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/ADT/Statistic.h" 20 #include "llvm/CodeGen/LiveInterval.h" 21 #include "llvm/CodeGen/LiveIntervals.h" 22 #include "llvm/CodeGen/LiveVariables.h" 23 #include "llvm/CodeGen/MachineBasicBlock.h" 24 #include "llvm/CodeGen/MachineFrameInfo.h" 25 #include "llvm/CodeGen/MachineFunction.h" 26 #include "llvm/CodeGen/MachineInstr.h" 27 #include "llvm/CodeGen/MachineMemOperand.h" 28 #include "llvm/CodeGen/MachineOperand.h" 29 #include "llvm/CodeGen/MachineRegisterInfo.h" 30 #include "llvm/CodeGen/SlotIndexes.h" 31 #include "llvm/CodeGen/TargetInstrInfo.h" 32 #include "llvm/CodeGen/TargetSubtargetInfo.h" 33 #include "llvm/MC/MCInstrDesc.h" 34 #include "llvm/MC/MCRegisterInfo.h" 35 #include "llvm/Support/BranchProbability.h" 36 #include "llvm/Support/ErrorHandling.h" 37 #include "llvm/Support/MathExtras.h" 38 #include "llvm/Target/TargetMachine.h" 39 #include <cassert> 40 #include <cstdint> 41 #include <iterator> 42 43 using namespace llvm; 44 45 #define GET_INSTRINFO_CTOR_DTOR 46 #define GET_INSTRMAP_INFO 47 #include "SystemZGenInstrInfo.inc" 48 49 #define DEBUG_TYPE "systemz-II" 50 STATISTIC(LOCRMuxJumps, "Number of LOCRMux jump-sequences (lower is better)"); 51 52 // Return a mask with Count low bits set. 53 static uint64_t allOnes(unsigned int Count) { 54 return Count == 0 ? 0 : (uint64_t(1) << (Count - 1) << 1) - 1; 55 } 56 57 // Reg should be a 32-bit GPR. Return true if it is a high register rather 58 // than a low register. 59 static bool isHighReg(unsigned int Reg) { 60 if (SystemZ::GRH32BitRegClass.contains(Reg)) 61 return true; 62 assert(SystemZ::GR32BitRegClass.contains(Reg) && "Invalid GRX32"); 63 return false; 64 } 65 66 // Pin the vtable to this file. 67 void SystemZInstrInfo::anchor() {} 68 69 SystemZInstrInfo::SystemZInstrInfo(SystemZSubtarget &sti) 70 : SystemZGenInstrInfo(SystemZ::ADJCALLSTACKDOWN, SystemZ::ADJCALLSTACKUP), 71 RI(), STI(sti) { 72 } 73 74 // MI is a 128-bit load or store. Split it into two 64-bit loads or stores, 75 // each having the opcode given by NewOpcode. 76 void SystemZInstrInfo::splitMove(MachineBasicBlock::iterator MI, 77 unsigned NewOpcode) const { 78 MachineBasicBlock *MBB = MI->getParent(); 79 MachineFunction &MF = *MBB->getParent(); 80 81 // Get two load or store instructions. Use the original instruction for one 82 // of them (arbitrarily the second here) and create a clone for the other. 83 MachineInstr *EarlierMI = MF.CloneMachineInstr(&*MI); 84 MBB->insert(MI, EarlierMI); 85 86 // Set up the two 64-bit registers and remember super reg and its flags. 87 MachineOperand &HighRegOp = EarlierMI->getOperand(0); 88 MachineOperand &LowRegOp = MI->getOperand(0); 89 unsigned Reg128 = LowRegOp.getReg(); 90 unsigned Reg128Killed = getKillRegState(LowRegOp.isKill()); 91 unsigned Reg128Undef = getUndefRegState(LowRegOp.isUndef()); 92 HighRegOp.setReg(RI.getSubReg(HighRegOp.getReg(), SystemZ::subreg_h64)); 93 LowRegOp.setReg(RI.getSubReg(LowRegOp.getReg(), SystemZ::subreg_l64)); 94 95 if (MI->mayStore()) { 96 // Add implicit uses of the super register in case one of the subregs is 97 // undefined. We could track liveness and skip storing an undefined 98 // subreg, but this is hopefully rare (discovered with llvm-stress). 99 // If Reg128 was killed, set kill flag on MI. 100 unsigned Reg128UndefImpl = (Reg128Undef | RegState::Implicit); 101 MachineInstrBuilder(MF, EarlierMI).addReg(Reg128, Reg128UndefImpl); 102 MachineInstrBuilder(MF, MI).addReg(Reg128, (Reg128UndefImpl | Reg128Killed)); 103 } 104 105 // The address in the first (high) instruction is already correct. 106 // Adjust the offset in the second (low) instruction. 107 MachineOperand &HighOffsetOp = EarlierMI->getOperand(2); 108 MachineOperand &LowOffsetOp = MI->getOperand(2); 109 LowOffsetOp.setImm(LowOffsetOp.getImm() + 8); 110 111 // Clear the kill flags on the registers in the first instruction. 112 if (EarlierMI->getOperand(0).isReg() && EarlierMI->getOperand(0).isUse()) 113 EarlierMI->getOperand(0).setIsKill(false); 114 EarlierMI->getOperand(1).setIsKill(false); 115 EarlierMI->getOperand(3).setIsKill(false); 116 117 // Set the opcodes. 118 unsigned HighOpcode = getOpcodeForOffset(NewOpcode, HighOffsetOp.getImm()); 119 unsigned LowOpcode = getOpcodeForOffset(NewOpcode, LowOffsetOp.getImm()); 120 assert(HighOpcode && LowOpcode && "Both offsets should be in range"); 121 122 EarlierMI->setDesc(get(HighOpcode)); 123 MI->setDesc(get(LowOpcode)); 124 } 125 126 // Split ADJDYNALLOC instruction MI. 127 void SystemZInstrInfo::splitAdjDynAlloc(MachineBasicBlock::iterator MI) const { 128 MachineBasicBlock *MBB = MI->getParent(); 129 MachineFunction &MF = *MBB->getParent(); 130 MachineFrameInfo &MFFrame = MF.getFrameInfo(); 131 MachineOperand &OffsetMO = MI->getOperand(2); 132 133 uint64_t Offset = (MFFrame.getMaxCallFrameSize() + 134 SystemZMC::CallFrameSize + 135 OffsetMO.getImm()); 136 unsigned NewOpcode = getOpcodeForOffset(SystemZ::LA, Offset); 137 assert(NewOpcode && "No support for huge argument lists yet"); 138 MI->setDesc(get(NewOpcode)); 139 OffsetMO.setImm(Offset); 140 } 141 142 // MI is an RI-style pseudo instruction. Replace it with LowOpcode 143 // if the first operand is a low GR32 and HighOpcode if the first operand 144 // is a high GR32. ConvertHigh is true if LowOpcode takes a signed operand 145 // and HighOpcode takes an unsigned 32-bit operand. In those cases, 146 // MI has the same kind of operand as LowOpcode, so needs to be converted 147 // if HighOpcode is used. 148 void SystemZInstrInfo::expandRIPseudo(MachineInstr &MI, unsigned LowOpcode, 149 unsigned HighOpcode, 150 bool ConvertHigh) const { 151 unsigned Reg = MI.getOperand(0).getReg(); 152 bool IsHigh = isHighReg(Reg); 153 MI.setDesc(get(IsHigh ? HighOpcode : LowOpcode)); 154 if (IsHigh && ConvertHigh) 155 MI.getOperand(1).setImm(uint32_t(MI.getOperand(1).getImm())); 156 } 157 158 // MI is a three-operand RIE-style pseudo instruction. Replace it with 159 // LowOpcodeK if the registers are both low GR32s, otherwise use a move 160 // followed by HighOpcode or LowOpcode, depending on whether the target 161 // is a high or low GR32. 162 void SystemZInstrInfo::expandRIEPseudo(MachineInstr &MI, unsigned LowOpcode, 163 unsigned LowOpcodeK, 164 unsigned HighOpcode) const { 165 unsigned DestReg = MI.getOperand(0).getReg(); 166 unsigned SrcReg = MI.getOperand(1).getReg(); 167 bool DestIsHigh = isHighReg(DestReg); 168 bool SrcIsHigh = isHighReg(SrcReg); 169 if (!DestIsHigh && !SrcIsHigh) 170 MI.setDesc(get(LowOpcodeK)); 171 else { 172 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), DestReg, SrcReg, 173 SystemZ::LR, 32, MI.getOperand(1).isKill(), 174 MI.getOperand(1).isUndef()); 175 MI.setDesc(get(DestIsHigh ? HighOpcode : LowOpcode)); 176 MI.getOperand(1).setReg(DestReg); 177 MI.tieOperands(0, 1); 178 } 179 } 180 181 // MI is an RXY-style pseudo instruction. Replace it with LowOpcode 182 // if the first operand is a low GR32 and HighOpcode if the first operand 183 // is a high GR32. 184 void SystemZInstrInfo::expandRXYPseudo(MachineInstr &MI, unsigned LowOpcode, 185 unsigned HighOpcode) const { 186 unsigned Reg = MI.getOperand(0).getReg(); 187 unsigned Opcode = getOpcodeForOffset(isHighReg(Reg) ? HighOpcode : LowOpcode, 188 MI.getOperand(2).getImm()); 189 MI.setDesc(get(Opcode)); 190 } 191 192 // MI is a load-on-condition pseudo instruction with a single register 193 // (source or destination) operand. Replace it with LowOpcode if the 194 // register is a low GR32 and HighOpcode if the register is a high GR32. 195 void SystemZInstrInfo::expandLOCPseudo(MachineInstr &MI, unsigned LowOpcode, 196 unsigned HighOpcode) const { 197 unsigned Reg = MI.getOperand(0).getReg(); 198 unsigned Opcode = isHighReg(Reg) ? HighOpcode : LowOpcode; 199 MI.setDesc(get(Opcode)); 200 } 201 202 // MI is a load-register-on-condition pseudo instruction. Replace it with 203 // LowOpcode if source and destination are both low GR32s and HighOpcode if 204 // source and destination are both high GR32s. 205 void SystemZInstrInfo::expandLOCRPseudo(MachineInstr &MI, unsigned LowOpcode, 206 unsigned HighOpcode) const { 207 unsigned DestReg = MI.getOperand(0).getReg(); 208 unsigned SrcReg = MI.getOperand(2).getReg(); 209 bool DestIsHigh = isHighReg(DestReg); 210 bool SrcIsHigh = isHighReg(SrcReg); 211 212 if (!DestIsHigh && !SrcIsHigh) 213 MI.setDesc(get(LowOpcode)); 214 else if (DestIsHigh && SrcIsHigh) 215 MI.setDesc(get(HighOpcode)); 216 else 217 LOCRMuxJumps++; 218 219 // If we were unable to implement the pseudo with a single instruction, we 220 // need to convert it back into a branch sequence. This cannot be done here 221 // since the caller of expandPostRAPseudo does not handle changes to the CFG 222 // correctly. This change is defered to the SystemZExpandPseudo pass. 223 } 224 225 // MI is an RR-style pseudo instruction that zero-extends the low Size bits 226 // of one GRX32 into another. Replace it with LowOpcode if both operands 227 // are low registers, otherwise use RISB[LH]G. 228 void SystemZInstrInfo::expandZExtPseudo(MachineInstr &MI, unsigned LowOpcode, 229 unsigned Size) const { 230 MachineInstrBuilder MIB = 231 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), 232 MI.getOperand(0).getReg(), MI.getOperand(1).getReg(), LowOpcode, 233 Size, MI.getOperand(1).isKill(), MI.getOperand(1).isUndef()); 234 235 // Keep the remaining operands as-is. 236 for (unsigned I = 2; I < MI.getNumOperands(); ++I) 237 MIB.add(MI.getOperand(I)); 238 239 MI.eraseFromParent(); 240 } 241 242 void SystemZInstrInfo::expandLoadStackGuard(MachineInstr *MI) const { 243 MachineBasicBlock *MBB = MI->getParent(); 244 MachineFunction &MF = *MBB->getParent(); 245 const unsigned Reg64 = MI->getOperand(0).getReg(); 246 const unsigned Reg32 = RI.getSubReg(Reg64, SystemZ::subreg_l32); 247 248 // EAR can only load the low subregister so us a shift for %a0 to produce 249 // the GR containing %a0 and %a1. 250 251 // ear <reg>, %a0 252 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::EAR), Reg32) 253 .addReg(SystemZ::A0) 254 .addReg(Reg64, RegState::ImplicitDefine); 255 256 // sllg <reg>, <reg>, 32 257 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::SLLG), Reg64) 258 .addReg(Reg64) 259 .addReg(0) 260 .addImm(32); 261 262 // ear <reg>, %a1 263 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::EAR), Reg32) 264 .addReg(SystemZ::A1); 265 266 // lg <reg>, 40(<reg>) 267 MI->setDesc(get(SystemZ::LG)); 268 MachineInstrBuilder(MF, MI).addReg(Reg64).addImm(40).addReg(0); 269 } 270 271 // Emit a zero-extending move from 32-bit GPR SrcReg to 32-bit GPR 272 // DestReg before MBBI in MBB. Use LowLowOpcode when both DestReg and SrcReg 273 // are low registers, otherwise use RISB[LH]G. Size is the number of bits 274 // taken from the low end of SrcReg (8 for LLCR, 16 for LLHR and 32 for LR). 275 // KillSrc is true if this move is the last use of SrcReg. 276 MachineInstrBuilder 277 SystemZInstrInfo::emitGRX32Move(MachineBasicBlock &MBB, 278 MachineBasicBlock::iterator MBBI, 279 const DebugLoc &DL, unsigned DestReg, 280 unsigned SrcReg, unsigned LowLowOpcode, 281 unsigned Size, bool KillSrc, 282 bool UndefSrc) const { 283 unsigned Opcode; 284 bool DestIsHigh = isHighReg(DestReg); 285 bool SrcIsHigh = isHighReg(SrcReg); 286 if (DestIsHigh && SrcIsHigh) 287 Opcode = SystemZ::RISBHH; 288 else if (DestIsHigh && !SrcIsHigh) 289 Opcode = SystemZ::RISBHL; 290 else if (!DestIsHigh && SrcIsHigh) 291 Opcode = SystemZ::RISBLH; 292 else { 293 return BuildMI(MBB, MBBI, DL, get(LowLowOpcode), DestReg) 294 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)); 295 } 296 unsigned Rotate = (DestIsHigh != SrcIsHigh ? 32 : 0); 297 return BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 298 .addReg(DestReg, RegState::Undef) 299 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)) 300 .addImm(32 - Size).addImm(128 + 31).addImm(Rotate); 301 } 302 303 MachineInstr *SystemZInstrInfo::commuteInstructionImpl(MachineInstr &MI, 304 bool NewMI, 305 unsigned OpIdx1, 306 unsigned OpIdx2) const { 307 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & { 308 if (NewMI) 309 return *MI.getParent()->getParent()->CloneMachineInstr(&MI); 310 return MI; 311 }; 312 313 switch (MI.getOpcode()) { 314 case SystemZ::LOCRMux: 315 case SystemZ::LOCFHR: 316 case SystemZ::LOCR: 317 case SystemZ::LOCGR: { 318 auto &WorkingMI = cloneIfNew(MI); 319 // Invert condition. 320 unsigned CCValid = WorkingMI.getOperand(3).getImm(); 321 unsigned CCMask = WorkingMI.getOperand(4).getImm(); 322 WorkingMI.getOperand(4).setImm(CCMask ^ CCValid); 323 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 324 OpIdx1, OpIdx2); 325 } 326 default: 327 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 328 } 329 } 330 331 // If MI is a simple load or store for a frame object, return the register 332 // it loads or stores and set FrameIndex to the index of the frame object. 333 // Return 0 otherwise. 334 // 335 // Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 336 static int isSimpleMove(const MachineInstr &MI, int &FrameIndex, 337 unsigned Flag) { 338 const MCInstrDesc &MCID = MI.getDesc(); 339 if ((MCID.TSFlags & Flag) && MI.getOperand(1).isFI() && 340 MI.getOperand(2).getImm() == 0 && MI.getOperand(3).getReg() == 0) { 341 FrameIndex = MI.getOperand(1).getIndex(); 342 return MI.getOperand(0).getReg(); 343 } 344 return 0; 345 } 346 347 unsigned SystemZInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 348 int &FrameIndex) const { 349 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXLoad); 350 } 351 352 unsigned SystemZInstrInfo::isStoreToStackSlot(const MachineInstr &MI, 353 int &FrameIndex) const { 354 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXStore); 355 } 356 357 bool SystemZInstrInfo::isStackSlotCopy(const MachineInstr &MI, 358 int &DestFrameIndex, 359 int &SrcFrameIndex) const { 360 // Check for MVC 0(Length,FI1),0(FI2) 361 const MachineFrameInfo &MFI = MI.getParent()->getParent()->getFrameInfo(); 362 if (MI.getOpcode() != SystemZ::MVC || !MI.getOperand(0).isFI() || 363 MI.getOperand(1).getImm() != 0 || !MI.getOperand(3).isFI() || 364 MI.getOperand(4).getImm() != 0) 365 return false; 366 367 // Check that Length covers the full slots. 368 int64_t Length = MI.getOperand(2).getImm(); 369 unsigned FI1 = MI.getOperand(0).getIndex(); 370 unsigned FI2 = MI.getOperand(3).getIndex(); 371 if (MFI.getObjectSize(FI1) != Length || 372 MFI.getObjectSize(FI2) != Length) 373 return false; 374 375 DestFrameIndex = FI1; 376 SrcFrameIndex = FI2; 377 return true; 378 } 379 380 bool SystemZInstrInfo::analyzeBranch(MachineBasicBlock &MBB, 381 MachineBasicBlock *&TBB, 382 MachineBasicBlock *&FBB, 383 SmallVectorImpl<MachineOperand> &Cond, 384 bool AllowModify) const { 385 // Most of the code and comments here are boilerplate. 386 387 // Start from the bottom of the block and work up, examining the 388 // terminator instructions. 389 MachineBasicBlock::iterator I = MBB.end(); 390 while (I != MBB.begin()) { 391 --I; 392 if (I->isDebugValue()) 393 continue; 394 395 // Working from the bottom, when we see a non-terminator instruction, we're 396 // done. 397 if (!isUnpredicatedTerminator(*I)) 398 break; 399 400 // A terminator that isn't a branch can't easily be handled by this 401 // analysis. 402 if (!I->isBranch()) 403 return true; 404 405 // Can't handle indirect branches. 406 SystemZII::Branch Branch(getBranchInfo(*I)); 407 if (!Branch.Target->isMBB()) 408 return true; 409 410 // Punt on compound branches. 411 if (Branch.Type != SystemZII::BranchNormal) 412 return true; 413 414 if (Branch.CCMask == SystemZ::CCMASK_ANY) { 415 // Handle unconditional branches. 416 if (!AllowModify) { 417 TBB = Branch.Target->getMBB(); 418 continue; 419 } 420 421 // If the block has any instructions after a JMP, delete them. 422 while (std::next(I) != MBB.end()) 423 std::next(I)->eraseFromParent(); 424 425 Cond.clear(); 426 FBB = nullptr; 427 428 // Delete the JMP if it's equivalent to a fall-through. 429 if (MBB.isLayoutSuccessor(Branch.Target->getMBB())) { 430 TBB = nullptr; 431 I->eraseFromParent(); 432 I = MBB.end(); 433 continue; 434 } 435 436 // TBB is used to indicate the unconditinal destination. 437 TBB = Branch.Target->getMBB(); 438 continue; 439 } 440 441 // Working from the bottom, handle the first conditional branch. 442 if (Cond.empty()) { 443 // FIXME: add X86-style branch swap 444 FBB = TBB; 445 TBB = Branch.Target->getMBB(); 446 Cond.push_back(MachineOperand::CreateImm(Branch.CCValid)); 447 Cond.push_back(MachineOperand::CreateImm(Branch.CCMask)); 448 continue; 449 } 450 451 // Handle subsequent conditional branches. 452 assert(Cond.size() == 2 && TBB && "Should have seen a conditional branch"); 453 454 // Only handle the case where all conditional branches branch to the same 455 // destination. 456 if (TBB != Branch.Target->getMBB()) 457 return true; 458 459 // If the conditions are the same, we can leave them alone. 460 unsigned OldCCValid = Cond[0].getImm(); 461 unsigned OldCCMask = Cond[1].getImm(); 462 if (OldCCValid == Branch.CCValid && OldCCMask == Branch.CCMask) 463 continue; 464 465 // FIXME: Try combining conditions like X86 does. Should be easy on Z! 466 return false; 467 } 468 469 return false; 470 } 471 472 unsigned SystemZInstrInfo::removeBranch(MachineBasicBlock &MBB, 473 int *BytesRemoved) const { 474 assert(!BytesRemoved && "code size not handled"); 475 476 // Most of the code and comments here are boilerplate. 477 MachineBasicBlock::iterator I = MBB.end(); 478 unsigned Count = 0; 479 480 while (I != MBB.begin()) { 481 --I; 482 if (I->isDebugValue()) 483 continue; 484 if (!I->isBranch()) 485 break; 486 if (!getBranchInfo(*I).Target->isMBB()) 487 break; 488 // Remove the branch. 489 I->eraseFromParent(); 490 I = MBB.end(); 491 ++Count; 492 } 493 494 return Count; 495 } 496 497 bool SystemZInstrInfo:: 498 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 499 assert(Cond.size() == 2 && "Invalid condition"); 500 Cond[1].setImm(Cond[1].getImm() ^ Cond[0].getImm()); 501 return false; 502 } 503 504 unsigned SystemZInstrInfo::insertBranch(MachineBasicBlock &MBB, 505 MachineBasicBlock *TBB, 506 MachineBasicBlock *FBB, 507 ArrayRef<MachineOperand> Cond, 508 const DebugLoc &DL, 509 int *BytesAdded) const { 510 // In this function we output 32-bit branches, which should always 511 // have enough range. They can be shortened and relaxed by later code 512 // in the pipeline, if desired. 513 514 // Shouldn't be a fall through. 515 assert(TBB && "insertBranch must not be told to insert a fallthrough"); 516 assert((Cond.size() == 2 || Cond.size() == 0) && 517 "SystemZ branch conditions have one component!"); 518 assert(!BytesAdded && "code size not handled"); 519 520 if (Cond.empty()) { 521 // Unconditional branch? 522 assert(!FBB && "Unconditional branch with multiple successors!"); 523 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(TBB); 524 return 1; 525 } 526 527 // Conditional branch. 528 unsigned Count = 0; 529 unsigned CCValid = Cond[0].getImm(); 530 unsigned CCMask = Cond[1].getImm(); 531 BuildMI(&MBB, DL, get(SystemZ::BRC)) 532 .addImm(CCValid).addImm(CCMask).addMBB(TBB); 533 ++Count; 534 535 if (FBB) { 536 // Two-way Conditional branch. Insert the second branch. 537 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(FBB); 538 ++Count; 539 } 540 return Count; 541 } 542 543 bool SystemZInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg, 544 unsigned &SrcReg2, int &Mask, 545 int &Value) const { 546 assert(MI.isCompare() && "Caller should have checked for a comparison"); 547 548 if (MI.getNumExplicitOperands() == 2 && MI.getOperand(0).isReg() && 549 MI.getOperand(1).isImm()) { 550 SrcReg = MI.getOperand(0).getReg(); 551 SrcReg2 = 0; 552 Value = MI.getOperand(1).getImm(); 553 Mask = ~0; 554 return true; 555 } 556 557 return false; 558 } 559 560 // If Reg is a virtual register, return its definition, otherwise return null. 561 static MachineInstr *getDef(unsigned Reg, 562 const MachineRegisterInfo *MRI) { 563 if (TargetRegisterInfo::isPhysicalRegister(Reg)) 564 return nullptr; 565 return MRI->getUniqueVRegDef(Reg); 566 } 567 568 // Return true if MI is a shift of type Opcode by Imm bits. 569 static bool isShift(MachineInstr *MI, unsigned Opcode, int64_t Imm) { 570 return (MI->getOpcode() == Opcode && 571 !MI->getOperand(2).getReg() && 572 MI->getOperand(3).getImm() == Imm); 573 } 574 575 // If the destination of MI has no uses, delete it as dead. 576 static void eraseIfDead(MachineInstr *MI, const MachineRegisterInfo *MRI) { 577 if (MRI->use_nodbg_empty(MI->getOperand(0).getReg())) 578 MI->eraseFromParent(); 579 } 580 581 // Compare compares SrcReg against zero. Check whether SrcReg contains 582 // the result of an IPM sequence whose input CC survives until Compare, 583 // and whether Compare is therefore redundant. Delete it and return 584 // true if so. 585 static bool removeIPMBasedCompare(MachineInstr &Compare, unsigned SrcReg, 586 const MachineRegisterInfo *MRI, 587 const TargetRegisterInfo *TRI) { 588 MachineInstr *LGFR = nullptr; 589 MachineInstr *RLL = getDef(SrcReg, MRI); 590 if (RLL && RLL->getOpcode() == SystemZ::LGFR) { 591 LGFR = RLL; 592 RLL = getDef(LGFR->getOperand(1).getReg(), MRI); 593 } 594 if (!RLL || !isShift(RLL, SystemZ::RLL, 31)) 595 return false; 596 597 MachineInstr *SRL = getDef(RLL->getOperand(1).getReg(), MRI); 598 if (!SRL || !isShift(SRL, SystemZ::SRL, SystemZ::IPM_CC)) 599 return false; 600 601 MachineInstr *IPM = getDef(SRL->getOperand(1).getReg(), MRI); 602 if (!IPM || IPM->getOpcode() != SystemZ::IPM) 603 return false; 604 605 // Check that there are no assignments to CC between the IPM and Compare, 606 if (IPM->getParent() != Compare.getParent()) 607 return false; 608 MachineBasicBlock::iterator MBBI = IPM, MBBE = Compare.getIterator(); 609 for (++MBBI; MBBI != MBBE; ++MBBI) { 610 MachineInstr &MI = *MBBI; 611 if (MI.modifiesRegister(SystemZ::CC, TRI)) 612 return false; 613 } 614 615 Compare.eraseFromParent(); 616 if (LGFR) 617 eraseIfDead(LGFR, MRI); 618 eraseIfDead(RLL, MRI); 619 eraseIfDead(SRL, MRI); 620 eraseIfDead(IPM, MRI); 621 622 return true; 623 } 624 625 bool SystemZInstrInfo::optimizeCompareInstr( 626 MachineInstr &Compare, unsigned SrcReg, unsigned SrcReg2, int Mask, 627 int Value, const MachineRegisterInfo *MRI) const { 628 assert(!SrcReg2 && "Only optimizing constant comparisons so far"); 629 bool IsLogical = (Compare.getDesc().TSFlags & SystemZII::IsLogical) != 0; 630 return Value == 0 && !IsLogical && 631 removeIPMBasedCompare(Compare, SrcReg, MRI, &RI); 632 } 633 634 bool SystemZInstrInfo::canInsertSelect(const MachineBasicBlock &MBB, 635 ArrayRef<MachineOperand> Pred, 636 unsigned TrueReg, unsigned FalseReg, 637 int &CondCycles, int &TrueCycles, 638 int &FalseCycles) const { 639 // Not all subtargets have LOCR instructions. 640 if (!STI.hasLoadStoreOnCond()) 641 return false; 642 if (Pred.size() != 2) 643 return false; 644 645 // Check register classes. 646 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 647 const TargetRegisterClass *RC = 648 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 649 if (!RC) 650 return false; 651 652 // We have LOCR instructions for 32 and 64 bit general purpose registers. 653 if ((STI.hasLoadStoreOnCond2() && 654 SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) || 655 SystemZ::GR32BitRegClass.hasSubClassEq(RC) || 656 SystemZ::GR64BitRegClass.hasSubClassEq(RC)) { 657 CondCycles = 2; 658 TrueCycles = 2; 659 FalseCycles = 2; 660 return true; 661 } 662 663 // Can't do anything else. 664 return false; 665 } 666 667 void SystemZInstrInfo::insertSelect(MachineBasicBlock &MBB, 668 MachineBasicBlock::iterator I, 669 const DebugLoc &DL, unsigned DstReg, 670 ArrayRef<MachineOperand> Pred, 671 unsigned TrueReg, 672 unsigned FalseReg) const { 673 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 674 const TargetRegisterClass *RC = MRI.getRegClass(DstReg); 675 676 assert(Pred.size() == 2 && "Invalid condition"); 677 unsigned CCValid = Pred[0].getImm(); 678 unsigned CCMask = Pred[1].getImm(); 679 680 unsigned Opc; 681 if (SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) { 682 if (STI.hasLoadStoreOnCond2()) 683 Opc = SystemZ::LOCRMux; 684 else { 685 Opc = SystemZ::LOCR; 686 MRI.constrainRegClass(DstReg, &SystemZ::GR32BitRegClass); 687 unsigned TReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass); 688 unsigned FReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass); 689 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), TReg).addReg(TrueReg); 690 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), FReg).addReg(FalseReg); 691 TrueReg = TReg; 692 FalseReg = FReg; 693 } 694 } else if (SystemZ::GR64BitRegClass.hasSubClassEq(RC)) 695 Opc = SystemZ::LOCGR; 696 else 697 llvm_unreachable("Invalid register class"); 698 699 BuildMI(MBB, I, DL, get(Opc), DstReg) 700 .addReg(FalseReg).addReg(TrueReg) 701 .addImm(CCValid).addImm(CCMask); 702 } 703 704 bool SystemZInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, 705 unsigned Reg, 706 MachineRegisterInfo *MRI) const { 707 unsigned DefOpc = DefMI.getOpcode(); 708 if (DefOpc != SystemZ::LHIMux && DefOpc != SystemZ::LHI && 709 DefOpc != SystemZ::LGHI) 710 return false; 711 if (DefMI.getOperand(0).getReg() != Reg) 712 return false; 713 int32_t ImmVal = (int32_t)DefMI.getOperand(1).getImm(); 714 715 unsigned UseOpc = UseMI.getOpcode(); 716 unsigned NewUseOpc; 717 unsigned UseIdx; 718 int CommuteIdx = -1; 719 switch (UseOpc) { 720 case SystemZ::LOCRMux: 721 if (!STI.hasLoadStoreOnCond2()) 722 return false; 723 NewUseOpc = SystemZ::LOCHIMux; 724 if (UseMI.getOperand(2).getReg() == Reg) 725 UseIdx = 2; 726 else if (UseMI.getOperand(1).getReg() == Reg) 727 UseIdx = 2, CommuteIdx = 1; 728 else 729 return false; 730 break; 731 case SystemZ::LOCGR: 732 if (!STI.hasLoadStoreOnCond2()) 733 return false; 734 NewUseOpc = SystemZ::LOCGHI; 735 if (UseMI.getOperand(2).getReg() == Reg) 736 UseIdx = 2; 737 else if (UseMI.getOperand(1).getReg() == Reg) 738 UseIdx = 2, CommuteIdx = 1; 739 else 740 return false; 741 break; 742 default: 743 return false; 744 } 745 746 if (CommuteIdx != -1) 747 if (!commuteInstruction(UseMI, false, CommuteIdx, UseIdx)) 748 return false; 749 750 bool DeleteDef = MRI->hasOneNonDBGUse(Reg); 751 UseMI.setDesc(get(NewUseOpc)); 752 UseMI.getOperand(UseIdx).ChangeToImmediate(ImmVal); 753 if (DeleteDef) 754 DefMI.eraseFromParent(); 755 756 return true; 757 } 758 759 bool SystemZInstrInfo::isPredicable(const MachineInstr &MI) const { 760 unsigned Opcode = MI.getOpcode(); 761 if (Opcode == SystemZ::Return || 762 Opcode == SystemZ::Trap || 763 Opcode == SystemZ::CallJG || 764 Opcode == SystemZ::CallBR) 765 return true; 766 return false; 767 } 768 769 bool SystemZInstrInfo:: 770 isProfitableToIfCvt(MachineBasicBlock &MBB, 771 unsigned NumCycles, unsigned ExtraPredCycles, 772 BranchProbability Probability) const { 773 // Avoid using conditional returns at the end of a loop (since then 774 // we'd need to emit an unconditional branch to the beginning anyway, 775 // making the loop body longer). This doesn't apply for low-probability 776 // loops (eg. compare-and-swap retry), so just decide based on branch 777 // probability instead of looping structure. 778 // However, since Compare and Trap instructions cost the same as a regular 779 // Compare instruction, we should allow the if conversion to convert this 780 // into a Conditional Compare regardless of the branch probability. 781 if (MBB.getLastNonDebugInstr()->getOpcode() != SystemZ::Trap && 782 MBB.succ_empty() && Probability < BranchProbability(1, 8)) 783 return false; 784 // For now only convert single instructions. 785 return NumCycles == 1; 786 } 787 788 bool SystemZInstrInfo:: 789 isProfitableToIfCvt(MachineBasicBlock &TMBB, 790 unsigned NumCyclesT, unsigned ExtraPredCyclesT, 791 MachineBasicBlock &FMBB, 792 unsigned NumCyclesF, unsigned ExtraPredCyclesF, 793 BranchProbability Probability) const { 794 // For now avoid converting mutually-exclusive cases. 795 return false; 796 } 797 798 bool SystemZInstrInfo:: 799 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles, 800 BranchProbability Probability) const { 801 // For now only duplicate single instructions. 802 return NumCycles == 1; 803 } 804 805 bool SystemZInstrInfo::PredicateInstruction( 806 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const { 807 assert(Pred.size() == 2 && "Invalid condition"); 808 unsigned CCValid = Pred[0].getImm(); 809 unsigned CCMask = Pred[1].getImm(); 810 assert(CCMask > 0 && CCMask < 15 && "Invalid predicate"); 811 unsigned Opcode = MI.getOpcode(); 812 if (Opcode == SystemZ::Trap) { 813 MI.setDesc(get(SystemZ::CondTrap)); 814 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 815 .addImm(CCValid).addImm(CCMask) 816 .addReg(SystemZ::CC, RegState::Implicit); 817 return true; 818 } 819 if (Opcode == SystemZ::Return) { 820 MI.setDesc(get(SystemZ::CondReturn)); 821 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 822 .addImm(CCValid).addImm(CCMask) 823 .addReg(SystemZ::CC, RegState::Implicit); 824 return true; 825 } 826 if (Opcode == SystemZ::CallJG) { 827 MachineOperand FirstOp = MI.getOperand(0); 828 const uint32_t *RegMask = MI.getOperand(1).getRegMask(); 829 MI.RemoveOperand(1); 830 MI.RemoveOperand(0); 831 MI.setDesc(get(SystemZ::CallBRCL)); 832 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 833 .addImm(CCValid) 834 .addImm(CCMask) 835 .add(FirstOp) 836 .addRegMask(RegMask) 837 .addReg(SystemZ::CC, RegState::Implicit); 838 return true; 839 } 840 if (Opcode == SystemZ::CallBR) { 841 const uint32_t *RegMask = MI.getOperand(0).getRegMask(); 842 MI.RemoveOperand(0); 843 MI.setDesc(get(SystemZ::CallBCR)); 844 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 845 .addImm(CCValid).addImm(CCMask) 846 .addRegMask(RegMask) 847 .addReg(SystemZ::CC, RegState::Implicit); 848 return true; 849 } 850 return false; 851 } 852 853 void SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 854 MachineBasicBlock::iterator MBBI, 855 const DebugLoc &DL, unsigned DestReg, 856 unsigned SrcReg, bool KillSrc) const { 857 // Split 128-bit GPR moves into two 64-bit moves. Add implicit uses of the 858 // super register in case one of the subregs is undefined. 859 // This handles ADDR128 too. 860 if (SystemZ::GR128BitRegClass.contains(DestReg, SrcReg)) { 861 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_h64), 862 RI.getSubReg(SrcReg, SystemZ::subreg_h64), KillSrc); 863 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI)) 864 .addReg(SrcReg, RegState::Implicit); 865 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_l64), 866 RI.getSubReg(SrcReg, SystemZ::subreg_l64), KillSrc); 867 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI)) 868 .addReg(SrcReg, (getKillRegState(KillSrc) | RegState::Implicit)); 869 return; 870 } 871 872 if (SystemZ::GRX32BitRegClass.contains(DestReg, SrcReg)) { 873 emitGRX32Move(MBB, MBBI, DL, DestReg, SrcReg, SystemZ::LR, 32, KillSrc, 874 false); 875 return; 876 } 877 878 // Move 128-bit floating-point values between VR128 and FP128. 879 if (SystemZ::VR128BitRegClass.contains(DestReg) && 880 SystemZ::FP128BitRegClass.contains(SrcReg)) { 881 unsigned SrcRegHi = 882 RI.getMatchingSuperReg(RI.getSubReg(SrcReg, SystemZ::subreg_h64), 883 SystemZ::subreg_r64, &SystemZ::VR128BitRegClass); 884 unsigned SrcRegLo = 885 RI.getMatchingSuperReg(RI.getSubReg(SrcReg, SystemZ::subreg_l64), 886 SystemZ::subreg_r64, &SystemZ::VR128BitRegClass); 887 888 BuildMI(MBB, MBBI, DL, get(SystemZ::VMRHG), DestReg) 889 .addReg(SrcRegHi, getKillRegState(KillSrc)) 890 .addReg(SrcRegLo, getKillRegState(KillSrc)); 891 return; 892 } 893 if (SystemZ::FP128BitRegClass.contains(DestReg) && 894 SystemZ::VR128BitRegClass.contains(SrcReg)) { 895 unsigned DestRegHi = 896 RI.getMatchingSuperReg(RI.getSubReg(DestReg, SystemZ::subreg_h64), 897 SystemZ::subreg_r64, &SystemZ::VR128BitRegClass); 898 unsigned DestRegLo = 899 RI.getMatchingSuperReg(RI.getSubReg(DestReg, SystemZ::subreg_l64), 900 SystemZ::subreg_r64, &SystemZ::VR128BitRegClass); 901 902 if (DestRegHi != SrcReg) 903 copyPhysReg(MBB, MBBI, DL, DestRegHi, SrcReg, false); 904 BuildMI(MBB, MBBI, DL, get(SystemZ::VREPG), DestRegLo) 905 .addReg(SrcReg, getKillRegState(KillSrc)).addImm(1); 906 return; 907 } 908 909 // Everything else needs only one instruction. 910 unsigned Opcode; 911 if (SystemZ::GR64BitRegClass.contains(DestReg, SrcReg)) 912 Opcode = SystemZ::LGR; 913 else if (SystemZ::FP32BitRegClass.contains(DestReg, SrcReg)) 914 // For z13 we prefer LDR over LER to avoid partial register dependencies. 915 Opcode = STI.hasVector() ? SystemZ::LDR32 : SystemZ::LER; 916 else if (SystemZ::FP64BitRegClass.contains(DestReg, SrcReg)) 917 Opcode = SystemZ::LDR; 918 else if (SystemZ::FP128BitRegClass.contains(DestReg, SrcReg)) 919 Opcode = SystemZ::LXR; 920 else if (SystemZ::VR32BitRegClass.contains(DestReg, SrcReg)) 921 Opcode = SystemZ::VLR32; 922 else if (SystemZ::VR64BitRegClass.contains(DestReg, SrcReg)) 923 Opcode = SystemZ::VLR64; 924 else if (SystemZ::VR128BitRegClass.contains(DestReg, SrcReg)) 925 Opcode = SystemZ::VLR; 926 else if (SystemZ::AR32BitRegClass.contains(DestReg, SrcReg)) 927 Opcode = SystemZ::CPYA; 928 else if (SystemZ::AR32BitRegClass.contains(DestReg) && 929 SystemZ::GR32BitRegClass.contains(SrcReg)) 930 Opcode = SystemZ::SAR; 931 else if (SystemZ::GR32BitRegClass.contains(DestReg) && 932 SystemZ::AR32BitRegClass.contains(SrcReg)) 933 Opcode = SystemZ::EAR; 934 else 935 llvm_unreachable("Impossible reg-to-reg copy"); 936 937 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 938 .addReg(SrcReg, getKillRegState(KillSrc)); 939 } 940 941 void SystemZInstrInfo::storeRegToStackSlot( 942 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned SrcReg, 943 bool isKill, int FrameIdx, const TargetRegisterClass *RC, 944 const TargetRegisterInfo *TRI) const { 945 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 946 947 // Callers may expect a single instruction, so keep 128-bit moves 948 // together for now and lower them after register allocation. 949 unsigned LoadOpcode, StoreOpcode; 950 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 951 addFrameReference(BuildMI(MBB, MBBI, DL, get(StoreOpcode)) 952 .addReg(SrcReg, getKillRegState(isKill)), 953 FrameIdx); 954 } 955 956 void SystemZInstrInfo::loadRegFromStackSlot( 957 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned DestReg, 958 int FrameIdx, const TargetRegisterClass *RC, 959 const TargetRegisterInfo *TRI) const { 960 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 961 962 // Callers may expect a single instruction, so keep 128-bit moves 963 // together for now and lower them after register allocation. 964 unsigned LoadOpcode, StoreOpcode; 965 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 966 addFrameReference(BuildMI(MBB, MBBI, DL, get(LoadOpcode), DestReg), 967 FrameIdx); 968 } 969 970 // Return true if MI is a simple load or store with a 12-bit displacement 971 // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 972 static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) { 973 const MCInstrDesc &MCID = MI->getDesc(); 974 return ((MCID.TSFlags & Flag) && 975 isUInt<12>(MI->getOperand(2).getImm()) && 976 MI->getOperand(3).getReg() == 0); 977 } 978 979 namespace { 980 981 struct LogicOp { 982 LogicOp() = default; 983 LogicOp(unsigned regSize, unsigned immLSB, unsigned immSize) 984 : RegSize(regSize), ImmLSB(immLSB), ImmSize(immSize) {} 985 986 explicit operator bool() const { return RegSize; } 987 988 unsigned RegSize = 0; 989 unsigned ImmLSB = 0; 990 unsigned ImmSize = 0; 991 }; 992 993 } // end anonymous namespace 994 995 static LogicOp interpretAndImmediate(unsigned Opcode) { 996 switch (Opcode) { 997 case SystemZ::NILMux: return LogicOp(32, 0, 16); 998 case SystemZ::NIHMux: return LogicOp(32, 16, 16); 999 case SystemZ::NILL64: return LogicOp(64, 0, 16); 1000 case SystemZ::NILH64: return LogicOp(64, 16, 16); 1001 case SystemZ::NIHL64: return LogicOp(64, 32, 16); 1002 case SystemZ::NIHH64: return LogicOp(64, 48, 16); 1003 case SystemZ::NIFMux: return LogicOp(32, 0, 32); 1004 case SystemZ::NILF64: return LogicOp(64, 0, 32); 1005 case SystemZ::NIHF64: return LogicOp(64, 32, 32); 1006 default: return LogicOp(); 1007 } 1008 } 1009 1010 static void transferDeadCC(MachineInstr *OldMI, MachineInstr *NewMI) { 1011 if (OldMI->registerDefIsDead(SystemZ::CC)) { 1012 MachineOperand *CCDef = NewMI->findRegisterDefOperand(SystemZ::CC); 1013 if (CCDef != nullptr) 1014 CCDef->setIsDead(true); 1015 } 1016 } 1017 1018 // Used to return from convertToThreeAddress after replacing two-address 1019 // instruction OldMI with three-address instruction NewMI. 1020 static MachineInstr *finishConvertToThreeAddress(MachineInstr *OldMI, 1021 MachineInstr *NewMI, 1022 LiveVariables *LV) { 1023 if (LV) { 1024 unsigned NumOps = OldMI->getNumOperands(); 1025 for (unsigned I = 1; I < NumOps; ++I) { 1026 MachineOperand &Op = OldMI->getOperand(I); 1027 if (Op.isReg() && Op.isKill()) 1028 LV->replaceKillInstruction(Op.getReg(), *OldMI, *NewMI); 1029 } 1030 } 1031 transferDeadCC(OldMI, NewMI); 1032 return NewMI; 1033 } 1034 1035 MachineInstr *SystemZInstrInfo::convertToThreeAddress( 1036 MachineFunction::iterator &MFI, MachineInstr &MI, LiveVariables *LV) const { 1037 MachineBasicBlock *MBB = MI.getParent(); 1038 MachineFunction *MF = MBB->getParent(); 1039 MachineRegisterInfo &MRI = MF->getRegInfo(); 1040 1041 unsigned Opcode = MI.getOpcode(); 1042 unsigned NumOps = MI.getNumOperands(); 1043 1044 // Try to convert something like SLL into SLLK, if supported. 1045 // We prefer to keep the two-operand form where possible both 1046 // because it tends to be shorter and because some instructions 1047 // have memory forms that can be used during spilling. 1048 if (STI.hasDistinctOps()) { 1049 MachineOperand &Dest = MI.getOperand(0); 1050 MachineOperand &Src = MI.getOperand(1); 1051 unsigned DestReg = Dest.getReg(); 1052 unsigned SrcReg = Src.getReg(); 1053 // AHIMux is only really a three-operand instruction when both operands 1054 // are low registers. Try to constrain both operands to be low if 1055 // possible. 1056 if (Opcode == SystemZ::AHIMux && 1057 TargetRegisterInfo::isVirtualRegister(DestReg) && 1058 TargetRegisterInfo::isVirtualRegister(SrcReg) && 1059 MRI.getRegClass(DestReg)->contains(SystemZ::R1L) && 1060 MRI.getRegClass(SrcReg)->contains(SystemZ::R1L)) { 1061 MRI.constrainRegClass(DestReg, &SystemZ::GR32BitRegClass); 1062 MRI.constrainRegClass(SrcReg, &SystemZ::GR32BitRegClass); 1063 } 1064 int ThreeOperandOpcode = SystemZ::getThreeOperandOpcode(Opcode); 1065 if (ThreeOperandOpcode >= 0) { 1066 // Create three address instruction without adding the implicit 1067 // operands. Those will instead be copied over from the original 1068 // instruction by the loop below. 1069 MachineInstrBuilder MIB( 1070 *MF, MF->CreateMachineInstr(get(ThreeOperandOpcode), MI.getDebugLoc(), 1071 /*NoImplicit=*/true)); 1072 MIB.add(Dest); 1073 // Keep the kill state, but drop the tied flag. 1074 MIB.addReg(Src.getReg(), getKillRegState(Src.isKill()), Src.getSubReg()); 1075 // Keep the remaining operands as-is. 1076 for (unsigned I = 2; I < NumOps; ++I) 1077 MIB.add(MI.getOperand(I)); 1078 MBB->insert(MI, MIB); 1079 return finishConvertToThreeAddress(&MI, MIB, LV); 1080 } 1081 } 1082 1083 // Try to convert an AND into an RISBG-type instruction. 1084 if (LogicOp And = interpretAndImmediate(Opcode)) { 1085 uint64_t Imm = MI.getOperand(2).getImm() << And.ImmLSB; 1086 // AND IMMEDIATE leaves the other bits of the register unchanged. 1087 Imm |= allOnes(And.RegSize) & ~(allOnes(And.ImmSize) << And.ImmLSB); 1088 unsigned Start, End; 1089 if (isRxSBGMask(Imm, And.RegSize, Start, End)) { 1090 unsigned NewOpcode; 1091 if (And.RegSize == 64) { 1092 NewOpcode = SystemZ::RISBG; 1093 // Prefer RISBGN if available, since it does not clobber CC. 1094 if (STI.hasMiscellaneousExtensions()) 1095 NewOpcode = SystemZ::RISBGN; 1096 } else { 1097 NewOpcode = SystemZ::RISBMux; 1098 Start &= 31; 1099 End &= 31; 1100 } 1101 MachineOperand &Dest = MI.getOperand(0); 1102 MachineOperand &Src = MI.getOperand(1); 1103 MachineInstrBuilder MIB = 1104 BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpcode)) 1105 .add(Dest) 1106 .addReg(0) 1107 .addReg(Src.getReg(), getKillRegState(Src.isKill()), 1108 Src.getSubReg()) 1109 .addImm(Start) 1110 .addImm(End + 128) 1111 .addImm(0); 1112 return finishConvertToThreeAddress(&MI, MIB, LV); 1113 } 1114 } 1115 return nullptr; 1116 } 1117 1118 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1119 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1120 MachineBasicBlock::iterator InsertPt, int FrameIndex, 1121 LiveIntervals *LIS) const { 1122 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 1123 const MachineFrameInfo &MFI = MF.getFrameInfo(); 1124 unsigned Size = MFI.getObjectSize(FrameIndex); 1125 unsigned Opcode = MI.getOpcode(); 1126 1127 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 1128 if (LIS != nullptr && (Opcode == SystemZ::LA || Opcode == SystemZ::LAY) && 1129 isInt<8>(MI.getOperand(2).getImm()) && !MI.getOperand(3).getReg()) { 1130 1131 // Check CC liveness, since new instruction introduces a dead 1132 // def of CC. 1133 MCRegUnitIterator CCUnit(SystemZ::CC, TRI); 1134 LiveRange &CCLiveRange = LIS->getRegUnit(*CCUnit); 1135 ++CCUnit; 1136 assert(!CCUnit.isValid() && "CC only has one reg unit."); 1137 SlotIndex MISlot = 1138 LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot(); 1139 if (!CCLiveRange.liveAt(MISlot)) { 1140 // LA(Y) %reg, CONST(%reg) -> AGSI %mem, CONST 1141 MachineInstr *BuiltMI = BuildMI(*InsertPt->getParent(), InsertPt, 1142 MI.getDebugLoc(), get(SystemZ::AGSI)) 1143 .addFrameIndex(FrameIndex) 1144 .addImm(0) 1145 .addImm(MI.getOperand(2).getImm()); 1146 BuiltMI->findRegisterDefOperand(SystemZ::CC)->setIsDead(true); 1147 CCLiveRange.createDeadDef(MISlot, LIS->getVNInfoAllocator()); 1148 return BuiltMI; 1149 } 1150 } 1151 return nullptr; 1152 } 1153 1154 // All other cases require a single operand. 1155 if (Ops.size() != 1) 1156 return nullptr; 1157 1158 unsigned OpNum = Ops[0]; 1159 assert(Size * 8 == 1160 TRI->getRegSizeInBits(*MF.getRegInfo() 1161 .getRegClass(MI.getOperand(OpNum).getReg())) && 1162 "Invalid size combination"); 1163 1164 if ((Opcode == SystemZ::AHI || Opcode == SystemZ::AGHI) && OpNum == 0 && 1165 isInt<8>(MI.getOperand(2).getImm())) { 1166 // A(G)HI %reg, CONST -> A(G)SI %mem, CONST 1167 Opcode = (Opcode == SystemZ::AHI ? SystemZ::ASI : SystemZ::AGSI); 1168 MachineInstr *BuiltMI = 1169 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode)) 1170 .addFrameIndex(FrameIndex) 1171 .addImm(0) 1172 .addImm(MI.getOperand(2).getImm()); 1173 transferDeadCC(&MI, BuiltMI); 1174 return BuiltMI; 1175 } 1176 1177 if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) { 1178 bool Op0IsGPR = (Opcode == SystemZ::LGDR); 1179 bool Op1IsGPR = (Opcode == SystemZ::LDGR); 1180 // If we're spilling the destination of an LDGR or LGDR, store the 1181 // source register instead. 1182 if (OpNum == 0) { 1183 unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD; 1184 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1185 get(StoreOpcode)) 1186 .add(MI.getOperand(1)) 1187 .addFrameIndex(FrameIndex) 1188 .addImm(0) 1189 .addReg(0); 1190 } 1191 // If we're spilling the source of an LDGR or LGDR, load the 1192 // destination register instead. 1193 if (OpNum == 1) { 1194 unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD; 1195 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1196 get(LoadOpcode)) 1197 .add(MI.getOperand(0)) 1198 .addFrameIndex(FrameIndex) 1199 .addImm(0) 1200 .addReg(0); 1201 } 1202 } 1203 1204 // Look for cases where the source of a simple store or the destination 1205 // of a simple load is being spilled. Try to use MVC instead. 1206 // 1207 // Although MVC is in practice a fast choice in these cases, it is still 1208 // logically a bytewise copy. This means that we cannot use it if the 1209 // load or store is volatile. We also wouldn't be able to use MVC if 1210 // the two memories partially overlap, but that case cannot occur here, 1211 // because we know that one of the memories is a full frame index. 1212 // 1213 // For performance reasons, we also want to avoid using MVC if the addresses 1214 // might be equal. We don't worry about that case here, because spill slot 1215 // coloring happens later, and because we have special code to remove 1216 // MVCs that turn out to be redundant. 1217 if (OpNum == 0 && MI.hasOneMemOperand()) { 1218 MachineMemOperand *MMO = *MI.memoperands_begin(); 1219 if (MMO->getSize() == Size && !MMO->isVolatile()) { 1220 // Handle conversion of loads. 1221 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXLoad)) { 1222 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1223 get(SystemZ::MVC)) 1224 .addFrameIndex(FrameIndex) 1225 .addImm(0) 1226 .addImm(Size) 1227 .add(MI.getOperand(1)) 1228 .addImm(MI.getOperand(2).getImm()) 1229 .addMemOperand(MMO); 1230 } 1231 // Handle conversion of stores. 1232 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXStore)) { 1233 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1234 get(SystemZ::MVC)) 1235 .add(MI.getOperand(1)) 1236 .addImm(MI.getOperand(2).getImm()) 1237 .addImm(Size) 1238 .addFrameIndex(FrameIndex) 1239 .addImm(0) 1240 .addMemOperand(MMO); 1241 } 1242 } 1243 } 1244 1245 // If the spilled operand is the final one, try to change <INSN>R 1246 // into <INSN>. 1247 int MemOpcode = SystemZ::getMemOpcode(Opcode); 1248 if (MemOpcode >= 0) { 1249 unsigned NumOps = MI.getNumExplicitOperands(); 1250 if (OpNum == NumOps - 1) { 1251 const MCInstrDesc &MemDesc = get(MemOpcode); 1252 uint64_t AccessBytes = SystemZII::getAccessSize(MemDesc.TSFlags); 1253 assert(AccessBytes != 0 && "Size of access should be known"); 1254 assert(AccessBytes <= Size && "Access outside the frame index"); 1255 uint64_t Offset = Size - AccessBytes; 1256 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt, 1257 MI.getDebugLoc(), get(MemOpcode)); 1258 for (unsigned I = 0; I < OpNum; ++I) 1259 MIB.add(MI.getOperand(I)); 1260 MIB.addFrameIndex(FrameIndex).addImm(Offset); 1261 if (MemDesc.TSFlags & SystemZII::HasIndex) 1262 MIB.addReg(0); 1263 transferDeadCC(&MI, MIB); 1264 return MIB; 1265 } 1266 } 1267 1268 return nullptr; 1269 } 1270 1271 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1272 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1273 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, 1274 LiveIntervals *LIS) const { 1275 return nullptr; 1276 } 1277 1278 bool SystemZInstrInfo::expandPostRAPseudo(MachineInstr &MI) const { 1279 switch (MI.getOpcode()) { 1280 case SystemZ::L128: 1281 splitMove(MI, SystemZ::LG); 1282 return true; 1283 1284 case SystemZ::ST128: 1285 splitMove(MI, SystemZ::STG); 1286 return true; 1287 1288 case SystemZ::LX: 1289 splitMove(MI, SystemZ::LD); 1290 return true; 1291 1292 case SystemZ::STX: 1293 splitMove(MI, SystemZ::STD); 1294 return true; 1295 1296 case SystemZ::LBMux: 1297 expandRXYPseudo(MI, SystemZ::LB, SystemZ::LBH); 1298 return true; 1299 1300 case SystemZ::LHMux: 1301 expandRXYPseudo(MI, SystemZ::LH, SystemZ::LHH); 1302 return true; 1303 1304 case SystemZ::LLCRMux: 1305 expandZExtPseudo(MI, SystemZ::LLCR, 8); 1306 return true; 1307 1308 case SystemZ::LLHRMux: 1309 expandZExtPseudo(MI, SystemZ::LLHR, 16); 1310 return true; 1311 1312 case SystemZ::LLCMux: 1313 expandRXYPseudo(MI, SystemZ::LLC, SystemZ::LLCH); 1314 return true; 1315 1316 case SystemZ::LLHMux: 1317 expandRXYPseudo(MI, SystemZ::LLH, SystemZ::LLHH); 1318 return true; 1319 1320 case SystemZ::LMux: 1321 expandRXYPseudo(MI, SystemZ::L, SystemZ::LFH); 1322 return true; 1323 1324 case SystemZ::LOCMux: 1325 expandLOCPseudo(MI, SystemZ::LOC, SystemZ::LOCFH); 1326 return true; 1327 1328 case SystemZ::LOCHIMux: 1329 expandLOCPseudo(MI, SystemZ::LOCHI, SystemZ::LOCHHI); 1330 return true; 1331 1332 case SystemZ::LOCRMux: 1333 expandLOCRPseudo(MI, SystemZ::LOCR, SystemZ::LOCFHR); 1334 return true; 1335 1336 case SystemZ::STCMux: 1337 expandRXYPseudo(MI, SystemZ::STC, SystemZ::STCH); 1338 return true; 1339 1340 case SystemZ::STHMux: 1341 expandRXYPseudo(MI, SystemZ::STH, SystemZ::STHH); 1342 return true; 1343 1344 case SystemZ::STMux: 1345 expandRXYPseudo(MI, SystemZ::ST, SystemZ::STFH); 1346 return true; 1347 1348 case SystemZ::STOCMux: 1349 expandLOCPseudo(MI, SystemZ::STOC, SystemZ::STOCFH); 1350 return true; 1351 1352 case SystemZ::LHIMux: 1353 expandRIPseudo(MI, SystemZ::LHI, SystemZ::IIHF, true); 1354 return true; 1355 1356 case SystemZ::IIFMux: 1357 expandRIPseudo(MI, SystemZ::IILF, SystemZ::IIHF, false); 1358 return true; 1359 1360 case SystemZ::IILMux: 1361 expandRIPseudo(MI, SystemZ::IILL, SystemZ::IIHL, false); 1362 return true; 1363 1364 case SystemZ::IIHMux: 1365 expandRIPseudo(MI, SystemZ::IILH, SystemZ::IIHH, false); 1366 return true; 1367 1368 case SystemZ::NIFMux: 1369 expandRIPseudo(MI, SystemZ::NILF, SystemZ::NIHF, false); 1370 return true; 1371 1372 case SystemZ::NILMux: 1373 expandRIPseudo(MI, SystemZ::NILL, SystemZ::NIHL, false); 1374 return true; 1375 1376 case SystemZ::NIHMux: 1377 expandRIPseudo(MI, SystemZ::NILH, SystemZ::NIHH, false); 1378 return true; 1379 1380 case SystemZ::OIFMux: 1381 expandRIPseudo(MI, SystemZ::OILF, SystemZ::OIHF, false); 1382 return true; 1383 1384 case SystemZ::OILMux: 1385 expandRIPseudo(MI, SystemZ::OILL, SystemZ::OIHL, false); 1386 return true; 1387 1388 case SystemZ::OIHMux: 1389 expandRIPseudo(MI, SystemZ::OILH, SystemZ::OIHH, false); 1390 return true; 1391 1392 case SystemZ::XIFMux: 1393 expandRIPseudo(MI, SystemZ::XILF, SystemZ::XIHF, false); 1394 return true; 1395 1396 case SystemZ::TMLMux: 1397 expandRIPseudo(MI, SystemZ::TMLL, SystemZ::TMHL, false); 1398 return true; 1399 1400 case SystemZ::TMHMux: 1401 expandRIPseudo(MI, SystemZ::TMLH, SystemZ::TMHH, false); 1402 return true; 1403 1404 case SystemZ::AHIMux: 1405 expandRIPseudo(MI, SystemZ::AHI, SystemZ::AIH, false); 1406 return true; 1407 1408 case SystemZ::AHIMuxK: 1409 expandRIEPseudo(MI, SystemZ::AHI, SystemZ::AHIK, SystemZ::AIH); 1410 return true; 1411 1412 case SystemZ::AFIMux: 1413 expandRIPseudo(MI, SystemZ::AFI, SystemZ::AIH, false); 1414 return true; 1415 1416 case SystemZ::CHIMux: 1417 expandRIPseudo(MI, SystemZ::CHI, SystemZ::CIH, false); 1418 return true; 1419 1420 case SystemZ::CFIMux: 1421 expandRIPseudo(MI, SystemZ::CFI, SystemZ::CIH, false); 1422 return true; 1423 1424 case SystemZ::CLFIMux: 1425 expandRIPseudo(MI, SystemZ::CLFI, SystemZ::CLIH, false); 1426 return true; 1427 1428 case SystemZ::CMux: 1429 expandRXYPseudo(MI, SystemZ::C, SystemZ::CHF); 1430 return true; 1431 1432 case SystemZ::CLMux: 1433 expandRXYPseudo(MI, SystemZ::CL, SystemZ::CLHF); 1434 return true; 1435 1436 case SystemZ::RISBMux: { 1437 bool DestIsHigh = isHighReg(MI.getOperand(0).getReg()); 1438 bool SrcIsHigh = isHighReg(MI.getOperand(2).getReg()); 1439 if (SrcIsHigh == DestIsHigh) 1440 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHH : SystemZ::RISBLL)); 1441 else { 1442 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHL : SystemZ::RISBLH)); 1443 MI.getOperand(5).setImm(MI.getOperand(5).getImm() ^ 32); 1444 } 1445 return true; 1446 } 1447 1448 case SystemZ::ADJDYNALLOC: 1449 splitAdjDynAlloc(MI); 1450 return true; 1451 1452 case TargetOpcode::LOAD_STACK_GUARD: 1453 expandLoadStackGuard(&MI); 1454 return true; 1455 1456 default: 1457 return false; 1458 } 1459 } 1460 1461 unsigned SystemZInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { 1462 if (MI.getOpcode() == TargetOpcode::INLINEASM) { 1463 const MachineFunction *MF = MI.getParent()->getParent(); 1464 const char *AsmStr = MI.getOperand(0).getSymbolName(); 1465 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo()); 1466 } 1467 return MI.getDesc().getSize(); 1468 } 1469 1470 SystemZII::Branch 1471 SystemZInstrInfo::getBranchInfo(const MachineInstr &MI) const { 1472 switch (MI.getOpcode()) { 1473 case SystemZ::BR: 1474 case SystemZ::BI: 1475 case SystemZ::J: 1476 case SystemZ::JG: 1477 return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY, 1478 SystemZ::CCMASK_ANY, &MI.getOperand(0)); 1479 1480 case SystemZ::BRC: 1481 case SystemZ::BRCL: 1482 return SystemZII::Branch(SystemZII::BranchNormal, MI.getOperand(0).getImm(), 1483 MI.getOperand(1).getImm(), &MI.getOperand(2)); 1484 1485 case SystemZ::BRCT: 1486 case SystemZ::BRCTH: 1487 return SystemZII::Branch(SystemZII::BranchCT, SystemZ::CCMASK_ICMP, 1488 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1489 1490 case SystemZ::BRCTG: 1491 return SystemZII::Branch(SystemZII::BranchCTG, SystemZ::CCMASK_ICMP, 1492 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1493 1494 case SystemZ::CIJ: 1495 case SystemZ::CRJ: 1496 return SystemZII::Branch(SystemZII::BranchC, SystemZ::CCMASK_ICMP, 1497 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1498 1499 case SystemZ::CLIJ: 1500 case SystemZ::CLRJ: 1501 return SystemZII::Branch(SystemZII::BranchCL, SystemZ::CCMASK_ICMP, 1502 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1503 1504 case SystemZ::CGIJ: 1505 case SystemZ::CGRJ: 1506 return SystemZII::Branch(SystemZII::BranchCG, SystemZ::CCMASK_ICMP, 1507 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1508 1509 case SystemZ::CLGIJ: 1510 case SystemZ::CLGRJ: 1511 return SystemZII::Branch(SystemZII::BranchCLG, SystemZ::CCMASK_ICMP, 1512 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1513 1514 default: 1515 llvm_unreachable("Unrecognized branch opcode"); 1516 } 1517 } 1518 1519 void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC, 1520 unsigned &LoadOpcode, 1521 unsigned &StoreOpcode) const { 1522 if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) { 1523 LoadOpcode = SystemZ::L; 1524 StoreOpcode = SystemZ::ST; 1525 } else if (RC == &SystemZ::GRH32BitRegClass) { 1526 LoadOpcode = SystemZ::LFH; 1527 StoreOpcode = SystemZ::STFH; 1528 } else if (RC == &SystemZ::GRX32BitRegClass) { 1529 LoadOpcode = SystemZ::LMux; 1530 StoreOpcode = SystemZ::STMux; 1531 } else if (RC == &SystemZ::GR64BitRegClass || 1532 RC == &SystemZ::ADDR64BitRegClass) { 1533 LoadOpcode = SystemZ::LG; 1534 StoreOpcode = SystemZ::STG; 1535 } else if (RC == &SystemZ::GR128BitRegClass || 1536 RC == &SystemZ::ADDR128BitRegClass) { 1537 LoadOpcode = SystemZ::L128; 1538 StoreOpcode = SystemZ::ST128; 1539 } else if (RC == &SystemZ::FP32BitRegClass) { 1540 LoadOpcode = SystemZ::LE; 1541 StoreOpcode = SystemZ::STE; 1542 } else if (RC == &SystemZ::FP64BitRegClass) { 1543 LoadOpcode = SystemZ::LD; 1544 StoreOpcode = SystemZ::STD; 1545 } else if (RC == &SystemZ::FP128BitRegClass) { 1546 LoadOpcode = SystemZ::LX; 1547 StoreOpcode = SystemZ::STX; 1548 } else if (RC == &SystemZ::VR32BitRegClass) { 1549 LoadOpcode = SystemZ::VL32; 1550 StoreOpcode = SystemZ::VST32; 1551 } else if (RC == &SystemZ::VR64BitRegClass) { 1552 LoadOpcode = SystemZ::VL64; 1553 StoreOpcode = SystemZ::VST64; 1554 } else if (RC == &SystemZ::VF128BitRegClass || 1555 RC == &SystemZ::VR128BitRegClass) { 1556 LoadOpcode = SystemZ::VL; 1557 StoreOpcode = SystemZ::VST; 1558 } else 1559 llvm_unreachable("Unsupported regclass to load or store"); 1560 } 1561 1562 unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode, 1563 int64_t Offset) const { 1564 const MCInstrDesc &MCID = get(Opcode); 1565 int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset); 1566 if (isUInt<12>(Offset) && isUInt<12>(Offset2)) { 1567 // Get the instruction to use for unsigned 12-bit displacements. 1568 int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode); 1569 if (Disp12Opcode >= 0) 1570 return Disp12Opcode; 1571 1572 // All address-related instructions can use unsigned 12-bit 1573 // displacements. 1574 return Opcode; 1575 } 1576 if (isInt<20>(Offset) && isInt<20>(Offset2)) { 1577 // Get the instruction to use for signed 20-bit displacements. 1578 int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode); 1579 if (Disp20Opcode >= 0) 1580 return Disp20Opcode; 1581 1582 // Check whether Opcode allows signed 20-bit displacements. 1583 if (MCID.TSFlags & SystemZII::Has20BitOffset) 1584 return Opcode; 1585 } 1586 return 0; 1587 } 1588 1589 unsigned SystemZInstrInfo::getLoadAndTest(unsigned Opcode) const { 1590 switch (Opcode) { 1591 case SystemZ::L: return SystemZ::LT; 1592 case SystemZ::LY: return SystemZ::LT; 1593 case SystemZ::LG: return SystemZ::LTG; 1594 case SystemZ::LGF: return SystemZ::LTGF; 1595 case SystemZ::LR: return SystemZ::LTR; 1596 case SystemZ::LGFR: return SystemZ::LTGFR; 1597 case SystemZ::LGR: return SystemZ::LTGR; 1598 case SystemZ::LER: return SystemZ::LTEBR; 1599 case SystemZ::LDR: return SystemZ::LTDBR; 1600 case SystemZ::LXR: return SystemZ::LTXBR; 1601 case SystemZ::LCDFR: return SystemZ::LCDBR; 1602 case SystemZ::LPDFR: return SystemZ::LPDBR; 1603 case SystemZ::LNDFR: return SystemZ::LNDBR; 1604 case SystemZ::LCDFR_32: return SystemZ::LCEBR; 1605 case SystemZ::LPDFR_32: return SystemZ::LPEBR; 1606 case SystemZ::LNDFR_32: return SystemZ::LNEBR; 1607 // On zEC12 we prefer to use RISBGN. But if there is a chance to 1608 // actually use the condition code, we may turn it back into RISGB. 1609 // Note that RISBG is not really a "load-and-test" instruction, 1610 // but sets the same condition code values, so is OK to use here. 1611 case SystemZ::RISBGN: return SystemZ::RISBG; 1612 default: return 0; 1613 } 1614 } 1615 1616 // Return true if Mask matches the regexp 0*1+0*, given that zero masks 1617 // have already been filtered out. Store the first set bit in LSB and 1618 // the number of set bits in Length if so. 1619 static bool isStringOfOnes(uint64_t Mask, unsigned &LSB, unsigned &Length) { 1620 unsigned First = findFirstSet(Mask); 1621 uint64_t Top = (Mask >> First) + 1; 1622 if ((Top & -Top) == Top) { 1623 LSB = First; 1624 Length = findFirstSet(Top); 1625 return true; 1626 } 1627 return false; 1628 } 1629 1630 bool SystemZInstrInfo::isRxSBGMask(uint64_t Mask, unsigned BitSize, 1631 unsigned &Start, unsigned &End) const { 1632 // Reject trivial all-zero masks. 1633 Mask &= allOnes(BitSize); 1634 if (Mask == 0) 1635 return false; 1636 1637 // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of 1638 // the msb and End specifies the index of the lsb. 1639 unsigned LSB, Length; 1640 if (isStringOfOnes(Mask, LSB, Length)) { 1641 Start = 63 - (LSB + Length - 1); 1642 End = 63 - LSB; 1643 return true; 1644 } 1645 1646 // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb 1647 // of the low 1s and End specifies the lsb of the high 1s. 1648 if (isStringOfOnes(Mask ^ allOnes(BitSize), LSB, Length)) { 1649 assert(LSB > 0 && "Bottom bit must be set"); 1650 assert(LSB + Length < BitSize && "Top bit must be set"); 1651 Start = 63 - (LSB - 1); 1652 End = 63 - (LSB + Length); 1653 return true; 1654 } 1655 1656 return false; 1657 } 1658 1659 unsigned SystemZInstrInfo::getFusedCompare(unsigned Opcode, 1660 SystemZII::FusedCompareType Type, 1661 const MachineInstr *MI) const { 1662 switch (Opcode) { 1663 case SystemZ::CHI: 1664 case SystemZ::CGHI: 1665 if (!(MI && isInt<8>(MI->getOperand(1).getImm()))) 1666 return 0; 1667 break; 1668 case SystemZ::CLFI: 1669 case SystemZ::CLGFI: 1670 if (!(MI && isUInt<8>(MI->getOperand(1).getImm()))) 1671 return 0; 1672 break; 1673 case SystemZ::CL: 1674 case SystemZ::CLG: 1675 if (!STI.hasMiscellaneousExtensions()) 1676 return 0; 1677 if (!(MI && MI->getOperand(3).getReg() == 0)) 1678 return 0; 1679 break; 1680 } 1681 switch (Type) { 1682 case SystemZII::CompareAndBranch: 1683 switch (Opcode) { 1684 case SystemZ::CR: 1685 return SystemZ::CRJ; 1686 case SystemZ::CGR: 1687 return SystemZ::CGRJ; 1688 case SystemZ::CHI: 1689 return SystemZ::CIJ; 1690 case SystemZ::CGHI: 1691 return SystemZ::CGIJ; 1692 case SystemZ::CLR: 1693 return SystemZ::CLRJ; 1694 case SystemZ::CLGR: 1695 return SystemZ::CLGRJ; 1696 case SystemZ::CLFI: 1697 return SystemZ::CLIJ; 1698 case SystemZ::CLGFI: 1699 return SystemZ::CLGIJ; 1700 default: 1701 return 0; 1702 } 1703 case SystemZII::CompareAndReturn: 1704 switch (Opcode) { 1705 case SystemZ::CR: 1706 return SystemZ::CRBReturn; 1707 case SystemZ::CGR: 1708 return SystemZ::CGRBReturn; 1709 case SystemZ::CHI: 1710 return SystemZ::CIBReturn; 1711 case SystemZ::CGHI: 1712 return SystemZ::CGIBReturn; 1713 case SystemZ::CLR: 1714 return SystemZ::CLRBReturn; 1715 case SystemZ::CLGR: 1716 return SystemZ::CLGRBReturn; 1717 case SystemZ::CLFI: 1718 return SystemZ::CLIBReturn; 1719 case SystemZ::CLGFI: 1720 return SystemZ::CLGIBReturn; 1721 default: 1722 return 0; 1723 } 1724 case SystemZII::CompareAndSibcall: 1725 switch (Opcode) { 1726 case SystemZ::CR: 1727 return SystemZ::CRBCall; 1728 case SystemZ::CGR: 1729 return SystemZ::CGRBCall; 1730 case SystemZ::CHI: 1731 return SystemZ::CIBCall; 1732 case SystemZ::CGHI: 1733 return SystemZ::CGIBCall; 1734 case SystemZ::CLR: 1735 return SystemZ::CLRBCall; 1736 case SystemZ::CLGR: 1737 return SystemZ::CLGRBCall; 1738 case SystemZ::CLFI: 1739 return SystemZ::CLIBCall; 1740 case SystemZ::CLGFI: 1741 return SystemZ::CLGIBCall; 1742 default: 1743 return 0; 1744 } 1745 case SystemZII::CompareAndTrap: 1746 switch (Opcode) { 1747 case SystemZ::CR: 1748 return SystemZ::CRT; 1749 case SystemZ::CGR: 1750 return SystemZ::CGRT; 1751 case SystemZ::CHI: 1752 return SystemZ::CIT; 1753 case SystemZ::CGHI: 1754 return SystemZ::CGIT; 1755 case SystemZ::CLR: 1756 return SystemZ::CLRT; 1757 case SystemZ::CLGR: 1758 return SystemZ::CLGRT; 1759 case SystemZ::CLFI: 1760 return SystemZ::CLFIT; 1761 case SystemZ::CLGFI: 1762 return SystemZ::CLGIT; 1763 case SystemZ::CL: 1764 return SystemZ::CLT; 1765 case SystemZ::CLG: 1766 return SystemZ::CLGT; 1767 default: 1768 return 0; 1769 } 1770 } 1771 return 0; 1772 } 1773 1774 unsigned SystemZInstrInfo::getLoadAndTrap(unsigned Opcode) const { 1775 if (!STI.hasLoadAndTrap()) 1776 return 0; 1777 switch (Opcode) { 1778 case SystemZ::L: 1779 case SystemZ::LY: 1780 return SystemZ::LAT; 1781 case SystemZ::LG: 1782 return SystemZ::LGAT; 1783 case SystemZ::LFH: 1784 return SystemZ::LFHAT; 1785 case SystemZ::LLGF: 1786 return SystemZ::LLGFAT; 1787 case SystemZ::LLGT: 1788 return SystemZ::LLGTAT; 1789 } 1790 return 0; 1791 } 1792 1793 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB, 1794 MachineBasicBlock::iterator MBBI, 1795 unsigned Reg, uint64_t Value) const { 1796 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 1797 unsigned Opcode; 1798 if (isInt<16>(Value)) 1799 Opcode = SystemZ::LGHI; 1800 else if (SystemZ::isImmLL(Value)) 1801 Opcode = SystemZ::LLILL; 1802 else if (SystemZ::isImmLH(Value)) { 1803 Opcode = SystemZ::LLILH; 1804 Value >>= 16; 1805 } else { 1806 assert(isInt<32>(Value) && "Huge values not handled yet"); 1807 Opcode = SystemZ::LGFI; 1808 } 1809 BuildMI(MBB, MBBI, DL, get(Opcode), Reg).addImm(Value); 1810 } 1811 1812 bool SystemZInstrInfo:: 1813 areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb, 1814 AliasAnalysis *AA) const { 1815 1816 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand()) 1817 return false; 1818 1819 // If mem-operands show that the same address Value is used by both 1820 // instructions, check for non-overlapping offsets and widths. Not 1821 // sure if a register based analysis would be an improvement... 1822 1823 MachineMemOperand *MMOa = *MIa.memoperands_begin(); 1824 MachineMemOperand *MMOb = *MIb.memoperands_begin(); 1825 const Value *VALa = MMOa->getValue(); 1826 const Value *VALb = MMOb->getValue(); 1827 bool SameVal = (VALa && VALb && (VALa == VALb)); 1828 if (!SameVal) { 1829 const PseudoSourceValue *PSVa = MMOa->getPseudoValue(); 1830 const PseudoSourceValue *PSVb = MMOb->getPseudoValue(); 1831 if (PSVa && PSVb && (PSVa == PSVb)) 1832 SameVal = true; 1833 } 1834 if (SameVal) { 1835 int OffsetA = MMOa->getOffset(), OffsetB = MMOb->getOffset(); 1836 int WidthA = MMOa->getSize(), WidthB = MMOb->getSize(); 1837 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB; 1838 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA; 1839 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; 1840 if (LowOffset + LowWidth <= HighOffset) 1841 return true; 1842 } 1843 1844 return false; 1845 } 1846