1 //===-- SystemZInstrInfo.cpp - SystemZ instruction information ------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains the SystemZ implementation of the TargetInstrInfo class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "MCTargetDesc/SystemZMCTargetDesc.h" 15 #include "SystemZ.h" 16 #include "SystemZInstrBuilder.h" 17 #include "SystemZInstrInfo.h" 18 #include "SystemZSubtarget.h" 19 #include "llvm/CodeGen/LiveInterval.h" 20 #include "llvm/CodeGen/LiveIntervalAnalysis.h" 21 #include "llvm/CodeGen/LiveVariables.h" 22 #include "llvm/CodeGen/MachineBasicBlock.h" 23 #include "llvm/CodeGen/MachineFrameInfo.h" 24 #include "llvm/CodeGen/MachineFunction.h" 25 #include "llvm/CodeGen/MachineInstr.h" 26 #include "llvm/CodeGen/MachineInstrBuilder.h" 27 #include "llvm/CodeGen/MachineMemOperand.h" 28 #include "llvm/CodeGen/MachineOperand.h" 29 #include "llvm/CodeGen/MachineRegisterInfo.h" 30 #include "llvm/CodeGen/SlotIndexes.h" 31 #include "llvm/MC/MCInstrDesc.h" 32 #include "llvm/MC/MCRegisterInfo.h" 33 #include "llvm/Support/BranchProbability.h" 34 #include "llvm/Support/ErrorHandling.h" 35 #include "llvm/Support/MathExtras.h" 36 #include "llvm/Target/TargetInstrInfo.h" 37 #include "llvm/Target/TargetMachine.h" 38 #include "llvm/Target/TargetSubtargetInfo.h" 39 #include <cassert> 40 #include <cstdint> 41 #include <iterator> 42 43 using namespace llvm; 44 45 #define GET_INSTRINFO_CTOR_DTOR 46 #define GET_INSTRMAP_INFO 47 #include "SystemZGenInstrInfo.inc" 48 49 // Return a mask with Count low bits set. 50 static uint64_t allOnes(unsigned int Count) { 51 return Count == 0 ? 0 : (uint64_t(1) << (Count - 1) << 1) - 1; 52 } 53 54 // Reg should be a 32-bit GPR. Return true if it is a high register rather 55 // than a low register. 56 static bool isHighReg(unsigned int Reg) { 57 if (SystemZ::GRH32BitRegClass.contains(Reg)) 58 return true; 59 assert(SystemZ::GR32BitRegClass.contains(Reg) && "Invalid GRX32"); 60 return false; 61 } 62 63 // Pin the vtable to this file. 64 void SystemZInstrInfo::anchor() {} 65 66 SystemZInstrInfo::SystemZInstrInfo(SystemZSubtarget &sti) 67 : SystemZGenInstrInfo(SystemZ::ADJCALLSTACKDOWN, SystemZ::ADJCALLSTACKUP), 68 RI(), STI(sti) { 69 } 70 71 // MI is a 128-bit load or store. Split it into two 64-bit loads or stores, 72 // each having the opcode given by NewOpcode. 73 void SystemZInstrInfo::splitMove(MachineBasicBlock::iterator MI, 74 unsigned NewOpcode) const { 75 MachineBasicBlock *MBB = MI->getParent(); 76 MachineFunction &MF = *MBB->getParent(); 77 78 // Get two load or store instructions. Use the original instruction for one 79 // of them (arbitrarily the second here) and create a clone for the other. 80 MachineInstr *EarlierMI = MF.CloneMachineInstr(&*MI); 81 MBB->insert(MI, EarlierMI); 82 83 // Set up the two 64-bit registers and remember super reg and its flags. 84 MachineOperand &HighRegOp = EarlierMI->getOperand(0); 85 MachineOperand &LowRegOp = MI->getOperand(0); 86 unsigned Reg128 = LowRegOp.getReg(); 87 unsigned Reg128Killed = getKillRegState(LowRegOp.isKill()); 88 unsigned Reg128Undef = getUndefRegState(LowRegOp.isUndef()); 89 HighRegOp.setReg(RI.getSubReg(HighRegOp.getReg(), SystemZ::subreg_h64)); 90 LowRegOp.setReg(RI.getSubReg(LowRegOp.getReg(), SystemZ::subreg_l64)); 91 92 if (MI->mayStore()) { 93 // Add implicit uses of the super register in case one of the subregs is 94 // undefined. We could track liveness and skip storing an undefined 95 // subreg, but this is hopefully rare (discovered with llvm-stress). 96 // If Reg128 was killed, set kill flag on MI. 97 unsigned Reg128UndefImpl = (Reg128Undef | RegState::Implicit); 98 MachineInstrBuilder(MF, EarlierMI).addReg(Reg128, Reg128UndefImpl); 99 MachineInstrBuilder(MF, MI).addReg(Reg128, (Reg128UndefImpl | Reg128Killed)); 100 } 101 102 // The address in the first (high) instruction is already correct. 103 // Adjust the offset in the second (low) instruction. 104 MachineOperand &HighOffsetOp = EarlierMI->getOperand(2); 105 MachineOperand &LowOffsetOp = MI->getOperand(2); 106 LowOffsetOp.setImm(LowOffsetOp.getImm() + 8); 107 108 // Clear the kill flags for the base and index registers in the first 109 // instruction. 110 EarlierMI->getOperand(1).setIsKill(false); 111 EarlierMI->getOperand(3).setIsKill(false); 112 113 // Set the opcodes. 114 unsigned HighOpcode = getOpcodeForOffset(NewOpcode, HighOffsetOp.getImm()); 115 unsigned LowOpcode = getOpcodeForOffset(NewOpcode, LowOffsetOp.getImm()); 116 assert(HighOpcode && LowOpcode && "Both offsets should be in range"); 117 118 EarlierMI->setDesc(get(HighOpcode)); 119 MI->setDesc(get(LowOpcode)); 120 } 121 122 // Split ADJDYNALLOC instruction MI. 123 void SystemZInstrInfo::splitAdjDynAlloc(MachineBasicBlock::iterator MI) const { 124 MachineBasicBlock *MBB = MI->getParent(); 125 MachineFunction &MF = *MBB->getParent(); 126 MachineFrameInfo &MFFrame = MF.getFrameInfo(); 127 MachineOperand &OffsetMO = MI->getOperand(2); 128 129 uint64_t Offset = (MFFrame.getMaxCallFrameSize() + 130 SystemZMC::CallFrameSize + 131 OffsetMO.getImm()); 132 unsigned NewOpcode = getOpcodeForOffset(SystemZ::LA, Offset); 133 assert(NewOpcode && "No support for huge argument lists yet"); 134 MI->setDesc(get(NewOpcode)); 135 OffsetMO.setImm(Offset); 136 } 137 138 // MI is an RI-style pseudo instruction. Replace it with LowOpcode 139 // if the first operand is a low GR32 and HighOpcode if the first operand 140 // is a high GR32. ConvertHigh is true if LowOpcode takes a signed operand 141 // and HighOpcode takes an unsigned 32-bit operand. In those cases, 142 // MI has the same kind of operand as LowOpcode, so needs to be converted 143 // if HighOpcode is used. 144 void SystemZInstrInfo::expandRIPseudo(MachineInstr &MI, unsigned LowOpcode, 145 unsigned HighOpcode, 146 bool ConvertHigh) const { 147 unsigned Reg = MI.getOperand(0).getReg(); 148 bool IsHigh = isHighReg(Reg); 149 MI.setDesc(get(IsHigh ? HighOpcode : LowOpcode)); 150 if (IsHigh && ConvertHigh) 151 MI.getOperand(1).setImm(uint32_t(MI.getOperand(1).getImm())); 152 } 153 154 // MI is a three-operand RIE-style pseudo instruction. Replace it with 155 // LowOpcodeK if the registers are both low GR32s, otherwise use a move 156 // followed by HighOpcode or LowOpcode, depending on whether the target 157 // is a high or low GR32. 158 void SystemZInstrInfo::expandRIEPseudo(MachineInstr &MI, unsigned LowOpcode, 159 unsigned LowOpcodeK, 160 unsigned HighOpcode) const { 161 unsigned DestReg = MI.getOperand(0).getReg(); 162 unsigned SrcReg = MI.getOperand(1).getReg(); 163 bool DestIsHigh = isHighReg(DestReg); 164 bool SrcIsHigh = isHighReg(SrcReg); 165 if (!DestIsHigh && !SrcIsHigh) 166 MI.setDesc(get(LowOpcodeK)); 167 else { 168 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), DestReg, SrcReg, 169 SystemZ::LR, 32, MI.getOperand(1).isKill(), 170 MI.getOperand(1).isUndef()); 171 MI.setDesc(get(DestIsHigh ? HighOpcode : LowOpcode)); 172 MI.getOperand(1).setReg(DestReg); 173 MI.tieOperands(0, 1); 174 } 175 } 176 177 // MI is an RXY-style pseudo instruction. Replace it with LowOpcode 178 // if the first operand is a low GR32 and HighOpcode if the first operand 179 // is a high GR32. 180 void SystemZInstrInfo::expandRXYPseudo(MachineInstr &MI, unsigned LowOpcode, 181 unsigned HighOpcode) const { 182 unsigned Reg = MI.getOperand(0).getReg(); 183 unsigned Opcode = getOpcodeForOffset(isHighReg(Reg) ? HighOpcode : LowOpcode, 184 MI.getOperand(2).getImm()); 185 MI.setDesc(get(Opcode)); 186 } 187 188 // MI is a load-on-condition pseudo instruction with a single register 189 // (source or destination) operand. Replace it with LowOpcode if the 190 // register is a low GR32 and HighOpcode if the register is a high GR32. 191 void SystemZInstrInfo::expandLOCPseudo(MachineInstr &MI, unsigned LowOpcode, 192 unsigned HighOpcode) const { 193 unsigned Reg = MI.getOperand(0).getReg(); 194 unsigned Opcode = isHighReg(Reg) ? HighOpcode : LowOpcode; 195 MI.setDesc(get(Opcode)); 196 } 197 198 // MI is a load-register-on-condition pseudo instruction. Replace it with 199 // LowOpcode if source and destination are both low GR32s and HighOpcode if 200 // source and destination are both high GR32s. 201 void SystemZInstrInfo::expandLOCRPseudo(MachineInstr &MI, unsigned LowOpcode, 202 unsigned HighOpcode) const { 203 unsigned DestReg = MI.getOperand(0).getReg(); 204 unsigned SrcReg = MI.getOperand(2).getReg(); 205 bool DestIsHigh = isHighReg(DestReg); 206 bool SrcIsHigh = isHighReg(SrcReg); 207 208 if (!DestIsHigh && !SrcIsHigh) 209 MI.setDesc(get(LowOpcode)); 210 else if (DestIsHigh && SrcIsHigh) 211 MI.setDesc(get(HighOpcode)); 212 213 // If we were unable to implement the pseudo with a single instruction, we 214 // need to convert it back into a branch sequence. This cannot be done here 215 // since the caller of expandPostRAPseudo does not handle changes to the CFG 216 // correctly. This change is defered to the SystemZExpandPseudo pass. 217 } 218 219 // MI is an RR-style pseudo instruction that zero-extends the low Size bits 220 // of one GRX32 into another. Replace it with LowOpcode if both operands 221 // are low registers, otherwise use RISB[LH]G. 222 void SystemZInstrInfo::expandZExtPseudo(MachineInstr &MI, unsigned LowOpcode, 223 unsigned Size) const { 224 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), 225 MI.getOperand(0).getReg(), MI.getOperand(1).getReg(), LowOpcode, 226 Size, MI.getOperand(1).isKill(), MI.getOperand(1).isUndef()); 227 MI.eraseFromParent(); 228 } 229 230 void SystemZInstrInfo::expandLoadStackGuard(MachineInstr *MI) const { 231 MachineBasicBlock *MBB = MI->getParent(); 232 MachineFunction &MF = *MBB->getParent(); 233 const unsigned Reg = MI->getOperand(0).getReg(); 234 235 // Conveniently, all 4 instructions are cloned from LOAD_STACK_GUARD, 236 // so they already have operand 0 set to reg. 237 238 // ear <reg>, %a0 239 MachineInstr *Ear1MI = MF.CloneMachineInstr(MI); 240 MBB->insert(MI, Ear1MI); 241 Ear1MI->setDesc(get(SystemZ::EAR)); 242 MachineInstrBuilder(MF, Ear1MI).addReg(SystemZ::A0); 243 244 // sllg <reg>, <reg>, 32 245 MachineInstr *SllgMI = MF.CloneMachineInstr(MI); 246 MBB->insert(MI, SllgMI); 247 SllgMI->setDesc(get(SystemZ::SLLG)); 248 MachineInstrBuilder(MF, SllgMI).addReg(Reg).addReg(0).addImm(32); 249 250 // ear <reg>, %a1 251 MachineInstr *Ear2MI = MF.CloneMachineInstr(MI); 252 MBB->insert(MI, Ear2MI); 253 Ear2MI->setDesc(get(SystemZ::EAR)); 254 MachineInstrBuilder(MF, Ear2MI).addReg(SystemZ::A1); 255 256 // lg <reg>, 40(<reg>) 257 MI->setDesc(get(SystemZ::LG)); 258 MachineInstrBuilder(MF, MI).addReg(Reg).addImm(40).addReg(0); 259 } 260 261 // Emit a zero-extending move from 32-bit GPR SrcReg to 32-bit GPR 262 // DestReg before MBBI in MBB. Use LowLowOpcode when both DestReg and SrcReg 263 // are low registers, otherwise use RISB[LH]G. Size is the number of bits 264 // taken from the low end of SrcReg (8 for LLCR, 16 for LLHR and 32 for LR). 265 // KillSrc is true if this move is the last use of SrcReg. 266 void SystemZInstrInfo::emitGRX32Move(MachineBasicBlock &MBB, 267 MachineBasicBlock::iterator MBBI, 268 const DebugLoc &DL, unsigned DestReg, 269 unsigned SrcReg, unsigned LowLowOpcode, 270 unsigned Size, bool KillSrc, 271 bool UndefSrc) const { 272 unsigned Opcode; 273 bool DestIsHigh = isHighReg(DestReg); 274 bool SrcIsHigh = isHighReg(SrcReg); 275 if (DestIsHigh && SrcIsHigh) 276 Opcode = SystemZ::RISBHH; 277 else if (DestIsHigh && !SrcIsHigh) 278 Opcode = SystemZ::RISBHL; 279 else if (!DestIsHigh && SrcIsHigh) 280 Opcode = SystemZ::RISBLH; 281 else { 282 BuildMI(MBB, MBBI, DL, get(LowLowOpcode), DestReg) 283 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)); 284 return; 285 } 286 unsigned Rotate = (DestIsHigh != SrcIsHigh ? 32 : 0); 287 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 288 .addReg(DestReg, RegState::Undef) 289 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)) 290 .addImm(32 - Size).addImm(128 + 31).addImm(Rotate); 291 } 292 293 MachineInstr *SystemZInstrInfo::commuteInstructionImpl(MachineInstr &MI, 294 bool NewMI, 295 unsigned OpIdx1, 296 unsigned OpIdx2) const { 297 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & { 298 if (NewMI) 299 return *MI.getParent()->getParent()->CloneMachineInstr(&MI); 300 return MI; 301 }; 302 303 switch (MI.getOpcode()) { 304 case SystemZ::LOCRMux: 305 case SystemZ::LOCFHR: 306 case SystemZ::LOCR: 307 case SystemZ::LOCGR: { 308 auto &WorkingMI = cloneIfNew(MI); 309 // Invert condition. 310 unsigned CCValid = WorkingMI.getOperand(3).getImm(); 311 unsigned CCMask = WorkingMI.getOperand(4).getImm(); 312 WorkingMI.getOperand(4).setImm(CCMask ^ CCValid); 313 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 314 OpIdx1, OpIdx2); 315 } 316 default: 317 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 318 } 319 } 320 321 // If MI is a simple load or store for a frame object, return the register 322 // it loads or stores and set FrameIndex to the index of the frame object. 323 // Return 0 otherwise. 324 // 325 // Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 326 static int isSimpleMove(const MachineInstr &MI, int &FrameIndex, 327 unsigned Flag) { 328 const MCInstrDesc &MCID = MI.getDesc(); 329 if ((MCID.TSFlags & Flag) && MI.getOperand(1).isFI() && 330 MI.getOperand(2).getImm() == 0 && MI.getOperand(3).getReg() == 0) { 331 FrameIndex = MI.getOperand(1).getIndex(); 332 return MI.getOperand(0).getReg(); 333 } 334 return 0; 335 } 336 337 unsigned SystemZInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 338 int &FrameIndex) const { 339 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXLoad); 340 } 341 342 unsigned SystemZInstrInfo::isStoreToStackSlot(const MachineInstr &MI, 343 int &FrameIndex) const { 344 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXStore); 345 } 346 347 bool SystemZInstrInfo::isStackSlotCopy(const MachineInstr &MI, 348 int &DestFrameIndex, 349 int &SrcFrameIndex) const { 350 // Check for MVC 0(Length,FI1),0(FI2) 351 const MachineFrameInfo &MFI = MI.getParent()->getParent()->getFrameInfo(); 352 if (MI.getOpcode() != SystemZ::MVC || !MI.getOperand(0).isFI() || 353 MI.getOperand(1).getImm() != 0 || !MI.getOperand(3).isFI() || 354 MI.getOperand(4).getImm() != 0) 355 return false; 356 357 // Check that Length covers the full slots. 358 int64_t Length = MI.getOperand(2).getImm(); 359 unsigned FI1 = MI.getOperand(0).getIndex(); 360 unsigned FI2 = MI.getOperand(3).getIndex(); 361 if (MFI.getObjectSize(FI1) != Length || 362 MFI.getObjectSize(FI2) != Length) 363 return false; 364 365 DestFrameIndex = FI1; 366 SrcFrameIndex = FI2; 367 return true; 368 } 369 370 bool SystemZInstrInfo::analyzeBranch(MachineBasicBlock &MBB, 371 MachineBasicBlock *&TBB, 372 MachineBasicBlock *&FBB, 373 SmallVectorImpl<MachineOperand> &Cond, 374 bool AllowModify) const { 375 // Most of the code and comments here are boilerplate. 376 377 // Start from the bottom of the block and work up, examining the 378 // terminator instructions. 379 MachineBasicBlock::iterator I = MBB.end(); 380 while (I != MBB.begin()) { 381 --I; 382 if (I->isDebugValue()) 383 continue; 384 385 // Working from the bottom, when we see a non-terminator instruction, we're 386 // done. 387 if (!isUnpredicatedTerminator(*I)) 388 break; 389 390 // A terminator that isn't a branch can't easily be handled by this 391 // analysis. 392 if (!I->isBranch()) 393 return true; 394 395 // Can't handle indirect branches. 396 SystemZII::Branch Branch(getBranchInfo(*I)); 397 if (!Branch.Target->isMBB()) 398 return true; 399 400 // Punt on compound branches. 401 if (Branch.Type != SystemZII::BranchNormal) 402 return true; 403 404 if (Branch.CCMask == SystemZ::CCMASK_ANY) { 405 // Handle unconditional branches. 406 if (!AllowModify) { 407 TBB = Branch.Target->getMBB(); 408 continue; 409 } 410 411 // If the block has any instructions after a JMP, delete them. 412 while (std::next(I) != MBB.end()) 413 std::next(I)->eraseFromParent(); 414 415 Cond.clear(); 416 FBB = nullptr; 417 418 // Delete the JMP if it's equivalent to a fall-through. 419 if (MBB.isLayoutSuccessor(Branch.Target->getMBB())) { 420 TBB = nullptr; 421 I->eraseFromParent(); 422 I = MBB.end(); 423 continue; 424 } 425 426 // TBB is used to indicate the unconditinal destination. 427 TBB = Branch.Target->getMBB(); 428 continue; 429 } 430 431 // Working from the bottom, handle the first conditional branch. 432 if (Cond.empty()) { 433 // FIXME: add X86-style branch swap 434 FBB = TBB; 435 TBB = Branch.Target->getMBB(); 436 Cond.push_back(MachineOperand::CreateImm(Branch.CCValid)); 437 Cond.push_back(MachineOperand::CreateImm(Branch.CCMask)); 438 continue; 439 } 440 441 // Handle subsequent conditional branches. 442 assert(Cond.size() == 2 && TBB && "Should have seen a conditional branch"); 443 444 // Only handle the case where all conditional branches branch to the same 445 // destination. 446 if (TBB != Branch.Target->getMBB()) 447 return true; 448 449 // If the conditions are the same, we can leave them alone. 450 unsigned OldCCValid = Cond[0].getImm(); 451 unsigned OldCCMask = Cond[1].getImm(); 452 if (OldCCValid == Branch.CCValid && OldCCMask == Branch.CCMask) 453 continue; 454 455 // FIXME: Try combining conditions like X86 does. Should be easy on Z! 456 return false; 457 } 458 459 return false; 460 } 461 462 unsigned SystemZInstrInfo::removeBranch(MachineBasicBlock &MBB, 463 int *BytesRemoved) const { 464 assert(!BytesRemoved && "code size not handled"); 465 466 // Most of the code and comments here are boilerplate. 467 MachineBasicBlock::iterator I = MBB.end(); 468 unsigned Count = 0; 469 470 while (I != MBB.begin()) { 471 --I; 472 if (I->isDebugValue()) 473 continue; 474 if (!I->isBranch()) 475 break; 476 if (!getBranchInfo(*I).Target->isMBB()) 477 break; 478 // Remove the branch. 479 I->eraseFromParent(); 480 I = MBB.end(); 481 ++Count; 482 } 483 484 return Count; 485 } 486 487 bool SystemZInstrInfo:: 488 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 489 assert(Cond.size() == 2 && "Invalid condition"); 490 Cond[1].setImm(Cond[1].getImm() ^ Cond[0].getImm()); 491 return false; 492 } 493 494 unsigned SystemZInstrInfo::insertBranch(MachineBasicBlock &MBB, 495 MachineBasicBlock *TBB, 496 MachineBasicBlock *FBB, 497 ArrayRef<MachineOperand> Cond, 498 const DebugLoc &DL, 499 int *BytesAdded) const { 500 // In this function we output 32-bit branches, which should always 501 // have enough range. They can be shortened and relaxed by later code 502 // in the pipeline, if desired. 503 504 // Shouldn't be a fall through. 505 assert(TBB && "insertBranch must not be told to insert a fallthrough"); 506 assert((Cond.size() == 2 || Cond.size() == 0) && 507 "SystemZ branch conditions have one component!"); 508 assert(!BytesAdded && "code size not handled"); 509 510 if (Cond.empty()) { 511 // Unconditional branch? 512 assert(!FBB && "Unconditional branch with multiple successors!"); 513 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(TBB); 514 return 1; 515 } 516 517 // Conditional branch. 518 unsigned Count = 0; 519 unsigned CCValid = Cond[0].getImm(); 520 unsigned CCMask = Cond[1].getImm(); 521 BuildMI(&MBB, DL, get(SystemZ::BRC)) 522 .addImm(CCValid).addImm(CCMask).addMBB(TBB); 523 ++Count; 524 525 if (FBB) { 526 // Two-way Conditional branch. Insert the second branch. 527 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(FBB); 528 ++Count; 529 } 530 return Count; 531 } 532 533 bool SystemZInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg, 534 unsigned &SrcReg2, int &Mask, 535 int &Value) const { 536 assert(MI.isCompare() && "Caller should have checked for a comparison"); 537 538 if (MI.getNumExplicitOperands() == 2 && MI.getOperand(0).isReg() && 539 MI.getOperand(1).isImm()) { 540 SrcReg = MI.getOperand(0).getReg(); 541 SrcReg2 = 0; 542 Value = MI.getOperand(1).getImm(); 543 Mask = ~0; 544 return true; 545 } 546 547 return false; 548 } 549 550 // If Reg is a virtual register, return its definition, otherwise return null. 551 static MachineInstr *getDef(unsigned Reg, 552 const MachineRegisterInfo *MRI) { 553 if (TargetRegisterInfo::isPhysicalRegister(Reg)) 554 return nullptr; 555 return MRI->getUniqueVRegDef(Reg); 556 } 557 558 // Return true if MI is a shift of type Opcode by Imm bits. 559 static bool isShift(MachineInstr *MI, unsigned Opcode, int64_t Imm) { 560 return (MI->getOpcode() == Opcode && 561 !MI->getOperand(2).getReg() && 562 MI->getOperand(3).getImm() == Imm); 563 } 564 565 // If the destination of MI has no uses, delete it as dead. 566 static void eraseIfDead(MachineInstr *MI, const MachineRegisterInfo *MRI) { 567 if (MRI->use_nodbg_empty(MI->getOperand(0).getReg())) 568 MI->eraseFromParent(); 569 } 570 571 // Compare compares SrcReg against zero. Check whether SrcReg contains 572 // the result of an IPM sequence whose input CC survives until Compare, 573 // and whether Compare is therefore redundant. Delete it and return 574 // true if so. 575 static bool removeIPMBasedCompare(MachineInstr &Compare, unsigned SrcReg, 576 const MachineRegisterInfo *MRI, 577 const TargetRegisterInfo *TRI) { 578 MachineInstr *LGFR = nullptr; 579 MachineInstr *RLL = getDef(SrcReg, MRI); 580 if (RLL && RLL->getOpcode() == SystemZ::LGFR) { 581 LGFR = RLL; 582 RLL = getDef(LGFR->getOperand(1).getReg(), MRI); 583 } 584 if (!RLL || !isShift(RLL, SystemZ::RLL, 31)) 585 return false; 586 587 MachineInstr *SRL = getDef(RLL->getOperand(1).getReg(), MRI); 588 if (!SRL || !isShift(SRL, SystemZ::SRL, SystemZ::IPM_CC)) 589 return false; 590 591 MachineInstr *IPM = getDef(SRL->getOperand(1).getReg(), MRI); 592 if (!IPM || IPM->getOpcode() != SystemZ::IPM) 593 return false; 594 595 // Check that there are no assignments to CC between the IPM and Compare, 596 if (IPM->getParent() != Compare.getParent()) 597 return false; 598 MachineBasicBlock::iterator MBBI = IPM, MBBE = Compare.getIterator(); 599 for (++MBBI; MBBI != MBBE; ++MBBI) { 600 MachineInstr &MI = *MBBI; 601 if (MI.modifiesRegister(SystemZ::CC, TRI)) 602 return false; 603 } 604 605 Compare.eraseFromParent(); 606 if (LGFR) 607 eraseIfDead(LGFR, MRI); 608 eraseIfDead(RLL, MRI); 609 eraseIfDead(SRL, MRI); 610 eraseIfDead(IPM, MRI); 611 612 return true; 613 } 614 615 bool SystemZInstrInfo::optimizeCompareInstr( 616 MachineInstr &Compare, unsigned SrcReg, unsigned SrcReg2, int Mask, 617 int Value, const MachineRegisterInfo *MRI) const { 618 assert(!SrcReg2 && "Only optimizing constant comparisons so far"); 619 bool IsLogical = (Compare.getDesc().TSFlags & SystemZII::IsLogical) != 0; 620 return Value == 0 && !IsLogical && 621 removeIPMBasedCompare(Compare, SrcReg, MRI, &RI); 622 } 623 624 bool SystemZInstrInfo::canInsertSelect(const MachineBasicBlock &MBB, 625 ArrayRef<MachineOperand> Pred, 626 unsigned TrueReg, unsigned FalseReg, 627 int &CondCycles, int &TrueCycles, 628 int &FalseCycles) const { 629 // Not all subtargets have LOCR instructions. 630 if (!STI.hasLoadStoreOnCond()) 631 return false; 632 if (Pred.size() != 2) 633 return false; 634 635 // Check register classes. 636 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 637 const TargetRegisterClass *RC = 638 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 639 if (!RC) 640 return false; 641 642 // We have LOCR instructions for 32 and 64 bit general purpose registers. 643 if ((STI.hasLoadStoreOnCond2() && 644 SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) || 645 SystemZ::GR32BitRegClass.hasSubClassEq(RC) || 646 SystemZ::GR64BitRegClass.hasSubClassEq(RC)) { 647 CondCycles = 2; 648 TrueCycles = 2; 649 FalseCycles = 2; 650 return true; 651 } 652 653 // Can't do anything else. 654 return false; 655 } 656 657 void SystemZInstrInfo::insertSelect(MachineBasicBlock &MBB, 658 MachineBasicBlock::iterator I, 659 const DebugLoc &DL, unsigned DstReg, 660 ArrayRef<MachineOperand> Pred, 661 unsigned TrueReg, 662 unsigned FalseReg) const { 663 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 664 const TargetRegisterClass *RC = MRI.getRegClass(DstReg); 665 666 assert(Pred.size() == 2 && "Invalid condition"); 667 unsigned CCValid = Pred[0].getImm(); 668 unsigned CCMask = Pred[1].getImm(); 669 670 unsigned Opc; 671 if (SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) { 672 if (STI.hasLoadStoreOnCond2()) 673 Opc = SystemZ::LOCRMux; 674 else { 675 Opc = SystemZ::LOCR; 676 MRI.constrainRegClass(DstReg, &SystemZ::GR32BitRegClass); 677 } 678 } else if (SystemZ::GR64BitRegClass.hasSubClassEq(RC)) 679 Opc = SystemZ::LOCGR; 680 else 681 llvm_unreachable("Invalid register class"); 682 683 BuildMI(MBB, I, DL, get(Opc), DstReg) 684 .addReg(FalseReg).addReg(TrueReg) 685 .addImm(CCValid).addImm(CCMask); 686 } 687 688 bool SystemZInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, 689 unsigned Reg, 690 MachineRegisterInfo *MRI) const { 691 unsigned DefOpc = DefMI.getOpcode(); 692 if (DefOpc != SystemZ::LHIMux && DefOpc != SystemZ::LHI && 693 DefOpc != SystemZ::LGHI) 694 return false; 695 if (DefMI.getOperand(0).getReg() != Reg) 696 return false; 697 int32_t ImmVal = (int32_t)DefMI.getOperand(1).getImm(); 698 699 unsigned UseOpc = UseMI.getOpcode(); 700 unsigned NewUseOpc; 701 unsigned UseIdx; 702 int CommuteIdx = -1; 703 switch (UseOpc) { 704 case SystemZ::LOCRMux: 705 if (!STI.hasLoadStoreOnCond2()) 706 return false; 707 NewUseOpc = SystemZ::LOCHIMux; 708 if (UseMI.getOperand(2).getReg() == Reg) 709 UseIdx = 2; 710 else if (UseMI.getOperand(1).getReg() == Reg) 711 UseIdx = 2, CommuteIdx = 1; 712 else 713 return false; 714 break; 715 case SystemZ::LOCGR: 716 if (!STI.hasLoadStoreOnCond2()) 717 return false; 718 NewUseOpc = SystemZ::LOCGHI; 719 if (UseMI.getOperand(2).getReg() == Reg) 720 UseIdx = 2; 721 else if (UseMI.getOperand(1).getReg() == Reg) 722 UseIdx = 2, CommuteIdx = 1; 723 else 724 return false; 725 break; 726 default: 727 return false; 728 } 729 730 if (CommuteIdx != -1) 731 if (!commuteInstruction(UseMI, false, CommuteIdx, UseIdx)) 732 return false; 733 734 bool DeleteDef = MRI->hasOneNonDBGUse(Reg); 735 UseMI.setDesc(get(NewUseOpc)); 736 UseMI.getOperand(UseIdx).ChangeToImmediate(ImmVal); 737 if (DeleteDef) 738 DefMI.eraseFromParent(); 739 740 return true; 741 } 742 743 bool SystemZInstrInfo::isPredicable(const MachineInstr &MI) const { 744 unsigned Opcode = MI.getOpcode(); 745 if (Opcode == SystemZ::Return || 746 Opcode == SystemZ::Trap || 747 Opcode == SystemZ::CallJG || 748 Opcode == SystemZ::CallBR) 749 return true; 750 return false; 751 } 752 753 bool SystemZInstrInfo:: 754 isProfitableToIfCvt(MachineBasicBlock &MBB, 755 unsigned NumCycles, unsigned ExtraPredCycles, 756 BranchProbability Probability) const { 757 // Avoid using conditional returns at the end of a loop (since then 758 // we'd need to emit an unconditional branch to the beginning anyway, 759 // making the loop body longer). This doesn't apply for low-probability 760 // loops (eg. compare-and-swap retry), so just decide based on branch 761 // probability instead of looping structure. 762 // However, since Compare and Trap instructions cost the same as a regular 763 // Compare instruction, we should allow the if conversion to convert this 764 // into a Conditional Compare regardless of the branch probability. 765 if (MBB.getLastNonDebugInstr()->getOpcode() != SystemZ::Trap && 766 MBB.succ_empty() && Probability < BranchProbability(1, 8)) 767 return false; 768 // For now only convert single instructions. 769 return NumCycles == 1; 770 } 771 772 bool SystemZInstrInfo:: 773 isProfitableToIfCvt(MachineBasicBlock &TMBB, 774 unsigned NumCyclesT, unsigned ExtraPredCyclesT, 775 MachineBasicBlock &FMBB, 776 unsigned NumCyclesF, unsigned ExtraPredCyclesF, 777 BranchProbability Probability) const { 778 // For now avoid converting mutually-exclusive cases. 779 return false; 780 } 781 782 bool SystemZInstrInfo:: 783 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles, 784 BranchProbability Probability) const { 785 // For now only duplicate single instructions. 786 return NumCycles == 1; 787 } 788 789 bool SystemZInstrInfo::PredicateInstruction( 790 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const { 791 assert(Pred.size() == 2 && "Invalid condition"); 792 unsigned CCValid = Pred[0].getImm(); 793 unsigned CCMask = Pred[1].getImm(); 794 assert(CCMask > 0 && CCMask < 15 && "Invalid predicate"); 795 unsigned Opcode = MI.getOpcode(); 796 if (Opcode == SystemZ::Trap) { 797 MI.setDesc(get(SystemZ::CondTrap)); 798 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 799 .addImm(CCValid).addImm(CCMask) 800 .addReg(SystemZ::CC, RegState::Implicit); 801 return true; 802 } 803 if (Opcode == SystemZ::Return) { 804 MI.setDesc(get(SystemZ::CondReturn)); 805 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 806 .addImm(CCValid).addImm(CCMask) 807 .addReg(SystemZ::CC, RegState::Implicit); 808 return true; 809 } 810 if (Opcode == SystemZ::CallJG) { 811 MachineOperand FirstOp = MI.getOperand(0); 812 const uint32_t *RegMask = MI.getOperand(1).getRegMask(); 813 MI.RemoveOperand(1); 814 MI.RemoveOperand(0); 815 MI.setDesc(get(SystemZ::CallBRCL)); 816 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 817 .addImm(CCValid) 818 .addImm(CCMask) 819 .add(FirstOp) 820 .addRegMask(RegMask) 821 .addReg(SystemZ::CC, RegState::Implicit); 822 return true; 823 } 824 if (Opcode == SystemZ::CallBR) { 825 const uint32_t *RegMask = MI.getOperand(0).getRegMask(); 826 MI.RemoveOperand(0); 827 MI.setDesc(get(SystemZ::CallBCR)); 828 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 829 .addImm(CCValid).addImm(CCMask) 830 .addRegMask(RegMask) 831 .addReg(SystemZ::CC, RegState::Implicit); 832 return true; 833 } 834 return false; 835 } 836 837 void SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 838 MachineBasicBlock::iterator MBBI, 839 const DebugLoc &DL, unsigned DestReg, 840 unsigned SrcReg, bool KillSrc) const { 841 // Split 128-bit GPR moves into two 64-bit moves. This handles ADDR128 too. 842 if (SystemZ::GR128BitRegClass.contains(DestReg, SrcReg)) { 843 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_h64), 844 RI.getSubReg(SrcReg, SystemZ::subreg_h64), KillSrc); 845 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_l64), 846 RI.getSubReg(SrcReg, SystemZ::subreg_l64), KillSrc); 847 return; 848 } 849 850 if (SystemZ::GRX32BitRegClass.contains(DestReg, SrcReg)) { 851 emitGRX32Move(MBB, MBBI, DL, DestReg, SrcReg, SystemZ::LR, 32, KillSrc, 852 false); 853 return; 854 } 855 856 // Everything else needs only one instruction. 857 unsigned Opcode; 858 if (SystemZ::GR64BitRegClass.contains(DestReg, SrcReg)) 859 Opcode = SystemZ::LGR; 860 else if (SystemZ::FP32BitRegClass.contains(DestReg, SrcReg)) 861 // For z13 we prefer LDR over LER to avoid partial register dependencies. 862 Opcode = STI.hasVector() ? SystemZ::LDR32 : SystemZ::LER; 863 else if (SystemZ::FP64BitRegClass.contains(DestReg, SrcReg)) 864 Opcode = SystemZ::LDR; 865 else if (SystemZ::FP128BitRegClass.contains(DestReg, SrcReg)) 866 Opcode = SystemZ::LXR; 867 else if (SystemZ::VR32BitRegClass.contains(DestReg, SrcReg)) 868 Opcode = SystemZ::VLR32; 869 else if (SystemZ::VR64BitRegClass.contains(DestReg, SrcReg)) 870 Opcode = SystemZ::VLR64; 871 else if (SystemZ::VR128BitRegClass.contains(DestReg, SrcReg)) 872 Opcode = SystemZ::VLR; 873 else if (SystemZ::AR32BitRegClass.contains(DestReg, SrcReg)) 874 Opcode = SystemZ::CPYA; 875 else if (SystemZ::AR32BitRegClass.contains(DestReg) && 876 SystemZ::GR32BitRegClass.contains(SrcReg)) 877 Opcode = SystemZ::SAR; 878 else if (SystemZ::GR32BitRegClass.contains(DestReg) && 879 SystemZ::AR32BitRegClass.contains(SrcReg)) 880 Opcode = SystemZ::EAR; 881 else 882 llvm_unreachable("Impossible reg-to-reg copy"); 883 884 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 885 .addReg(SrcReg, getKillRegState(KillSrc)); 886 } 887 888 void SystemZInstrInfo::storeRegToStackSlot( 889 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned SrcReg, 890 bool isKill, int FrameIdx, const TargetRegisterClass *RC, 891 const TargetRegisterInfo *TRI) const { 892 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 893 894 // Callers may expect a single instruction, so keep 128-bit moves 895 // together for now and lower them after register allocation. 896 unsigned LoadOpcode, StoreOpcode; 897 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 898 addFrameReference(BuildMI(MBB, MBBI, DL, get(StoreOpcode)) 899 .addReg(SrcReg, getKillRegState(isKill)), 900 FrameIdx); 901 } 902 903 void SystemZInstrInfo::loadRegFromStackSlot( 904 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned DestReg, 905 int FrameIdx, const TargetRegisterClass *RC, 906 const TargetRegisterInfo *TRI) const { 907 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 908 909 // Callers may expect a single instruction, so keep 128-bit moves 910 // together for now and lower them after register allocation. 911 unsigned LoadOpcode, StoreOpcode; 912 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 913 addFrameReference(BuildMI(MBB, MBBI, DL, get(LoadOpcode), DestReg), 914 FrameIdx); 915 } 916 917 // Return true if MI is a simple load or store with a 12-bit displacement 918 // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 919 static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) { 920 const MCInstrDesc &MCID = MI->getDesc(); 921 return ((MCID.TSFlags & Flag) && 922 isUInt<12>(MI->getOperand(2).getImm()) && 923 MI->getOperand(3).getReg() == 0); 924 } 925 926 namespace { 927 928 struct LogicOp { 929 LogicOp() = default; 930 LogicOp(unsigned regSize, unsigned immLSB, unsigned immSize) 931 : RegSize(regSize), ImmLSB(immLSB), ImmSize(immSize) {} 932 933 explicit operator bool() const { return RegSize; } 934 935 unsigned RegSize = 0; 936 unsigned ImmLSB = 0; 937 unsigned ImmSize = 0; 938 }; 939 940 } // end anonymous namespace 941 942 static LogicOp interpretAndImmediate(unsigned Opcode) { 943 switch (Opcode) { 944 case SystemZ::NILMux: return LogicOp(32, 0, 16); 945 case SystemZ::NIHMux: return LogicOp(32, 16, 16); 946 case SystemZ::NILL64: return LogicOp(64, 0, 16); 947 case SystemZ::NILH64: return LogicOp(64, 16, 16); 948 case SystemZ::NIHL64: return LogicOp(64, 32, 16); 949 case SystemZ::NIHH64: return LogicOp(64, 48, 16); 950 case SystemZ::NIFMux: return LogicOp(32, 0, 32); 951 case SystemZ::NILF64: return LogicOp(64, 0, 32); 952 case SystemZ::NIHF64: return LogicOp(64, 32, 32); 953 default: return LogicOp(); 954 } 955 } 956 957 static void transferDeadCC(MachineInstr *OldMI, MachineInstr *NewMI) { 958 if (OldMI->registerDefIsDead(SystemZ::CC)) { 959 MachineOperand *CCDef = NewMI->findRegisterDefOperand(SystemZ::CC); 960 if (CCDef != nullptr) 961 CCDef->setIsDead(true); 962 } 963 } 964 965 // Used to return from convertToThreeAddress after replacing two-address 966 // instruction OldMI with three-address instruction NewMI. 967 static MachineInstr *finishConvertToThreeAddress(MachineInstr *OldMI, 968 MachineInstr *NewMI, 969 LiveVariables *LV) { 970 if (LV) { 971 unsigned NumOps = OldMI->getNumOperands(); 972 for (unsigned I = 1; I < NumOps; ++I) { 973 MachineOperand &Op = OldMI->getOperand(I); 974 if (Op.isReg() && Op.isKill()) 975 LV->replaceKillInstruction(Op.getReg(), *OldMI, *NewMI); 976 } 977 } 978 transferDeadCC(OldMI, NewMI); 979 return NewMI; 980 } 981 982 MachineInstr *SystemZInstrInfo::convertToThreeAddress( 983 MachineFunction::iterator &MFI, MachineInstr &MI, LiveVariables *LV) const { 984 MachineBasicBlock *MBB = MI.getParent(); 985 MachineFunction *MF = MBB->getParent(); 986 MachineRegisterInfo &MRI = MF->getRegInfo(); 987 988 unsigned Opcode = MI.getOpcode(); 989 unsigned NumOps = MI.getNumOperands(); 990 991 // Try to convert something like SLL into SLLK, if supported. 992 // We prefer to keep the two-operand form where possible both 993 // because it tends to be shorter and because some instructions 994 // have memory forms that can be used during spilling. 995 if (STI.hasDistinctOps()) { 996 MachineOperand &Dest = MI.getOperand(0); 997 MachineOperand &Src = MI.getOperand(1); 998 unsigned DestReg = Dest.getReg(); 999 unsigned SrcReg = Src.getReg(); 1000 // AHIMux is only really a three-operand instruction when both operands 1001 // are low registers. Try to constrain both operands to be low if 1002 // possible. 1003 if (Opcode == SystemZ::AHIMux && 1004 TargetRegisterInfo::isVirtualRegister(DestReg) && 1005 TargetRegisterInfo::isVirtualRegister(SrcReg) && 1006 MRI.getRegClass(DestReg)->contains(SystemZ::R1L) && 1007 MRI.getRegClass(SrcReg)->contains(SystemZ::R1L)) { 1008 MRI.constrainRegClass(DestReg, &SystemZ::GR32BitRegClass); 1009 MRI.constrainRegClass(SrcReg, &SystemZ::GR32BitRegClass); 1010 } 1011 int ThreeOperandOpcode = SystemZ::getThreeOperandOpcode(Opcode); 1012 if (ThreeOperandOpcode >= 0) { 1013 // Create three address instruction without adding the implicit 1014 // operands. Those will instead be copied over from the original 1015 // instruction by the loop below. 1016 MachineInstrBuilder MIB( 1017 *MF, MF->CreateMachineInstr(get(ThreeOperandOpcode), MI.getDebugLoc(), 1018 /*NoImplicit=*/true)); 1019 MIB.add(Dest); 1020 // Keep the kill state, but drop the tied flag. 1021 MIB.addReg(Src.getReg(), getKillRegState(Src.isKill()), Src.getSubReg()); 1022 // Keep the remaining operands as-is. 1023 for (unsigned I = 2; I < NumOps; ++I) 1024 MIB.add(MI.getOperand(I)); 1025 MBB->insert(MI, MIB); 1026 return finishConvertToThreeAddress(&MI, MIB, LV); 1027 } 1028 } 1029 1030 // Try to convert an AND into an RISBG-type instruction. 1031 if (LogicOp And = interpretAndImmediate(Opcode)) { 1032 uint64_t Imm = MI.getOperand(2).getImm() << And.ImmLSB; 1033 // AND IMMEDIATE leaves the other bits of the register unchanged. 1034 Imm |= allOnes(And.RegSize) & ~(allOnes(And.ImmSize) << And.ImmLSB); 1035 unsigned Start, End; 1036 if (isRxSBGMask(Imm, And.RegSize, Start, End)) { 1037 unsigned NewOpcode; 1038 if (And.RegSize == 64) { 1039 NewOpcode = SystemZ::RISBG; 1040 // Prefer RISBGN if available, since it does not clobber CC. 1041 if (STI.hasMiscellaneousExtensions()) 1042 NewOpcode = SystemZ::RISBGN; 1043 } else { 1044 NewOpcode = SystemZ::RISBMux; 1045 Start &= 31; 1046 End &= 31; 1047 } 1048 MachineOperand &Dest = MI.getOperand(0); 1049 MachineOperand &Src = MI.getOperand(1); 1050 MachineInstrBuilder MIB = 1051 BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpcode)) 1052 .add(Dest) 1053 .addReg(0) 1054 .addReg(Src.getReg(), getKillRegState(Src.isKill()), 1055 Src.getSubReg()) 1056 .addImm(Start) 1057 .addImm(End + 128) 1058 .addImm(0); 1059 return finishConvertToThreeAddress(&MI, MIB, LV); 1060 } 1061 } 1062 return nullptr; 1063 } 1064 1065 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1066 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1067 MachineBasicBlock::iterator InsertPt, int FrameIndex, 1068 LiveIntervals *LIS) const { 1069 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 1070 const MachineFrameInfo &MFI = MF.getFrameInfo(); 1071 unsigned Size = MFI.getObjectSize(FrameIndex); 1072 unsigned Opcode = MI.getOpcode(); 1073 1074 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 1075 if (LIS != nullptr && (Opcode == SystemZ::LA || Opcode == SystemZ::LAY) && 1076 isInt<8>(MI.getOperand(2).getImm()) && !MI.getOperand(3).getReg()) { 1077 1078 // Check CC liveness, since new instruction introduces a dead 1079 // def of CC. 1080 MCRegUnitIterator CCUnit(SystemZ::CC, TRI); 1081 LiveRange &CCLiveRange = LIS->getRegUnit(*CCUnit); 1082 ++CCUnit; 1083 assert(!CCUnit.isValid() && "CC only has one reg unit."); 1084 SlotIndex MISlot = 1085 LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot(); 1086 if (!CCLiveRange.liveAt(MISlot)) { 1087 // LA(Y) %reg, CONST(%reg) -> AGSI %mem, CONST 1088 MachineInstr *BuiltMI = BuildMI(*InsertPt->getParent(), InsertPt, 1089 MI.getDebugLoc(), get(SystemZ::AGSI)) 1090 .addFrameIndex(FrameIndex) 1091 .addImm(0) 1092 .addImm(MI.getOperand(2).getImm()); 1093 BuiltMI->findRegisterDefOperand(SystemZ::CC)->setIsDead(true); 1094 CCLiveRange.createDeadDef(MISlot, LIS->getVNInfoAllocator()); 1095 return BuiltMI; 1096 } 1097 } 1098 return nullptr; 1099 } 1100 1101 // All other cases require a single operand. 1102 if (Ops.size() != 1) 1103 return nullptr; 1104 1105 unsigned OpNum = Ops[0]; 1106 assert(Size == 1107 MF.getRegInfo() 1108 .getRegClass(MI.getOperand(OpNum).getReg()) 1109 ->getSize() && 1110 "Invalid size combination"); 1111 1112 if ((Opcode == SystemZ::AHI || Opcode == SystemZ::AGHI) && OpNum == 0 && 1113 isInt<8>(MI.getOperand(2).getImm())) { 1114 // A(G)HI %reg, CONST -> A(G)SI %mem, CONST 1115 Opcode = (Opcode == SystemZ::AHI ? SystemZ::ASI : SystemZ::AGSI); 1116 MachineInstr *BuiltMI = 1117 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode)) 1118 .addFrameIndex(FrameIndex) 1119 .addImm(0) 1120 .addImm(MI.getOperand(2).getImm()); 1121 transferDeadCC(&MI, BuiltMI); 1122 return BuiltMI; 1123 } 1124 1125 if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) { 1126 bool Op0IsGPR = (Opcode == SystemZ::LGDR); 1127 bool Op1IsGPR = (Opcode == SystemZ::LDGR); 1128 // If we're spilling the destination of an LDGR or LGDR, store the 1129 // source register instead. 1130 if (OpNum == 0) { 1131 unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD; 1132 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1133 get(StoreOpcode)) 1134 .add(MI.getOperand(1)) 1135 .addFrameIndex(FrameIndex) 1136 .addImm(0) 1137 .addReg(0); 1138 } 1139 // If we're spilling the source of an LDGR or LGDR, load the 1140 // destination register instead. 1141 if (OpNum == 1) { 1142 unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD; 1143 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1144 get(LoadOpcode)) 1145 .add(MI.getOperand(0)) 1146 .addFrameIndex(FrameIndex) 1147 .addImm(0) 1148 .addReg(0); 1149 } 1150 } 1151 1152 // Look for cases where the source of a simple store or the destination 1153 // of a simple load is being spilled. Try to use MVC instead. 1154 // 1155 // Although MVC is in practice a fast choice in these cases, it is still 1156 // logically a bytewise copy. This means that we cannot use it if the 1157 // load or store is volatile. We also wouldn't be able to use MVC if 1158 // the two memories partially overlap, but that case cannot occur here, 1159 // because we know that one of the memories is a full frame index. 1160 // 1161 // For performance reasons, we also want to avoid using MVC if the addresses 1162 // might be equal. We don't worry about that case here, because spill slot 1163 // coloring happens later, and because we have special code to remove 1164 // MVCs that turn out to be redundant. 1165 if (OpNum == 0 && MI.hasOneMemOperand()) { 1166 MachineMemOperand *MMO = *MI.memoperands_begin(); 1167 if (MMO->getSize() == Size && !MMO->isVolatile()) { 1168 // Handle conversion of loads. 1169 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXLoad)) { 1170 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1171 get(SystemZ::MVC)) 1172 .addFrameIndex(FrameIndex) 1173 .addImm(0) 1174 .addImm(Size) 1175 .add(MI.getOperand(1)) 1176 .addImm(MI.getOperand(2).getImm()) 1177 .addMemOperand(MMO); 1178 } 1179 // Handle conversion of stores. 1180 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXStore)) { 1181 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1182 get(SystemZ::MVC)) 1183 .add(MI.getOperand(1)) 1184 .addImm(MI.getOperand(2).getImm()) 1185 .addImm(Size) 1186 .addFrameIndex(FrameIndex) 1187 .addImm(0) 1188 .addMemOperand(MMO); 1189 } 1190 } 1191 } 1192 1193 // If the spilled operand is the final one, try to change <INSN>R 1194 // into <INSN>. 1195 int MemOpcode = SystemZ::getMemOpcode(Opcode); 1196 if (MemOpcode >= 0) { 1197 unsigned NumOps = MI.getNumExplicitOperands(); 1198 if (OpNum == NumOps - 1) { 1199 const MCInstrDesc &MemDesc = get(MemOpcode); 1200 uint64_t AccessBytes = SystemZII::getAccessSize(MemDesc.TSFlags); 1201 assert(AccessBytes != 0 && "Size of access should be known"); 1202 assert(AccessBytes <= Size && "Access outside the frame index"); 1203 uint64_t Offset = Size - AccessBytes; 1204 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt, 1205 MI.getDebugLoc(), get(MemOpcode)); 1206 for (unsigned I = 0; I < OpNum; ++I) 1207 MIB.add(MI.getOperand(I)); 1208 MIB.addFrameIndex(FrameIndex).addImm(Offset); 1209 if (MemDesc.TSFlags & SystemZII::HasIndex) 1210 MIB.addReg(0); 1211 transferDeadCC(&MI, MIB); 1212 return MIB; 1213 } 1214 } 1215 1216 return nullptr; 1217 } 1218 1219 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1220 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1221 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, 1222 LiveIntervals *LIS) const { 1223 return nullptr; 1224 } 1225 1226 bool SystemZInstrInfo::expandPostRAPseudo(MachineInstr &MI) const { 1227 switch (MI.getOpcode()) { 1228 case SystemZ::L128: 1229 splitMove(MI, SystemZ::LG); 1230 return true; 1231 1232 case SystemZ::ST128: 1233 splitMove(MI, SystemZ::STG); 1234 return true; 1235 1236 case SystemZ::LX: 1237 splitMove(MI, SystemZ::LD); 1238 return true; 1239 1240 case SystemZ::STX: 1241 splitMove(MI, SystemZ::STD); 1242 return true; 1243 1244 case SystemZ::LBMux: 1245 expandRXYPseudo(MI, SystemZ::LB, SystemZ::LBH); 1246 return true; 1247 1248 case SystemZ::LHMux: 1249 expandRXYPseudo(MI, SystemZ::LH, SystemZ::LHH); 1250 return true; 1251 1252 case SystemZ::LLCRMux: 1253 expandZExtPseudo(MI, SystemZ::LLCR, 8); 1254 return true; 1255 1256 case SystemZ::LLHRMux: 1257 expandZExtPseudo(MI, SystemZ::LLHR, 16); 1258 return true; 1259 1260 case SystemZ::LLCMux: 1261 expandRXYPseudo(MI, SystemZ::LLC, SystemZ::LLCH); 1262 return true; 1263 1264 case SystemZ::LLHMux: 1265 expandRXYPseudo(MI, SystemZ::LLH, SystemZ::LLHH); 1266 return true; 1267 1268 case SystemZ::LMux: 1269 expandRXYPseudo(MI, SystemZ::L, SystemZ::LFH); 1270 return true; 1271 1272 case SystemZ::LOCMux: 1273 expandLOCPseudo(MI, SystemZ::LOC, SystemZ::LOCFH); 1274 return true; 1275 1276 case SystemZ::LOCHIMux: 1277 expandLOCPseudo(MI, SystemZ::LOCHI, SystemZ::LOCHHI); 1278 return true; 1279 1280 case SystemZ::LOCRMux: 1281 expandLOCRPseudo(MI, SystemZ::LOCR, SystemZ::LOCFHR); 1282 return true; 1283 1284 case SystemZ::STCMux: 1285 expandRXYPseudo(MI, SystemZ::STC, SystemZ::STCH); 1286 return true; 1287 1288 case SystemZ::STHMux: 1289 expandRXYPseudo(MI, SystemZ::STH, SystemZ::STHH); 1290 return true; 1291 1292 case SystemZ::STMux: 1293 expandRXYPseudo(MI, SystemZ::ST, SystemZ::STFH); 1294 return true; 1295 1296 case SystemZ::STOCMux: 1297 expandLOCPseudo(MI, SystemZ::STOC, SystemZ::STOCFH); 1298 return true; 1299 1300 case SystemZ::LHIMux: 1301 expandRIPseudo(MI, SystemZ::LHI, SystemZ::IIHF, true); 1302 return true; 1303 1304 case SystemZ::IIFMux: 1305 expandRIPseudo(MI, SystemZ::IILF, SystemZ::IIHF, false); 1306 return true; 1307 1308 case SystemZ::IILMux: 1309 expandRIPseudo(MI, SystemZ::IILL, SystemZ::IIHL, false); 1310 return true; 1311 1312 case SystemZ::IIHMux: 1313 expandRIPseudo(MI, SystemZ::IILH, SystemZ::IIHH, false); 1314 return true; 1315 1316 case SystemZ::NIFMux: 1317 expandRIPseudo(MI, SystemZ::NILF, SystemZ::NIHF, false); 1318 return true; 1319 1320 case SystemZ::NILMux: 1321 expandRIPseudo(MI, SystemZ::NILL, SystemZ::NIHL, false); 1322 return true; 1323 1324 case SystemZ::NIHMux: 1325 expandRIPseudo(MI, SystemZ::NILH, SystemZ::NIHH, false); 1326 return true; 1327 1328 case SystemZ::OIFMux: 1329 expandRIPseudo(MI, SystemZ::OILF, SystemZ::OIHF, false); 1330 return true; 1331 1332 case SystemZ::OILMux: 1333 expandRIPseudo(MI, SystemZ::OILL, SystemZ::OIHL, false); 1334 return true; 1335 1336 case SystemZ::OIHMux: 1337 expandRIPseudo(MI, SystemZ::OILH, SystemZ::OIHH, false); 1338 return true; 1339 1340 case SystemZ::XIFMux: 1341 expandRIPseudo(MI, SystemZ::XILF, SystemZ::XIHF, false); 1342 return true; 1343 1344 case SystemZ::TMLMux: 1345 expandRIPseudo(MI, SystemZ::TMLL, SystemZ::TMHL, false); 1346 return true; 1347 1348 case SystemZ::TMHMux: 1349 expandRIPseudo(MI, SystemZ::TMLH, SystemZ::TMHH, false); 1350 return true; 1351 1352 case SystemZ::AHIMux: 1353 expandRIPseudo(MI, SystemZ::AHI, SystemZ::AIH, false); 1354 return true; 1355 1356 case SystemZ::AHIMuxK: 1357 expandRIEPseudo(MI, SystemZ::AHI, SystemZ::AHIK, SystemZ::AIH); 1358 return true; 1359 1360 case SystemZ::AFIMux: 1361 expandRIPseudo(MI, SystemZ::AFI, SystemZ::AIH, false); 1362 return true; 1363 1364 case SystemZ::CHIMux: 1365 expandRIPseudo(MI, SystemZ::CHI, SystemZ::CIH, false); 1366 return true; 1367 1368 case SystemZ::CFIMux: 1369 expandRIPseudo(MI, SystemZ::CFI, SystemZ::CIH, false); 1370 return true; 1371 1372 case SystemZ::CLFIMux: 1373 expandRIPseudo(MI, SystemZ::CLFI, SystemZ::CLIH, false); 1374 return true; 1375 1376 case SystemZ::CMux: 1377 expandRXYPseudo(MI, SystemZ::C, SystemZ::CHF); 1378 return true; 1379 1380 case SystemZ::CLMux: 1381 expandRXYPseudo(MI, SystemZ::CL, SystemZ::CLHF); 1382 return true; 1383 1384 case SystemZ::RISBMux: { 1385 bool DestIsHigh = isHighReg(MI.getOperand(0).getReg()); 1386 bool SrcIsHigh = isHighReg(MI.getOperand(2).getReg()); 1387 if (SrcIsHigh == DestIsHigh) 1388 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHH : SystemZ::RISBLL)); 1389 else { 1390 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHL : SystemZ::RISBLH)); 1391 MI.getOperand(5).setImm(MI.getOperand(5).getImm() ^ 32); 1392 } 1393 return true; 1394 } 1395 1396 case SystemZ::ADJDYNALLOC: 1397 splitAdjDynAlloc(MI); 1398 return true; 1399 1400 case TargetOpcode::LOAD_STACK_GUARD: 1401 expandLoadStackGuard(&MI); 1402 return true; 1403 1404 default: 1405 return false; 1406 } 1407 } 1408 1409 unsigned SystemZInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { 1410 if (MI.getOpcode() == TargetOpcode::INLINEASM) { 1411 const MachineFunction *MF = MI.getParent()->getParent(); 1412 const char *AsmStr = MI.getOperand(0).getSymbolName(); 1413 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo()); 1414 } 1415 return MI.getDesc().getSize(); 1416 } 1417 1418 SystemZII::Branch 1419 SystemZInstrInfo::getBranchInfo(const MachineInstr &MI) const { 1420 switch (MI.getOpcode()) { 1421 case SystemZ::BR: 1422 case SystemZ::J: 1423 case SystemZ::JG: 1424 return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY, 1425 SystemZ::CCMASK_ANY, &MI.getOperand(0)); 1426 1427 case SystemZ::BRC: 1428 case SystemZ::BRCL: 1429 return SystemZII::Branch(SystemZII::BranchNormal, MI.getOperand(0).getImm(), 1430 MI.getOperand(1).getImm(), &MI.getOperand(2)); 1431 1432 case SystemZ::BRCT: 1433 case SystemZ::BRCTH: 1434 return SystemZII::Branch(SystemZII::BranchCT, SystemZ::CCMASK_ICMP, 1435 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1436 1437 case SystemZ::BRCTG: 1438 return SystemZII::Branch(SystemZII::BranchCTG, SystemZ::CCMASK_ICMP, 1439 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1440 1441 case SystemZ::CIJ: 1442 case SystemZ::CRJ: 1443 return SystemZII::Branch(SystemZII::BranchC, SystemZ::CCMASK_ICMP, 1444 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1445 1446 case SystemZ::CLIJ: 1447 case SystemZ::CLRJ: 1448 return SystemZII::Branch(SystemZII::BranchCL, SystemZ::CCMASK_ICMP, 1449 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1450 1451 case SystemZ::CGIJ: 1452 case SystemZ::CGRJ: 1453 return SystemZII::Branch(SystemZII::BranchCG, SystemZ::CCMASK_ICMP, 1454 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1455 1456 case SystemZ::CLGIJ: 1457 case SystemZ::CLGRJ: 1458 return SystemZII::Branch(SystemZII::BranchCLG, SystemZ::CCMASK_ICMP, 1459 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1460 1461 default: 1462 llvm_unreachable("Unrecognized branch opcode"); 1463 } 1464 } 1465 1466 void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC, 1467 unsigned &LoadOpcode, 1468 unsigned &StoreOpcode) const { 1469 if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) { 1470 LoadOpcode = SystemZ::L; 1471 StoreOpcode = SystemZ::ST; 1472 } else if (RC == &SystemZ::GRH32BitRegClass) { 1473 LoadOpcode = SystemZ::LFH; 1474 StoreOpcode = SystemZ::STFH; 1475 } else if (RC == &SystemZ::GRX32BitRegClass) { 1476 LoadOpcode = SystemZ::LMux; 1477 StoreOpcode = SystemZ::STMux; 1478 } else if (RC == &SystemZ::GR64BitRegClass || 1479 RC == &SystemZ::ADDR64BitRegClass) { 1480 LoadOpcode = SystemZ::LG; 1481 StoreOpcode = SystemZ::STG; 1482 } else if (RC == &SystemZ::GR128BitRegClass || 1483 RC == &SystemZ::ADDR128BitRegClass) { 1484 LoadOpcode = SystemZ::L128; 1485 StoreOpcode = SystemZ::ST128; 1486 } else if (RC == &SystemZ::FP32BitRegClass) { 1487 LoadOpcode = SystemZ::LE; 1488 StoreOpcode = SystemZ::STE; 1489 } else if (RC == &SystemZ::FP64BitRegClass) { 1490 LoadOpcode = SystemZ::LD; 1491 StoreOpcode = SystemZ::STD; 1492 } else if (RC == &SystemZ::FP128BitRegClass) { 1493 LoadOpcode = SystemZ::LX; 1494 StoreOpcode = SystemZ::STX; 1495 } else if (RC == &SystemZ::VR32BitRegClass) { 1496 LoadOpcode = SystemZ::VL32; 1497 StoreOpcode = SystemZ::VST32; 1498 } else if (RC == &SystemZ::VR64BitRegClass) { 1499 LoadOpcode = SystemZ::VL64; 1500 StoreOpcode = SystemZ::VST64; 1501 } else if (RC == &SystemZ::VF128BitRegClass || 1502 RC == &SystemZ::VR128BitRegClass) { 1503 LoadOpcode = SystemZ::VL; 1504 StoreOpcode = SystemZ::VST; 1505 } else 1506 llvm_unreachable("Unsupported regclass to load or store"); 1507 } 1508 1509 unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode, 1510 int64_t Offset) const { 1511 const MCInstrDesc &MCID = get(Opcode); 1512 int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset); 1513 if (isUInt<12>(Offset) && isUInt<12>(Offset2)) { 1514 // Get the instruction to use for unsigned 12-bit displacements. 1515 int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode); 1516 if (Disp12Opcode >= 0) 1517 return Disp12Opcode; 1518 1519 // All address-related instructions can use unsigned 12-bit 1520 // displacements. 1521 return Opcode; 1522 } 1523 if (isInt<20>(Offset) && isInt<20>(Offset2)) { 1524 // Get the instruction to use for signed 20-bit displacements. 1525 int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode); 1526 if (Disp20Opcode >= 0) 1527 return Disp20Opcode; 1528 1529 // Check whether Opcode allows signed 20-bit displacements. 1530 if (MCID.TSFlags & SystemZII::Has20BitOffset) 1531 return Opcode; 1532 } 1533 return 0; 1534 } 1535 1536 unsigned SystemZInstrInfo::getLoadAndTest(unsigned Opcode) const { 1537 switch (Opcode) { 1538 case SystemZ::L: return SystemZ::LT; 1539 case SystemZ::LY: return SystemZ::LT; 1540 case SystemZ::LG: return SystemZ::LTG; 1541 case SystemZ::LGF: return SystemZ::LTGF; 1542 case SystemZ::LR: return SystemZ::LTR; 1543 case SystemZ::LGFR: return SystemZ::LTGFR; 1544 case SystemZ::LGR: return SystemZ::LTGR; 1545 case SystemZ::LER: return SystemZ::LTEBR; 1546 case SystemZ::LDR: return SystemZ::LTDBR; 1547 case SystemZ::LXR: return SystemZ::LTXBR; 1548 case SystemZ::LCDFR: return SystemZ::LCDBR; 1549 case SystemZ::LPDFR: return SystemZ::LPDBR; 1550 case SystemZ::LNDFR: return SystemZ::LNDBR; 1551 case SystemZ::LCDFR_32: return SystemZ::LCEBR; 1552 case SystemZ::LPDFR_32: return SystemZ::LPEBR; 1553 case SystemZ::LNDFR_32: return SystemZ::LNEBR; 1554 // On zEC12 we prefer to use RISBGN. But if there is a chance to 1555 // actually use the condition code, we may turn it back into RISGB. 1556 // Note that RISBG is not really a "load-and-test" instruction, 1557 // but sets the same condition code values, so is OK to use here. 1558 case SystemZ::RISBGN: return SystemZ::RISBG; 1559 default: return 0; 1560 } 1561 } 1562 1563 // Return true if Mask matches the regexp 0*1+0*, given that zero masks 1564 // have already been filtered out. Store the first set bit in LSB and 1565 // the number of set bits in Length if so. 1566 static bool isStringOfOnes(uint64_t Mask, unsigned &LSB, unsigned &Length) { 1567 unsigned First = findFirstSet(Mask); 1568 uint64_t Top = (Mask >> First) + 1; 1569 if ((Top & -Top) == Top) { 1570 LSB = First; 1571 Length = findFirstSet(Top); 1572 return true; 1573 } 1574 return false; 1575 } 1576 1577 bool SystemZInstrInfo::isRxSBGMask(uint64_t Mask, unsigned BitSize, 1578 unsigned &Start, unsigned &End) const { 1579 // Reject trivial all-zero masks. 1580 Mask &= allOnes(BitSize); 1581 if (Mask == 0) 1582 return false; 1583 1584 // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of 1585 // the msb and End specifies the index of the lsb. 1586 unsigned LSB, Length; 1587 if (isStringOfOnes(Mask, LSB, Length)) { 1588 Start = 63 - (LSB + Length - 1); 1589 End = 63 - LSB; 1590 return true; 1591 } 1592 1593 // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb 1594 // of the low 1s and End specifies the lsb of the high 1s. 1595 if (isStringOfOnes(Mask ^ allOnes(BitSize), LSB, Length)) { 1596 assert(LSB > 0 && "Bottom bit must be set"); 1597 assert(LSB + Length < BitSize && "Top bit must be set"); 1598 Start = 63 - (LSB - 1); 1599 End = 63 - (LSB + Length); 1600 return true; 1601 } 1602 1603 return false; 1604 } 1605 1606 unsigned SystemZInstrInfo::getFusedCompare(unsigned Opcode, 1607 SystemZII::FusedCompareType Type, 1608 const MachineInstr *MI) const { 1609 switch (Opcode) { 1610 case SystemZ::CHI: 1611 case SystemZ::CGHI: 1612 if (!(MI && isInt<8>(MI->getOperand(1).getImm()))) 1613 return 0; 1614 break; 1615 case SystemZ::CLFI: 1616 case SystemZ::CLGFI: 1617 if (!(MI && isUInt<8>(MI->getOperand(1).getImm()))) 1618 return 0; 1619 break; 1620 case SystemZ::CL: 1621 case SystemZ::CLG: 1622 if (!STI.hasMiscellaneousExtensions()) 1623 return 0; 1624 if (!(MI && MI->getOperand(3).getReg() == 0)) 1625 return 0; 1626 break; 1627 } 1628 switch (Type) { 1629 case SystemZII::CompareAndBranch: 1630 switch (Opcode) { 1631 case SystemZ::CR: 1632 return SystemZ::CRJ; 1633 case SystemZ::CGR: 1634 return SystemZ::CGRJ; 1635 case SystemZ::CHI: 1636 return SystemZ::CIJ; 1637 case SystemZ::CGHI: 1638 return SystemZ::CGIJ; 1639 case SystemZ::CLR: 1640 return SystemZ::CLRJ; 1641 case SystemZ::CLGR: 1642 return SystemZ::CLGRJ; 1643 case SystemZ::CLFI: 1644 return SystemZ::CLIJ; 1645 case SystemZ::CLGFI: 1646 return SystemZ::CLGIJ; 1647 default: 1648 return 0; 1649 } 1650 case SystemZII::CompareAndReturn: 1651 switch (Opcode) { 1652 case SystemZ::CR: 1653 return SystemZ::CRBReturn; 1654 case SystemZ::CGR: 1655 return SystemZ::CGRBReturn; 1656 case SystemZ::CHI: 1657 return SystemZ::CIBReturn; 1658 case SystemZ::CGHI: 1659 return SystemZ::CGIBReturn; 1660 case SystemZ::CLR: 1661 return SystemZ::CLRBReturn; 1662 case SystemZ::CLGR: 1663 return SystemZ::CLGRBReturn; 1664 case SystemZ::CLFI: 1665 return SystemZ::CLIBReturn; 1666 case SystemZ::CLGFI: 1667 return SystemZ::CLGIBReturn; 1668 default: 1669 return 0; 1670 } 1671 case SystemZII::CompareAndSibcall: 1672 switch (Opcode) { 1673 case SystemZ::CR: 1674 return SystemZ::CRBCall; 1675 case SystemZ::CGR: 1676 return SystemZ::CGRBCall; 1677 case SystemZ::CHI: 1678 return SystemZ::CIBCall; 1679 case SystemZ::CGHI: 1680 return SystemZ::CGIBCall; 1681 case SystemZ::CLR: 1682 return SystemZ::CLRBCall; 1683 case SystemZ::CLGR: 1684 return SystemZ::CLGRBCall; 1685 case SystemZ::CLFI: 1686 return SystemZ::CLIBCall; 1687 case SystemZ::CLGFI: 1688 return SystemZ::CLGIBCall; 1689 default: 1690 return 0; 1691 } 1692 case SystemZII::CompareAndTrap: 1693 switch (Opcode) { 1694 case SystemZ::CR: 1695 return SystemZ::CRT; 1696 case SystemZ::CGR: 1697 return SystemZ::CGRT; 1698 case SystemZ::CHI: 1699 return SystemZ::CIT; 1700 case SystemZ::CGHI: 1701 return SystemZ::CGIT; 1702 case SystemZ::CLR: 1703 return SystemZ::CLRT; 1704 case SystemZ::CLGR: 1705 return SystemZ::CLGRT; 1706 case SystemZ::CLFI: 1707 return SystemZ::CLFIT; 1708 case SystemZ::CLGFI: 1709 return SystemZ::CLGIT; 1710 case SystemZ::CL: 1711 return SystemZ::CLT; 1712 case SystemZ::CLG: 1713 return SystemZ::CLGT; 1714 default: 1715 return 0; 1716 } 1717 } 1718 return 0; 1719 } 1720 1721 unsigned SystemZInstrInfo::getLoadAndTrap(unsigned Opcode) const { 1722 if (!STI.hasLoadAndTrap()) 1723 return 0; 1724 switch (Opcode) { 1725 case SystemZ::L: 1726 case SystemZ::LY: 1727 return SystemZ::LAT; 1728 case SystemZ::LG: 1729 return SystemZ::LGAT; 1730 case SystemZ::LFH: 1731 return SystemZ::LFHAT; 1732 case SystemZ::LLGF: 1733 return SystemZ::LLGFAT; 1734 case SystemZ::LLGT: 1735 return SystemZ::LLGTAT; 1736 } 1737 return 0; 1738 } 1739 1740 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB, 1741 MachineBasicBlock::iterator MBBI, 1742 unsigned Reg, uint64_t Value) const { 1743 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 1744 unsigned Opcode; 1745 if (isInt<16>(Value)) 1746 Opcode = SystemZ::LGHI; 1747 else if (SystemZ::isImmLL(Value)) 1748 Opcode = SystemZ::LLILL; 1749 else if (SystemZ::isImmLH(Value)) { 1750 Opcode = SystemZ::LLILH; 1751 Value >>= 16; 1752 } else { 1753 assert(isInt<32>(Value) && "Huge values not handled yet"); 1754 Opcode = SystemZ::LGFI; 1755 } 1756 BuildMI(MBB, MBBI, DL, get(Opcode), Reg).addImm(Value); 1757 } 1758 1759 bool SystemZInstrInfo:: 1760 areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb, 1761 AliasAnalysis *AA) const { 1762 1763 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand()) 1764 return false; 1765 1766 // If mem-operands show that the same address Value is used by both 1767 // instructions, check for non-overlapping offsets and widths. Not 1768 // sure if a register based analysis would be an improvement... 1769 1770 MachineMemOperand *MMOa = *MIa.memoperands_begin(); 1771 MachineMemOperand *MMOb = *MIb.memoperands_begin(); 1772 const Value *VALa = MMOa->getValue(); 1773 const Value *VALb = MMOb->getValue(); 1774 bool SameVal = (VALa && VALb && (VALa == VALb)); 1775 if (!SameVal) { 1776 const PseudoSourceValue *PSVa = MMOa->getPseudoValue(); 1777 const PseudoSourceValue *PSVb = MMOb->getPseudoValue(); 1778 if (PSVa && PSVb && (PSVa == PSVb)) 1779 SameVal = true; 1780 } 1781 if (SameVal) { 1782 int OffsetA = MMOa->getOffset(), OffsetB = MMOb->getOffset(); 1783 int WidthA = MMOa->getSize(), WidthB = MMOb->getSize(); 1784 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB; 1785 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA; 1786 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; 1787 if (LowOffset + LowWidth <= HighOffset) 1788 return true; 1789 } 1790 1791 return false; 1792 } 1793