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 // Move 128-bit floating-point values between VR128 and FP128. 873 if (SystemZ::VR128BitRegClass.contains(DestReg) && 874 SystemZ::FP128BitRegClass.contains(SrcReg)) { 875 unsigned SrcRegHi = 876 RI.getMatchingSuperReg(RI.getSubReg(SrcReg, SystemZ::subreg_h64), 877 SystemZ::subreg_r64, &SystemZ::VR128BitRegClass); 878 unsigned SrcRegLo = 879 RI.getMatchingSuperReg(RI.getSubReg(SrcReg, SystemZ::subreg_l64), 880 SystemZ::subreg_r64, &SystemZ::VR128BitRegClass); 881 882 BuildMI(MBB, MBBI, DL, get(SystemZ::VMRHG), DestReg) 883 .addReg(SrcRegHi, getKillRegState(KillSrc)) 884 .addReg(SrcRegLo, getKillRegState(KillSrc)); 885 return; 886 } 887 if (SystemZ::FP128BitRegClass.contains(DestReg) && 888 SystemZ::VR128BitRegClass.contains(SrcReg)) { 889 unsigned DestRegHi = 890 RI.getMatchingSuperReg(RI.getSubReg(DestReg, SystemZ::subreg_h64), 891 SystemZ::subreg_r64, &SystemZ::VR128BitRegClass); 892 unsigned DestRegLo = 893 RI.getMatchingSuperReg(RI.getSubReg(DestReg, SystemZ::subreg_l64), 894 SystemZ::subreg_r64, &SystemZ::VR128BitRegClass); 895 896 if (DestRegHi != SrcReg) 897 copyPhysReg(MBB, MBBI, DL, DestRegHi, SrcReg, false); 898 BuildMI(MBB, MBBI, DL, get(SystemZ::VREPG), DestRegLo) 899 .addReg(SrcReg, getKillRegState(KillSrc)).addImm(1); 900 return; 901 } 902 903 // Everything else needs only one instruction. 904 unsigned Opcode; 905 if (SystemZ::GR64BitRegClass.contains(DestReg, SrcReg)) 906 Opcode = SystemZ::LGR; 907 else if (SystemZ::FP32BitRegClass.contains(DestReg, SrcReg)) 908 // For z13 we prefer LDR over LER to avoid partial register dependencies. 909 Opcode = STI.hasVector() ? SystemZ::LDR32 : SystemZ::LER; 910 else if (SystemZ::FP64BitRegClass.contains(DestReg, SrcReg)) 911 Opcode = SystemZ::LDR; 912 else if (SystemZ::FP128BitRegClass.contains(DestReg, SrcReg)) 913 Opcode = SystemZ::LXR; 914 else if (SystemZ::VR32BitRegClass.contains(DestReg, SrcReg)) 915 Opcode = SystemZ::VLR32; 916 else if (SystemZ::VR64BitRegClass.contains(DestReg, SrcReg)) 917 Opcode = SystemZ::VLR64; 918 else if (SystemZ::VR128BitRegClass.contains(DestReg, SrcReg)) 919 Opcode = SystemZ::VLR; 920 else if (SystemZ::AR32BitRegClass.contains(DestReg, SrcReg)) 921 Opcode = SystemZ::CPYA; 922 else if (SystemZ::AR32BitRegClass.contains(DestReg) && 923 SystemZ::GR32BitRegClass.contains(SrcReg)) 924 Opcode = SystemZ::SAR; 925 else if (SystemZ::GR32BitRegClass.contains(DestReg) && 926 SystemZ::AR32BitRegClass.contains(SrcReg)) 927 Opcode = SystemZ::EAR; 928 else 929 llvm_unreachable("Impossible reg-to-reg copy"); 930 931 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 932 .addReg(SrcReg, getKillRegState(KillSrc)); 933 } 934 935 void SystemZInstrInfo::storeRegToStackSlot( 936 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned SrcReg, 937 bool isKill, int FrameIdx, const TargetRegisterClass *RC, 938 const TargetRegisterInfo *TRI) const { 939 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 940 941 // Callers may expect a single instruction, so keep 128-bit moves 942 // together for now and lower them after register allocation. 943 unsigned LoadOpcode, StoreOpcode; 944 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 945 addFrameReference(BuildMI(MBB, MBBI, DL, get(StoreOpcode)) 946 .addReg(SrcReg, getKillRegState(isKill)), 947 FrameIdx); 948 } 949 950 void SystemZInstrInfo::loadRegFromStackSlot( 951 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned DestReg, 952 int FrameIdx, const TargetRegisterClass *RC, 953 const TargetRegisterInfo *TRI) const { 954 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 955 956 // Callers may expect a single instruction, so keep 128-bit moves 957 // together for now and lower them after register allocation. 958 unsigned LoadOpcode, StoreOpcode; 959 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 960 addFrameReference(BuildMI(MBB, MBBI, DL, get(LoadOpcode), DestReg), 961 FrameIdx); 962 } 963 964 // Return true if MI is a simple load or store with a 12-bit displacement 965 // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 966 static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) { 967 const MCInstrDesc &MCID = MI->getDesc(); 968 return ((MCID.TSFlags & Flag) && 969 isUInt<12>(MI->getOperand(2).getImm()) && 970 MI->getOperand(3).getReg() == 0); 971 } 972 973 namespace { 974 975 struct LogicOp { 976 LogicOp() = default; 977 LogicOp(unsigned regSize, unsigned immLSB, unsigned immSize) 978 : RegSize(regSize), ImmLSB(immLSB), ImmSize(immSize) {} 979 980 explicit operator bool() const { return RegSize; } 981 982 unsigned RegSize = 0; 983 unsigned ImmLSB = 0; 984 unsigned ImmSize = 0; 985 }; 986 987 } // end anonymous namespace 988 989 static LogicOp interpretAndImmediate(unsigned Opcode) { 990 switch (Opcode) { 991 case SystemZ::NILMux: return LogicOp(32, 0, 16); 992 case SystemZ::NIHMux: return LogicOp(32, 16, 16); 993 case SystemZ::NILL64: return LogicOp(64, 0, 16); 994 case SystemZ::NILH64: return LogicOp(64, 16, 16); 995 case SystemZ::NIHL64: return LogicOp(64, 32, 16); 996 case SystemZ::NIHH64: return LogicOp(64, 48, 16); 997 case SystemZ::NIFMux: return LogicOp(32, 0, 32); 998 case SystemZ::NILF64: return LogicOp(64, 0, 32); 999 case SystemZ::NIHF64: return LogicOp(64, 32, 32); 1000 default: return LogicOp(); 1001 } 1002 } 1003 1004 static void transferDeadCC(MachineInstr *OldMI, MachineInstr *NewMI) { 1005 if (OldMI->registerDefIsDead(SystemZ::CC)) { 1006 MachineOperand *CCDef = NewMI->findRegisterDefOperand(SystemZ::CC); 1007 if (CCDef != nullptr) 1008 CCDef->setIsDead(true); 1009 } 1010 } 1011 1012 // Used to return from convertToThreeAddress after replacing two-address 1013 // instruction OldMI with three-address instruction NewMI. 1014 static MachineInstr *finishConvertToThreeAddress(MachineInstr *OldMI, 1015 MachineInstr *NewMI, 1016 LiveVariables *LV) { 1017 if (LV) { 1018 unsigned NumOps = OldMI->getNumOperands(); 1019 for (unsigned I = 1; I < NumOps; ++I) { 1020 MachineOperand &Op = OldMI->getOperand(I); 1021 if (Op.isReg() && Op.isKill()) 1022 LV->replaceKillInstruction(Op.getReg(), *OldMI, *NewMI); 1023 } 1024 } 1025 transferDeadCC(OldMI, NewMI); 1026 return NewMI; 1027 } 1028 1029 MachineInstr *SystemZInstrInfo::convertToThreeAddress( 1030 MachineFunction::iterator &MFI, MachineInstr &MI, LiveVariables *LV) const { 1031 MachineBasicBlock *MBB = MI.getParent(); 1032 MachineFunction *MF = MBB->getParent(); 1033 MachineRegisterInfo &MRI = MF->getRegInfo(); 1034 1035 unsigned Opcode = MI.getOpcode(); 1036 unsigned NumOps = MI.getNumOperands(); 1037 1038 // Try to convert something like SLL into SLLK, if supported. 1039 // We prefer to keep the two-operand form where possible both 1040 // because it tends to be shorter and because some instructions 1041 // have memory forms that can be used during spilling. 1042 if (STI.hasDistinctOps()) { 1043 MachineOperand &Dest = MI.getOperand(0); 1044 MachineOperand &Src = MI.getOperand(1); 1045 unsigned DestReg = Dest.getReg(); 1046 unsigned SrcReg = Src.getReg(); 1047 // AHIMux is only really a three-operand instruction when both operands 1048 // are low registers. Try to constrain both operands to be low if 1049 // possible. 1050 if (Opcode == SystemZ::AHIMux && 1051 TargetRegisterInfo::isVirtualRegister(DestReg) && 1052 TargetRegisterInfo::isVirtualRegister(SrcReg) && 1053 MRI.getRegClass(DestReg)->contains(SystemZ::R1L) && 1054 MRI.getRegClass(SrcReg)->contains(SystemZ::R1L)) { 1055 MRI.constrainRegClass(DestReg, &SystemZ::GR32BitRegClass); 1056 MRI.constrainRegClass(SrcReg, &SystemZ::GR32BitRegClass); 1057 } 1058 int ThreeOperandOpcode = SystemZ::getThreeOperandOpcode(Opcode); 1059 if (ThreeOperandOpcode >= 0) { 1060 // Create three address instruction without adding the implicit 1061 // operands. Those will instead be copied over from the original 1062 // instruction by the loop below. 1063 MachineInstrBuilder MIB( 1064 *MF, MF->CreateMachineInstr(get(ThreeOperandOpcode), MI.getDebugLoc(), 1065 /*NoImplicit=*/true)); 1066 MIB.add(Dest); 1067 // Keep the kill state, but drop the tied flag. 1068 MIB.addReg(Src.getReg(), getKillRegState(Src.isKill()), Src.getSubReg()); 1069 // Keep the remaining operands as-is. 1070 for (unsigned I = 2; I < NumOps; ++I) 1071 MIB.add(MI.getOperand(I)); 1072 MBB->insert(MI, MIB); 1073 return finishConvertToThreeAddress(&MI, MIB, LV); 1074 } 1075 } 1076 1077 // Try to convert an AND into an RISBG-type instruction. 1078 if (LogicOp And = interpretAndImmediate(Opcode)) { 1079 uint64_t Imm = MI.getOperand(2).getImm() << And.ImmLSB; 1080 // AND IMMEDIATE leaves the other bits of the register unchanged. 1081 Imm |= allOnes(And.RegSize) & ~(allOnes(And.ImmSize) << And.ImmLSB); 1082 unsigned Start, End; 1083 if (isRxSBGMask(Imm, And.RegSize, Start, End)) { 1084 unsigned NewOpcode; 1085 if (And.RegSize == 64) { 1086 NewOpcode = SystemZ::RISBG; 1087 // Prefer RISBGN if available, since it does not clobber CC. 1088 if (STI.hasMiscellaneousExtensions()) 1089 NewOpcode = SystemZ::RISBGN; 1090 } else { 1091 NewOpcode = SystemZ::RISBMux; 1092 Start &= 31; 1093 End &= 31; 1094 } 1095 MachineOperand &Dest = MI.getOperand(0); 1096 MachineOperand &Src = MI.getOperand(1); 1097 MachineInstrBuilder MIB = 1098 BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpcode)) 1099 .add(Dest) 1100 .addReg(0) 1101 .addReg(Src.getReg(), getKillRegState(Src.isKill()), 1102 Src.getSubReg()) 1103 .addImm(Start) 1104 .addImm(End + 128) 1105 .addImm(0); 1106 return finishConvertToThreeAddress(&MI, MIB, LV); 1107 } 1108 } 1109 return nullptr; 1110 } 1111 1112 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1113 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1114 MachineBasicBlock::iterator InsertPt, int FrameIndex, 1115 LiveIntervals *LIS) const { 1116 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 1117 const MachineFrameInfo &MFI = MF.getFrameInfo(); 1118 unsigned Size = MFI.getObjectSize(FrameIndex); 1119 unsigned Opcode = MI.getOpcode(); 1120 1121 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 1122 if (LIS != nullptr && (Opcode == SystemZ::LA || Opcode == SystemZ::LAY) && 1123 isInt<8>(MI.getOperand(2).getImm()) && !MI.getOperand(3).getReg()) { 1124 1125 // Check CC liveness, since new instruction introduces a dead 1126 // def of CC. 1127 MCRegUnitIterator CCUnit(SystemZ::CC, TRI); 1128 LiveRange &CCLiveRange = LIS->getRegUnit(*CCUnit); 1129 ++CCUnit; 1130 assert(!CCUnit.isValid() && "CC only has one reg unit."); 1131 SlotIndex MISlot = 1132 LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot(); 1133 if (!CCLiveRange.liveAt(MISlot)) { 1134 // LA(Y) %reg, CONST(%reg) -> AGSI %mem, CONST 1135 MachineInstr *BuiltMI = BuildMI(*InsertPt->getParent(), InsertPt, 1136 MI.getDebugLoc(), get(SystemZ::AGSI)) 1137 .addFrameIndex(FrameIndex) 1138 .addImm(0) 1139 .addImm(MI.getOperand(2).getImm()); 1140 BuiltMI->findRegisterDefOperand(SystemZ::CC)->setIsDead(true); 1141 CCLiveRange.createDeadDef(MISlot, LIS->getVNInfoAllocator()); 1142 return BuiltMI; 1143 } 1144 } 1145 return nullptr; 1146 } 1147 1148 // All other cases require a single operand. 1149 if (Ops.size() != 1) 1150 return nullptr; 1151 1152 unsigned OpNum = Ops[0]; 1153 assert(Size * 8 == 1154 TRI->getRegSizeInBits(*MF.getRegInfo() 1155 .getRegClass(MI.getOperand(OpNum).getReg())) && 1156 "Invalid size combination"); 1157 1158 if ((Opcode == SystemZ::AHI || Opcode == SystemZ::AGHI) && OpNum == 0 && 1159 isInt<8>(MI.getOperand(2).getImm())) { 1160 // A(G)HI %reg, CONST -> A(G)SI %mem, CONST 1161 Opcode = (Opcode == SystemZ::AHI ? SystemZ::ASI : SystemZ::AGSI); 1162 MachineInstr *BuiltMI = 1163 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode)) 1164 .addFrameIndex(FrameIndex) 1165 .addImm(0) 1166 .addImm(MI.getOperand(2).getImm()); 1167 transferDeadCC(&MI, BuiltMI); 1168 return BuiltMI; 1169 } 1170 1171 if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) { 1172 bool Op0IsGPR = (Opcode == SystemZ::LGDR); 1173 bool Op1IsGPR = (Opcode == SystemZ::LDGR); 1174 // If we're spilling the destination of an LDGR or LGDR, store the 1175 // source register instead. 1176 if (OpNum == 0) { 1177 unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD; 1178 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1179 get(StoreOpcode)) 1180 .add(MI.getOperand(1)) 1181 .addFrameIndex(FrameIndex) 1182 .addImm(0) 1183 .addReg(0); 1184 } 1185 // If we're spilling the source of an LDGR or LGDR, load the 1186 // destination register instead. 1187 if (OpNum == 1) { 1188 unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD; 1189 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1190 get(LoadOpcode)) 1191 .add(MI.getOperand(0)) 1192 .addFrameIndex(FrameIndex) 1193 .addImm(0) 1194 .addReg(0); 1195 } 1196 } 1197 1198 // Look for cases where the source of a simple store or the destination 1199 // of a simple load is being spilled. Try to use MVC instead. 1200 // 1201 // Although MVC is in practice a fast choice in these cases, it is still 1202 // logically a bytewise copy. This means that we cannot use it if the 1203 // load or store is volatile. We also wouldn't be able to use MVC if 1204 // the two memories partially overlap, but that case cannot occur here, 1205 // because we know that one of the memories is a full frame index. 1206 // 1207 // For performance reasons, we also want to avoid using MVC if the addresses 1208 // might be equal. We don't worry about that case here, because spill slot 1209 // coloring happens later, and because we have special code to remove 1210 // MVCs that turn out to be redundant. 1211 if (OpNum == 0 && MI.hasOneMemOperand()) { 1212 MachineMemOperand *MMO = *MI.memoperands_begin(); 1213 if (MMO->getSize() == Size && !MMO->isVolatile()) { 1214 // Handle conversion of loads. 1215 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXLoad)) { 1216 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1217 get(SystemZ::MVC)) 1218 .addFrameIndex(FrameIndex) 1219 .addImm(0) 1220 .addImm(Size) 1221 .add(MI.getOperand(1)) 1222 .addImm(MI.getOperand(2).getImm()) 1223 .addMemOperand(MMO); 1224 } 1225 // Handle conversion of stores. 1226 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXStore)) { 1227 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1228 get(SystemZ::MVC)) 1229 .add(MI.getOperand(1)) 1230 .addImm(MI.getOperand(2).getImm()) 1231 .addImm(Size) 1232 .addFrameIndex(FrameIndex) 1233 .addImm(0) 1234 .addMemOperand(MMO); 1235 } 1236 } 1237 } 1238 1239 // If the spilled operand is the final one, try to change <INSN>R 1240 // into <INSN>. 1241 int MemOpcode = SystemZ::getMemOpcode(Opcode); 1242 if (MemOpcode >= 0) { 1243 unsigned NumOps = MI.getNumExplicitOperands(); 1244 if (OpNum == NumOps - 1) { 1245 const MCInstrDesc &MemDesc = get(MemOpcode); 1246 uint64_t AccessBytes = SystemZII::getAccessSize(MemDesc.TSFlags); 1247 assert(AccessBytes != 0 && "Size of access should be known"); 1248 assert(AccessBytes <= Size && "Access outside the frame index"); 1249 uint64_t Offset = Size - AccessBytes; 1250 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt, 1251 MI.getDebugLoc(), get(MemOpcode)); 1252 for (unsigned I = 0; I < OpNum; ++I) 1253 MIB.add(MI.getOperand(I)); 1254 MIB.addFrameIndex(FrameIndex).addImm(Offset); 1255 if (MemDesc.TSFlags & SystemZII::HasIndex) 1256 MIB.addReg(0); 1257 transferDeadCC(&MI, MIB); 1258 return MIB; 1259 } 1260 } 1261 1262 return nullptr; 1263 } 1264 1265 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1266 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1267 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, 1268 LiveIntervals *LIS) const { 1269 return nullptr; 1270 } 1271 1272 bool SystemZInstrInfo::expandPostRAPseudo(MachineInstr &MI) const { 1273 switch (MI.getOpcode()) { 1274 case SystemZ::L128: 1275 splitMove(MI, SystemZ::LG); 1276 return true; 1277 1278 case SystemZ::ST128: 1279 splitMove(MI, SystemZ::STG); 1280 return true; 1281 1282 case SystemZ::LX: 1283 splitMove(MI, SystemZ::LD); 1284 return true; 1285 1286 case SystemZ::STX: 1287 splitMove(MI, SystemZ::STD); 1288 return true; 1289 1290 case SystemZ::LBMux: 1291 expandRXYPseudo(MI, SystemZ::LB, SystemZ::LBH); 1292 return true; 1293 1294 case SystemZ::LHMux: 1295 expandRXYPseudo(MI, SystemZ::LH, SystemZ::LHH); 1296 return true; 1297 1298 case SystemZ::LLCRMux: 1299 expandZExtPseudo(MI, SystemZ::LLCR, 8); 1300 return true; 1301 1302 case SystemZ::LLHRMux: 1303 expandZExtPseudo(MI, SystemZ::LLHR, 16); 1304 return true; 1305 1306 case SystemZ::LLCMux: 1307 expandRXYPseudo(MI, SystemZ::LLC, SystemZ::LLCH); 1308 return true; 1309 1310 case SystemZ::LLHMux: 1311 expandRXYPseudo(MI, SystemZ::LLH, SystemZ::LLHH); 1312 return true; 1313 1314 case SystemZ::LMux: 1315 expandRXYPseudo(MI, SystemZ::L, SystemZ::LFH); 1316 return true; 1317 1318 case SystemZ::LOCMux: 1319 expandLOCPseudo(MI, SystemZ::LOC, SystemZ::LOCFH); 1320 return true; 1321 1322 case SystemZ::LOCHIMux: 1323 expandLOCPseudo(MI, SystemZ::LOCHI, SystemZ::LOCHHI); 1324 return true; 1325 1326 case SystemZ::LOCRMux: 1327 expandLOCRPseudo(MI, SystemZ::LOCR, SystemZ::LOCFHR); 1328 return true; 1329 1330 case SystemZ::STCMux: 1331 expandRXYPseudo(MI, SystemZ::STC, SystemZ::STCH); 1332 return true; 1333 1334 case SystemZ::STHMux: 1335 expandRXYPseudo(MI, SystemZ::STH, SystemZ::STHH); 1336 return true; 1337 1338 case SystemZ::STMux: 1339 expandRXYPseudo(MI, SystemZ::ST, SystemZ::STFH); 1340 return true; 1341 1342 case SystemZ::STOCMux: 1343 expandLOCPseudo(MI, SystemZ::STOC, SystemZ::STOCFH); 1344 return true; 1345 1346 case SystemZ::LHIMux: 1347 expandRIPseudo(MI, SystemZ::LHI, SystemZ::IIHF, true); 1348 return true; 1349 1350 case SystemZ::IIFMux: 1351 expandRIPseudo(MI, SystemZ::IILF, SystemZ::IIHF, false); 1352 return true; 1353 1354 case SystemZ::IILMux: 1355 expandRIPseudo(MI, SystemZ::IILL, SystemZ::IIHL, false); 1356 return true; 1357 1358 case SystemZ::IIHMux: 1359 expandRIPseudo(MI, SystemZ::IILH, SystemZ::IIHH, false); 1360 return true; 1361 1362 case SystemZ::NIFMux: 1363 expandRIPseudo(MI, SystemZ::NILF, SystemZ::NIHF, false); 1364 return true; 1365 1366 case SystemZ::NILMux: 1367 expandRIPseudo(MI, SystemZ::NILL, SystemZ::NIHL, false); 1368 return true; 1369 1370 case SystemZ::NIHMux: 1371 expandRIPseudo(MI, SystemZ::NILH, SystemZ::NIHH, false); 1372 return true; 1373 1374 case SystemZ::OIFMux: 1375 expandRIPseudo(MI, SystemZ::OILF, SystemZ::OIHF, false); 1376 return true; 1377 1378 case SystemZ::OILMux: 1379 expandRIPseudo(MI, SystemZ::OILL, SystemZ::OIHL, false); 1380 return true; 1381 1382 case SystemZ::OIHMux: 1383 expandRIPseudo(MI, SystemZ::OILH, SystemZ::OIHH, false); 1384 return true; 1385 1386 case SystemZ::XIFMux: 1387 expandRIPseudo(MI, SystemZ::XILF, SystemZ::XIHF, false); 1388 return true; 1389 1390 case SystemZ::TMLMux: 1391 expandRIPseudo(MI, SystemZ::TMLL, SystemZ::TMHL, false); 1392 return true; 1393 1394 case SystemZ::TMHMux: 1395 expandRIPseudo(MI, SystemZ::TMLH, SystemZ::TMHH, false); 1396 return true; 1397 1398 case SystemZ::AHIMux: 1399 expandRIPseudo(MI, SystemZ::AHI, SystemZ::AIH, false); 1400 return true; 1401 1402 case SystemZ::AHIMuxK: 1403 expandRIEPseudo(MI, SystemZ::AHI, SystemZ::AHIK, SystemZ::AIH); 1404 return true; 1405 1406 case SystemZ::AFIMux: 1407 expandRIPseudo(MI, SystemZ::AFI, SystemZ::AIH, false); 1408 return true; 1409 1410 case SystemZ::CHIMux: 1411 expandRIPseudo(MI, SystemZ::CHI, SystemZ::CIH, false); 1412 return true; 1413 1414 case SystemZ::CFIMux: 1415 expandRIPseudo(MI, SystemZ::CFI, SystemZ::CIH, false); 1416 return true; 1417 1418 case SystemZ::CLFIMux: 1419 expandRIPseudo(MI, SystemZ::CLFI, SystemZ::CLIH, false); 1420 return true; 1421 1422 case SystemZ::CMux: 1423 expandRXYPseudo(MI, SystemZ::C, SystemZ::CHF); 1424 return true; 1425 1426 case SystemZ::CLMux: 1427 expandRXYPseudo(MI, SystemZ::CL, SystemZ::CLHF); 1428 return true; 1429 1430 case SystemZ::RISBMux: { 1431 bool DestIsHigh = isHighReg(MI.getOperand(0).getReg()); 1432 bool SrcIsHigh = isHighReg(MI.getOperand(2).getReg()); 1433 if (SrcIsHigh == DestIsHigh) 1434 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHH : SystemZ::RISBLL)); 1435 else { 1436 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHL : SystemZ::RISBLH)); 1437 MI.getOperand(5).setImm(MI.getOperand(5).getImm() ^ 32); 1438 } 1439 return true; 1440 } 1441 1442 case SystemZ::ADJDYNALLOC: 1443 splitAdjDynAlloc(MI); 1444 return true; 1445 1446 case TargetOpcode::LOAD_STACK_GUARD: 1447 expandLoadStackGuard(&MI); 1448 return true; 1449 1450 default: 1451 return false; 1452 } 1453 } 1454 1455 unsigned SystemZInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { 1456 if (MI.getOpcode() == TargetOpcode::INLINEASM) { 1457 const MachineFunction *MF = MI.getParent()->getParent(); 1458 const char *AsmStr = MI.getOperand(0).getSymbolName(); 1459 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo()); 1460 } 1461 return MI.getDesc().getSize(); 1462 } 1463 1464 SystemZII::Branch 1465 SystemZInstrInfo::getBranchInfo(const MachineInstr &MI) const { 1466 switch (MI.getOpcode()) { 1467 case SystemZ::BR: 1468 case SystemZ::BI: 1469 case SystemZ::J: 1470 case SystemZ::JG: 1471 return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY, 1472 SystemZ::CCMASK_ANY, &MI.getOperand(0)); 1473 1474 case SystemZ::BRC: 1475 case SystemZ::BRCL: 1476 return SystemZII::Branch(SystemZII::BranchNormal, MI.getOperand(0).getImm(), 1477 MI.getOperand(1).getImm(), &MI.getOperand(2)); 1478 1479 case SystemZ::BRCT: 1480 case SystemZ::BRCTH: 1481 return SystemZII::Branch(SystemZII::BranchCT, SystemZ::CCMASK_ICMP, 1482 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1483 1484 case SystemZ::BRCTG: 1485 return SystemZII::Branch(SystemZII::BranchCTG, SystemZ::CCMASK_ICMP, 1486 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1487 1488 case SystemZ::CIJ: 1489 case SystemZ::CRJ: 1490 return SystemZII::Branch(SystemZII::BranchC, SystemZ::CCMASK_ICMP, 1491 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1492 1493 case SystemZ::CLIJ: 1494 case SystemZ::CLRJ: 1495 return SystemZII::Branch(SystemZII::BranchCL, SystemZ::CCMASK_ICMP, 1496 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1497 1498 case SystemZ::CGIJ: 1499 case SystemZ::CGRJ: 1500 return SystemZII::Branch(SystemZII::BranchCG, SystemZ::CCMASK_ICMP, 1501 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1502 1503 case SystemZ::CLGIJ: 1504 case SystemZ::CLGRJ: 1505 return SystemZII::Branch(SystemZII::BranchCLG, SystemZ::CCMASK_ICMP, 1506 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1507 1508 default: 1509 llvm_unreachable("Unrecognized branch opcode"); 1510 } 1511 } 1512 1513 void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC, 1514 unsigned &LoadOpcode, 1515 unsigned &StoreOpcode) const { 1516 if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) { 1517 LoadOpcode = SystemZ::L; 1518 StoreOpcode = SystemZ::ST; 1519 } else if (RC == &SystemZ::GRH32BitRegClass) { 1520 LoadOpcode = SystemZ::LFH; 1521 StoreOpcode = SystemZ::STFH; 1522 } else if (RC == &SystemZ::GRX32BitRegClass) { 1523 LoadOpcode = SystemZ::LMux; 1524 StoreOpcode = SystemZ::STMux; 1525 } else if (RC == &SystemZ::GR64BitRegClass || 1526 RC == &SystemZ::ADDR64BitRegClass) { 1527 LoadOpcode = SystemZ::LG; 1528 StoreOpcode = SystemZ::STG; 1529 } else if (RC == &SystemZ::GR128BitRegClass || 1530 RC == &SystemZ::ADDR128BitRegClass) { 1531 LoadOpcode = SystemZ::L128; 1532 StoreOpcode = SystemZ::ST128; 1533 } else if (RC == &SystemZ::FP32BitRegClass) { 1534 LoadOpcode = SystemZ::LE; 1535 StoreOpcode = SystemZ::STE; 1536 } else if (RC == &SystemZ::FP64BitRegClass) { 1537 LoadOpcode = SystemZ::LD; 1538 StoreOpcode = SystemZ::STD; 1539 } else if (RC == &SystemZ::FP128BitRegClass) { 1540 LoadOpcode = SystemZ::LX; 1541 StoreOpcode = SystemZ::STX; 1542 } else if (RC == &SystemZ::VR32BitRegClass) { 1543 LoadOpcode = SystemZ::VL32; 1544 StoreOpcode = SystemZ::VST32; 1545 } else if (RC == &SystemZ::VR64BitRegClass) { 1546 LoadOpcode = SystemZ::VL64; 1547 StoreOpcode = SystemZ::VST64; 1548 } else if (RC == &SystemZ::VF128BitRegClass || 1549 RC == &SystemZ::VR128BitRegClass) { 1550 LoadOpcode = SystemZ::VL; 1551 StoreOpcode = SystemZ::VST; 1552 } else 1553 llvm_unreachable("Unsupported regclass to load or store"); 1554 } 1555 1556 unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode, 1557 int64_t Offset) const { 1558 const MCInstrDesc &MCID = get(Opcode); 1559 int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset); 1560 if (isUInt<12>(Offset) && isUInt<12>(Offset2)) { 1561 // Get the instruction to use for unsigned 12-bit displacements. 1562 int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode); 1563 if (Disp12Opcode >= 0) 1564 return Disp12Opcode; 1565 1566 // All address-related instructions can use unsigned 12-bit 1567 // displacements. 1568 return Opcode; 1569 } 1570 if (isInt<20>(Offset) && isInt<20>(Offset2)) { 1571 // Get the instruction to use for signed 20-bit displacements. 1572 int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode); 1573 if (Disp20Opcode >= 0) 1574 return Disp20Opcode; 1575 1576 // Check whether Opcode allows signed 20-bit displacements. 1577 if (MCID.TSFlags & SystemZII::Has20BitOffset) 1578 return Opcode; 1579 } 1580 return 0; 1581 } 1582 1583 unsigned SystemZInstrInfo::getLoadAndTest(unsigned Opcode) const { 1584 switch (Opcode) { 1585 case SystemZ::L: return SystemZ::LT; 1586 case SystemZ::LY: return SystemZ::LT; 1587 case SystemZ::LG: return SystemZ::LTG; 1588 case SystemZ::LGF: return SystemZ::LTGF; 1589 case SystemZ::LR: return SystemZ::LTR; 1590 case SystemZ::LGFR: return SystemZ::LTGFR; 1591 case SystemZ::LGR: return SystemZ::LTGR; 1592 case SystemZ::LER: return SystemZ::LTEBR; 1593 case SystemZ::LDR: return SystemZ::LTDBR; 1594 case SystemZ::LXR: return SystemZ::LTXBR; 1595 case SystemZ::LCDFR: return SystemZ::LCDBR; 1596 case SystemZ::LPDFR: return SystemZ::LPDBR; 1597 case SystemZ::LNDFR: return SystemZ::LNDBR; 1598 case SystemZ::LCDFR_32: return SystemZ::LCEBR; 1599 case SystemZ::LPDFR_32: return SystemZ::LPEBR; 1600 case SystemZ::LNDFR_32: return SystemZ::LNEBR; 1601 // On zEC12 we prefer to use RISBGN. But if there is a chance to 1602 // actually use the condition code, we may turn it back into RISGB. 1603 // Note that RISBG is not really a "load-and-test" instruction, 1604 // but sets the same condition code values, so is OK to use here. 1605 case SystemZ::RISBGN: return SystemZ::RISBG; 1606 default: return 0; 1607 } 1608 } 1609 1610 // Return true if Mask matches the regexp 0*1+0*, given that zero masks 1611 // have already been filtered out. Store the first set bit in LSB and 1612 // the number of set bits in Length if so. 1613 static bool isStringOfOnes(uint64_t Mask, unsigned &LSB, unsigned &Length) { 1614 unsigned First = findFirstSet(Mask); 1615 uint64_t Top = (Mask >> First) + 1; 1616 if ((Top & -Top) == Top) { 1617 LSB = First; 1618 Length = findFirstSet(Top); 1619 return true; 1620 } 1621 return false; 1622 } 1623 1624 bool SystemZInstrInfo::isRxSBGMask(uint64_t Mask, unsigned BitSize, 1625 unsigned &Start, unsigned &End) const { 1626 // Reject trivial all-zero masks. 1627 Mask &= allOnes(BitSize); 1628 if (Mask == 0) 1629 return false; 1630 1631 // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of 1632 // the msb and End specifies the index of the lsb. 1633 unsigned LSB, Length; 1634 if (isStringOfOnes(Mask, LSB, Length)) { 1635 Start = 63 - (LSB + Length - 1); 1636 End = 63 - LSB; 1637 return true; 1638 } 1639 1640 // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb 1641 // of the low 1s and End specifies the lsb of the high 1s. 1642 if (isStringOfOnes(Mask ^ allOnes(BitSize), LSB, Length)) { 1643 assert(LSB > 0 && "Bottom bit must be set"); 1644 assert(LSB + Length < BitSize && "Top bit must be set"); 1645 Start = 63 - (LSB - 1); 1646 End = 63 - (LSB + Length); 1647 return true; 1648 } 1649 1650 return false; 1651 } 1652 1653 unsigned SystemZInstrInfo::getFusedCompare(unsigned Opcode, 1654 SystemZII::FusedCompareType Type, 1655 const MachineInstr *MI) const { 1656 switch (Opcode) { 1657 case SystemZ::CHI: 1658 case SystemZ::CGHI: 1659 if (!(MI && isInt<8>(MI->getOperand(1).getImm()))) 1660 return 0; 1661 break; 1662 case SystemZ::CLFI: 1663 case SystemZ::CLGFI: 1664 if (!(MI && isUInt<8>(MI->getOperand(1).getImm()))) 1665 return 0; 1666 break; 1667 case SystemZ::CL: 1668 case SystemZ::CLG: 1669 if (!STI.hasMiscellaneousExtensions()) 1670 return 0; 1671 if (!(MI && MI->getOperand(3).getReg() == 0)) 1672 return 0; 1673 break; 1674 } 1675 switch (Type) { 1676 case SystemZII::CompareAndBranch: 1677 switch (Opcode) { 1678 case SystemZ::CR: 1679 return SystemZ::CRJ; 1680 case SystemZ::CGR: 1681 return SystemZ::CGRJ; 1682 case SystemZ::CHI: 1683 return SystemZ::CIJ; 1684 case SystemZ::CGHI: 1685 return SystemZ::CGIJ; 1686 case SystemZ::CLR: 1687 return SystemZ::CLRJ; 1688 case SystemZ::CLGR: 1689 return SystemZ::CLGRJ; 1690 case SystemZ::CLFI: 1691 return SystemZ::CLIJ; 1692 case SystemZ::CLGFI: 1693 return SystemZ::CLGIJ; 1694 default: 1695 return 0; 1696 } 1697 case SystemZII::CompareAndReturn: 1698 switch (Opcode) { 1699 case SystemZ::CR: 1700 return SystemZ::CRBReturn; 1701 case SystemZ::CGR: 1702 return SystemZ::CGRBReturn; 1703 case SystemZ::CHI: 1704 return SystemZ::CIBReturn; 1705 case SystemZ::CGHI: 1706 return SystemZ::CGIBReturn; 1707 case SystemZ::CLR: 1708 return SystemZ::CLRBReturn; 1709 case SystemZ::CLGR: 1710 return SystemZ::CLGRBReturn; 1711 case SystemZ::CLFI: 1712 return SystemZ::CLIBReturn; 1713 case SystemZ::CLGFI: 1714 return SystemZ::CLGIBReturn; 1715 default: 1716 return 0; 1717 } 1718 case SystemZII::CompareAndSibcall: 1719 switch (Opcode) { 1720 case SystemZ::CR: 1721 return SystemZ::CRBCall; 1722 case SystemZ::CGR: 1723 return SystemZ::CGRBCall; 1724 case SystemZ::CHI: 1725 return SystemZ::CIBCall; 1726 case SystemZ::CGHI: 1727 return SystemZ::CGIBCall; 1728 case SystemZ::CLR: 1729 return SystemZ::CLRBCall; 1730 case SystemZ::CLGR: 1731 return SystemZ::CLGRBCall; 1732 case SystemZ::CLFI: 1733 return SystemZ::CLIBCall; 1734 case SystemZ::CLGFI: 1735 return SystemZ::CLGIBCall; 1736 default: 1737 return 0; 1738 } 1739 case SystemZII::CompareAndTrap: 1740 switch (Opcode) { 1741 case SystemZ::CR: 1742 return SystemZ::CRT; 1743 case SystemZ::CGR: 1744 return SystemZ::CGRT; 1745 case SystemZ::CHI: 1746 return SystemZ::CIT; 1747 case SystemZ::CGHI: 1748 return SystemZ::CGIT; 1749 case SystemZ::CLR: 1750 return SystemZ::CLRT; 1751 case SystemZ::CLGR: 1752 return SystemZ::CLGRT; 1753 case SystemZ::CLFI: 1754 return SystemZ::CLFIT; 1755 case SystemZ::CLGFI: 1756 return SystemZ::CLGIT; 1757 case SystemZ::CL: 1758 return SystemZ::CLT; 1759 case SystemZ::CLG: 1760 return SystemZ::CLGT; 1761 default: 1762 return 0; 1763 } 1764 } 1765 return 0; 1766 } 1767 1768 unsigned SystemZInstrInfo::getLoadAndTrap(unsigned Opcode) const { 1769 if (!STI.hasLoadAndTrap()) 1770 return 0; 1771 switch (Opcode) { 1772 case SystemZ::L: 1773 case SystemZ::LY: 1774 return SystemZ::LAT; 1775 case SystemZ::LG: 1776 return SystemZ::LGAT; 1777 case SystemZ::LFH: 1778 return SystemZ::LFHAT; 1779 case SystemZ::LLGF: 1780 return SystemZ::LLGFAT; 1781 case SystemZ::LLGT: 1782 return SystemZ::LLGTAT; 1783 } 1784 return 0; 1785 } 1786 1787 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB, 1788 MachineBasicBlock::iterator MBBI, 1789 unsigned Reg, uint64_t Value) const { 1790 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 1791 unsigned Opcode; 1792 if (isInt<16>(Value)) 1793 Opcode = SystemZ::LGHI; 1794 else if (SystemZ::isImmLL(Value)) 1795 Opcode = SystemZ::LLILL; 1796 else if (SystemZ::isImmLH(Value)) { 1797 Opcode = SystemZ::LLILH; 1798 Value >>= 16; 1799 } else { 1800 assert(isInt<32>(Value) && "Huge values not handled yet"); 1801 Opcode = SystemZ::LGFI; 1802 } 1803 BuildMI(MBB, MBBI, DL, get(Opcode), Reg).addImm(Value); 1804 } 1805 1806 bool SystemZInstrInfo:: 1807 areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb, 1808 AliasAnalysis *AA) const { 1809 1810 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand()) 1811 return false; 1812 1813 // If mem-operands show that the same address Value is used by both 1814 // instructions, check for non-overlapping offsets and widths. Not 1815 // sure if a register based analysis would be an improvement... 1816 1817 MachineMemOperand *MMOa = *MIa.memoperands_begin(); 1818 MachineMemOperand *MMOb = *MIb.memoperands_begin(); 1819 const Value *VALa = MMOa->getValue(); 1820 const Value *VALb = MMOb->getValue(); 1821 bool SameVal = (VALa && VALb && (VALa == VALb)); 1822 if (!SameVal) { 1823 const PseudoSourceValue *PSVa = MMOa->getPseudoValue(); 1824 const PseudoSourceValue *PSVb = MMOb->getPseudoValue(); 1825 if (PSVa && PSVb && (PSVa == PSVb)) 1826 SameVal = true; 1827 } 1828 if (SameVal) { 1829 int OffsetA = MMOa->getOffset(), OffsetB = MMOb->getOffset(); 1830 int WidthA = MMOa->getSize(), WidthB = MMOb->getSize(); 1831 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB; 1832 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA; 1833 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; 1834 if (LowOffset + LowWidth <= HighOffset) 1835 return true; 1836 } 1837 1838 return false; 1839 } 1840