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