOperand(1); 1138 const MachineOperand &Op2 = MI->getOperand(2); 1139 const MachineOperand &Op3 = MI->getOperand(3); 1140 unsigned Rd = Op0.getReg(); 1141 unsigned Pu = Op1.getReg(); 1142 unsigned Rs = Op2.getReg(); 1143 unsigned Rt = Op3.getReg(); 1144 DebugLoc DL = MI->getDebugLoc(); 1145 unsigned K1 = getKillRegState(Op1.isKill()); 1146 unsigned K2 = getKillRegState(Op2.isKill()); 1147 unsigned K3 = getKillRegState(Op3.isKill()); 1148 if (Rd != Rs) 1149 BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpt), Rd) 1150 .addReg(Pu, (Rd == Rt) ? K1 : 0) 1151 .addReg(Rs, K2); 1152 if (Rd != Rt) 1153 BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpf), Rd) 1154 .addReg(Pu, K1) 1155 .addReg(Rt, K3); 1156 MBB.erase(MI); 1157 return true; 1158 } 1159 case Hexagon::VSelectPseudo_V6: { 1160 const MachineOperand &Op0 = MI->getOperand(0); 1161 const MachineOperand &Op1 = MI->getOperand(1); 1162 const MachineOperand &Op2 = MI->getOperand(2); 1163 const MachineOperand &Op3 = MI->getOperand(3); 1164 BuildMI(MBB, MI, DL, get(Hexagon::V6_vcmov)) 1165 .addOperand(Op0) 1166 .addOperand(Op1) 1167 .addOperand(Op2); 1168 BuildMI(MBB, MI, DL, get(Hexagon::V6_vncmov)) 1169 .addOperand(Op0) 1170 .addOperand(Op1) 1171 .addOperand(Op3); 1172 MBB.erase(MI); 1173 return true; 1174 } 1175 case Hexagon::VSelectDblPseudo_V6: { 1176 MachineOperand &Op0 = MI->getOperand(0); 1177 MachineOperand &Op1 = MI->getOperand(1); 1178 MachineOperand &Op2 = MI->getOperand(2); 1179 MachineOperand &Op3 = MI->getOperand(3); 1180 unsigned SrcLo = HRI.getSubReg(Op2.getReg(), Hexagon::subreg_loreg); 1181 unsigned SrcHi = HRI.getSubReg(Op2.getReg(), Hexagon::subreg_hireg); 1182 BuildMI(MBB, MI, DL, get(Hexagon::V6_vccombine)) 1183 .addOperand(Op0) 1184 .addOperand(Op1) 1185 .addReg(SrcHi) 1186 .addReg(SrcLo); 1187 SrcLo = HRI.getSubReg(Op3.getReg(), Hexagon::subreg_loreg); 1188 SrcHi = HRI.getSubReg(Op3.getReg(), Hexagon::subreg_hireg); 1189 BuildMI(MBB, MI, DL, get(Hexagon::V6_vnccombine)) 1190 .addOperand(Op0) 1191 .addOperand(Op1) 1192 .addReg(SrcHi) 1193 .addReg(SrcLo); 1194 MBB.erase(MI); 1195 return true; 1196 } 1197 case Hexagon::TCRETURNi: 1198 MI->setDesc(get(Hexagon::J2_jump)); 1199 return true; 1200 case Hexagon::TCRETURNr: 1201 MI->setDesc(get(Hexagon::J2_jumpr)); 1202 return true; 1203 case Hexagon::TFRI_f: 1204 case Hexagon::TFRI_cPt_f: 1205 case Hexagon::TFRI_cNotPt_f: { 1206 unsigned Opx = (Opc == Hexagon::TFRI_f) ? 1 : 2; 1207 APFloat FVal = MI->getOperand(Opx).getFPImm()->getValueAPF(); 1208 APInt IVal = FVal.bitcastToAPInt(); 1209 MI->RemoveOperand(Opx); 1210 unsigned NewOpc = (Opc == Hexagon::TFRI_f) ? Hexagon::A2_tfrsi : 1211 (Opc == Hexagon::TFRI_cPt_f) ? Hexagon::C2_cmoveit : 1212 Hexagon::C2_cmoveif; 1213 MI->setDesc(get(NewOpc)); 1214 MI->addOperand(MachineOperand::CreateImm(IVal.getZExtValue())); 1215 return true; 1216 } 1217 } 1218 1219 return false; 1220 } 1221 1222 1223 // We indicate that we want to reverse the branch by 1224 // inserting the reversed branching opcode. 1225 bool HexagonInstrInfo::ReverseBranchCondition( 1226 SmallVectorImpl<MachineOperand> &Cond) const { 1227 if (Cond.empty()) 1228 return true; 1229 assert(Cond[0].isImm() && "First entry in the cond vector not imm-val"); 1230 unsigned opcode = Cond[0].getImm(); 1231 //unsigned temp; 1232 assert(get(opcode).isBranch() && "Should be a branching condition."); 1233 if (isEndLoopN(opcode)) 1234 return true; 1235 unsigned NewOpcode = getInvertedPredicatedOpcode(opcode); 1236 Cond[0].setImm(NewOpcode); 1237 return false; 1238 } 1239 1240 1241 void HexagonInstrInfo::insertNoop(MachineBasicBlock &MBB, 1242 MachineBasicBlock::iterator MI) const { 1243 DebugLoc DL; 1244 BuildMI(MBB, MI, DL, get(Hexagon::A2_nop)); 1245 } 1246 1247 1248 // Returns true if an instruction is predicated irrespective of the predicate 1249 // sense. For example, all of the following will return true. 1250 // if (p0) R1 = add(R2, R3) 1251 // if (!p0) R1 = add(R2, R3) 1252 // if (p0.new) R1 = add(R2, R3) 1253 // if (!p0.new) R1 = add(R2, R3) 1254 // Note: New-value stores are not included here as in the current 1255 // implementation, we don't need to check their predicate sense. 1256 bool HexagonInstrInfo::isPredicated(const MachineInstr &MI) const { 1257 const uint64_t F = MI.getDesc().TSFlags; 1258 return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask; 1259 } 1260 1261 1262 bool HexagonInstrInfo::PredicateInstruction( 1263 MachineInstr &MI, ArrayRef<MachineOperand> Cond) const { 1264 if (Cond.empty() || isNewValueJump(Cond[0].getImm()) || 1265 isEndLoopN(Cond[0].getImm())) { 1266 DEBUG(dbgs() << "\nCannot predicate:"; MI.dump();); 1267 return false; 1268 } 1269 int Opc = MI.getOpcode(); 1270 assert (isPredicable(MI) && "Expected predicable instruction"); 1271 bool invertJump = predOpcodeHasNot(Cond); 1272 1273 // We have to predicate MI "in place", i.e. after this function returns, 1274 // MI will need to be transformed into a predicated form. To avoid com- 1275 // plicated manipulations with the operands (handling tied operands, 1276 // etc.), build a new temporary instruction, then overwrite MI with it. 1277 1278 MachineBasicBlock &B = *MI.getParent(); 1279 DebugLoc DL = MI.getDebugLoc(); 1280 unsigned PredOpc = getCondOpcode(Opc, invertJump); 1281 MachineInstrBuilder T = BuildMI(B, MI, DL, get(PredOpc)); 1282 unsigned NOp = 0, NumOps = MI.getNumOperands(); 1283 while (NOp < NumOps) { 1284 MachineOperand &Op = MI.getOperand(NOp); 1285 if (!Op.isReg() || !Op.isDef() || Op.isImplicit()) 1286 break; 1287 T.addOperand(Op); 1288 NOp++; 1289 } 1290 1291 unsigned PredReg, PredRegPos, PredRegFlags; 1292 bool GotPredReg = getPredReg(Cond, PredReg, PredRegPos, PredRegFlags); 1293 (void)GotPredReg; 1294 assert(GotPredReg); 1295 T.addReg(PredReg, PredRegFlags); 1296 while (NOp < NumOps) 1297 T.addOperand(MI.getOperand(NOp++)); 1298 1299 MI.setDesc(get(PredOpc)); 1300 while (unsigned n = MI.getNumOperands()) 1301 MI.RemoveOperand(n-1); 1302 for (unsigned i = 0, n = T->getNumOperands(); i < n; ++i) 1303 MI.addOperand(T->getOperand(i)); 1304 1305 MachineBasicBlock::instr_iterator TI = T->getIterator(); 1306 B.erase(TI); 1307 1308 MachineRegisterInfo &MRI = B.getParent()->getRegInfo(); 1309 MRI.clearKillFlags(PredReg); 1310 return true; 1311 } 1312 1313 1314 bool HexagonInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1, 1315 ArrayRef<MachineOperand> Pred2) const { 1316 // TODO: Fix this 1317 return false; 1318 } 1319 1320 1321 bool HexagonInstrInfo::DefinesPredicate( 1322 MachineInstr &MI, std::vector<MachineOperand> &Pred) const { 1323 auto &HRI = getRegisterInfo(); 1324 for (unsigned oper = 0; oper < MI.getNumOperands(); ++oper) { 1325 MachineOperand MO = MI.getOperand(oper); 1326 if (MO.isReg() && MO.isDef()) { 1327 const TargetRegisterClass* RC = HRI.getMinimalPhysRegClass(MO.getReg()); 1328 if (RC == &Hexagon::PredRegsRegClass) { 1329 Pred.push_back(MO); 1330 return true; 1331 } 1332 } 1333 } 1334 return false; 1335 } 1336 1337 1338 bool HexagonInstrInfo::isPredicable(MachineInstr &MI) const { 1339 return MI.getDesc().isPredicable(); 1340 } 1341 1342 1343 bool HexagonInstrInfo::isSchedulingBoundary(const MachineInstr *MI, 1344 const MachineBasicBlock *MBB, const MachineFunction &MF) const { 1345 // Debug info is never a scheduling boundary. It's necessary to be explicit 1346 // due to the special treatment of IT instructions below, otherwise a 1347 // dbg_value followed by an IT will result in the IT instruction being 1348 // considered a scheduling hazard, which is wrong. It should be the actual 1349 // instruction preceding the dbg_value instruction(s), just like it is 1350 // when debug info is not present. 1351 if (MI->isDebugValue()) 1352 return false; 1353 1354 // Throwing call is a boundary. 1355 if (MI->isCall()) { 1356 // If any of the block's successors is a landing pad, this could be a 1357 // throwing call. 1358 for (auto I : MBB->successors()) 1359 if (I->isEHPad()) 1360 return true; 1361 } 1362 1363 // Don't mess around with no return calls. 1364 if (MI->getOpcode() == Hexagon::CALLv3nr) 1365 return true; 1366 1367 // Terminators and labels can't be scheduled around. 1368 if (MI->getDesc().isTerminator() || MI->isPosition()) 1369 return true; 1370 1371 if (MI->isInlineAsm() && !ScheduleInlineAsm) 1372 return true; 1373 1374 return false; 1375 } 1376 1377 1378 /// Measure the specified inline asm to determine an approximation of its 1379 /// length. 1380 /// Comments (which run till the next SeparatorString or newline) do not 1381 /// count as an instruction. 1382 /// Any other non-whitespace text is considered an instruction, with 1383 /// multiple instructions separated by SeparatorString or newlines. 1384 /// Variable-length instructions are not handled here; this function 1385 /// may be overloaded in the target code to do that. 1386 /// Hexagon counts the number of ##'s and adjust for that many 1387 /// constant exenders. 1388 unsigned HexagonInstrInfo::getInlineAsmLength(const char *Str, 1389 const MCAsmInfo &MAI) const { 1390 StringRef AStr(Str); 1391 // Count the number of instructions in the asm. 1392 bool atInsnStart = true; 1393 unsigned Length = 0; 1394 for (; *Str; ++Str) { 1395 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(), 1396 strlen(MAI.getSeparatorString())) == 0) 1397 atInsnStart = true; 1398 if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) { 1399 Length += MAI.getMaxInstLength(); 1400 atInsnStart = false; 1401 } 1402 if (atInsnStart && strncmp(Str, MAI.getCommentString(), 1403 strlen(MAI.getCommentString())) == 0) 1404 atInsnStart = false; 1405 } 1406 1407 // Add to size number of constant extenders seen * 4. 1408 StringRef Occ("##"); 1409 Length += AStr.count(Occ)*4; 1410 return Length; 1411 } 1412 1413 1414 ScheduleHazardRecognizer* 1415 HexagonInstrInfo::CreateTargetPostRAHazardRecognizer( 1416 const InstrItineraryData *II, const ScheduleDAG *DAG) const { 1417 return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG); 1418 } 1419 1420 1421 /// \brief For a comparison instruction, return the source registers in 1422 /// \p SrcReg and \p SrcReg2 if having two register operands, and the value it 1423 /// compares against in CmpValue. Return true if the comparison instruction 1424 /// can be analyzed. 1425 bool HexagonInstrInfo::analyzeCompare(const MachineInstr *MI, 1426 unsigned &SrcReg, unsigned &SrcReg2, int &Mask, int &Value) const { 1427 unsigned Opc = MI->getOpcode(); 1428 1429 // Set mask and the first source register. 1430 switch (Opc) { 1431 case Hexagon::C2_cmpeq: 1432 case Hexagon::C2_cmpeqp: 1433 case Hexagon::C2_cmpgt: 1434 case Hexagon::C2_cmpgtp: 1435 case Hexagon::C2_cmpgtu: 1436 case Hexagon::C2_cmpgtup: 1437 case Hexagon::C4_cmpneq: 1438 case Hexagon::C4_cmplte: 1439 case Hexagon::C4_cmplteu: 1440 case Hexagon::C2_cmpeqi: 1441 case Hexagon::C2_cmpgti: 1442 case Hexagon::C2_cmpgtui: 1443 case Hexagon::C4_cmpneqi: 1444 case Hexagon::C4_cmplteui: 1445 case Hexagon::C4_cmpltei: 1446 SrcReg = MI->getOperand(1).getReg(); 1447 Mask = ~0; 1448 break; 1449 case Hexagon::A4_cmpbeq: 1450 case Hexagon::A4_cmpbgt: 1451 case Hexagon::A4_cmpbgtu: 1452 case Hexagon::A4_cmpbeqi: 1453 case Hexagon::A4_cmpbgti: 1454 case Hexagon::A4_cmpbgtui: 1455 SrcReg = MI->getOperand(1).getReg(); 1456 Mask = 0xFF; 1457 break; 1458 case Hexagon::A4_cmpheq: 1459 case Hexagon::A4_cmphgt: 1460 case Hexagon::A4_cmphgtu: 1461 case Hexagon::A4_cmpheqi: 1462 case Hexagon::A4_cmphgti: 1463 case Hexagon::A4_cmphgtui: 1464 SrcReg = MI->getOperand(1).getReg(); 1465 Mask = 0xFFFF; 1466 break; 1467 } 1468 1469 // Set the value/second source register. 1470 switch (Opc) { 1471 case Hexagon::C2_cmpeq: 1472 case Hexagon::C2_cmpeqp: 1473 case Hexagon::C2_cmpgt: 1474 case Hexagon::C2_cmpgtp: 1475 case Hexagon::C2_cmpgtu: 1476 case Hexagon::C2_cmpgtup: 1477 case Hexagon::A4_cmpbeq: 1478 case Hexagon::A4_cmpbgt: 1479 case Hexagon::A4_cmpbgtu: 1480 case Hexagon::A4_cmpheq: 1481 case Hexagon::A4_cmphgt: 1482 case Hexagon::A4_cmphgtu: 1483 case Hexagon::C4_cmpneq: 1484 case Hexagon::C4_cmplte: 1485 case Hexagon::C4_cmplteu: 1486 SrcReg2 = MI->getOperand(2).getReg(); 1487 return true; 1488 1489 case Hexagon::C2_cmpeqi: 1490 case Hexagon::C2_cmpgtui: 1491 case Hexagon::C2_cmpgti: 1492 case Hexagon::C4_cmpneqi: 1493 case Hexagon::C4_cmplteui: 1494 case Hexagon::C4_cmpltei: 1495 case Hexagon::A4_cmpbeqi: 1496 case Hexagon::A4_cmpbgti: 1497 case Hexagon::A4_cmpbgtui: 1498 case Hexagon::A4_cmpheqi: 1499 case Hexagon::A4_cmphgti: 1500 case Hexagon::A4_cmphgtui: 1501 SrcReg2 = 0; 1502 Value = MI->getOperand(2).getImm(); 1503 return true; 1504 } 1505 1506 return false; 1507 } 1508 1509 1510 unsigned HexagonInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 1511 const MachineInstr *MI, unsigned *PredCost) const { 1512 return getInstrTimingClassLatency(ItinData, MI); 1513 } 1514 1515 1516 DFAPacketizer *HexagonInstrInfo::CreateTargetScheduleState( 1517 const TargetSubtargetInfo &STI) const { 1518 const InstrItineraryData *II = STI.getInstrItineraryData(); 1519 return static_cast<const HexagonSubtarget&>(STI).createDFAPacketizer(II); 1520 } 1521 1522 1523 // Inspired by this pair: 1524 // %R13<def> = L2_loadri_io %R29, 136; mem:LD4[FixedStack0] 1525 // S2_storeri_io %R29, 132, %R1<kill>; flags: mem:ST4[FixedStack1] 1526 // Currently AA considers the addresses in these instructions to be aliasing. 1527 bool HexagonInstrInfo::areMemAccessesTriviallyDisjoint(MachineInstr *MIa, 1528 MachineInstr *MIb, AliasAnalysis *AA) const { 1529 int OffsetA = 0, OffsetB = 0; 1530 unsigned SizeA = 0, SizeB = 0; 1531 1532 if (MIa->hasUnmodeledSideEffects() || MIb->hasUnmodeledSideEffects() || 1533 MIa->hasOrderedMemoryRef() || MIb->hasOrderedMemoryRef()) 1534 return false; 1535 1536 // Instructions that are pure loads, not loads and stores like memops are not 1537 // dependent. 1538 if (MIa->mayLoad() && !isMemOp(MIa) && MIb->mayLoad() && !isMemOp(MIb)) 1539 return true; 1540 1541 // Get base, offset, and access size in MIa. 1542 unsigned BaseRegA = getBaseAndOffset(MIa, OffsetA, SizeA); 1543 if (!BaseRegA || !SizeA) 1544 return false; 1545 1546 // Get base, offset, and access size in MIb. 1547 unsigned BaseRegB = getBaseAndOffset(MIb, OffsetB, SizeB); 1548 if (!BaseRegB || !SizeB) 1549 return false; 1550 1551 if (BaseRegA != BaseRegB) 1552 return false; 1553 1554 // This is a mem access with the same base register and known offsets from it. 1555 // Reason about it. 1556 if (OffsetA > OffsetB) { 1557 uint64_t offDiff = (uint64_t)((int64_t)OffsetA - (int64_t)OffsetB); 1558 return (SizeB <= offDiff); 1559 } else if (OffsetA < OffsetB) { 1560 uint64_t offDiff = (uint64_t)((int64_t)OffsetB - (int64_t)OffsetA); 1561 return (SizeA <= offDiff); 1562 } 1563 1564 return false; 1565 } 1566 1567 1568 unsigned HexagonInstrInfo::createVR(MachineFunction* MF, MVT VT) const { 1569 MachineRegisterInfo &MRI = MF->getRegInfo(); 1570 const TargetRegisterClass *TRC; 1571 if (VT == MVT::i1) { 1572 TRC = &Hexagon::PredRegsRegClass; 1573 } else if (VT == MVT::i32 || VT == MVT::f32) { 1574 TRC = &Hexagon::IntRegsRegClass; 1575 } else if (VT == MVT::i64 || VT == MVT::f64) { 1576 TRC = &Hexagon::DoubleRegsRegClass; 1577 } else { 1578 llvm_unreachable("Cannot handle this register class"); 1579 } 1580 1581 unsigned NewReg = MRI.createVirtualRegister(TRC); 1582 return NewReg; 1583 } 1584 1585 1586 bool HexagonInstrInfo::isAbsoluteSet(const MachineInstr* MI) const { 1587 return (getAddrMode(MI) == HexagonII::AbsoluteSet); 1588 } 1589 1590 1591 bool HexagonInstrInfo::isAccumulator(const MachineInstr *MI) const { 1592 const uint64_t F = MI->getDesc().TSFlags; 1593 return((F >> HexagonII::AccumulatorPos) & HexagonII::AccumulatorMask); 1594 } 1595 1596 1597 bool HexagonInstrInfo::isComplex(const MachineInstr *MI) const { 1598 const MachineFunction *MF = MI->getParent()->getParent(); 1599 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); 1600 const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; 1601 1602 if (!(isTC1(MI)) 1603 && !(QII->isTC2Early(MI)) 1604 && !(MI->getDesc().mayLoad()) 1605 && !(MI->getDesc().mayStore()) 1606 && (MI->getDesc().getOpcode() != Hexagon::S2_allocframe) 1607 && (MI->getDesc().getOpcode() != Hexagon::L2_deallocframe) 1608 && !(QII->isMemOp(MI)) 1609 && !(MI->isBranch()) 1610 && !(MI->isReturn()) 1611 && !MI->isCall()) 1612 return true; 1613 1614 return false; 1615 } 1616 1617 1618 // Return true if the instruction is a compund branch instruction. 1619 bool HexagonInstrInfo::isCompoundBranchInstr(const MachineInstr *MI) const { 1620 return (getType(MI) == HexagonII::TypeCOMPOUND && MI->isBranch()); 1621 } 1622 1623 1624 bool HexagonInstrInfo::isCondInst(const MachineInstr *MI) const { 1625 return (MI->isBranch() && isPredicated(*MI)) || 1626 isConditionalTransfer(MI) || 1627 isConditionalALU32(MI) || 1628 isConditionalLoad(MI) || 1629 // Predicated stores which don't have a .new on any operands. 1630 (MI->mayStore() && isPredicated(*MI) && !isNewValueStore(MI) && 1631 !isPredicatedNew(*MI)); 1632 } 1633 1634 1635 bool HexagonInstrInfo::isConditionalALU32(const MachineInstr* MI) const { 1636 switch (MI->getOpcode()) { 1637 case Hexagon::A2_paddf: 1638 case Hexagon::A2_paddfnew: 1639 case Hexagon::A2_paddif: 1640 case Hexagon::A2_paddifnew: 1641 case Hexagon::A2_paddit: 1642 case Hexagon::A2_padditnew: 1643 case Hexagon::A2_paddt: 1644 case Hexagon::A2_paddtnew: 1645 case Hexagon::A2_pandf: 1646 case Hexagon::A2_pandfnew: 1647 case Hexagon::A2_pandt: 1648 case Hexagon::A2_pandtnew: 1649 case Hexagon::A2_porf: 1650 case Hexagon::A2_porfnew: 1651 case Hexagon::A2_port: 1652 case Hexagon::A2_portnew: 1653 case Hexagon::A2_psubf: 1654 case Hexagon::A2_psubfnew: 1655 case Hexagon::A2_psubt: 1656 case Hexagon::A2_psubtnew: 1657 case Hexagon::A2_pxorf: 1658 case Hexagon::A2_pxorfnew: 1659 case Hexagon::A2_pxort: 1660 case Hexagon::A2_pxortnew: 1661 case Hexagon::A4_paslhf: 1662 case Hexagon::A4_paslhfnew: 1663 case Hexagon::A4_paslht: 1664 case Hexagon::A4_paslhtnew: 1665 case Hexagon::A4_pasrhf: 1666 case Hexagon::A4_pasrhfnew: 1667 case Hexagon::A4_pasrht: 1668 case Hexagon::A4_pasrhtnew: 1669 case Hexagon::A4_psxtbf: 1670 case Hexagon::A4_psxtbfnew: 1671 case Hexagon::A4_psxtbt: 1672 case Hexagon::A4_psxtbtnew: 1673 case Hexagon::A4_psxthf: 1674 case Hexagon::A4_psxthfnew: 1675 case Hexagon::A4_psxtht: 1676 case Hexagon::A4_psxthtnew: 1677 case Hexagon::A4_pzxtbf: 1678 case Hexagon::A4_pzxtbfnew: 1679 case Hexagon::A4_pzxtbt: 1680 case Hexagon::A4_pzxtbtnew: 1681 case Hexagon::A4_pzxthf: 1682 case Hexagon::A4_pzxthfnew: 1683 case Hexagon::A4_pzxtht: 1684 case Hexagon::A4_pzxthtnew: 1685 case Hexagon::C2_ccombinewf: 1686 case Hexagon::C2_ccombinewt: 1687 return true; 1688 } 1689 return false; 1690 } 1691 1692 1693 // FIXME - Function name and it's functionality don't match. 1694 // It should be renamed to hasPredNewOpcode() 1695 bool HexagonInstrInfo::isConditionalLoad(const MachineInstr* MI) const { 1696 if (!MI->getDesc().mayLoad() || !isPredicated(*MI)) 1697 return false; 1698 1699 int PNewOpcode = Hexagon::getPredNewOpcode(MI->getOpcode()); 1700 // Instruction with valid predicated-new opcode can be promoted to .new. 1701 return PNewOpcode >= 0; 1702 } 1703 1704 1705 // Returns true if an instruction is a conditional store. 1706 // 1707 // Note: It doesn't include conditional new-value stores as they can't be 1708 // converted to .new predicate. 1709 bool HexagonInstrInfo::isConditionalStore(const MachineInstr* MI) const { 1710 switch (MI->getOpcode()) { 1711 default: return false; 1712 case Hexagon::S4_storeirbt_io: 1713 case Hexagon::S4_storeirbf_io: 1714 case Hexagon::S4_pstorerbt_rr: 1715 case Hexagon::S4_pstorerbf_rr: 1716 case Hexagon::S2_pstorerbt_io: 1717 case Hexagon::S2_pstorerbf_io: 1718 case Hexagon::S2_pstorerbt_pi: 1719 case Hexagon::S2_pstorerbf_pi: 1720 case Hexagon::S2_pstorerdt_io: 1721 case Hexagon::S2_pstorerdf_io: 1722 case Hexagon::S4_pstorerdt_rr: 1723 case Hexagon::S4_pstorerdf_rr: 1724 case Hexagon::S2_pstorerdt_pi: 1725 case Hexagon::S2_pstorerdf_pi: 1726 case Hexagon::S2_pstorerht_io: 1727 case Hexagon::S2_pstorerhf_io: 1728 case Hexagon::S4_storeirht_io: 1729 case Hexagon::S4_storeirhf_io: 1730 case Hexagon::S4_pstorerht_rr: 1731 case Hexagon::S4_pstorerhf_rr: 1732 case Hexagon::S2_pstorerht_pi: 1733 case Hexagon::S2_pstorerhf_pi: 1734 case Hexagon::S2_pstorerit_io: 1735 case Hexagon::S2_pstorerif_io: 1736 case Hexagon::S4_storeirit_io: 1737 case Hexagon::S4_storeirif_io: 1738 case Hexagon::S4_pstorerit_rr: 1739 case Hexagon::S4_pstorerif_rr: 1740 case Hexagon::S2_pstorerit_pi: 1741 case Hexagon::S2_pstorerif_pi: 1742 1743 // V4 global address store before promoting to dot new. 1744 case Hexagon::S4_pstorerdt_abs: 1745 case Hexagon::S4_pstorerdf_abs: 1746 case Hexagon::S4_pstorerbt_abs: 1747 case Hexagon::S4_pstorerbf_abs: 1748 case Hexagon::S4_pstorerht_abs: 1749 case Hexagon::S4_pstorerhf_abs: 1750 case Hexagon::S4_pstorerit_abs: 1751 case Hexagon::S4_pstorerif_abs: 1752 return true; 1753 1754 // Predicated new value stores (i.e. if (p0) memw(..)=r0.new) are excluded 1755 // from the "Conditional Store" list. Because a predicated new value store 1756 // would NOT be promoted to a double dot new store. 1757 // This function returns yes for those stores that are predicated but not 1758 // yet promoted to predicate dot new instructions. 1759 } 1760 } 1761 1762 1763 bool HexagonInstrInfo::isConditionalTransfer(const MachineInstr *MI) const { 1764 switch (MI->getOpcode()) { 1765 case Hexagon::A2_tfrt: 1766 case Hexagon::A2_tfrf: 1767 case Hexagon::C2_cmoveit: 1768 case Hexagon::C2_cmoveif: 1769 case Hexagon::A2_tfrtnew: 1770 case Hexagon::A2_tfrfnew: 1771 case Hexagon::C2_cmovenewit: 1772 case Hexagon::C2_cmovenewif: 1773 case Hexagon::A2_tfrpt: 1774 case Hexagon::A2_tfrpf: 1775 return true; 1776 1777 default: 1778 return false; 1779 } 1780 return false; 1781 } 1782 1783 1784 // TODO: In order to have isExtendable for fpimm/f32Ext, we need to handle 1785 // isFPImm and later getFPImm as well. 1786 bool HexagonInstrInfo::isConstExtended(const MachineInstr *MI) const { 1787 const uint64_t F = MI->getDesc().TSFlags; 1788 unsigned isExtended = (F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask; 1789 if (isExtended) // Instruction must be extended. 1790 return true; 1791 1792 unsigned isExtendable = 1793 (F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask; 1794 if (!isExtendable) 1795 return false; 1796 1797 if (MI->isCall()) 1798 return false; 1799 1800 short ExtOpNum = getCExtOpNum(MI); 1801 const MachineOperand &MO = MI->getOperand(ExtOpNum); 1802 // Use MO operand flags to determine if MO 1803 // has the HMOTF_ConstExtended flag set. 1804 if (MO.getTargetFlags() && HexagonII::HMOTF_ConstExtended) 1805 return true; 1806 // If this is a Machine BB address we are talking about, and it is 1807 // not marked as extended, say so. 1808 if (MO.isMBB()) 1809 return false; 1810 1811 // We could be using an instruction with an extendable immediate and shoehorn 1812 // a global address into it. If it is a global address it will be constant 1813 // extended. We do this for COMBINE. 1814 // We currently only handle isGlobal() because it is the only kind of 1815 // object we are going to end up with here for now. 1816 // In the future we probably should add isSymbol(), etc. 1817 if (MO.isGlobal() || MO.isSymbol() || MO.isBlockAddress() || 1818 MO.isJTI() || MO.isCPI()) 1819 return true; 1820 1821 // If the extendable operand is not 'Immediate' type, the instruction should 1822 // have 'isExtended' flag set. 1823 assert(MO.isImm() && "Extendable operand must be Immediate type"); 1824 1825 int MinValue = getMinValue(MI); 1826 int MaxValue = getMaxValue(MI); 1827 int ImmValue = MO.getImm(); 1828 1829 return (ImmValue < MinValue || ImmValue > MaxValue); 1830 } 1831 1832 1833 bool HexagonInstrInfo::isDeallocRet(const MachineInstr *MI) const { 1834 switch (MI->getOpcode()) { 1835 case Hexagon::L4_return : 1836 case Hexagon::L4_return_t : 1837 case Hexagon::L4_return_f : 1838 case Hexagon::L4_return_tnew_pnt : 1839 case Hexagon::L4_return_fnew_pnt : 1840 case Hexagon::L4_return_tnew_pt : 1841 case Hexagon::L4_return_fnew_pt : 1842 return true; 1843 } 1844 return false; 1845 } 1846 1847 1848 // Return true when ConsMI uses a register defined by ProdMI. 1849 bool HexagonInstrInfo::isDependent(const MachineInstr *ProdMI, 1850 const MachineInstr *ConsMI) const { 1851 const MCInstrDesc &ProdMCID = ProdMI->getDesc(); 1852 if (!ProdMCID.getNumDefs()) 1853 return false; 1854 1855 auto &HRI = getRegisterInfo(); 1856 1857 SmallVector<unsigned, 4> DefsA; 1858 SmallVector<unsigned, 4> DefsB; 1859 SmallVector<unsigned, 8> UsesA; 1860 SmallVector<unsigned, 8> UsesB; 1861 1862 parseOperands(ProdMI, DefsA, UsesA); 1863 parseOperands(ConsMI, DefsB, UsesB); 1864 1865 for (auto &RegA : DefsA) 1866 for (auto &RegB : UsesB) { 1867 // True data dependency. 1868 if (RegA == RegB) 1869 return true; 1870 1871 if (Hexagon::DoubleRegsRegClass.contains(RegA)) 1872 for (MCSubRegIterator SubRegs(RegA, &HRI); SubRegs.isValid(); ++SubRegs) 1873 if (RegB == *SubRegs) 1874 return true; 1875 1876 if (Hexagon::DoubleRegsRegClass.contains(RegB)) 1877 for (MCSubRegIterator SubRegs(RegB, &HRI); SubRegs.isValid(); ++SubRegs) 1878 if (RegA == *SubRegs) 1879 return true; 1880 } 1881 1882 return false; 1883 } 1884 1885 1886 // Returns true if the instruction is alread a .cur. 1887 bool HexagonInstrInfo::isDotCurInst(const MachineInstr* MI) const { 1888 switch (MI->getOpcode()) { 1889 case Hexagon::V6_vL32b_cur_pi: 1890 case Hexagon::V6_vL32b_cur_ai: 1891 case Hexagon::V6_vL32b_cur_pi_128B: 1892 case Hexagon::V6_vL32b_cur_ai_128B: 1893 return true; 1894 } 1895 return false; 1896 } 1897 1898 1899 // Returns true, if any one of the operands is a dot new 1900 // insn, whether it is predicated dot new or register dot new. 1901 bool HexagonInstrInfo::isDotNewInst(const MachineInstr* MI) const { 1902 if (isNewValueInst(MI) || (isPredicated(*MI) && isPredicatedNew(*MI))) 1903 return true; 1904 1905 return false; 1906 } 1907 1908 1909 /// Symmetrical. See if these two instructions are fit for duplex pair. 1910 bool HexagonInstrInfo::isDuplexPair(const MachineInstr *MIa, 1911 const MachineInstr *MIb) const { 1912 HexagonII::SubInstructionGroup MIaG = getDuplexCandidateGroup(MIa); 1913 HexagonII::SubInstructionGroup MIbG = getDuplexCandidateGroup(MIb); 1914 return (isDuplexPairMatch(MIaG, MIbG) || isDuplexPairMatch(MIbG, MIaG)); 1915 } 1916 1917 1918 bool HexagonInstrInfo::isEarlySourceInstr(const MachineInstr *MI) const { 1919 if (!MI) 1920 return false; 1921 1922 if (MI->mayLoad() || MI->mayStore() || MI->isCompare()) 1923 return true; 1924 1925 // Multiply 1926 unsigned SchedClass = MI->getDesc().getSchedClass(); 1927 if (SchedClass == Hexagon::Sched::M_tc_3or4x_SLOT23) 1928 return true; 1929 return false; 1930 } 1931 1932 1933 bool HexagonInstrInfo::isEndLoopN(unsigned Opcode) const { 1934 return (Opcode == Hexagon::ENDLOOP0 || 1935 Opcode == Hexagon::ENDLOOP1); 1936 } 1937 1938 1939 bool HexagonInstrInfo::isExpr(unsigned OpType) const { 1940 switch(OpType) { 1941 case MachineOperand::MO_MachineBasicBlock: 1942 case MachineOperand::MO_GlobalAddress: 1943 case MachineOperand::MO_ExternalSymbol: 1944 case MachineOperand::MO_JumpTableIndex: 1945 case MachineOperand::MO_ConstantPoolIndex: 1946 case MachineOperand::MO_BlockAddress: 1947 return true; 1948 default: 1949 return false; 1950 } 1951 } 1952 1953 1954 bool HexagonInstrInfo::isExtendable(const MachineInstr *MI) const { 1955 const MCInstrDesc &MID = MI->getDesc(); 1956 const uint64_t F = MID.TSFlags; 1957 if ((F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask) 1958 return true; 1959 1960 // TODO: This is largely obsolete now. Will need to be removed 1961 // in consecutive patches. 1962 switch(MI->getOpcode()) { 1963 // TFR_FI Remains a special case. 1964 case Hexagon::TFR_FI: 1965 return true; 1966 default: 1967 return false; 1968 } 1969 return false; 1970 } 1971 1972 1973 // This returns true in two cases: 1974 // - The OP code itself indicates that this is an extended instruction. 1975 // - One of MOs has been marked with HMOTF_ConstExtended flag. 1976 bool HexagonInstrInfo::isExtended(const MachineInstr *MI) const { 1977 // First check if this is permanently extended op code. 1978 const uint64_t F = MI->getDesc().TSFlags; 1979 if ((F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask) 1980 return true; 1981 // Use MO operand flags to determine if one of MI's operands 1982 // has HMOTF_ConstExtended flag set. 1983 for (MachineInstr::const_mop_iterator I = MI->operands_begin(), 1984 E = MI->operands_end(); I != E; ++I) { 1985 if (I->getTargetFlags() && HexagonII::HMOTF_ConstExtended) 1986 return true; 1987 } 1988 return false; 1989 } 1990 1991 1992 bool HexagonInstrInfo::isFloat(const MachineInstr *MI) const { 1993 unsigned Opcode = MI->getOpcode(); 1994 const uint64_t F = get(Opcode).TSFlags; 1995 return (F >> HexagonII::FPPos) & HexagonII::FPMask; 1996 } 1997 1998 1999 // No V60 HVX VMEM with A_INDIRECT. 2000 bool HexagonInstrInfo::isHVXMemWithAIndirect(const MachineInstr *I, 2001 const MachineInstr *J) const { 2002 if (!isV60VectorInstruction(I)) 2003 return false; 2004 if (!I->mayLoad() && !I->mayStore()) 2005 return false; 2006 return J->isIndirectBranch() || isIndirectCall(J) || isIndirectL4Return(J); 2007 } 2008 2009 2010 bool HexagonInstrInfo::isIndirectCall(const MachineInstr *MI) const { 2011 switch (MI->getOpcode()) { 2012 case Hexagon::J2_callr : 2013 case Hexagon::J2_callrf : 2014 case Hexagon::J2_callrt : 2015 return true; 2016 } 2017 return false; 2018 } 2019 2020 2021 bool HexagonInstrInfo::isIndirectL4Return(const MachineInstr *MI) const { 2022 switch (MI->getOpcode()) { 2023 case Hexagon::L4_return : 2024 case Hexagon::L4_return_t : 2025 case Hexagon::L4_return_f : 2026 case Hexagon::L4_return_fnew_pnt : 2027 case Hexagon::L4_return_fnew_pt : 2028 case Hexagon::L4_return_tnew_pnt : 2029 case Hexagon::L4_return_tnew_pt : 2030 return true; 2031 } 2032 return false; 2033 } 2034 2035 2036 bool HexagonInstrInfo::isJumpR(const MachineInstr *MI) const { 2037 switch (MI->getOpcode()) { 2038 case Hexagon::J2_jumpr : 2039 case Hexagon::J2_jumprt : 2040 case Hexagon::J2_jumprf : 2041 case Hexagon::J2_jumprtnewpt : 2042 case Hexagon::J2_jumprfnewpt : 2043 case Hexagon::J2_jumprtnew : 2044 case Hexagon::J2_jumprfnew : 2045 return true; 2046 } 2047 return false; 2048 } 2049 2050 2051 // Return true if a given MI can accomodate given offset. 2052 // Use abs estimate as oppose to the exact number. 2053 // TODO: This will need to be changed to use MC level 2054 // definition of instruction extendable field size. 2055 bool HexagonInstrInfo::isJumpWithinBranchRange(const MachineInstr *MI, 2056 unsigned offset) const { 2057 // This selection of jump instructions matches to that what 2058 // AnalyzeBranch can parse, plus NVJ. 2059 if (isNewValueJump(MI)) // r9:2 2060 return isInt<11>(offset); 2061 2062 switch (MI->getOpcode()) { 2063 // Still missing Jump to address condition on register value. 2064 default: 2065 return false; 2066 case Hexagon::J2_jump: // bits<24> dst; // r22:2 2067 case Hexagon::J2_call: 2068 case Hexagon::CALLv3nr: 2069 return isInt<24>(offset); 2070 case Hexagon::J2_jumpt: //bits<17> dst; // r15:2 2071 case Hexagon::J2_jumpf: 2072 case Hexagon::J2_jumptnew: 2073 case Hexagon::J2_jumptnewpt: 2074 case Hexagon::J2_jumpfnew: 2075 case Hexagon::J2_jumpfnewpt: 2076 case Hexagon::J2_callt: 2077 case Hexagon::J2_callf: 2078 return isInt<17>(offset); 2079 case Hexagon::J2_loop0i: 2080 case Hexagon::J2_loop0iext: 2081 case Hexagon::J2_loop0r: 2082 case Hexagon::J2_loop0rext: 2083 case Hexagon::J2_loop1i: 2084 case Hexagon::J2_loop1iext: 2085 case Hexagon::J2_loop1r: 2086 case Hexagon::J2_loop1rext: 2087 return isInt<9>(offset); 2088 // TODO: Add all the compound branches here. Can we do this in Relation model? 2089 case Hexagon::J4_cmpeqi_tp0_jump_nt: 2090 case Hexagon::J4_cmpeqi_tp1_jump_nt: 2091 return isInt<11>(offset); 2092 } 2093 } 2094 2095 2096 bool HexagonInstrInfo::isLateInstrFeedsEarlyInstr(const MachineInstr *LRMI, 2097 const MachineInstr *ESMI) const { 2098 if (!LRMI || !ESMI) 2099 return false; 2100 2101 bool isLate = isLateResultInstr(LRMI); 2102 bool isEarly = isEarlySourceInstr(ESMI); 2103 2104 DEBUG(dbgs() << "V60" << (isLate ? "-LR " : " -- ")); 2105 DEBUG(LRMI->dump()); 2106 DEBUG(dbgs() << "V60" << (isEarly ? "-ES " : " -- ")); 2107 DEBUG(ESMI->dump()); 2108 2109 if (isLate && isEarly) { 2110 DEBUG(dbgs() << "++Is Late Result feeding Early Source\n"); 2111 return true; 2112 } 2113 2114 return false; 2115 } 2116 2117 2118 bool HexagonInstrInfo::isLateResultInstr(const MachineInstr *MI) const { 2119 if (!MI) 2120 return false; 2121 2122 switch (MI->getOpcode()) { 2123 case TargetOpcode::EXTRACT_SUBREG: 2124 case TargetOpcode::INSERT_SUBREG: 2125 case TargetOpcode::SUBREG_TO_REG: 2126 case TargetOpcode::REG_SEQUENCE: 2127 case TargetOpcode::IMPLICIT_DEF: 2128 case TargetOpcode::COPY: 2129 case TargetOpcode::INLINEASM: 2130 case TargetOpcode::PHI: 2131 return false; 2132 default: 2133 break; 2134 } 2135 2136 unsigned SchedClass = MI->getDesc().getSchedClass(); 2137 2138 switch (SchedClass) { 2139 case Hexagon::Sched::ALU32_2op_tc_1_SLOT0123: 2140 case Hexagon::Sched::ALU32_3op_tc_1_SLOT0123: 2141 case Hexagon::Sched::ALU32_ADDI_tc_1_SLOT0123: 2142 case Hexagon::Sched::ALU64_tc_1_SLOT23: 2143 case Hexagon::Sched::EXTENDER_tc_1_SLOT0123: 2144 case Hexagon::Sched::S_2op_tc_1_SLOT23: 2145 case Hexagon::Sched::S_3op_tc_1_SLOT23: 2146 case Hexagon::Sched::V2LDST_tc_ld_SLOT01: 2147 case Hexagon::Sched::V2LDST_tc_st_SLOT0: 2148 case Hexagon::Sched::V2LDST_tc_st_SLOT01: 2149 case Hexagon::Sched::V4LDST_tc_ld_SLOT01: 2150 case Hexagon::Sched::V4LDST_tc_st_SLOT0: 2151 case Hexagon::Sched::V4LDST_tc_st_SLOT01: 2152 return false; 2153 } 2154 return true; 2155 } 2156 2157 2158 bool HexagonInstrInfo::isLateSourceInstr(const MachineInstr *MI) const { 2159 if (!MI) 2160 return false; 2161 2162 // Instructions with iclass A_CVI_VX and attribute A_CVI_LATE uses a multiply 2163 // resource, but all operands can be received late like an ALU instruction. 2164 return MI->getDesc().getSchedClass() == Hexagon::Sched::CVI_VX_LATE; 2165 } 2166 2167 2168 bool HexagonInstrInfo::isLoopN(const MachineInstr *MI) const { 2169 unsigned Opcode = MI->getOpcode(); 2170 return Opcode == Hexagon::J2_loop0i || 2171 Opcode == Hexagon::J2_loop0r || 2172 Opcode == Hexagon::J2_loop0iext || 2173 Opcode == Hexagon::J2_loop0rext || 2174 Opcode == Hexagon::J2_loop1i || 2175 Opcode == Hexagon::J2_loop1r || 2176 Opcode == Hexagon::J2_loop1iext || 2177 Opcode == Hexagon::J2_loop1rext; 2178 } 2179 2180 2181 bool HexagonInstrInfo::isMemOp(const MachineInstr *MI) const { 2182 switch (MI->getOpcode()) { 2183 default: return false; 2184 case Hexagon::L4_iadd_memopw_io : 2185 case Hexagon::L4_isub_memopw_io : 2186 case Hexagon::L4_add_memopw_io : 2187 case Hexagon::L4_sub_memopw_io : 2188 case Hexagon::L4_and_memopw_io : 2189 case Hexagon::L4_or_memopw_io : 2190 case Hexagon::L4_iadd_memoph_io : 2191 case Hexagon::L4_isub_memoph_io : 2192 case Hexagon::L4_add_memoph_io : 2193 case Hexagon::L4_sub_memoph_io : 2194 case Hexagon::L4_and_memoph_io : 2195 case Hexagon::L4_or_memoph_io : 2196 case Hexagon::L4_iadd_memopb_io : 2197 case Hexagon::L4_isub_memopb_io : 2198 case Hexagon::L4_add_memopb_io : 2199 case Hexagon::L4_sub_memopb_io : 2200 case Hexagon::L4_and_memopb_io : 2201 case Hexagon::L4_or_memopb_io : 2202 case Hexagon::L4_ior_memopb_io: 2203 case Hexagon::L4_ior_memoph_io: 2204 case Hexagon::L4_ior_memopw_io: 2205 case Hexagon::L4_iand_memopb_io: 2206 case Hexagon::L4_iand_memoph_io: 2207 case Hexagon::L4_iand_memopw_io: 2208 return true; 2209 } 2210 return false; 2211 } 2212 2213 2214 bool HexagonInstrInfo::isNewValue(const MachineInstr* MI) const { 2215 const uint64_t F = MI->getDesc().TSFlags; 2216 return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask; 2217 } 2218 2219 2220 bool HexagonInstrInfo::isNewValue(unsigned Opcode) const { 2221 const uint64_t F = get(Opcode).TSFlags; 2222 return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask; 2223 } 2224 2225 2226 bool HexagonInstrInfo::isNewValueInst(const MachineInstr *MI) const { 2227 return isNewValueJump(MI) || isNewValueStore(MI); 2228 } 2229 2230 2231 bool HexagonInstrInfo::isNewValueJump(const MachineInstr *MI) const { 2232 return isNewValue(MI) && MI->isBranch(); 2233 } 2234 2235 2236 bool HexagonInstrInfo::isNewValueJump(unsigned Opcode) const { 2237 return isNewValue(Opcode) && get(Opcode).isBranch() && isPredicated(Opcode); 2238 } 2239 2240 2241 bool HexagonInstrInfo::isNewValueStore(const MachineInstr *MI) const { 2242 const uint64_t F = MI->getDesc().TSFlags; 2243 return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask; 2244 } 2245 2246 2247 bool HexagonInstrInfo::isNewValueStore(unsigned Opcode) const { 2248 const uint64_t F = get(Opcode).TSFlags; 2249 return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask; 2250 } 2251 2252 2253 // Returns true if a particular operand is extendable for an instruction. 2254 bool HexagonInstrInfo::isOperandExtended(const MachineInstr *MI, 2255 unsigned OperandNum) const { 2256 const uint64_t F = MI->getDesc().TSFlags; 2257 return ((F >> HexagonII::ExtendableOpPos) & HexagonII::ExtendableOpMask) 2258 == OperandNum; 2259 } 2260 2261 2262 bool HexagonInstrInfo::isPostIncrement(const MachineInstr* MI) const { 2263 return getAddrMode(MI) == HexagonII::PostInc; 2264 } 2265 2266 2267 bool HexagonInstrInfo::isPredicatedNew(const MachineInstr &MI) const { 2268 const uint64_t F = MI.getDesc().TSFlags; 2269 assert(isPredicated(MI)); 2270 return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask; 2271 } 2272 2273 2274 bool HexagonInstrInfo::isPredicatedNew(unsigned Opcode) const { 2275 const uint64_t F = get(Opcode).TSFlags; 2276 assert(isPredicated(Opcode)); 2277 return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask; 2278 } 2279 2280 2281 bool HexagonInstrInfo::isPredicatedTrue(const MachineInstr &MI) const { 2282 const uint64_t F = MI.getDesc().TSFlags; 2283 return !((F >> HexagonII::PredicatedFalsePos) & 2284 HexagonII::PredicatedFalseMask); 2285 } 2286 2287 2288 bool HexagonInstrInfo::isPredicatedTrue(unsigned Opcode) const { 2289 const uint64_t F = get(Opcode).TSFlags; 2290 // Make sure that the instruction is predicated. 2291 assert((F>> HexagonII::PredicatedPos) & HexagonII::PredicatedMask); 2292 return !((F >> HexagonII::PredicatedFalsePos) & 2293 HexagonII::PredicatedFalseMask); 2294 } 2295 2296 2297 bool HexagonInstrInfo::isPredicated(unsigned Opcode) const { 2298 const uint64_t F = get(Opcode).TSFlags; 2299 return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask; 2300 } 2301 2302 2303 bool HexagonInstrInfo::isPredicateLate(unsigned Opcode) const { 2304 const uint64_t F = get(Opcode).TSFlags; 2305 return ~(F >> HexagonII::PredicateLatePos) & HexagonII::PredicateLateMask; 2306 } 2307 2308 2309 bool HexagonInstrInfo::isPredictedTaken(unsigned Opcode) const { 2310 const uint64_t F = get(Opcode).TSFlags; 2311 assert(get(Opcode).isBranch() && 2312 (isPredicatedNew(Opcode) || isNewValue(Opcode))); 2313 return (F >> HexagonII::TakenPos) & HexagonII::TakenMask; 2314 } 2315 2316 2317 bool HexagonInstrInfo::isSaveCalleeSavedRegsCall(const MachineInstr *MI) const { 2318 return MI->getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4 || 2319 MI->getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT || 2320 MI->getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_PIC || 2321 MI->getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT_PIC; 2322 } 2323 2324 2325 bool HexagonInstrInfo::isSignExtendingLoad(const MachineInstr* MI) const { 2326 switch (MI->getOpcode()) { 2327 // Byte 2328 case Hexagon::L2_loadrb_io: 2329 case Hexagon::L4_loadrb_ur: 2330 case Hexagon::L4_loadrb_ap: 2331 case Hexagon::L2_loadrb_pr: 2332 case Hexagon::L2_loadrb_pbr: 2333 case Hexagon::L2_loadrb_pi: 2334 case Hexagon::L2_loadrb_pci: 2335 case Hexagon::L2_loadrb_pcr: 2336 case Hexagon::L2_loadbsw2_io: 2337 case Hexagon::L4_loadbsw2_ur: 2338 case Hexagon::L4_loadbsw2_ap: 2339 case Hexagon::L2_loadbsw2_pr: 2340 case Hexagon::L2_loadbsw2_pbr: 2341 case Hexagon::L2_loadbsw2_pi: 2342 case Hexagon::L2_loadbsw2_pci: 2343 case Hexagon::L2_loadbsw2_pcr: 2344 case Hexagon::L2_loadbsw4_io: 2345 case Hexagon::L4_loadbsw4_ur: 2346 case Hexagon::L4_loadbsw4_ap: 2347 case Hexagon::L2_loadbsw4_pr: 2348 case Hexagon::L2_loadbsw4_pbr: 2349 case Hexagon::L2_loadbsw4_pi: 2350 case Hexagon::L2_loadbsw4_pci: 2351 case Hexagon::L2_loadbsw4_pcr: 2352 case Hexagon::L4_loadrb_rr: 2353 case Hexagon::L2_ploadrbt_io: 2354 case Hexagon::L2_ploadrbt_pi: 2355 case Hexagon::L2_ploadrbf_io: 2356 case Hexagon::L2_ploadrbf_pi: 2357 case Hexagon::L2_ploadrbtnew_io: 2358 case Hexagon::L2_ploadrbfnew_io: 2359 case Hexagon::L4_ploadrbt_rr: 2360 case Hexagon::L4_ploadrbf_rr: 2361 case Hexagon::L4_ploadrbtnew_rr: 2362 case Hexagon::L4_ploadrbfnew_rr: 2363 case Hexagon::L2_ploadrbtnew_pi: 2364 case Hexagon::L2_ploadrbfnew_pi: 2365 case Hexagon::L4_ploadrbt_abs: 2366 case Hexagon::L4_ploadrbf_abs: 2367 case Hexagon::L4_ploadrbtnew_abs: 2368 case Hexagon::L4_ploadrbfnew_abs: 2369 case Hexagon::L2_loadrbgp: 2370 // Half 2371 case Hexagon::L2_loadrh_io: 2372 case Hexagon::L4_loadrh_ur: 2373 case Hexagon::L4_loadrh_ap: 2374 case Hexagon::L2_loadrh_pr: 2375 case Hexagon::L2_loadrh_pbr: 2376 case Hexagon::L2_loadrh_pi: 2377 case Hexagon::L2_loadrh_pci: 2378 case Hexagon::L2_loadrh_pcr: 2379 case Hexagon::L4_loadrh_rr: 2380 case Hexagon::L2_ploadrht_io: 2381 case Hexagon::L2_ploadrht_pi: 2382 case Hexagon::L2_ploadrhf_io: 2383 case Hexagon::L2_ploadrhf_pi: 2384 case Hexagon::L2_ploadrhtnew_io: 2385 case Hexagon::L2_ploadrhfnew_io: 2386 case Hexagon::L4_ploadrht_rr: 2387 case Hexagon::L4_ploadrhf_rr: 2388 case Hexagon::L4_ploadrhtnew_rr: 2389 case Hexagon::L4_ploadrhfnew_rr: 2390 case Hexagon::L2_ploadrhtnew_pi: 2391 case Hexagon::L2_ploadrhfnew_pi: 2392 case Hexagon::L4_ploadrht_abs: 2393 case Hexagon::L4_ploadrhf_abs: 2394 case Hexagon::L4_ploadrhtnew_abs: 2395 case Hexagon::L4_ploadrhfnew_abs: 2396 case Hexagon::L2_loadrhgp: 2397 return true; 2398 default: 2399 return false; 2400 } 2401 } 2402 2403 2404 bool HexagonInstrInfo::isSolo(const MachineInstr* MI) const { 2405 const uint64_t F = MI->getDesc().TSFlags; 2406 return (F >> HexagonII::SoloPos) & HexagonII::SoloMask; 2407 } 2408 2409 2410 bool HexagonInstrInfo::isSpillPredRegOp(const MachineInstr *MI) const { 2411 switch (MI->getOpcode()) { 2412 case Hexagon::STriw_pred : 2413 case Hexagon::LDriw_pred : 2414 return true; 2415 default: 2416 return false; 2417 } 2418 } 2419 2420 2421 // Returns true when SU has a timing class TC1. 2422 bool HexagonInstrInfo::isTC1(const MachineInstr *MI) const { 2423 unsigned SchedClass = MI->getDesc().getSchedClass(); 2424 switch (SchedClass) { 2425 case Hexagon::Sched::ALU32_2op_tc_1_SLOT0123: 2426 case Hexagon::Sched::ALU32_3op_tc_1_SLOT0123: 2427 case Hexagon::Sched::ALU32_ADDI_tc_1_SLOT0123: 2428 case Hexagon::Sched::ALU64_tc_1_SLOT23: 2429 case Hexagon::Sched::EXTENDER_tc_1_SLOT0123: 2430 //case Hexagon::Sched::M_tc_1_SLOT23: 2431 case Hexagon::Sched::S_2op_tc_1_SLOT23: 2432 case Hexagon::Sched::S_3op_tc_1_SLOT23: 2433 return true; 2434 2435 default: 2436 return false; 2437 } 2438 } 2439 2440 2441 bool HexagonInstrInfo::isTC2(const MachineInstr *MI) const { 2442 unsigned SchedClass = MI->getDesc().getSchedClass(); 2443 switch (SchedClass) { 2444 case Hexagon::Sched::ALU32_3op_tc_2_SLOT0123: 2445 case Hexagon::Sched::ALU64_tc_2_SLOT23: 2446 case Hexagon::Sched::CR_tc_2_SLOT3: 2447 case Hexagon::Sched::M_tc_2_SLOT23: 2448 case Hexagon::Sched::S_2op_tc_2_SLOT23: 2449 case Hexagon::Sched::S_3op_tc_2_SLOT23: 2450 return true; 2451 2452 default: 2453 return false; 2454 } 2455 } 2456 2457 2458 bool HexagonInstrInfo::isTC2Early(const MachineInstr *MI) const { 2459 unsigned SchedClass = MI->getDesc().getSchedClass(); 2460 switch (SchedClass) { 2461 case Hexagon::Sched::ALU32_2op_tc_2early_SLOT0123: 2462 case Hexagon::Sched::ALU32_3op_tc_2early_SLOT0123: 2463 case Hexagon::Sched::ALU64_tc_2early_SLOT23: 2464 case Hexagon::Sched::CR_tc_2early_SLOT23: 2465 case Hexagon::Sched::CR_tc_2early_SLOT3: 2466 case Hexagon::Sched::J_tc_2early_SLOT0123: 2467 case Hexagon::Sched::J_tc_2early_SLOT2: 2468 case Hexagon::Sched::J_tc_2early_SLOT23: 2469 case Hexagon::Sched::S_2op_tc_2early_SLOT23: 2470 case Hexagon::Sched::S_3op_tc_2early_SLOT23: 2471 return true; 2472 2473 default: 2474 return false; 2475 } 2476 } 2477 2478 2479 bool HexagonInstrInfo::isTC4x(const MachineInstr *MI) const { 2480 if (!MI) 2481 return false; 2482 2483 unsigned SchedClass = MI->getDesc().getSchedClass(); 2484 return SchedClass == Hexagon::Sched::M_tc_3or4x_SLOT23; 2485 } 2486 2487 2488 bool HexagonInstrInfo::isV60VectorInstruction(const MachineInstr *MI) const { 2489 if (!MI) 2490 return false; 2491 2492 const uint64_t V = getType(MI); 2493 return HexagonII::TypeCVI_FIRST <= V && V <= HexagonII::TypeCVI_LAST; 2494 } 2495 2496 2497 // Check if the Offset is a valid auto-inc imm by Load/Store Type. 2498 // 2499 bool HexagonInstrInfo::isValidAutoIncImm(const EVT VT, const int Offset) const { 2500 if (VT == MVT::v16i32 || VT == MVT::v8i64 || 2501 VT == MVT::v32i16 || VT == MVT::v64i8) { 2502 return (Offset >= Hexagon_MEMV_AUTOINC_MIN && 2503 Offset <= Hexagon_MEMV_AUTOINC_MAX && 2504 (Offset & 0x3f) == 0); 2505 } 2506 // 128B 2507 if (VT == MVT::v32i32 || VT == MVT::v16i64 || 2508 VT == MVT::v64i16 || VT == MVT::v128i8) { 2509 return (Offset >= Hexagon_MEMV_AUTOINC_MIN_128B && 2510 Offset <= Hexagon_MEMV_AUTOINC_MAX_128B && 2511 (Offset & 0x7f) == 0); 2512 } 2513 if (VT == MVT::i64) { 2514 return (Offset >= Hexagon_MEMD_AUTOINC_MIN && 2515 Offset <= Hexagon_MEMD_AUTOINC_MAX && 2516 (Offset & 0x7) == 0); 2517 } 2518 if (VT == MVT::i32) { 2519 return (Offset >= Hexagon_MEMW_AUTOINC_MIN && 2520 Offset <= Hexagon_MEMW_AUTOINC_MAX && 2521 (Offset & 0x3) == 0); 2522 } 2523 if (VT == MVT::i16) { 2524 return (Offset >= Hexagon_MEMH_AUTOINC_MIN && 2525 Offset <= Hexagon_MEMH_AUTOINC_MAX && 2526 (Offset & 0x1) == 0); 2527 } 2528 if (VT == MVT::i8) { 2529 return (Offset >= Hexagon_MEMB_AUTOINC_MIN && 2530 Offset <= Hexagon_MEMB_AUTOINC_MAX); 2531 } 2532 llvm_unreachable("Not an auto-inc opc!"); 2533 } 2534 2535 2536 bool HexagonInstrInfo::isValidOffset(unsigned Opcode, int Offset, 2537 bool Extend) const { 2538 // This function is to check whether the "Offset" is in the correct range of 2539 // the given "Opcode". If "Offset" is not in the correct range, "A2_addi" is 2540 // inserted to calculate the final address. Due to this reason, the function 2541 // assumes that the "Offset" has correct alignment. 2542 // We used to assert if the offset was not properly aligned, however, 2543 // there are cases where a misaligned pointer recast can cause this 2544 // problem, and we need to allow for it. The front end warns of such 2545 // misaligns with respect to load size. 2546 2547 switch (Opcode) { 2548 case Hexagon::STriq_pred_V6: 2549 case Hexagon::STriq_pred_vec_V6: 2550 case Hexagon::STriv_pseudo_V6: 2551 case Hexagon::STrivv_pseudo_V6: 2552 case Hexagon::LDriq_pred_V6: 2553 case Hexagon::LDriq_pred_vec_V6: 2554 case Hexagon::LDriv_pseudo_V6: 2555 case Hexagon::LDrivv_pseudo_V6: 2556 case Hexagon::LDrivv_indexed: 2557 case Hexagon::STrivv_indexed: 2558 case Hexagon::V6_vL32b_ai: 2559 case Hexagon::V6_vS32b_ai: 2560 case Hexagon::V6_vL32Ub_ai: 2561 case Hexagon::V6_vS32Ub_ai: 2562 return (Offset >= Hexagon_MEMV_OFFSET_MIN) && 2563 (Offset <= Hexagon_MEMV_OFFSET_MAX); 2564 2565 case Hexagon::STriq_pred_V6_128B: 2566 case Hexagon::STriq_pred_vec_V6_128B: 2567 case Hexagon::STriv_pseudo_V6_128B: 2568 case Hexagon::STrivv_pseudo_V6_128B: 2569 case Hexagon::LDriq_pred_V6_128B: 2570 case Hexagon::LDriq_pred_vec_V6_128B: 2571 case Hexagon::LDriv_pseudo_V6_128B: 2572 case Hexagon::LDrivv_pseudo_V6_128B: 2573 case Hexagon::LDrivv_indexed_128B: 2574 case Hexagon::STrivv_indexed_128B: 2575 case Hexagon::V6_vL32b_ai_128B: 2576 case Hexagon::V6_vS32b_ai_128B: 2577 case Hexagon::V6_vL32Ub_ai_128B: 2578 case Hexagon::V6_vS32Ub_ai_128B: 2579 return (Offset >= Hexagon_MEMV_OFFSET_MIN_128B) && 2580 (Offset <= Hexagon_MEMV_OFFSET_MAX_128B); 2581 2582 case Hexagon::J2_loop0i: 2583 case Hexagon::J2_loop1i: 2584 return isUInt<10>(Offset); 2585 } 2586 2587 if (Extend) 2588 return true; 2589 2590 switch (Opcode) { 2591 case Hexagon::L2_loadri_io: 2592 case Hexagon::S2_storeri_io: 2593 return (Offset >= Hexagon_MEMW_OFFSET_MIN) && 2594 (Offset <= Hexagon_MEMW_OFFSET_MAX); 2595 2596 case Hexagon::L2_loadrd_io: 2597 case Hexagon::S2_storerd_io: 2598 return (Offset >= Hexagon_MEMD_OFFSET_MIN) && 2599 (Offset <= Hexagon_MEMD_OFFSET_MAX); 2600 2601 case Hexagon::L2_loadrh_io: 2602 case Hexagon::L2_loadruh_io: 2603 case Hexagon::S2_storerh_io: 2604 return (Offset >= Hexagon_MEMH_OFFSET_MIN) && 2605 (Offset <= Hexagon_MEMH_OFFSET_MAX); 2606 2607 case Hexagon::L2_loadrb_io: 2608 case Hexagon::L2_loadrub_io: 2609 case Hexagon::S2_storerb_io: 2610 return (Offset >= Hexagon_MEMB_OFFSET_MIN) && 2611 (Offset <= Hexagon_MEMB_OFFSET_MAX); 2612 2613 case Hexagon::A2_addi: 2614 return (Offset >= Hexagon_ADDI_OFFSET_MIN) && 2615 (Offset <= Hexagon_ADDI_OFFSET_MAX); 2616 2617 case Hexagon::L4_iadd_memopw_io : 2618 case Hexagon::L4_isub_memopw_io : 2619 case Hexagon::L4_add_memopw_io : 2620 case Hexagon::L4_sub_memopw_io : 2621 case Hexagon::L4_and_memopw_io : 2622 case Hexagon::L4_or_memopw_io : 2623 return (0 <= Offset && Offset <= 255); 2624 2625 case Hexagon::L4_iadd_memoph_io : 2626 case Hexagon::L4_isub_memoph_io : 2627 case Hexagon::L4_add_memoph_io : 2628 case Hexagon::L4_sub_memoph_io : 2629 case Hexagon::L4_and_memoph_io : 2630 case Hexagon::L4_or_memoph_io : 2631 return (0 <= Offset && Offset <= 127); 2632 2633 case Hexagon::L4_iadd_memopb_io : 2634 case Hexagon::L4_isub_memopb_io : 2635 case Hexagon::L4_add_memopb_io : 2636 case Hexagon::L4_sub_memopb_io : 2637 case Hexagon::L4_and_memopb_io : 2638 case Hexagon::L4_or_memopb_io : 2639 return (0 <= Offset && Offset <= 63); 2640 2641 // LDriw_xxx and STriw_xxx are pseudo operations, so it has to take offset of 2642 // any size. Later pass knows how to handle it. 2643 case Hexagon::STriw_pred: 2644 case Hexagon::LDriw_pred: 2645 case Hexagon::STriw_mod: 2646 case Hexagon::LDriw_mod: 2647 return true; 2648 2649 case Hexagon::TFR_FI: 2650 case Hexagon::TFR_FIA: 2651 case Hexagon::INLINEASM: 2652 return true; 2653 2654 case Hexagon::L2_ploadrbt_io: 2655 case Hexagon::L2_ploadrbf_io: 2656 case Hexagon::L2_ploadrubt_io: 2657 case Hexagon::L2_ploadrubf_io: 2658 case Hexagon::S2_pstorerbt_io: 2659 case Hexagon::S2_pstorerbf_io: 2660 case Hexagon::S4_storeirb_io: 2661 case Hexagon::S4_storeirbt_io: 2662 case Hexagon::S4_storeirbf_io: 2663 return isUInt<6>(Offset); 2664 2665 case Hexagon::L2_ploadrht_io: 2666 case Hexagon::L2_ploadrhf_io: 2667 case Hexagon::L2_ploadruht_io: 2668 case Hexagon::L2_ploadruhf_io: 2669 case Hexagon::S2_pstorerht_io: 2670 case Hexagon::S2_pstorerhf_io: 2671 case Hexagon::S4_storeirh_io: 2672 case Hexagon::S4_storeirht_io: 2673 case Hexagon::S4_storeirhf_io: 2674 return isShiftedUInt<6,1>(Offset); 2675 2676 case Hexagon::L2_ploadrit_io: 2677 case Hexagon::L2_ploadrif_io: 2678 case Hexagon::S2_pstorerit_io: 2679 case Hexagon::S2_pstorerif_io: 2680 case Hexagon::S4_storeiri_io: 2681 case Hexagon::S4_storeirit_io: 2682 case Hexagon::S4_storeirif_io: 2683 return isShiftedUInt<6,2>(Offset); 2684 2685 case Hexagon::L2_ploadrdt_io: 2686 case Hexagon::L2_ploadrdf_io: 2687 case Hexagon::S2_pstorerdt_io: 2688 case Hexagon::S2_pstorerdf_io: 2689 return isShiftedUInt<6,3>(Offset); 2690 } // switch 2691 2692 llvm_unreachable("No offset range is defined for this opcode. " 2693 "Please define it in the above switch statement!"); 2694 } 2695 2696 2697 bool HexagonInstrInfo::isVecAcc(const MachineInstr *MI) const { 2698 return MI && isV60VectorInstruction(MI) && isAccumulator(MI); 2699 } 2700 2701 2702 bool HexagonInstrInfo::isVecALU(const MachineInstr *MI) const { 2703 if (!MI) 2704 return false; 2705 const uint64_t F = get(MI->getOpcode()).TSFlags; 2706 const uint64_t V = ((F >> HexagonII::TypePos) & HexagonII::TypeMask); 2707 return 2708 V == HexagonII::TypeCVI_VA || 2709 V == HexagonII::TypeCVI_VA_DV; 2710 } 2711 2712 2713 bool HexagonInstrInfo::isVecUsableNextPacket(const MachineInstr *ProdMI, 2714 const MachineInstr *ConsMI) const { 2715 if (EnableACCForwarding && isVecAcc(ProdMI) && isVecAcc(ConsMI)) 2716 return true; 2717 2718 if (EnableALUForwarding && (isVecALU(ConsMI) || isLateSourceInstr(ConsMI))) 2719 return true; 2720 2721 if (mayBeNewStore(ConsMI)) 2722 return true; 2723 2724 return false; 2725 } 2726 2727 2728 bool HexagonInstrInfo::isZeroExtendingLoad(const MachineInstr* MI) const { 2729 switch (MI->getOpcode()) { 2730 // Byte 2731 case Hexagon::L2_loadrub_io: 2732 case Hexagon::L4_loadrub_ur: 2733 case Hexagon::L4_loadrub_ap: 2734 case Hexagon::L2_loadrub_pr: 2735 case Hexagon::L2_loadrub_pbr: 2736 case Hexagon::L2_loadrub_pi: 2737 case Hexagon::L2_loadrub_pci: 2738 case Hexagon::L2_loadrub_pcr: 2739 case Hexagon::L2_loadbzw2_io: 2740 case Hexagon::L4_loadbzw2_ur: 2741 case Hexagon::L4_loadbzw2_ap: 2742 case Hexagon::L2_loadbzw2_pr: 2743 case Hexagon::L2_loadbzw2_pbr: 2744 case Hexagon::L2_loadbzw2_pi: 2745 case Hexagon::L2_loadbzw2_pci: 2746 case Hexagon::L2_loadbzw2_pcr: 2747 case Hexagon::L2_loadbzw4_io: 2748 case Hexagon::L4_loadbzw4_ur: 2749 case Hexagon::L4_loadbzw4_ap: 2750 case Hexagon::L2_loadbzw4_pr: 2751 case Hexagon::L2_loadbzw4_pbr: 2752 case Hexagon::L2_loadbzw4_pi: 2753 case Hexagon::L2_loadbzw4_pci: 2754 case Hexagon::L2_loadbzw4_pcr: 2755 case Hexagon::L4_loadrub_rr: 2756 case Hexagon::L2_ploadrubt_io: 2757 case Hexagon::L2_ploadrubt_pi: 2758 case Hexagon::L2_ploadrubf_io: 2759 case Hexagon::L2_ploadrubf_pi: 2760 case Hexagon::L2_ploadrubtnew_io: 2761 case Hexagon::L2_ploadrubfnew_io: 2762 case Hexagon::L4_ploadrubt_rr: 2763 case Hexagon::L4_ploadrubf_rr: 2764 case Hexagon::L4_ploadrubtnew_rr: 2765 case Hexagon::L4_ploadrubfnew_rr: 2766 case Hexagon::L2_ploadrubtnew_pi: 2767 case Hexagon::L2_ploadrubfnew_pi: 2768 case Hexagon::L4_ploadrubt_abs: 2769 case Hexagon::L4_ploadrubf_abs: 2770 case Hexagon::L4_ploadrubtnew_abs: 2771 case Hexagon::L4_ploadrubfnew_abs: 2772 case Hexagon::L2_loadrubgp: 2773 // Half 2774 case Hexagon::L2_loadruh_io: 2775 case Hexagon::L4_loadruh_ur: 2776 case Hexagon::L4_loadruh_ap: 2777 case Hexagon::L2_loadruh_pr: 2778 case Hexagon::L2_loadruh_pbr: 2779 case Hexagon::L2_loadruh_pi: 2780 case Hexagon::L2_loadruh_pci: 2781 case Hexagon::L2_loadruh_pcr: 2782 case Hexagon::L4_loadruh_rr: 2783 case Hexagon::L2_ploadruht_io: 2784 case Hexagon::L2_ploadruht_pi: 2785 case Hexagon::L2_ploadruhf_io: 2786 case Hexagon::L2_ploadruhf_pi: 2787 case Hexagon::L2_ploadruhtnew_io: 2788 case Hexagon::L2_ploadruhfnew_io: 2789 case Hexagon::L4_ploadruht_rr: 2790 case Hexagon::L4_ploadruhf_rr: 2791 case Hexagon::L4_ploadruhtnew_rr: 2792 case Hexagon::L4_ploadruhfnew_rr: 2793 case Hexagon::L2_ploadruhtnew_pi: 2794 case Hexagon::L2_ploadruhfnew_pi: 2795 case Hexagon::L4_ploadruht_abs: 2796 case Hexagon::L4_ploadruhf_abs: 2797 case Hexagon::L4_ploadruhtnew_abs: 2798 case Hexagon::L4_ploadruhfnew_abs: 2799 case Hexagon::L2_loadruhgp: 2800 return true; 2801 default: 2802 return false; 2803 } 2804 } 2805 2806 2807 /// \brief Can these instructions execute at the same time in a bundle. 2808 bool HexagonInstrInfo::canExecuteInBundle(const MachineInstr *First, 2809 const MachineInstr *Second) const { 2810 if (DisableNVSchedule) 2811 return false; 2812 if (mayBeNewStore(Second)) { 2813 // Make sure the definition of the first instruction is the value being 2814 // stored. 2815 const MachineOperand &Stored = 2816 Second->getOperand(Second->getNumOperands() - 1); 2817 if (!Stored.isReg()) 2818 return false; 2819 for (unsigned i = 0, e = First->getNumOperands(); i < e; ++i) { 2820 const MachineOperand &Op = First->getOperand(i); 2821 if (Op.isReg() && Op.isDef() && Op.getReg() == Stored.getReg()) 2822 return true; 2823 } 2824 } 2825 return false; 2826 } 2827 2828 2829 bool HexagonInstrInfo::hasEHLabel(const MachineBasicBlock *B) const { 2830 for (auto &I : *B) 2831 if (I.isEHLabel()) 2832 return true; 2833 return false; 2834 } 2835 2836 2837 // Returns true if an instruction can be converted into a non-extended 2838 // equivalent instruction. 2839 bool HexagonInstrInfo::hasNonExtEquivalent(const MachineInstr *MI) const { 2840 short NonExtOpcode; 2841 // Check if the instruction has a register form that uses register in place 2842 // of the extended operand, if so return that as the non-extended form. 2843 if (Hexagon::getRegForm(MI->getOpcode()) >= 0) 2844 return true; 2845 2846 if (MI->getDesc().mayLoad() || MI->getDesc().mayStore()) { 2847 // Check addressing mode and retrieve non-ext equivalent instruction. 2848 2849 switch (getAddrMode(MI)) { 2850 case HexagonII::Absolute : 2851 // Load/store with absolute addressing mode can be converted into 2852 // base+offset mode. 2853 NonExtOpcode = Hexagon::getBaseWithImmOffset(MI->getOpcode()); 2854 break; 2855 case HexagonII::BaseImmOffset : 2856 // Load/store with base+offset addressing mode can be converted into 2857 // base+register offset addressing mode. However left shift operand should 2858 // be set to 0. 2859 NonExtOpcode = Hexagon::getBaseWithRegOffset(MI->getOpcode()); 2860 break; 2861 case HexagonII::BaseLongOffset: 2862 NonExtOpcode = Hexagon::getRegShlForm(MI->getOpcode()); 2863 break; 2864 default: 2865 return false; 2866 } 2867 if (NonExtOpcode < 0) 2868 return false; 2869 return true; 2870 } 2871 return false; 2872 } 2873 2874 2875 bool HexagonInstrInfo::hasPseudoInstrPair(const MachineInstr *MI) const { 2876 return Hexagon::getRealHWInstr(MI->getOpcode(), 2877 Hexagon::InstrType_Pseudo) >= 0; 2878 } 2879 2880 2881 bool HexagonInstrInfo::hasUncondBranch(const MachineBasicBlock *B) 2882 const { 2883 MachineBasicBlock::const_iterator I = B->getFirstTerminator(), E = B->end(); 2884 while (I != E) { 2885 if (I->isBarrier()) 2886 return true; 2887 ++I; 2888 } 2889 return false; 2890 } 2891 2892 2893 // Returns true, if a LD insn can be promoted to a cur load. 2894 bool HexagonInstrInfo::mayBeCurLoad(const MachineInstr *MI) const { 2895 auto &HST = MI->getParent()->getParent()->getSubtarget<HexagonSubtarget>(); 2896 const uint64_t F = MI->getDesc().TSFlags; 2897 return ((F >> HexagonII::mayCVLoadPos) & HexagonII::mayCVLoadMask) && 2898 HST.hasV60TOps(); 2899 } 2900 2901 2902 // Returns true, if a ST insn can be promoted to a new-value store. 2903 bool HexagonInstrInfo::mayBeNewStore(const MachineInstr *MI) const { 2904 const uint64_t F = MI->getDesc().TSFlags; 2905 return (F >> HexagonII::mayNVStorePos) & HexagonII::mayNVStoreMask; 2906 } 2907 2908 2909 bool HexagonInstrInfo::producesStall(const MachineInstr *ProdMI, 2910 const MachineInstr *ConsMI) const { 2911 // There is no stall when ProdMI is not a V60 vector. 2912 if (!isV60VectorInstruction(ProdMI)) 2913 return false; 2914 2915 // There is no stall when ProdMI and ConsMI are not dependent. 2916 if (!isDependent(ProdMI, ConsMI)) 2917 return false; 2918 2919 // When Forward Scheduling is enabled, there is no stall if ProdMI and ConsMI 2920 // are scheduled in consecutive packets. 2921 if (isVecUsableNextPacket(ProdMI, ConsMI)) 2922 return false; 2923 2924 return true; 2925 } 2926 2927 2928 bool HexagonInstrInfo::producesStall(const MachineInstr *MI, 2929 MachineBasicBlock::const_instr_iterator BII) const { 2930 // There is no stall when I is not a V60 vector. 2931 if (!isV60VectorInstruction(MI)) 2932 return false; 2933 2934 MachineBasicBlock::const_instr_iterator MII = BII; 2935 MachineBasicBlock::const_instr_iterator MIE = MII->getParent()->instr_end(); 2936 2937 if (!(*MII).isBundle()) { 2938 const MachineInstr *J = &*MII; 2939 if (!isV60VectorInstruction(J)) 2940 return false; 2941 else if (isVecUsableNextPacket(J, MI)) 2942 return false; 2943 return true; 2944 } 2945 2946 for (++MII; MII != MIE && MII->isInsideBundle(); ++MII) { 2947 const MachineInstr *J = &*MII; 2948 if (producesStall(J, MI)) 2949 return true; 2950 } 2951 return false; 2952 } 2953 2954 2955 bool HexagonInstrInfo::predCanBeUsedAsDotNew(const MachineInstr *MI, 2956 unsigned PredReg) const { 2957 for (unsigned opNum = 0; opNum < MI->getNumOperands(); opNum++) { 2958 const MachineOperand &MO = MI->getOperand(opNum); 2959 if (MO.isReg() && MO.isDef() && MO.isImplicit() && (MO.getReg() == PredReg)) 2960 return false; // Predicate register must be explicitly defined. 2961 } 2962 2963 // Hexagon Programmer's Reference says that decbin, memw_locked, and 2964 // memd_locked cannot be used as .new as well, 2965 // but we don't seem to have these instructions defined. 2966 return MI->getOpcode() != Hexagon::A4_tlbmatch; 2967 } 2968 2969 2970 bool HexagonInstrInfo::PredOpcodeHasJMP_c(unsigned Opcode) const { 2971 return (Opcode == Hexagon::J2_jumpt) || 2972 (Opcode == Hexagon::J2_jumpf) || 2973 (Opcode == Hexagon::J2_jumptnew) || 2974 (Opcode == Hexagon::J2_jumpfnew) || 2975 (Opcode == Hexagon::J2_jumptnewpt) || 2976 (Opcode == Hexagon::J2_jumpfnewpt); 2977 } 2978 2979 2980 bool HexagonInstrInfo::predOpcodeHasNot(ArrayRef<MachineOperand> Cond) const { 2981 if (Cond.empty() || !isPredicated(Cond[0].getImm())) 2982 return false; 2983 return !isPredicatedTrue(Cond[0].getImm()); 2984 } 2985 2986 2987 short HexagonInstrInfo::getAbsoluteForm(const MachineInstr *MI) const { 2988 return Hexagon::getAbsoluteForm(MI->getOpcode()); 2989 } 2990 2991 2992 unsigned HexagonInstrInfo::getAddrMode(const MachineInstr* MI) const { 2993 const uint64_t F = MI->getDesc().TSFlags; 2994 return (F >> HexagonII::AddrModePos) & HexagonII::AddrModeMask; 2995 } 2996 2997 2998 // Returns the base register in a memory access (load/store). The offset is 2999 // returned in Offset and the access size is returned in AccessSize. 3000 unsigned HexagonInstrInfo::getBaseAndOffset(const MachineInstr *MI, 3001 int &Offset, unsigned &AccessSize) const { 3002 // Return if it is not a base+offset type instruction or a MemOp. 3003 if (getAddrMode(MI) != HexagonII::BaseImmOffset && 3004 getAddrMode(MI) != HexagonII::BaseLongOffset && 3005 !isMemOp(MI) && !isPostIncrement(MI)) 3006 return 0; 3007 3008 // Since it is a memory access instruction, getMemAccessSize() should never 3009 // return 0. 3010 assert (getMemAccessSize(MI) && 3011 "BaseImmOffset or BaseLongOffset or MemOp without accessSize"); 3012 3013 // Return Values of getMemAccessSize() are 3014 // 0 - Checked in the assert above. 3015 // 1, 2, 3, 4 & 7, 8 - The statement below is correct for all these. 3016 // MemAccessSize is represented as 1+log2(N) where N is size in bits. 3017 AccessSize = (1U << (getMemAccessSize(MI) - 1)); 3018 3019 unsigned basePos = 0, offsetPos = 0; 3020 if (!getBaseAndOffsetPosition(MI, basePos, offsetPos)) 3021 return 0; 3022 3023 // Post increment updates its EA after the mem access, 3024 // so we need to treat its offset as zero. 3025 if (isPostIncrement(MI)) 3026 Offset = 0; 3027 else { 3028 Offset = MI->getOperand(offsetPos).getImm(); 3029 } 3030 3031 return MI->getOperand(basePos).getReg(); 3032 } 3033 3034 3035 /// Return the position of the base and offset operands for this instruction. 3036 bool HexagonInstrInfo::getBaseAndOffsetPosition(const MachineInstr *MI, 3037 unsigned &BasePos, unsigned &OffsetPos) const { 3038 // Deal with memops first. 3039 if (isMemOp(MI)) { 3040 assert (MI->getOperand(0).isReg() && MI->getOperand(1).isImm() && 3041 "Bad Memop."); 3042 BasePos = 0; 3043 OffsetPos = 1; 3044 } else if (MI->mayStore()) { 3045 BasePos = 0; 3046 OffsetPos = 1; 3047 } else if (MI->mayLoad()) { 3048 BasePos = 1; 3049 OffsetPos = 2; 3050 } else 3051 return false; 3052 3053 if (isPredicated(*MI)) { 3054 BasePos++; 3055 OffsetPos++; 3056 } 3057 if (isPostIncrement(MI)) { 3058 BasePos++; 3059 OffsetPos++; 3060 } 3061 3062 if (!MI->getOperand(BasePos).isReg() || !MI->getOperand(OffsetPos).isImm()) 3063 return false; 3064 3065 return true; 3066 } 3067 3068 3069 // Inserts branching instructions in reverse order of their occurence. 3070 // e.g. jump_t t1 (i1) 3071 // jump t2 (i2) 3072 // Jumpers = {i2, i1} 3073 SmallVector<MachineInstr*, 2> HexagonInstrInfo::getBranchingInstrs( 3074 MachineBasicBlock& MBB) const { 3075 SmallVector<MachineInstr*, 2> Jumpers; 3076 // If the block has no terminators, it just falls into the block after it. 3077 MachineBasicBlock::instr_iterator I = MBB.instr_end(); 3078 if (I == MBB.instr_begin()) 3079 return Jumpers; 3080 3081 // A basic block may looks like this: 3082 // 3083 // [ insn 3084 // EH_LABEL 3085 // insn 3086 // insn 3087 // insn 3088 // EH_LABEL 3089 // insn ] 3090 // 3091 // It has two succs but does not have a terminator 3092 // Don't know how to handle it. 3093 do { 3094 --I; 3095 if (I->isEHLabel()) 3096 return Jumpers; 3097 } while (I != MBB.instr_begin()); 3098 3099 I = MBB.instr_end(); 3100 --I; 3101 3102 while (I->isDebugValue()) { 3103 if (I == MBB.instr_begin()) 3104 return Jumpers; 3105 --I; 3106 } 3107 if (!isUnpredicatedTerminator(*I)) 3108 return Jumpers; 3109 3110 // Get the last instruction in the block. 3111 MachineInstr *LastInst = &*I; 3112 Jumpers.push_back(LastInst); 3113 MachineInstr *SecondLastInst = nullptr; 3114 // Find one more terminator if present. 3115 do { 3116 if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(*I)) { 3117 if (!SecondLastInst) { 3118 SecondLastInst = &*I; 3119 Jumpers.push_back(SecondLastInst); 3120 } else // This is a third branch. 3121 return Jumpers; 3122 } 3123 if (I == MBB.instr_begin()) 3124 break; 3125 --I; 3126 } while (true); 3127 return Jumpers; 3128 } 3129 3130 3131 short HexagonInstrInfo::getBaseWithLongOffset(short Opcode) const { 3132 if (Opcode < 0) 3133 return -1; 3134 return Hexagon::getBaseWithLongOffset(Opcode); 3135 } 3136 3137 3138 short HexagonInstrInfo::getBaseWithLongOffset(const MachineInstr *MI) const { 3139 return Hexagon::getBaseWithLongOffset(MI->getOpcode()); 3140 } 3141 3142 3143 short HexagonInstrInfo::getBaseWithRegOffset(const MachineInstr *MI) const { 3144 return Hexagon::getBaseWithRegOffset(MI->getOpcode()); 3145 } 3146 3147 3148 // Returns Operand Index for the constant extended instruction. 3149 unsigned HexagonInstrInfo::getCExtOpNum(const MachineInstr *MI) const { 3150 const uint64_t F = MI->getDesc().TSFlags; 3151 return (F >> HexagonII::ExtendableOpPos) & HexagonII::ExtendableOpMask; 3152 } 3153 3154 // See if instruction could potentially be a duplex candidate. 3155 // If so, return its group. Zero otherwise. 3156 HexagonII::CompoundGroup HexagonInstrInfo::getCompoundCandidateGroup( 3157 const MachineInstr *MI) const { 3158 unsigned DstReg, SrcReg, Src1Reg, Src2Reg; 3159 3160 switch (MI->getOpcode()) { 3161 default: 3162 return HexagonII::HCG_None; 3163 // 3164 // Compound pairs. 3165 // "p0=cmp.eq(Rs16,Rt16); if (p0.new) jump:nt #r9:2" 3166 // "Rd16=#U6 ; jump #r9:2" 3167 // "Rd16=Rs16 ; jump #r9:2" 3168 // 3169 case Hexagon::C2_cmpeq: 3170 case Hexagon::C2_cmpgt: 3171 case Hexagon::C2_cmpgtu: 3172 DstReg = MI->getOperand(0).getReg(); 3173 Src1Reg = MI->getOperand(1).getReg(); 3174 Src2Reg = MI->getOperand(2).getReg(); 3175 if (Hexagon::PredRegsRegClass.contains(DstReg) && 3176 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) && 3177 isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg)) 3178 return HexagonII::HCG_A; 3179 break; 3180 case Hexagon::C2_cmpeqi: 3181 case Hexagon::C2_cmpgti: 3182 case Hexagon::C2_cmpgtui: 3183 // P0 = cmp.eq(Rs,#u2) 3184 DstReg = MI->getOperand(0).getReg(); 3185 SrcReg = MI->getOperand(1).getReg(); 3186 if (Hexagon::PredRegsRegClass.contains(DstReg) && 3187 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) && 3188 isIntRegForSubInst(SrcReg) && MI->getOperand(2).isImm() && 3189 ((isUInt<5>(MI->getOperand(2).getImm())) || 3190 (MI->getOperand(2).getImm() == -1))) 3191 return HexagonII::HCG_A; 3192 break; 3193 case Hexagon::A2_tfr: 3194 // Rd = Rs 3195 DstReg = MI->getOperand(0).getReg(); 3196 SrcReg = MI->getOperand(1).getReg(); 3197 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg)) 3198 return HexagonII::HCG_A; 3199 break; 3200 case Hexagon::A2_tfrsi: 3201 // Rd = #u6 3202 // Do not test for #u6 size since the const is getting extended 3203 // regardless and compound could be formed. 3204 DstReg = MI->getOperand(0).getReg(); 3205 if (isIntRegForSubInst(DstReg)) 3206 return HexagonII::HCG_A; 3207 break; 3208 case Hexagon::S2_tstbit_i: 3209 DstReg = MI->getOperand(0).getReg(); 3210 Src1Reg = MI->getOperand(1).getReg(); 3211 if (Hexagon::PredRegsRegClass.contains(DstReg) && 3212 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) && 3213 MI->getOperand(2).isImm() && 3214 isIntRegForSubInst(Src1Reg) && (MI->getOperand(2).getImm() == 0)) 3215 return HexagonII::HCG_A; 3216 break; 3217 // The fact that .new form is used pretty much guarantees 3218 // that predicate register will match. Nevertheless, 3219 // there could be some false positives without additional 3220 // checking. 3221 case Hexagon::J2_jumptnew: 3222 case Hexagon::J2_jumpfnew: 3223 case Hexagon::J2_jumptnewpt: 3224 case Hexagon::J2_jumpfnewpt: 3225 Src1Reg = MI->getOperand(0).getReg(); 3226 if (Hexagon::PredRegsRegClass.contains(Src1Reg) && 3227 (Hexagon::P0 == Src1Reg || Hexagon::P1 == Src1Reg)) 3228 return HexagonII::HCG_B; 3229 break; 3230 // Transfer and jump: 3231 // Rd=#U6 ; jump #r9:2 3232 // Rd=Rs ; jump #r9:2 3233 // Do not test for jump range here. 3234 case Hexagon::J2_jump: 3235 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4: 3236 return HexagonII::HCG_C; 3237 break; 3238 } 3239 3240 return HexagonII::HCG_None; 3241 } 3242 3243 3244 // Returns -1 when there is no opcode found. 3245 unsigned HexagonInstrInfo::getCompoundOpcode(const MachineInstr *GA, 3246 const MachineInstr *GB) const { 3247 assert(getCompoundCandidateGroup(GA) == HexagonII::HCG_A); 3248 assert(getCompoundCandidateGroup(GB) == HexagonII::HCG_B); 3249 if ((GA->getOpcode() != Hexagon::C2_cmpeqi) || 3250 (GB->getOpcode() != Hexagon::J2_jumptnew)) 3251 return -1; 3252 unsigned DestReg = GA->getOperand(0).getReg(); 3253 if (!GB->readsRegister(DestReg)) 3254 return -1; 3255 if (DestReg == Hexagon::P0) 3256 return Hexagon::J4_cmpeqi_tp0_jump_nt; 3257 if (DestReg == Hexagon::P1) 3258 return Hexagon::J4_cmpeqi_tp1_jump_nt; 3259 return -1; 3260 } 3261 3262 3263 int HexagonInstrInfo::getCondOpcode(int Opc, bool invertPredicate) const { 3264 enum Hexagon::PredSense inPredSense; 3265 inPredSense = invertPredicate ? Hexagon::PredSense_false : 3266 Hexagon::PredSense_true; 3267 int CondOpcode = Hexagon::getPredOpcode(Opc, inPredSense); 3268 if (CondOpcode >= 0) // Valid Conditional opcode/instruction 3269 return CondOpcode; 3270 3271 // This switch case will be removed once all the instructions have been 3272 // modified to use relation maps. 3273 switch(Opc) { 3274 case Hexagon::TFRI_f: 3275 return !invertPredicate ? Hexagon::TFRI_cPt_f : 3276 Hexagon::TFRI_cNotPt_f; 3277 } 3278 3279 llvm_unreachable("Unexpected predicable instruction"); 3280 } 3281 3282 3283 // Return the cur value instruction for a given store. 3284 int HexagonInstrInfo::getDotCurOp(const MachineInstr* MI) const { 3285 switch (MI->getOpcode()) { 3286 default: llvm_unreachable("Unknown .cur type"); 3287 case Hexagon::V6_vL32b_pi: 3288 return Hexagon::V6_vL32b_cur_pi; 3289 case Hexagon::V6_vL32b_ai: 3290 return Hexagon::V6_vL32b_cur_ai; 3291 //128B 3292 case Hexagon::V6_vL32b_pi_128B: 3293 return Hexagon::V6_vL32b_cur_pi_128B; 3294 case Hexagon::V6_vL32b_ai_128B: 3295 return Hexagon::V6_vL32b_cur_ai_128B; 3296 } 3297 return 0; 3298 } 3299 3300 3301 3302 // The diagram below shows the steps involved in the conversion of a predicated 3303 // store instruction to its .new predicated new-value form. 3304 // 3305 // p.new NV store [ if(p0.new)memw(R0+#0)=R2.new ] 3306 // ^ ^ 3307 // / \ (not OK. it will cause new-value store to be 3308 // / X conditional on p0.new while R2 producer is 3309 // / \ on p0) 3310 // / \. 3311 // p.new store p.old NV store 3312 // [if(p0.new)memw(R0+#0)=R2] [if(p0)memw(R0+#0)=R2.new] 3313 // ^ ^ 3314 // \ / 3315 // \ / 3316 // \ / 3317 // p.old store 3318 // [if (p0)memw(R0+#0)=R2] 3319 // 3320 // 3321 // The following set of instructions further explains the scenario where 3322 // conditional new-value store becomes invalid when promoted to .new predicate 3323 // form. 3324 // 3325 // { 1) if (p0) r0 = add(r1, r2) 3326 // 2) p0 = cmp.eq(r3, #0) } 3327 // 3328 // 3) if (p0) memb(r1+#0) = r0 --> this instruction can't be grouped with 3329 // the first two instructions because in instr 1, r0 is conditional on old value 3330 // of p0 but its use in instr 3 is conditional on p0 modified by instr 2 which 3331 // is not valid for new-value stores. 3332 // Predicated new value stores (i.e. if (p0) memw(..)=r0.new) are excluded 3333 // from the "Conditional Store" list. Because a predicated new value store 3334 // would NOT be promoted to a double dot new store. See diagram below: 3335 // This function returns yes for those stores that are predicated but not 3336 // yet promoted to predicate dot new instructions. 3337 // 3338 // +---------------------+ 3339 // /-----| if (p0) memw(..)=r0 |---------\~ 3340 // || +---------------------+ || 3341 // promote || /\ /\ || promote 3342 // || /||\ /||\ || 3343 // \||/ demote || \||/ 3344 // \/ || || \/ 3345 // +-------------------------+ || +-------------------------+ 3346 // | if (p0.new) memw(..)=r0 | || | if (p0) memw(..)=r0.new | 3347 // +-------------------------+ || +-------------------------+ 3348 // || || || 3349 // || demote \||/ 3350 // promote || \/ NOT possible 3351 // || || /\~ 3352 // \||/ || /||\~ 3353 // \/ || || 3354 // +-----------------------------+ 3355 // | if (p0.new) memw(..)=r0.new | 3356 // +-----------------------------+ 3357 // Double Dot New Store 3358 // 3359 // Returns the most basic instruction for the .new predicated instructions and 3360 // new-value stores. 3361 // For example, all of the following instructions will be converted back to the 3362 // same instruction: 3363 // 1) if (p0.new) memw(R0+#0) = R1.new ---> 3364 // 2) if (p0) memw(R0+#0)= R1.new -------> if (p0) memw(R0+#0) = R1 3365 // 3) if (p0.new) memw(R0+#0) = R1 ---> 3366 // 3367 // To understand the translation of instruction 1 to its original form, consider 3368 // a packet with 3 instructions. 3369 // { p0 = cmp.eq(R0,R1) 3370 // if (p0.new) R2 = add(R3, R4) 3371 // R5 = add (R3, R1) 3372 // } 3373 // if (p0) memw(R5+#0) = R2 <--- trying to include it in the previous packet 3374 // 3375 // This instruction can be part of the previous packet only if both p0 and R2 3376 // are promoted to .new values. This promotion happens in steps, first 3377 // predicate register is promoted to .new and in the next iteration R2 is 3378 // promoted. Therefore, in case of dependence check failure (due to R5) during 3379 // next iteration, it should be converted back to its most basic form. 3380 3381 3382 // Return the new value instruction for a given store. 3383 int HexagonInstrInfo::getDotNewOp(const MachineInstr* MI) const { 3384 int NVOpcode = Hexagon::getNewValueOpcode(MI->getOpcode()); 3385 if (NVOpcode >= 0) // Valid new-value store instruction. 3386 return NVOpcode; 3387 3388 switch (MI->getOpcode()) { 3389 default: llvm_unreachable("Unknown .new type"); 3390 case Hexagon::S4_storerb_ur: 3391 return Hexagon::S4_storerbnew_ur; 3392 3393 case Hexagon::S2_storerb_pci: 3394 return Hexagon::S2_storerb_pci; 3395 3396 case Hexagon::S2_storeri_pci: 3397 return Hexagon::S2_storeri_pci; 3398 3399 case Hexagon::S2_storerh_pci: 3400 return Hexagon::S2_storerh_pci; 3401 3402 case Hexagon::S2_storerd_pci: 3403 return Hexagon::S2_storerd_pci; 3404 3405 case Hexagon::S2_storerf_pci: 3406 return Hexagon::S2_storerf_pci; 3407 3408 case Hexagon::V6_vS32b_ai: 3409 return Hexagon::V6_vS32b_new_ai; 3410 3411 case Hexagon::V6_vS32b_pi: 3412 return Hexagon::V6_vS32b_new_pi; 3413 3414 // 128B 3415 case Hexagon::V6_vS32b_ai_128B: 3416 return Hexagon::V6_vS32b_new_ai_128B; 3417 3418 case Hexagon::V6_vS32b_pi_128B: 3419 return Hexagon::V6_vS32b_new_pi_128B; 3420 } 3421 return 0; 3422 } 3423 3424 3425 // Returns the opcode to use when converting MI, which is a conditional jump, 3426 // into a conditional instruction which uses the .new value of the predicate. 3427 // We also use branch probabilities to add a hint to the jump. 3428 int HexagonInstrInfo::getDotNewPredJumpOp(const MachineInstr *MI, 3429 const MachineBranchProbabilityInfo *MBPI) const { 3430 // We assume that block can have at most two successors. 3431 bool taken = false; 3432 const MachineBasicBlock *Src = MI->getParent(); 3433 const MachineOperand *BrTarget = &MI->getOperand(1); 3434 const MachineBasicBlock *Dst = BrTarget->getMBB(); 3435 3436 const BranchProbability Prediction = MBPI->getEdgeProbability(Src, Dst); 3437 if (Prediction >= BranchProbability(1,2)) 3438 taken = true; 3439 3440 switch (MI->getOpcode()) { 3441 case Hexagon::J2_jumpt: 3442 return taken ? Hexagon::J2_jumptnewpt : Hexagon::J2_jumptnew; 3443 case Hexagon::J2_jumpf: 3444 return taken ? Hexagon::J2_jumpfnewpt : Hexagon::J2_jumpfnew; 3445 3446 default: 3447 llvm_unreachable("Unexpected jump instruction."); 3448 } 3449 } 3450 3451 3452 // Return .new predicate version for an instruction. 3453 int HexagonInstrInfo::getDotNewPredOp(const MachineInstr *MI, 3454 const MachineBranchProbabilityInfo *MBPI) const { 3455 int NewOpcode = Hexagon::getPredNewOpcode(MI->getOpcode()); 3456 if (NewOpcode >= 0) // Valid predicate new instruction 3457 return NewOpcode; 3458 3459 switch (MI->getOpcode()) { 3460 // Condtional Jumps 3461 case Hexagon::J2_jumpt: 3462 case Hexagon::J2_jumpf: 3463 return getDotNewPredJumpOp(MI, MBPI); 3464 3465 default: 3466 assert(0 && "Unknown .new type"); 3467 } 3468 return 0; 3469 } 3470 3471 3472 int HexagonInstrInfo::getDotOldOp(const int opc) const { 3473 int NewOp = opc; 3474 if (isPredicated(NewOp) && isPredicatedNew(NewOp)) { // Get predicate old form 3475 NewOp = Hexagon::getPredOldOpcode(NewOp); 3476 assert(NewOp >= 0 && 3477 "Couldn't change predicate new instruction to its old form."); 3478 } 3479 3480 if (isNewValueStore(NewOp)) { // Convert into non-new-value format 3481 NewOp = Hexagon::getNonNVStore(NewOp); 3482 assert(NewOp >= 0 && "Couldn't change new-value store to its old form."); 3483 } 3484 return NewOp; 3485 } 3486 3487 3488 // See if instruction could potentially be a duplex candidate. 3489 // If so, return its group. Zero otherwise. 3490 HexagonII::SubInstructionGroup HexagonInstrInfo::getDuplexCandidateGroup( 3491 const MachineInstr *MI) const { 3492 unsigned DstReg, SrcReg, Src1Reg, Src2Reg; 3493 auto &HRI = getRegisterInfo(); 3494 3495 switch (MI->getOpcode()) { 3496 default: 3497 return HexagonII::HSIG_None; 3498 // 3499 // Group L1: 3500 // 3501 // Rd = memw(Rs+#u4:2) 3502 // Rd = memub(Rs+#u4:0) 3503 case Hexagon::L2_loadri_io: 3504 DstReg = MI->getOperand(0).getReg(); 3505 SrcReg = MI->getOperand(1).getReg(); 3506 // Special case this one from Group L2. 3507 // Rd = memw(r29+#u5:2) 3508 if (isIntRegForSubInst(DstReg)) { 3509 if (Hexagon::IntRegsRegClass.contains(SrcReg) && 3510 HRI.getStackRegister() == SrcReg && 3511 MI->getOperand(2).isImm() && 3512 isShiftedUInt<5,2>(MI->getOperand(2).getImm())) 3513 return HexagonII::HSIG_L2; 3514 // Rd = memw(Rs+#u4:2) 3515 if (isIntRegForSubInst(SrcReg) && 3516 (MI->getOperand(2).isImm() && 3517 isShiftedUInt<4,2>(MI->getOperand(2).getImm()))) 3518 return HexagonII::HSIG_L1; 3519 } 3520 break; 3521 case Hexagon::L2_loadrub_io: 3522 // Rd = memub(Rs+#u4:0) 3523 DstReg = MI->getOperand(0).getReg(); 3524 SrcReg = MI->getOperand(1).getReg(); 3525 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 3526 MI->getOperand(2).isImm() && isUInt<4>(MI->getOperand(2).getImm())) 3527 return HexagonII::HSIG_L1; 3528 break; 3529 // 3530 // Group L2: 3531 // 3532 // Rd = memh/memuh(Rs+#u3:1) 3533 // Rd = memb(Rs+#u3:0) 3534 // Rd = memw(r29+#u5:2) - Handled above. 3535 // Rdd = memd(r29+#u5:3) 3536 // deallocframe 3537 // [if ([!]p0[.new])] dealloc_return 3538 // [if ([!]p0[.new])] jumpr r31 3539 case Hexagon::L2_loadrh_io: 3540 case Hexagon::L2_loadruh_io: 3541 // Rd = memh/memuh(Rs+#u3:1) 3542 DstReg = MI->getOperand(0).getReg(); 3543 SrcReg = MI->getOperand(1).getReg(); 3544 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 3545 MI->getOperand(2).isImm() && 3546 isShiftedUInt<3,1>(MI->getOperand(2).getImm())) 3547 return HexagonII::HSIG_L2; 3548 break; 3549 case Hexagon::L2_loadrb_io: 3550 // Rd = memb(Rs+#u3:0) 3551 DstReg = MI->getOperand(0).getReg(); 3552 SrcReg = MI->getOperand(1).getReg(); 3553 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 3554 MI->getOperand(2).isImm() && 3555 isUInt<3>(MI->getOperand(2).getImm())) 3556 return HexagonII::HSIG_L2; 3557 break; 3558 case Hexagon::L2_loadrd_io: 3559 // Rdd = memd(r29+#u5:3) 3560 DstReg = MI->getOperand(0).getReg(); 3561 SrcReg = MI->getOperand(1).getReg(); 3562 if (isDblRegForSubInst(DstReg, HRI) && 3563 Hexagon::IntRegsRegClass.contains(SrcReg) && 3564 HRI.getStackRegister() == SrcReg && 3565 MI->getOperand(2).isImm() && 3566 isShiftedUInt<5,3>(MI->getOperand(2).getImm())) 3567 return HexagonII::HSIG_L2; 3568 break; 3569 // dealloc_return is not documented in Hexagon Manual, but marked 3570 // with A_SUBINSN attribute in iset_v4classic.py. 3571 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4: 3572 case Hexagon::L4_return: 3573 case Hexagon::L2_deallocframe: 3574 return HexagonII::HSIG_L2; 3575 case Hexagon::EH_RETURN_JMPR: 3576 case Hexagon::JMPret : 3577 // jumpr r31 3578 // Actual form JMPR %PC<imp-def>, %R31<imp-use>, %R0<imp-use,internal>. 3579 DstReg = MI->getOperand(0).getReg(); 3580 if (Hexagon::IntRegsRegClass.contains(DstReg) && (Hexagon::R31 == DstReg)) 3581 return HexagonII::HSIG_L2; 3582 break; 3583 case Hexagon::JMPrett: 3584 case Hexagon::JMPretf: 3585 case Hexagon::JMPrettnewpt: 3586 case Hexagon::JMPretfnewpt : 3587 case Hexagon::JMPrettnew : 3588 case Hexagon::JMPretfnew : 3589 DstReg = MI->getOperand(1).getReg(); 3590 SrcReg = MI->getOperand(0).getReg(); 3591 // [if ([!]p0[.new])] jumpr r31 3592 if ((Hexagon::PredRegsRegClass.contains(SrcReg) && 3593 (Hexagon::P0 == SrcReg)) && 3594 (Hexagon::IntRegsRegClass.contains(DstReg) && (Hexagon::R31 == DstReg))) 3595 return HexagonII::HSIG_L2; 3596 break; 3597 case Hexagon::L4_return_t : 3598 case Hexagon::L4_return_f : 3599 case Hexagon::L4_return_tnew_pnt : 3600 case Hexagon::L4_return_fnew_pnt : 3601 case Hexagon::L4_return_tnew_pt : 3602 case Hexagon::L4_return_fnew_pt : 3603 // [if ([!]p0[.new])] dealloc_return 3604 SrcReg = MI->getOperand(0).getReg(); 3605 if (Hexagon::PredRegsRegClass.contains(SrcReg) && (Hexagon::P0 == SrcReg)) 3606 return HexagonII::HSIG_L2; 3607 break; 3608 // 3609 // Group S1: 3610 // 3611 // memw(Rs+#u4:2) = Rt 3612 // memb(Rs+#u4:0) = Rt 3613 case Hexagon::S2_storeri_io: 3614 // Special case this one from Group S2. 3615 // memw(r29+#u5:2) = Rt 3616 Src1Reg = MI->getOperand(0).getReg(); 3617 Src2Reg = MI->getOperand(2).getReg(); 3618 if (Hexagon::IntRegsRegClass.contains(Src1Reg) && 3619 isIntRegForSubInst(Src2Reg) && 3620 HRI.getStackRegister() == Src1Reg && MI->getOperand(1).isImm() && 3621 isShiftedUInt<5,2>(MI->getOperand(1).getImm())) 3622 return HexagonII::HSIG_S2; 3623 // memw(Rs+#u4:2) = Rt 3624 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) && 3625 MI->getOperand(1).isImm() && 3626 isShiftedUInt<4,2>(MI->getOperand(1).getImm())) 3627 return HexagonII::HSIG_S1; 3628 break; 3629 case Hexagon::S2_storerb_io: 3630 // memb(Rs+#u4:0) = Rt 3631 Src1Reg = MI->getOperand(0).getReg(); 3632 Src2Reg = MI->getOperand(2).getReg(); 3633 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) && 3634 MI->getOperand(1).isImm() && isUInt<4>(MI->getOperand(1).getImm())) 3635 return HexagonII::HSIG_S1; 3636 break; 3637 // 3638 // Group S2: 3639 // 3640 // memh(Rs+#u3:1) = Rt 3641 // memw(r29+#u5:2) = Rt 3642 // memd(r29+#s6:3) = Rtt 3643 // memw(Rs+#u4:2) = #U1 3644 // memb(Rs+#u4) = #U1 3645 // allocframe(#u5:3) 3646 case Hexagon::S2_storerh_io: 3647 // memh(Rs+#u3:1) = Rt 3648 Src1Reg = MI->getOperand(0).getReg(); 3649 Src2Reg = MI->getOperand(2).getReg(); 3650 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) && 3651 MI->getOperand(1).isImm() && 3652 isShiftedUInt<3,1>(MI->getOperand(1).getImm())) 3653 return HexagonII::HSIG_S1; 3654 break; 3655 case Hexagon::S2_storerd_io: 3656 // memd(r29+#s6:3) = Rtt 3657 Src1Reg = MI->getOperand(0).getReg(); 3658 Src2Reg = MI->getOperand(2).getReg(); 3659 if (isDblRegForSubInst(Src2Reg, HRI) && 3660 Hexagon::IntRegsRegClass.contains(Src1Reg) && 3661 HRI.getStackRegister() == Src1Reg && MI->getOperand(1).isImm() && 3662 isShiftedInt<6,3>(MI->getOperand(1).getImm())) 3663 return HexagonII::HSIG_S2; 3664 break; 3665 case Hexagon::S4_storeiri_io: 3666 // memw(Rs+#u4:2) = #U1 3667 Src1Reg = MI->getOperand(0).getReg(); 3668 if (isIntRegForSubInst(Src1Reg) && MI->getOperand(1).isImm() && 3669 isShiftedUInt<4,2>(MI->getOperand(1).getImm()) && 3670 MI->getOperand(2).isImm() && isUInt<1>(MI->getOperand(2).getImm())) 3671 return HexagonII::HSIG_S2; 3672 break; 3673 case Hexagon::S4_storeirb_io: 3674 // memb(Rs+#u4) = #U1 3675 Src1Reg = MI->getOperand(0).getReg(); 3676 if (isIntRegForSubInst(Src1Reg) && 3677 MI->getOperand(1).isImm() && isUInt<4>(MI->getOperand(1).getImm()) && 3678 MI->getOperand(2).isImm() && isUInt<1>(MI->getOperand(2).getImm())) 3679 return HexagonII::HSIG_S2; 3680 break; 3681 case Hexagon::S2_allocframe: 3682 if (MI->getOperand(0).isImm() && 3683 isShiftedUInt<5,3>(MI->getOperand(0).getImm())) 3684 return HexagonII::HSIG_S1; 3685 break; 3686 // 3687 // Group A: 3688 // 3689 // Rx = add(Rx,#s7) 3690 // Rd = Rs 3691 // Rd = #u6 3692 // Rd = #-1 3693 // if ([!]P0[.new]) Rd = #0 3694 // Rd = add(r29,#u6:2) 3695 // Rx = add(Rx,Rs) 3696 // P0 = cmp.eq(Rs,#u2) 3697 // Rdd = combine(#0,Rs) 3698 // Rdd = combine(Rs,#0) 3699 // Rdd = combine(#u2,#U2) 3700 // Rd = add(Rs,#1) 3701 // Rd = add(Rs,#-1) 3702 // Rd = sxth/sxtb/zxtb/zxth(Rs) 3703 // Rd = and(Rs,#1) 3704 case Hexagon::A2_addi: 3705 DstReg = MI->getOperand(0).getReg(); 3706 SrcReg = MI->getOperand(1).getReg(); 3707 if (isIntRegForSubInst(DstReg)) { 3708 // Rd = add(r29,#u6:2) 3709 if (Hexagon::IntRegsRegClass.contains(SrcReg) && 3710 HRI.getStackRegister() == SrcReg && MI->getOperand(2).isImm() && 3711 isShiftedUInt<6,2>(MI->getOperand(2).getImm())) 3712 return HexagonII::HSIG_A; 3713 // Rx = add(Rx,#s7) 3714 if ((DstReg == SrcReg) && MI->getOperand(2).isImm() && 3715 isInt<7>(MI->getOperand(2).getImm())) 3716 return HexagonII::HSIG_A; 3717 // Rd = add(Rs,#1) 3718 // Rd = add(Rs,#-1) 3719 if (isIntRegForSubInst(SrcReg) && MI->getOperand(2).isImm() && 3720 ((MI->getOperand(2).getImm() == 1) || 3721 (MI->getOperand(2).getImm() == -1))) 3722 return HexagonII::HSIG_A; 3723 } 3724 break; 3725 case Hexagon::A2_add: 3726 // Rx = add(Rx,Rs) 3727 DstReg = MI->getOperand(0).getReg(); 3728 Src1Reg = MI->getOperand(1).getReg(); 3729 Src2Reg = MI->getOperand(2).getReg(); 3730 if (isIntRegForSubInst(DstReg) && (DstReg == Src1Reg) && 3731 isIntRegForSubInst(Src2Reg)) 3732 return HexagonII::HSIG_A; 3733 break; 3734 case Hexagon::A2_andir: 3735 // Same as zxtb. 3736 // Rd16=and(Rs16,#255) 3737 // Rd16=and(Rs16,#1) 3738 DstReg = MI->getOperand(0).getReg(); 3739 SrcReg = MI->getOperand(1).getReg(); 3740 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 3741 MI->getOperand(2).isImm() && 3742 ((MI->getOperand(2).getImm() == 1) || 3743 (MI->getOperand(2).getImm() == 255))) 3744 return HexagonII::HSIG_A; 3745 break; 3746 case Hexagon::A2_tfr: 3747 // Rd = Rs 3748 DstReg = MI->getOperand(0).getReg(); 3749 SrcReg = MI->getOperand(1).getReg(); 3750 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg)) 3751 return HexagonII::HSIG_A; 3752 break; 3753 case Hexagon::A2_tfrsi: 3754 // Rd = #u6 3755 // Do not test for #u6 size since the const is getting extended 3756 // regardless and compound could be formed. 3757 // Rd = #-1 3758 DstReg = MI->getOperand(0).getReg(); 3759 if (isIntRegForSubInst(DstReg)) 3760 return HexagonII::HSIG_A; 3761 break; 3762 case Hexagon::C2_cmoveit: 3763 case Hexagon::C2_cmovenewit: 3764 case Hexagon::C2_cmoveif: 3765 case Hexagon::C2_cmovenewif: 3766 // if ([!]P0[.new]) Rd = #0 3767 // Actual form: 3768 // %R16<def> = C2_cmovenewit %P0<internal>, 0, %R16<imp-use,undef>; 3769 DstReg = MI->getOperand(0).getReg(); 3770 SrcReg = MI->getOperand(1).getReg(); 3771 if (isIntRegForSubInst(DstReg) && 3772 Hexagon::PredRegsRegClass.contains(SrcReg) && Hexagon::P0 == SrcReg && 3773 MI->getOperand(2).isImm() && MI->getOperand(2).getImm() == 0) 3774 return HexagonII::HSIG_A; 3775 break; 3776 case Hexagon::C2_cmpeqi: 3777 // P0 = cmp.eq(Rs,#u2) 3778 DstReg = MI->getOperand(0).getReg(); 3779 SrcReg = MI->getOperand(1).getReg(); 3780 if (Hexagon::PredRegsRegClass.contains(DstReg) && 3781 Hexagon::P0 == DstReg && isIntRegForSubInst(SrcReg) && 3782 MI->getOperand(2).isImm() && isUInt<2>(MI->getOperand(2).getImm())) 3783 return HexagonII::HSIG_A; 3784 break; 3785 case Hexagon::A2_combineii: 3786 case Hexagon::A4_combineii: 3787 // Rdd = combine(#u2,#U2) 3788 DstReg = MI->getOperand(0).getReg(); 3789 if (isDblRegForSubInst(DstReg, HRI) && 3790 ((MI->getOperand(1).isImm() && isUInt<2>(MI->getOperand(1).getImm())) || 3791 (MI->getOperand(1).isGlobal() && 3792 isUInt<2>(MI->getOperand(1).getOffset()))) && 3793 ((MI->getOperand(2).isImm() && isUInt<2>(MI->getOperand(2).getImm())) || 3794 (MI->getOperand(2).isGlobal() && 3795 isUInt<2>(MI->getOperand(2).getOffset())))) 3796 return HexagonII::HSIG_A; 3797 break; 3798 case Hexagon::A4_combineri: 3799 // Rdd = combine(Rs,#0) 3800 DstReg = MI->getOperand(0).getReg(); 3801 SrcReg = MI->getOperand(1).getReg(); 3802 if (isDblRegForSubInst(DstReg, HRI) && isIntRegForSubInst(SrcReg) && 3803 ((MI->getOperand(2).isImm() && MI->getOperand(2).getImm() == 0) || 3804 (MI->getOperand(2).isGlobal() && MI->getOperand(2).getOffset() == 0))) 3805 return HexagonII::HSIG_A; 3806 break; 3807 case Hexagon::A4_combineir: 3808 // Rdd = combine(#0,Rs) 3809 DstReg = MI->getOperand(0).getReg(); 3810 SrcReg = MI->getOperand(2).getReg(); 3811 if (isDblRegForSubInst(DstReg, HRI) && isIntRegForSubInst(SrcReg) && 3812 ((MI->getOperand(1).isImm() && MI->getOperand(1).getImm() == 0) || 3813 (MI->getOperand(1).isGlobal() && MI->getOperand(1).getOffset() == 0))) 3814 return HexagonII::HSIG_A; 3815 break; 3816 case Hexagon::A2_sxtb: 3817 case Hexagon::A2_sxth: 3818 case Hexagon::A2_zxtb: 3819 case Hexagon::A2_zxth: 3820 // Rd = sxth/sxtb/zxtb/zxth(Rs) 3821 DstReg = MI->getOperand(0).getReg(); 3822 SrcReg = MI->getOperand(1).getReg(); 3823 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg)) 3824 return HexagonII::HSIG_A; 3825 break; 3826 } 3827 3828 return HexagonII::HSIG_None; 3829 } 3830 3831 3832 short HexagonInstrInfo::getEquivalentHWInstr(const MachineInstr *MI) const { 3833 return Hexagon::getRealHWInstr(MI->getOpcode(), Hexagon::InstrType_Real); 3834 } 3835 3836 3837 // Return first non-debug instruction in the basic block. 3838 MachineInstr *HexagonInstrInfo::getFirstNonDbgInst(MachineBasicBlock *BB) 3839 const { 3840 for (auto MII = BB->instr_begin(), End = BB->instr_end(); MII != End; MII++) { 3841 MachineInstr *MI = &*MII; 3842 if (MI->isDebugValue()) 3843 continue; 3844 return MI; 3845 } 3846 return nullptr; 3847 } 3848 3849 3850 unsigned HexagonInstrInfo::getInstrTimingClassLatency( 3851 const InstrItineraryData *ItinData, const MachineInstr *MI) const { 3852 // Default to one cycle for no itinerary. However, an "empty" itinerary may 3853 // still have a MinLatency property, which getStageLatency checks. 3854 if (!ItinData) 3855 return getInstrLatency(ItinData, MI); 3856 3857 // Get the latency embedded in the itinerary. If we're not using timing class 3858 // latencies or if we using BSB scheduling, then restrict the maximum latency 3859 // to 1 (that is, either 0 or 1). 3860 if (MI->isTransient()) 3861 return 0; 3862 unsigned Latency = ItinData->getStageLatency(MI->getDesc().getSchedClass()); 3863 if (!EnableTimingClassLatency || 3864 MI->getParent()->getParent()->getSubtarget<HexagonSubtarget>(). 3865 useBSBScheduling()) 3866 if (Latency > 1) 3867 Latency = 1; 3868 return Latency; 3869 } 3870 3871 3872 // inverts the predication logic. 3873 // p -> NotP 3874 // NotP -> P 3875 bool HexagonInstrInfo::getInvertedPredSense( 3876 SmallVectorImpl<MachineOperand> &Cond) const { 3877 if (Cond.empty()) 3878 return false; 3879 unsigned Opc = getInvertedPredicatedOpcode(Cond[0].getImm()); 3880 Cond[0].setImm(Opc); 3881 return true; 3882 } 3883 3884 3885 unsigned HexagonInstrInfo::getInvertedPredicatedOpcode(const int Opc) const { 3886 int InvPredOpcode; 3887 InvPredOpcode = isPredicatedTrue(Opc) ? Hexagon::getFalsePredOpcode(Opc) 3888 : Hexagon::getTruePredOpcode(Opc); 3889 if (InvPredOpcode >= 0) // Valid instruction with the inverted predicate. 3890 return InvPredOpcode; 3891 3892 llvm_unreachable("Unexpected predicated instruction"); 3893 } 3894 3895 3896 // Returns the max value that doesn't need to be extended. 3897 int HexagonInstrInfo::getMaxValue(const MachineInstr *MI) const { 3898 const uint64_t F = MI->getDesc().TSFlags; 3899 unsigned isSigned = (F >> HexagonII::ExtentSignedPos) 3900 & HexagonII::ExtentSignedMask; 3901 unsigned bits = (F >> HexagonII::ExtentBitsPos) 3902 & HexagonII::ExtentBitsMask; 3903 3904 if (isSigned) // if value is signed 3905 return ~(-1U << (bits - 1)); 3906 else 3907 return ~(-1U << bits); 3908 } 3909 3910 3911 unsigned HexagonInstrInfo::getMemAccessSize(const MachineInstr* MI) const { 3912 const uint64_t F = MI->getDesc().TSFlags; 3913 return (F >> HexagonII::MemAccessSizePos) & HexagonII::MemAccesSizeMask; 3914 } 3915 3916 3917 // Returns the min value that doesn't need to be extended. 3918 int HexagonInstrInfo::getMinValue(const MachineInstr *MI) const { 3919 const uint64_t F = MI->getDesc().TSFlags; 3920 unsigned isSigned = (F >> HexagonII::ExtentSignedPos) 3921 & HexagonII::ExtentSignedMask; 3922 unsigned bits = (F >> HexagonII::ExtentBitsPos) 3923 & HexagonII::ExtentBitsMask; 3924 3925 if (isSigned) // if value is signed 3926 return -1U << (bits - 1); 3927 else 3928 return 0; 3929 } 3930 3931 3932 // Returns opcode of the non-extended equivalent instruction. 3933 short HexagonInstrInfo::getNonExtOpcode(const MachineInstr *MI) const { 3934 // Check if the instruction has a register form that uses register in place 3935 // of the extended operand, if so return that as the non-extended form. 3936 short NonExtOpcode = Hexagon::getRegForm(MI->getOpcode()); 3937 if (NonExtOpcode >= 0) 3938 return NonExtOpcode; 3939 3940 if (MI->getDesc().mayLoad() || MI->getDesc().mayStore()) { 3941 // Check addressing mode and retrieve non-ext equivalent instruction. 3942 switch (getAddrMode(MI)) { 3943 case HexagonII::Absolute : 3944 return Hexagon::getBaseWithImmOffset(MI->getOpcode()); 3945 case HexagonII::BaseImmOffset : 3946 return Hexagon::getBaseWithRegOffset(MI->getOpcode()); 3947 case HexagonII::BaseLongOffset: 3948 return Hexagon::getRegShlForm(MI->getOpcode()); 3949 3950 default: 3951 return -1; 3952 } 3953 } 3954 return -1; 3955 } 3956 3957 3958 bool HexagonInstrInfo::getPredReg(ArrayRef<MachineOperand> Cond, 3959 unsigned &PredReg, unsigned &PredRegPos, unsigned &PredRegFlags) const { 3960 if (Cond.empty()) 3961 return false; 3962 assert(Cond.size() == 2); 3963 if (isNewValueJump(Cond[0].getImm()) || Cond[1].isMBB()) { 3964 DEBUG(dbgs() << "No predregs for new-value jumps/endloop"); 3965 return false; 3966 } 3967 PredReg = Cond[1].getReg(); 3968 PredRegPos = 1; 3969 // See IfConversion.cpp why we add RegState::Implicit | RegState::Undef 3970 PredRegFlags = 0; 3971 if (Cond[1].isImplicit()) 3972 PredRegFlags = RegState::Implicit; 3973 if (Cond[1].isUndef()) 3974 PredRegFlags |= RegState::Undef; 3975 return true; 3976 } 3977 3978 3979 short HexagonInstrInfo::getPseudoInstrPair(const MachineInstr *MI) const { 3980 return Hexagon::getRealHWInstr(MI->getOpcode(), Hexagon::InstrType_Pseudo); 3981 } 3982 3983 3984 short HexagonInstrInfo::getRegForm(const MachineInstr *MI) const { 3985 return Hexagon::getRegForm(MI->getOpcode()); 3986 } 3987 3988 3989 // Return the number of bytes required to encode the instruction. 3990 // Hexagon instructions are fixed length, 4 bytes, unless they 3991 // use a constant extender, which requires another 4 bytes. 3992 // For debug instructions and prolog labels, return 0. 3993 unsigned HexagonInstrInfo::getSize(const MachineInstr *MI) const { 3994 if (MI->isDebugValue() || MI->isPosition()) 3995 return 0; 3996 3997 unsigned Size = MI->getDesc().getSize(); 3998 if (!Size) 3999 // Assume the default insn size in case it cannot be determined 4000 // for whatever reason. 4001 Size = HEXAGON_INSTR_SIZE; 4002 4003 if (isConstExtended(MI) || isExtended(MI)) 4004 Size += HEXAGON_INSTR_SIZE; 4005 4006 // Try and compute number of instructions in asm. 4007 if (BranchRelaxAsmLarge && MI->getOpcode() == Hexagon::INLINEASM) { 4008 const MachineBasicBlock &MBB = *MI->getParent(); 4009 const MachineFunction *MF = MBB.getParent(); 4010 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo(); 4011 4012 // Count the number of register definitions to find the asm string. 4013 unsigned NumDefs = 0; 4014 for (; MI->getOperand(NumDefs).isReg() && MI->getOperand(NumDefs).isDef(); 4015 ++NumDefs) 4016 assert(NumDefs != MI->getNumOperands()-2 && "No asm string?"); 4017 4018 assert(MI->getOperand(NumDefs).isSymbol() && "No asm string?"); 4019 // Disassemble the AsmStr and approximate number of instructions. 4020 const char *AsmStr = MI->getOperand(NumDefs).getSymbolName(); 4021 Size = getInlineAsmLength(AsmStr, *MAI); 4022 } 4023 4024 return Size; 4025 } 4026 4027 4028 uint64_t HexagonInstrInfo::getType(const MachineInstr* MI) const { 4029 const uint64_t F = MI->getDesc().TSFlags; 4030 return (F >> HexagonII::TypePos) & HexagonII::TypeMask; 4031 } 4032 4033 4034 unsigned HexagonInstrInfo::getUnits(const MachineInstr* MI) const { 4035 const TargetSubtargetInfo &ST = MI->getParent()->getParent()->getSubtarget(); 4036 const InstrItineraryData &II = *ST.getInstrItineraryData(); 4037 const InstrStage &IS = *II.beginStage(MI->getDesc().getSchedClass()); 4038 4039 return IS.getUnits(); 4040 } 4041 4042 4043 unsigned HexagonInstrInfo::getValidSubTargets(const unsigned Opcode) const { 4044 const uint64_t F = get(Opcode).TSFlags; 4045 return (F >> HexagonII::validSubTargetPos) & HexagonII::validSubTargetMask; 4046 } 4047 4048 4049 // Calculate size of the basic block without debug instructions. 4050 unsigned HexagonInstrInfo::nonDbgBBSize(const MachineBasicBlock *BB) const { 4051 return nonDbgMICount(BB->instr_begin(), BB->instr_end()); 4052 } 4053 4054 4055 unsigned HexagonInstrInfo::nonDbgBundleSize( 4056 MachineBasicBlock::const_iterator BundleHead) const { 4057 assert(BundleHead->isBundle() && "Not a bundle header"); 4058 auto MII = BundleHead.getInstrIterator(); 4059 // Skip the bundle header. 4060 return nonDbgMICount(++MII, getBundleEnd(*BundleHead)); 4061 } 4062 4063 4064 /// immediateExtend - Changes the instruction in place to one using an immediate 4065 /// extender. 4066 void HexagonInstrInfo::immediateExtend(MachineInstr *MI) const { 4067 assert((isExtendable(MI)||isConstExtended(MI)) && 4068 "Instruction must be extendable"); 4069 // Find which operand is extendable. 4070 short ExtOpNum = getCExtOpNum(MI); 4071 MachineOperand &MO = MI->getOperand(ExtOpNum); 4072 // This needs to be something we understand. 4073 assert((MO.isMBB() || MO.isImm()) && 4074 "Branch with unknown extendable field type"); 4075 // Mark given operand as extended. 4076 MO.addTargetFlag(HexagonII::HMOTF_ConstExtended); 4077 } 4078 4079 4080 bool HexagonInstrInfo::invertAndChangeJumpTarget( 4081 MachineInstr* MI, MachineBasicBlock* NewTarget) const { 4082 DEBUG(dbgs() << "\n[invertAndChangeJumpTarget] to BB#" 4083 << NewTarget->getNumber(); MI->dump();); 4084 assert(MI->isBranch()); 4085 unsigned NewOpcode = getInvertedPredicatedOpcode(MI->getOpcode()); 4086 int TargetPos = MI->getNumOperands() - 1; 4087 // In general branch target is the last operand, 4088 // but some implicit defs added at the end might change it. 4089 while ((TargetPos > -1) && !MI->getOperand(TargetPos).isMBB()) 4090 --TargetPos; 4091 assert((TargetPos >= 0) && MI->getOperand(TargetPos).isMBB()); 4092 MI->getOperand(TargetPos).setMBB(NewTarget); 4093 if (EnableBranchPrediction && isPredicatedNew(*MI)) { 4094 NewOpcode = reversePrediction(NewOpcode); 4095 } 4096 MI->setDesc(get(NewOpcode)); 4097 return true; 4098 } 4099 4100 4101 void HexagonInstrInfo::genAllInsnTimingClasses(MachineFunction &MF) const { 4102 /* +++ The code below is used to generate complete set of Hexagon Insn +++ */ 4103 MachineFunction::iterator A = MF.begin(); 4104 MachineBasicBlock &B = *A; 4105 MachineBasicBlock::iterator I = B.begin(); 4106 MachineInstr *MI = &*I; 4107 DebugLoc DL = MI->getDebugLoc(); 4108 MachineInstr *NewMI; 4109 4110 for (unsigned insn = TargetOpcode::GENERIC_OP_END+1; 4111 insn < Hexagon::INSTRUCTION_LIST_END; ++insn) { 4112 NewMI = BuildMI(B, MI, DL, get(insn)); 4113 DEBUG(dbgs() << "\n" << getName(NewMI->getOpcode()) << 4114 " Class: " << NewMI->getDesc().getSchedClass()); 4115 NewMI->eraseFromParent(); 4116 } 4117 /* --- The code above is used to generate complete set of Hexagon Insn --- */ 4118 } 4119 4120 4121 // inverts the predication logic. 4122 // p -> NotP 4123 // NotP -> P 4124 bool HexagonInstrInfo::reversePredSense(MachineInstr* MI) const { 4125 DEBUG(dbgs() << "\nTrying to reverse pred. sense of:"; MI->dump()); 4126 MI->setDesc(get(getInvertedPredicatedOpcode(MI->getOpcode()))); 4127 return true; 4128 } 4129 4130 4131 // Reverse the branch prediction. 4132 unsigned HexagonInstrInfo::reversePrediction(unsigned Opcode) const { 4133 int PredRevOpcode = -1; 4134 if (isPredictedTaken(Opcode)) 4135 PredRevOpcode = Hexagon::notTakenBranchPrediction(Opcode); 4136 else 4137 PredRevOpcode = Hexagon::takenBranchPrediction(Opcode); 4138 assert(PredRevOpcode > 0); 4139 return PredRevOpcode; 4140 } 4141 4142 4143 // TODO: Add more rigorous validation. 4144 bool HexagonInstrInfo::validateBranchCond(const ArrayRef<MachineOperand> &Cond) 4145 const { 4146 return Cond.empty() || (Cond[0].isImm() && (Cond.size() != 1)); 4147 } 4148 4149 4150 short HexagonInstrInfo::xformRegToImmOffset(const MachineInstr *MI) const { 4151 return Hexagon::xformRegToImmOffset(MI->getOpcode()); 4152 } 4153