1 //===- HexagonInstrInfo.cpp - Hexagon Instruction Information -------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains the Hexagon implementation of the TargetInstrInfo class. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "HexagonInstrInfo.h" 14 #include "Hexagon.h" 15 #include "HexagonFrameLowering.h" 16 #include "HexagonHazardRecognizer.h" 17 #include "HexagonRegisterInfo.h" 18 #include "HexagonSubtarget.h" 19 #include "llvm/ADT/ArrayRef.h" 20 #include "llvm/ADT/SmallPtrSet.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/ADT/StringRef.h" 23 #include "llvm/CodeGen/DFAPacketizer.h" 24 #include "llvm/CodeGen/LivePhysRegs.h" 25 #include "llvm/CodeGen/MachineBasicBlock.h" 26 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 27 #include "llvm/CodeGen/MachineFrameInfo.h" 28 #include "llvm/CodeGen/MachineFunction.h" 29 #include "llvm/CodeGen/MachineInstr.h" 30 #include "llvm/CodeGen/MachineInstrBuilder.h" 31 #include "llvm/CodeGen/MachineInstrBundle.h" 32 #include "llvm/CodeGen/MachineLoopInfo.h" 33 #include "llvm/CodeGen/MachineMemOperand.h" 34 #include "llvm/CodeGen/MachineOperand.h" 35 #include "llvm/CodeGen/MachineRegisterInfo.h" 36 #include "llvm/CodeGen/ScheduleDAG.h" 37 #include "llvm/CodeGen/TargetInstrInfo.h" 38 #include "llvm/CodeGen/TargetOpcodes.h" 39 #include "llvm/CodeGen/TargetRegisterInfo.h" 40 #include "llvm/CodeGen/TargetSubtargetInfo.h" 41 #include "llvm/IR/DebugLoc.h" 42 #include "llvm/MC/MCAsmInfo.h" 43 #include "llvm/MC/MCInstrDesc.h" 44 #include "llvm/MC/MCInstrItineraries.h" 45 #include "llvm/MC/MCRegisterInfo.h" 46 #include "llvm/Support/BranchProbability.h" 47 #include "llvm/Support/CommandLine.h" 48 #include "llvm/Support/Debug.h" 49 #include "llvm/Support/ErrorHandling.h" 50 #include "llvm/Support/MachineValueType.h" 51 #include "llvm/Support/MathExtras.h" 52 #include "llvm/Support/raw_ostream.h" 53 #include "llvm/Target/TargetMachine.h" 54 #include <cassert> 55 #include <cctype> 56 #include <cstdint> 57 #include <cstring> 58 #include <iterator> 59 #include <string> 60 #include <utility> 61 62 using namespace llvm; 63 64 #define DEBUG_TYPE "hexagon-instrinfo" 65 66 #define GET_INSTRINFO_CTOR_DTOR 67 #define GET_INSTRMAP_INFO 68 #include "HexagonDepTimingClasses.h" 69 #include "HexagonGenDFAPacketizer.inc" 70 #include "HexagonGenInstrInfo.inc" 71 72 cl::opt<bool> ScheduleInlineAsm("hexagon-sched-inline-asm", cl::Hidden, 73 cl::init(false), cl::desc("Do not consider inline-asm a scheduling/" 74 "packetization boundary.")); 75 76 static cl::opt<bool> EnableBranchPrediction("hexagon-enable-branch-prediction", 77 cl::Hidden, cl::init(true), cl::desc("Enable branch prediction")); 78 79 static cl::opt<bool> DisableNVSchedule("disable-hexagon-nv-schedule", 80 cl::Hidden, cl::ZeroOrMore, cl::init(false), 81 cl::desc("Disable schedule adjustment for new value stores.")); 82 83 static cl::opt<bool> EnableTimingClassLatency( 84 "enable-timing-class-latency", cl::Hidden, cl::init(false), 85 cl::desc("Enable timing class latency")); 86 87 static cl::opt<bool> EnableALUForwarding( 88 "enable-alu-forwarding", cl::Hidden, cl::init(true), 89 cl::desc("Enable vec alu forwarding")); 90 91 static cl::opt<bool> EnableACCForwarding( 92 "enable-acc-forwarding", cl::Hidden, cl::init(true), 93 cl::desc("Enable vec acc forwarding")); 94 95 static cl::opt<bool> BranchRelaxAsmLarge("branch-relax-asm-large", 96 cl::init(true), cl::Hidden, cl::ZeroOrMore, cl::desc("branch relax asm")); 97 98 static cl::opt<bool> UseDFAHazardRec("dfa-hazard-rec", 99 cl::init(true), cl::Hidden, cl::ZeroOrMore, 100 cl::desc("Use the DFA based hazard recognizer.")); 101 102 /// Constants for Hexagon instructions. 103 const int Hexagon_MEMW_OFFSET_MAX = 4095; 104 const int Hexagon_MEMW_OFFSET_MIN = -4096; 105 const int Hexagon_MEMD_OFFSET_MAX = 8191; 106 const int Hexagon_MEMD_OFFSET_MIN = -8192; 107 const int Hexagon_MEMH_OFFSET_MAX = 2047; 108 const int Hexagon_MEMH_OFFSET_MIN = -2048; 109 const int Hexagon_MEMB_OFFSET_MAX = 1023; 110 const int Hexagon_MEMB_OFFSET_MIN = -1024; 111 const int Hexagon_ADDI_OFFSET_MAX = 32767; 112 const int Hexagon_ADDI_OFFSET_MIN = -32768; 113 114 // Pin the vtable to this file. 115 void HexagonInstrInfo::anchor() {} 116 117 HexagonInstrInfo::HexagonInstrInfo(HexagonSubtarget &ST) 118 : HexagonGenInstrInfo(Hexagon::ADJCALLSTACKDOWN, Hexagon::ADJCALLSTACKUP), 119 Subtarget(ST) {} 120 121 namespace llvm { 122 namespace HexagonFUnits { 123 bool isSlot0Only(unsigned units); 124 } 125 } 126 127 static bool isIntRegForSubInst(unsigned Reg) { 128 return (Reg >= Hexagon::R0 && Reg <= Hexagon::R7) || 129 (Reg >= Hexagon::R16 && Reg <= Hexagon::R23); 130 } 131 132 static bool isDblRegForSubInst(unsigned Reg, const HexagonRegisterInfo &HRI) { 133 return isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::isub_lo)) && 134 isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::isub_hi)); 135 } 136 137 /// Calculate number of instructions excluding the debug instructions. 138 static unsigned nonDbgMICount(MachineBasicBlock::const_instr_iterator MIB, 139 MachineBasicBlock::const_instr_iterator MIE) { 140 unsigned Count = 0; 141 for (; MIB != MIE; ++MIB) { 142 if (!MIB->isDebugInstr()) 143 ++Count; 144 } 145 return Count; 146 } 147 148 /// Find the hardware loop instruction used to set-up the specified loop. 149 /// On Hexagon, we have two instructions used to set-up the hardware loop 150 /// (LOOP0, LOOP1) with corresponding endloop (ENDLOOP0, ENDLOOP1) instructions 151 /// to indicate the end of a loop. 152 MachineInstr *HexagonInstrInfo::findLoopInstr(MachineBasicBlock *BB, 153 unsigned EndLoopOp, MachineBasicBlock *TargetBB, 154 SmallPtrSet<MachineBasicBlock *, 8> &Visited) const { 155 unsigned LOOPi; 156 unsigned LOOPr; 157 if (EndLoopOp == Hexagon::ENDLOOP0) { 158 LOOPi = Hexagon::J2_loop0i; 159 LOOPr = Hexagon::J2_loop0r; 160 } else { // EndLoopOp == Hexagon::EndLOOP1 161 LOOPi = Hexagon::J2_loop1i; 162 LOOPr = Hexagon::J2_loop1r; 163 } 164 165 // The loop set-up instruction will be in a predecessor block 166 for (MachineBasicBlock *PB : BB->predecessors()) { 167 // If this has been visited, already skip it. 168 if (!Visited.insert(PB).second) 169 continue; 170 if (PB == BB) 171 continue; 172 for (auto I = PB->instr_rbegin(), E = PB->instr_rend(); I != E; ++I) { 173 unsigned Opc = I->getOpcode(); 174 if (Opc == LOOPi || Opc == LOOPr) 175 return &*I; 176 // We've reached a different loop, which means the loop01 has been 177 // removed. 178 if (Opc == EndLoopOp && I->getOperand(0).getMBB() != TargetBB) 179 return nullptr; 180 } 181 // Check the predecessors for the LOOP instruction. 182 if (MachineInstr *Loop = findLoopInstr(PB, EndLoopOp, TargetBB, Visited)) 183 return Loop; 184 } 185 return nullptr; 186 } 187 188 /// Gather register def/uses from MI. 189 /// This treats possible (predicated) defs as actually happening ones 190 /// (conservatively). 191 static inline void parseOperands(const MachineInstr &MI, 192 SmallVector<unsigned, 4> &Defs, SmallVector<unsigned, 8> &Uses) { 193 Defs.clear(); 194 Uses.clear(); 195 196 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 197 const MachineOperand &MO = MI.getOperand(i); 198 199 if (!MO.isReg()) 200 continue; 201 202 Register Reg = MO.getReg(); 203 if (!Reg) 204 continue; 205 206 if (MO.isUse()) 207 Uses.push_back(MO.getReg()); 208 209 if (MO.isDef()) 210 Defs.push_back(MO.getReg()); 211 } 212 } 213 214 // Position dependent, so check twice for swap. 215 static bool isDuplexPairMatch(unsigned Ga, unsigned Gb) { 216 switch (Ga) { 217 case HexagonII::HSIG_None: 218 default: 219 return false; 220 case HexagonII::HSIG_L1: 221 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_A); 222 case HexagonII::HSIG_L2: 223 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 || 224 Gb == HexagonII::HSIG_A); 225 case HexagonII::HSIG_S1: 226 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 || 227 Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_A); 228 case HexagonII::HSIG_S2: 229 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 || 230 Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_S2 || 231 Gb == HexagonII::HSIG_A); 232 case HexagonII::HSIG_A: 233 return (Gb == HexagonII::HSIG_A); 234 case HexagonII::HSIG_Compound: 235 return (Gb == HexagonII::HSIG_Compound); 236 } 237 return false; 238 } 239 240 /// isLoadFromStackSlot - If the specified machine instruction is a direct 241 /// load from a stack slot, return the virtual or physical register number of 242 /// the destination along with the FrameIndex of the loaded stack slot. If 243 /// not, return 0. This predicate must return 0 if the instruction has 244 /// any side effects other than loading from the stack slot. 245 unsigned HexagonInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 246 int &FrameIndex) const { 247 switch (MI.getOpcode()) { 248 default: 249 break; 250 case Hexagon::L2_loadri_io: 251 case Hexagon::L2_loadrd_io: 252 case Hexagon::V6_vL32b_ai: 253 case Hexagon::V6_vL32b_nt_ai: 254 case Hexagon::V6_vL32Ub_ai: 255 case Hexagon::LDriw_pred: 256 case Hexagon::LDriw_ctr: 257 case Hexagon::PS_vloadrq_ai: 258 case Hexagon::PS_vloadrw_ai: 259 case Hexagon::PS_vloadrw_nt_ai: { 260 const MachineOperand OpFI = MI.getOperand(1); 261 if (!OpFI.isFI()) 262 return 0; 263 const MachineOperand OpOff = MI.getOperand(2); 264 if (!OpOff.isImm() || OpOff.getImm() != 0) 265 return 0; 266 FrameIndex = OpFI.getIndex(); 267 return MI.getOperand(0).getReg(); 268 } 269 270 case Hexagon::L2_ploadrit_io: 271 case Hexagon::L2_ploadrif_io: 272 case Hexagon::L2_ploadrdt_io: 273 case Hexagon::L2_ploadrdf_io: { 274 const MachineOperand OpFI = MI.getOperand(2); 275 if (!OpFI.isFI()) 276 return 0; 277 const MachineOperand OpOff = MI.getOperand(3); 278 if (!OpOff.isImm() || OpOff.getImm() != 0) 279 return 0; 280 FrameIndex = OpFI.getIndex(); 281 return MI.getOperand(0).getReg(); 282 } 283 } 284 285 return 0; 286 } 287 288 /// isStoreToStackSlot - If the specified machine instruction is a direct 289 /// store to a stack slot, return the virtual or physical register number of 290 /// the source reg along with the FrameIndex of the loaded stack slot. If 291 /// not, return 0. This predicate must return 0 if the instruction has 292 /// any side effects other than storing to the stack slot. 293 unsigned HexagonInstrInfo::isStoreToStackSlot(const MachineInstr &MI, 294 int &FrameIndex) const { 295 switch (MI.getOpcode()) { 296 default: 297 break; 298 case Hexagon::S2_storerb_io: 299 case Hexagon::S2_storerh_io: 300 case Hexagon::S2_storeri_io: 301 case Hexagon::S2_storerd_io: 302 case Hexagon::V6_vS32b_ai: 303 case Hexagon::V6_vS32Ub_ai: 304 case Hexagon::STriw_pred: 305 case Hexagon::STriw_ctr: 306 case Hexagon::PS_vstorerq_ai: 307 case Hexagon::PS_vstorerw_ai: { 308 const MachineOperand &OpFI = MI.getOperand(0); 309 if (!OpFI.isFI()) 310 return 0; 311 const MachineOperand &OpOff = MI.getOperand(1); 312 if (!OpOff.isImm() || OpOff.getImm() != 0) 313 return 0; 314 FrameIndex = OpFI.getIndex(); 315 return MI.getOperand(2).getReg(); 316 } 317 318 case Hexagon::S2_pstorerbt_io: 319 case Hexagon::S2_pstorerbf_io: 320 case Hexagon::S2_pstorerht_io: 321 case Hexagon::S2_pstorerhf_io: 322 case Hexagon::S2_pstorerit_io: 323 case Hexagon::S2_pstorerif_io: 324 case Hexagon::S2_pstorerdt_io: 325 case Hexagon::S2_pstorerdf_io: { 326 const MachineOperand &OpFI = MI.getOperand(1); 327 if (!OpFI.isFI()) 328 return 0; 329 const MachineOperand &OpOff = MI.getOperand(2); 330 if (!OpOff.isImm() || OpOff.getImm() != 0) 331 return 0; 332 FrameIndex = OpFI.getIndex(); 333 return MI.getOperand(3).getReg(); 334 } 335 } 336 337 return 0; 338 } 339 340 /// This function checks if the instruction or bundle of instructions 341 /// has load from stack slot and returns frameindex and machine memory 342 /// operand of that instruction if true. 343 bool HexagonInstrInfo::hasLoadFromStackSlot( 344 const MachineInstr &MI, 345 SmallVectorImpl<const MachineMemOperand *> &Accesses) const { 346 if (MI.isBundle()) { 347 const MachineBasicBlock *MBB = MI.getParent(); 348 MachineBasicBlock::const_instr_iterator MII = MI.getIterator(); 349 for (++MII; MII != MBB->instr_end() && MII->isInsideBundle(); ++MII) 350 if (TargetInstrInfo::hasLoadFromStackSlot(*MII, Accesses)) 351 return true; 352 return false; 353 } 354 355 return TargetInstrInfo::hasLoadFromStackSlot(MI, Accesses); 356 } 357 358 /// This function checks if the instruction or bundle of instructions 359 /// has store to stack slot and returns frameindex and machine memory 360 /// operand of that instruction if true. 361 bool HexagonInstrInfo::hasStoreToStackSlot( 362 const MachineInstr &MI, 363 SmallVectorImpl<const MachineMemOperand *> &Accesses) const { 364 if (MI.isBundle()) { 365 const MachineBasicBlock *MBB = MI.getParent(); 366 MachineBasicBlock::const_instr_iterator MII = MI.getIterator(); 367 for (++MII; MII != MBB->instr_end() && MII->isInsideBundle(); ++MII) 368 if (TargetInstrInfo::hasStoreToStackSlot(*MII, Accesses)) 369 return true; 370 return false; 371 } 372 373 return TargetInstrInfo::hasStoreToStackSlot(MI, Accesses); 374 } 375 376 /// This function can analyze one/two way branching only and should (mostly) be 377 /// called by target independent side. 378 /// First entry is always the opcode of the branching instruction, except when 379 /// the Cond vector is supposed to be empty, e.g., when analyzeBranch fails, a 380 /// BB with only unconditional jump. Subsequent entries depend upon the opcode, 381 /// e.g. Jump_c p will have 382 /// Cond[0] = Jump_c 383 /// Cond[1] = p 384 /// HW-loop ENDLOOP: 385 /// Cond[0] = ENDLOOP 386 /// Cond[1] = MBB 387 /// New value jump: 388 /// Cond[0] = Hexagon::CMPEQri_f_Jumpnv_t_V4 -- specific opcode 389 /// Cond[1] = R 390 /// Cond[2] = Imm 391 bool HexagonInstrInfo::analyzeBranch(MachineBasicBlock &MBB, 392 MachineBasicBlock *&TBB, 393 MachineBasicBlock *&FBB, 394 SmallVectorImpl<MachineOperand> &Cond, 395 bool AllowModify) const { 396 TBB = nullptr; 397 FBB = nullptr; 398 Cond.clear(); 399 400 // If the block has no terminators, it just falls into the block after it. 401 MachineBasicBlock::instr_iterator I = MBB.instr_end(); 402 if (I == MBB.instr_begin()) 403 return false; 404 405 // A basic block may looks like this: 406 // 407 // [ insn 408 // EH_LABEL 409 // insn 410 // insn 411 // insn 412 // EH_LABEL 413 // insn ] 414 // 415 // It has two succs but does not have a terminator 416 // Don't know how to handle it. 417 do { 418 --I; 419 if (I->isEHLabel()) 420 // Don't analyze EH branches. 421 return true; 422 } while (I != MBB.instr_begin()); 423 424 I = MBB.instr_end(); 425 --I; 426 427 while (I->isDebugInstr()) { 428 if (I == MBB.instr_begin()) 429 return false; 430 --I; 431 } 432 433 bool JumpToBlock = I->getOpcode() == Hexagon::J2_jump && 434 I->getOperand(0).isMBB(); 435 // Delete the J2_jump if it's equivalent to a fall-through. 436 if (AllowModify && JumpToBlock && 437 MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) { 438 LLVM_DEBUG(dbgs() << "\nErasing the jump to successor block\n";); 439 I->eraseFromParent(); 440 I = MBB.instr_end(); 441 if (I == MBB.instr_begin()) 442 return false; 443 --I; 444 } 445 if (!isUnpredicatedTerminator(*I)) 446 return false; 447 448 // Get the last instruction in the block. 449 MachineInstr *LastInst = &*I; 450 MachineInstr *SecondLastInst = nullptr; 451 // Find one more terminator if present. 452 while (true) { 453 if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(*I)) { 454 if (!SecondLastInst) 455 SecondLastInst = &*I; 456 else 457 // This is a third branch. 458 return true; 459 } 460 if (I == MBB.instr_begin()) 461 break; 462 --I; 463 } 464 465 int LastOpcode = LastInst->getOpcode(); 466 int SecLastOpcode = SecondLastInst ? SecondLastInst->getOpcode() : 0; 467 // If the branch target is not a basic block, it could be a tail call. 468 // (It is, if the target is a function.) 469 if (LastOpcode == Hexagon::J2_jump && !LastInst->getOperand(0).isMBB()) 470 return true; 471 if (SecLastOpcode == Hexagon::J2_jump && 472 !SecondLastInst->getOperand(0).isMBB()) 473 return true; 474 475 bool LastOpcodeHasJMP_c = PredOpcodeHasJMP_c(LastOpcode); 476 bool LastOpcodeHasNVJump = isNewValueJump(*LastInst); 477 478 if (LastOpcodeHasJMP_c && !LastInst->getOperand(1).isMBB()) 479 return true; 480 481 // If there is only one terminator instruction, process it. 482 if (LastInst && !SecondLastInst) { 483 if (LastOpcode == Hexagon::J2_jump) { 484 TBB = LastInst->getOperand(0).getMBB(); 485 return false; 486 } 487 if (isEndLoopN(LastOpcode)) { 488 TBB = LastInst->getOperand(0).getMBB(); 489 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode())); 490 Cond.push_back(LastInst->getOperand(0)); 491 return false; 492 } 493 if (LastOpcodeHasJMP_c) { 494 TBB = LastInst->getOperand(1).getMBB(); 495 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode())); 496 Cond.push_back(LastInst->getOperand(0)); 497 return false; 498 } 499 // Only supporting rr/ri versions of new-value jumps. 500 if (LastOpcodeHasNVJump && (LastInst->getNumExplicitOperands() == 3)) { 501 TBB = LastInst->getOperand(2).getMBB(); 502 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode())); 503 Cond.push_back(LastInst->getOperand(0)); 504 Cond.push_back(LastInst->getOperand(1)); 505 return false; 506 } 507 LLVM_DEBUG(dbgs() << "\nCant analyze " << printMBBReference(MBB) 508 << " with one jump\n";); 509 // Otherwise, don't know what this is. 510 return true; 511 } 512 513 bool SecLastOpcodeHasJMP_c = PredOpcodeHasJMP_c(SecLastOpcode); 514 bool SecLastOpcodeHasNVJump = isNewValueJump(*SecondLastInst); 515 if (SecLastOpcodeHasJMP_c && (LastOpcode == Hexagon::J2_jump)) { 516 if (!SecondLastInst->getOperand(1).isMBB()) 517 return true; 518 TBB = SecondLastInst->getOperand(1).getMBB(); 519 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode())); 520 Cond.push_back(SecondLastInst->getOperand(0)); 521 FBB = LastInst->getOperand(0).getMBB(); 522 return false; 523 } 524 525 // Only supporting rr/ri versions of new-value jumps. 526 if (SecLastOpcodeHasNVJump && 527 (SecondLastInst->getNumExplicitOperands() == 3) && 528 (LastOpcode == Hexagon::J2_jump)) { 529 TBB = SecondLastInst->getOperand(2).getMBB(); 530 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode())); 531 Cond.push_back(SecondLastInst->getOperand(0)); 532 Cond.push_back(SecondLastInst->getOperand(1)); 533 FBB = LastInst->getOperand(0).getMBB(); 534 return false; 535 } 536 537 // If the block ends with two Hexagon:JMPs, handle it. The second one is not 538 // executed, so remove it. 539 if (SecLastOpcode == Hexagon::J2_jump && LastOpcode == Hexagon::J2_jump) { 540 TBB = SecondLastInst->getOperand(0).getMBB(); 541 I = LastInst->getIterator(); 542 if (AllowModify) 543 I->eraseFromParent(); 544 return false; 545 } 546 547 // If the block ends with an ENDLOOP, and J2_jump, handle it. 548 if (isEndLoopN(SecLastOpcode) && LastOpcode == Hexagon::J2_jump) { 549 TBB = SecondLastInst->getOperand(0).getMBB(); 550 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode())); 551 Cond.push_back(SecondLastInst->getOperand(0)); 552 FBB = LastInst->getOperand(0).getMBB(); 553 return false; 554 } 555 LLVM_DEBUG(dbgs() << "\nCant analyze " << printMBBReference(MBB) 556 << " with two jumps";); 557 // Otherwise, can't handle this. 558 return true; 559 } 560 561 unsigned HexagonInstrInfo::removeBranch(MachineBasicBlock &MBB, 562 int *BytesRemoved) const { 563 assert(!BytesRemoved && "code size not handled"); 564 565 LLVM_DEBUG(dbgs() << "\nRemoving branches out of " << printMBBReference(MBB)); 566 MachineBasicBlock::iterator I = MBB.end(); 567 unsigned Count = 0; 568 while (I != MBB.begin()) { 569 --I; 570 if (I->isDebugInstr()) 571 continue; 572 // Only removing branches from end of MBB. 573 if (!I->isBranch()) 574 return Count; 575 if (Count && (I->getOpcode() == Hexagon::J2_jump)) 576 llvm_unreachable("Malformed basic block: unconditional branch not last"); 577 MBB.erase(&MBB.back()); 578 I = MBB.end(); 579 ++Count; 580 } 581 return Count; 582 } 583 584 unsigned HexagonInstrInfo::insertBranch(MachineBasicBlock &MBB, 585 MachineBasicBlock *TBB, 586 MachineBasicBlock *FBB, 587 ArrayRef<MachineOperand> Cond, 588 const DebugLoc &DL, 589 int *BytesAdded) const { 590 unsigned BOpc = Hexagon::J2_jump; 591 unsigned BccOpc = Hexagon::J2_jumpt; 592 assert(validateBranchCond(Cond) && "Invalid branching condition"); 593 assert(TBB && "insertBranch must not be told to insert a fallthrough"); 594 assert(!BytesAdded && "code size not handled"); 595 596 // Check if reverseBranchCondition has asked to reverse this branch 597 // If we want to reverse the branch an odd number of times, we want 598 // J2_jumpf. 599 if (!Cond.empty() && Cond[0].isImm()) 600 BccOpc = Cond[0].getImm(); 601 602 if (!FBB) { 603 if (Cond.empty()) { 604 // Due to a bug in TailMerging/CFG Optimization, we need to add a 605 // special case handling of a predicated jump followed by an 606 // unconditional jump. If not, Tail Merging and CFG Optimization go 607 // into an infinite loop. 608 MachineBasicBlock *NewTBB, *NewFBB; 609 SmallVector<MachineOperand, 4> Cond; 610 auto Term = MBB.getFirstTerminator(); 611 if (Term != MBB.end() && isPredicated(*Term) && 612 !analyzeBranch(MBB, NewTBB, NewFBB, Cond, false) && 613 MachineFunction::iterator(NewTBB) == ++MBB.getIterator()) { 614 reverseBranchCondition(Cond); 615 removeBranch(MBB); 616 return insertBranch(MBB, TBB, nullptr, Cond, DL); 617 } 618 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB); 619 } else if (isEndLoopN(Cond[0].getImm())) { 620 int EndLoopOp = Cond[0].getImm(); 621 assert(Cond[1].isMBB()); 622 // Since we're adding an ENDLOOP, there better be a LOOP instruction. 623 // Check for it, and change the BB target if needed. 624 SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs; 625 MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, Cond[1].getMBB(), 626 VisitedBBs); 627 assert(Loop != nullptr && "Inserting an ENDLOOP without a LOOP"); 628 Loop->getOperand(0).setMBB(TBB); 629 // Add the ENDLOOP after the finding the LOOP0. 630 BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB); 631 } else if (isNewValueJump(Cond[0].getImm())) { 632 assert((Cond.size() == 3) && "Only supporting rr/ri version of nvjump"); 633 // New value jump 634 // (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset) 635 // (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset) 636 unsigned Flags1 = getUndefRegState(Cond[1].isUndef()); 637 LLVM_DEBUG(dbgs() << "\nInserting NVJump for " 638 << printMBBReference(MBB);); 639 if (Cond[2].isReg()) { 640 unsigned Flags2 = getUndefRegState(Cond[2].isUndef()); 641 BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1). 642 addReg(Cond[2].getReg(), Flags2).addMBB(TBB); 643 } else if(Cond[2].isImm()) { 644 BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1). 645 addImm(Cond[2].getImm()).addMBB(TBB); 646 } else 647 llvm_unreachable("Invalid condition for branching"); 648 } else { 649 assert((Cond.size() == 2) && "Malformed cond vector"); 650 const MachineOperand &RO = Cond[1]; 651 unsigned Flags = getUndefRegState(RO.isUndef()); 652 BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB); 653 } 654 return 1; 655 } 656 assert((!Cond.empty()) && 657 "Cond. cannot be empty when multiple branchings are required"); 658 assert((!isNewValueJump(Cond[0].getImm())) && 659 "NV-jump cannot be inserted with another branch"); 660 // Special case for hardware loops. The condition is a basic block. 661 if (isEndLoopN(Cond[0].getImm())) { 662 int EndLoopOp = Cond[0].getImm(); 663 assert(Cond[1].isMBB()); 664 // Since we're adding an ENDLOOP, there better be a LOOP instruction. 665 // Check for it, and change the BB target if needed. 666 SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs; 667 MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, Cond[1].getMBB(), 668 VisitedBBs); 669 assert(Loop != nullptr && "Inserting an ENDLOOP without a LOOP"); 670 Loop->getOperand(0).setMBB(TBB); 671 // Add the ENDLOOP after the finding the LOOP0. 672 BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB); 673 } else { 674 const MachineOperand &RO = Cond[1]; 675 unsigned Flags = getUndefRegState(RO.isUndef()); 676 BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB); 677 } 678 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB); 679 680 return 2; 681 } 682 683 namespace { 684 class HexagonPipelinerLoopInfo : public TargetInstrInfo::PipelinerLoopInfo { 685 MachineInstr *Loop, *EndLoop; 686 MachineFunction *MF; 687 const HexagonInstrInfo *TII; 688 int64_t TripCount; 689 Register LoopCount; 690 DebugLoc DL; 691 692 public: 693 HexagonPipelinerLoopInfo(MachineInstr *Loop, MachineInstr *EndLoop) 694 : Loop(Loop), EndLoop(EndLoop), MF(Loop->getParent()->getParent()), 695 TII(MF->getSubtarget<HexagonSubtarget>().getInstrInfo()), 696 DL(Loop->getDebugLoc()) { 697 // Inspect the Loop instruction up-front, as it may be deleted when we call 698 // createTripCountGreaterCondition. 699 TripCount = Loop->getOpcode() == Hexagon::J2_loop0r 700 ? -1 701 : Loop->getOperand(1).getImm(); 702 if (TripCount == -1) 703 LoopCount = Loop->getOperand(1).getReg(); 704 } 705 706 bool shouldIgnoreForPipelining(const MachineInstr *MI) const override { 707 // Only ignore the terminator. 708 return MI == EndLoop; 709 } 710 711 Optional<bool> 712 createTripCountGreaterCondition(int TC, MachineBasicBlock &MBB, 713 SmallVectorImpl<MachineOperand> &Cond) override { 714 if (TripCount == -1) { 715 // Check if we're done with the loop. 716 unsigned Done = TII->createVR(MF, MVT::i1); 717 MachineInstr *NewCmp = BuildMI(&MBB, DL, 718 TII->get(Hexagon::C2_cmpgtui), Done) 719 .addReg(LoopCount) 720 .addImm(TC); 721 Cond.push_back(MachineOperand::CreateImm(Hexagon::J2_jumpf)); 722 Cond.push_back(NewCmp->getOperand(0)); 723 return {}; 724 } 725 726 return TripCount > TC; 727 } 728 729 void setPreheader(MachineBasicBlock *NewPreheader) override { 730 NewPreheader->splice(NewPreheader->getFirstTerminator(), Loop->getParent(), 731 Loop); 732 } 733 734 void adjustTripCount(int TripCountAdjust) override { 735 // If the loop trip count is a compile-time value, then just change the 736 // value. 737 if (Loop->getOpcode() == Hexagon::J2_loop0i || 738 Loop->getOpcode() == Hexagon::J2_loop1i) { 739 int64_t TripCount = Loop->getOperand(1).getImm() + TripCountAdjust; 740 assert(TripCount > 0 && "Can't create an empty or negative loop!"); 741 Loop->getOperand(1).setImm(TripCount); 742 return; 743 } 744 745 // The loop trip count is a run-time value. We generate code to subtract 746 // one from the trip count, and update the loop instruction. 747 Register LoopCount = Loop->getOperand(1).getReg(); 748 Register NewLoopCount = TII->createVR(MF, MVT::i32); 749 BuildMI(*Loop->getParent(), Loop, Loop->getDebugLoc(), 750 TII->get(Hexagon::A2_addi), NewLoopCount) 751 .addReg(LoopCount) 752 .addImm(TripCountAdjust); 753 Loop->getOperand(1).setReg(NewLoopCount); 754 } 755 756 void disposed() override { Loop->eraseFromParent(); } 757 }; 758 } // namespace 759 760 std::unique_ptr<TargetInstrInfo::PipelinerLoopInfo> 761 HexagonInstrInfo::analyzeLoopForPipelining(MachineBasicBlock *LoopBB) const { 762 // We really "analyze" only hardware loops right now. 763 MachineBasicBlock::iterator I = LoopBB->getFirstTerminator(); 764 765 if (I != LoopBB->end() && isEndLoopN(I->getOpcode())) { 766 SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs; 767 MachineInstr *LoopInst = findLoopInstr( 768 LoopBB, I->getOpcode(), I->getOperand(0).getMBB(), VisitedBBs); 769 if (LoopInst) 770 return std::make_unique<HexagonPipelinerLoopInfo>(LoopInst, &*I); 771 } 772 return nullptr; 773 } 774 775 bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB, 776 unsigned NumCycles, unsigned ExtraPredCycles, 777 BranchProbability Probability) const { 778 return nonDbgBBSize(&MBB) <= 3; 779 } 780 781 bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB, 782 unsigned NumTCycles, unsigned ExtraTCycles, MachineBasicBlock &FMBB, 783 unsigned NumFCycles, unsigned ExtraFCycles, BranchProbability Probability) 784 const { 785 return nonDbgBBSize(&TMBB) <= 3 && nonDbgBBSize(&FMBB) <= 3; 786 } 787 788 bool HexagonInstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB, 789 unsigned NumInstrs, BranchProbability Probability) const { 790 return NumInstrs <= 4; 791 } 792 793 static void getLiveInRegsAt(LivePhysRegs &Regs, const MachineInstr &MI) { 794 SmallVector<std::pair<MCPhysReg, const MachineOperand*>,2> Clobbers; 795 const MachineBasicBlock &B = *MI.getParent(); 796 Regs.addLiveIns(B); 797 auto E = MachineBasicBlock::const_iterator(MI.getIterator()); 798 for (auto I = B.begin(); I != E; ++I) { 799 Clobbers.clear(); 800 Regs.stepForward(*I, Clobbers); 801 } 802 } 803 804 static void getLiveOutRegsAt(LivePhysRegs &Regs, const MachineInstr &MI) { 805 const MachineBasicBlock &B = *MI.getParent(); 806 Regs.addLiveOuts(B); 807 auto E = ++MachineBasicBlock::const_iterator(MI.getIterator()).getReverse(); 808 for (auto I = B.rbegin(); I != E; ++I) 809 Regs.stepBackward(*I); 810 } 811 812 void HexagonInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 813 MachineBasicBlock::iterator I, 814 const DebugLoc &DL, MCRegister DestReg, 815 MCRegister SrcReg, bool KillSrc) const { 816 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 817 unsigned KillFlag = getKillRegState(KillSrc); 818 819 if (Hexagon::IntRegsRegClass.contains(SrcReg, DestReg)) { 820 BuildMI(MBB, I, DL, get(Hexagon::A2_tfr), DestReg) 821 .addReg(SrcReg, KillFlag); 822 return; 823 } 824 if (Hexagon::DoubleRegsRegClass.contains(SrcReg, DestReg)) { 825 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrp), DestReg) 826 .addReg(SrcReg, KillFlag); 827 return; 828 } 829 if (Hexagon::PredRegsRegClass.contains(SrcReg, DestReg)) { 830 // Map Pd = Ps to Pd = or(Ps, Ps). 831 BuildMI(MBB, I, DL, get(Hexagon::C2_or), DestReg) 832 .addReg(SrcReg).addReg(SrcReg, KillFlag); 833 return; 834 } 835 if (Hexagon::CtrRegsRegClass.contains(DestReg) && 836 Hexagon::IntRegsRegClass.contains(SrcReg)) { 837 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg) 838 .addReg(SrcReg, KillFlag); 839 return; 840 } 841 if (Hexagon::IntRegsRegClass.contains(DestReg) && 842 Hexagon::CtrRegsRegClass.contains(SrcReg)) { 843 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrcrr), DestReg) 844 .addReg(SrcReg, KillFlag); 845 return; 846 } 847 if (Hexagon::ModRegsRegClass.contains(DestReg) && 848 Hexagon::IntRegsRegClass.contains(SrcReg)) { 849 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg) 850 .addReg(SrcReg, KillFlag); 851 return; 852 } 853 if (Hexagon::PredRegsRegClass.contains(SrcReg) && 854 Hexagon::IntRegsRegClass.contains(DestReg)) { 855 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg) 856 .addReg(SrcReg, KillFlag); 857 return; 858 } 859 if (Hexagon::IntRegsRegClass.contains(SrcReg) && 860 Hexagon::PredRegsRegClass.contains(DestReg)) { 861 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrrp), DestReg) 862 .addReg(SrcReg, KillFlag); 863 return; 864 } 865 if (Hexagon::PredRegsRegClass.contains(SrcReg) && 866 Hexagon::IntRegsRegClass.contains(DestReg)) { 867 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg) 868 .addReg(SrcReg, KillFlag); 869 return; 870 } 871 if (Hexagon::HvxVRRegClass.contains(SrcReg, DestReg)) { 872 BuildMI(MBB, I, DL, get(Hexagon::V6_vassign), DestReg). 873 addReg(SrcReg, KillFlag); 874 return; 875 } 876 if (Hexagon::HvxWRRegClass.contains(SrcReg, DestReg)) { 877 LivePhysRegs LiveAtMI(HRI); 878 getLiveInRegsAt(LiveAtMI, *I); 879 Register SrcLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo); 880 Register SrcHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi); 881 unsigned UndefLo = getUndefRegState(!LiveAtMI.contains(SrcLo)); 882 unsigned UndefHi = getUndefRegState(!LiveAtMI.contains(SrcHi)); 883 BuildMI(MBB, I, DL, get(Hexagon::V6_vcombine), DestReg) 884 .addReg(SrcHi, KillFlag | UndefHi) 885 .addReg(SrcLo, KillFlag | UndefLo); 886 return; 887 } 888 if (Hexagon::HvxQRRegClass.contains(SrcReg, DestReg)) { 889 BuildMI(MBB, I, DL, get(Hexagon::V6_pred_and), DestReg) 890 .addReg(SrcReg) 891 .addReg(SrcReg, KillFlag); 892 return; 893 } 894 if (Hexagon::HvxQRRegClass.contains(SrcReg) && 895 Hexagon::HvxVRRegClass.contains(DestReg)) { 896 llvm_unreachable("Unimplemented pred to vec"); 897 return; 898 } 899 if (Hexagon::HvxQRRegClass.contains(DestReg) && 900 Hexagon::HvxVRRegClass.contains(SrcReg)) { 901 llvm_unreachable("Unimplemented vec to pred"); 902 return; 903 } 904 905 #ifndef NDEBUG 906 // Show the invalid registers to ease debugging. 907 dbgs() << "Invalid registers for copy in " << printMBBReference(MBB) << ": " 908 << printReg(DestReg, &HRI) << " = " << printReg(SrcReg, &HRI) << '\n'; 909 #endif 910 llvm_unreachable("Unimplemented"); 911 } 912 913 void HexagonInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 914 MachineBasicBlock::iterator I, Register SrcReg, bool isKill, int FI, 915 const TargetRegisterClass *RC, const TargetRegisterInfo *TRI) const { 916 DebugLoc DL = MBB.findDebugLoc(I); 917 MachineFunction &MF = *MBB.getParent(); 918 MachineFrameInfo &MFI = MF.getFrameInfo(); 919 unsigned SlotAlign = MFI.getObjectAlignment(FI); 920 unsigned KillFlag = getKillRegState(isKill); 921 922 MachineMemOperand *MMO = MF.getMachineMemOperand( 923 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore, 924 MFI.getObjectSize(FI), SlotAlign); 925 926 if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) { 927 BuildMI(MBB, I, DL, get(Hexagon::S2_storeri_io)) 928 .addFrameIndex(FI).addImm(0) 929 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 930 } else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) { 931 BuildMI(MBB, I, DL, get(Hexagon::S2_storerd_io)) 932 .addFrameIndex(FI).addImm(0) 933 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 934 } else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) { 935 BuildMI(MBB, I, DL, get(Hexagon::STriw_pred)) 936 .addFrameIndex(FI).addImm(0) 937 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 938 } else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) { 939 BuildMI(MBB, I, DL, get(Hexagon::STriw_ctr)) 940 .addFrameIndex(FI).addImm(0) 941 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 942 } else if (Hexagon::HvxQRRegClass.hasSubClassEq(RC)) { 943 BuildMI(MBB, I, DL, get(Hexagon::PS_vstorerq_ai)) 944 .addFrameIndex(FI).addImm(0) 945 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 946 } else if (Hexagon::HvxVRRegClass.hasSubClassEq(RC)) { 947 BuildMI(MBB, I, DL, get(Hexagon::PS_vstorerv_ai)) 948 .addFrameIndex(FI).addImm(0) 949 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 950 } else if (Hexagon::HvxWRRegClass.hasSubClassEq(RC)) { 951 BuildMI(MBB, I, DL, get(Hexagon::PS_vstorerw_ai)) 952 .addFrameIndex(FI).addImm(0) 953 .addReg(SrcReg, KillFlag).addMemOperand(MMO); 954 } else { 955 llvm_unreachable("Unimplemented"); 956 } 957 } 958 959 void HexagonInstrInfo::loadRegFromStackSlot( 960 MachineBasicBlock &MBB, MachineBasicBlock::iterator I, Register DestReg, 961 int FI, const TargetRegisterClass *RC, 962 const TargetRegisterInfo *TRI) const { 963 DebugLoc DL = MBB.findDebugLoc(I); 964 MachineFunction &MF = *MBB.getParent(); 965 MachineFrameInfo &MFI = MF.getFrameInfo(); 966 unsigned SlotAlign = MFI.getObjectAlignment(FI); 967 968 MachineMemOperand *MMO = MF.getMachineMemOperand( 969 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad, 970 MFI.getObjectSize(FI), SlotAlign); 971 972 if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) { 973 BuildMI(MBB, I, DL, get(Hexagon::L2_loadri_io), DestReg) 974 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 975 } else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) { 976 BuildMI(MBB, I, DL, get(Hexagon::L2_loadrd_io), DestReg) 977 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 978 } else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) { 979 BuildMI(MBB, I, DL, get(Hexagon::LDriw_pred), DestReg) 980 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 981 } else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) { 982 BuildMI(MBB, I, DL, get(Hexagon::LDriw_ctr), DestReg) 983 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 984 } else if (Hexagon::HvxQRRegClass.hasSubClassEq(RC)) { 985 BuildMI(MBB, I, DL, get(Hexagon::PS_vloadrq_ai), DestReg) 986 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 987 } else if (Hexagon::HvxVRRegClass.hasSubClassEq(RC)) { 988 BuildMI(MBB, I, DL, get(Hexagon::PS_vloadrv_ai), DestReg) 989 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 990 } else if (Hexagon::HvxWRRegClass.hasSubClassEq(RC)) { 991 BuildMI(MBB, I, DL, get(Hexagon::PS_vloadrw_ai), DestReg) 992 .addFrameIndex(FI).addImm(0).addMemOperand(MMO); 993 } else { 994 llvm_unreachable("Can't store this register to stack slot"); 995 } 996 } 997 998 /// expandPostRAPseudo - This function is called for all pseudo instructions 999 /// that remain after register allocation. Many pseudo instructions are 1000 /// created to help register allocation. This is the place to convert them 1001 /// into real instructions. The target can edit MI in place, or it can insert 1002 /// new instructions and erase MI. The function should return true if 1003 /// anything was changed. 1004 bool HexagonInstrInfo::expandPostRAPseudo(MachineInstr &MI) const { 1005 MachineBasicBlock &MBB = *MI.getParent(); 1006 MachineFunction &MF = *MBB.getParent(); 1007 MachineRegisterInfo &MRI = MF.getRegInfo(); 1008 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 1009 LivePhysRegs LiveIn(HRI), LiveOut(HRI); 1010 DebugLoc DL = MI.getDebugLoc(); 1011 unsigned Opc = MI.getOpcode(); 1012 1013 auto RealCirc = [&](unsigned Opc, bool HasImm, unsigned MxOp) { 1014 Register Mx = MI.getOperand(MxOp).getReg(); 1015 unsigned CSx = (Mx == Hexagon::M0 ? Hexagon::CS0 : Hexagon::CS1); 1016 BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrrcr), CSx) 1017 .add(MI.getOperand((HasImm ? 5 : 4))); 1018 auto MIB = BuildMI(MBB, MI, DL, get(Opc)).add(MI.getOperand(0)) 1019 .add(MI.getOperand(1)).add(MI.getOperand(2)).add(MI.getOperand(3)); 1020 if (HasImm) 1021 MIB.add(MI.getOperand(4)); 1022 MIB.addReg(CSx, RegState::Implicit); 1023 MBB.erase(MI); 1024 return true; 1025 }; 1026 1027 auto UseAligned = [&] (const MachineInstr &MI, unsigned NeedAlign) { 1028 if (MI.memoperands().empty()) 1029 return false; 1030 return all_of(MI.memoperands(), [NeedAlign](const MachineMemOperand *MMO) { 1031 return MMO->getAlign() >= NeedAlign; 1032 }); 1033 }; 1034 1035 switch (Opc) { 1036 case TargetOpcode::COPY: { 1037 MachineOperand &MD = MI.getOperand(0); 1038 MachineOperand &MS = MI.getOperand(1); 1039 MachineBasicBlock::iterator MBBI = MI.getIterator(); 1040 if (MD.getReg() != MS.getReg() && !MS.isUndef()) { 1041 copyPhysReg(MBB, MI, DL, MD.getReg(), MS.getReg(), MS.isKill()); 1042 std::prev(MBBI)->copyImplicitOps(*MBB.getParent(), MI); 1043 } 1044 MBB.erase(MBBI); 1045 return true; 1046 } 1047 case Hexagon::PS_aligna: 1048 BuildMI(MBB, MI, DL, get(Hexagon::A2_andir), MI.getOperand(0).getReg()) 1049 .addReg(HRI.getFrameRegister()) 1050 .addImm(-MI.getOperand(1).getImm()); 1051 MBB.erase(MI); 1052 return true; 1053 case Hexagon::V6_vassignp: { 1054 Register SrcReg = MI.getOperand(1).getReg(); 1055 Register DstReg = MI.getOperand(0).getReg(); 1056 Register SrcLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo); 1057 Register SrcHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi); 1058 getLiveInRegsAt(LiveIn, MI); 1059 unsigned UndefLo = getUndefRegState(!LiveIn.contains(SrcLo)); 1060 unsigned UndefHi = getUndefRegState(!LiveIn.contains(SrcHi)); 1061 unsigned Kill = getKillRegState(MI.getOperand(1).isKill()); 1062 BuildMI(MBB, MI, DL, get(Hexagon::V6_vcombine), DstReg) 1063 .addReg(SrcHi, UndefHi) 1064 .addReg(SrcLo, Kill | UndefLo); 1065 MBB.erase(MI); 1066 return true; 1067 } 1068 case Hexagon::V6_lo: { 1069 Register SrcReg = MI.getOperand(1).getReg(); 1070 Register DstReg = MI.getOperand(0).getReg(); 1071 Register SrcSubLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo); 1072 copyPhysReg(MBB, MI, DL, DstReg, SrcSubLo, MI.getOperand(1).isKill()); 1073 MBB.erase(MI); 1074 MRI.clearKillFlags(SrcSubLo); 1075 return true; 1076 } 1077 case Hexagon::V6_hi: { 1078 Register SrcReg = MI.getOperand(1).getReg(); 1079 Register DstReg = MI.getOperand(0).getReg(); 1080 Register SrcSubHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi); 1081 copyPhysReg(MBB, MI, DL, DstReg, SrcSubHi, MI.getOperand(1).isKill()); 1082 MBB.erase(MI); 1083 MRI.clearKillFlags(SrcSubHi); 1084 return true; 1085 } 1086 case Hexagon::PS_vloadrv_ai: { 1087 Register DstReg = MI.getOperand(0).getReg(); 1088 const MachineOperand &BaseOp = MI.getOperand(1); 1089 assert(BaseOp.getSubReg() == 0); 1090 int Offset = MI.getOperand(2).getImm(); 1091 unsigned NeedAlign = HRI.getSpillAlignment(Hexagon::HvxVRRegClass); 1092 unsigned NewOpc = UseAligned(MI, NeedAlign) ? Hexagon::V6_vL32b_ai 1093 : Hexagon::V6_vL32Ub_ai; 1094 BuildMI(MBB, MI, DL, get(NewOpc), DstReg) 1095 .addReg(BaseOp.getReg(), getRegState(BaseOp)) 1096 .addImm(Offset) 1097 .cloneMemRefs(MI); 1098 MBB.erase(MI); 1099 return true; 1100 } 1101 case Hexagon::PS_vloadrw_ai: { 1102 Register DstReg = MI.getOperand(0).getReg(); 1103 const MachineOperand &BaseOp = MI.getOperand(1); 1104 assert(BaseOp.getSubReg() == 0); 1105 int Offset = MI.getOperand(2).getImm(); 1106 unsigned VecOffset = HRI.getSpillSize(Hexagon::HvxVRRegClass); 1107 unsigned NeedAlign = HRI.getSpillAlignment(Hexagon::HvxVRRegClass); 1108 unsigned NewOpc = UseAligned(MI, NeedAlign) ? Hexagon::V6_vL32b_ai 1109 : Hexagon::V6_vL32Ub_ai; 1110 BuildMI(MBB, MI, DL, get(NewOpc), 1111 HRI.getSubReg(DstReg, Hexagon::vsub_lo)) 1112 .addReg(BaseOp.getReg(), getRegState(BaseOp) & ~RegState::Kill) 1113 .addImm(Offset) 1114 .cloneMemRefs(MI); 1115 BuildMI(MBB, MI, DL, get(NewOpc), 1116 HRI.getSubReg(DstReg, Hexagon::vsub_hi)) 1117 .addReg(BaseOp.getReg(), getRegState(BaseOp)) 1118 .addImm(Offset + VecOffset) 1119 .cloneMemRefs(MI); 1120 MBB.erase(MI); 1121 return true; 1122 } 1123 case Hexagon::PS_vstorerv_ai: { 1124 const MachineOperand &SrcOp = MI.getOperand(2); 1125 assert(SrcOp.getSubReg() == 0); 1126 const MachineOperand &BaseOp = MI.getOperand(0); 1127 assert(BaseOp.getSubReg() == 0); 1128 int Offset = MI.getOperand(1).getImm(); 1129 unsigned NeedAlign = HRI.getSpillAlignment(Hexagon::HvxVRRegClass); 1130 unsigned NewOpc = UseAligned(MI, NeedAlign) ? Hexagon::V6_vS32b_ai 1131 : Hexagon::V6_vS32Ub_ai; 1132 BuildMI(MBB, MI, DL, get(NewOpc)) 1133 .addReg(BaseOp.getReg(), getRegState(BaseOp)) 1134 .addImm(Offset) 1135 .addReg(SrcOp.getReg(), getRegState(SrcOp)) 1136 .cloneMemRefs(MI); 1137 MBB.erase(MI); 1138 return true; 1139 } 1140 case Hexagon::PS_vstorerw_ai: { 1141 Register SrcReg = MI.getOperand(2).getReg(); 1142 const MachineOperand &BaseOp = MI.getOperand(0); 1143 assert(BaseOp.getSubReg() == 0); 1144 int Offset = MI.getOperand(1).getImm(); 1145 unsigned VecOffset = HRI.getSpillSize(Hexagon::HvxVRRegClass); 1146 unsigned NeedAlign = HRI.getSpillAlignment(Hexagon::HvxVRRegClass); 1147 unsigned NewOpc = UseAligned(MI, NeedAlign) ? Hexagon::V6_vS32b_ai 1148 : Hexagon::V6_vS32Ub_ai; 1149 BuildMI(MBB, MI, DL, get(NewOpc)) 1150 .addReg(BaseOp.getReg(), getRegState(BaseOp) & ~RegState::Kill) 1151 .addImm(Offset) 1152 .addReg(HRI.getSubReg(SrcReg, Hexagon::vsub_lo)) 1153 .cloneMemRefs(MI); 1154 BuildMI(MBB, MI, DL, get(NewOpc)) 1155 .addReg(BaseOp.getReg(), getRegState(BaseOp)) 1156 .addImm(Offset + VecOffset) 1157 .addReg(HRI.getSubReg(SrcReg, Hexagon::vsub_hi)) 1158 .cloneMemRefs(MI); 1159 MBB.erase(MI); 1160 return true; 1161 } 1162 case Hexagon::PS_true: { 1163 Register Reg = MI.getOperand(0).getReg(); 1164 BuildMI(MBB, MI, DL, get(Hexagon::C2_orn), Reg) 1165 .addReg(Reg, RegState::Undef) 1166 .addReg(Reg, RegState::Undef); 1167 MBB.erase(MI); 1168 return true; 1169 } 1170 case Hexagon::PS_false: { 1171 Register Reg = MI.getOperand(0).getReg(); 1172 BuildMI(MBB, MI, DL, get(Hexagon::C2_andn), Reg) 1173 .addReg(Reg, RegState::Undef) 1174 .addReg(Reg, RegState::Undef); 1175 MBB.erase(MI); 1176 return true; 1177 } 1178 case Hexagon::PS_qtrue: { 1179 BuildMI(MBB, MI, DL, get(Hexagon::V6_veqw), MI.getOperand(0).getReg()) 1180 .addReg(Hexagon::V0, RegState::Undef) 1181 .addReg(Hexagon::V0, RegState::Undef); 1182 MBB.erase(MI); 1183 return true; 1184 } 1185 case Hexagon::PS_qfalse: { 1186 BuildMI(MBB, MI, DL, get(Hexagon::V6_vgtw), MI.getOperand(0).getReg()) 1187 .addReg(Hexagon::V0, RegState::Undef) 1188 .addReg(Hexagon::V0, RegState::Undef); 1189 MBB.erase(MI); 1190 return true; 1191 } 1192 case Hexagon::PS_vdd0: { 1193 Register Vd = MI.getOperand(0).getReg(); 1194 BuildMI(MBB, MI, DL, get(Hexagon::V6_vsubw_dv), Vd) 1195 .addReg(Vd, RegState::Undef) 1196 .addReg(Vd, RegState::Undef); 1197 MBB.erase(MI); 1198 return true; 1199 } 1200 case Hexagon::PS_vmulw: { 1201 // Expand a 64-bit vector multiply into 2 32-bit scalar multiplies. 1202 Register DstReg = MI.getOperand(0).getReg(); 1203 Register Src1Reg = MI.getOperand(1).getReg(); 1204 Register Src2Reg = MI.getOperand(2).getReg(); 1205 Register Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::isub_hi); 1206 Register Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::isub_lo); 1207 Register Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::isub_hi); 1208 Register Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::isub_lo); 1209 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_mpyi), 1210 HRI.getSubReg(DstReg, Hexagon::isub_hi)) 1211 .addReg(Src1SubHi) 1212 .addReg(Src2SubHi); 1213 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_mpyi), 1214 HRI.getSubReg(DstReg, Hexagon::isub_lo)) 1215 .addReg(Src1SubLo) 1216 .addReg(Src2SubLo); 1217 MBB.erase(MI); 1218 MRI.clearKillFlags(Src1SubHi); 1219 MRI.clearKillFlags(Src1SubLo); 1220 MRI.clearKillFlags(Src2SubHi); 1221 MRI.clearKillFlags(Src2SubLo); 1222 return true; 1223 } 1224 case Hexagon::PS_vmulw_acc: { 1225 // Expand 64-bit vector multiply with addition into 2 scalar multiplies. 1226 Register DstReg = MI.getOperand(0).getReg(); 1227 Register Src1Reg = MI.getOperand(1).getReg(); 1228 Register Src2Reg = MI.getOperand(2).getReg(); 1229 Register Src3Reg = MI.getOperand(3).getReg(); 1230 Register Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::isub_hi); 1231 Register Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::isub_lo); 1232 Register Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::isub_hi); 1233 Register Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::isub_lo); 1234 Register Src3SubHi = HRI.getSubReg(Src3Reg, Hexagon::isub_hi); 1235 Register Src3SubLo = HRI.getSubReg(Src3Reg, Hexagon::isub_lo); 1236 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_maci), 1237 HRI.getSubReg(DstReg, Hexagon::isub_hi)) 1238 .addReg(Src1SubHi) 1239 .addReg(Src2SubHi) 1240 .addReg(Src3SubHi); 1241 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_maci), 1242 HRI.getSubReg(DstReg, Hexagon::isub_lo)) 1243 .addReg(Src1SubLo) 1244 .addReg(Src2SubLo) 1245 .addReg(Src3SubLo); 1246 MBB.erase(MI); 1247 MRI.clearKillFlags(Src1SubHi); 1248 MRI.clearKillFlags(Src1SubLo); 1249 MRI.clearKillFlags(Src2SubHi); 1250 MRI.clearKillFlags(Src2SubLo); 1251 MRI.clearKillFlags(Src3SubHi); 1252 MRI.clearKillFlags(Src3SubLo); 1253 return true; 1254 } 1255 case Hexagon::PS_pselect: { 1256 const MachineOperand &Op0 = MI.getOperand(0); 1257 const MachineOperand &Op1 = MI.getOperand(1); 1258 const MachineOperand &Op2 = MI.getOperand(2); 1259 const MachineOperand &Op3 = MI.getOperand(3); 1260 Register Rd = Op0.getReg(); 1261 Register Pu = Op1.getReg(); 1262 Register Rs = Op2.getReg(); 1263 Register Rt = Op3.getReg(); 1264 DebugLoc DL = MI.getDebugLoc(); 1265 unsigned K1 = getKillRegState(Op1.isKill()); 1266 unsigned K2 = getKillRegState(Op2.isKill()); 1267 unsigned K3 = getKillRegState(Op3.isKill()); 1268 if (Rd != Rs) 1269 BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpt), Rd) 1270 .addReg(Pu, (Rd == Rt) ? K1 : 0) 1271 .addReg(Rs, K2); 1272 if (Rd != Rt) 1273 BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpf), Rd) 1274 .addReg(Pu, K1) 1275 .addReg(Rt, K3); 1276 MBB.erase(MI); 1277 return true; 1278 } 1279 case Hexagon::PS_vselect: { 1280 const MachineOperand &Op0 = MI.getOperand(0); 1281 const MachineOperand &Op1 = MI.getOperand(1); 1282 const MachineOperand &Op2 = MI.getOperand(2); 1283 const MachineOperand &Op3 = MI.getOperand(3); 1284 getLiveOutRegsAt(LiveOut, MI); 1285 bool IsDestLive = !LiveOut.available(MRI, Op0.getReg()); 1286 Register PReg = Op1.getReg(); 1287 assert(Op1.getSubReg() == 0); 1288 unsigned PState = getRegState(Op1); 1289 1290 if (Op0.getReg() != Op2.getReg()) { 1291 unsigned S = Op0.getReg() != Op3.getReg() ? PState & ~RegState::Kill 1292 : PState; 1293 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vcmov)) 1294 .add(Op0) 1295 .addReg(PReg, S) 1296 .add(Op2); 1297 if (IsDestLive) 1298 T.addReg(Op0.getReg(), RegState::Implicit); 1299 IsDestLive = true; 1300 } 1301 if (Op0.getReg() != Op3.getReg()) { 1302 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vncmov)) 1303 .add(Op0) 1304 .addReg(PReg, PState) 1305 .add(Op3); 1306 if (IsDestLive) 1307 T.addReg(Op0.getReg(), RegState::Implicit); 1308 } 1309 MBB.erase(MI); 1310 return true; 1311 } 1312 case Hexagon::PS_wselect: { 1313 MachineOperand &Op0 = MI.getOperand(0); 1314 MachineOperand &Op1 = MI.getOperand(1); 1315 MachineOperand &Op2 = MI.getOperand(2); 1316 MachineOperand &Op3 = MI.getOperand(3); 1317 getLiveOutRegsAt(LiveOut, MI); 1318 bool IsDestLive = !LiveOut.available(MRI, Op0.getReg()); 1319 Register PReg = Op1.getReg(); 1320 assert(Op1.getSubReg() == 0); 1321 unsigned PState = getRegState(Op1); 1322 1323 if (Op0.getReg() != Op2.getReg()) { 1324 unsigned S = Op0.getReg() != Op3.getReg() ? PState & ~RegState::Kill 1325 : PState; 1326 Register SrcLo = HRI.getSubReg(Op2.getReg(), Hexagon::vsub_lo); 1327 Register SrcHi = HRI.getSubReg(Op2.getReg(), Hexagon::vsub_hi); 1328 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vccombine)) 1329 .add(Op0) 1330 .addReg(PReg, S) 1331 .addReg(SrcHi) 1332 .addReg(SrcLo); 1333 if (IsDestLive) 1334 T.addReg(Op0.getReg(), RegState::Implicit); 1335 IsDestLive = true; 1336 } 1337 if (Op0.getReg() != Op3.getReg()) { 1338 Register SrcLo = HRI.getSubReg(Op3.getReg(), Hexagon::vsub_lo); 1339 Register SrcHi = HRI.getSubReg(Op3.getReg(), Hexagon::vsub_hi); 1340 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vnccombine)) 1341 .add(Op0) 1342 .addReg(PReg, PState) 1343 .addReg(SrcHi) 1344 .addReg(SrcLo); 1345 if (IsDestLive) 1346 T.addReg(Op0.getReg(), RegState::Implicit); 1347 } 1348 MBB.erase(MI); 1349 return true; 1350 } 1351 1352 case Hexagon::PS_crash: { 1353 // Generate a misaligned load that is guaranteed to cause a crash. 1354 class CrashPseudoSourceValue : public PseudoSourceValue { 1355 public: 1356 CrashPseudoSourceValue(const TargetInstrInfo &TII) 1357 : PseudoSourceValue(TargetCustom, TII) {} 1358 1359 bool isConstant(const MachineFrameInfo *) const override { 1360 return false; 1361 } 1362 bool isAliased(const MachineFrameInfo *) const override { 1363 return false; 1364 } 1365 bool mayAlias(const MachineFrameInfo *) const override { 1366 return false; 1367 } 1368 void printCustom(raw_ostream &OS) const override { 1369 OS << "MisalignedCrash"; 1370 } 1371 }; 1372 1373 static const CrashPseudoSourceValue CrashPSV(*this); 1374 MachineMemOperand *MMO = MF.getMachineMemOperand( 1375 MachinePointerInfo(&CrashPSV), 1376 MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile, 8, 1); 1377 BuildMI(MBB, MI, DL, get(Hexagon::PS_loadrdabs), Hexagon::D13) 1378 .addImm(0xBADC0FEE) // Misaligned load. 1379 .addMemOperand(MMO); 1380 MBB.erase(MI); 1381 return true; 1382 } 1383 1384 case Hexagon::PS_tailcall_i: 1385 MI.setDesc(get(Hexagon::J2_jump)); 1386 return true; 1387 case Hexagon::PS_tailcall_r: 1388 case Hexagon::PS_jmpret: 1389 MI.setDesc(get(Hexagon::J2_jumpr)); 1390 return true; 1391 case Hexagon::PS_jmprett: 1392 MI.setDesc(get(Hexagon::J2_jumprt)); 1393 return true; 1394 case Hexagon::PS_jmpretf: 1395 MI.setDesc(get(Hexagon::J2_jumprf)); 1396 return true; 1397 case Hexagon::PS_jmprettnewpt: 1398 MI.setDesc(get(Hexagon::J2_jumprtnewpt)); 1399 return true; 1400 case Hexagon::PS_jmpretfnewpt: 1401 MI.setDesc(get(Hexagon::J2_jumprfnewpt)); 1402 return true; 1403 case Hexagon::PS_jmprettnew: 1404 MI.setDesc(get(Hexagon::J2_jumprtnew)); 1405 return true; 1406 case Hexagon::PS_jmpretfnew: 1407 MI.setDesc(get(Hexagon::J2_jumprfnew)); 1408 return true; 1409 1410 case Hexagon::PS_loadrub_pci: 1411 return RealCirc(Hexagon::L2_loadrub_pci, /*HasImm*/true, /*MxOp*/4); 1412 case Hexagon::PS_loadrb_pci: 1413 return RealCirc(Hexagon::L2_loadrb_pci, /*HasImm*/true, /*MxOp*/4); 1414 case Hexagon::PS_loadruh_pci: 1415 return RealCirc(Hexagon::L2_loadruh_pci, /*HasImm*/true, /*MxOp*/4); 1416 case Hexagon::PS_loadrh_pci: 1417 return RealCirc(Hexagon::L2_loadrh_pci, /*HasImm*/true, /*MxOp*/4); 1418 case Hexagon::PS_loadri_pci: 1419 return RealCirc(Hexagon::L2_loadri_pci, /*HasImm*/true, /*MxOp*/4); 1420 case Hexagon::PS_loadrd_pci: 1421 return RealCirc(Hexagon::L2_loadrd_pci, /*HasImm*/true, /*MxOp*/4); 1422 case Hexagon::PS_loadrub_pcr: 1423 return RealCirc(Hexagon::L2_loadrub_pcr, /*HasImm*/false, /*MxOp*/3); 1424 case Hexagon::PS_loadrb_pcr: 1425 return RealCirc(Hexagon::L2_loadrb_pcr, /*HasImm*/false, /*MxOp*/3); 1426 case Hexagon::PS_loadruh_pcr: 1427 return RealCirc(Hexagon::L2_loadruh_pcr, /*HasImm*/false, /*MxOp*/3); 1428 case Hexagon::PS_loadrh_pcr: 1429 return RealCirc(Hexagon::L2_loadrh_pcr, /*HasImm*/false, /*MxOp*/3); 1430 case Hexagon::PS_loadri_pcr: 1431 return RealCirc(Hexagon::L2_loadri_pcr, /*HasImm*/false, /*MxOp*/3); 1432 case Hexagon::PS_loadrd_pcr: 1433 return RealCirc(Hexagon::L2_loadrd_pcr, /*HasImm*/false, /*MxOp*/3); 1434 case Hexagon::PS_storerb_pci: 1435 return RealCirc(Hexagon::S2_storerb_pci, /*HasImm*/true, /*MxOp*/3); 1436 case Hexagon::PS_storerh_pci: 1437 return RealCirc(Hexagon::S2_storerh_pci, /*HasImm*/true, /*MxOp*/3); 1438 case Hexagon::PS_storerf_pci: 1439 return RealCirc(Hexagon::S2_storerf_pci, /*HasImm*/true, /*MxOp*/3); 1440 case Hexagon::PS_storeri_pci: 1441 return RealCirc(Hexagon::S2_storeri_pci, /*HasImm*/true, /*MxOp*/3); 1442 case Hexagon::PS_storerd_pci: 1443 return RealCirc(Hexagon::S2_storerd_pci, /*HasImm*/true, /*MxOp*/3); 1444 case Hexagon::PS_storerb_pcr: 1445 return RealCirc(Hexagon::S2_storerb_pcr, /*HasImm*/false, /*MxOp*/2); 1446 case Hexagon::PS_storerh_pcr: 1447 return RealCirc(Hexagon::S2_storerh_pcr, /*HasImm*/false, /*MxOp*/2); 1448 case Hexagon::PS_storerf_pcr: 1449 return RealCirc(Hexagon::S2_storerf_pcr, /*HasImm*/false, /*MxOp*/2); 1450 case Hexagon::PS_storeri_pcr: 1451 return RealCirc(Hexagon::S2_storeri_pcr, /*HasImm*/false, /*MxOp*/2); 1452 case Hexagon::PS_storerd_pcr: 1453 return RealCirc(Hexagon::S2_storerd_pcr, /*HasImm*/false, /*MxOp*/2); 1454 } 1455 1456 return false; 1457 } 1458 1459 MachineBasicBlock::instr_iterator 1460 HexagonInstrInfo::expandVGatherPseudo(MachineInstr &MI) const { 1461 MachineBasicBlock &MBB = *MI.getParent(); 1462 const DebugLoc &DL = MI.getDebugLoc(); 1463 unsigned Opc = MI.getOpcode(); 1464 MachineBasicBlock::iterator First; 1465 1466 switch (Opc) { 1467 case Hexagon::V6_vgathermh_pseudo: 1468 First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermh)) 1469 .add(MI.getOperand(1)) 1470 .add(MI.getOperand(2)) 1471 .add(MI.getOperand(3)); 1472 BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) 1473 .add(MI.getOperand(0)) 1474 .addImm(0) 1475 .addReg(Hexagon::VTMP); 1476 MBB.erase(MI); 1477 return First.getInstrIterator(); 1478 1479 case Hexagon::V6_vgathermw_pseudo: 1480 First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermw)) 1481 .add(MI.getOperand(1)) 1482 .add(MI.getOperand(2)) 1483 .add(MI.getOperand(3)); 1484 BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) 1485 .add(MI.getOperand(0)) 1486 .addImm(0) 1487 .addReg(Hexagon::VTMP); 1488 MBB.erase(MI); 1489 return First.getInstrIterator(); 1490 1491 case Hexagon::V6_vgathermhw_pseudo: 1492 First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhw)) 1493 .add(MI.getOperand(1)) 1494 .add(MI.getOperand(2)) 1495 .add(MI.getOperand(3)); 1496 BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) 1497 .add(MI.getOperand(0)) 1498 .addImm(0) 1499 .addReg(Hexagon::VTMP); 1500 MBB.erase(MI); 1501 return First.getInstrIterator(); 1502 1503 case Hexagon::V6_vgathermhq_pseudo: 1504 First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhq)) 1505 .add(MI.getOperand(1)) 1506 .add(MI.getOperand(2)) 1507 .add(MI.getOperand(3)) 1508 .add(MI.getOperand(4)); 1509 BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) 1510 .add(MI.getOperand(0)) 1511 .addImm(0) 1512 .addReg(Hexagon::VTMP); 1513 MBB.erase(MI); 1514 return First.getInstrIterator(); 1515 1516 case Hexagon::V6_vgathermwq_pseudo: 1517 First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermwq)) 1518 .add(MI.getOperand(1)) 1519 .add(MI.getOperand(2)) 1520 .add(MI.getOperand(3)) 1521 .add(MI.getOperand(4)); 1522 BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) 1523 .add(MI.getOperand(0)) 1524 .addImm(0) 1525 .addReg(Hexagon::VTMP); 1526 MBB.erase(MI); 1527 return First.getInstrIterator(); 1528 1529 case Hexagon::V6_vgathermhwq_pseudo: 1530 First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhwq)) 1531 .add(MI.getOperand(1)) 1532 .add(MI.getOperand(2)) 1533 .add(MI.getOperand(3)) 1534 .add(MI.getOperand(4)); 1535 BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai)) 1536 .add(MI.getOperand(0)) 1537 .addImm(0) 1538 .addReg(Hexagon::VTMP); 1539 MBB.erase(MI); 1540 return First.getInstrIterator(); 1541 } 1542 1543 return MI.getIterator(); 1544 } 1545 1546 // We indicate that we want to reverse the branch by 1547 // inserting the reversed branching opcode. 1548 bool HexagonInstrInfo::reverseBranchCondition( 1549 SmallVectorImpl<MachineOperand> &Cond) const { 1550 if (Cond.empty()) 1551 return true; 1552 assert(Cond[0].isImm() && "First entry in the cond vector not imm-val"); 1553 unsigned opcode = Cond[0].getImm(); 1554 //unsigned temp; 1555 assert(get(opcode).isBranch() && "Should be a branching condition."); 1556 if (isEndLoopN(opcode)) 1557 return true; 1558 unsigned NewOpcode = getInvertedPredicatedOpcode(opcode); 1559 Cond[0].setImm(NewOpcode); 1560 return false; 1561 } 1562 1563 void HexagonInstrInfo::insertNoop(MachineBasicBlock &MBB, 1564 MachineBasicBlock::iterator MI) const { 1565 DebugLoc DL; 1566 BuildMI(MBB, MI, DL, get(Hexagon::A2_nop)); 1567 } 1568 1569 bool HexagonInstrInfo::isPostIncrement(const MachineInstr &MI) const { 1570 return getAddrMode(MI) == HexagonII::PostInc; 1571 } 1572 1573 // Returns true if an instruction is predicated irrespective of the predicate 1574 // sense. For example, all of the following will return true. 1575 // if (p0) R1 = add(R2, R3) 1576 // if (!p0) R1 = add(R2, R3) 1577 // if (p0.new) R1 = add(R2, R3) 1578 // if (!p0.new) R1 = add(R2, R3) 1579 // Note: New-value stores are not included here as in the current 1580 // implementation, we don't need to check their predicate sense. 1581 bool HexagonInstrInfo::isPredicated(const MachineInstr &MI) const { 1582 const uint64_t F = MI.getDesc().TSFlags; 1583 return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask; 1584 } 1585 1586 bool HexagonInstrInfo::PredicateInstruction( 1587 MachineInstr &MI, ArrayRef<MachineOperand> Cond) const { 1588 if (Cond.empty() || isNewValueJump(Cond[0].getImm()) || 1589 isEndLoopN(Cond[0].getImm())) { 1590 LLVM_DEBUG(dbgs() << "\nCannot predicate:"; MI.dump();); 1591 return false; 1592 } 1593 int Opc = MI.getOpcode(); 1594 assert (isPredicable(MI) && "Expected predicable instruction"); 1595 bool invertJump = predOpcodeHasNot(Cond); 1596 1597 // We have to predicate MI "in place", i.e. after this function returns, 1598 // MI will need to be transformed into a predicated form. To avoid com- 1599 // plicated manipulations with the operands (handling tied operands, 1600 // etc.), build a new temporary instruction, then overwrite MI with it. 1601 1602 MachineBasicBlock &B = *MI.getParent(); 1603 DebugLoc DL = MI.getDebugLoc(); 1604 unsigned PredOpc = getCondOpcode(Opc, invertJump); 1605 MachineInstrBuilder T = BuildMI(B, MI, DL, get(PredOpc)); 1606 unsigned NOp = 0, NumOps = MI.getNumOperands(); 1607 while (NOp < NumOps) { 1608 MachineOperand &Op = MI.getOperand(NOp); 1609 if (!Op.isReg() || !Op.isDef() || Op.isImplicit()) 1610 break; 1611 T.add(Op); 1612 NOp++; 1613 } 1614 1615 unsigned PredReg, PredRegPos, PredRegFlags; 1616 bool GotPredReg = getPredReg(Cond, PredReg, PredRegPos, PredRegFlags); 1617 (void)GotPredReg; 1618 assert(GotPredReg); 1619 T.addReg(PredReg, PredRegFlags); 1620 while (NOp < NumOps) 1621 T.add(MI.getOperand(NOp++)); 1622 1623 MI.setDesc(get(PredOpc)); 1624 while (unsigned n = MI.getNumOperands()) 1625 MI.RemoveOperand(n-1); 1626 for (unsigned i = 0, n = T->getNumOperands(); i < n; ++i) 1627 MI.addOperand(T->getOperand(i)); 1628 1629 MachineBasicBlock::instr_iterator TI = T->getIterator(); 1630 B.erase(TI); 1631 1632 MachineRegisterInfo &MRI = B.getParent()->getRegInfo(); 1633 MRI.clearKillFlags(PredReg); 1634 return true; 1635 } 1636 1637 bool HexagonInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1, 1638 ArrayRef<MachineOperand> Pred2) const { 1639 // TODO: Fix this 1640 return false; 1641 } 1642 1643 bool HexagonInstrInfo::DefinesPredicate(MachineInstr &MI, 1644 std::vector<MachineOperand> &Pred) const { 1645 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 1646 1647 for (unsigned oper = 0; oper < MI.getNumOperands(); ++oper) { 1648 MachineOperand MO = MI.getOperand(oper); 1649 if (MO.isReg()) { 1650 if (!MO.isDef()) 1651 continue; 1652 const TargetRegisterClass* RC = HRI.getMinimalPhysRegClass(MO.getReg()); 1653 if (RC == &Hexagon::PredRegsRegClass) { 1654 Pred.push_back(MO); 1655 return true; 1656 } 1657 continue; 1658 } else if (MO.isRegMask()) { 1659 for (unsigned PR : Hexagon::PredRegsRegClass) { 1660 if (!MI.modifiesRegister(PR, &HRI)) 1661 continue; 1662 Pred.push_back(MO); 1663 return true; 1664 } 1665 } 1666 } 1667 return false; 1668 } 1669 1670 bool HexagonInstrInfo::isPredicable(const MachineInstr &MI) const { 1671 if (!MI.getDesc().isPredicable()) 1672 return false; 1673 1674 if (MI.isCall() || isTailCall(MI)) { 1675 if (!Subtarget.usePredicatedCalls()) 1676 return false; 1677 } 1678 1679 // HVX loads are not predicable on v60, but are on v62. 1680 if (!Subtarget.hasV62Ops()) { 1681 switch (MI.getOpcode()) { 1682 case Hexagon::V6_vL32b_ai: 1683 case Hexagon::V6_vL32b_pi: 1684 case Hexagon::V6_vL32b_ppu: 1685 case Hexagon::V6_vL32b_cur_ai: 1686 case Hexagon::V6_vL32b_cur_pi: 1687 case Hexagon::V6_vL32b_cur_ppu: 1688 case Hexagon::V6_vL32b_nt_ai: 1689 case Hexagon::V6_vL32b_nt_pi: 1690 case Hexagon::V6_vL32b_nt_ppu: 1691 case Hexagon::V6_vL32b_tmp_ai: 1692 case Hexagon::V6_vL32b_tmp_pi: 1693 case Hexagon::V6_vL32b_tmp_ppu: 1694 case Hexagon::V6_vL32b_nt_cur_ai: 1695 case Hexagon::V6_vL32b_nt_cur_pi: 1696 case Hexagon::V6_vL32b_nt_cur_ppu: 1697 case Hexagon::V6_vL32b_nt_tmp_ai: 1698 case Hexagon::V6_vL32b_nt_tmp_pi: 1699 case Hexagon::V6_vL32b_nt_tmp_ppu: 1700 return false; 1701 } 1702 } 1703 return true; 1704 } 1705 1706 bool HexagonInstrInfo::isSchedulingBoundary(const MachineInstr &MI, 1707 const MachineBasicBlock *MBB, 1708 const MachineFunction &MF) const { 1709 // Debug info is never a scheduling boundary. It's necessary to be explicit 1710 // due to the special treatment of IT instructions below, otherwise a 1711 // dbg_value followed by an IT will result in the IT instruction being 1712 // considered a scheduling hazard, which is wrong. It should be the actual 1713 // instruction preceding the dbg_value instruction(s), just like it is 1714 // when debug info is not present. 1715 if (MI.isDebugInstr()) 1716 return false; 1717 1718 // Throwing call is a boundary. 1719 if (MI.isCall()) { 1720 // Don't mess around with no return calls. 1721 if (doesNotReturn(MI)) 1722 return true; 1723 // If any of the block's successors is a landing pad, this could be a 1724 // throwing call. 1725 for (auto I : MBB->successors()) 1726 if (I->isEHPad()) 1727 return true; 1728 } 1729 1730 // Terminators and labels can't be scheduled around. 1731 if (MI.getDesc().isTerminator() || MI.isPosition()) 1732 return true; 1733 1734 if (MI.isInlineAsm() && !ScheduleInlineAsm) 1735 return true; 1736 1737 return false; 1738 } 1739 1740 /// Measure the specified inline asm to determine an approximation of its 1741 /// length. 1742 /// Comments (which run till the next SeparatorString or newline) do not 1743 /// count as an instruction. 1744 /// Any other non-whitespace text is considered an instruction, with 1745 /// multiple instructions separated by SeparatorString or newlines. 1746 /// Variable-length instructions are not handled here; this function 1747 /// may be overloaded in the target code to do that. 1748 /// Hexagon counts the number of ##'s and adjust for that many 1749 /// constant exenders. 1750 unsigned HexagonInstrInfo::getInlineAsmLength(const char *Str, 1751 const MCAsmInfo &MAI, 1752 const TargetSubtargetInfo *STI) const { 1753 StringRef AStr(Str); 1754 // Count the number of instructions in the asm. 1755 bool atInsnStart = true; 1756 unsigned Length = 0; 1757 const unsigned MaxInstLength = MAI.getMaxInstLength(STI); 1758 for (; *Str; ++Str) { 1759 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(), 1760 strlen(MAI.getSeparatorString())) == 0) 1761 atInsnStart = true; 1762 if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) { 1763 Length += MaxInstLength; 1764 atInsnStart = false; 1765 } 1766 if (atInsnStart && strncmp(Str, MAI.getCommentString().data(), 1767 MAI.getCommentString().size()) == 0) 1768 atInsnStart = false; 1769 } 1770 1771 // Add to size number of constant extenders seen * 4. 1772 StringRef Occ("##"); 1773 Length += AStr.count(Occ)*4; 1774 return Length; 1775 } 1776 1777 ScheduleHazardRecognizer* 1778 HexagonInstrInfo::CreateTargetPostRAHazardRecognizer( 1779 const InstrItineraryData *II, const ScheduleDAG *DAG) const { 1780 if (UseDFAHazardRec) 1781 return new HexagonHazardRecognizer(II, this, Subtarget); 1782 return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG); 1783 } 1784 1785 /// For a comparison instruction, return the source registers in 1786 /// \p SrcReg and \p SrcReg2 if having two register operands, and the value it 1787 /// compares against in CmpValue. Return true if the comparison instruction 1788 /// can be analyzed. 1789 bool HexagonInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg, 1790 unsigned &SrcReg2, int &Mask, 1791 int &Value) const { 1792 unsigned Opc = MI.getOpcode(); 1793 1794 // Set mask and the first source register. 1795 switch (Opc) { 1796 case Hexagon::C2_cmpeq: 1797 case Hexagon::C2_cmpeqp: 1798 case Hexagon::C2_cmpgt: 1799 case Hexagon::C2_cmpgtp: 1800 case Hexagon::C2_cmpgtu: 1801 case Hexagon::C2_cmpgtup: 1802 case Hexagon::C4_cmpneq: 1803 case Hexagon::C4_cmplte: 1804 case Hexagon::C4_cmplteu: 1805 case Hexagon::C2_cmpeqi: 1806 case Hexagon::C2_cmpgti: 1807 case Hexagon::C2_cmpgtui: 1808 case Hexagon::C4_cmpneqi: 1809 case Hexagon::C4_cmplteui: 1810 case Hexagon::C4_cmpltei: 1811 SrcReg = MI.getOperand(1).getReg(); 1812 Mask = ~0; 1813 break; 1814 case Hexagon::A4_cmpbeq: 1815 case Hexagon::A4_cmpbgt: 1816 case Hexagon::A4_cmpbgtu: 1817 case Hexagon::A4_cmpbeqi: 1818 case Hexagon::A4_cmpbgti: 1819 case Hexagon::A4_cmpbgtui: 1820 SrcReg = MI.getOperand(1).getReg(); 1821 Mask = 0xFF; 1822 break; 1823 case Hexagon::A4_cmpheq: 1824 case Hexagon::A4_cmphgt: 1825 case Hexagon::A4_cmphgtu: 1826 case Hexagon::A4_cmpheqi: 1827 case Hexagon::A4_cmphgti: 1828 case Hexagon::A4_cmphgtui: 1829 SrcReg = MI.getOperand(1).getReg(); 1830 Mask = 0xFFFF; 1831 break; 1832 } 1833 1834 // Set the value/second source register. 1835 switch (Opc) { 1836 case Hexagon::C2_cmpeq: 1837 case Hexagon::C2_cmpeqp: 1838 case Hexagon::C2_cmpgt: 1839 case Hexagon::C2_cmpgtp: 1840 case Hexagon::C2_cmpgtu: 1841 case Hexagon::C2_cmpgtup: 1842 case Hexagon::A4_cmpbeq: 1843 case Hexagon::A4_cmpbgt: 1844 case Hexagon::A4_cmpbgtu: 1845 case Hexagon::A4_cmpheq: 1846 case Hexagon::A4_cmphgt: 1847 case Hexagon::A4_cmphgtu: 1848 case Hexagon::C4_cmpneq: 1849 case Hexagon::C4_cmplte: 1850 case Hexagon::C4_cmplteu: 1851 SrcReg2 = MI.getOperand(2).getReg(); 1852 return true; 1853 1854 case Hexagon::C2_cmpeqi: 1855 case Hexagon::C2_cmpgtui: 1856 case Hexagon::C2_cmpgti: 1857 case Hexagon::C4_cmpneqi: 1858 case Hexagon::C4_cmplteui: 1859 case Hexagon::C4_cmpltei: 1860 case Hexagon::A4_cmpbeqi: 1861 case Hexagon::A4_cmpbgti: 1862 case Hexagon::A4_cmpbgtui: 1863 case Hexagon::A4_cmpheqi: 1864 case Hexagon::A4_cmphgti: 1865 case Hexagon::A4_cmphgtui: { 1866 SrcReg2 = 0; 1867 const MachineOperand &Op2 = MI.getOperand(2); 1868 if (!Op2.isImm()) 1869 return false; 1870 Value = MI.getOperand(2).getImm(); 1871 return true; 1872 } 1873 } 1874 1875 return false; 1876 } 1877 1878 unsigned HexagonInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, 1879 const MachineInstr &MI, 1880 unsigned *PredCost) const { 1881 return getInstrTimingClassLatency(ItinData, MI); 1882 } 1883 1884 DFAPacketizer *HexagonInstrInfo::CreateTargetScheduleState( 1885 const TargetSubtargetInfo &STI) const { 1886 const InstrItineraryData *II = STI.getInstrItineraryData(); 1887 return static_cast<const HexagonSubtarget&>(STI).createDFAPacketizer(II); 1888 } 1889 1890 // Inspired by this pair: 1891 // %r13 = L2_loadri_io %r29, 136; mem:LD4[FixedStack0] 1892 // S2_storeri_io %r29, 132, killed %r1; flags: mem:ST4[FixedStack1] 1893 // Currently AA considers the addresses in these instructions to be aliasing. 1894 bool HexagonInstrInfo::areMemAccessesTriviallyDisjoint( 1895 const MachineInstr &MIa, const MachineInstr &MIb) const { 1896 if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects() || 1897 MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef()) 1898 return false; 1899 1900 // Instructions that are pure loads, not loads and stores like memops are not 1901 // dependent. 1902 if (MIa.mayLoad() && !isMemOp(MIa) && MIb.mayLoad() && !isMemOp(MIb)) 1903 return true; 1904 1905 // Get the base register in MIa. 1906 unsigned BasePosA, OffsetPosA; 1907 if (!getBaseAndOffsetPosition(MIa, BasePosA, OffsetPosA)) 1908 return false; 1909 const MachineOperand &BaseA = MIa.getOperand(BasePosA); 1910 Register BaseRegA = BaseA.getReg(); 1911 unsigned BaseSubA = BaseA.getSubReg(); 1912 1913 // Get the base register in MIb. 1914 unsigned BasePosB, OffsetPosB; 1915 if (!getBaseAndOffsetPosition(MIb, BasePosB, OffsetPosB)) 1916 return false; 1917 const MachineOperand &BaseB = MIb.getOperand(BasePosB); 1918 Register BaseRegB = BaseB.getReg(); 1919 unsigned BaseSubB = BaseB.getSubReg(); 1920 1921 if (BaseRegA != BaseRegB || BaseSubA != BaseSubB) 1922 return false; 1923 1924 // Get the access sizes. 1925 unsigned SizeA = getMemAccessSize(MIa); 1926 unsigned SizeB = getMemAccessSize(MIb); 1927 1928 // Get the offsets. Handle immediates only for now. 1929 const MachineOperand &OffA = MIa.getOperand(OffsetPosA); 1930 const MachineOperand &OffB = MIb.getOperand(OffsetPosB); 1931 if (!MIa.getOperand(OffsetPosA).isImm() || 1932 !MIb.getOperand(OffsetPosB).isImm()) 1933 return false; 1934 int OffsetA = isPostIncrement(MIa) ? 0 : OffA.getImm(); 1935 int OffsetB = isPostIncrement(MIb) ? 0 : OffB.getImm(); 1936 1937 // This is a mem access with the same base register and known offsets from it. 1938 // Reason about it. 1939 if (OffsetA > OffsetB) { 1940 uint64_t OffDiff = (uint64_t)((int64_t)OffsetA - (int64_t)OffsetB); 1941 return SizeB <= OffDiff; 1942 } 1943 if (OffsetA < OffsetB) { 1944 uint64_t OffDiff = (uint64_t)((int64_t)OffsetB - (int64_t)OffsetA); 1945 return SizeA <= OffDiff; 1946 } 1947 1948 return false; 1949 } 1950 1951 /// If the instruction is an increment of a constant value, return the amount. 1952 bool HexagonInstrInfo::getIncrementValue(const MachineInstr &MI, 1953 int &Value) const { 1954 if (isPostIncrement(MI)) { 1955 unsigned BasePos = 0, OffsetPos = 0; 1956 if (!getBaseAndOffsetPosition(MI, BasePos, OffsetPos)) 1957 return false; 1958 const MachineOperand &OffsetOp = MI.getOperand(OffsetPos); 1959 if (OffsetOp.isImm()) { 1960 Value = OffsetOp.getImm(); 1961 return true; 1962 } 1963 } else if (MI.getOpcode() == Hexagon::A2_addi) { 1964 const MachineOperand &AddOp = MI.getOperand(2); 1965 if (AddOp.isImm()) { 1966 Value = AddOp.getImm(); 1967 return true; 1968 } 1969 } 1970 1971 return false; 1972 } 1973 1974 std::pair<unsigned, unsigned> 1975 HexagonInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const { 1976 return std::make_pair(TF & ~HexagonII::MO_Bitmasks, 1977 TF & HexagonII::MO_Bitmasks); 1978 } 1979 1980 ArrayRef<std::pair<unsigned, const char*>> 1981 HexagonInstrInfo::getSerializableDirectMachineOperandTargetFlags() const { 1982 using namespace HexagonII; 1983 1984 static const std::pair<unsigned, const char*> Flags[] = { 1985 {MO_PCREL, "hexagon-pcrel"}, 1986 {MO_GOT, "hexagon-got"}, 1987 {MO_LO16, "hexagon-lo16"}, 1988 {MO_HI16, "hexagon-hi16"}, 1989 {MO_GPREL, "hexagon-gprel"}, 1990 {MO_GDGOT, "hexagon-gdgot"}, 1991 {MO_GDPLT, "hexagon-gdplt"}, 1992 {MO_IE, "hexagon-ie"}, 1993 {MO_IEGOT, "hexagon-iegot"}, 1994 {MO_TPREL, "hexagon-tprel"} 1995 }; 1996 return makeArrayRef(Flags); 1997 } 1998 1999 ArrayRef<std::pair<unsigned, const char*>> 2000 HexagonInstrInfo::getSerializableBitmaskMachineOperandTargetFlags() const { 2001 using namespace HexagonII; 2002 2003 static const std::pair<unsigned, const char*> Flags[] = { 2004 {HMOTF_ConstExtended, "hexagon-ext"} 2005 }; 2006 return makeArrayRef(Flags); 2007 } 2008 2009 unsigned HexagonInstrInfo::createVR(MachineFunction *MF, MVT VT) const { 2010 MachineRegisterInfo &MRI = MF->getRegInfo(); 2011 const TargetRegisterClass *TRC; 2012 if (VT == MVT::i1) { 2013 TRC = &Hexagon::PredRegsRegClass; 2014 } else if (VT == MVT::i32 || VT == MVT::f32) { 2015 TRC = &Hexagon::IntRegsRegClass; 2016 } else if (VT == MVT::i64 || VT == MVT::f64) { 2017 TRC = &Hexagon::DoubleRegsRegClass; 2018 } else { 2019 llvm_unreachable("Cannot handle this register class"); 2020 } 2021 2022 Register NewReg = MRI.createVirtualRegister(TRC); 2023 return NewReg; 2024 } 2025 2026 bool HexagonInstrInfo::isAbsoluteSet(const MachineInstr &MI) const { 2027 return (getAddrMode(MI) == HexagonII::AbsoluteSet); 2028 } 2029 2030 bool HexagonInstrInfo::isAccumulator(const MachineInstr &MI) const { 2031 const uint64_t F = MI.getDesc().TSFlags; 2032 return((F >> HexagonII::AccumulatorPos) & HexagonII::AccumulatorMask); 2033 } 2034 2035 bool HexagonInstrInfo::isBaseImmOffset(const MachineInstr &MI) const { 2036 return getAddrMode(MI) == HexagonII::BaseImmOffset; 2037 } 2038 2039 bool HexagonInstrInfo::isComplex(const MachineInstr &MI) const { 2040 return !isTC1(MI) && !isTC2Early(MI) && !MI.getDesc().mayLoad() && 2041 !MI.getDesc().mayStore() && 2042 MI.getDesc().getOpcode() != Hexagon::S2_allocframe && 2043 MI.getDesc().getOpcode() != Hexagon::L2_deallocframe && 2044 !isMemOp(MI) && !MI.isBranch() && !MI.isReturn() && !MI.isCall(); 2045 } 2046 2047 // Return true if the instruction is a compund branch instruction. 2048 bool HexagonInstrInfo::isCompoundBranchInstr(const MachineInstr &MI) const { 2049 return getType(MI) == HexagonII::TypeCJ && MI.isBranch(); 2050 } 2051 2052 // TODO: In order to have isExtendable for fpimm/f32Ext, we need to handle 2053 // isFPImm and later getFPImm as well. 2054 bool HexagonInstrInfo::isConstExtended(const MachineInstr &MI) const { 2055 const uint64_t F = MI.getDesc().TSFlags; 2056 unsigned isExtended = (F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask; 2057 if (isExtended) // Instruction must be extended. 2058 return true; 2059 2060 unsigned isExtendable = 2061 (F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask; 2062 if (!isExtendable) 2063 return false; 2064 2065 if (MI.isCall()) 2066 return false; 2067 2068 short ExtOpNum = getCExtOpNum(MI); 2069 const MachineOperand &MO = MI.getOperand(ExtOpNum); 2070 // Use MO operand flags to determine if MO 2071 // has the HMOTF_ConstExtended flag set. 2072 if (MO.getTargetFlags() & HexagonII::HMOTF_ConstExtended) 2073 return true; 2074 // If this is a Machine BB address we are talking about, and it is 2075 // not marked as extended, say so. 2076 if (MO.isMBB()) 2077 return false; 2078 2079 // We could be using an instruction with an extendable immediate and shoehorn 2080 // a global address into it. If it is a global address it will be constant 2081 // extended. We do this for COMBINE. 2082 if (MO.isGlobal() || MO.isSymbol() || MO.isBlockAddress() || 2083 MO.isJTI() || MO.isCPI() || MO.isFPImm()) 2084 return true; 2085 2086 // If the extendable operand is not 'Immediate' type, the instruction should 2087 // have 'isExtended' flag set. 2088 assert(MO.isImm() && "Extendable operand must be Immediate type"); 2089 2090 int MinValue = getMinValue(MI); 2091 int MaxValue = getMaxValue(MI); 2092 int ImmValue = MO.getImm(); 2093 2094 return (ImmValue < MinValue || ImmValue > MaxValue); 2095 } 2096 2097 bool HexagonInstrInfo::isDeallocRet(const MachineInstr &MI) const { 2098 switch (MI.getOpcode()) { 2099 case Hexagon::L4_return: 2100 case Hexagon::L4_return_t: 2101 case Hexagon::L4_return_f: 2102 case Hexagon::L4_return_tnew_pnt: 2103 case Hexagon::L4_return_fnew_pnt: 2104 case Hexagon::L4_return_tnew_pt: 2105 case Hexagon::L4_return_fnew_pt: 2106 return true; 2107 } 2108 return false; 2109 } 2110 2111 // Return true when ConsMI uses a register defined by ProdMI. 2112 bool HexagonInstrInfo::isDependent(const MachineInstr &ProdMI, 2113 const MachineInstr &ConsMI) const { 2114 if (!ProdMI.getDesc().getNumDefs()) 2115 return false; 2116 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 2117 2118 SmallVector<unsigned, 4> DefsA; 2119 SmallVector<unsigned, 4> DefsB; 2120 SmallVector<unsigned, 8> UsesA; 2121 SmallVector<unsigned, 8> UsesB; 2122 2123 parseOperands(ProdMI, DefsA, UsesA); 2124 parseOperands(ConsMI, DefsB, UsesB); 2125 2126 for (auto &RegA : DefsA) 2127 for (auto &RegB : UsesB) { 2128 // True data dependency. 2129 if (RegA == RegB) 2130 return true; 2131 2132 if (Register::isPhysicalRegister(RegA)) 2133 for (MCSubRegIterator SubRegs(RegA, &HRI); SubRegs.isValid(); ++SubRegs) 2134 if (RegB == *SubRegs) 2135 return true; 2136 2137 if (Register::isPhysicalRegister(RegB)) 2138 for (MCSubRegIterator SubRegs(RegB, &HRI); SubRegs.isValid(); ++SubRegs) 2139 if (RegA == *SubRegs) 2140 return true; 2141 } 2142 2143 return false; 2144 } 2145 2146 // Returns true if the instruction is alread a .cur. 2147 bool HexagonInstrInfo::isDotCurInst(const MachineInstr &MI) const { 2148 switch (MI.getOpcode()) { 2149 case Hexagon::V6_vL32b_cur_pi: 2150 case Hexagon::V6_vL32b_cur_ai: 2151 return true; 2152 } 2153 return false; 2154 } 2155 2156 // Returns true, if any one of the operands is a dot new 2157 // insn, whether it is predicated dot new or register dot new. 2158 bool HexagonInstrInfo::isDotNewInst(const MachineInstr &MI) const { 2159 if (isNewValueInst(MI) || (isPredicated(MI) && isPredicatedNew(MI))) 2160 return true; 2161 2162 return false; 2163 } 2164 2165 /// Symmetrical. See if these two instructions are fit for duplex pair. 2166 bool HexagonInstrInfo::isDuplexPair(const MachineInstr &MIa, 2167 const MachineInstr &MIb) const { 2168 HexagonII::SubInstructionGroup MIaG = getDuplexCandidateGroup(MIa); 2169 HexagonII::SubInstructionGroup MIbG = getDuplexCandidateGroup(MIb); 2170 return (isDuplexPairMatch(MIaG, MIbG) || isDuplexPairMatch(MIbG, MIaG)); 2171 } 2172 2173 bool HexagonInstrInfo::isEarlySourceInstr(const MachineInstr &MI) const { 2174 if (MI.mayLoadOrStore() || MI.isCompare()) 2175 return true; 2176 2177 // Multiply 2178 unsigned SchedClass = MI.getDesc().getSchedClass(); 2179 return is_TC4x(SchedClass) || is_TC3x(SchedClass); 2180 } 2181 2182 bool HexagonInstrInfo::isEndLoopN(unsigned Opcode) const { 2183 return (Opcode == Hexagon::ENDLOOP0 || 2184 Opcode == Hexagon::ENDLOOP1); 2185 } 2186 2187 bool HexagonInstrInfo::isExpr(unsigned OpType) const { 2188 switch(OpType) { 2189 case MachineOperand::MO_MachineBasicBlock: 2190 case MachineOperand::MO_GlobalAddress: 2191 case MachineOperand::MO_ExternalSymbol: 2192 case MachineOperand::MO_JumpTableIndex: 2193 case MachineOperand::MO_ConstantPoolIndex: 2194 case MachineOperand::MO_BlockAddress: 2195 return true; 2196 default: 2197 return false; 2198 } 2199 } 2200 2201 bool HexagonInstrInfo::isExtendable(const MachineInstr &MI) const { 2202 const MCInstrDesc &MID = MI.getDesc(); 2203 const uint64_t F = MID.TSFlags; 2204 if ((F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask) 2205 return true; 2206 2207 // TODO: This is largely obsolete now. Will need to be removed 2208 // in consecutive patches. 2209 switch (MI.getOpcode()) { 2210 // PS_fi and PS_fia remain special cases. 2211 case Hexagon::PS_fi: 2212 case Hexagon::PS_fia: 2213 return true; 2214 default: 2215 return false; 2216 } 2217 return false; 2218 } 2219 2220 // This returns true in two cases: 2221 // - The OP code itself indicates that this is an extended instruction. 2222 // - One of MOs has been marked with HMOTF_ConstExtended flag. 2223 bool HexagonInstrInfo::isExtended(const MachineInstr &MI) const { 2224 // First check if this is permanently extended op code. 2225 const uint64_t F = MI.getDesc().TSFlags; 2226 if ((F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask) 2227 return true; 2228 // Use MO operand flags to determine if one of MI's operands 2229 // has HMOTF_ConstExtended flag set. 2230 for (const MachineOperand &MO : MI.operands()) 2231 if (MO.getTargetFlags() & HexagonII::HMOTF_ConstExtended) 2232 return true; 2233 return false; 2234 } 2235 2236 bool HexagonInstrInfo::isFloat(const MachineInstr &MI) const { 2237 unsigned Opcode = MI.getOpcode(); 2238 const uint64_t F = get(Opcode).TSFlags; 2239 return (F >> HexagonII::FPPos) & HexagonII::FPMask; 2240 } 2241 2242 // No V60 HVX VMEM with A_INDIRECT. 2243 bool HexagonInstrInfo::isHVXMemWithAIndirect(const MachineInstr &I, 2244 const MachineInstr &J) const { 2245 if (!isHVXVec(I)) 2246 return false; 2247 if (!I.mayLoad() && !I.mayStore()) 2248 return false; 2249 return J.isIndirectBranch() || isIndirectCall(J) || isIndirectL4Return(J); 2250 } 2251 2252 bool HexagonInstrInfo::isIndirectCall(const MachineInstr &MI) const { 2253 switch (MI.getOpcode()) { 2254 case Hexagon::J2_callr: 2255 case Hexagon::J2_callrf: 2256 case Hexagon::J2_callrt: 2257 case Hexagon::PS_call_nr: 2258 return true; 2259 } 2260 return false; 2261 } 2262 2263 bool HexagonInstrInfo::isIndirectL4Return(const MachineInstr &MI) const { 2264 switch (MI.getOpcode()) { 2265 case Hexagon::L4_return: 2266 case Hexagon::L4_return_t: 2267 case Hexagon::L4_return_f: 2268 case Hexagon::L4_return_fnew_pnt: 2269 case Hexagon::L4_return_fnew_pt: 2270 case Hexagon::L4_return_tnew_pnt: 2271 case Hexagon::L4_return_tnew_pt: 2272 return true; 2273 } 2274 return false; 2275 } 2276 2277 bool HexagonInstrInfo::isJumpR(const MachineInstr &MI) const { 2278 switch (MI.getOpcode()) { 2279 case Hexagon::J2_jumpr: 2280 case Hexagon::J2_jumprt: 2281 case Hexagon::J2_jumprf: 2282 case Hexagon::J2_jumprtnewpt: 2283 case Hexagon::J2_jumprfnewpt: 2284 case Hexagon::J2_jumprtnew: 2285 case Hexagon::J2_jumprfnew: 2286 return true; 2287 } 2288 return false; 2289 } 2290 2291 // Return true if a given MI can accommodate given offset. 2292 // Use abs estimate as oppose to the exact number. 2293 // TODO: This will need to be changed to use MC level 2294 // definition of instruction extendable field size. 2295 bool HexagonInstrInfo::isJumpWithinBranchRange(const MachineInstr &MI, 2296 unsigned offset) const { 2297 // This selection of jump instructions matches to that what 2298 // analyzeBranch can parse, plus NVJ. 2299 if (isNewValueJump(MI)) // r9:2 2300 return isInt<11>(offset); 2301 2302 switch (MI.getOpcode()) { 2303 // Still missing Jump to address condition on register value. 2304 default: 2305 return false; 2306 case Hexagon::J2_jump: // bits<24> dst; // r22:2 2307 case Hexagon::J2_call: 2308 case Hexagon::PS_call_nr: 2309 return isInt<24>(offset); 2310 case Hexagon::J2_jumpt: //bits<17> dst; // r15:2 2311 case Hexagon::J2_jumpf: 2312 case Hexagon::J2_jumptnew: 2313 case Hexagon::J2_jumptnewpt: 2314 case Hexagon::J2_jumpfnew: 2315 case Hexagon::J2_jumpfnewpt: 2316 case Hexagon::J2_callt: 2317 case Hexagon::J2_callf: 2318 return isInt<17>(offset); 2319 case Hexagon::J2_loop0i: 2320 case Hexagon::J2_loop0iext: 2321 case Hexagon::J2_loop0r: 2322 case Hexagon::J2_loop0rext: 2323 case Hexagon::J2_loop1i: 2324 case Hexagon::J2_loop1iext: 2325 case Hexagon::J2_loop1r: 2326 case Hexagon::J2_loop1rext: 2327 return isInt<9>(offset); 2328 // TODO: Add all the compound branches here. Can we do this in Relation model? 2329 case Hexagon::J4_cmpeqi_tp0_jump_nt: 2330 case Hexagon::J4_cmpeqi_tp1_jump_nt: 2331 case Hexagon::J4_cmpeqn1_tp0_jump_nt: 2332 case Hexagon::J4_cmpeqn1_tp1_jump_nt: 2333 return isInt<11>(offset); 2334 } 2335 } 2336 2337 bool HexagonInstrInfo::isLateInstrFeedsEarlyInstr(const MachineInstr &LRMI, 2338 const MachineInstr &ESMI) const { 2339 bool isLate = isLateResultInstr(LRMI); 2340 bool isEarly = isEarlySourceInstr(ESMI); 2341 2342 LLVM_DEBUG(dbgs() << "V60" << (isLate ? "-LR " : " -- ")); 2343 LLVM_DEBUG(LRMI.dump()); 2344 LLVM_DEBUG(dbgs() << "V60" << (isEarly ? "-ES " : " -- ")); 2345 LLVM_DEBUG(ESMI.dump()); 2346 2347 if (isLate && isEarly) { 2348 LLVM_DEBUG(dbgs() << "++Is Late Result feeding Early Source\n"); 2349 return true; 2350 } 2351 2352 return false; 2353 } 2354 2355 bool HexagonInstrInfo::isLateResultInstr(const MachineInstr &MI) const { 2356 switch (MI.getOpcode()) { 2357 case TargetOpcode::EXTRACT_SUBREG: 2358 case TargetOpcode::INSERT_SUBREG: 2359 case TargetOpcode::SUBREG_TO_REG: 2360 case TargetOpcode::REG_SEQUENCE: 2361 case TargetOpcode::IMPLICIT_DEF: 2362 case TargetOpcode::COPY: 2363 case TargetOpcode::INLINEASM: 2364 case TargetOpcode::PHI: 2365 return false; 2366 default: 2367 break; 2368 } 2369 2370 unsigned SchedClass = MI.getDesc().getSchedClass(); 2371 return !is_TC1(SchedClass); 2372 } 2373 2374 bool HexagonInstrInfo::isLateSourceInstr(const MachineInstr &MI) const { 2375 // Instructions with iclass A_CVI_VX and attribute A_CVI_LATE uses a multiply 2376 // resource, but all operands can be received late like an ALU instruction. 2377 return getType(MI) == HexagonII::TypeCVI_VX_LATE; 2378 } 2379 2380 bool HexagonInstrInfo::isLoopN(const MachineInstr &MI) const { 2381 unsigned Opcode = MI.getOpcode(); 2382 return Opcode == Hexagon::J2_loop0i || 2383 Opcode == Hexagon::J2_loop0r || 2384 Opcode == Hexagon::J2_loop0iext || 2385 Opcode == Hexagon::J2_loop0rext || 2386 Opcode == Hexagon::J2_loop1i || 2387 Opcode == Hexagon::J2_loop1r || 2388 Opcode == Hexagon::J2_loop1iext || 2389 Opcode == Hexagon::J2_loop1rext; 2390 } 2391 2392 bool HexagonInstrInfo::isMemOp(const MachineInstr &MI) const { 2393 switch (MI.getOpcode()) { 2394 default: return false; 2395 case Hexagon::L4_iadd_memopw_io: 2396 case Hexagon::L4_isub_memopw_io: 2397 case Hexagon::L4_add_memopw_io: 2398 case Hexagon::L4_sub_memopw_io: 2399 case Hexagon::L4_and_memopw_io: 2400 case Hexagon::L4_or_memopw_io: 2401 case Hexagon::L4_iadd_memoph_io: 2402 case Hexagon::L4_isub_memoph_io: 2403 case Hexagon::L4_add_memoph_io: 2404 case Hexagon::L4_sub_memoph_io: 2405 case Hexagon::L4_and_memoph_io: 2406 case Hexagon::L4_or_memoph_io: 2407 case Hexagon::L4_iadd_memopb_io: 2408 case Hexagon::L4_isub_memopb_io: 2409 case Hexagon::L4_add_memopb_io: 2410 case Hexagon::L4_sub_memopb_io: 2411 case Hexagon::L4_and_memopb_io: 2412 case Hexagon::L4_or_memopb_io: 2413 case Hexagon::L4_ior_memopb_io: 2414 case Hexagon::L4_ior_memoph_io: 2415 case Hexagon::L4_ior_memopw_io: 2416 case Hexagon::L4_iand_memopb_io: 2417 case Hexagon::L4_iand_memoph_io: 2418 case Hexagon::L4_iand_memopw_io: 2419 return true; 2420 } 2421 return false; 2422 } 2423 2424 bool HexagonInstrInfo::isNewValue(const MachineInstr &MI) const { 2425 const uint64_t F = MI.getDesc().TSFlags; 2426 return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask; 2427 } 2428 2429 bool HexagonInstrInfo::isNewValue(unsigned Opcode) const { 2430 const uint64_t F = get(Opcode).TSFlags; 2431 return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask; 2432 } 2433 2434 bool HexagonInstrInfo::isNewValueInst(const MachineInstr &MI) const { 2435 return isNewValueJump(MI) || isNewValueStore(MI); 2436 } 2437 2438 bool HexagonInstrInfo::isNewValueJump(const MachineInstr &MI) const { 2439 return isNewValue(MI) && MI.isBranch(); 2440 } 2441 2442 bool HexagonInstrInfo::isNewValueJump(unsigned Opcode) const { 2443 return isNewValue(Opcode) && get(Opcode).isBranch() && isPredicated(Opcode); 2444 } 2445 2446 bool HexagonInstrInfo::isNewValueStore(const MachineInstr &MI) const { 2447 const uint64_t F = MI.getDesc().TSFlags; 2448 return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask; 2449 } 2450 2451 bool HexagonInstrInfo::isNewValueStore(unsigned Opcode) const { 2452 const uint64_t F = get(Opcode).TSFlags; 2453 return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask; 2454 } 2455 2456 // Returns true if a particular operand is extendable for an instruction. 2457 bool HexagonInstrInfo::isOperandExtended(const MachineInstr &MI, 2458 unsigned OperandNum) const { 2459 const uint64_t F = MI.getDesc().TSFlags; 2460 return ((F >> HexagonII::ExtendableOpPos) & HexagonII::ExtendableOpMask) 2461 == OperandNum; 2462 } 2463 2464 bool HexagonInstrInfo::isPredicatedNew(const MachineInstr &MI) const { 2465 const uint64_t F = MI.getDesc().TSFlags; 2466 assert(isPredicated(MI)); 2467 return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask; 2468 } 2469 2470 bool HexagonInstrInfo::isPredicatedNew(unsigned Opcode) const { 2471 const uint64_t F = get(Opcode).TSFlags; 2472 assert(isPredicated(Opcode)); 2473 return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask; 2474 } 2475 2476 bool HexagonInstrInfo::isPredicatedTrue(const MachineInstr &MI) const { 2477 const uint64_t F = MI.getDesc().TSFlags; 2478 return !((F >> HexagonII::PredicatedFalsePos) & 2479 HexagonII::PredicatedFalseMask); 2480 } 2481 2482 bool HexagonInstrInfo::isPredicatedTrue(unsigned Opcode) const { 2483 const uint64_t F = get(Opcode).TSFlags; 2484 // Make sure that the instruction is predicated. 2485 assert((F>> HexagonII::PredicatedPos) & HexagonII::PredicatedMask); 2486 return !((F >> HexagonII::PredicatedFalsePos) & 2487 HexagonII::PredicatedFalseMask); 2488 } 2489 2490 bool HexagonInstrInfo::isPredicated(unsigned Opcode) const { 2491 const uint64_t F = get(Opcode).TSFlags; 2492 return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask; 2493 } 2494 2495 bool HexagonInstrInfo::isPredicateLate(unsigned Opcode) const { 2496 const uint64_t F = get(Opcode).TSFlags; 2497 return (F >> HexagonII::PredicateLatePos) & HexagonII::PredicateLateMask; 2498 } 2499 2500 bool HexagonInstrInfo::isPredictedTaken(unsigned Opcode) const { 2501 const uint64_t F = get(Opcode).TSFlags; 2502 assert(get(Opcode).isBranch() && 2503 (isPredicatedNew(Opcode) || isNewValue(Opcode))); 2504 return (F >> HexagonII::TakenPos) & HexagonII::TakenMask; 2505 } 2506 2507 bool HexagonInstrInfo::isSaveCalleeSavedRegsCall(const MachineInstr &MI) const { 2508 return MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4 || 2509 MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT || 2510 MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_PIC || 2511 MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT_PIC; 2512 } 2513 2514 bool HexagonInstrInfo::isSignExtendingLoad(const MachineInstr &MI) const { 2515 switch (MI.getOpcode()) { 2516 // Byte 2517 case Hexagon::L2_loadrb_io: 2518 case Hexagon::L4_loadrb_ur: 2519 case Hexagon::L4_loadrb_ap: 2520 case Hexagon::L2_loadrb_pr: 2521 case Hexagon::L2_loadrb_pbr: 2522 case Hexagon::L2_loadrb_pi: 2523 case Hexagon::L2_loadrb_pci: 2524 case Hexagon::L2_loadrb_pcr: 2525 case Hexagon::L2_loadbsw2_io: 2526 case Hexagon::L4_loadbsw2_ur: 2527 case Hexagon::L4_loadbsw2_ap: 2528 case Hexagon::L2_loadbsw2_pr: 2529 case Hexagon::L2_loadbsw2_pbr: 2530 case Hexagon::L2_loadbsw2_pi: 2531 case Hexagon::L2_loadbsw2_pci: 2532 case Hexagon::L2_loadbsw2_pcr: 2533 case Hexagon::L2_loadbsw4_io: 2534 case Hexagon::L4_loadbsw4_ur: 2535 case Hexagon::L4_loadbsw4_ap: 2536 case Hexagon::L2_loadbsw4_pr: 2537 case Hexagon::L2_loadbsw4_pbr: 2538 case Hexagon::L2_loadbsw4_pi: 2539 case Hexagon::L2_loadbsw4_pci: 2540 case Hexagon::L2_loadbsw4_pcr: 2541 case Hexagon::L4_loadrb_rr: 2542 case Hexagon::L2_ploadrbt_io: 2543 case Hexagon::L2_ploadrbt_pi: 2544 case Hexagon::L2_ploadrbf_io: 2545 case Hexagon::L2_ploadrbf_pi: 2546 case Hexagon::L2_ploadrbtnew_io: 2547 case Hexagon::L2_ploadrbfnew_io: 2548 case Hexagon::L4_ploadrbt_rr: 2549 case Hexagon::L4_ploadrbf_rr: 2550 case Hexagon::L4_ploadrbtnew_rr: 2551 case Hexagon::L4_ploadrbfnew_rr: 2552 case Hexagon::L2_ploadrbtnew_pi: 2553 case Hexagon::L2_ploadrbfnew_pi: 2554 case Hexagon::L4_ploadrbt_abs: 2555 case Hexagon::L4_ploadrbf_abs: 2556 case Hexagon::L4_ploadrbtnew_abs: 2557 case Hexagon::L4_ploadrbfnew_abs: 2558 case Hexagon::L2_loadrbgp: 2559 // Half 2560 case Hexagon::L2_loadrh_io: 2561 case Hexagon::L4_loadrh_ur: 2562 case Hexagon::L4_loadrh_ap: 2563 case Hexagon::L2_loadrh_pr: 2564 case Hexagon::L2_loadrh_pbr: 2565 case Hexagon::L2_loadrh_pi: 2566 case Hexagon::L2_loadrh_pci: 2567 case Hexagon::L2_loadrh_pcr: 2568 case Hexagon::L4_loadrh_rr: 2569 case Hexagon::L2_ploadrht_io: 2570 case Hexagon::L2_ploadrht_pi: 2571 case Hexagon::L2_ploadrhf_io: 2572 case Hexagon::L2_ploadrhf_pi: 2573 case Hexagon::L2_ploadrhtnew_io: 2574 case Hexagon::L2_ploadrhfnew_io: 2575 case Hexagon::L4_ploadrht_rr: 2576 case Hexagon::L4_ploadrhf_rr: 2577 case Hexagon::L4_ploadrhtnew_rr: 2578 case Hexagon::L4_ploadrhfnew_rr: 2579 case Hexagon::L2_ploadrhtnew_pi: 2580 case Hexagon::L2_ploadrhfnew_pi: 2581 case Hexagon::L4_ploadrht_abs: 2582 case Hexagon::L4_ploadrhf_abs: 2583 case Hexagon::L4_ploadrhtnew_abs: 2584 case Hexagon::L4_ploadrhfnew_abs: 2585 case Hexagon::L2_loadrhgp: 2586 return true; 2587 default: 2588 return false; 2589 } 2590 } 2591 2592 bool HexagonInstrInfo::isSolo(const MachineInstr &MI) const { 2593 const uint64_t F = MI.getDesc().TSFlags; 2594 return (F >> HexagonII::SoloPos) & HexagonII::SoloMask; 2595 } 2596 2597 bool HexagonInstrInfo::isSpillPredRegOp(const MachineInstr &MI) const { 2598 switch (MI.getOpcode()) { 2599 case Hexagon::STriw_pred: 2600 case Hexagon::LDriw_pred: 2601 return true; 2602 default: 2603 return false; 2604 } 2605 } 2606 2607 bool HexagonInstrInfo::isTailCall(const MachineInstr &MI) const { 2608 if (!MI.isBranch()) 2609 return false; 2610 2611 for (auto &Op : MI.operands()) 2612 if (Op.isGlobal() || Op.isSymbol()) 2613 return true; 2614 return false; 2615 } 2616 2617 // Returns true when SU has a timing class TC1. 2618 bool HexagonInstrInfo::isTC1(const MachineInstr &MI) const { 2619 unsigned SchedClass = MI.getDesc().getSchedClass(); 2620 return is_TC1(SchedClass); 2621 } 2622 2623 bool HexagonInstrInfo::isTC2(const MachineInstr &MI) const { 2624 unsigned SchedClass = MI.getDesc().getSchedClass(); 2625 return is_TC2(SchedClass); 2626 } 2627 2628 bool HexagonInstrInfo::isTC2Early(const MachineInstr &MI) const { 2629 unsigned SchedClass = MI.getDesc().getSchedClass(); 2630 return is_TC2early(SchedClass); 2631 } 2632 2633 bool HexagonInstrInfo::isTC4x(const MachineInstr &MI) const { 2634 unsigned SchedClass = MI.getDesc().getSchedClass(); 2635 return is_TC4x(SchedClass); 2636 } 2637 2638 // Schedule this ASAP. 2639 bool HexagonInstrInfo::isToBeScheduledASAP(const MachineInstr &MI1, 2640 const MachineInstr &MI2) const { 2641 if (mayBeCurLoad(MI1)) { 2642 // if (result of SU is used in Next) return true; 2643 Register DstReg = MI1.getOperand(0).getReg(); 2644 int N = MI2.getNumOperands(); 2645 for (int I = 0; I < N; I++) 2646 if (MI2.getOperand(I).isReg() && DstReg == MI2.getOperand(I).getReg()) 2647 return true; 2648 } 2649 if (mayBeNewStore(MI2)) 2650 if (MI2.getOpcode() == Hexagon::V6_vS32b_pi) 2651 if (MI1.getOperand(0).isReg() && MI2.getOperand(3).isReg() && 2652 MI1.getOperand(0).getReg() == MI2.getOperand(3).getReg()) 2653 return true; 2654 return false; 2655 } 2656 2657 bool HexagonInstrInfo::isHVXVec(const MachineInstr &MI) const { 2658 const uint64_t V = getType(MI); 2659 return HexagonII::TypeCVI_FIRST <= V && V <= HexagonII::TypeCVI_LAST; 2660 } 2661 2662 // Check if the Offset is a valid auto-inc imm by Load/Store Type. 2663 bool HexagonInstrInfo::isValidAutoIncImm(const EVT VT, int Offset) const { 2664 int Size = VT.getSizeInBits() / 8; 2665 if (Offset % Size != 0) 2666 return false; 2667 int Count = Offset / Size; 2668 2669 switch (VT.getSimpleVT().SimpleTy) { 2670 // For scalars the auto-inc is s4 2671 case MVT::i8: 2672 case MVT::i16: 2673 case MVT::i32: 2674 case MVT::i64: 2675 case MVT::f32: 2676 case MVT::f64: 2677 case MVT::v2i16: 2678 case MVT::v2i32: 2679 case MVT::v4i8: 2680 case MVT::v4i16: 2681 case MVT::v8i8: 2682 return isInt<4>(Count); 2683 // For HVX vectors the auto-inc is s3 2684 case MVT::v64i8: 2685 case MVT::v32i16: 2686 case MVT::v16i32: 2687 case MVT::v8i64: 2688 case MVT::v128i8: 2689 case MVT::v64i16: 2690 case MVT::v32i32: 2691 case MVT::v16i64: 2692 return isInt<3>(Count); 2693 default: 2694 break; 2695 } 2696 2697 llvm_unreachable("Not an valid type!"); 2698 } 2699 2700 bool HexagonInstrInfo::isValidOffset(unsigned Opcode, int Offset, 2701 const TargetRegisterInfo *TRI, bool Extend) const { 2702 // This function is to check whether the "Offset" is in the correct range of 2703 // the given "Opcode". If "Offset" is not in the correct range, "A2_addi" is 2704 // inserted to calculate the final address. Due to this reason, the function 2705 // assumes that the "Offset" has correct alignment. 2706 // We used to assert if the offset was not properly aligned, however, 2707 // there are cases where a misaligned pointer recast can cause this 2708 // problem, and we need to allow for it. The front end warns of such 2709 // misaligns with respect to load size. 2710 switch (Opcode) { 2711 case Hexagon::PS_vstorerq_ai: 2712 case Hexagon::PS_vstorerv_ai: 2713 case Hexagon::PS_vstorerw_ai: 2714 case Hexagon::PS_vstorerw_nt_ai: 2715 case Hexagon::PS_vloadrq_ai: 2716 case Hexagon::PS_vloadrv_ai: 2717 case Hexagon::PS_vloadrw_ai: 2718 case Hexagon::PS_vloadrw_nt_ai: 2719 case Hexagon::V6_vL32b_ai: 2720 case Hexagon::V6_vS32b_ai: 2721 case Hexagon::V6_vL32b_nt_ai: 2722 case Hexagon::V6_vS32b_nt_ai: 2723 case Hexagon::V6_vL32Ub_ai: 2724 case Hexagon::V6_vS32Ub_ai: { 2725 unsigned VectorSize = TRI->getSpillSize(Hexagon::HvxVRRegClass); 2726 assert(isPowerOf2_32(VectorSize)); 2727 if (Offset & (VectorSize-1)) 2728 return false; 2729 return isInt<4>(Offset >> Log2_32(VectorSize)); 2730 } 2731 2732 case Hexagon::J2_loop0i: 2733 case Hexagon::J2_loop1i: 2734 return isUInt<10>(Offset); 2735 2736 case Hexagon::S4_storeirb_io: 2737 case Hexagon::S4_storeirbt_io: 2738 case Hexagon::S4_storeirbf_io: 2739 return isUInt<6>(Offset); 2740 2741 case Hexagon::S4_storeirh_io: 2742 case Hexagon::S4_storeirht_io: 2743 case Hexagon::S4_storeirhf_io: 2744 return isShiftedUInt<6,1>(Offset); 2745 2746 case Hexagon::S4_storeiri_io: 2747 case Hexagon::S4_storeirit_io: 2748 case Hexagon::S4_storeirif_io: 2749 return isShiftedUInt<6,2>(Offset); 2750 } 2751 2752 if (Extend) 2753 return true; 2754 2755 switch (Opcode) { 2756 case Hexagon::L2_loadri_io: 2757 case Hexagon::S2_storeri_io: 2758 return (Offset >= Hexagon_MEMW_OFFSET_MIN) && 2759 (Offset <= Hexagon_MEMW_OFFSET_MAX); 2760 2761 case Hexagon::L2_loadrd_io: 2762 case Hexagon::S2_storerd_io: 2763 return (Offset >= Hexagon_MEMD_OFFSET_MIN) && 2764 (Offset <= Hexagon_MEMD_OFFSET_MAX); 2765 2766 case Hexagon::L2_loadrh_io: 2767 case Hexagon::L2_loadruh_io: 2768 case Hexagon::S2_storerh_io: 2769 case Hexagon::S2_storerf_io: 2770 return (Offset >= Hexagon_MEMH_OFFSET_MIN) && 2771 (Offset <= Hexagon_MEMH_OFFSET_MAX); 2772 2773 case Hexagon::L2_loadrb_io: 2774 case Hexagon::L2_loadrub_io: 2775 case Hexagon::S2_storerb_io: 2776 return (Offset >= Hexagon_MEMB_OFFSET_MIN) && 2777 (Offset <= Hexagon_MEMB_OFFSET_MAX); 2778 2779 case Hexagon::A2_addi: 2780 return (Offset >= Hexagon_ADDI_OFFSET_MIN) && 2781 (Offset <= Hexagon_ADDI_OFFSET_MAX); 2782 2783 case Hexagon::L4_iadd_memopw_io: 2784 case Hexagon::L4_isub_memopw_io: 2785 case Hexagon::L4_add_memopw_io: 2786 case Hexagon::L4_sub_memopw_io: 2787 case Hexagon::L4_and_memopw_io: 2788 case Hexagon::L4_or_memopw_io: 2789 return (0 <= Offset && Offset <= 255); 2790 2791 case Hexagon::L4_iadd_memoph_io: 2792 case Hexagon::L4_isub_memoph_io: 2793 case Hexagon::L4_add_memoph_io: 2794 case Hexagon::L4_sub_memoph_io: 2795 case Hexagon::L4_and_memoph_io: 2796 case Hexagon::L4_or_memoph_io: 2797 return (0 <= Offset && Offset <= 127); 2798 2799 case Hexagon::L4_iadd_memopb_io: 2800 case Hexagon::L4_isub_memopb_io: 2801 case Hexagon::L4_add_memopb_io: 2802 case Hexagon::L4_sub_memopb_io: 2803 case Hexagon::L4_and_memopb_io: 2804 case Hexagon::L4_or_memopb_io: 2805 return (0 <= Offset && Offset <= 63); 2806 2807 // LDriw_xxx and STriw_xxx are pseudo operations, so it has to take offset of 2808 // any size. Later pass knows how to handle it. 2809 case Hexagon::STriw_pred: 2810 case Hexagon::LDriw_pred: 2811 case Hexagon::STriw_ctr: 2812 case Hexagon::LDriw_ctr: 2813 return true; 2814 2815 case Hexagon::PS_fi: 2816 case Hexagon::PS_fia: 2817 case Hexagon::INLINEASM: 2818 return true; 2819 2820 case Hexagon::L2_ploadrbt_io: 2821 case Hexagon::L2_ploadrbf_io: 2822 case Hexagon::L2_ploadrubt_io: 2823 case Hexagon::L2_ploadrubf_io: 2824 case Hexagon::S2_pstorerbt_io: 2825 case Hexagon::S2_pstorerbf_io: 2826 return isUInt<6>(Offset); 2827 2828 case Hexagon::L2_ploadrht_io: 2829 case Hexagon::L2_ploadrhf_io: 2830 case Hexagon::L2_ploadruht_io: 2831 case Hexagon::L2_ploadruhf_io: 2832 case Hexagon::S2_pstorerht_io: 2833 case Hexagon::S2_pstorerhf_io: 2834 return isShiftedUInt<6,1>(Offset); 2835 2836 case Hexagon::L2_ploadrit_io: 2837 case Hexagon::L2_ploadrif_io: 2838 case Hexagon::S2_pstorerit_io: 2839 case Hexagon::S2_pstorerif_io: 2840 return isShiftedUInt<6,2>(Offset); 2841 2842 case Hexagon::L2_ploadrdt_io: 2843 case Hexagon::L2_ploadrdf_io: 2844 case Hexagon::S2_pstorerdt_io: 2845 case Hexagon::S2_pstorerdf_io: 2846 return isShiftedUInt<6,3>(Offset); 2847 } // switch 2848 2849 llvm_unreachable("No offset range is defined for this opcode. " 2850 "Please define it in the above switch statement!"); 2851 } 2852 2853 bool HexagonInstrInfo::isVecAcc(const MachineInstr &MI) const { 2854 return isHVXVec(MI) && isAccumulator(MI); 2855 } 2856 2857 bool HexagonInstrInfo::isVecALU(const MachineInstr &MI) const { 2858 const uint64_t F = get(MI.getOpcode()).TSFlags; 2859 const uint64_t V = ((F >> HexagonII::TypePos) & HexagonII::TypeMask); 2860 return 2861 V == HexagonII::TypeCVI_VA || 2862 V == HexagonII::TypeCVI_VA_DV; 2863 } 2864 2865 bool HexagonInstrInfo::isVecUsableNextPacket(const MachineInstr &ProdMI, 2866 const MachineInstr &ConsMI) const { 2867 if (EnableACCForwarding && isVecAcc(ProdMI) && isVecAcc(ConsMI)) 2868 return true; 2869 2870 if (EnableALUForwarding && (isVecALU(ConsMI) || isLateSourceInstr(ConsMI))) 2871 return true; 2872 2873 if (mayBeNewStore(ConsMI)) 2874 return true; 2875 2876 return false; 2877 } 2878 2879 bool HexagonInstrInfo::isZeroExtendingLoad(const MachineInstr &MI) const { 2880 switch (MI.getOpcode()) { 2881 // Byte 2882 case Hexagon::L2_loadrub_io: 2883 case Hexagon::L4_loadrub_ur: 2884 case Hexagon::L4_loadrub_ap: 2885 case Hexagon::L2_loadrub_pr: 2886 case Hexagon::L2_loadrub_pbr: 2887 case Hexagon::L2_loadrub_pi: 2888 case Hexagon::L2_loadrub_pci: 2889 case Hexagon::L2_loadrub_pcr: 2890 case Hexagon::L2_loadbzw2_io: 2891 case Hexagon::L4_loadbzw2_ur: 2892 case Hexagon::L4_loadbzw2_ap: 2893 case Hexagon::L2_loadbzw2_pr: 2894 case Hexagon::L2_loadbzw2_pbr: 2895 case Hexagon::L2_loadbzw2_pi: 2896 case Hexagon::L2_loadbzw2_pci: 2897 case Hexagon::L2_loadbzw2_pcr: 2898 case Hexagon::L2_loadbzw4_io: 2899 case Hexagon::L4_loadbzw4_ur: 2900 case Hexagon::L4_loadbzw4_ap: 2901 case Hexagon::L2_loadbzw4_pr: 2902 case Hexagon::L2_loadbzw4_pbr: 2903 case Hexagon::L2_loadbzw4_pi: 2904 case Hexagon::L2_loadbzw4_pci: 2905 case Hexagon::L2_loadbzw4_pcr: 2906 case Hexagon::L4_loadrub_rr: 2907 case Hexagon::L2_ploadrubt_io: 2908 case Hexagon::L2_ploadrubt_pi: 2909 case Hexagon::L2_ploadrubf_io: 2910 case Hexagon::L2_ploadrubf_pi: 2911 case Hexagon::L2_ploadrubtnew_io: 2912 case Hexagon::L2_ploadrubfnew_io: 2913 case Hexagon::L4_ploadrubt_rr: 2914 case Hexagon::L4_ploadrubf_rr: 2915 case Hexagon::L4_ploadrubtnew_rr: 2916 case Hexagon::L4_ploadrubfnew_rr: 2917 case Hexagon::L2_ploadrubtnew_pi: 2918 case Hexagon::L2_ploadrubfnew_pi: 2919 case Hexagon::L4_ploadrubt_abs: 2920 case Hexagon::L4_ploadrubf_abs: 2921 case Hexagon::L4_ploadrubtnew_abs: 2922 case Hexagon::L4_ploadrubfnew_abs: 2923 case Hexagon::L2_loadrubgp: 2924 // Half 2925 case Hexagon::L2_loadruh_io: 2926 case Hexagon::L4_loadruh_ur: 2927 case Hexagon::L4_loadruh_ap: 2928 case Hexagon::L2_loadruh_pr: 2929 case Hexagon::L2_loadruh_pbr: 2930 case Hexagon::L2_loadruh_pi: 2931 case Hexagon::L2_loadruh_pci: 2932 case Hexagon::L2_loadruh_pcr: 2933 case Hexagon::L4_loadruh_rr: 2934 case Hexagon::L2_ploadruht_io: 2935 case Hexagon::L2_ploadruht_pi: 2936 case Hexagon::L2_ploadruhf_io: 2937 case Hexagon::L2_ploadruhf_pi: 2938 case Hexagon::L2_ploadruhtnew_io: 2939 case Hexagon::L2_ploadruhfnew_io: 2940 case Hexagon::L4_ploadruht_rr: 2941 case Hexagon::L4_ploadruhf_rr: 2942 case Hexagon::L4_ploadruhtnew_rr: 2943 case Hexagon::L4_ploadruhfnew_rr: 2944 case Hexagon::L2_ploadruhtnew_pi: 2945 case Hexagon::L2_ploadruhfnew_pi: 2946 case Hexagon::L4_ploadruht_abs: 2947 case Hexagon::L4_ploadruhf_abs: 2948 case Hexagon::L4_ploadruhtnew_abs: 2949 case Hexagon::L4_ploadruhfnew_abs: 2950 case Hexagon::L2_loadruhgp: 2951 return true; 2952 default: 2953 return false; 2954 } 2955 } 2956 2957 // Add latency to instruction. 2958 bool HexagonInstrInfo::addLatencyToSchedule(const MachineInstr &MI1, 2959 const MachineInstr &MI2) const { 2960 if (isHVXVec(MI1) && isHVXVec(MI2)) 2961 if (!isVecUsableNextPacket(MI1, MI2)) 2962 return true; 2963 return false; 2964 } 2965 2966 /// Get the base register and byte offset of a load/store instr. 2967 bool HexagonInstrInfo::getMemOperandsWithOffset( 2968 const MachineInstr &LdSt, SmallVectorImpl<const MachineOperand *> &BaseOps, 2969 int64_t &Offset, bool &OffsetIsScalable, const TargetRegisterInfo *TRI) const { 2970 unsigned AccessSize = 0; 2971 OffsetIsScalable = false; 2972 const MachineOperand *BaseOp = getBaseAndOffset(LdSt, Offset, AccessSize); 2973 if (!BaseOp || !BaseOp->isReg()) 2974 return false; 2975 BaseOps.push_back(BaseOp); 2976 return true; 2977 } 2978 2979 /// Can these instructions execute at the same time in a bundle. 2980 bool HexagonInstrInfo::canExecuteInBundle(const MachineInstr &First, 2981 const MachineInstr &Second) const { 2982 if (Second.mayStore() && First.getOpcode() == Hexagon::S2_allocframe) { 2983 const MachineOperand &Op = Second.getOperand(0); 2984 if (Op.isReg() && Op.isUse() && Op.getReg() == Hexagon::R29) 2985 return true; 2986 } 2987 if (DisableNVSchedule) 2988 return false; 2989 if (mayBeNewStore(Second)) { 2990 // Make sure the definition of the first instruction is the value being 2991 // stored. 2992 const MachineOperand &Stored = 2993 Second.getOperand(Second.getNumOperands() - 1); 2994 if (!Stored.isReg()) 2995 return false; 2996 for (unsigned i = 0, e = First.getNumOperands(); i < e; ++i) { 2997 const MachineOperand &Op = First.getOperand(i); 2998 if (Op.isReg() && Op.isDef() && Op.getReg() == Stored.getReg()) 2999 return true; 3000 } 3001 } 3002 return false; 3003 } 3004 3005 bool HexagonInstrInfo::doesNotReturn(const MachineInstr &CallMI) const { 3006 unsigned Opc = CallMI.getOpcode(); 3007 return Opc == Hexagon::PS_call_nr || Opc == Hexagon::PS_callr_nr; 3008 } 3009 3010 bool HexagonInstrInfo::hasEHLabel(const MachineBasicBlock *B) const { 3011 for (auto &I : *B) 3012 if (I.isEHLabel()) 3013 return true; 3014 return false; 3015 } 3016 3017 // Returns true if an instruction can be converted into a non-extended 3018 // equivalent instruction. 3019 bool HexagonInstrInfo::hasNonExtEquivalent(const MachineInstr &MI) const { 3020 short NonExtOpcode; 3021 // Check if the instruction has a register form that uses register in place 3022 // of the extended operand, if so return that as the non-extended form. 3023 if (Hexagon::getRegForm(MI.getOpcode()) >= 0) 3024 return true; 3025 3026 if (MI.getDesc().mayLoad() || MI.getDesc().mayStore()) { 3027 // Check addressing mode and retrieve non-ext equivalent instruction. 3028 3029 switch (getAddrMode(MI)) { 3030 case HexagonII::Absolute: 3031 // Load/store with absolute addressing mode can be converted into 3032 // base+offset mode. 3033 NonExtOpcode = Hexagon::changeAddrMode_abs_io(MI.getOpcode()); 3034 break; 3035 case HexagonII::BaseImmOffset: 3036 // Load/store with base+offset addressing mode can be converted into 3037 // base+register offset addressing mode. However left shift operand should 3038 // be set to 0. 3039 NonExtOpcode = Hexagon::changeAddrMode_io_rr(MI.getOpcode()); 3040 break; 3041 case HexagonII::BaseLongOffset: 3042 NonExtOpcode = Hexagon::changeAddrMode_ur_rr(MI.getOpcode()); 3043 break; 3044 default: 3045 return false; 3046 } 3047 if (NonExtOpcode < 0) 3048 return false; 3049 return true; 3050 } 3051 return false; 3052 } 3053 3054 bool HexagonInstrInfo::hasPseudoInstrPair(const MachineInstr &MI) const { 3055 return Hexagon::getRealHWInstr(MI.getOpcode(), 3056 Hexagon::InstrType_Pseudo) >= 0; 3057 } 3058 3059 bool HexagonInstrInfo::hasUncondBranch(const MachineBasicBlock *B) 3060 const { 3061 MachineBasicBlock::const_iterator I = B->getFirstTerminator(), E = B->end(); 3062 while (I != E) { 3063 if (I->isBarrier()) 3064 return true; 3065 ++I; 3066 } 3067 return false; 3068 } 3069 3070 // Returns true, if a LD insn can be promoted to a cur load. 3071 bool HexagonInstrInfo::mayBeCurLoad(const MachineInstr &MI) const { 3072 const uint64_t F = MI.getDesc().TSFlags; 3073 return ((F >> HexagonII::mayCVLoadPos) & HexagonII::mayCVLoadMask) && 3074 Subtarget.hasV60Ops(); 3075 } 3076 3077 // Returns true, if a ST insn can be promoted to a new-value store. 3078 bool HexagonInstrInfo::mayBeNewStore(const MachineInstr &MI) const { 3079 if (MI.mayStore() && !Subtarget.useNewValueStores()) 3080 return false; 3081 3082 const uint64_t F = MI.getDesc().TSFlags; 3083 return (F >> HexagonII::mayNVStorePos) & HexagonII::mayNVStoreMask; 3084 } 3085 3086 bool HexagonInstrInfo::producesStall(const MachineInstr &ProdMI, 3087 const MachineInstr &ConsMI) const { 3088 // There is no stall when ProdMI is not a V60 vector. 3089 if (!isHVXVec(ProdMI)) 3090 return false; 3091 3092 // There is no stall when ProdMI and ConsMI are not dependent. 3093 if (!isDependent(ProdMI, ConsMI)) 3094 return false; 3095 3096 // When Forward Scheduling is enabled, there is no stall if ProdMI and ConsMI 3097 // are scheduled in consecutive packets. 3098 if (isVecUsableNextPacket(ProdMI, ConsMI)) 3099 return false; 3100 3101 return true; 3102 } 3103 3104 bool HexagonInstrInfo::producesStall(const MachineInstr &MI, 3105 MachineBasicBlock::const_instr_iterator BII) const { 3106 // There is no stall when I is not a V60 vector. 3107 if (!isHVXVec(MI)) 3108 return false; 3109 3110 MachineBasicBlock::const_instr_iterator MII = BII; 3111 MachineBasicBlock::const_instr_iterator MIE = MII->getParent()->instr_end(); 3112 3113 if (!MII->isBundle()) 3114 return producesStall(*MII, MI); 3115 3116 for (++MII; MII != MIE && MII->isInsideBundle(); ++MII) { 3117 const MachineInstr &J = *MII; 3118 if (producesStall(J, MI)) 3119 return true; 3120 } 3121 return false; 3122 } 3123 3124 bool HexagonInstrInfo::predCanBeUsedAsDotNew(const MachineInstr &MI, 3125 unsigned PredReg) const { 3126 for (const MachineOperand &MO : MI.operands()) { 3127 // Predicate register must be explicitly defined. 3128 if (MO.isRegMask() && MO.clobbersPhysReg(PredReg)) 3129 return false; 3130 if (MO.isReg() && MO.isDef() && MO.isImplicit() && (MO.getReg() == PredReg)) 3131 return false; 3132 } 3133 3134 // Instruction that produce late predicate cannot be used as sources of 3135 // dot-new. 3136 switch (MI.getOpcode()) { 3137 case Hexagon::A4_addp_c: 3138 case Hexagon::A4_subp_c: 3139 case Hexagon::A4_tlbmatch: 3140 case Hexagon::A5_ACS: 3141 case Hexagon::F2_sfinvsqrta: 3142 case Hexagon::F2_sfrecipa: 3143 case Hexagon::J2_endloop0: 3144 case Hexagon::J2_endloop01: 3145 case Hexagon::J2_ploop1si: 3146 case Hexagon::J2_ploop1sr: 3147 case Hexagon::J2_ploop2si: 3148 case Hexagon::J2_ploop2sr: 3149 case Hexagon::J2_ploop3si: 3150 case Hexagon::J2_ploop3sr: 3151 case Hexagon::S2_cabacdecbin: 3152 case Hexagon::S2_storew_locked: 3153 case Hexagon::S4_stored_locked: 3154 return false; 3155 } 3156 return true; 3157 } 3158 3159 bool HexagonInstrInfo::PredOpcodeHasJMP_c(unsigned Opcode) const { 3160 return Opcode == Hexagon::J2_jumpt || 3161 Opcode == Hexagon::J2_jumptpt || 3162 Opcode == Hexagon::J2_jumpf || 3163 Opcode == Hexagon::J2_jumpfpt || 3164 Opcode == Hexagon::J2_jumptnew || 3165 Opcode == Hexagon::J2_jumpfnew || 3166 Opcode == Hexagon::J2_jumptnewpt || 3167 Opcode == Hexagon::J2_jumpfnewpt; 3168 } 3169 3170 bool HexagonInstrInfo::predOpcodeHasNot(ArrayRef<MachineOperand> Cond) const { 3171 if (Cond.empty() || !isPredicated(Cond[0].getImm())) 3172 return false; 3173 return !isPredicatedTrue(Cond[0].getImm()); 3174 } 3175 3176 unsigned HexagonInstrInfo::getAddrMode(const MachineInstr &MI) const { 3177 const uint64_t F = MI.getDesc().TSFlags; 3178 return (F >> HexagonII::AddrModePos) & HexagonII::AddrModeMask; 3179 } 3180 3181 // Returns the base register in a memory access (load/store). The offset is 3182 // returned in Offset and the access size is returned in AccessSize. 3183 // If the base operand has a subregister or the offset field does not contain 3184 // an immediate value, return nullptr. 3185 MachineOperand *HexagonInstrInfo::getBaseAndOffset(const MachineInstr &MI, 3186 int64_t &Offset, 3187 unsigned &AccessSize) const { 3188 // Return if it is not a base+offset type instruction or a MemOp. 3189 if (getAddrMode(MI) != HexagonII::BaseImmOffset && 3190 getAddrMode(MI) != HexagonII::BaseLongOffset && 3191 !isMemOp(MI) && !isPostIncrement(MI)) 3192 return nullptr; 3193 3194 AccessSize = getMemAccessSize(MI); 3195 3196 unsigned BasePos = 0, OffsetPos = 0; 3197 if (!getBaseAndOffsetPosition(MI, BasePos, OffsetPos)) 3198 return nullptr; 3199 3200 // Post increment updates its EA after the mem access, 3201 // so we need to treat its offset as zero. 3202 if (isPostIncrement(MI)) { 3203 Offset = 0; 3204 } else { 3205 const MachineOperand &OffsetOp = MI.getOperand(OffsetPos); 3206 if (!OffsetOp.isImm()) 3207 return nullptr; 3208 Offset = OffsetOp.getImm(); 3209 } 3210 3211 const MachineOperand &BaseOp = MI.getOperand(BasePos); 3212 if (BaseOp.getSubReg() != 0) 3213 return nullptr; 3214 return &const_cast<MachineOperand&>(BaseOp); 3215 } 3216 3217 /// Return the position of the base and offset operands for this instruction. 3218 bool HexagonInstrInfo::getBaseAndOffsetPosition(const MachineInstr &MI, 3219 unsigned &BasePos, unsigned &OffsetPos) const { 3220 if (!isAddrModeWithOffset(MI) && !isPostIncrement(MI)) 3221 return false; 3222 3223 // Deal with memops first. 3224 if (isMemOp(MI)) { 3225 BasePos = 0; 3226 OffsetPos = 1; 3227 } else if (MI.mayStore()) { 3228 BasePos = 0; 3229 OffsetPos = 1; 3230 } else if (MI.mayLoad()) { 3231 BasePos = 1; 3232 OffsetPos = 2; 3233 } else 3234 return false; 3235 3236 if (isPredicated(MI)) { 3237 BasePos++; 3238 OffsetPos++; 3239 } 3240 if (isPostIncrement(MI)) { 3241 BasePos++; 3242 OffsetPos++; 3243 } 3244 3245 if (!MI.getOperand(BasePos).isReg() || !MI.getOperand(OffsetPos).isImm()) 3246 return false; 3247 3248 return true; 3249 } 3250 3251 // Inserts branching instructions in reverse order of their occurrence. 3252 // e.g. jump_t t1 (i1) 3253 // jump t2 (i2) 3254 // Jumpers = {i2, i1} 3255 SmallVector<MachineInstr*, 2> HexagonInstrInfo::getBranchingInstrs( 3256 MachineBasicBlock& MBB) const { 3257 SmallVector<MachineInstr*, 2> Jumpers; 3258 // If the block has no terminators, it just falls into the block after it. 3259 MachineBasicBlock::instr_iterator I = MBB.instr_end(); 3260 if (I == MBB.instr_begin()) 3261 return Jumpers; 3262 3263 // A basic block may looks like this: 3264 // 3265 // [ insn 3266 // EH_LABEL 3267 // insn 3268 // insn 3269 // insn 3270 // EH_LABEL 3271 // insn ] 3272 // 3273 // It has two succs but does not have a terminator 3274 // Don't know how to handle it. 3275 do { 3276 --I; 3277 if (I->isEHLabel()) 3278 return Jumpers; 3279 } while (I != MBB.instr_begin()); 3280 3281 I = MBB.instr_end(); 3282 --I; 3283 3284 while (I->isDebugInstr()) { 3285 if (I == MBB.instr_begin()) 3286 return Jumpers; 3287 --I; 3288 } 3289 if (!isUnpredicatedTerminator(*I)) 3290 return Jumpers; 3291 3292 // Get the last instruction in the block. 3293 MachineInstr *LastInst = &*I; 3294 Jumpers.push_back(LastInst); 3295 MachineInstr *SecondLastInst = nullptr; 3296 // Find one more terminator if present. 3297 do { 3298 if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(*I)) { 3299 if (!SecondLastInst) { 3300 SecondLastInst = &*I; 3301 Jumpers.push_back(SecondLastInst); 3302 } else // This is a third branch. 3303 return Jumpers; 3304 } 3305 if (I == MBB.instr_begin()) 3306 break; 3307 --I; 3308 } while (true); 3309 return Jumpers; 3310 } 3311 3312 // Returns Operand Index for the constant extended instruction. 3313 unsigned HexagonInstrInfo::getCExtOpNum(const MachineInstr &MI) const { 3314 const uint64_t F = MI.getDesc().TSFlags; 3315 return (F >> HexagonII::ExtendableOpPos) & HexagonII::ExtendableOpMask; 3316 } 3317 3318 // See if instruction could potentially be a duplex candidate. 3319 // If so, return its group. Zero otherwise. 3320 HexagonII::CompoundGroup HexagonInstrInfo::getCompoundCandidateGroup( 3321 const MachineInstr &MI) const { 3322 unsigned DstReg, SrcReg, Src1Reg, Src2Reg; 3323 3324 switch (MI.getOpcode()) { 3325 default: 3326 return HexagonII::HCG_None; 3327 // 3328 // Compound pairs. 3329 // "p0=cmp.eq(Rs16,Rt16); if (p0.new) jump:nt #r9:2" 3330 // "Rd16=#U6 ; jump #r9:2" 3331 // "Rd16=Rs16 ; jump #r9:2" 3332 // 3333 case Hexagon::C2_cmpeq: 3334 case Hexagon::C2_cmpgt: 3335 case Hexagon::C2_cmpgtu: 3336 DstReg = MI.getOperand(0).getReg(); 3337 Src1Reg = MI.getOperand(1).getReg(); 3338 Src2Reg = MI.getOperand(2).getReg(); 3339 if (Hexagon::PredRegsRegClass.contains(DstReg) && 3340 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) && 3341 isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg)) 3342 return HexagonII::HCG_A; 3343 break; 3344 case Hexagon::C2_cmpeqi: 3345 case Hexagon::C2_cmpgti: 3346 case Hexagon::C2_cmpgtui: 3347 // P0 = cmp.eq(Rs,#u2) 3348 DstReg = MI.getOperand(0).getReg(); 3349 SrcReg = MI.getOperand(1).getReg(); 3350 if (Hexagon::PredRegsRegClass.contains(DstReg) && 3351 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) && 3352 isIntRegForSubInst(SrcReg) && MI.getOperand(2).isImm() && 3353 ((isUInt<5>(MI.getOperand(2).getImm())) || 3354 (MI.getOperand(2).getImm() == -1))) 3355 return HexagonII::HCG_A; 3356 break; 3357 case Hexagon::A2_tfr: 3358 // Rd = Rs 3359 DstReg = MI.getOperand(0).getReg(); 3360 SrcReg = MI.getOperand(1).getReg(); 3361 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg)) 3362 return HexagonII::HCG_A; 3363 break; 3364 case Hexagon::A2_tfrsi: 3365 // Rd = #u6 3366 // Do not test for #u6 size since the const is getting extended 3367 // regardless and compound could be formed. 3368 DstReg = MI.getOperand(0).getReg(); 3369 if (isIntRegForSubInst(DstReg)) 3370 return HexagonII::HCG_A; 3371 break; 3372 case Hexagon::S2_tstbit_i: 3373 DstReg = MI.getOperand(0).getReg(); 3374 Src1Reg = MI.getOperand(1).getReg(); 3375 if (Hexagon::PredRegsRegClass.contains(DstReg) && 3376 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) && 3377 MI.getOperand(2).isImm() && 3378 isIntRegForSubInst(Src1Reg) && (MI.getOperand(2).getImm() == 0)) 3379 return HexagonII::HCG_A; 3380 break; 3381 // The fact that .new form is used pretty much guarantees 3382 // that predicate register will match. Nevertheless, 3383 // there could be some false positives without additional 3384 // checking. 3385 case Hexagon::J2_jumptnew: 3386 case Hexagon::J2_jumpfnew: 3387 case Hexagon::J2_jumptnewpt: 3388 case Hexagon::J2_jumpfnewpt: 3389 Src1Reg = MI.getOperand(0).getReg(); 3390 if (Hexagon::PredRegsRegClass.contains(Src1Reg) && 3391 (Hexagon::P0 == Src1Reg || Hexagon::P1 == Src1Reg)) 3392 return HexagonII::HCG_B; 3393 break; 3394 // Transfer and jump: 3395 // Rd=#U6 ; jump #r9:2 3396 // Rd=Rs ; jump #r9:2 3397 // Do not test for jump range here. 3398 case Hexagon::J2_jump: 3399 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4: 3400 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4_PIC: 3401 return HexagonII::HCG_C; 3402 } 3403 3404 return HexagonII::HCG_None; 3405 } 3406 3407 // Returns -1 when there is no opcode found. 3408 unsigned HexagonInstrInfo::getCompoundOpcode(const MachineInstr &GA, 3409 const MachineInstr &GB) const { 3410 assert(getCompoundCandidateGroup(GA) == HexagonII::HCG_A); 3411 assert(getCompoundCandidateGroup(GB) == HexagonII::HCG_B); 3412 if ((GA.getOpcode() != Hexagon::C2_cmpeqi) || 3413 (GB.getOpcode() != Hexagon::J2_jumptnew)) 3414 return -1u; 3415 Register DestReg = GA.getOperand(0).getReg(); 3416 if (!GB.readsRegister(DestReg)) 3417 return -1u; 3418 if (DestReg != Hexagon::P0 && DestReg != Hexagon::P1) 3419 return -1u; 3420 // The value compared against must be either u5 or -1. 3421 const MachineOperand &CmpOp = GA.getOperand(2); 3422 if (!CmpOp.isImm()) 3423 return -1u; 3424 int V = CmpOp.getImm(); 3425 if (V == -1) 3426 return DestReg == Hexagon::P0 ? Hexagon::J4_cmpeqn1_tp0_jump_nt 3427 : Hexagon::J4_cmpeqn1_tp1_jump_nt; 3428 if (!isUInt<5>(V)) 3429 return -1u; 3430 return DestReg == Hexagon::P0 ? Hexagon::J4_cmpeqi_tp0_jump_nt 3431 : Hexagon::J4_cmpeqi_tp1_jump_nt; 3432 } 3433 3434 // Returns -1 if there is no opcode found. 3435 int HexagonInstrInfo::getDuplexOpcode(const MachineInstr &MI, 3436 bool ForBigCore) const { 3437 // Static table to switch the opcodes across Tiny Core and Big Core. 3438 // dup_ opcodes are Big core opcodes. 3439 // NOTE: There are special instructions that need to handled later. 3440 // L4_return* instructions, they will only occupy SLOT0 (on big core too). 3441 // PS_jmpret - This pseudo translates to J2_jumpr which occupies only SLOT2. 3442 // The compiler need to base the root instruction to L6_return_map_to_raw 3443 // which can go any slot. 3444 static const std::map<unsigned, unsigned> DupMap = { 3445 {Hexagon::A2_add, Hexagon::dup_A2_add}, 3446 {Hexagon::A2_addi, Hexagon::dup_A2_addi}, 3447 {Hexagon::A2_andir, Hexagon::dup_A2_andir}, 3448 {Hexagon::A2_combineii, Hexagon::dup_A2_combineii}, 3449 {Hexagon::A2_sxtb, Hexagon::dup_A2_sxtb}, 3450 {Hexagon::A2_sxth, Hexagon::dup_A2_sxth}, 3451 {Hexagon::A2_tfr, Hexagon::dup_A2_tfr}, 3452 {Hexagon::A2_tfrsi, Hexagon::dup_A2_tfrsi}, 3453 {Hexagon::A2_zxtb, Hexagon::dup_A2_zxtb}, 3454 {Hexagon::A2_zxth, Hexagon::dup_A2_zxth}, 3455 {Hexagon::A4_combineii, Hexagon::dup_A4_combineii}, 3456 {Hexagon::A4_combineir, Hexagon::dup_A4_combineir}, 3457 {Hexagon::A4_combineri, Hexagon::dup_A4_combineri}, 3458 {Hexagon::C2_cmoveif, Hexagon::dup_C2_cmoveif}, 3459 {Hexagon::C2_cmoveit, Hexagon::dup_C2_cmoveit}, 3460 {Hexagon::C2_cmovenewif, Hexagon::dup_C2_cmovenewif}, 3461 {Hexagon::C2_cmovenewit, Hexagon::dup_C2_cmovenewit}, 3462 {Hexagon::C2_cmpeqi, Hexagon::dup_C2_cmpeqi}, 3463 {Hexagon::L2_deallocframe, Hexagon::dup_L2_deallocframe}, 3464 {Hexagon::L2_loadrb_io, Hexagon::dup_L2_loadrb_io}, 3465 {Hexagon::L2_loadrd_io, Hexagon::dup_L2_loadrd_io}, 3466 {Hexagon::L2_loadrh_io, Hexagon::dup_L2_loadrh_io}, 3467 {Hexagon::L2_loadri_io, Hexagon::dup_L2_loadri_io}, 3468 {Hexagon::L2_loadrub_io, Hexagon::dup_L2_loadrub_io}, 3469 {Hexagon::L2_loadruh_io, Hexagon::dup_L2_loadruh_io}, 3470 {Hexagon::S2_allocframe, Hexagon::dup_S2_allocframe}, 3471 {Hexagon::S2_storerb_io, Hexagon::dup_S2_storerb_io}, 3472 {Hexagon::S2_storerd_io, Hexagon::dup_S2_storerd_io}, 3473 {Hexagon::S2_storerh_io, Hexagon::dup_S2_storerh_io}, 3474 {Hexagon::S2_storeri_io, Hexagon::dup_S2_storeri_io}, 3475 {Hexagon::S4_storeirb_io, Hexagon::dup_S4_storeirb_io}, 3476 {Hexagon::S4_storeiri_io, Hexagon::dup_S4_storeiri_io}, 3477 }; 3478 unsigned OpNum = MI.getOpcode(); 3479 // Conversion to Big core. 3480 if (ForBigCore) { 3481 auto Iter = DupMap.find(OpNum); 3482 if (Iter != DupMap.end()) 3483 return Iter->second; 3484 } else { // Conversion to Tiny core. 3485 for (auto Iter = DupMap.begin(), End = DupMap.end(); Iter != End; ++Iter) 3486 if (Iter->second == OpNum) 3487 return Iter->first; 3488 } 3489 return -1; 3490 } 3491 3492 int HexagonInstrInfo::getCondOpcode(int Opc, bool invertPredicate) const { 3493 enum Hexagon::PredSense inPredSense; 3494 inPredSense = invertPredicate ? Hexagon::PredSense_false : 3495 Hexagon::PredSense_true; 3496 int CondOpcode = Hexagon::getPredOpcode(Opc, inPredSense); 3497 if (CondOpcode >= 0) // Valid Conditional opcode/instruction 3498 return CondOpcode; 3499 3500 llvm_unreachable("Unexpected predicable instruction"); 3501 } 3502 3503 // Return the cur value instruction for a given store. 3504 int HexagonInstrInfo::getDotCurOp(const MachineInstr &MI) const { 3505 switch (MI.getOpcode()) { 3506 default: llvm_unreachable("Unknown .cur type"); 3507 case Hexagon::V6_vL32b_pi: 3508 return Hexagon::V6_vL32b_cur_pi; 3509 case Hexagon::V6_vL32b_ai: 3510 return Hexagon::V6_vL32b_cur_ai; 3511 case Hexagon::V6_vL32b_nt_pi: 3512 return Hexagon::V6_vL32b_nt_cur_pi; 3513 case Hexagon::V6_vL32b_nt_ai: 3514 return Hexagon::V6_vL32b_nt_cur_ai; 3515 } 3516 return 0; 3517 } 3518 3519 // Return the regular version of the .cur instruction. 3520 int HexagonInstrInfo::getNonDotCurOp(const MachineInstr &MI) const { 3521 switch (MI.getOpcode()) { 3522 default: llvm_unreachable("Unknown .cur type"); 3523 case Hexagon::V6_vL32b_cur_pi: 3524 return Hexagon::V6_vL32b_pi; 3525 case Hexagon::V6_vL32b_cur_ai: 3526 return Hexagon::V6_vL32b_ai; 3527 case Hexagon::V6_vL32b_nt_cur_pi: 3528 return Hexagon::V6_vL32b_nt_pi; 3529 case Hexagon::V6_vL32b_nt_cur_ai: 3530 return Hexagon::V6_vL32b_nt_ai; 3531 } 3532 return 0; 3533 } 3534 3535 // The diagram below shows the steps involved in the conversion of a predicated 3536 // store instruction to its .new predicated new-value form. 3537 // 3538 // Note: It doesn't include conditional new-value stores as they can't be 3539 // converted to .new predicate. 3540 // 3541 // p.new NV store [ if(p0.new)memw(R0+#0)=R2.new ] 3542 // ^ ^ 3543 // / \ (not OK. it will cause new-value store to be 3544 // / X conditional on p0.new while R2 producer is 3545 // / \ on p0) 3546 // / \. 3547 // p.new store p.old NV store 3548 // [if(p0.new)memw(R0+#0)=R2] [if(p0)memw(R0+#0)=R2.new] 3549 // ^ ^ 3550 // \ / 3551 // \ / 3552 // \ / 3553 // p.old store 3554 // [if (p0)memw(R0+#0)=R2] 3555 // 3556 // The following set of instructions further explains the scenario where 3557 // conditional new-value store becomes invalid when promoted to .new predicate 3558 // form. 3559 // 3560 // { 1) if (p0) r0 = add(r1, r2) 3561 // 2) p0 = cmp.eq(r3, #0) } 3562 // 3563 // 3) if (p0) memb(r1+#0) = r0 --> this instruction can't be grouped with 3564 // the first two instructions because in instr 1, r0 is conditional on old value 3565 // of p0 but its use in instr 3 is conditional on p0 modified by instr 2 which 3566 // is not valid for new-value stores. 3567 // Predicated new value stores (i.e. if (p0) memw(..)=r0.new) are excluded 3568 // from the "Conditional Store" list. Because a predicated new value store 3569 // would NOT be promoted to a double dot new store. See diagram below: 3570 // This function returns yes for those stores that are predicated but not 3571 // yet promoted to predicate dot new instructions. 3572 // 3573 // +---------------------+ 3574 // /-----| if (p0) memw(..)=r0 |---------\~ 3575 // || +---------------------+ || 3576 // promote || /\ /\ || promote 3577 // || /||\ /||\ || 3578 // \||/ demote || \||/ 3579 // \/ || || \/ 3580 // +-------------------------+ || +-------------------------+ 3581 // | if (p0.new) memw(..)=r0 | || | if (p0) memw(..)=r0.new | 3582 // +-------------------------+ || +-------------------------+ 3583 // || || || 3584 // || demote \||/ 3585 // promote || \/ NOT possible 3586 // || || /\~ 3587 // \||/ || /||\~ 3588 // \/ || || 3589 // +-----------------------------+ 3590 // | if (p0.new) memw(..)=r0.new | 3591 // +-----------------------------+ 3592 // Double Dot New Store 3593 // 3594 // Returns the most basic instruction for the .new predicated instructions and 3595 // new-value stores. 3596 // For example, all of the following instructions will be converted back to the 3597 // same instruction: 3598 // 1) if (p0.new) memw(R0+#0) = R1.new ---> 3599 // 2) if (p0) memw(R0+#0)= R1.new -------> if (p0) memw(R0+#0) = R1 3600 // 3) if (p0.new) memw(R0+#0) = R1 ---> 3601 // 3602 // To understand the translation of instruction 1 to its original form, consider 3603 // a packet with 3 instructions. 3604 // { p0 = cmp.eq(R0,R1) 3605 // if (p0.new) R2 = add(R3, R4) 3606 // R5 = add (R3, R1) 3607 // } 3608 // if (p0) memw(R5+#0) = R2 <--- trying to include it in the previous packet 3609 // 3610 // This instruction can be part of the previous packet only if both p0 and R2 3611 // are promoted to .new values. This promotion happens in steps, first 3612 // predicate register is promoted to .new and in the next iteration R2 is 3613 // promoted. Therefore, in case of dependence check failure (due to R5) during 3614 // next iteration, it should be converted back to its most basic form. 3615 3616 // Return the new value instruction for a given store. 3617 int HexagonInstrInfo::getDotNewOp(const MachineInstr &MI) const { 3618 int NVOpcode = Hexagon::getNewValueOpcode(MI.getOpcode()); 3619 if (NVOpcode >= 0) // Valid new-value store instruction. 3620 return NVOpcode; 3621 3622 switch (MI.getOpcode()) { 3623 default: 3624 report_fatal_error(std::string("Unknown .new type: ") + 3625 std::to_string(MI.getOpcode())); 3626 case Hexagon::S4_storerb_ur: 3627 return Hexagon::S4_storerbnew_ur; 3628 3629 case Hexagon::S2_storerb_pci: 3630 return Hexagon::S2_storerb_pci; 3631 3632 case Hexagon::S2_storeri_pci: 3633 return Hexagon::S2_storeri_pci; 3634 3635 case Hexagon::S2_storerh_pci: 3636 return Hexagon::S2_storerh_pci; 3637 3638 case Hexagon::S2_storerd_pci: 3639 return Hexagon::S2_storerd_pci; 3640 3641 case Hexagon::S2_storerf_pci: 3642 return Hexagon::S2_storerf_pci; 3643 3644 case Hexagon::V6_vS32b_ai: 3645 return Hexagon::V6_vS32b_new_ai; 3646 3647 case Hexagon::V6_vS32b_pi: 3648 return Hexagon::V6_vS32b_new_pi; 3649 } 3650 return 0; 3651 } 3652 3653 // Returns the opcode to use when converting MI, which is a conditional jump, 3654 // into a conditional instruction which uses the .new value of the predicate. 3655 // We also use branch probabilities to add a hint to the jump. 3656 // If MBPI is null, all edges will be treated as equally likely for the 3657 // purposes of establishing a predication hint. 3658 int HexagonInstrInfo::getDotNewPredJumpOp(const MachineInstr &MI, 3659 const MachineBranchProbabilityInfo *MBPI) const { 3660 // We assume that block can have at most two successors. 3661 const MachineBasicBlock *Src = MI.getParent(); 3662 const MachineOperand &BrTarget = MI.getOperand(1); 3663 bool Taken = false; 3664 const BranchProbability OneHalf(1, 2); 3665 3666 auto getEdgeProbability = [MBPI] (const MachineBasicBlock *Src, 3667 const MachineBasicBlock *Dst) { 3668 if (MBPI) 3669 return MBPI->getEdgeProbability(Src, Dst); 3670 return BranchProbability(1, Src->succ_size()); 3671 }; 3672 3673 if (BrTarget.isMBB()) { 3674 const MachineBasicBlock *Dst = BrTarget.getMBB(); 3675 Taken = getEdgeProbability(Src, Dst) >= OneHalf; 3676 } else { 3677 // The branch target is not a basic block (most likely a function). 3678 // Since BPI only gives probabilities for targets that are basic blocks, 3679 // try to identify another target of this branch (potentially a fall- 3680 // -through) and check the probability of that target. 3681 // 3682 // The only handled branch combinations are: 3683 // - one conditional branch, 3684 // - one conditional branch followed by one unconditional branch. 3685 // Otherwise, assume not-taken. 3686 assert(MI.isConditionalBranch()); 3687 const MachineBasicBlock &B = *MI.getParent(); 3688 bool SawCond = false, Bad = false; 3689 for (const MachineInstr &I : B) { 3690 if (!I.isBranch()) 3691 continue; 3692 if (I.isConditionalBranch()) { 3693 SawCond = true; 3694 if (&I != &MI) { 3695 Bad = true; 3696 break; 3697 } 3698 } 3699 if (I.isUnconditionalBranch() && !SawCond) { 3700 Bad = true; 3701 break; 3702 } 3703 } 3704 if (!Bad) { 3705 MachineBasicBlock::const_instr_iterator It(MI); 3706 MachineBasicBlock::const_instr_iterator NextIt = std::next(It); 3707 if (NextIt == B.instr_end()) { 3708 // If this branch is the last, look for the fall-through block. 3709 for (const MachineBasicBlock *SB : B.successors()) { 3710 if (!B.isLayoutSuccessor(SB)) 3711 continue; 3712 Taken = getEdgeProbability(Src, SB) < OneHalf; 3713 break; 3714 } 3715 } else { 3716 assert(NextIt->isUnconditionalBranch()); 3717 // Find the first MBB operand and assume it's the target. 3718 const MachineBasicBlock *BT = nullptr; 3719 for (const MachineOperand &Op : NextIt->operands()) { 3720 if (!Op.isMBB()) 3721 continue; 3722 BT = Op.getMBB(); 3723 break; 3724 } 3725 Taken = BT && getEdgeProbability(Src, BT) < OneHalf; 3726 } 3727 } // if (!Bad) 3728 } 3729 3730 // The Taken flag should be set to something reasonable by this point. 3731 3732 switch (MI.getOpcode()) { 3733 case Hexagon::J2_jumpt: 3734 return Taken ? Hexagon::J2_jumptnewpt : Hexagon::J2_jumptnew; 3735 case Hexagon::J2_jumpf: 3736 return Taken ? Hexagon::J2_jumpfnewpt : Hexagon::J2_jumpfnew; 3737 3738 default: 3739 llvm_unreachable("Unexpected jump instruction."); 3740 } 3741 } 3742 3743 // Return .new predicate version for an instruction. 3744 int HexagonInstrInfo::getDotNewPredOp(const MachineInstr &MI, 3745 const MachineBranchProbabilityInfo *MBPI) const { 3746 switch (MI.getOpcode()) { 3747 // Condtional Jumps 3748 case Hexagon::J2_jumpt: 3749 case Hexagon::J2_jumpf: 3750 return getDotNewPredJumpOp(MI, MBPI); 3751 } 3752 3753 int NewOpcode = Hexagon::getPredNewOpcode(MI.getOpcode()); 3754 if (NewOpcode >= 0) 3755 return NewOpcode; 3756 return 0; 3757 } 3758 3759 int HexagonInstrInfo::getDotOldOp(const MachineInstr &MI) const { 3760 int NewOp = MI.getOpcode(); 3761 if (isPredicated(NewOp) && isPredicatedNew(NewOp)) { // Get predicate old form 3762 NewOp = Hexagon::getPredOldOpcode(NewOp); 3763 // All Hexagon architectures have prediction bits on dot-new branches, 3764 // but only Hexagon V60+ has prediction bits on dot-old ones. Make sure 3765 // to pick the right opcode when converting back to dot-old. 3766 if (!Subtarget.getFeatureBits()[Hexagon::ArchV60]) { 3767 switch (NewOp) { 3768 case Hexagon::J2_jumptpt: 3769 NewOp = Hexagon::J2_jumpt; 3770 break; 3771 case Hexagon::J2_jumpfpt: 3772 NewOp = Hexagon::J2_jumpf; 3773 break; 3774 case Hexagon::J2_jumprtpt: 3775 NewOp = Hexagon::J2_jumprt; 3776 break; 3777 case Hexagon::J2_jumprfpt: 3778 NewOp = Hexagon::J2_jumprf; 3779 break; 3780 } 3781 } 3782 assert(NewOp >= 0 && 3783 "Couldn't change predicate new instruction to its old form."); 3784 } 3785 3786 if (isNewValueStore(NewOp)) { // Convert into non-new-value format 3787 NewOp = Hexagon::getNonNVStore(NewOp); 3788 assert(NewOp >= 0 && "Couldn't change new-value store to its old form."); 3789 } 3790 3791 if (Subtarget.hasV60Ops()) 3792 return NewOp; 3793 3794 // Subtargets prior to V60 didn't support 'taken' forms of predicated jumps. 3795 switch (NewOp) { 3796 case Hexagon::J2_jumpfpt: 3797 return Hexagon::J2_jumpf; 3798 case Hexagon::J2_jumptpt: 3799 return Hexagon::J2_jumpt; 3800 case Hexagon::J2_jumprfpt: 3801 return Hexagon::J2_jumprf; 3802 case Hexagon::J2_jumprtpt: 3803 return Hexagon::J2_jumprt; 3804 } 3805 return NewOp; 3806 } 3807 3808 // See if instruction could potentially be a duplex candidate. 3809 // If so, return its group. Zero otherwise. 3810 HexagonII::SubInstructionGroup HexagonInstrInfo::getDuplexCandidateGroup( 3811 const MachineInstr &MI) const { 3812 unsigned DstReg, SrcReg, Src1Reg, Src2Reg; 3813 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 3814 3815 switch (MI.getOpcode()) { 3816 default: 3817 return HexagonII::HSIG_None; 3818 // 3819 // Group L1: 3820 // 3821 // Rd = memw(Rs+#u4:2) 3822 // Rd = memub(Rs+#u4:0) 3823 case Hexagon::L2_loadri_io: 3824 case Hexagon::dup_L2_loadri_io: 3825 DstReg = MI.getOperand(0).getReg(); 3826 SrcReg = MI.getOperand(1).getReg(); 3827 // Special case this one from Group L2. 3828 // Rd = memw(r29+#u5:2) 3829 if (isIntRegForSubInst(DstReg)) { 3830 if (Hexagon::IntRegsRegClass.contains(SrcReg) && 3831 HRI.getStackRegister() == SrcReg && 3832 MI.getOperand(2).isImm() && 3833 isShiftedUInt<5,2>(MI.getOperand(2).getImm())) 3834 return HexagonII::HSIG_L2; 3835 // Rd = memw(Rs+#u4:2) 3836 if (isIntRegForSubInst(SrcReg) && 3837 (MI.getOperand(2).isImm() && 3838 isShiftedUInt<4,2>(MI.getOperand(2).getImm()))) 3839 return HexagonII::HSIG_L1; 3840 } 3841 break; 3842 case Hexagon::L2_loadrub_io: 3843 case Hexagon::dup_L2_loadrub_io: 3844 // Rd = memub(Rs+#u4:0) 3845 DstReg = MI.getOperand(0).getReg(); 3846 SrcReg = MI.getOperand(1).getReg(); 3847 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 3848 MI.getOperand(2).isImm() && isUInt<4>(MI.getOperand(2).getImm())) 3849 return HexagonII::HSIG_L1; 3850 break; 3851 // 3852 // Group L2: 3853 // 3854 // Rd = memh/memuh(Rs+#u3:1) 3855 // Rd = memb(Rs+#u3:0) 3856 // Rd = memw(r29+#u5:2) - Handled above. 3857 // Rdd = memd(r29+#u5:3) 3858 // deallocframe 3859 // [if ([!]p0[.new])] dealloc_return 3860 // [if ([!]p0[.new])] jumpr r31 3861 case Hexagon::L2_loadrh_io: 3862 case Hexagon::L2_loadruh_io: 3863 case Hexagon::dup_L2_loadrh_io: 3864 case Hexagon::dup_L2_loadruh_io: 3865 // Rd = memh/memuh(Rs+#u3:1) 3866 DstReg = MI.getOperand(0).getReg(); 3867 SrcReg = MI.getOperand(1).getReg(); 3868 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 3869 MI.getOperand(2).isImm() && 3870 isShiftedUInt<3,1>(MI.getOperand(2).getImm())) 3871 return HexagonII::HSIG_L2; 3872 break; 3873 case Hexagon::L2_loadrb_io: 3874 case Hexagon::dup_L2_loadrb_io: 3875 // Rd = memb(Rs+#u3:0) 3876 DstReg = MI.getOperand(0).getReg(); 3877 SrcReg = MI.getOperand(1).getReg(); 3878 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 3879 MI.getOperand(2).isImm() && 3880 isUInt<3>(MI.getOperand(2).getImm())) 3881 return HexagonII::HSIG_L2; 3882 break; 3883 case Hexagon::L2_loadrd_io: 3884 case Hexagon::dup_L2_loadrd_io: 3885 // Rdd = memd(r29+#u5:3) 3886 DstReg = MI.getOperand(0).getReg(); 3887 SrcReg = MI.getOperand(1).getReg(); 3888 if (isDblRegForSubInst(DstReg, HRI) && 3889 Hexagon::IntRegsRegClass.contains(SrcReg) && 3890 HRI.getStackRegister() == SrcReg && 3891 MI.getOperand(2).isImm() && 3892 isShiftedUInt<5,3>(MI.getOperand(2).getImm())) 3893 return HexagonII::HSIG_L2; 3894 break; 3895 // dealloc_return is not documented in Hexagon Manual, but marked 3896 // with A_SUBINSN attribute in iset_v4classic.py. 3897 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4: 3898 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4_PIC: 3899 case Hexagon::L4_return: 3900 case Hexagon::L2_deallocframe: 3901 case Hexagon::dup_L2_deallocframe: 3902 return HexagonII::HSIG_L2; 3903 case Hexagon::EH_RETURN_JMPR: 3904 case Hexagon::PS_jmpret: 3905 case Hexagon::SL2_jumpr31: 3906 // jumpr r31 3907 // Actual form JMPR implicit-def %pc, implicit %r31, implicit internal %r0 3908 DstReg = MI.getOperand(0).getReg(); 3909 if (Hexagon::IntRegsRegClass.contains(DstReg) && (Hexagon::R31 == DstReg)) 3910 return HexagonII::HSIG_L2; 3911 break; 3912 case Hexagon::PS_jmprett: 3913 case Hexagon::PS_jmpretf: 3914 case Hexagon::PS_jmprettnewpt: 3915 case Hexagon::PS_jmpretfnewpt: 3916 case Hexagon::PS_jmprettnew: 3917 case Hexagon::PS_jmpretfnew: 3918 case Hexagon::SL2_jumpr31_t: 3919 case Hexagon::SL2_jumpr31_f: 3920 case Hexagon::SL2_jumpr31_tnew: 3921 case Hexagon::SL2_jumpr31_fnew: 3922 DstReg = MI.getOperand(1).getReg(); 3923 SrcReg = MI.getOperand(0).getReg(); 3924 // [if ([!]p0[.new])] jumpr r31 3925 if ((Hexagon::PredRegsRegClass.contains(SrcReg) && 3926 (Hexagon::P0 == SrcReg)) && 3927 (Hexagon::IntRegsRegClass.contains(DstReg) && (Hexagon::R31 == DstReg))) 3928 return HexagonII::HSIG_L2; 3929 break; 3930 case Hexagon::L4_return_t: 3931 case Hexagon::L4_return_f: 3932 case Hexagon::L4_return_tnew_pnt: 3933 case Hexagon::L4_return_fnew_pnt: 3934 case Hexagon::L4_return_tnew_pt: 3935 case Hexagon::L4_return_fnew_pt: 3936 // [if ([!]p0[.new])] dealloc_return 3937 SrcReg = MI.getOperand(0).getReg(); 3938 if (Hexagon::PredRegsRegClass.contains(SrcReg) && (Hexagon::P0 == SrcReg)) 3939 return HexagonII::HSIG_L2; 3940 break; 3941 // 3942 // Group S1: 3943 // 3944 // memw(Rs+#u4:2) = Rt 3945 // memb(Rs+#u4:0) = Rt 3946 case Hexagon::S2_storeri_io: 3947 case Hexagon::dup_S2_storeri_io: 3948 // Special case this one from Group S2. 3949 // memw(r29+#u5:2) = Rt 3950 Src1Reg = MI.getOperand(0).getReg(); 3951 Src2Reg = MI.getOperand(2).getReg(); 3952 if (Hexagon::IntRegsRegClass.contains(Src1Reg) && 3953 isIntRegForSubInst(Src2Reg) && 3954 HRI.getStackRegister() == Src1Reg && MI.getOperand(1).isImm() && 3955 isShiftedUInt<5,2>(MI.getOperand(1).getImm())) 3956 return HexagonII::HSIG_S2; 3957 // memw(Rs+#u4:2) = Rt 3958 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) && 3959 MI.getOperand(1).isImm() && 3960 isShiftedUInt<4,2>(MI.getOperand(1).getImm())) 3961 return HexagonII::HSIG_S1; 3962 break; 3963 case Hexagon::S2_storerb_io: 3964 case Hexagon::dup_S2_storerb_io: 3965 // memb(Rs+#u4:0) = Rt 3966 Src1Reg = MI.getOperand(0).getReg(); 3967 Src2Reg = MI.getOperand(2).getReg(); 3968 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) && 3969 MI.getOperand(1).isImm() && isUInt<4>(MI.getOperand(1).getImm())) 3970 return HexagonII::HSIG_S1; 3971 break; 3972 // 3973 // Group S2: 3974 // 3975 // memh(Rs+#u3:1) = Rt 3976 // memw(r29+#u5:2) = Rt 3977 // memd(r29+#s6:3) = Rtt 3978 // memw(Rs+#u4:2) = #U1 3979 // memb(Rs+#u4) = #U1 3980 // allocframe(#u5:3) 3981 case Hexagon::S2_storerh_io: 3982 case Hexagon::dup_S2_storerh_io: 3983 // memh(Rs+#u3:1) = Rt 3984 Src1Reg = MI.getOperand(0).getReg(); 3985 Src2Reg = MI.getOperand(2).getReg(); 3986 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) && 3987 MI.getOperand(1).isImm() && 3988 isShiftedUInt<3,1>(MI.getOperand(1).getImm())) 3989 return HexagonII::HSIG_S1; 3990 break; 3991 case Hexagon::S2_storerd_io: 3992 case Hexagon::dup_S2_storerd_io: 3993 // memd(r29+#s6:3) = Rtt 3994 Src1Reg = MI.getOperand(0).getReg(); 3995 Src2Reg = MI.getOperand(2).getReg(); 3996 if (isDblRegForSubInst(Src2Reg, HRI) && 3997 Hexagon::IntRegsRegClass.contains(Src1Reg) && 3998 HRI.getStackRegister() == Src1Reg && MI.getOperand(1).isImm() && 3999 isShiftedInt<6,3>(MI.getOperand(1).getImm())) 4000 return HexagonII::HSIG_S2; 4001 break; 4002 case Hexagon::S4_storeiri_io: 4003 case Hexagon::dup_S4_storeiri_io: 4004 // memw(Rs+#u4:2) = #U1 4005 Src1Reg = MI.getOperand(0).getReg(); 4006 if (isIntRegForSubInst(Src1Reg) && MI.getOperand(1).isImm() && 4007 isShiftedUInt<4,2>(MI.getOperand(1).getImm()) && 4008 MI.getOperand(2).isImm() && isUInt<1>(MI.getOperand(2).getImm())) 4009 return HexagonII::HSIG_S2; 4010 break; 4011 case Hexagon::S4_storeirb_io: 4012 case Hexagon::dup_S4_storeirb_io: 4013 // memb(Rs+#u4) = #U1 4014 Src1Reg = MI.getOperand(0).getReg(); 4015 if (isIntRegForSubInst(Src1Reg) && 4016 MI.getOperand(1).isImm() && isUInt<4>(MI.getOperand(1).getImm()) && 4017 MI.getOperand(2).isImm() && isUInt<1>(MI.getOperand(2).getImm())) 4018 return HexagonII::HSIG_S2; 4019 break; 4020 case Hexagon::S2_allocframe: 4021 case Hexagon::dup_S2_allocframe: 4022 if (MI.getOperand(2).isImm() && 4023 isShiftedUInt<5,3>(MI.getOperand(2).getImm())) 4024 return HexagonII::HSIG_S1; 4025 break; 4026 // 4027 // Group A: 4028 // 4029 // Rx = add(Rx,#s7) 4030 // Rd = Rs 4031 // Rd = #u6 4032 // Rd = #-1 4033 // if ([!]P0[.new]) Rd = #0 4034 // Rd = add(r29,#u6:2) 4035 // Rx = add(Rx,Rs) 4036 // P0 = cmp.eq(Rs,#u2) 4037 // Rdd = combine(#0,Rs) 4038 // Rdd = combine(Rs,#0) 4039 // Rdd = combine(#u2,#U2) 4040 // Rd = add(Rs,#1) 4041 // Rd = add(Rs,#-1) 4042 // Rd = sxth/sxtb/zxtb/zxth(Rs) 4043 // Rd = and(Rs,#1) 4044 case Hexagon::A2_addi: 4045 case Hexagon::dup_A2_addi: 4046 DstReg = MI.getOperand(0).getReg(); 4047 SrcReg = MI.getOperand(1).getReg(); 4048 if (isIntRegForSubInst(DstReg)) { 4049 // Rd = add(r29,#u6:2) 4050 if (Hexagon::IntRegsRegClass.contains(SrcReg) && 4051 HRI.getStackRegister() == SrcReg && MI.getOperand(2).isImm() && 4052 isShiftedUInt<6,2>(MI.getOperand(2).getImm())) 4053 return HexagonII::HSIG_A; 4054 // Rx = add(Rx,#s7) 4055 if ((DstReg == SrcReg) && MI.getOperand(2).isImm() && 4056 isInt<7>(MI.getOperand(2).getImm())) 4057 return HexagonII::HSIG_A; 4058 // Rd = add(Rs,#1) 4059 // Rd = add(Rs,#-1) 4060 if (isIntRegForSubInst(SrcReg) && MI.getOperand(2).isImm() && 4061 ((MI.getOperand(2).getImm() == 1) || 4062 (MI.getOperand(2).getImm() == -1))) 4063 return HexagonII::HSIG_A; 4064 } 4065 break; 4066 case Hexagon::A2_add: 4067 case Hexagon::dup_A2_add: 4068 // Rx = add(Rx,Rs) 4069 DstReg = MI.getOperand(0).getReg(); 4070 Src1Reg = MI.getOperand(1).getReg(); 4071 Src2Reg = MI.getOperand(2).getReg(); 4072 if (isIntRegForSubInst(DstReg) && (DstReg == Src1Reg) && 4073 isIntRegForSubInst(Src2Reg)) 4074 return HexagonII::HSIG_A; 4075 break; 4076 case Hexagon::A2_andir: 4077 case Hexagon::dup_A2_andir: 4078 // Same as zxtb. 4079 // Rd16=and(Rs16,#255) 4080 // Rd16=and(Rs16,#1) 4081 DstReg = MI.getOperand(0).getReg(); 4082 SrcReg = MI.getOperand(1).getReg(); 4083 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) && 4084 MI.getOperand(2).isImm() && 4085 ((MI.getOperand(2).getImm() == 1) || 4086 (MI.getOperand(2).getImm() == 255))) 4087 return HexagonII::HSIG_A; 4088 break; 4089 case Hexagon::A2_tfr: 4090 case Hexagon::dup_A2_tfr: 4091 // Rd = Rs 4092 DstReg = MI.getOperand(0).getReg(); 4093 SrcReg = MI.getOperand(1).getReg(); 4094 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg)) 4095 return HexagonII::HSIG_A; 4096 break; 4097 case Hexagon::A2_tfrsi: 4098 case Hexagon::dup_A2_tfrsi: 4099 // Rd = #u6 4100 // Do not test for #u6 size since the const is getting extended 4101 // regardless and compound could be formed. 4102 // Rd = #-1 4103 DstReg = MI.getOperand(0).getReg(); 4104 if (isIntRegForSubInst(DstReg)) 4105 return HexagonII::HSIG_A; 4106 break; 4107 case Hexagon::C2_cmoveit: 4108 case Hexagon::C2_cmovenewit: 4109 case Hexagon::C2_cmoveif: 4110 case Hexagon::C2_cmovenewif: 4111 case Hexagon::dup_C2_cmoveit: 4112 case Hexagon::dup_C2_cmovenewit: 4113 case Hexagon::dup_C2_cmoveif: 4114 case Hexagon::dup_C2_cmovenewif: 4115 // if ([!]P0[.new]) Rd = #0 4116 // Actual form: 4117 // %r16 = C2_cmovenewit internal %p0, 0, implicit undef %r16; 4118 DstReg = MI.getOperand(0).getReg(); 4119 SrcReg = MI.getOperand(1).getReg(); 4120 if (isIntRegForSubInst(DstReg) && 4121 Hexagon::PredRegsRegClass.contains(SrcReg) && Hexagon::P0 == SrcReg && 4122 MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0) 4123 return HexagonII::HSIG_A; 4124 break; 4125 case Hexagon::C2_cmpeqi: 4126 case Hexagon::dup_C2_cmpeqi: 4127 // P0 = cmp.eq(Rs,#u2) 4128 DstReg = MI.getOperand(0).getReg(); 4129 SrcReg = MI.getOperand(1).getReg(); 4130 if (Hexagon::PredRegsRegClass.contains(DstReg) && 4131 Hexagon::P0 == DstReg && isIntRegForSubInst(SrcReg) && 4132 MI.getOperand(2).isImm() && isUInt<2>(MI.getOperand(2).getImm())) 4133 return HexagonII::HSIG_A; 4134 break; 4135 case Hexagon::A2_combineii: 4136 case Hexagon::A4_combineii: 4137 case Hexagon::dup_A2_combineii: 4138 case Hexagon::dup_A4_combineii: 4139 // Rdd = combine(#u2,#U2) 4140 DstReg = MI.getOperand(0).getReg(); 4141 if (isDblRegForSubInst(DstReg, HRI) && 4142 ((MI.getOperand(1).isImm() && isUInt<2>(MI.getOperand(1).getImm())) || 4143 (MI.getOperand(1).isGlobal() && 4144 isUInt<2>(MI.getOperand(1).getOffset()))) && 4145 ((MI.getOperand(2).isImm() && isUInt<2>(MI.getOperand(2).getImm())) || 4146 (MI.getOperand(2).isGlobal() && 4147 isUInt<2>(MI.getOperand(2).getOffset())))) 4148 return HexagonII::HSIG_A; 4149 break; 4150 case Hexagon::A4_combineri: 4151 case Hexagon::dup_A4_combineri: 4152 // Rdd = combine(Rs,#0) 4153 // Rdd = combine(Rs,#0) 4154 DstReg = MI.getOperand(0).getReg(); 4155 SrcReg = MI.getOperand(1).getReg(); 4156 if (isDblRegForSubInst(DstReg, HRI) && isIntRegForSubInst(SrcReg) && 4157 ((MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0) || 4158 (MI.getOperand(2).isGlobal() && MI.getOperand(2).getOffset() == 0))) 4159 return HexagonII::HSIG_A; 4160 break; 4161 case Hexagon::A4_combineir: 4162 case Hexagon::dup_A4_combineir: 4163 // Rdd = combine(#0,Rs) 4164 DstReg = MI.getOperand(0).getReg(); 4165 SrcReg = MI.getOperand(2).getReg(); 4166 if (isDblRegForSubInst(DstReg, HRI) && isIntRegForSubInst(SrcReg) && 4167 ((MI.getOperand(1).isImm() && MI.getOperand(1).getImm() == 0) || 4168 (MI.getOperand(1).isGlobal() && MI.getOperand(1).getOffset() == 0))) 4169 return HexagonII::HSIG_A; 4170 break; 4171 case Hexagon::A2_sxtb: 4172 case Hexagon::A2_sxth: 4173 case Hexagon::A2_zxtb: 4174 case Hexagon::A2_zxth: 4175 case Hexagon::dup_A2_sxtb: 4176 case Hexagon::dup_A2_sxth: 4177 case Hexagon::dup_A2_zxtb: 4178 case Hexagon::dup_A2_zxth: 4179 // Rd = sxth/sxtb/zxtb/zxth(Rs) 4180 DstReg = MI.getOperand(0).getReg(); 4181 SrcReg = MI.getOperand(1).getReg(); 4182 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg)) 4183 return HexagonII::HSIG_A; 4184 break; 4185 } 4186 4187 return HexagonII::HSIG_None; 4188 } 4189 4190 short HexagonInstrInfo::getEquivalentHWInstr(const MachineInstr &MI) const { 4191 return Hexagon::getRealHWInstr(MI.getOpcode(), Hexagon::InstrType_Real); 4192 } 4193 4194 unsigned HexagonInstrInfo::getInstrTimingClassLatency( 4195 const InstrItineraryData *ItinData, const MachineInstr &MI) const { 4196 // Default to one cycle for no itinerary. However, an "empty" itinerary may 4197 // still have a MinLatency property, which getStageLatency checks. 4198 if (!ItinData) 4199 return getInstrLatency(ItinData, MI); 4200 4201 if (MI.isTransient()) 4202 return 0; 4203 return ItinData->getStageLatency(MI.getDesc().getSchedClass()); 4204 } 4205 4206 /// getOperandLatency - Compute and return the use operand latency of a given 4207 /// pair of def and use. 4208 /// In most cases, the static scheduling itinerary was enough to determine the 4209 /// operand latency. But it may not be possible for instructions with variable 4210 /// number of defs / uses. 4211 /// 4212 /// This is a raw interface to the itinerary that may be directly overriden by 4213 /// a target. Use computeOperandLatency to get the best estimate of latency. 4214 int HexagonInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 4215 const MachineInstr &DefMI, 4216 unsigned DefIdx, 4217 const MachineInstr &UseMI, 4218 unsigned UseIdx) const { 4219 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 4220 4221 // Get DefIdx and UseIdx for super registers. 4222 const MachineOperand &DefMO = DefMI.getOperand(DefIdx); 4223 4224 if (DefMO.isReg() && Register::isPhysicalRegister(DefMO.getReg())) { 4225 if (DefMO.isImplicit()) { 4226 for (MCSuperRegIterator SR(DefMO.getReg(), &HRI); SR.isValid(); ++SR) { 4227 int Idx = DefMI.findRegisterDefOperandIdx(*SR, false, false, &HRI); 4228 if (Idx != -1) { 4229 DefIdx = Idx; 4230 break; 4231 } 4232 } 4233 } 4234 4235 const MachineOperand &UseMO = UseMI.getOperand(UseIdx); 4236 if (UseMO.isImplicit()) { 4237 for (MCSuperRegIterator SR(UseMO.getReg(), &HRI); SR.isValid(); ++SR) { 4238 int Idx = UseMI.findRegisterUseOperandIdx(*SR, false, &HRI); 4239 if (Idx != -1) { 4240 UseIdx = Idx; 4241 break; 4242 } 4243 } 4244 } 4245 } 4246 4247 int Latency = TargetInstrInfo::getOperandLatency(ItinData, DefMI, DefIdx, 4248 UseMI, UseIdx); 4249 if (!Latency) 4250 // We should never have 0 cycle latency between two instructions unless 4251 // they can be packetized together. However, this decision can't be made 4252 // here. 4253 Latency = 1; 4254 return Latency; 4255 } 4256 4257 // inverts the predication logic. 4258 // p -> NotP 4259 // NotP -> P 4260 bool HexagonInstrInfo::getInvertedPredSense( 4261 SmallVectorImpl<MachineOperand> &Cond) const { 4262 if (Cond.empty()) 4263 return false; 4264 unsigned Opc = getInvertedPredicatedOpcode(Cond[0].getImm()); 4265 Cond[0].setImm(Opc); 4266 return true; 4267 } 4268 4269 unsigned HexagonInstrInfo::getInvertedPredicatedOpcode(const int Opc) const { 4270 int InvPredOpcode; 4271 InvPredOpcode = isPredicatedTrue(Opc) ? Hexagon::getFalsePredOpcode(Opc) 4272 : Hexagon::getTruePredOpcode(Opc); 4273 if (InvPredOpcode >= 0) // Valid instruction with the inverted predicate. 4274 return InvPredOpcode; 4275 4276 llvm_unreachable("Unexpected predicated instruction"); 4277 } 4278 4279 // Returns the max value that doesn't need to be extended. 4280 int HexagonInstrInfo::getMaxValue(const MachineInstr &MI) const { 4281 const uint64_t F = MI.getDesc().TSFlags; 4282 unsigned isSigned = (F >> HexagonII::ExtentSignedPos) 4283 & HexagonII::ExtentSignedMask; 4284 unsigned bits = (F >> HexagonII::ExtentBitsPos) 4285 & HexagonII::ExtentBitsMask; 4286 4287 if (isSigned) // if value is signed 4288 return ~(-1U << (bits - 1)); 4289 else 4290 return ~(-1U << bits); 4291 } 4292 4293 4294 bool HexagonInstrInfo::isAddrModeWithOffset(const MachineInstr &MI) const { 4295 switch (MI.getOpcode()) { 4296 case Hexagon::L2_loadrbgp: 4297 case Hexagon::L2_loadrdgp: 4298 case Hexagon::L2_loadrhgp: 4299 case Hexagon::L2_loadrigp: 4300 case Hexagon::L2_loadrubgp: 4301 case Hexagon::L2_loadruhgp: 4302 case Hexagon::S2_storerbgp: 4303 case Hexagon::S2_storerbnewgp: 4304 case Hexagon::S2_storerhgp: 4305 case Hexagon::S2_storerhnewgp: 4306 case Hexagon::S2_storerigp: 4307 case Hexagon::S2_storerinewgp: 4308 case Hexagon::S2_storerdgp: 4309 case Hexagon::S2_storerfgp: 4310 return true; 4311 } 4312 const uint64_t F = MI.getDesc().TSFlags; 4313 unsigned addrMode = 4314 ((F >> HexagonII::AddrModePos) & HexagonII::AddrModeMask); 4315 // Disallow any base+offset instruction. The assembler does not yet reorder 4316 // based up any zero offset instruction. 4317 return (addrMode == HexagonII::BaseRegOffset || 4318 addrMode == HexagonII::BaseImmOffset || 4319 addrMode == HexagonII::BaseLongOffset); 4320 } 4321 4322 bool HexagonInstrInfo::isPureSlot0(const MachineInstr &MI) const { 4323 // Workaround for the Global Scheduler. Sometimes, it creates 4324 // A4_ext as a Pseudo instruction and calls this function to see if 4325 // it can be added to an existing bundle. Since the instruction doesn't 4326 // belong to any BB yet, we can't use getUnits API. 4327 if (MI.getOpcode() == Hexagon::A4_ext) 4328 return false; 4329 4330 unsigned FuncUnits = getUnits(MI); 4331 return HexagonFUnits::isSlot0Only(FuncUnits); 4332 } 4333 4334 bool HexagonInstrInfo::isRestrictNoSlot1Store(const MachineInstr &MI) const { 4335 const uint64_t F = MI.getDesc().TSFlags; 4336 return ((F >> HexagonII::RestrictNoSlot1StorePos) & 4337 HexagonII::RestrictNoSlot1StoreMask); 4338 } 4339 4340 void HexagonInstrInfo::changeDuplexOpcode(MachineBasicBlock::instr_iterator MII, 4341 bool ToBigInstrs) const { 4342 int Opcode = -1; 4343 if (ToBigInstrs) { // To BigCore Instr. 4344 // Check if the instruction can form a Duplex. 4345 if (getDuplexCandidateGroup(*MII)) 4346 // Get the opcode marked "dup_*" tag. 4347 Opcode = getDuplexOpcode(*MII, ToBigInstrs); 4348 } else // To TinyCore Instr. 4349 Opcode = getDuplexOpcode(*MII, ToBigInstrs); 4350 4351 // Change the opcode of the instruction. 4352 if (Opcode >= 0) 4353 MII->setDesc(get(Opcode)); 4354 } 4355 4356 // This function is used to translate instructions to facilitate generating 4357 // Duplexes on TinyCore. 4358 void HexagonInstrInfo::translateInstrsForDup(MachineFunction &MF, 4359 bool ToBigInstrs) const { 4360 for (auto &MB : MF) 4361 for (MachineBasicBlock::instr_iterator Instr = MB.instr_begin(), 4362 End = MB.instr_end(); 4363 Instr != End; ++Instr) 4364 changeDuplexOpcode(Instr, ToBigInstrs); 4365 } 4366 4367 // This is a specialized form of above function. 4368 void HexagonInstrInfo::translateInstrsForDup( 4369 MachineBasicBlock::instr_iterator MII, bool ToBigInstrs) const { 4370 MachineBasicBlock *MBB = MII->getParent(); 4371 while ((MII != MBB->instr_end()) && MII->isInsideBundle()) { 4372 changeDuplexOpcode(MII, ToBigInstrs); 4373 ++MII; 4374 } 4375 } 4376 4377 unsigned HexagonInstrInfo::getMemAccessSize(const MachineInstr &MI) const { 4378 using namespace HexagonII; 4379 4380 const uint64_t F = MI.getDesc().TSFlags; 4381 unsigned S = (F >> MemAccessSizePos) & MemAccesSizeMask; 4382 unsigned Size = getMemAccessSizeInBytes(MemAccessSize(S)); 4383 if (Size != 0) 4384 return Size; 4385 4386 // Handle vector access sizes. 4387 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); 4388 switch (S) { 4389 case HexagonII::HVXVectorAccess: 4390 return HRI.getSpillSize(Hexagon::HvxVRRegClass); 4391 default: 4392 llvm_unreachable("Unexpected instruction"); 4393 } 4394 } 4395 4396 // Returns the min value that doesn't need to be extended. 4397 int HexagonInstrInfo::getMinValue(const MachineInstr &MI) const { 4398 const uint64_t F = MI.getDesc().TSFlags; 4399 unsigned isSigned = (F >> HexagonII::ExtentSignedPos) 4400 & HexagonII::ExtentSignedMask; 4401 unsigned bits = (F >> HexagonII::ExtentBitsPos) 4402 & HexagonII::ExtentBitsMask; 4403 4404 if (isSigned) // if value is signed 4405 return -1U << (bits - 1); 4406 else 4407 return 0; 4408 } 4409 4410 // Returns opcode of the non-extended equivalent instruction. 4411 short HexagonInstrInfo::getNonExtOpcode(const MachineInstr &MI) const { 4412 // Check if the instruction has a register form that uses register in place 4413 // of the extended operand, if so return that as the non-extended form. 4414 short NonExtOpcode = Hexagon::getRegForm(MI.getOpcode()); 4415 if (NonExtOpcode >= 0) 4416 return NonExtOpcode; 4417 4418 if (MI.getDesc().mayLoad() || MI.getDesc().mayStore()) { 4419 // Check addressing mode and retrieve non-ext equivalent instruction. 4420 switch (getAddrMode(MI)) { 4421 case HexagonII::Absolute: 4422 return Hexagon::changeAddrMode_abs_io(MI.getOpcode()); 4423 case HexagonII::BaseImmOffset: 4424 return Hexagon::changeAddrMode_io_rr(MI.getOpcode()); 4425 case HexagonII::BaseLongOffset: 4426 return Hexagon::changeAddrMode_ur_rr(MI.getOpcode()); 4427 4428 default: 4429 return -1; 4430 } 4431 } 4432 return -1; 4433 } 4434 4435 bool HexagonInstrInfo::getPredReg(ArrayRef<MachineOperand> Cond, 4436 unsigned &PredReg, unsigned &PredRegPos, unsigned &PredRegFlags) const { 4437 if (Cond.empty()) 4438 return false; 4439 assert(Cond.size() == 2); 4440 if (isNewValueJump(Cond[0].getImm()) || Cond[1].isMBB()) { 4441 LLVM_DEBUG(dbgs() << "No predregs for new-value jumps/endloop"); 4442 return false; 4443 } 4444 PredReg = Cond[1].getReg(); 4445 PredRegPos = 1; 4446 // See IfConversion.cpp why we add RegState::Implicit | RegState::Undef 4447 PredRegFlags = 0; 4448 if (Cond[1].isImplicit()) 4449 PredRegFlags = RegState::Implicit; 4450 if (Cond[1].isUndef()) 4451 PredRegFlags |= RegState::Undef; 4452 return true; 4453 } 4454 4455 short HexagonInstrInfo::getPseudoInstrPair(const MachineInstr &MI) const { 4456 return Hexagon::getRealHWInstr(MI.getOpcode(), Hexagon::InstrType_Pseudo); 4457 } 4458 4459 short HexagonInstrInfo::getRegForm(const MachineInstr &MI) const { 4460 return Hexagon::getRegForm(MI.getOpcode()); 4461 } 4462 4463 // Return the number of bytes required to encode the instruction. 4464 // Hexagon instructions are fixed length, 4 bytes, unless they 4465 // use a constant extender, which requires another 4 bytes. 4466 // For debug instructions and prolog labels, return 0. 4467 unsigned HexagonInstrInfo::getSize(const MachineInstr &MI) const { 4468 if (MI.isDebugInstr() || MI.isPosition()) 4469 return 0; 4470 4471 unsigned Size = MI.getDesc().getSize(); 4472 if (!Size) 4473 // Assume the default insn size in case it cannot be determined 4474 // for whatever reason. 4475 Size = HEXAGON_INSTR_SIZE; 4476 4477 if (isConstExtended(MI) || isExtended(MI)) 4478 Size += HEXAGON_INSTR_SIZE; 4479 4480 // Try and compute number of instructions in asm. 4481 if (BranchRelaxAsmLarge && MI.getOpcode() == Hexagon::INLINEASM) { 4482 const MachineBasicBlock &MBB = *MI.getParent(); 4483 const MachineFunction *MF = MBB.getParent(); 4484 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo(); 4485 4486 // Count the number of register definitions to find the asm string. 4487 unsigned NumDefs = 0; 4488 for (; MI.getOperand(NumDefs).isReg() && MI.getOperand(NumDefs).isDef(); 4489 ++NumDefs) 4490 assert(NumDefs != MI.getNumOperands()-2 && "No asm string?"); 4491 4492 assert(MI.getOperand(NumDefs).isSymbol() && "No asm string?"); 4493 // Disassemble the AsmStr and approximate number of instructions. 4494 const char *AsmStr = MI.getOperand(NumDefs).getSymbolName(); 4495 Size = getInlineAsmLength(AsmStr, *MAI); 4496 } 4497 4498 return Size; 4499 } 4500 4501 uint64_t HexagonInstrInfo::getType(const MachineInstr &MI) const { 4502 const uint64_t F = MI.getDesc().TSFlags; 4503 return (F >> HexagonII::TypePos) & HexagonII::TypeMask; 4504 } 4505 4506 InstrStage::FuncUnits HexagonInstrInfo::getUnits(const MachineInstr &MI) const { 4507 const InstrItineraryData &II = *Subtarget.getInstrItineraryData(); 4508 const InstrStage &IS = *II.beginStage(MI.getDesc().getSchedClass()); 4509 4510 return IS.getUnits(); 4511 } 4512 4513 // Calculate size of the basic block without debug instructions. 4514 unsigned HexagonInstrInfo::nonDbgBBSize(const MachineBasicBlock *BB) const { 4515 return nonDbgMICount(BB->instr_begin(), BB->instr_end()); 4516 } 4517 4518 unsigned HexagonInstrInfo::nonDbgBundleSize( 4519 MachineBasicBlock::const_iterator BundleHead) const { 4520 assert(BundleHead->isBundle() && "Not a bundle header"); 4521 auto MII = BundleHead.getInstrIterator(); 4522 // Skip the bundle header. 4523 return nonDbgMICount(++MII, getBundleEnd(BundleHead.getInstrIterator())); 4524 } 4525 4526 /// immediateExtend - Changes the instruction in place to one using an immediate 4527 /// extender. 4528 void HexagonInstrInfo::immediateExtend(MachineInstr &MI) const { 4529 assert((isExtendable(MI)||isConstExtended(MI)) && 4530 "Instruction must be extendable"); 4531 // Find which operand is extendable. 4532 short ExtOpNum = getCExtOpNum(MI); 4533 MachineOperand &MO = MI.getOperand(ExtOpNum); 4534 // This needs to be something we understand. 4535 assert((MO.isMBB() || MO.isImm()) && 4536 "Branch with unknown extendable field type"); 4537 // Mark given operand as extended. 4538 MO.addTargetFlag(HexagonII::HMOTF_ConstExtended); 4539 } 4540 4541 bool HexagonInstrInfo::invertAndChangeJumpTarget( 4542 MachineInstr &MI, MachineBasicBlock *NewTarget) const { 4543 LLVM_DEBUG(dbgs() << "\n[invertAndChangeJumpTarget] to " 4544 << printMBBReference(*NewTarget); 4545 MI.dump();); 4546 assert(MI.isBranch()); 4547 unsigned NewOpcode = getInvertedPredicatedOpcode(MI.getOpcode()); 4548 int TargetPos = MI.getNumOperands() - 1; 4549 // In general branch target is the last operand, 4550 // but some implicit defs added at the end might change it. 4551 while ((TargetPos > -1) && !MI.getOperand(TargetPos).isMBB()) 4552 --TargetPos; 4553 assert((TargetPos >= 0) && MI.getOperand(TargetPos).isMBB()); 4554 MI.getOperand(TargetPos).setMBB(NewTarget); 4555 if (EnableBranchPrediction && isPredicatedNew(MI)) { 4556 NewOpcode = reversePrediction(NewOpcode); 4557 } 4558 MI.setDesc(get(NewOpcode)); 4559 return true; 4560 } 4561 4562 void HexagonInstrInfo::genAllInsnTimingClasses(MachineFunction &MF) const { 4563 /* +++ The code below is used to generate complete set of Hexagon Insn +++ */ 4564 MachineFunction::iterator A = MF.begin(); 4565 MachineBasicBlock &B = *A; 4566 MachineBasicBlock::iterator I = B.begin(); 4567 DebugLoc DL = I->getDebugLoc(); 4568 MachineInstr *NewMI; 4569 4570 for (unsigned insn = TargetOpcode::GENERIC_OP_END+1; 4571 insn < Hexagon::INSTRUCTION_LIST_END; ++insn) { 4572 NewMI = BuildMI(B, I, DL, get(insn)); 4573 LLVM_DEBUG(dbgs() << "\n" 4574 << getName(NewMI->getOpcode()) 4575 << " Class: " << NewMI->getDesc().getSchedClass()); 4576 NewMI->eraseFromParent(); 4577 } 4578 /* --- The code above is used to generate complete set of Hexagon Insn --- */ 4579 } 4580 4581 // inverts the predication logic. 4582 // p -> NotP 4583 // NotP -> P 4584 bool HexagonInstrInfo::reversePredSense(MachineInstr &MI) const { 4585 LLVM_DEBUG(dbgs() << "\nTrying to reverse pred. sense of:"; MI.dump()); 4586 MI.setDesc(get(getInvertedPredicatedOpcode(MI.getOpcode()))); 4587 return true; 4588 } 4589 4590 // Reverse the branch prediction. 4591 unsigned HexagonInstrInfo::reversePrediction(unsigned Opcode) const { 4592 int PredRevOpcode = -1; 4593 if (isPredictedTaken(Opcode)) 4594 PredRevOpcode = Hexagon::notTakenBranchPrediction(Opcode); 4595 else 4596 PredRevOpcode = Hexagon::takenBranchPrediction(Opcode); 4597 assert(PredRevOpcode > 0); 4598 return PredRevOpcode; 4599 } 4600 4601 // TODO: Add more rigorous validation. 4602 bool HexagonInstrInfo::validateBranchCond(const ArrayRef<MachineOperand> &Cond) 4603 const { 4604 return Cond.empty() || (Cond[0].isImm() && (Cond.size() != 1)); 4605 } 4606 4607 void HexagonInstrInfo:: 4608 setBundleNoShuf(MachineBasicBlock::instr_iterator MIB) const { 4609 assert(MIB->isBundle()); 4610 MachineOperand &Operand = MIB->getOperand(0); 4611 if (Operand.isImm()) 4612 Operand.setImm(Operand.getImm() | memShufDisabledMask); 4613 else 4614 MIB->addOperand(MachineOperand::CreateImm(memShufDisabledMask)); 4615 } 4616 4617 bool HexagonInstrInfo::getBundleNoShuf(const MachineInstr &MIB) const { 4618 assert(MIB.isBundle()); 4619 const MachineOperand &Operand = MIB.getOperand(0); 4620 return (Operand.isImm() && (Operand.getImm() & memShufDisabledMask) != 0); 4621 } 4622 4623 // Addressing mode relations. 4624 short HexagonInstrInfo::changeAddrMode_abs_io(short Opc) const { 4625 return Opc >= 0 ? Hexagon::changeAddrMode_abs_io(Opc) : Opc; 4626 } 4627 4628 short HexagonInstrInfo::changeAddrMode_io_abs(short Opc) const { 4629 return Opc >= 0 ? Hexagon::changeAddrMode_io_abs(Opc) : Opc; 4630 } 4631 4632 short HexagonInstrInfo::changeAddrMode_io_pi(short Opc) const { 4633 return Opc >= 0 ? Hexagon::changeAddrMode_io_pi(Opc) : Opc; 4634 } 4635 4636 short HexagonInstrInfo::changeAddrMode_io_rr(short Opc) const { 4637 return Opc >= 0 ? Hexagon::changeAddrMode_io_rr(Opc) : Opc; 4638 } 4639 4640 short HexagonInstrInfo::changeAddrMode_pi_io(short Opc) const { 4641 return Opc >= 0 ? Hexagon::changeAddrMode_pi_io(Opc) : Opc; 4642 } 4643 4644 short HexagonInstrInfo::changeAddrMode_rr_io(short Opc) const { 4645 return Opc >= 0 ? Hexagon::changeAddrMode_rr_io(Opc) : Opc; 4646 } 4647 4648 short HexagonInstrInfo::changeAddrMode_rr_ur(short Opc) const { 4649 return Opc >= 0 ? Hexagon::changeAddrMode_rr_ur(Opc) : Opc; 4650 } 4651 4652 short HexagonInstrInfo::changeAddrMode_ur_rr(short Opc) const { 4653 return Opc >= 0 ? Hexagon::changeAddrMode_ur_rr(Opc) : Opc; 4654 } 4655