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/MachineMemOperand.h" 27 #include "llvm/CodeGen/MachineOperand.h" 28 #include "llvm/CodeGen/MachineRegisterInfo.h" 29 #include "llvm/CodeGen/SlotIndexes.h" 30 #include "llvm/MC/MCInstrDesc.h" 31 #include "llvm/MC/MCRegisterInfo.h" 32 #include "llvm/Support/BranchProbability.h" 33 #include "llvm/Support/ErrorHandling.h" 34 #include "llvm/Support/MathExtras.h" 35 #include "llvm/Target/TargetInstrInfo.h" 36 #include "llvm/Target/TargetMachine.h" 37 #include "llvm/Target/TargetSubtargetInfo.h" 38 #include <cassert> 39 #include <cstdint> 40 #include <iterator> 41 42 using namespace llvm; 43 44 #define GET_INSTRINFO_CTOR_DTOR 45 #define GET_INSTRMAP_INFO 46 #include "SystemZGenInstrInfo.inc" 47 48 // Return a mask with Count low bits set. 49 static uint64_t allOnes(unsigned int Count) { 50 return Count == 0 ? 0 : (uint64_t(1) << (Count - 1) << 1) - 1; 51 } 52 53 // Reg should be a 32-bit GPR. Return true if it is a high register rather 54 // than a low register. 55 static bool isHighReg(unsigned int Reg) { 56 if (SystemZ::GRH32BitRegClass.contains(Reg)) 57 return true; 58 assert(SystemZ::GR32BitRegClass.contains(Reg) && "Invalid GRX32"); 59 return false; 60 } 61 62 // Pin the vtable to this file. 63 void SystemZInstrInfo::anchor() {} 64 65 SystemZInstrInfo::SystemZInstrInfo(SystemZSubtarget &sti) 66 : SystemZGenInstrInfo(SystemZ::ADJCALLSTACKDOWN, SystemZ::ADJCALLSTACKUP), 67 RI(), STI(sti) { 68 } 69 70 // MI is a 128-bit load or store. Split it into two 64-bit loads or stores, 71 // each having the opcode given by NewOpcode. 72 void SystemZInstrInfo::splitMove(MachineBasicBlock::iterator MI, 73 unsigned NewOpcode) const { 74 MachineBasicBlock *MBB = MI->getParent(); 75 MachineFunction &MF = *MBB->getParent(); 76 77 // Get two load or store instructions. Use the original instruction for one 78 // of them (arbitrarily the second here) and create a clone for the other. 79 MachineInstr *EarlierMI = MF.CloneMachineInstr(&*MI); 80 MBB->insert(MI, EarlierMI); 81 82 // Set up the two 64-bit registers and remember super reg and its flags. 83 MachineOperand &HighRegOp = EarlierMI->getOperand(0); 84 MachineOperand &LowRegOp = MI->getOperand(0); 85 unsigned Reg128 = LowRegOp.getReg(); 86 unsigned Reg128Killed = getKillRegState(LowRegOp.isKill()); 87 unsigned Reg128Undef = getUndefRegState(LowRegOp.isUndef()); 88 HighRegOp.setReg(RI.getSubReg(HighRegOp.getReg(), SystemZ::subreg_h64)); 89 LowRegOp.setReg(RI.getSubReg(LowRegOp.getReg(), SystemZ::subreg_l64)); 90 91 if (MI->mayStore()) { 92 // Add implicit uses of the super register in case one of the subregs is 93 // undefined. We could track liveness and skip storing an undefined 94 // subreg, but this is hopefully rare (discovered with llvm-stress). 95 // If Reg128 was killed, set kill flag on MI. 96 unsigned Reg128UndefImpl = (Reg128Undef | RegState::Implicit); 97 MachineInstrBuilder(MF, EarlierMI).addReg(Reg128, Reg128UndefImpl); 98 MachineInstrBuilder(MF, MI).addReg(Reg128, (Reg128UndefImpl | Reg128Killed)); 99 } 100 101 // The address in the first (high) instruction is already correct. 102 // Adjust the offset in the second (low) instruction. 103 MachineOperand &HighOffsetOp = EarlierMI->getOperand(2); 104 MachineOperand &LowOffsetOp = MI->getOperand(2); 105 LowOffsetOp.setImm(LowOffsetOp.getImm() + 8); 106 107 // Clear the kill flags on the registers in the first instruction. 108 if (EarlierMI->getOperand(0).isReg() && EarlierMI->getOperand(0).isUse()) 109 EarlierMI->getOperand(0).setIsKill(false); 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 MachineInstrBuilder MIB = 225 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), 226 MI.getOperand(0).getReg(), MI.getOperand(1).getReg(), LowOpcode, 227 Size, MI.getOperand(1).isKill(), MI.getOperand(1).isUndef()); 228 229 // Keep the remaining operands as-is. 230 for (unsigned I = 2; I < MI.getNumOperands(); ++I) 231 MIB.add(MI.getOperand(I)); 232 233 MI.eraseFromParent(); 234 } 235 236 void SystemZInstrInfo::expandLoadStackGuard(MachineInstr *MI) const { 237 MachineBasicBlock *MBB = MI->getParent(); 238 MachineFunction &MF = *MBB->getParent(); 239 const unsigned Reg = MI->getOperand(0).getReg(); 240 241 // Conveniently, all 4 instructions are cloned from LOAD_STACK_GUARD, 242 // so they already have operand 0 set to reg. 243 244 // ear <reg>, %a0 245 MachineInstr *Ear1MI = MF.CloneMachineInstr(MI); 246 MBB->insert(MI, Ear1MI); 247 Ear1MI->setDesc(get(SystemZ::EAR)); 248 MachineInstrBuilder(MF, Ear1MI).addReg(SystemZ::A0); 249 250 // sllg <reg>, <reg>, 32 251 MachineInstr *SllgMI = MF.CloneMachineInstr(MI); 252 MBB->insert(MI, SllgMI); 253 SllgMI->setDesc(get(SystemZ::SLLG)); 254 MachineInstrBuilder(MF, SllgMI).addReg(Reg).addReg(0).addImm(32); 255 256 // ear <reg>, %a1 257 MachineInstr *Ear2MI = MF.CloneMachineInstr(MI); 258 MBB->insert(MI, Ear2MI); 259 Ear2MI->setDesc(get(SystemZ::EAR)); 260 MachineInstrBuilder(MF, Ear2MI).addReg(SystemZ::A1); 261 262 // lg <reg>, 40(<reg>) 263 MI->setDesc(get(SystemZ::LG)); 264 MachineInstrBuilder(MF, MI).addReg(Reg).addImm(40).addReg(0); 265 } 266 267 // Emit a zero-extending move from 32-bit GPR SrcReg to 32-bit GPR 268 // DestReg before MBBI in MBB. Use LowLowOpcode when both DestReg and SrcReg 269 // are low registers, otherwise use RISB[LH]G. Size is the number of bits 270 // taken from the low end of SrcReg (8 for LLCR, 16 for LLHR and 32 for LR). 271 // KillSrc is true if this move is the last use of SrcReg. 272 MachineInstrBuilder 273 SystemZInstrInfo::emitGRX32Move(MachineBasicBlock &MBB, 274 MachineBasicBlock::iterator MBBI, 275 const DebugLoc &DL, unsigned DestReg, 276 unsigned SrcReg, unsigned LowLowOpcode, 277 unsigned Size, bool KillSrc, 278 bool UndefSrc) const { 279 unsigned Opcode; 280 bool DestIsHigh = isHighReg(DestReg); 281 bool SrcIsHigh = isHighReg(SrcReg); 282 if (DestIsHigh && SrcIsHigh) 283 Opcode = SystemZ::RISBHH; 284 else if (DestIsHigh && !SrcIsHigh) 285 Opcode = SystemZ::RISBHL; 286 else if (!DestIsHigh && SrcIsHigh) 287 Opcode = SystemZ::RISBLH; 288 else { 289 return BuildMI(MBB, MBBI, DL, get(LowLowOpcode), DestReg) 290 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)); 291 } 292 unsigned Rotate = (DestIsHigh != SrcIsHigh ? 32 : 0); 293 return BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 294 .addReg(DestReg, RegState::Undef) 295 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)) 296 .addImm(32 - Size).addImm(128 + 31).addImm(Rotate); 297 } 298 299 MachineInstr *SystemZInstrInfo::commuteInstructionImpl(MachineInstr &MI, 300 bool NewMI, 301 unsigned OpIdx1, 302 unsigned OpIdx2) const { 303 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & { 304 if (NewMI) 305 return *MI.getParent()->getParent()->CloneMachineInstr(&MI); 306 return MI; 307 }; 308 309 switch (MI.getOpcode()) { 310 case SystemZ::LOCRMux: 311 case SystemZ::LOCFHR: 312 case SystemZ::LOCR: 313 case SystemZ::LOCGR: { 314 auto &WorkingMI = cloneIfNew(MI); 315 // Invert condition. 316 unsigned CCValid = WorkingMI.getOperand(3).getImm(); 317 unsigned CCMask = WorkingMI.getOperand(4).getImm(); 318 WorkingMI.getOperand(4).setImm(CCMask ^ CCValid); 319 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 320 OpIdx1, OpIdx2); 321 } 322 default: 323 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 324 } 325 } 326 327 // If MI is a simple load or store for a frame object, return the register 328 // it loads or stores and set FrameIndex to the index of the frame object. 329 // Return 0 otherwise. 330 // 331 // Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 332 static int isSimpleMove(const MachineInstr &MI, int &FrameIndex, 333 unsigned Flag) { 334 const MCInstrDesc &MCID = MI.getDesc(); 335 if ((MCID.TSFlags & Flag) && MI.getOperand(1).isFI() && 336 MI.getOperand(2).getImm() == 0 && MI.getOperand(3).getReg() == 0) { 337 FrameIndex = MI.getOperand(1).getIndex(); 338 return MI.getOperand(0).getReg(); 339 } 340 return 0; 341 } 342 343 unsigned SystemZInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 344 int &FrameIndex) const { 345 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXLoad); 346 } 347 348 unsigned SystemZInstrInfo::isStoreToStackSlot(const MachineInstr &MI, 349 int &FrameIndex) const { 350 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXStore); 351 } 352 353 bool SystemZInstrInfo::isStackSlotCopy(const MachineInstr &MI, 354 int &DestFrameIndex, 355 int &SrcFrameIndex) const { 356 // Check for MVC 0(Length,FI1),0(FI2) 357 const MachineFrameInfo &MFI = MI.getParent()->getParent()->getFrameInfo(); 358 if (MI.getOpcode() != SystemZ::MVC || !MI.getOperand(0).isFI() || 359 MI.getOperand(1).getImm() != 0 || !MI.getOperand(3).isFI() || 360 MI.getOperand(4).getImm() != 0) 361 return false; 362 363 // Check that Length covers the full slots. 364 int64_t Length = MI.getOperand(2).getImm(); 365 unsigned FI1 = MI.getOperand(0).getIndex(); 366 unsigned FI2 = MI.getOperand(3).getIndex(); 367 if (MFI.getObjectSize(FI1) != Length || 368 MFI.getObjectSize(FI2) != Length) 369 return false; 370 371 DestFrameIndex = FI1; 372 SrcFrameIndex = FI2; 373 return true; 374 } 375 376 bool SystemZInstrInfo::analyzeBranch(MachineBasicBlock &MBB, 377 MachineBasicBlock *&TBB, 378 MachineBasicBlock *&FBB, 379 SmallVectorImpl<MachineOperand> &Cond, 380 bool AllowModify) const { 381 // Most of the code and comments here are boilerplate. 382 383 // Start from the bottom of the block and work up, examining the 384 // terminator instructions. 385 MachineBasicBlock::iterator I = MBB.end(); 386 while (I != MBB.begin()) { 387 --I; 388 if (I->isDebugValue()) 389 continue; 390 391 // Working from the bottom, when we see a non-terminator instruction, we're 392 // done. 393 if (!isUnpredicatedTerminator(*I)) 394 break; 395 396 // A terminator that isn't a branch can't easily be handled by this 397 // analysis. 398 if (!I->isBranch()) 399 return true; 400 401 // Can't handle indirect branches. 402 SystemZII::Branch Branch(getBranchInfo(*I)); 403 if (!Branch.Target->isMBB()) 404 return true; 405 406 // Punt on compound branches. 407 if (Branch.Type != SystemZII::BranchNormal) 408 return true; 409 410 if (Branch.CCMask == SystemZ::CCMASK_ANY) { 411 // Handle unconditional branches. 412 if (!AllowModify) { 413 TBB = Branch.Target->getMBB(); 414 continue; 415 } 416 417 // If the block has any instructions after a JMP, delete them. 418 while (std::next(I) != MBB.end()) 419 std::next(I)->eraseFromParent(); 420 421 Cond.clear(); 422 FBB = nullptr; 423 424 // Delete the JMP if it's equivalent to a fall-through. 425 if (MBB.isLayoutSuccessor(Branch.Target->getMBB())) { 426 TBB = nullptr; 427 I->eraseFromParent(); 428 I = MBB.end(); 429 continue; 430 } 431 432 // TBB is used to indicate the unconditinal destination. 433 TBB = Branch.Target->getMBB(); 434 continue; 435 } 436 437 // Working from the bottom, handle the first conditional branch. 438 if (Cond.empty()) { 439 // FIXME: add X86-style branch swap 440 FBB = TBB; 441 TBB = Branch.Target->getMBB(); 442 Cond.push_back(MachineOperand::CreateImm(Branch.CCValid)); 443 Cond.push_back(MachineOperand::CreateImm(Branch.CCMask)); 444 continue; 445 } 446 447 // Handle subsequent conditional branches. 448 assert(Cond.size() == 2 && TBB && "Should have seen a conditional branch"); 449 450 // Only handle the case where all conditional branches branch to the same 451 // destination. 452 if (TBB != Branch.Target->getMBB()) 453 return true; 454 455 // If the conditions are the same, we can leave them alone. 456 unsigned OldCCValid = Cond[0].getImm(); 457 unsigned OldCCMask = Cond[1].getImm(); 458 if (OldCCValid == Branch.CCValid && OldCCMask == Branch.CCMask) 459 continue; 460 461 // FIXME: Try combining conditions like X86 does. Should be easy on Z! 462 return false; 463 } 464 465 return false; 466 } 467 468 unsigned SystemZInstrInfo::removeBranch(MachineBasicBlock &MBB, 469 int *BytesRemoved) const { 470 assert(!BytesRemoved && "code size not handled"); 471 472 // Most of the code and comments here are boilerplate. 473 MachineBasicBlock::iterator I = MBB.end(); 474 unsigned Count = 0; 475 476 while (I != MBB.begin()) { 477 --I; 478 if (I->isDebugValue()) 479 continue; 480 if (!I->isBranch()) 481 break; 482 if (!getBranchInfo(*I).Target->isMBB()) 483 break; 484 // Remove the branch. 485 I->eraseFromParent(); 486 I = MBB.end(); 487 ++Count; 488 } 489 490 return Count; 491 } 492 493 bool SystemZInstrInfo:: 494 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 495 assert(Cond.size() == 2 && "Invalid condition"); 496 Cond[1].setImm(Cond[1].getImm() ^ Cond[0].getImm()); 497 return false; 498 } 499 500 unsigned SystemZInstrInfo::insertBranch(MachineBasicBlock &MBB, 501 MachineBasicBlock *TBB, 502 MachineBasicBlock *FBB, 503 ArrayRef<MachineOperand> Cond, 504 const DebugLoc &DL, 505 int *BytesAdded) const { 506 // In this function we output 32-bit branches, which should always 507 // have enough range. They can be shortened and relaxed by later code 508 // in the pipeline, if desired. 509 510 // Shouldn't be a fall through. 511 assert(TBB && "insertBranch must not be told to insert a fallthrough"); 512 assert((Cond.size() == 2 || Cond.size() == 0) && 513 "SystemZ branch conditions have one component!"); 514 assert(!BytesAdded && "code size not handled"); 515 516 if (Cond.empty()) { 517 // Unconditional branch? 518 assert(!FBB && "Unconditional branch with multiple successors!"); 519 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(TBB); 520 return 1; 521 } 522 523 // Conditional branch. 524 unsigned Count = 0; 525 unsigned CCValid = Cond[0].getImm(); 526 unsigned CCMask = Cond[1].getImm(); 527 BuildMI(&MBB, DL, get(SystemZ::BRC)) 528 .addImm(CCValid).addImm(CCMask).addMBB(TBB); 529 ++Count; 530 531 if (FBB) { 532 // Two-way Conditional branch. Insert the second branch. 533 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(FBB); 534 ++Count; 535 } 536 return Count; 537 } 538 539 bool SystemZInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg, 540 unsigned &SrcReg2, int &Mask, 541 int &Value) const { 542 assert(MI.isCompare() && "Caller should have checked for a comparison"); 543 544 if (MI.getNumExplicitOperands() == 2 && MI.getOperand(0).isReg() && 545 MI.getOperand(1).isImm()) { 546 SrcReg = MI.getOperand(0).getReg(); 547 SrcReg2 = 0; 548 Value = MI.getOperand(1).getImm(); 549 Mask = ~0; 550 return true; 551 } 552 553 return false; 554 } 555 556 // If Reg is a virtual register, return its definition, otherwise return null. 557 static MachineInstr *getDef(unsigned Reg, 558 const MachineRegisterInfo *MRI) { 559 if (TargetRegisterInfo::isPhysicalRegister(Reg)) 560 return nullptr; 561 return MRI->getUniqueVRegDef(Reg); 562 } 563 564 // Return true if MI is a shift of type Opcode by Imm bits. 565 static bool isShift(MachineInstr *MI, unsigned Opcode, int64_t Imm) { 566 return (MI->getOpcode() == Opcode && 567 !MI->getOperand(2).getReg() && 568 MI->getOperand(3).getImm() == Imm); 569 } 570 571 // If the destination of MI has no uses, delete it as dead. 572 static void eraseIfDead(MachineInstr *MI, const MachineRegisterInfo *MRI) { 573 if (MRI->use_nodbg_empty(MI->getOperand(0).getReg())) 574 MI->eraseFromParent(); 575 } 576 577 // Compare compares SrcReg against zero. Check whether SrcReg contains 578 // the result of an IPM sequence whose input CC survives until Compare, 579 // and whether Compare is therefore redundant. Delete it and return 580 // true if so. 581 static bool removeIPMBasedCompare(MachineInstr &Compare, unsigned SrcReg, 582 const MachineRegisterInfo *MRI, 583 const TargetRegisterInfo *TRI) { 584 MachineInstr *LGFR = nullptr; 585 MachineInstr *RLL = getDef(SrcReg, MRI); 586 if (RLL && RLL->getOpcode() == SystemZ::LGFR) { 587 LGFR = RLL; 588 RLL = getDef(LGFR->getOperand(1).getReg(), MRI); 589 } 590 if (!RLL || !isShift(RLL, SystemZ::RLL, 31)) 591 return false; 592 593 MachineInstr *SRL = getDef(RLL->getOperand(1).getReg(), MRI); 594 if (!SRL || !isShift(SRL, SystemZ::SRL, SystemZ::IPM_CC)) 595 return false; 596 597 MachineInstr *IPM = getDef(SRL->getOperand(1).getReg(), MRI); 598 if (!IPM || IPM->getOpcode() != SystemZ::IPM) 599 return false; 600 601 // Check that there are no assignments to CC between the IPM and Compare, 602 if (IPM->getParent() != Compare.getParent()) 603 return false; 604 MachineBasicBlock::iterator MBBI = IPM, MBBE = Compare.getIterator(); 605 for (++MBBI; MBBI != MBBE; ++MBBI) { 606 MachineInstr &MI = *MBBI; 607 if (MI.modifiesRegister(SystemZ::CC, TRI)) 608 return false; 609 } 610 611 Compare.eraseFromParent(); 612 if (LGFR) 613 eraseIfDead(LGFR, MRI); 614 eraseIfDead(RLL, MRI); 615 eraseIfDead(SRL, MRI); 616 eraseIfDead(IPM, MRI); 617 618 return true; 619 } 620 621 bool SystemZInstrInfo::optimizeCompareInstr( 622 MachineInstr &Compare, unsigned SrcReg, unsigned SrcReg2, int Mask, 623 int Value, const MachineRegisterInfo *MRI) const { 624 assert(!SrcReg2 && "Only optimizing constant comparisons so far"); 625 bool IsLogical = (Compare.getDesc().TSFlags & SystemZII::IsLogical) != 0; 626 return Value == 0 && !IsLogical && 627 removeIPMBasedCompare(Compare, SrcReg, MRI, &RI); 628 } 629 630 bool SystemZInstrInfo::canInsertSelect(const MachineBasicBlock &MBB, 631 ArrayRef<MachineOperand> Pred, 632 unsigned TrueReg, unsigned FalseReg, 633 int &CondCycles, int &TrueCycles, 634 int &FalseCycles) const { 635 // Not all subtargets have LOCR instructions. 636 if (!STI.hasLoadStoreOnCond()) 637 return false; 638 if (Pred.size() != 2) 639 return false; 640 641 // Check register classes. 642 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 643 const TargetRegisterClass *RC = 644 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 645 if (!RC) 646 return false; 647 648 // We have LOCR instructions for 32 and 64 bit general purpose registers. 649 if ((STI.hasLoadStoreOnCond2() && 650 SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) || 651 SystemZ::GR32BitRegClass.hasSubClassEq(RC) || 652 SystemZ::GR64BitRegClass.hasSubClassEq(RC)) { 653 CondCycles = 2; 654 TrueCycles = 2; 655 FalseCycles = 2; 656 return true; 657 } 658 659 // Can't do anything else. 660 return false; 661 } 662 663 void SystemZInstrInfo::insertSelect(MachineBasicBlock &MBB, 664 MachineBasicBlock::iterator I, 665 const DebugLoc &DL, unsigned DstReg, 666 ArrayRef<MachineOperand> Pred, 667 unsigned TrueReg, 668 unsigned FalseReg) const { 669 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 670 const TargetRegisterClass *RC = MRI.getRegClass(DstReg); 671 672 assert(Pred.size() == 2 && "Invalid condition"); 673 unsigned CCValid = Pred[0].getImm(); 674 unsigned CCMask = Pred[1].getImm(); 675 676 unsigned Opc; 677 if (SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) { 678 if (STI.hasLoadStoreOnCond2()) 679 Opc = SystemZ::LOCRMux; 680 else { 681 Opc = SystemZ::LOCR; 682 MRI.constrainRegClass(DstReg, &SystemZ::GR32BitRegClass); 683 unsigned TReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass); 684 unsigned FReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass); 685 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), TReg).addReg(TrueReg); 686 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), FReg).addReg(FalseReg); 687 TrueReg = TReg; 688 FalseReg = FReg; 689 } 690 } else if (SystemZ::GR64BitRegClass.hasSubClassEq(RC)) 691 Opc = SystemZ::LOCGR; 692 else 693 llvm_unreachable("Invalid register class"); 694 695 BuildMI(MBB, I, DL, get(Opc), DstReg) 696 .addReg(FalseReg).addReg(TrueReg) 697 .addImm(CCValid).addImm(CCMask); 698 } 699 700 bool SystemZInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, 701 unsigned Reg, 702 MachineRegisterInfo *MRI) const { 703 unsigned DefOpc = DefMI.getOpcode(); 704 if (DefOpc != SystemZ::LHIMux && DefOpc != SystemZ::LHI && 705 DefOpc != SystemZ::LGHI) 706 return false; 707 if (DefMI.getOperand(0).getReg() != Reg) 708 return false; 709 int32_t ImmVal = (int32_t)DefMI.getOperand(1).getImm(); 710 711 unsigned UseOpc = UseMI.getOpcode(); 712 unsigned NewUseOpc; 713 unsigned UseIdx; 714 int CommuteIdx = -1; 715 switch (UseOpc) { 716 case SystemZ::LOCRMux: 717 if (!STI.hasLoadStoreOnCond2()) 718 return false; 719 NewUseOpc = SystemZ::LOCHIMux; 720 if (UseMI.getOperand(2).getReg() == Reg) 721 UseIdx = 2; 722 else if (UseMI.getOperand(1).getReg() == Reg) 723 UseIdx = 2, CommuteIdx = 1; 724 else 725 return false; 726 break; 727 case SystemZ::LOCGR: 728 if (!STI.hasLoadStoreOnCond2()) 729 return false; 730 NewUseOpc = SystemZ::LOCGHI; 731 if (UseMI.getOperand(2).getReg() == Reg) 732 UseIdx = 2; 733 else if (UseMI.getOperand(1).getReg() == Reg) 734 UseIdx = 2, CommuteIdx = 1; 735 else 736 return false; 737 break; 738 default: 739 return false; 740 } 741 742 if (CommuteIdx != -1) 743 if (!commuteInstruction(UseMI, false, CommuteIdx, UseIdx)) 744 return false; 745 746 bool DeleteDef = MRI->hasOneNonDBGUse(Reg); 747 UseMI.setDesc(get(NewUseOpc)); 748 UseMI.getOperand(UseIdx).ChangeToImmediate(ImmVal); 749 if (DeleteDef) 750 DefMI.eraseFromParent(); 751 752 return true; 753 } 754 755 bool SystemZInstrInfo::isPredicable(const MachineInstr &MI) const { 756 unsigned Opcode = MI.getOpcode(); 757 if (Opcode == SystemZ::Return || 758 Opcode == SystemZ::Trap || 759 Opcode == SystemZ::CallJG || 760 Opcode == SystemZ::CallBR) 761 return true; 762 return false; 763 } 764 765 bool SystemZInstrInfo:: 766 isProfitableToIfCvt(MachineBasicBlock &MBB, 767 unsigned NumCycles, unsigned ExtraPredCycles, 768 BranchProbability Probability) const { 769 // Avoid using conditional returns at the end of a loop (since then 770 // we'd need to emit an unconditional branch to the beginning anyway, 771 // making the loop body longer). This doesn't apply for low-probability 772 // loops (eg. compare-and-swap retry), so just decide based on branch 773 // probability instead of looping structure. 774 // However, since Compare and Trap instructions cost the same as a regular 775 // Compare instruction, we should allow the if conversion to convert this 776 // into a Conditional Compare regardless of the branch probability. 777 if (MBB.getLastNonDebugInstr()->getOpcode() != SystemZ::Trap && 778 MBB.succ_empty() && Probability < BranchProbability(1, 8)) 779 return false; 780 // For now only convert single instructions. 781 return NumCycles == 1; 782 } 783 784 bool SystemZInstrInfo:: 785 isProfitableToIfCvt(MachineBasicBlock &TMBB, 786 unsigned NumCyclesT, unsigned ExtraPredCyclesT, 787 MachineBasicBlock &FMBB, 788 unsigned NumCyclesF, unsigned ExtraPredCyclesF, 789 BranchProbability Probability) const { 790 // For now avoid converting mutually-exclusive cases. 791 return false; 792 } 793 794 bool SystemZInstrInfo:: 795 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles, 796 BranchProbability Probability) const { 797 // For now only duplicate single instructions. 798 return NumCycles == 1; 799 } 800 801 bool SystemZInstrInfo::PredicateInstruction( 802 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const { 803 assert(Pred.size() == 2 && "Invalid condition"); 804 unsigned CCValid = Pred[0].getImm(); 805 unsigned CCMask = Pred[1].getImm(); 806 assert(CCMask > 0 && CCMask < 15 && "Invalid predicate"); 807 unsigned Opcode = MI.getOpcode(); 808 if (Opcode == SystemZ::Trap) { 809 MI.setDesc(get(SystemZ::CondTrap)); 810 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 811 .addImm(CCValid).addImm(CCMask) 812 .addReg(SystemZ::CC, RegState::Implicit); 813 return true; 814 } 815 if (Opcode == SystemZ::Return) { 816 MI.setDesc(get(SystemZ::CondReturn)); 817 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 818 .addImm(CCValid).addImm(CCMask) 819 .addReg(SystemZ::CC, RegState::Implicit); 820 return true; 821 } 822 if (Opcode == SystemZ::CallJG) { 823 MachineOperand FirstOp = MI.getOperand(0); 824 const uint32_t *RegMask = MI.getOperand(1).getRegMask(); 825 MI.RemoveOperand(1); 826 MI.RemoveOperand(0); 827 MI.setDesc(get(SystemZ::CallBRCL)); 828 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 829 .addImm(CCValid) 830 .addImm(CCMask) 831 .add(FirstOp) 832 .addRegMask(RegMask) 833 .addReg(SystemZ::CC, RegState::Implicit); 834 return true; 835 } 836 if (Opcode == SystemZ::CallBR) { 837 const uint32_t *RegMask = MI.getOperand(0).getRegMask(); 838 MI.RemoveOperand(0); 839 MI.setDesc(get(SystemZ::CallBCR)); 840 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 841 .addImm(CCValid).addImm(CCMask) 842 .addRegMask(RegMask) 843 .addReg(SystemZ::CC, RegState::Implicit); 844 return true; 845 } 846 return false; 847 } 848 849 void SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 850 MachineBasicBlock::iterator MBBI, 851 const DebugLoc &DL, unsigned DestReg, 852 unsigned SrcReg, bool KillSrc) const { 853 // Split 128-bit GPR moves into two 64-bit moves. This handles ADDR128 too. 854 if (SystemZ::GR128BitRegClass.contains(DestReg, SrcReg)) { 855 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_h64), 856 RI.getSubReg(SrcReg, SystemZ::subreg_h64), KillSrc); 857 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_l64), 858 RI.getSubReg(SrcReg, SystemZ::subreg_l64), KillSrc); 859 return; 860 } 861 862 if (SystemZ::GRX32BitRegClass.contains(DestReg, SrcReg)) { 863 emitGRX32Move(MBB, MBBI, DL, DestReg, SrcReg, SystemZ::LR, 32, KillSrc, 864 false); 865 return; 866 } 867 868 // Everything else needs only one instruction. 869 unsigned Opcode; 870 if (SystemZ::GR64BitRegClass.contains(DestReg, SrcReg)) 871 Opcode = SystemZ::LGR; 872 else if (SystemZ::FP32BitRegClass.contains(DestReg, SrcReg)) 873 // For z13 we prefer LDR over LER to avoid partial register dependencies. 874 Opcode = STI.hasVector() ? SystemZ::LDR32 : SystemZ::LER; 875 else if (SystemZ::FP64BitRegClass.contains(DestReg, SrcReg)) 876 Opcode = SystemZ::LDR; 877 else if (SystemZ::FP128BitRegClass.contains(DestReg, SrcReg)) 878 Opcode = SystemZ::LXR; 879 else if (SystemZ::VR32BitRegClass.contains(DestReg, SrcReg)) 880 Opcode = SystemZ::VLR32; 881 else if (SystemZ::VR64BitRegClass.contains(DestReg, SrcReg)) 882 Opcode = SystemZ::VLR64; 883 else if (SystemZ::VR128BitRegClass.contains(DestReg, SrcReg)) 884 Opcode = SystemZ::VLR; 885 else if (SystemZ::AR32BitRegClass.contains(DestReg, SrcReg)) 886 Opcode = SystemZ::CPYA; 887 else if (SystemZ::AR32BitRegClass.contains(DestReg) && 888 SystemZ::GR32BitRegClass.contains(SrcReg)) 889 Opcode = SystemZ::SAR; 890 else if (SystemZ::GR32BitRegClass.contains(DestReg) && 891 SystemZ::AR32BitRegClass.contains(SrcReg)) 892 Opcode = SystemZ::EAR; 893 else 894 llvm_unreachable("Impossible reg-to-reg copy"); 895 896 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 897 .addReg(SrcReg, getKillRegState(KillSrc)); 898 } 899 900 void SystemZInstrInfo::storeRegToStackSlot( 901 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned SrcReg, 902 bool isKill, int FrameIdx, const TargetRegisterClass *RC, 903 const TargetRegisterInfo *TRI) const { 904 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 905 906 // Callers may expect a single instruction, so keep 128-bit moves 907 // together for now and lower them after register allocation. 908 unsigned LoadOpcode, StoreOpcode; 909 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 910 addFrameReference(BuildMI(MBB, MBBI, DL, get(StoreOpcode)) 911 .addReg(SrcReg, getKillRegState(isKill)), 912 FrameIdx); 913 } 914 915 void SystemZInstrInfo::loadRegFromStackSlot( 916 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned DestReg, 917 int FrameIdx, const TargetRegisterClass *RC, 918 const TargetRegisterInfo *TRI) const { 919 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 920 921 // Callers may expect a single instruction, so keep 128-bit moves 922 // together for now and lower them after register allocation. 923 unsigned LoadOpcode, StoreOpcode; 924 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 925 addFrameReference(BuildMI(MBB, MBBI, DL, get(LoadOpcode), DestReg), 926 FrameIdx); 927 } 928 929 // Return true if MI is a simple load or store with a 12-bit displacement 930 // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 931 static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) { 932 const MCInstrDesc &MCID = MI->getDesc(); 933 return ((MCID.TSFlags & Flag) && 934 isUInt<12>(MI->getOperand(2).getImm()) && 935 MI->getOperand(3).getReg() == 0); 936 } 937 938 namespace { 939 940 struct LogicOp { 941 LogicOp() = default; 942 LogicOp(unsigned regSize, unsigned immLSB, unsigned immSize) 943 : RegSize(regSize), ImmLSB(immLSB), ImmSize(immSize) {} 944 945 explicit operator bool() const { return RegSize; } 946 947 unsigned RegSize = 0; 948 unsigned ImmLSB = 0; 949 unsigned ImmSize = 0; 950 }; 951 952 } // end anonymous namespace 953 954 static LogicOp interpretAndImmediate(unsigned Opcode) { 955 switch (Opcode) { 956 case SystemZ::NILMux: return LogicOp(32, 0, 16); 957 case SystemZ::NIHMux: return LogicOp(32, 16, 16); 958 case SystemZ::NILL64: return LogicOp(64, 0, 16); 959 case SystemZ::NILH64: return LogicOp(64, 16, 16); 960 case SystemZ::NIHL64: return LogicOp(64, 32, 16); 961 case SystemZ::NIHH64: return LogicOp(64, 48, 16); 962 case SystemZ::NIFMux: return LogicOp(32, 0, 32); 963 case SystemZ::NILF64: return LogicOp(64, 0, 32); 964 case SystemZ::NIHF64: return LogicOp(64, 32, 32); 965 default: return LogicOp(); 966 } 967 } 968 969 static void transferDeadCC(MachineInstr *OldMI, MachineInstr *NewMI) { 970 if (OldMI->registerDefIsDead(SystemZ::CC)) { 971 MachineOperand *CCDef = NewMI->findRegisterDefOperand(SystemZ::CC); 972 if (CCDef != nullptr) 973 CCDef->setIsDead(true); 974 } 975 } 976 977 // Used to return from convertToThreeAddress after replacing two-address 978 // instruction OldMI with three-address instruction NewMI. 979 static MachineInstr *finishConvertToThreeAddress(MachineInstr *OldMI, 980 MachineInstr *NewMI, 981 LiveVariables *LV) { 982 if (LV) { 983 unsigned NumOps = OldMI->getNumOperands(); 984 for (unsigned I = 1; I < NumOps; ++I) { 985 MachineOperand &Op = OldMI->getOperand(I); 986 if (Op.isReg() && Op.isKill()) 987 LV->replaceKillInstruction(Op.getReg(), *OldMI, *NewMI); 988 } 989 } 990 transferDeadCC(OldMI, NewMI); 991 return NewMI; 992 } 993 994 MachineInstr *SystemZInstrInfo::convertToThreeAddress( 995 MachineFunction::iterator &MFI, MachineInstr &MI, LiveVariables *LV) const { 996 MachineBasicBlock *MBB = MI.getParent(); 997 MachineFunction *MF = MBB->getParent(); 998 MachineRegisterInfo &MRI = MF->getRegInfo(); 999 1000 unsigned Opcode = MI.getOpcode(); 1001 unsigned NumOps = MI.getNumOperands(); 1002 1003 // Try to convert something like SLL into SLLK, if supported. 1004 // We prefer to keep the two-operand form where possible both 1005 // because it tends to be shorter and because some instructions 1006 // have memory forms that can be used during spilling. 1007 if (STI.hasDistinctOps()) { 1008 MachineOperand &Dest = MI.getOperand(0); 1009 MachineOperand &Src = MI.getOperand(1); 1010 unsigned DestReg = Dest.getReg(); 1011 unsigned SrcReg = Src.getReg(); 1012 // AHIMux is only really a three-operand instruction when both operands 1013 // are low registers. Try to constrain both operands to be low if 1014 // possible. 1015 if (Opcode == SystemZ::AHIMux && 1016 TargetRegisterInfo::isVirtualRegister(DestReg) && 1017 TargetRegisterInfo::isVirtualRegister(SrcReg) && 1018 MRI.getRegClass(DestReg)->contains(SystemZ::R1L) && 1019 MRI.getRegClass(SrcReg)->contains(SystemZ::R1L)) { 1020 MRI.constrainRegClass(DestReg, &SystemZ::GR32BitRegClass); 1021 MRI.constrainRegClass(SrcReg, &SystemZ::GR32BitRegClass); 1022 } 1023 int ThreeOperandOpcode = SystemZ::getThreeOperandOpcode(Opcode); 1024 if (ThreeOperandOpcode >= 0) { 1025 // Create three address instruction without adding the implicit 1026 // operands. Those will instead be copied over from the original 1027 // instruction by the loop below. 1028 MachineInstrBuilder MIB( 1029 *MF, MF->CreateMachineInstr(get(ThreeOperandOpcode), MI.getDebugLoc(), 1030 /*NoImplicit=*/true)); 1031 MIB.add(Dest); 1032 // Keep the kill state, but drop the tied flag. 1033 MIB.addReg(Src.getReg(), getKillRegState(Src.isKill()), Src.getSubReg()); 1034 // Keep the remaining operands as-is. 1035 for (unsigned I = 2; I < NumOps; ++I) 1036 MIB.add(MI.getOperand(I)); 1037 MBB->insert(MI, MIB); 1038 return finishConvertToThreeAddress(&MI, MIB, LV); 1039 } 1040 } 1041 1042 // Try to convert an AND into an RISBG-type instruction. 1043 if (LogicOp And = interpretAndImmediate(Opcode)) { 1044 uint64_t Imm = MI.getOperand(2).getImm() << And.ImmLSB; 1045 // AND IMMEDIATE leaves the other bits of the register unchanged. 1046 Imm |= allOnes(And.RegSize) & ~(allOnes(And.ImmSize) << And.ImmLSB); 1047 unsigned Start, End; 1048 if (isRxSBGMask(Imm, And.RegSize, Start, End)) { 1049 unsigned NewOpcode; 1050 if (And.RegSize == 64) { 1051 NewOpcode = SystemZ::RISBG; 1052 // Prefer RISBGN if available, since it does not clobber CC. 1053 if (STI.hasMiscellaneousExtensions()) 1054 NewOpcode = SystemZ::RISBGN; 1055 } else { 1056 NewOpcode = SystemZ::RISBMux; 1057 Start &= 31; 1058 End &= 31; 1059 } 1060 MachineOperand &Dest = MI.getOperand(0); 1061 MachineOperand &Src = MI.getOperand(1); 1062 MachineInstrBuilder MIB = 1063 BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpcode)) 1064 .add(Dest) 1065 .addReg(0) 1066 .addReg(Src.getReg(), getKillRegState(Src.isKill()), 1067 Src.getSubReg()) 1068 .addImm(Start) 1069 .addImm(End + 128) 1070 .addImm(0); 1071 return finishConvertToThreeAddress(&MI, MIB, LV); 1072 } 1073 } 1074 return nullptr; 1075 } 1076 1077 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1078 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1079 MachineBasicBlock::iterator InsertPt, int FrameIndex, 1080 LiveIntervals *LIS) const { 1081 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 1082 const MachineFrameInfo &MFI = MF.getFrameInfo(); 1083 unsigned Size = MFI.getObjectSize(FrameIndex); 1084 unsigned Opcode = MI.getOpcode(); 1085 1086 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 1087 if (LIS != nullptr && (Opcode == SystemZ::LA || Opcode == SystemZ::LAY) && 1088 isInt<8>(MI.getOperand(2).getImm()) && !MI.getOperand(3).getReg()) { 1089 1090 // Check CC liveness, since new instruction introduces a dead 1091 // def of CC. 1092 MCRegUnitIterator CCUnit(SystemZ::CC, TRI); 1093 LiveRange &CCLiveRange = LIS->getRegUnit(*CCUnit); 1094 ++CCUnit; 1095 assert(!CCUnit.isValid() && "CC only has one reg unit."); 1096 SlotIndex MISlot = 1097 LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot(); 1098 if (!CCLiveRange.liveAt(MISlot)) { 1099 // LA(Y) %reg, CONST(%reg) -> AGSI %mem, CONST 1100 MachineInstr *BuiltMI = BuildMI(*InsertPt->getParent(), InsertPt, 1101 MI.getDebugLoc(), get(SystemZ::AGSI)) 1102 .addFrameIndex(FrameIndex) 1103 .addImm(0) 1104 .addImm(MI.getOperand(2).getImm()); 1105 BuiltMI->findRegisterDefOperand(SystemZ::CC)->setIsDead(true); 1106 CCLiveRange.createDeadDef(MISlot, LIS->getVNInfoAllocator()); 1107 return BuiltMI; 1108 } 1109 } 1110 return nullptr; 1111 } 1112 1113 // All other cases require a single operand. 1114 if (Ops.size() != 1) 1115 return nullptr; 1116 1117 unsigned OpNum = Ops[0]; 1118 assert(Size * 8 == 1119 TRI->getRegSizeInBits(*MF.getRegInfo() 1120 .getRegClass(MI.getOperand(OpNum).getReg())) && 1121 "Invalid size combination"); 1122 1123 if ((Opcode == SystemZ::AHI || Opcode == SystemZ::AGHI) && OpNum == 0 && 1124 isInt<8>(MI.getOperand(2).getImm())) { 1125 // A(G)HI %reg, CONST -> A(G)SI %mem, CONST 1126 Opcode = (Opcode == SystemZ::AHI ? SystemZ::ASI : SystemZ::AGSI); 1127 MachineInstr *BuiltMI = 1128 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode)) 1129 .addFrameIndex(FrameIndex) 1130 .addImm(0) 1131 .addImm(MI.getOperand(2).getImm()); 1132 transferDeadCC(&MI, BuiltMI); 1133 return BuiltMI; 1134 } 1135 1136 if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) { 1137 bool Op0IsGPR = (Opcode == SystemZ::LGDR); 1138 bool Op1IsGPR = (Opcode == SystemZ::LDGR); 1139 // If we're spilling the destination of an LDGR or LGDR, store the 1140 // source register instead. 1141 if (OpNum == 0) { 1142 unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD; 1143 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1144 get(StoreOpcode)) 1145 .add(MI.getOperand(1)) 1146 .addFrameIndex(FrameIndex) 1147 .addImm(0) 1148 .addReg(0); 1149 } 1150 // If we're spilling the source of an LDGR or LGDR, load the 1151 // destination register instead. 1152 if (OpNum == 1) { 1153 unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD; 1154 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1155 get(LoadOpcode)) 1156 .add(MI.getOperand(0)) 1157 .addFrameIndex(FrameIndex) 1158 .addImm(0) 1159 .addReg(0); 1160 } 1161 } 1162 1163 // Look for cases where the source of a simple store or the destination 1164 // of a simple load is being spilled. Try to use MVC instead. 1165 // 1166 // Although MVC is in practice a fast choice in these cases, it is still 1167 // logically a bytewise copy. This means that we cannot use it if the 1168 // load or store is volatile. We also wouldn't be able to use MVC if 1169 // the two memories partially overlap, but that case cannot occur here, 1170 // because we know that one of the memories is a full frame index. 1171 // 1172 // For performance reasons, we also want to avoid using MVC if the addresses 1173 // might be equal. We don't worry about that case here, because spill slot 1174 // coloring happens later, and because we have special code to remove 1175 // MVCs that turn out to be redundant. 1176 if (OpNum == 0 && MI.hasOneMemOperand()) { 1177 MachineMemOperand *MMO = *MI.memoperands_begin(); 1178 if (MMO->getSize() == Size && !MMO->isVolatile()) { 1179 // Handle conversion of loads. 1180 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXLoad)) { 1181 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1182 get(SystemZ::MVC)) 1183 .addFrameIndex(FrameIndex) 1184 .addImm(0) 1185 .addImm(Size) 1186 .add(MI.getOperand(1)) 1187 .addImm(MI.getOperand(2).getImm()) 1188 .addMemOperand(MMO); 1189 } 1190 // Handle conversion of stores. 1191 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXStore)) { 1192 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1193 get(SystemZ::MVC)) 1194 .add(MI.getOperand(1)) 1195 .addImm(MI.getOperand(2).getImm()) 1196 .addImm(Size) 1197 .addFrameIndex(FrameIndex) 1198 .addImm(0) 1199 .addMemOperand(MMO); 1200 } 1201 } 1202 } 1203 1204 // If the spilled operand is the final one, try to change <INSN>R 1205 // into <INSN>. 1206 int MemOpcode = SystemZ::getMemOpcode(Opcode); 1207 if (MemOpcode >= 0) { 1208 unsigned NumOps = MI.getNumExplicitOperands(); 1209 if (OpNum == NumOps - 1) { 1210 const MCInstrDesc &MemDesc = get(MemOpcode); 1211 uint64_t AccessBytes = SystemZII::getAccessSize(MemDesc.TSFlags); 1212 assert(AccessBytes != 0 && "Size of access should be known"); 1213 assert(AccessBytes <= Size && "Access outside the frame index"); 1214 uint64_t Offset = Size - AccessBytes; 1215 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt, 1216 MI.getDebugLoc(), get(MemOpcode)); 1217 for (unsigned I = 0; I < OpNum; ++I) 1218 MIB.add(MI.getOperand(I)); 1219 MIB.addFrameIndex(FrameIndex).addImm(Offset); 1220 if (MemDesc.TSFlags & SystemZII::HasIndex) 1221 MIB.addReg(0); 1222 transferDeadCC(&MI, MIB); 1223 return MIB; 1224 } 1225 } 1226 1227 return nullptr; 1228 } 1229 1230 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1231 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1232 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, 1233 LiveIntervals *LIS) const { 1234 return nullptr; 1235 } 1236 1237 bool SystemZInstrInfo::expandPostRAPseudo(MachineInstr &MI) const { 1238 switch (MI.getOpcode()) { 1239 case SystemZ::L128: 1240 splitMove(MI, SystemZ::LG); 1241 return true; 1242 1243 case SystemZ::ST128: 1244 splitMove(MI, SystemZ::STG); 1245 return true; 1246 1247 case SystemZ::LX: 1248 splitMove(MI, SystemZ::LD); 1249 return true; 1250 1251 case SystemZ::STX: 1252 splitMove(MI, SystemZ::STD); 1253 return true; 1254 1255 case SystemZ::LBMux: 1256 expandRXYPseudo(MI, SystemZ::LB, SystemZ::LBH); 1257 return true; 1258 1259 case SystemZ::LHMux: 1260 expandRXYPseudo(MI, SystemZ::LH, SystemZ::LHH); 1261 return true; 1262 1263 case SystemZ::LLCRMux: 1264 expandZExtPseudo(MI, SystemZ::LLCR, 8); 1265 return true; 1266 1267 case SystemZ::LLHRMux: 1268 expandZExtPseudo(MI, SystemZ::LLHR, 16); 1269 return true; 1270 1271 case SystemZ::LLCMux: 1272 expandRXYPseudo(MI, SystemZ::LLC, SystemZ::LLCH); 1273 return true; 1274 1275 case SystemZ::LLHMux: 1276 expandRXYPseudo(MI, SystemZ::LLH, SystemZ::LLHH); 1277 return true; 1278 1279 case SystemZ::LMux: 1280 expandRXYPseudo(MI, SystemZ::L, SystemZ::LFH); 1281 return true; 1282 1283 case SystemZ::LOCMux: 1284 expandLOCPseudo(MI, SystemZ::LOC, SystemZ::LOCFH); 1285 return true; 1286 1287 case SystemZ::LOCHIMux: 1288 expandLOCPseudo(MI, SystemZ::LOCHI, SystemZ::LOCHHI); 1289 return true; 1290 1291 case SystemZ::LOCRMux: 1292 expandLOCRPseudo(MI, SystemZ::LOCR, SystemZ::LOCFHR); 1293 return true; 1294 1295 case SystemZ::STCMux: 1296 expandRXYPseudo(MI, SystemZ::STC, SystemZ::STCH); 1297 return true; 1298 1299 case SystemZ::STHMux: 1300 expandRXYPseudo(MI, SystemZ::STH, SystemZ::STHH); 1301 return true; 1302 1303 case SystemZ::STMux: 1304 expandRXYPseudo(MI, SystemZ::ST, SystemZ::STFH); 1305 return true; 1306 1307 case SystemZ::STOCMux: 1308 expandLOCPseudo(MI, SystemZ::STOC, SystemZ::STOCFH); 1309 return true; 1310 1311 case SystemZ::LHIMux: 1312 expandRIPseudo(MI, SystemZ::LHI, SystemZ::IIHF, true); 1313 return true; 1314 1315 case SystemZ::IIFMux: 1316 expandRIPseudo(MI, SystemZ::IILF, SystemZ::IIHF, false); 1317 return true; 1318 1319 case SystemZ::IILMux: 1320 expandRIPseudo(MI, SystemZ::IILL, SystemZ::IIHL, false); 1321 return true; 1322 1323 case SystemZ::IIHMux: 1324 expandRIPseudo(MI, SystemZ::IILH, SystemZ::IIHH, false); 1325 return true; 1326 1327 case SystemZ::NIFMux: 1328 expandRIPseudo(MI, SystemZ::NILF, SystemZ::NIHF, false); 1329 return true; 1330 1331 case SystemZ::NILMux: 1332 expandRIPseudo(MI, SystemZ::NILL, SystemZ::NIHL, false); 1333 return true; 1334 1335 case SystemZ::NIHMux: 1336 expandRIPseudo(MI, SystemZ::NILH, SystemZ::NIHH, false); 1337 return true; 1338 1339 case SystemZ::OIFMux: 1340 expandRIPseudo(MI, SystemZ::OILF, SystemZ::OIHF, false); 1341 return true; 1342 1343 case SystemZ::OILMux: 1344 expandRIPseudo(MI, SystemZ::OILL, SystemZ::OIHL, false); 1345 return true; 1346 1347 case SystemZ::OIHMux: 1348 expandRIPseudo(MI, SystemZ::OILH, SystemZ::OIHH, false); 1349 return true; 1350 1351 case SystemZ::XIFMux: 1352 expandRIPseudo(MI, SystemZ::XILF, SystemZ::XIHF, false); 1353 return true; 1354 1355 case SystemZ::TMLMux: 1356 expandRIPseudo(MI, SystemZ::TMLL, SystemZ::TMHL, false); 1357 return true; 1358 1359 case SystemZ::TMHMux: 1360 expandRIPseudo(MI, SystemZ::TMLH, SystemZ::TMHH, false); 1361 return true; 1362 1363 case SystemZ::AHIMux: 1364 expandRIPseudo(MI, SystemZ::AHI, SystemZ::AIH, false); 1365 return true; 1366 1367 case SystemZ::AHIMuxK: 1368 expandRIEPseudo(MI, SystemZ::AHI, SystemZ::AHIK, SystemZ::AIH); 1369 return true; 1370 1371 case SystemZ::AFIMux: 1372 expandRIPseudo(MI, SystemZ::AFI, SystemZ::AIH, false); 1373 return true; 1374 1375 case SystemZ::CHIMux: 1376 expandRIPseudo(MI, SystemZ::CHI, SystemZ::CIH, false); 1377 return true; 1378 1379 case SystemZ::CFIMux: 1380 expandRIPseudo(MI, SystemZ::CFI, SystemZ::CIH, false); 1381 return true; 1382 1383 case SystemZ::CLFIMux: 1384 expandRIPseudo(MI, SystemZ::CLFI, SystemZ::CLIH, false); 1385 return true; 1386 1387 case SystemZ::CMux: 1388 expandRXYPseudo(MI, SystemZ::C, SystemZ::CHF); 1389 return true; 1390 1391 case SystemZ::CLMux: 1392 expandRXYPseudo(MI, SystemZ::CL, SystemZ::CLHF); 1393 return true; 1394 1395 case SystemZ::RISBMux: { 1396 bool DestIsHigh = isHighReg(MI.getOperand(0).getReg()); 1397 bool SrcIsHigh = isHighReg(MI.getOperand(2).getReg()); 1398 if (SrcIsHigh == DestIsHigh) 1399 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHH : SystemZ::RISBLL)); 1400 else { 1401 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHL : SystemZ::RISBLH)); 1402 MI.getOperand(5).setImm(MI.getOperand(5).getImm() ^ 32); 1403 } 1404 return true; 1405 } 1406 1407 case SystemZ::ADJDYNALLOC: 1408 splitAdjDynAlloc(MI); 1409 return true; 1410 1411 case TargetOpcode::LOAD_STACK_GUARD: 1412 expandLoadStackGuard(&MI); 1413 return true; 1414 1415 default: 1416 return false; 1417 } 1418 } 1419 1420 unsigned SystemZInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { 1421 if (MI.getOpcode() == TargetOpcode::INLINEASM) { 1422 const MachineFunction *MF = MI.getParent()->getParent(); 1423 const char *AsmStr = MI.getOperand(0).getSymbolName(); 1424 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo()); 1425 } 1426 return MI.getDesc().getSize(); 1427 } 1428 1429 SystemZII::Branch 1430 SystemZInstrInfo::getBranchInfo(const MachineInstr &MI) const { 1431 switch (MI.getOpcode()) { 1432 case SystemZ::BR: 1433 case SystemZ::J: 1434 case SystemZ::JG: 1435 return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY, 1436 SystemZ::CCMASK_ANY, &MI.getOperand(0)); 1437 1438 case SystemZ::BRC: 1439 case SystemZ::BRCL: 1440 return SystemZII::Branch(SystemZII::BranchNormal, MI.getOperand(0).getImm(), 1441 MI.getOperand(1).getImm(), &MI.getOperand(2)); 1442 1443 case SystemZ::BRCT: 1444 case SystemZ::BRCTH: 1445 return SystemZII::Branch(SystemZII::BranchCT, SystemZ::CCMASK_ICMP, 1446 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1447 1448 case SystemZ::BRCTG: 1449 return SystemZII::Branch(SystemZII::BranchCTG, SystemZ::CCMASK_ICMP, 1450 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1451 1452 case SystemZ::CIJ: 1453 case SystemZ::CRJ: 1454 return SystemZII::Branch(SystemZII::BranchC, SystemZ::CCMASK_ICMP, 1455 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1456 1457 case SystemZ::CLIJ: 1458 case SystemZ::CLRJ: 1459 return SystemZII::Branch(SystemZII::BranchCL, SystemZ::CCMASK_ICMP, 1460 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1461 1462 case SystemZ::CGIJ: 1463 case SystemZ::CGRJ: 1464 return SystemZII::Branch(SystemZII::BranchCG, SystemZ::CCMASK_ICMP, 1465 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1466 1467 case SystemZ::CLGIJ: 1468 case SystemZ::CLGRJ: 1469 return SystemZII::Branch(SystemZII::BranchCLG, SystemZ::CCMASK_ICMP, 1470 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1471 1472 default: 1473 llvm_unreachable("Unrecognized branch opcode"); 1474 } 1475 } 1476 1477 void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC, 1478 unsigned &LoadOpcode, 1479 unsigned &StoreOpcode) const { 1480 if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) { 1481 LoadOpcode = SystemZ::L; 1482 StoreOpcode = SystemZ::ST; 1483 } else if (RC == &SystemZ::GRH32BitRegClass) { 1484 LoadOpcode = SystemZ::LFH; 1485 StoreOpcode = SystemZ::STFH; 1486 } else if (RC == &SystemZ::GRX32BitRegClass) { 1487 LoadOpcode = SystemZ::LMux; 1488 StoreOpcode = SystemZ::STMux; 1489 } else if (RC == &SystemZ::GR64BitRegClass || 1490 RC == &SystemZ::ADDR64BitRegClass) { 1491 LoadOpcode = SystemZ::LG; 1492 StoreOpcode = SystemZ::STG; 1493 } else if (RC == &SystemZ::GR128BitRegClass || 1494 RC == &SystemZ::ADDR128BitRegClass) { 1495 LoadOpcode = SystemZ::L128; 1496 StoreOpcode = SystemZ::ST128; 1497 } else if (RC == &SystemZ::FP32BitRegClass) { 1498 LoadOpcode = SystemZ::LE; 1499 StoreOpcode = SystemZ::STE; 1500 } else if (RC == &SystemZ::FP64BitRegClass) { 1501 LoadOpcode = SystemZ::LD; 1502 StoreOpcode = SystemZ::STD; 1503 } else if (RC == &SystemZ::FP128BitRegClass) { 1504 LoadOpcode = SystemZ::LX; 1505 StoreOpcode = SystemZ::STX; 1506 } else if (RC == &SystemZ::VR32BitRegClass) { 1507 LoadOpcode = SystemZ::VL32; 1508 StoreOpcode = SystemZ::VST32; 1509 } else if (RC == &SystemZ::VR64BitRegClass) { 1510 LoadOpcode = SystemZ::VL64; 1511 StoreOpcode = SystemZ::VST64; 1512 } else if (RC == &SystemZ::VF128BitRegClass || 1513 RC == &SystemZ::VR128BitRegClass) { 1514 LoadOpcode = SystemZ::VL; 1515 StoreOpcode = SystemZ::VST; 1516 } else 1517 llvm_unreachable("Unsupported regclass to load or store"); 1518 } 1519 1520 unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode, 1521 int64_t Offset) const { 1522 const MCInstrDesc &MCID = get(Opcode); 1523 int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset); 1524 if (isUInt<12>(Offset) && isUInt<12>(Offset2)) { 1525 // Get the instruction to use for unsigned 12-bit displacements. 1526 int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode); 1527 if (Disp12Opcode >= 0) 1528 return Disp12Opcode; 1529 1530 // All address-related instructions can use unsigned 12-bit 1531 // displacements. 1532 return Opcode; 1533 } 1534 if (isInt<20>(Offset) && isInt<20>(Offset2)) { 1535 // Get the instruction to use for signed 20-bit displacements. 1536 int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode); 1537 if (Disp20Opcode >= 0) 1538 return Disp20Opcode; 1539 1540 // Check whether Opcode allows signed 20-bit displacements. 1541 if (MCID.TSFlags & SystemZII::Has20BitOffset) 1542 return Opcode; 1543 } 1544 return 0; 1545 } 1546 1547 unsigned SystemZInstrInfo::getLoadAndTest(unsigned Opcode) const { 1548 switch (Opcode) { 1549 case SystemZ::L: return SystemZ::LT; 1550 case SystemZ::LY: return SystemZ::LT; 1551 case SystemZ::LG: return SystemZ::LTG; 1552 case SystemZ::LGF: return SystemZ::LTGF; 1553 case SystemZ::LR: return SystemZ::LTR; 1554 case SystemZ::LGFR: return SystemZ::LTGFR; 1555 case SystemZ::LGR: return SystemZ::LTGR; 1556 case SystemZ::LER: return SystemZ::LTEBR; 1557 case SystemZ::LDR: return SystemZ::LTDBR; 1558 case SystemZ::LXR: return SystemZ::LTXBR; 1559 case SystemZ::LCDFR: return SystemZ::LCDBR; 1560 case SystemZ::LPDFR: return SystemZ::LPDBR; 1561 case SystemZ::LNDFR: return SystemZ::LNDBR; 1562 case SystemZ::LCDFR_32: return SystemZ::LCEBR; 1563 case SystemZ::LPDFR_32: return SystemZ::LPEBR; 1564 case SystemZ::LNDFR_32: return SystemZ::LNEBR; 1565 // On zEC12 we prefer to use RISBGN. But if there is a chance to 1566 // actually use the condition code, we may turn it back into RISGB. 1567 // Note that RISBG is not really a "load-and-test" instruction, 1568 // but sets the same condition code values, so is OK to use here. 1569 case SystemZ::RISBGN: return SystemZ::RISBG; 1570 default: return 0; 1571 } 1572 } 1573 1574 // Return true if Mask matches the regexp 0*1+0*, given that zero masks 1575 // have already been filtered out. Store the first set bit in LSB and 1576 // the number of set bits in Length if so. 1577 static bool isStringOfOnes(uint64_t Mask, unsigned &LSB, unsigned &Length) { 1578 unsigned First = findFirstSet(Mask); 1579 uint64_t Top = (Mask >> First) + 1; 1580 if ((Top & -Top) == Top) { 1581 LSB = First; 1582 Length = findFirstSet(Top); 1583 return true; 1584 } 1585 return false; 1586 } 1587 1588 bool SystemZInstrInfo::isRxSBGMask(uint64_t Mask, unsigned BitSize, 1589 unsigned &Start, unsigned &End) const { 1590 // Reject trivial all-zero masks. 1591 Mask &= allOnes(BitSize); 1592 if (Mask == 0) 1593 return false; 1594 1595 // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of 1596 // the msb and End specifies the index of the lsb. 1597 unsigned LSB, Length; 1598 if (isStringOfOnes(Mask, LSB, Length)) { 1599 Start = 63 - (LSB + Length - 1); 1600 End = 63 - LSB; 1601 return true; 1602 } 1603 1604 // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb 1605 // of the low 1s and End specifies the lsb of the high 1s. 1606 if (isStringOfOnes(Mask ^ allOnes(BitSize), LSB, Length)) { 1607 assert(LSB > 0 && "Bottom bit must be set"); 1608 assert(LSB + Length < BitSize && "Top bit must be set"); 1609 Start = 63 - (LSB - 1); 1610 End = 63 - (LSB + Length); 1611 return true; 1612 } 1613 1614 return false; 1615 } 1616 1617 unsigned SystemZInstrInfo::getFusedCompare(unsigned Opcode, 1618 SystemZII::FusedCompareType Type, 1619 const MachineInstr *MI) const { 1620 switch (Opcode) { 1621 case SystemZ::CHI: 1622 case SystemZ::CGHI: 1623 if (!(MI && isInt<8>(MI->getOperand(1).getImm()))) 1624 return 0; 1625 break; 1626 case SystemZ::CLFI: 1627 case SystemZ::CLGFI: 1628 if (!(MI && isUInt<8>(MI->getOperand(1).getImm()))) 1629 return 0; 1630 break; 1631 case SystemZ::CL: 1632 case SystemZ::CLG: 1633 if (!STI.hasMiscellaneousExtensions()) 1634 return 0; 1635 if (!(MI && MI->getOperand(3).getReg() == 0)) 1636 return 0; 1637 break; 1638 } 1639 switch (Type) { 1640 case SystemZII::CompareAndBranch: 1641 switch (Opcode) { 1642 case SystemZ::CR: 1643 return SystemZ::CRJ; 1644 case SystemZ::CGR: 1645 return SystemZ::CGRJ; 1646 case SystemZ::CHI: 1647 return SystemZ::CIJ; 1648 case SystemZ::CGHI: 1649 return SystemZ::CGIJ; 1650 case SystemZ::CLR: 1651 return SystemZ::CLRJ; 1652 case SystemZ::CLGR: 1653 return SystemZ::CLGRJ; 1654 case SystemZ::CLFI: 1655 return SystemZ::CLIJ; 1656 case SystemZ::CLGFI: 1657 return SystemZ::CLGIJ; 1658 default: 1659 return 0; 1660 } 1661 case SystemZII::CompareAndReturn: 1662 switch (Opcode) { 1663 case SystemZ::CR: 1664 return SystemZ::CRBReturn; 1665 case SystemZ::CGR: 1666 return SystemZ::CGRBReturn; 1667 case SystemZ::CHI: 1668 return SystemZ::CIBReturn; 1669 case SystemZ::CGHI: 1670 return SystemZ::CGIBReturn; 1671 case SystemZ::CLR: 1672 return SystemZ::CLRBReturn; 1673 case SystemZ::CLGR: 1674 return SystemZ::CLGRBReturn; 1675 case SystemZ::CLFI: 1676 return SystemZ::CLIBReturn; 1677 case SystemZ::CLGFI: 1678 return SystemZ::CLGIBReturn; 1679 default: 1680 return 0; 1681 } 1682 case SystemZII::CompareAndSibcall: 1683 switch (Opcode) { 1684 case SystemZ::CR: 1685 return SystemZ::CRBCall; 1686 case SystemZ::CGR: 1687 return SystemZ::CGRBCall; 1688 case SystemZ::CHI: 1689 return SystemZ::CIBCall; 1690 case SystemZ::CGHI: 1691 return SystemZ::CGIBCall; 1692 case SystemZ::CLR: 1693 return SystemZ::CLRBCall; 1694 case SystemZ::CLGR: 1695 return SystemZ::CLGRBCall; 1696 case SystemZ::CLFI: 1697 return SystemZ::CLIBCall; 1698 case SystemZ::CLGFI: 1699 return SystemZ::CLGIBCall; 1700 default: 1701 return 0; 1702 } 1703 case SystemZII::CompareAndTrap: 1704 switch (Opcode) { 1705 case SystemZ::CR: 1706 return SystemZ::CRT; 1707 case SystemZ::CGR: 1708 return SystemZ::CGRT; 1709 case SystemZ::CHI: 1710 return SystemZ::CIT; 1711 case SystemZ::CGHI: 1712 return SystemZ::CGIT; 1713 case SystemZ::CLR: 1714 return SystemZ::CLRT; 1715 case SystemZ::CLGR: 1716 return SystemZ::CLGRT; 1717 case SystemZ::CLFI: 1718 return SystemZ::CLFIT; 1719 case SystemZ::CLGFI: 1720 return SystemZ::CLGIT; 1721 case SystemZ::CL: 1722 return SystemZ::CLT; 1723 case SystemZ::CLG: 1724 return SystemZ::CLGT; 1725 default: 1726 return 0; 1727 } 1728 } 1729 return 0; 1730 } 1731 1732 unsigned SystemZInstrInfo::getLoadAndTrap(unsigned Opcode) const { 1733 if (!STI.hasLoadAndTrap()) 1734 return 0; 1735 switch (Opcode) { 1736 case SystemZ::L: 1737 case SystemZ::LY: 1738 return SystemZ::LAT; 1739 case SystemZ::LG: 1740 return SystemZ::LGAT; 1741 case SystemZ::LFH: 1742 return SystemZ::LFHAT; 1743 case SystemZ::LLGF: 1744 return SystemZ::LLGFAT; 1745 case SystemZ::LLGT: 1746 return SystemZ::LLGTAT; 1747 } 1748 return 0; 1749 } 1750 1751 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB, 1752 MachineBasicBlock::iterator MBBI, 1753 unsigned Reg, uint64_t Value) const { 1754 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 1755 unsigned Opcode; 1756 if (isInt<16>(Value)) 1757 Opcode = SystemZ::LGHI; 1758 else if (SystemZ::isImmLL(Value)) 1759 Opcode = SystemZ::LLILL; 1760 else if (SystemZ::isImmLH(Value)) { 1761 Opcode = SystemZ::LLILH; 1762 Value >>= 16; 1763 } else { 1764 assert(isInt<32>(Value) && "Huge values not handled yet"); 1765 Opcode = SystemZ::LGFI; 1766 } 1767 BuildMI(MBB, MBBI, DL, get(Opcode), Reg).addImm(Value); 1768 } 1769 1770 bool SystemZInstrInfo:: 1771 areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb, 1772 AliasAnalysis *AA) const { 1773 1774 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand()) 1775 return false; 1776 1777 // If mem-operands show that the same address Value is used by both 1778 // instructions, check for non-overlapping offsets and widths. Not 1779 // sure if a register based analysis would be an improvement... 1780 1781 MachineMemOperand *MMOa = *MIa.memoperands_begin(); 1782 MachineMemOperand *MMOb = *MIb.memoperands_begin(); 1783 const Value *VALa = MMOa->getValue(); 1784 const Value *VALb = MMOb->getValue(); 1785 bool SameVal = (VALa && VALb && (VALa == VALb)); 1786 if (!SameVal) { 1787 const PseudoSourceValue *PSVa = MMOa->getPseudoValue(); 1788 const PseudoSourceValue *PSVb = MMOb->getPseudoValue(); 1789 if (PSVa && PSVb && (PSVa == PSVb)) 1790 SameVal = true; 1791 } 1792 if (SameVal) { 1793 int OffsetA = MMOa->getOffset(), OffsetB = MMOb->getOffset(); 1794 int WidthA = MMOa->getSize(), WidthB = MMOb->getSize(); 1795 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB; 1796 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA; 1797 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; 1798 if (LowOffset + LowWidth <= HighOffset) 1799 return true; 1800 } 1801 1802 return false; 1803 } 1804