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 "SystemZInstrInfo.h" 15 #include "MCTargetDesc/SystemZMCTargetDesc.h" 16 #include "SystemZ.h" 17 #include "SystemZInstrBuilder.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 Reg64 = MI->getOperand(0).getReg(); 240 const unsigned Reg32 = RI.getSubReg(Reg64, SystemZ::subreg_l32); 241 242 // EAR can only load the low subregister so us a shift for %a0 to produce 243 // the GR containing %a0 and %a1. 244 245 // ear <reg>, %a0 246 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::EAR), Reg32) 247 .addReg(SystemZ::A0) 248 .addReg(Reg64, RegState::ImplicitDefine); 249 250 // sllg <reg>, <reg>, 32 251 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::SLLG), Reg64) 252 .addReg(Reg64) 253 .addReg(0) 254 .addImm(32); 255 256 // ear <reg>, %a1 257 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::EAR), Reg32) 258 .addReg(SystemZ::A1); 259 260 // lg <reg>, 40(<reg>) 261 MI->setDesc(get(SystemZ::LG)); 262 MachineInstrBuilder(MF, MI).addReg(Reg64).addImm(40).addReg(0); 263 } 264 265 // Emit a zero-extending move from 32-bit GPR SrcReg to 32-bit GPR 266 // DestReg before MBBI in MBB. Use LowLowOpcode when both DestReg and SrcReg 267 // are low registers, otherwise use RISB[LH]G. Size is the number of bits 268 // taken from the low end of SrcReg (8 for LLCR, 16 for LLHR and 32 for LR). 269 // KillSrc is true if this move is the last use of SrcReg. 270 MachineInstrBuilder 271 SystemZInstrInfo::emitGRX32Move(MachineBasicBlock &MBB, 272 MachineBasicBlock::iterator MBBI, 273 const DebugLoc &DL, unsigned DestReg, 274 unsigned SrcReg, unsigned LowLowOpcode, 275 unsigned Size, bool KillSrc, 276 bool UndefSrc) const { 277 unsigned Opcode; 278 bool DestIsHigh = isHighReg(DestReg); 279 bool SrcIsHigh = isHighReg(SrcReg); 280 if (DestIsHigh && SrcIsHigh) 281 Opcode = SystemZ::RISBHH; 282 else if (DestIsHigh && !SrcIsHigh) 283 Opcode = SystemZ::RISBHL; 284 else if (!DestIsHigh && SrcIsHigh) 285 Opcode = SystemZ::RISBLH; 286 else { 287 return BuildMI(MBB, MBBI, DL, get(LowLowOpcode), DestReg) 288 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)); 289 } 290 unsigned Rotate = (DestIsHigh != SrcIsHigh ? 32 : 0); 291 return BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 292 .addReg(DestReg, RegState::Undef) 293 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)) 294 .addImm(32 - Size).addImm(128 + 31).addImm(Rotate); 295 } 296 297 MachineInstr *SystemZInstrInfo::commuteInstructionImpl(MachineInstr &MI, 298 bool NewMI, 299 unsigned OpIdx1, 300 unsigned OpIdx2) const { 301 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & { 302 if (NewMI) 303 return *MI.getParent()->getParent()->CloneMachineInstr(&MI); 304 return MI; 305 }; 306 307 switch (MI.getOpcode()) { 308 case SystemZ::LOCRMux: 309 case SystemZ::LOCFHR: 310 case SystemZ::LOCR: 311 case SystemZ::LOCGR: { 312 auto &WorkingMI = cloneIfNew(MI); 313 // Invert condition. 314 unsigned CCValid = WorkingMI.getOperand(3).getImm(); 315 unsigned CCMask = WorkingMI.getOperand(4).getImm(); 316 WorkingMI.getOperand(4).setImm(CCMask ^ CCValid); 317 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 318 OpIdx1, OpIdx2); 319 } 320 default: 321 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 322 } 323 } 324 325 // If MI is a simple load or store for a frame object, return the register 326 // it loads or stores and set FrameIndex to the index of the frame object. 327 // Return 0 otherwise. 328 // 329 // Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 330 static int isSimpleMove(const MachineInstr &MI, int &FrameIndex, 331 unsigned Flag) { 332 const MCInstrDesc &MCID = MI.getDesc(); 333 if ((MCID.TSFlags & Flag) && MI.getOperand(1).isFI() && 334 MI.getOperand(2).getImm() == 0 && MI.getOperand(3).getReg() == 0) { 335 FrameIndex = MI.getOperand(1).getIndex(); 336 return MI.getOperand(0).getReg(); 337 } 338 return 0; 339 } 340 341 unsigned SystemZInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 342 int &FrameIndex) const { 343 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXLoad); 344 } 345 346 unsigned SystemZInstrInfo::isStoreToStackSlot(const MachineInstr &MI, 347 int &FrameIndex) const { 348 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXStore); 349 } 350 351 bool SystemZInstrInfo::isStackSlotCopy(const MachineInstr &MI, 352 int &DestFrameIndex, 353 int &SrcFrameIndex) const { 354 // Check for MVC 0(Length,FI1),0(FI2) 355 const MachineFrameInfo &MFI = MI.getParent()->getParent()->getFrameInfo(); 356 if (MI.getOpcode() != SystemZ::MVC || !MI.getOperand(0).isFI() || 357 MI.getOperand(1).getImm() != 0 || !MI.getOperand(3).isFI() || 358 MI.getOperand(4).getImm() != 0) 359 return false; 360 361 // Check that Length covers the full slots. 362 int64_t Length = MI.getOperand(2).getImm(); 363 unsigned FI1 = MI.getOperand(0).getIndex(); 364 unsigned FI2 = MI.getOperand(3).getIndex(); 365 if (MFI.getObjectSize(FI1) != Length || 366 MFI.getObjectSize(FI2) != Length) 367 return false; 368 369 DestFrameIndex = FI1; 370 SrcFrameIndex = FI2; 371 return true; 372 } 373 374 bool SystemZInstrInfo::analyzeBranch(MachineBasicBlock &MBB, 375 MachineBasicBlock *&TBB, 376 MachineBasicBlock *&FBB, 377 SmallVectorImpl<MachineOperand> &Cond, 378 bool AllowModify) const { 379 // Most of the code and comments here are boilerplate. 380 381 // Start from the bottom of the block and work up, examining the 382 // terminator instructions. 383 MachineBasicBlock::iterator I = MBB.end(); 384 while (I != MBB.begin()) { 385 --I; 386 if (I->isDebugValue()) 387 continue; 388 389 // Working from the bottom, when we see a non-terminator instruction, we're 390 // done. 391 if (!isUnpredicatedTerminator(*I)) 392 break; 393 394 // A terminator that isn't a branch can't easily be handled by this 395 // analysis. 396 if (!I->isBranch()) 397 return true; 398 399 // Can't handle indirect branches. 400 SystemZII::Branch Branch(getBranchInfo(*I)); 401 if (!Branch.Target->isMBB()) 402 return true; 403 404 // Punt on compound branches. 405 if (Branch.Type != SystemZII::BranchNormal) 406 return true; 407 408 if (Branch.CCMask == SystemZ::CCMASK_ANY) { 409 // Handle unconditional branches. 410 if (!AllowModify) { 411 TBB = Branch.Target->getMBB(); 412 continue; 413 } 414 415 // If the block has any instructions after a JMP, delete them. 416 while (std::next(I) != MBB.end()) 417 std::next(I)->eraseFromParent(); 418 419 Cond.clear(); 420 FBB = nullptr; 421 422 // Delete the JMP if it's equivalent to a fall-through. 423 if (MBB.isLayoutSuccessor(Branch.Target->getMBB())) { 424 TBB = nullptr; 425 I->eraseFromParent(); 426 I = MBB.end(); 427 continue; 428 } 429 430 // TBB is used to indicate the unconditinal destination. 431 TBB = Branch.Target->getMBB(); 432 continue; 433 } 434 435 // Working from the bottom, handle the first conditional branch. 436 if (Cond.empty()) { 437 // FIXME: add X86-style branch swap 438 FBB = TBB; 439 TBB = Branch.Target->getMBB(); 440 Cond.push_back(MachineOperand::CreateImm(Branch.CCValid)); 441 Cond.push_back(MachineOperand::CreateImm(Branch.CCMask)); 442 continue; 443 } 444 445 // Handle subsequent conditional branches. 446 assert(Cond.size() == 2 && TBB && "Should have seen a conditional branch"); 447 448 // Only handle the case where all conditional branches branch to the same 449 // destination. 450 if (TBB != Branch.Target->getMBB()) 451 return true; 452 453 // If the conditions are the same, we can leave them alone. 454 unsigned OldCCValid = Cond[0].getImm(); 455 unsigned OldCCMask = Cond[1].getImm(); 456 if (OldCCValid == Branch.CCValid && OldCCMask == Branch.CCMask) 457 continue; 458 459 // FIXME: Try combining conditions like X86 does. Should be easy on Z! 460 return false; 461 } 462 463 return false; 464 } 465 466 unsigned SystemZInstrInfo::removeBranch(MachineBasicBlock &MBB, 467 int *BytesRemoved) const { 468 assert(!BytesRemoved && "code size not handled"); 469 470 // Most of the code and comments here are boilerplate. 471 MachineBasicBlock::iterator I = MBB.end(); 472 unsigned Count = 0; 473 474 while (I != MBB.begin()) { 475 --I; 476 if (I->isDebugValue()) 477 continue; 478 if (!I->isBranch()) 479 break; 480 if (!getBranchInfo(*I).Target->isMBB()) 481 break; 482 // Remove the branch. 483 I->eraseFromParent(); 484 I = MBB.end(); 485 ++Count; 486 } 487 488 return Count; 489 } 490 491 bool SystemZInstrInfo:: 492 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 493 assert(Cond.size() == 2 && "Invalid condition"); 494 Cond[1].setImm(Cond[1].getImm() ^ Cond[0].getImm()); 495 return false; 496 } 497 498 unsigned SystemZInstrInfo::insertBranch(MachineBasicBlock &MBB, 499 MachineBasicBlock *TBB, 500 MachineBasicBlock *FBB, 501 ArrayRef<MachineOperand> Cond, 502 const DebugLoc &DL, 503 int *BytesAdded) const { 504 // In this function we output 32-bit branches, which should always 505 // have enough range. They can be shortened and relaxed by later code 506 // in the pipeline, if desired. 507 508 // Shouldn't be a fall through. 509 assert(TBB && "insertBranch must not be told to insert a fallthrough"); 510 assert((Cond.size() == 2 || Cond.size() == 0) && 511 "SystemZ branch conditions have one component!"); 512 assert(!BytesAdded && "code size not handled"); 513 514 if (Cond.empty()) { 515 // Unconditional branch? 516 assert(!FBB && "Unconditional branch with multiple successors!"); 517 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(TBB); 518 return 1; 519 } 520 521 // Conditional branch. 522 unsigned Count = 0; 523 unsigned CCValid = Cond[0].getImm(); 524 unsigned CCMask = Cond[1].getImm(); 525 BuildMI(&MBB, DL, get(SystemZ::BRC)) 526 .addImm(CCValid).addImm(CCMask).addMBB(TBB); 527 ++Count; 528 529 if (FBB) { 530 // Two-way Conditional branch. Insert the second branch. 531 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(FBB); 532 ++Count; 533 } 534 return Count; 535 } 536 537 bool SystemZInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg, 538 unsigned &SrcReg2, int &Mask, 539 int &Value) const { 540 assert(MI.isCompare() && "Caller should have checked for a comparison"); 541 542 if (MI.getNumExplicitOperands() == 2 && MI.getOperand(0).isReg() && 543 MI.getOperand(1).isImm()) { 544 SrcReg = MI.getOperand(0).getReg(); 545 SrcReg2 = 0; 546 Value = MI.getOperand(1).getImm(); 547 Mask = ~0; 548 return true; 549 } 550 551 return false; 552 } 553 554 // If Reg is a virtual register, return its definition, otherwise return null. 555 static MachineInstr *getDef(unsigned Reg, 556 const MachineRegisterInfo *MRI) { 557 if (TargetRegisterInfo::isPhysicalRegister(Reg)) 558 return nullptr; 559 return MRI->getUniqueVRegDef(Reg); 560 } 561 562 // Return true if MI is a shift of type Opcode by Imm bits. 563 static bool isShift(MachineInstr *MI, unsigned Opcode, int64_t Imm) { 564 return (MI->getOpcode() == Opcode && 565 !MI->getOperand(2).getReg() && 566 MI->getOperand(3).getImm() == Imm); 567 } 568 569 // If the destination of MI has no uses, delete it as dead. 570 static void eraseIfDead(MachineInstr *MI, const MachineRegisterInfo *MRI) { 571 if (MRI->use_nodbg_empty(MI->getOperand(0).getReg())) 572 MI->eraseFromParent(); 573 } 574 575 // Compare compares SrcReg against zero. Check whether SrcReg contains 576 // the result of an IPM sequence whose input CC survives until Compare, 577 // and whether Compare is therefore redundant. Delete it and return 578 // true if so. 579 static bool removeIPMBasedCompare(MachineInstr &Compare, unsigned SrcReg, 580 const MachineRegisterInfo *MRI, 581 const TargetRegisterInfo *TRI) { 582 MachineInstr *LGFR = nullptr; 583 MachineInstr *RLL = getDef(SrcReg, MRI); 584 if (RLL && RLL->getOpcode() == SystemZ::LGFR) { 585 LGFR = RLL; 586 RLL = getDef(LGFR->getOperand(1).getReg(), MRI); 587 } 588 if (!RLL || !isShift(RLL, SystemZ::RLL, 31)) 589 return false; 590 591 MachineInstr *SRL = getDef(RLL->getOperand(1).getReg(), MRI); 592 if (!SRL || !isShift(SRL, SystemZ::SRL, SystemZ::IPM_CC)) 593 return false; 594 595 MachineInstr *IPM = getDef(SRL->getOperand(1).getReg(), MRI); 596 if (!IPM || IPM->getOpcode() != SystemZ::IPM) 597 return false; 598 599 // Check that there are no assignments to CC between the IPM and Compare, 600 if (IPM->getParent() != Compare.getParent()) 601 return false; 602 MachineBasicBlock::iterator MBBI = IPM, MBBE = Compare.getIterator(); 603 for (++MBBI; MBBI != MBBE; ++MBBI) { 604 MachineInstr &MI = *MBBI; 605 if (MI.modifiesRegister(SystemZ::CC, TRI)) 606 return false; 607 } 608 609 Compare.eraseFromParent(); 610 if (LGFR) 611 eraseIfDead(LGFR, MRI); 612 eraseIfDead(RLL, MRI); 613 eraseIfDead(SRL, MRI); 614 eraseIfDead(IPM, MRI); 615 616 return true; 617 } 618 619 bool SystemZInstrInfo::optimizeCompareInstr( 620 MachineInstr &Compare, unsigned SrcReg, unsigned SrcReg2, int Mask, 621 int Value, const MachineRegisterInfo *MRI) const { 622 assert(!SrcReg2 && "Only optimizing constant comparisons so far"); 623 bool IsLogical = (Compare.getDesc().TSFlags & SystemZII::IsLogical) != 0; 624 return Value == 0 && !IsLogical && 625 removeIPMBasedCompare(Compare, SrcReg, MRI, &RI); 626 } 627 628 bool SystemZInstrInfo::canInsertSelect(const MachineBasicBlock &MBB, 629 ArrayRef<MachineOperand> Pred, 630 unsigned TrueReg, unsigned FalseReg, 631 int &CondCycles, int &TrueCycles, 632 int &FalseCycles) const { 633 // Not all subtargets have LOCR instructions. 634 if (!STI.hasLoadStoreOnCond()) 635 return false; 636 if (Pred.size() != 2) 637 return false; 638 639 // Check register classes. 640 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 641 const TargetRegisterClass *RC = 642 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 643 if (!RC) 644 return false; 645 646 // We have LOCR instructions for 32 and 64 bit general purpose registers. 647 if ((STI.hasLoadStoreOnCond2() && 648 SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) || 649 SystemZ::GR32BitRegClass.hasSubClassEq(RC) || 650 SystemZ::GR64BitRegClass.hasSubClassEq(RC)) { 651 CondCycles = 2; 652 TrueCycles = 2; 653 FalseCycles = 2; 654 return true; 655 } 656 657 // Can't do anything else. 658 return false; 659 } 660 661 void SystemZInstrInfo::insertSelect(MachineBasicBlock &MBB, 662 MachineBasicBlock::iterator I, 663 const DebugLoc &DL, unsigned DstReg, 664 ArrayRef<MachineOperand> Pred, 665 unsigned TrueReg, 666 unsigned FalseReg) const { 667 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 668 const TargetRegisterClass *RC = MRI.getRegClass(DstReg); 669 670 assert(Pred.size() == 2 && "Invalid condition"); 671 unsigned CCValid = Pred[0].getImm(); 672 unsigned CCMask = Pred[1].getImm(); 673 674 unsigned Opc; 675 if (SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) { 676 if (STI.hasLoadStoreOnCond2()) 677 Opc = SystemZ::LOCRMux; 678 else { 679 Opc = SystemZ::LOCR; 680 MRI.constrainRegClass(DstReg, &SystemZ::GR32BitRegClass); 681 unsigned TReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass); 682 unsigned FReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass); 683 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), TReg).addReg(TrueReg); 684 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), FReg).addReg(FalseReg); 685 TrueReg = TReg; 686 FalseReg = FReg; 687 } 688 } else if (SystemZ::GR64BitRegClass.hasSubClassEq(RC)) 689 Opc = SystemZ::LOCGR; 690 else 691 llvm_unreachable("Invalid register class"); 692 693 BuildMI(MBB, I, DL, get(Opc), DstReg) 694 .addReg(FalseReg).addReg(TrueReg) 695 .addImm(CCValid).addImm(CCMask); 696 } 697 698 bool SystemZInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, 699 unsigned Reg, 700 MachineRegisterInfo *MRI) const { 701 unsigned DefOpc = DefMI.getOpcode(); 702 if (DefOpc != SystemZ::LHIMux && DefOpc != SystemZ::LHI && 703 DefOpc != SystemZ::LGHI) 704 return false; 705 if (DefMI.getOperand(0).getReg() != Reg) 706 return false; 707 int32_t ImmVal = (int32_t)DefMI.getOperand(1).getImm(); 708 709 unsigned UseOpc = UseMI.getOpcode(); 710 unsigned NewUseOpc; 711 unsigned UseIdx; 712 int CommuteIdx = -1; 713 switch (UseOpc) { 714 case SystemZ::LOCRMux: 715 if (!STI.hasLoadStoreOnCond2()) 716 return false; 717 NewUseOpc = SystemZ::LOCHIMux; 718 if (UseMI.getOperand(2).getReg() == Reg) 719 UseIdx = 2; 720 else if (UseMI.getOperand(1).getReg() == Reg) 721 UseIdx = 2, CommuteIdx = 1; 722 else 723 return false; 724 break; 725 case SystemZ::LOCGR: 726 if (!STI.hasLoadStoreOnCond2()) 727 return false; 728 NewUseOpc = SystemZ::LOCGHI; 729 if (UseMI.getOperand(2).getReg() == Reg) 730 UseIdx = 2; 731 else if (UseMI.getOperand(1).getReg() == Reg) 732 UseIdx = 2, CommuteIdx = 1; 733 else 734 return false; 735 break; 736 default: 737 return false; 738 } 739 740 if (CommuteIdx != -1) 741 if (!commuteInstruction(UseMI, false, CommuteIdx, UseIdx)) 742 return false; 743 744 bool DeleteDef = MRI->hasOneNonDBGUse(Reg); 745 UseMI.setDesc(get(NewUseOpc)); 746 UseMI.getOperand(UseIdx).ChangeToImmediate(ImmVal); 747 if (DeleteDef) 748 DefMI.eraseFromParent(); 749 750 return true; 751 } 752 753 bool SystemZInstrInfo::isPredicable(const MachineInstr &MI) const { 754 unsigned Opcode = MI.getOpcode(); 755 if (Opcode == SystemZ::Return || 756 Opcode == SystemZ::Trap || 757 Opcode == SystemZ::CallJG || 758 Opcode == SystemZ::CallBR) 759 return true; 760 return false; 761 } 762 763 bool SystemZInstrInfo:: 764 isProfitableToIfCvt(MachineBasicBlock &MBB, 765 unsigned NumCycles, unsigned ExtraPredCycles, 766 BranchProbability Probability) const { 767 // Avoid using conditional returns at the end of a loop (since then 768 // we'd need to emit an unconditional branch to the beginning anyway, 769 // making the loop body longer). This doesn't apply for low-probability 770 // loops (eg. compare-and-swap retry), so just decide based on branch 771 // probability instead of looping structure. 772 // However, since Compare and Trap instructions cost the same as a regular 773 // Compare instruction, we should allow the if conversion to convert this 774 // into a Conditional Compare regardless of the branch probability. 775 if (MBB.getLastNonDebugInstr()->getOpcode() != SystemZ::Trap && 776 MBB.succ_empty() && Probability < BranchProbability(1, 8)) 777 return false; 778 // For now only convert single instructions. 779 return NumCycles == 1; 780 } 781 782 bool SystemZInstrInfo:: 783 isProfitableToIfCvt(MachineBasicBlock &TMBB, 784 unsigned NumCyclesT, unsigned ExtraPredCyclesT, 785 MachineBasicBlock &FMBB, 786 unsigned NumCyclesF, unsigned ExtraPredCyclesF, 787 BranchProbability Probability) const { 788 // For now avoid converting mutually-exclusive cases. 789 return false; 790 } 791 792 bool SystemZInstrInfo:: 793 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles, 794 BranchProbability Probability) const { 795 // For now only duplicate single instructions. 796 return NumCycles == 1; 797 } 798 799 bool SystemZInstrInfo::PredicateInstruction( 800 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const { 801 assert(Pred.size() == 2 && "Invalid condition"); 802 unsigned CCValid = Pred[0].getImm(); 803 unsigned CCMask = Pred[1].getImm(); 804 assert(CCMask > 0 && CCMask < 15 && "Invalid predicate"); 805 unsigned Opcode = MI.getOpcode(); 806 if (Opcode == SystemZ::Trap) { 807 MI.setDesc(get(SystemZ::CondTrap)); 808 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 809 .addImm(CCValid).addImm(CCMask) 810 .addReg(SystemZ::CC, RegState::Implicit); 811 return true; 812 } 813 if (Opcode == SystemZ::Return) { 814 MI.setDesc(get(SystemZ::CondReturn)); 815 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 816 .addImm(CCValid).addImm(CCMask) 817 .addReg(SystemZ::CC, RegState::Implicit); 818 return true; 819 } 820 if (Opcode == SystemZ::CallJG) { 821 MachineOperand FirstOp = MI.getOperand(0); 822 const uint32_t *RegMask = MI.getOperand(1).getRegMask(); 823 MI.RemoveOperand(1); 824 MI.RemoveOperand(0); 825 MI.setDesc(get(SystemZ::CallBRCL)); 826 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 827 .addImm(CCValid) 828 .addImm(CCMask) 829 .add(FirstOp) 830 .addRegMask(RegMask) 831 .addReg(SystemZ::CC, RegState::Implicit); 832 return true; 833 } 834 if (Opcode == SystemZ::CallBR) { 835 const uint32_t *RegMask = MI.getOperand(0).getRegMask(); 836 MI.RemoveOperand(0); 837 MI.setDesc(get(SystemZ::CallBCR)); 838 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 839 .addImm(CCValid).addImm(CCMask) 840 .addRegMask(RegMask) 841 .addReg(SystemZ::CC, RegState::Implicit); 842 return true; 843 } 844 return false; 845 } 846 847 void SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 848 MachineBasicBlock::iterator MBBI, 849 const DebugLoc &DL, unsigned DestReg, 850 unsigned SrcReg, bool KillSrc) const { 851 // Split 128-bit GPR moves into two 64-bit moves. Add implicit uses of the 852 // super register in case one of the subregs is undefined. 853 // 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 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI)) 858 .addReg(SrcReg, RegState::Implicit); 859 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_l64), 860 RI.getSubReg(SrcReg, SystemZ::subreg_l64), KillSrc); 861 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI)) 862 .addReg(SrcReg, (getKillRegState(KillSrc) | RegState::Implicit)); 863 return; 864 } 865 866 if (SystemZ::GRX32BitRegClass.contains(DestReg, SrcReg)) { 867 emitGRX32Move(MBB, MBBI, DL, DestReg, SrcReg, SystemZ::LR, 32, KillSrc, 868 false); 869 return; 870 } 871 872 // Everything else needs only one instruction. 873 unsigned Opcode; 874 if (SystemZ::GR64BitRegClass.contains(DestReg, SrcReg)) 875 Opcode = SystemZ::LGR; 876 else if (SystemZ::FP32BitRegClass.contains(DestReg, SrcReg)) 877 // For z13 we prefer LDR over LER to avoid partial register dependencies. 878 Opcode = STI.hasVector() ? SystemZ::LDR32 : SystemZ::LER; 879 else if (SystemZ::FP64BitRegClass.contains(DestReg, SrcReg)) 880 Opcode = SystemZ::LDR; 881 else if (SystemZ::FP128BitRegClass.contains(DestReg, SrcReg)) 882 Opcode = SystemZ::LXR; 883 else if (SystemZ::VR32BitRegClass.contains(DestReg, SrcReg)) 884 Opcode = SystemZ::VLR32; 885 else if (SystemZ::VR64BitRegClass.contains(DestReg, SrcReg)) 886 Opcode = SystemZ::VLR64; 887 else if (SystemZ::VR128BitRegClass.contains(DestReg, SrcReg)) 888 Opcode = SystemZ::VLR; 889 else if (SystemZ::AR32BitRegClass.contains(DestReg, SrcReg)) 890 Opcode = SystemZ::CPYA; 891 else if (SystemZ::AR32BitRegClass.contains(DestReg) && 892 SystemZ::GR32BitRegClass.contains(SrcReg)) 893 Opcode = SystemZ::SAR; 894 else if (SystemZ::GR32BitRegClass.contains(DestReg) && 895 SystemZ::AR32BitRegClass.contains(SrcReg)) 896 Opcode = SystemZ::EAR; 897 else 898 llvm_unreachable("Impossible reg-to-reg copy"); 899 900 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 901 .addReg(SrcReg, getKillRegState(KillSrc)); 902 } 903 904 void SystemZInstrInfo::storeRegToStackSlot( 905 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned SrcReg, 906 bool isKill, int FrameIdx, const TargetRegisterClass *RC, 907 const TargetRegisterInfo *TRI) const { 908 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 909 910 // Callers may expect a single instruction, so keep 128-bit moves 911 // together for now and lower them after register allocation. 912 unsigned LoadOpcode, StoreOpcode; 913 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 914 addFrameReference(BuildMI(MBB, MBBI, DL, get(StoreOpcode)) 915 .addReg(SrcReg, getKillRegState(isKill)), 916 FrameIdx); 917 } 918 919 void SystemZInstrInfo::loadRegFromStackSlot( 920 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned DestReg, 921 int FrameIdx, const TargetRegisterClass *RC, 922 const TargetRegisterInfo *TRI) const { 923 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 924 925 // Callers may expect a single instruction, so keep 128-bit moves 926 // together for now and lower them after register allocation. 927 unsigned LoadOpcode, StoreOpcode; 928 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 929 addFrameReference(BuildMI(MBB, MBBI, DL, get(LoadOpcode), DestReg), 930 FrameIdx); 931 } 932 933 // Return true if MI is a simple load or store with a 12-bit displacement 934 // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 935 static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) { 936 const MCInstrDesc &MCID = MI->getDesc(); 937 return ((MCID.TSFlags & Flag) && 938 isUInt<12>(MI->getOperand(2).getImm()) && 939 MI->getOperand(3).getReg() == 0); 940 } 941 942 namespace { 943 944 struct LogicOp { 945 LogicOp() = default; 946 LogicOp(unsigned regSize, unsigned immLSB, unsigned immSize) 947 : RegSize(regSize), ImmLSB(immLSB), ImmSize(immSize) {} 948 949 explicit operator bool() const { return RegSize; } 950 951 unsigned RegSize = 0; 952 unsigned ImmLSB = 0; 953 unsigned ImmSize = 0; 954 }; 955 956 } // end anonymous namespace 957 958 static LogicOp interpretAndImmediate(unsigned Opcode) { 959 switch (Opcode) { 960 case SystemZ::NILMux: return LogicOp(32, 0, 16); 961 case SystemZ::NIHMux: return LogicOp(32, 16, 16); 962 case SystemZ::NILL64: return LogicOp(64, 0, 16); 963 case SystemZ::NILH64: return LogicOp(64, 16, 16); 964 case SystemZ::NIHL64: return LogicOp(64, 32, 16); 965 case SystemZ::NIHH64: return LogicOp(64, 48, 16); 966 case SystemZ::NIFMux: return LogicOp(32, 0, 32); 967 case SystemZ::NILF64: return LogicOp(64, 0, 32); 968 case SystemZ::NIHF64: return LogicOp(64, 32, 32); 969 default: return LogicOp(); 970 } 971 } 972 973 static void transferDeadCC(MachineInstr *OldMI, MachineInstr *NewMI) { 974 if (OldMI->registerDefIsDead(SystemZ::CC)) { 975 MachineOperand *CCDef = NewMI->findRegisterDefOperand(SystemZ::CC); 976 if (CCDef != nullptr) 977 CCDef->setIsDead(true); 978 } 979 } 980 981 // Used to return from convertToThreeAddress after replacing two-address 982 // instruction OldMI with three-address instruction NewMI. 983 static MachineInstr *finishConvertToThreeAddress(MachineInstr *OldMI, 984 MachineInstr *NewMI, 985 LiveVariables *LV) { 986 if (LV) { 987 unsigned NumOps = OldMI->getNumOperands(); 988 for (unsigned I = 1; I < NumOps; ++I) { 989 MachineOperand &Op = OldMI->getOperand(I); 990 if (Op.isReg() && Op.isKill()) 991 LV->replaceKillInstruction(Op.getReg(), *OldMI, *NewMI); 992 } 993 } 994 transferDeadCC(OldMI, NewMI); 995 return NewMI; 996 } 997 998 MachineInstr *SystemZInstrInfo::convertToThreeAddress( 999 MachineFunction::iterator &MFI, MachineInstr &MI, LiveVariables *LV) const { 1000 MachineBasicBlock *MBB = MI.getParent(); 1001 MachineFunction *MF = MBB->getParent(); 1002 MachineRegisterInfo &MRI = MF->getRegInfo(); 1003 1004 unsigned Opcode = MI.getOpcode(); 1005 unsigned NumOps = MI.getNumOperands(); 1006 1007 // Try to convert something like SLL into SLLK, if supported. 1008 // We prefer to keep the two-operand form where possible both 1009 // because it tends to be shorter and because some instructions 1010 // have memory forms that can be used during spilling. 1011 if (STI.hasDistinctOps()) { 1012 MachineOperand &Dest = MI.getOperand(0); 1013 MachineOperand &Src = MI.getOperand(1); 1014 unsigned DestReg = Dest.getReg(); 1015 unsigned SrcReg = Src.getReg(); 1016 // AHIMux is only really a three-operand instruction when both operands 1017 // are low registers. Try to constrain both operands to be low if 1018 // possible. 1019 if (Opcode == SystemZ::AHIMux && 1020 TargetRegisterInfo::isVirtualRegister(DestReg) && 1021 TargetRegisterInfo::isVirtualRegister(SrcReg) && 1022 MRI.getRegClass(DestReg)->contains(SystemZ::R1L) && 1023 MRI.getRegClass(SrcReg)->contains(SystemZ::R1L)) { 1024 MRI.constrainRegClass(DestReg, &SystemZ::GR32BitRegClass); 1025 MRI.constrainRegClass(SrcReg, &SystemZ::GR32BitRegClass); 1026 } 1027 int ThreeOperandOpcode = SystemZ::getThreeOperandOpcode(Opcode); 1028 if (ThreeOperandOpcode >= 0) { 1029 // Create three address instruction without adding the implicit 1030 // operands. Those will instead be copied over from the original 1031 // instruction by the loop below. 1032 MachineInstrBuilder MIB( 1033 *MF, MF->CreateMachineInstr(get(ThreeOperandOpcode), MI.getDebugLoc(), 1034 /*NoImplicit=*/true)); 1035 MIB.add(Dest); 1036 // Keep the kill state, but drop the tied flag. 1037 MIB.addReg(Src.getReg(), getKillRegState(Src.isKill()), Src.getSubReg()); 1038 // Keep the remaining operands as-is. 1039 for (unsigned I = 2; I < NumOps; ++I) 1040 MIB.add(MI.getOperand(I)); 1041 MBB->insert(MI, MIB); 1042 return finishConvertToThreeAddress(&MI, MIB, LV); 1043 } 1044 } 1045 1046 // Try to convert an AND into an RISBG-type instruction. 1047 if (LogicOp And = interpretAndImmediate(Opcode)) { 1048 uint64_t Imm = MI.getOperand(2).getImm() << And.ImmLSB; 1049 // AND IMMEDIATE leaves the other bits of the register unchanged. 1050 Imm |= allOnes(And.RegSize) & ~(allOnes(And.ImmSize) << And.ImmLSB); 1051 unsigned Start, End; 1052 if (isRxSBGMask(Imm, And.RegSize, Start, End)) { 1053 unsigned NewOpcode; 1054 if (And.RegSize == 64) { 1055 NewOpcode = SystemZ::RISBG; 1056 // Prefer RISBGN if available, since it does not clobber CC. 1057 if (STI.hasMiscellaneousExtensions()) 1058 NewOpcode = SystemZ::RISBGN; 1059 } else { 1060 NewOpcode = SystemZ::RISBMux; 1061 Start &= 31; 1062 End &= 31; 1063 } 1064 MachineOperand &Dest = MI.getOperand(0); 1065 MachineOperand &Src = MI.getOperand(1); 1066 MachineInstrBuilder MIB = 1067 BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpcode)) 1068 .add(Dest) 1069 .addReg(0) 1070 .addReg(Src.getReg(), getKillRegState(Src.isKill()), 1071 Src.getSubReg()) 1072 .addImm(Start) 1073 .addImm(End + 128) 1074 .addImm(0); 1075 return finishConvertToThreeAddress(&MI, MIB, LV); 1076 } 1077 } 1078 return nullptr; 1079 } 1080 1081 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1082 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1083 MachineBasicBlock::iterator InsertPt, int FrameIndex, 1084 LiveIntervals *LIS) const { 1085 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 1086 const MachineFrameInfo &MFI = MF.getFrameInfo(); 1087 unsigned Size = MFI.getObjectSize(FrameIndex); 1088 unsigned Opcode = MI.getOpcode(); 1089 1090 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 1091 if (LIS != nullptr && (Opcode == SystemZ::LA || Opcode == SystemZ::LAY) && 1092 isInt<8>(MI.getOperand(2).getImm()) && !MI.getOperand(3).getReg()) { 1093 1094 // Check CC liveness, since new instruction introduces a dead 1095 // def of CC. 1096 MCRegUnitIterator CCUnit(SystemZ::CC, TRI); 1097 LiveRange &CCLiveRange = LIS->getRegUnit(*CCUnit); 1098 ++CCUnit; 1099 assert(!CCUnit.isValid() && "CC only has one reg unit."); 1100 SlotIndex MISlot = 1101 LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot(); 1102 if (!CCLiveRange.liveAt(MISlot)) { 1103 // LA(Y) %reg, CONST(%reg) -> AGSI %mem, CONST 1104 MachineInstr *BuiltMI = BuildMI(*InsertPt->getParent(), InsertPt, 1105 MI.getDebugLoc(), get(SystemZ::AGSI)) 1106 .addFrameIndex(FrameIndex) 1107 .addImm(0) 1108 .addImm(MI.getOperand(2).getImm()); 1109 BuiltMI->findRegisterDefOperand(SystemZ::CC)->setIsDead(true); 1110 CCLiveRange.createDeadDef(MISlot, LIS->getVNInfoAllocator()); 1111 return BuiltMI; 1112 } 1113 } 1114 return nullptr; 1115 } 1116 1117 // All other cases require a single operand. 1118 if (Ops.size() != 1) 1119 return nullptr; 1120 1121 unsigned OpNum = Ops[0]; 1122 assert(Size * 8 == 1123 TRI->getRegSizeInBits(*MF.getRegInfo() 1124 .getRegClass(MI.getOperand(OpNum).getReg())) && 1125 "Invalid size combination"); 1126 1127 if ((Opcode == SystemZ::AHI || Opcode == SystemZ::AGHI) && OpNum == 0 && 1128 isInt<8>(MI.getOperand(2).getImm())) { 1129 // A(G)HI %reg, CONST -> A(G)SI %mem, CONST 1130 Opcode = (Opcode == SystemZ::AHI ? SystemZ::ASI : SystemZ::AGSI); 1131 MachineInstr *BuiltMI = 1132 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode)) 1133 .addFrameIndex(FrameIndex) 1134 .addImm(0) 1135 .addImm(MI.getOperand(2).getImm()); 1136 transferDeadCC(&MI, BuiltMI); 1137 return BuiltMI; 1138 } 1139 1140 if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) { 1141 bool Op0IsGPR = (Opcode == SystemZ::LGDR); 1142 bool Op1IsGPR = (Opcode == SystemZ::LDGR); 1143 // If we're spilling the destination of an LDGR or LGDR, store the 1144 // source register instead. 1145 if (OpNum == 0) { 1146 unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD; 1147 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1148 get(StoreOpcode)) 1149 .add(MI.getOperand(1)) 1150 .addFrameIndex(FrameIndex) 1151 .addImm(0) 1152 .addReg(0); 1153 } 1154 // If we're spilling the source of an LDGR or LGDR, load the 1155 // destination register instead. 1156 if (OpNum == 1) { 1157 unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD; 1158 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1159 get(LoadOpcode)) 1160 .add(MI.getOperand(0)) 1161 .addFrameIndex(FrameIndex) 1162 .addImm(0) 1163 .addReg(0); 1164 } 1165 } 1166 1167 // Look for cases where the source of a simple store or the destination 1168 // of a simple load is being spilled. Try to use MVC instead. 1169 // 1170 // Although MVC is in practice a fast choice in these cases, it is still 1171 // logically a bytewise copy. This means that we cannot use it if the 1172 // load or store is volatile. We also wouldn't be able to use MVC if 1173 // the two memories partially overlap, but that case cannot occur here, 1174 // because we know that one of the memories is a full frame index. 1175 // 1176 // For performance reasons, we also want to avoid using MVC if the addresses 1177 // might be equal. We don't worry about that case here, because spill slot 1178 // coloring happens later, and because we have special code to remove 1179 // MVCs that turn out to be redundant. 1180 if (OpNum == 0 && MI.hasOneMemOperand()) { 1181 MachineMemOperand *MMO = *MI.memoperands_begin(); 1182 if (MMO->getSize() == Size && !MMO->isVolatile()) { 1183 // Handle conversion of loads. 1184 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXLoad)) { 1185 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1186 get(SystemZ::MVC)) 1187 .addFrameIndex(FrameIndex) 1188 .addImm(0) 1189 .addImm(Size) 1190 .add(MI.getOperand(1)) 1191 .addImm(MI.getOperand(2).getImm()) 1192 .addMemOperand(MMO); 1193 } 1194 // Handle conversion of stores. 1195 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXStore)) { 1196 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1197 get(SystemZ::MVC)) 1198 .add(MI.getOperand(1)) 1199 .addImm(MI.getOperand(2).getImm()) 1200 .addImm(Size) 1201 .addFrameIndex(FrameIndex) 1202 .addImm(0) 1203 .addMemOperand(MMO); 1204 } 1205 } 1206 } 1207 1208 // If the spilled operand is the final one, try to change <INSN>R 1209 // into <INSN>. 1210 int MemOpcode = SystemZ::getMemOpcode(Opcode); 1211 if (MemOpcode >= 0) { 1212 unsigned NumOps = MI.getNumExplicitOperands(); 1213 if (OpNum == NumOps - 1) { 1214 const MCInstrDesc &MemDesc = get(MemOpcode); 1215 uint64_t AccessBytes = SystemZII::getAccessSize(MemDesc.TSFlags); 1216 assert(AccessBytes != 0 && "Size of access should be known"); 1217 assert(AccessBytes <= Size && "Access outside the frame index"); 1218 uint64_t Offset = Size - AccessBytes; 1219 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt, 1220 MI.getDebugLoc(), get(MemOpcode)); 1221 for (unsigned I = 0; I < OpNum; ++I) 1222 MIB.add(MI.getOperand(I)); 1223 MIB.addFrameIndex(FrameIndex).addImm(Offset); 1224 if (MemDesc.TSFlags & SystemZII::HasIndex) 1225 MIB.addReg(0); 1226 transferDeadCC(&MI, MIB); 1227 return MIB; 1228 } 1229 } 1230 1231 return nullptr; 1232 } 1233 1234 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1235 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1236 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, 1237 LiveIntervals *LIS) const { 1238 return nullptr; 1239 } 1240 1241 bool SystemZInstrInfo::expandPostRAPseudo(MachineInstr &MI) const { 1242 switch (MI.getOpcode()) { 1243 case SystemZ::L128: 1244 splitMove(MI, SystemZ::LG); 1245 return true; 1246 1247 case SystemZ::ST128: 1248 splitMove(MI, SystemZ::STG); 1249 return true; 1250 1251 case SystemZ::LX: 1252 splitMove(MI, SystemZ::LD); 1253 return true; 1254 1255 case SystemZ::STX: 1256 splitMove(MI, SystemZ::STD); 1257 return true; 1258 1259 case SystemZ::LBMux: 1260 expandRXYPseudo(MI, SystemZ::LB, SystemZ::LBH); 1261 return true; 1262 1263 case SystemZ::LHMux: 1264 expandRXYPseudo(MI, SystemZ::LH, SystemZ::LHH); 1265 return true; 1266 1267 case SystemZ::LLCRMux: 1268 expandZExtPseudo(MI, SystemZ::LLCR, 8); 1269 return true; 1270 1271 case SystemZ::LLHRMux: 1272 expandZExtPseudo(MI, SystemZ::LLHR, 16); 1273 return true; 1274 1275 case SystemZ::LLCMux: 1276 expandRXYPseudo(MI, SystemZ::LLC, SystemZ::LLCH); 1277 return true; 1278 1279 case SystemZ::LLHMux: 1280 expandRXYPseudo(MI, SystemZ::LLH, SystemZ::LLHH); 1281 return true; 1282 1283 case SystemZ::LMux: 1284 expandRXYPseudo(MI, SystemZ::L, SystemZ::LFH); 1285 return true; 1286 1287 case SystemZ::LOCMux: 1288 expandLOCPseudo(MI, SystemZ::LOC, SystemZ::LOCFH); 1289 return true; 1290 1291 case SystemZ::LOCHIMux: 1292 expandLOCPseudo(MI, SystemZ::LOCHI, SystemZ::LOCHHI); 1293 return true; 1294 1295 case SystemZ::LOCRMux: 1296 expandLOCRPseudo(MI, SystemZ::LOCR, SystemZ::LOCFHR); 1297 return true; 1298 1299 case SystemZ::STCMux: 1300 expandRXYPseudo(MI, SystemZ::STC, SystemZ::STCH); 1301 return true; 1302 1303 case SystemZ::STHMux: 1304 expandRXYPseudo(MI, SystemZ::STH, SystemZ::STHH); 1305 return true; 1306 1307 case SystemZ::STMux: 1308 expandRXYPseudo(MI, SystemZ::ST, SystemZ::STFH); 1309 return true; 1310 1311 case SystemZ::STOCMux: 1312 expandLOCPseudo(MI, SystemZ::STOC, SystemZ::STOCFH); 1313 return true; 1314 1315 case SystemZ::LHIMux: 1316 expandRIPseudo(MI, SystemZ::LHI, SystemZ::IIHF, true); 1317 return true; 1318 1319 case SystemZ::IIFMux: 1320 expandRIPseudo(MI, SystemZ::IILF, SystemZ::IIHF, false); 1321 return true; 1322 1323 case SystemZ::IILMux: 1324 expandRIPseudo(MI, SystemZ::IILL, SystemZ::IIHL, false); 1325 return true; 1326 1327 case SystemZ::IIHMux: 1328 expandRIPseudo(MI, SystemZ::IILH, SystemZ::IIHH, false); 1329 return true; 1330 1331 case SystemZ::NIFMux: 1332 expandRIPseudo(MI, SystemZ::NILF, SystemZ::NIHF, false); 1333 return true; 1334 1335 case SystemZ::NILMux: 1336 expandRIPseudo(MI, SystemZ::NILL, SystemZ::NIHL, false); 1337 return true; 1338 1339 case SystemZ::NIHMux: 1340 expandRIPseudo(MI, SystemZ::NILH, SystemZ::NIHH, false); 1341 return true; 1342 1343 case SystemZ::OIFMux: 1344 expandRIPseudo(MI, SystemZ::OILF, SystemZ::OIHF, false); 1345 return true; 1346 1347 case SystemZ::OILMux: 1348 expandRIPseudo(MI, SystemZ::OILL, SystemZ::OIHL, false); 1349 return true; 1350 1351 case SystemZ::OIHMux: 1352 expandRIPseudo(MI, SystemZ::OILH, SystemZ::OIHH, false); 1353 return true; 1354 1355 case SystemZ::XIFMux: 1356 expandRIPseudo(MI, SystemZ::XILF, SystemZ::XIHF, false); 1357 return true; 1358 1359 case SystemZ::TMLMux: 1360 expandRIPseudo(MI, SystemZ::TMLL, SystemZ::TMHL, false); 1361 return true; 1362 1363 case SystemZ::TMHMux: 1364 expandRIPseudo(MI, SystemZ::TMLH, SystemZ::TMHH, false); 1365 return true; 1366 1367 case SystemZ::AHIMux: 1368 expandRIPseudo(MI, SystemZ::AHI, SystemZ::AIH, false); 1369 return true; 1370 1371 case SystemZ::AHIMuxK: 1372 expandRIEPseudo(MI, SystemZ::AHI, SystemZ::AHIK, SystemZ::AIH); 1373 return true; 1374 1375 case SystemZ::AFIMux: 1376 expandRIPseudo(MI, SystemZ::AFI, SystemZ::AIH, false); 1377 return true; 1378 1379 case SystemZ::CHIMux: 1380 expandRIPseudo(MI, SystemZ::CHI, SystemZ::CIH, false); 1381 return true; 1382 1383 case SystemZ::CFIMux: 1384 expandRIPseudo(MI, SystemZ::CFI, SystemZ::CIH, false); 1385 return true; 1386 1387 case SystemZ::CLFIMux: 1388 expandRIPseudo(MI, SystemZ::CLFI, SystemZ::CLIH, false); 1389 return true; 1390 1391 case SystemZ::CMux: 1392 expandRXYPseudo(MI, SystemZ::C, SystemZ::CHF); 1393 return true; 1394 1395 case SystemZ::CLMux: 1396 expandRXYPseudo(MI, SystemZ::CL, SystemZ::CLHF); 1397 return true; 1398 1399 case SystemZ::RISBMux: { 1400 bool DestIsHigh = isHighReg(MI.getOperand(0).getReg()); 1401 bool SrcIsHigh = isHighReg(MI.getOperand(2).getReg()); 1402 if (SrcIsHigh == DestIsHigh) 1403 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHH : SystemZ::RISBLL)); 1404 else { 1405 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHL : SystemZ::RISBLH)); 1406 MI.getOperand(5).setImm(MI.getOperand(5).getImm() ^ 32); 1407 } 1408 return true; 1409 } 1410 1411 case SystemZ::ADJDYNALLOC: 1412 splitAdjDynAlloc(MI); 1413 return true; 1414 1415 case TargetOpcode::LOAD_STACK_GUARD: 1416 expandLoadStackGuard(&MI); 1417 return true; 1418 1419 default: 1420 return false; 1421 } 1422 } 1423 1424 unsigned SystemZInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { 1425 if (MI.getOpcode() == TargetOpcode::INLINEASM) { 1426 const MachineFunction *MF = MI.getParent()->getParent(); 1427 const char *AsmStr = MI.getOperand(0).getSymbolName(); 1428 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo()); 1429 } 1430 return MI.getDesc().getSize(); 1431 } 1432 1433 SystemZII::Branch 1434 SystemZInstrInfo::getBranchInfo(const MachineInstr &MI) const { 1435 switch (MI.getOpcode()) { 1436 case SystemZ::BR: 1437 case SystemZ::J: 1438 case SystemZ::JG: 1439 return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY, 1440 SystemZ::CCMASK_ANY, &MI.getOperand(0)); 1441 1442 case SystemZ::BRC: 1443 case SystemZ::BRCL: 1444 return SystemZII::Branch(SystemZII::BranchNormal, MI.getOperand(0).getImm(), 1445 MI.getOperand(1).getImm(), &MI.getOperand(2)); 1446 1447 case SystemZ::BRCT: 1448 case SystemZ::BRCTH: 1449 return SystemZII::Branch(SystemZII::BranchCT, SystemZ::CCMASK_ICMP, 1450 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1451 1452 case SystemZ::BRCTG: 1453 return SystemZII::Branch(SystemZII::BranchCTG, SystemZ::CCMASK_ICMP, 1454 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1455 1456 case SystemZ::CIJ: 1457 case SystemZ::CRJ: 1458 return SystemZII::Branch(SystemZII::BranchC, SystemZ::CCMASK_ICMP, 1459 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1460 1461 case SystemZ::CLIJ: 1462 case SystemZ::CLRJ: 1463 return SystemZII::Branch(SystemZII::BranchCL, SystemZ::CCMASK_ICMP, 1464 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1465 1466 case SystemZ::CGIJ: 1467 case SystemZ::CGRJ: 1468 return SystemZII::Branch(SystemZII::BranchCG, SystemZ::CCMASK_ICMP, 1469 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1470 1471 case SystemZ::CLGIJ: 1472 case SystemZ::CLGRJ: 1473 return SystemZII::Branch(SystemZII::BranchCLG, SystemZ::CCMASK_ICMP, 1474 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1475 1476 default: 1477 llvm_unreachable("Unrecognized branch opcode"); 1478 } 1479 } 1480 1481 void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC, 1482 unsigned &LoadOpcode, 1483 unsigned &StoreOpcode) const { 1484 if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) { 1485 LoadOpcode = SystemZ::L; 1486 StoreOpcode = SystemZ::ST; 1487 } else if (RC == &SystemZ::GRH32BitRegClass) { 1488 LoadOpcode = SystemZ::LFH; 1489 StoreOpcode = SystemZ::STFH; 1490 } else if (RC == &SystemZ::GRX32BitRegClass) { 1491 LoadOpcode = SystemZ::LMux; 1492 StoreOpcode = SystemZ::STMux; 1493 } else if (RC == &SystemZ::GR64BitRegClass || 1494 RC == &SystemZ::ADDR64BitRegClass) { 1495 LoadOpcode = SystemZ::LG; 1496 StoreOpcode = SystemZ::STG; 1497 } else if (RC == &SystemZ::GR128BitRegClass || 1498 RC == &SystemZ::ADDR128BitRegClass) { 1499 LoadOpcode = SystemZ::L128; 1500 StoreOpcode = SystemZ::ST128; 1501 } else if (RC == &SystemZ::FP32BitRegClass) { 1502 LoadOpcode = SystemZ::LE; 1503 StoreOpcode = SystemZ::STE; 1504 } else if (RC == &SystemZ::FP64BitRegClass) { 1505 LoadOpcode = SystemZ::LD; 1506 StoreOpcode = SystemZ::STD; 1507 } else if (RC == &SystemZ::FP128BitRegClass) { 1508 LoadOpcode = SystemZ::LX; 1509 StoreOpcode = SystemZ::STX; 1510 } else if (RC == &SystemZ::VR32BitRegClass) { 1511 LoadOpcode = SystemZ::VL32; 1512 StoreOpcode = SystemZ::VST32; 1513 } else if (RC == &SystemZ::VR64BitRegClass) { 1514 LoadOpcode = SystemZ::VL64; 1515 StoreOpcode = SystemZ::VST64; 1516 } else if (RC == &SystemZ::VF128BitRegClass || 1517 RC == &SystemZ::VR128BitRegClass) { 1518 LoadOpcode = SystemZ::VL; 1519 StoreOpcode = SystemZ::VST; 1520 } else 1521 llvm_unreachable("Unsupported regclass to load or store"); 1522 } 1523 1524 unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode, 1525 int64_t Offset) const { 1526 const MCInstrDesc &MCID = get(Opcode); 1527 int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset); 1528 if (isUInt<12>(Offset) && isUInt<12>(Offset2)) { 1529 // Get the instruction to use for unsigned 12-bit displacements. 1530 int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode); 1531 if (Disp12Opcode >= 0) 1532 return Disp12Opcode; 1533 1534 // All address-related instructions can use unsigned 12-bit 1535 // displacements. 1536 return Opcode; 1537 } 1538 if (isInt<20>(Offset) && isInt<20>(Offset2)) { 1539 // Get the instruction to use for signed 20-bit displacements. 1540 int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode); 1541 if (Disp20Opcode >= 0) 1542 return Disp20Opcode; 1543 1544 // Check whether Opcode allows signed 20-bit displacements. 1545 if (MCID.TSFlags & SystemZII::Has20BitOffset) 1546 return Opcode; 1547 } 1548 return 0; 1549 } 1550 1551 unsigned SystemZInstrInfo::getLoadAndTest(unsigned Opcode) const { 1552 switch (Opcode) { 1553 case SystemZ::L: return SystemZ::LT; 1554 case SystemZ::LY: return SystemZ::LT; 1555 case SystemZ::LG: return SystemZ::LTG; 1556 case SystemZ::LGF: return SystemZ::LTGF; 1557 case SystemZ::LR: return SystemZ::LTR; 1558 case SystemZ::LGFR: return SystemZ::LTGFR; 1559 case SystemZ::LGR: return SystemZ::LTGR; 1560 case SystemZ::LER: return SystemZ::LTEBR; 1561 case SystemZ::LDR: return SystemZ::LTDBR; 1562 case SystemZ::LXR: return SystemZ::LTXBR; 1563 case SystemZ::LCDFR: return SystemZ::LCDBR; 1564 case SystemZ::LPDFR: return SystemZ::LPDBR; 1565 case SystemZ::LNDFR: return SystemZ::LNDBR; 1566 case SystemZ::LCDFR_32: return SystemZ::LCEBR; 1567 case SystemZ::LPDFR_32: return SystemZ::LPEBR; 1568 case SystemZ::LNDFR_32: return SystemZ::LNEBR; 1569 // On zEC12 we prefer to use RISBGN. But if there is a chance to 1570 // actually use the condition code, we may turn it back into RISGB. 1571 // Note that RISBG is not really a "load-and-test" instruction, 1572 // but sets the same condition code values, so is OK to use here. 1573 case SystemZ::RISBGN: return SystemZ::RISBG; 1574 default: return 0; 1575 } 1576 } 1577 1578 // Return true if Mask matches the regexp 0*1+0*, given that zero masks 1579 // have already been filtered out. Store the first set bit in LSB and 1580 // the number of set bits in Length if so. 1581 static bool isStringOfOnes(uint64_t Mask, unsigned &LSB, unsigned &Length) { 1582 unsigned First = findFirstSet(Mask); 1583 uint64_t Top = (Mask >> First) + 1; 1584 if ((Top & -Top) == Top) { 1585 LSB = First; 1586 Length = findFirstSet(Top); 1587 return true; 1588 } 1589 return false; 1590 } 1591 1592 bool SystemZInstrInfo::isRxSBGMask(uint64_t Mask, unsigned BitSize, 1593 unsigned &Start, unsigned &End) const { 1594 // Reject trivial all-zero masks. 1595 Mask &= allOnes(BitSize); 1596 if (Mask == 0) 1597 return false; 1598 1599 // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of 1600 // the msb and End specifies the index of the lsb. 1601 unsigned LSB, Length; 1602 if (isStringOfOnes(Mask, LSB, Length)) { 1603 Start = 63 - (LSB + Length - 1); 1604 End = 63 - LSB; 1605 return true; 1606 } 1607 1608 // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb 1609 // of the low 1s and End specifies the lsb of the high 1s. 1610 if (isStringOfOnes(Mask ^ allOnes(BitSize), LSB, Length)) { 1611 assert(LSB > 0 && "Bottom bit must be set"); 1612 assert(LSB + Length < BitSize && "Top bit must be set"); 1613 Start = 63 - (LSB - 1); 1614 End = 63 - (LSB + Length); 1615 return true; 1616 } 1617 1618 return false; 1619 } 1620 1621 unsigned SystemZInstrInfo::getFusedCompare(unsigned Opcode, 1622 SystemZII::FusedCompareType Type, 1623 const MachineInstr *MI) const { 1624 switch (Opcode) { 1625 case SystemZ::CHI: 1626 case SystemZ::CGHI: 1627 if (!(MI && isInt<8>(MI->getOperand(1).getImm()))) 1628 return 0; 1629 break; 1630 case SystemZ::CLFI: 1631 case SystemZ::CLGFI: 1632 if (!(MI && isUInt<8>(MI->getOperand(1).getImm()))) 1633 return 0; 1634 break; 1635 case SystemZ::CL: 1636 case SystemZ::CLG: 1637 if (!STI.hasMiscellaneousExtensions()) 1638 return 0; 1639 if (!(MI && MI->getOperand(3).getReg() == 0)) 1640 return 0; 1641 break; 1642 } 1643 switch (Type) { 1644 case SystemZII::CompareAndBranch: 1645 switch (Opcode) { 1646 case SystemZ::CR: 1647 return SystemZ::CRJ; 1648 case SystemZ::CGR: 1649 return SystemZ::CGRJ; 1650 case SystemZ::CHI: 1651 return SystemZ::CIJ; 1652 case SystemZ::CGHI: 1653 return SystemZ::CGIJ; 1654 case SystemZ::CLR: 1655 return SystemZ::CLRJ; 1656 case SystemZ::CLGR: 1657 return SystemZ::CLGRJ; 1658 case SystemZ::CLFI: 1659 return SystemZ::CLIJ; 1660 case SystemZ::CLGFI: 1661 return SystemZ::CLGIJ; 1662 default: 1663 return 0; 1664 } 1665 case SystemZII::CompareAndReturn: 1666 switch (Opcode) { 1667 case SystemZ::CR: 1668 return SystemZ::CRBReturn; 1669 case SystemZ::CGR: 1670 return SystemZ::CGRBReturn; 1671 case SystemZ::CHI: 1672 return SystemZ::CIBReturn; 1673 case SystemZ::CGHI: 1674 return SystemZ::CGIBReturn; 1675 case SystemZ::CLR: 1676 return SystemZ::CLRBReturn; 1677 case SystemZ::CLGR: 1678 return SystemZ::CLGRBReturn; 1679 case SystemZ::CLFI: 1680 return SystemZ::CLIBReturn; 1681 case SystemZ::CLGFI: 1682 return SystemZ::CLGIBReturn; 1683 default: 1684 return 0; 1685 } 1686 case SystemZII::CompareAndSibcall: 1687 switch (Opcode) { 1688 case SystemZ::CR: 1689 return SystemZ::CRBCall; 1690 case SystemZ::CGR: 1691 return SystemZ::CGRBCall; 1692 case SystemZ::CHI: 1693 return SystemZ::CIBCall; 1694 case SystemZ::CGHI: 1695 return SystemZ::CGIBCall; 1696 case SystemZ::CLR: 1697 return SystemZ::CLRBCall; 1698 case SystemZ::CLGR: 1699 return SystemZ::CLGRBCall; 1700 case SystemZ::CLFI: 1701 return SystemZ::CLIBCall; 1702 case SystemZ::CLGFI: 1703 return SystemZ::CLGIBCall; 1704 default: 1705 return 0; 1706 } 1707 case SystemZII::CompareAndTrap: 1708 switch (Opcode) { 1709 case SystemZ::CR: 1710 return SystemZ::CRT; 1711 case SystemZ::CGR: 1712 return SystemZ::CGRT; 1713 case SystemZ::CHI: 1714 return SystemZ::CIT; 1715 case SystemZ::CGHI: 1716 return SystemZ::CGIT; 1717 case SystemZ::CLR: 1718 return SystemZ::CLRT; 1719 case SystemZ::CLGR: 1720 return SystemZ::CLGRT; 1721 case SystemZ::CLFI: 1722 return SystemZ::CLFIT; 1723 case SystemZ::CLGFI: 1724 return SystemZ::CLGIT; 1725 case SystemZ::CL: 1726 return SystemZ::CLT; 1727 case SystemZ::CLG: 1728 return SystemZ::CLGT; 1729 default: 1730 return 0; 1731 } 1732 } 1733 return 0; 1734 } 1735 1736 unsigned SystemZInstrInfo::getLoadAndTrap(unsigned Opcode) const { 1737 if (!STI.hasLoadAndTrap()) 1738 return 0; 1739 switch (Opcode) { 1740 case SystemZ::L: 1741 case SystemZ::LY: 1742 return SystemZ::LAT; 1743 case SystemZ::LG: 1744 return SystemZ::LGAT; 1745 case SystemZ::LFH: 1746 return SystemZ::LFHAT; 1747 case SystemZ::LLGF: 1748 return SystemZ::LLGFAT; 1749 case SystemZ::LLGT: 1750 return SystemZ::LLGTAT; 1751 } 1752 return 0; 1753 } 1754 1755 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB, 1756 MachineBasicBlock::iterator MBBI, 1757 unsigned Reg, uint64_t Value) const { 1758 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 1759 unsigned Opcode; 1760 if (isInt<16>(Value)) 1761 Opcode = SystemZ::LGHI; 1762 else if (SystemZ::isImmLL(Value)) 1763 Opcode = SystemZ::LLILL; 1764 else if (SystemZ::isImmLH(Value)) { 1765 Opcode = SystemZ::LLILH; 1766 Value >>= 16; 1767 } else { 1768 assert(isInt<32>(Value) && "Huge values not handled yet"); 1769 Opcode = SystemZ::LGFI; 1770 } 1771 BuildMI(MBB, MBBI, DL, get(Opcode), Reg).addImm(Value); 1772 } 1773 1774 bool SystemZInstrInfo:: 1775 areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb, 1776 AliasAnalysis *AA) const { 1777 1778 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand()) 1779 return false; 1780 1781 // If mem-operands show that the same address Value is used by both 1782 // instructions, check for non-overlapping offsets and widths. Not 1783 // sure if a register based analysis would be an improvement... 1784 1785 MachineMemOperand *MMOa = *MIa.memoperands_begin(); 1786 MachineMemOperand *MMOb = *MIb.memoperands_begin(); 1787 const Value *VALa = MMOa->getValue(); 1788 const Value *VALb = MMOb->getValue(); 1789 bool SameVal = (VALa && VALb && (VALa == VALb)); 1790 if (!SameVal) { 1791 const PseudoSourceValue *PSVa = MMOa->getPseudoValue(); 1792 const PseudoSourceValue *PSVb = MMOb->getPseudoValue(); 1793 if (PSVa && PSVb && (PSVa == PSVb)) 1794 SameVal = true; 1795 } 1796 if (SameVal) { 1797 int OffsetA = MMOa->getOffset(), OffsetB = MMOb->getOffset(); 1798 int WidthA = MMOa->getSize(), WidthB = MMOb->getSize(); 1799 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB; 1800 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA; 1801 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; 1802 if (LowOffset + LowWidth <= HighOffset) 1803 return true; 1804 } 1805 1806 return false; 1807 } 1808