1 //===-- SystemZInstrInfo.cpp - SystemZ instruction information ------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains the SystemZ implementation of the TargetInstrInfo class. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "SystemZInstrInfo.h" 14 #include "MCTargetDesc/SystemZMCTargetDesc.h" 15 #include "SystemZ.h" 16 #include "SystemZInstrBuilder.h" 17 #include "SystemZSubtarget.h" 18 #include "llvm/ADT/Statistic.h" 19 #include "llvm/CodeGen/LiveInterval.h" 20 #include "llvm/CodeGen/LiveIntervals.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/CodeGen/TargetInstrInfo.h" 31 #include "llvm/CodeGen/TargetSubtargetInfo.h" 32 #include "llvm/MC/MCInstrDesc.h" 33 #include "llvm/MC/MCRegisterInfo.h" 34 #include "llvm/Support/BranchProbability.h" 35 #include "llvm/Support/ErrorHandling.h" 36 #include "llvm/Support/MathExtras.h" 37 #include "llvm/Target/TargetMachine.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 #define DEBUG_TYPE "systemz-II" 49 50 // Return a mask with Count low bits set. 51 static uint64_t allOnes(unsigned int Count) { 52 return Count == 0 ? 0 : (uint64_t(1) << (Count - 1) << 1) - 1; 53 } 54 55 // Pin the vtable to this file. 56 void SystemZInstrInfo::anchor() {} 57 58 SystemZInstrInfo::SystemZInstrInfo(SystemZSubtarget &sti) 59 : SystemZGenInstrInfo(SystemZ::ADJCALLSTACKDOWN, SystemZ::ADJCALLSTACKUP), 60 RI(), STI(sti) { 61 } 62 63 // MI is a 128-bit load or store. Split it into two 64-bit loads or stores, 64 // each having the opcode given by NewOpcode. 65 void SystemZInstrInfo::splitMove(MachineBasicBlock::iterator MI, 66 unsigned NewOpcode) const { 67 MachineBasicBlock *MBB = MI->getParent(); 68 MachineFunction &MF = *MBB->getParent(); 69 70 // Get two load or store instructions. Use the original instruction for one 71 // of them (arbitrarily the second here) and create a clone for the other. 72 MachineInstr *EarlierMI = MF.CloneMachineInstr(&*MI); 73 MBB->insert(MI, EarlierMI); 74 75 // Set up the two 64-bit registers and remember super reg and its flags. 76 MachineOperand &HighRegOp = EarlierMI->getOperand(0); 77 MachineOperand &LowRegOp = MI->getOperand(0); 78 Register Reg128 = LowRegOp.getReg(); 79 unsigned Reg128Killed = getKillRegState(LowRegOp.isKill()); 80 unsigned Reg128Undef = getUndefRegState(LowRegOp.isUndef()); 81 HighRegOp.setReg(RI.getSubReg(HighRegOp.getReg(), SystemZ::subreg_h64)); 82 LowRegOp.setReg(RI.getSubReg(LowRegOp.getReg(), SystemZ::subreg_l64)); 83 84 if (MI->mayStore()) { 85 // Add implicit uses of the super register in case one of the subregs is 86 // undefined. We could track liveness and skip storing an undefined 87 // subreg, but this is hopefully rare (discovered with llvm-stress). 88 // If Reg128 was killed, set kill flag on MI. 89 unsigned Reg128UndefImpl = (Reg128Undef | RegState::Implicit); 90 MachineInstrBuilder(MF, EarlierMI).addReg(Reg128, Reg128UndefImpl); 91 MachineInstrBuilder(MF, MI).addReg(Reg128, (Reg128UndefImpl | Reg128Killed)); 92 } 93 94 // The address in the first (high) instruction is already correct. 95 // Adjust the offset in the second (low) instruction. 96 MachineOperand &HighOffsetOp = EarlierMI->getOperand(2); 97 MachineOperand &LowOffsetOp = MI->getOperand(2); 98 LowOffsetOp.setImm(LowOffsetOp.getImm() + 8); 99 100 // Clear the kill flags on the registers in the first instruction. 101 if (EarlierMI->getOperand(0).isReg() && EarlierMI->getOperand(0).isUse()) 102 EarlierMI->getOperand(0).setIsKill(false); 103 EarlierMI->getOperand(1).setIsKill(false); 104 EarlierMI->getOperand(3).setIsKill(false); 105 106 // Set the opcodes. 107 unsigned HighOpcode = getOpcodeForOffset(NewOpcode, HighOffsetOp.getImm()); 108 unsigned LowOpcode = getOpcodeForOffset(NewOpcode, LowOffsetOp.getImm()); 109 assert(HighOpcode && LowOpcode && "Both offsets should be in range"); 110 111 EarlierMI->setDesc(get(HighOpcode)); 112 MI->setDesc(get(LowOpcode)); 113 } 114 115 // Split ADJDYNALLOC instruction MI. 116 void SystemZInstrInfo::splitAdjDynAlloc(MachineBasicBlock::iterator MI) const { 117 MachineBasicBlock *MBB = MI->getParent(); 118 MachineFunction &MF = *MBB->getParent(); 119 MachineFrameInfo &MFFrame = MF.getFrameInfo(); 120 MachineOperand &OffsetMO = MI->getOperand(2); 121 122 uint64_t Offset = (MFFrame.getMaxCallFrameSize() + 123 SystemZMC::CallFrameSize + 124 OffsetMO.getImm()); 125 unsigned NewOpcode = getOpcodeForOffset(SystemZ::LA, Offset); 126 assert(NewOpcode && "No support for huge argument lists yet"); 127 MI->setDesc(get(NewOpcode)); 128 OffsetMO.setImm(Offset); 129 } 130 131 // MI is an RI-style pseudo instruction. Replace it with LowOpcode 132 // if the first operand is a low GR32 and HighOpcode if the first operand 133 // is a high GR32. ConvertHigh is true if LowOpcode takes a signed operand 134 // and HighOpcode takes an unsigned 32-bit operand. In those cases, 135 // MI has the same kind of operand as LowOpcode, so needs to be converted 136 // if HighOpcode is used. 137 void SystemZInstrInfo::expandRIPseudo(MachineInstr &MI, unsigned LowOpcode, 138 unsigned HighOpcode, 139 bool ConvertHigh) const { 140 Register Reg = MI.getOperand(0).getReg(); 141 bool IsHigh = SystemZ::isHighReg(Reg); 142 MI.setDesc(get(IsHigh ? HighOpcode : LowOpcode)); 143 if (IsHigh && ConvertHigh) 144 MI.getOperand(1).setImm(uint32_t(MI.getOperand(1).getImm())); 145 } 146 147 // MI is a three-operand RIE-style pseudo instruction. Replace it with 148 // LowOpcodeK if the registers are both low GR32s, otherwise use a move 149 // followed by HighOpcode or LowOpcode, depending on whether the target 150 // is a high or low GR32. 151 void SystemZInstrInfo::expandRIEPseudo(MachineInstr &MI, unsigned LowOpcode, 152 unsigned LowOpcodeK, 153 unsigned HighOpcode) const { 154 Register DestReg = MI.getOperand(0).getReg(); 155 Register SrcReg = MI.getOperand(1).getReg(); 156 bool DestIsHigh = SystemZ::isHighReg(DestReg); 157 bool SrcIsHigh = SystemZ::isHighReg(SrcReg); 158 if (!DestIsHigh && !SrcIsHigh) 159 MI.setDesc(get(LowOpcodeK)); 160 else { 161 if (DestReg != SrcReg) { 162 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), DestReg, SrcReg, 163 SystemZ::LR, 32, MI.getOperand(1).isKill(), 164 MI.getOperand(1).isUndef()); 165 MI.getOperand(1).setReg(DestReg); 166 } 167 MI.setDesc(get(DestIsHigh ? HighOpcode : LowOpcode)); 168 MI.tieOperands(0, 1); 169 } 170 } 171 172 // MI is an RXY-style pseudo instruction. Replace it with LowOpcode 173 // if the first operand is a low GR32 and HighOpcode if the first operand 174 // is a high GR32. 175 void SystemZInstrInfo::expandRXYPseudo(MachineInstr &MI, unsigned LowOpcode, 176 unsigned HighOpcode) const { 177 Register Reg = MI.getOperand(0).getReg(); 178 unsigned Opcode = getOpcodeForOffset( 179 SystemZ::isHighReg(Reg) ? HighOpcode : LowOpcode, 180 MI.getOperand(2).getImm()); 181 MI.setDesc(get(Opcode)); 182 } 183 184 // MI is a load-on-condition pseudo instruction with a single register 185 // (source or destination) operand. Replace it with LowOpcode if the 186 // register is a low GR32 and HighOpcode if the register is a high GR32. 187 void SystemZInstrInfo::expandLOCPseudo(MachineInstr &MI, unsigned LowOpcode, 188 unsigned HighOpcode) const { 189 Register Reg = MI.getOperand(0).getReg(); 190 unsigned Opcode = SystemZ::isHighReg(Reg) ? HighOpcode : LowOpcode; 191 MI.setDesc(get(Opcode)); 192 } 193 194 // MI is an RR-style pseudo instruction that zero-extends the low Size bits 195 // of one GRX32 into another. Replace it with LowOpcode if both operands 196 // are low registers, otherwise use RISB[LH]G. 197 void SystemZInstrInfo::expandZExtPseudo(MachineInstr &MI, unsigned LowOpcode, 198 unsigned Size) const { 199 MachineInstrBuilder MIB = 200 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), 201 MI.getOperand(0).getReg(), MI.getOperand(1).getReg(), LowOpcode, 202 Size, MI.getOperand(1).isKill(), MI.getOperand(1).isUndef()); 203 204 // Keep the remaining operands as-is. 205 for (unsigned I = 2; I < MI.getNumOperands(); ++I) 206 MIB.add(MI.getOperand(I)); 207 208 MI.eraseFromParent(); 209 } 210 211 void SystemZInstrInfo::expandLoadStackGuard(MachineInstr *MI) const { 212 MachineBasicBlock *MBB = MI->getParent(); 213 MachineFunction &MF = *MBB->getParent(); 214 const Register Reg64 = MI->getOperand(0).getReg(); 215 const Register Reg32 = RI.getSubReg(Reg64, SystemZ::subreg_l32); 216 217 // EAR can only load the low subregister so us a shift for %a0 to produce 218 // the GR containing %a0 and %a1. 219 220 // ear <reg>, %a0 221 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::EAR), Reg32) 222 .addReg(SystemZ::A0) 223 .addReg(Reg64, RegState::ImplicitDefine); 224 225 // sllg <reg>, <reg>, 32 226 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::SLLG), Reg64) 227 .addReg(Reg64) 228 .addReg(0) 229 .addImm(32); 230 231 // ear <reg>, %a1 232 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::EAR), Reg32) 233 .addReg(SystemZ::A1); 234 235 // lg <reg>, 40(<reg>) 236 MI->setDesc(get(SystemZ::LG)); 237 MachineInstrBuilder(MF, MI).addReg(Reg64).addImm(40).addReg(0); 238 } 239 240 // Emit a zero-extending move from 32-bit GPR SrcReg to 32-bit GPR 241 // DestReg before MBBI in MBB. Use LowLowOpcode when both DestReg and SrcReg 242 // are low registers, otherwise use RISB[LH]G. Size is the number of bits 243 // taken from the low end of SrcReg (8 for LLCR, 16 for LLHR and 32 for LR). 244 // KillSrc is true if this move is the last use of SrcReg. 245 MachineInstrBuilder 246 SystemZInstrInfo::emitGRX32Move(MachineBasicBlock &MBB, 247 MachineBasicBlock::iterator MBBI, 248 const DebugLoc &DL, unsigned DestReg, 249 unsigned SrcReg, unsigned LowLowOpcode, 250 unsigned Size, bool KillSrc, 251 bool UndefSrc) const { 252 unsigned Opcode; 253 bool DestIsHigh = SystemZ::isHighReg(DestReg); 254 bool SrcIsHigh = SystemZ::isHighReg(SrcReg); 255 if (DestIsHigh && SrcIsHigh) 256 Opcode = SystemZ::RISBHH; 257 else if (DestIsHigh && !SrcIsHigh) 258 Opcode = SystemZ::RISBHL; 259 else if (!DestIsHigh && SrcIsHigh) 260 Opcode = SystemZ::RISBLH; 261 else { 262 return BuildMI(MBB, MBBI, DL, get(LowLowOpcode), DestReg) 263 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)); 264 } 265 unsigned Rotate = (DestIsHigh != SrcIsHigh ? 32 : 0); 266 return BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 267 .addReg(DestReg, RegState::Undef) 268 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)) 269 .addImm(32 - Size).addImm(128 + 31).addImm(Rotate); 270 } 271 272 MachineInstr *SystemZInstrInfo::commuteInstructionImpl(MachineInstr &MI, 273 bool NewMI, 274 unsigned OpIdx1, 275 unsigned OpIdx2) const { 276 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & { 277 if (NewMI) 278 return *MI.getParent()->getParent()->CloneMachineInstr(&MI); 279 return MI; 280 }; 281 282 switch (MI.getOpcode()) { 283 case SystemZ::SELRMux: 284 case SystemZ::SELFHR: 285 case SystemZ::SELR: 286 case SystemZ::SELGR: 287 case SystemZ::LOCRMux: 288 case SystemZ::LOCFHR: 289 case SystemZ::LOCR: 290 case SystemZ::LOCGR: { 291 auto &WorkingMI = cloneIfNew(MI); 292 // Invert condition. 293 unsigned CCValid = WorkingMI.getOperand(3).getImm(); 294 unsigned CCMask = WorkingMI.getOperand(4).getImm(); 295 WorkingMI.getOperand(4).setImm(CCMask ^ CCValid); 296 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 297 OpIdx1, OpIdx2); 298 } 299 default: 300 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 301 } 302 } 303 304 // If MI is a simple load or store for a frame object, return the register 305 // it loads or stores and set FrameIndex to the index of the frame object. 306 // Return 0 otherwise. 307 // 308 // Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 309 static int isSimpleMove(const MachineInstr &MI, int &FrameIndex, 310 unsigned Flag) { 311 const MCInstrDesc &MCID = MI.getDesc(); 312 if ((MCID.TSFlags & Flag) && MI.getOperand(1).isFI() && 313 MI.getOperand(2).getImm() == 0 && MI.getOperand(3).getReg() == 0) { 314 FrameIndex = MI.getOperand(1).getIndex(); 315 return MI.getOperand(0).getReg(); 316 } 317 return 0; 318 } 319 320 unsigned SystemZInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 321 int &FrameIndex) const { 322 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXLoad); 323 } 324 325 unsigned SystemZInstrInfo::isStoreToStackSlot(const MachineInstr &MI, 326 int &FrameIndex) const { 327 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXStore); 328 } 329 330 bool SystemZInstrInfo::isStackSlotCopy(const MachineInstr &MI, 331 int &DestFrameIndex, 332 int &SrcFrameIndex) const { 333 // Check for MVC 0(Length,FI1),0(FI2) 334 const MachineFrameInfo &MFI = MI.getParent()->getParent()->getFrameInfo(); 335 if (MI.getOpcode() != SystemZ::MVC || !MI.getOperand(0).isFI() || 336 MI.getOperand(1).getImm() != 0 || !MI.getOperand(3).isFI() || 337 MI.getOperand(4).getImm() != 0) 338 return false; 339 340 // Check that Length covers the full slots. 341 int64_t Length = MI.getOperand(2).getImm(); 342 unsigned FI1 = MI.getOperand(0).getIndex(); 343 unsigned FI2 = MI.getOperand(3).getIndex(); 344 if (MFI.getObjectSize(FI1) != Length || 345 MFI.getObjectSize(FI2) != Length) 346 return false; 347 348 DestFrameIndex = FI1; 349 SrcFrameIndex = FI2; 350 return true; 351 } 352 353 bool SystemZInstrInfo::analyzeBranch(MachineBasicBlock &MBB, 354 MachineBasicBlock *&TBB, 355 MachineBasicBlock *&FBB, 356 SmallVectorImpl<MachineOperand> &Cond, 357 bool AllowModify) const { 358 // Most of the code and comments here are boilerplate. 359 360 // Start from the bottom of the block and work up, examining the 361 // terminator instructions. 362 MachineBasicBlock::iterator I = MBB.end(); 363 while (I != MBB.begin()) { 364 --I; 365 if (I->isDebugInstr()) 366 continue; 367 368 // Working from the bottom, when we see a non-terminator instruction, we're 369 // done. 370 if (!isUnpredicatedTerminator(*I)) 371 break; 372 373 // A terminator that isn't a branch can't easily be handled by this 374 // analysis. 375 if (!I->isBranch()) 376 return true; 377 378 // Can't handle indirect branches. 379 SystemZII::Branch Branch(getBranchInfo(*I)); 380 if (!Branch.hasMBBTarget()) 381 return true; 382 383 // Punt on compound branches. 384 if (Branch.Type != SystemZII::BranchNormal) 385 return true; 386 387 if (Branch.CCMask == SystemZ::CCMASK_ANY) { 388 // Handle unconditional branches. 389 if (!AllowModify) { 390 TBB = Branch.getMBBTarget(); 391 continue; 392 } 393 394 // If the block has any instructions after a JMP, delete them. 395 while (std::next(I) != MBB.end()) 396 std::next(I)->eraseFromParent(); 397 398 Cond.clear(); 399 FBB = nullptr; 400 401 // Delete the JMP if it's equivalent to a fall-through. 402 if (MBB.isLayoutSuccessor(Branch.getMBBTarget())) { 403 TBB = nullptr; 404 I->eraseFromParent(); 405 I = MBB.end(); 406 continue; 407 } 408 409 // TBB is used to indicate the unconditinal destination. 410 TBB = Branch.getMBBTarget(); 411 continue; 412 } 413 414 // Working from the bottom, handle the first conditional branch. 415 if (Cond.empty()) { 416 // FIXME: add X86-style branch swap 417 FBB = TBB; 418 TBB = Branch.getMBBTarget(); 419 Cond.push_back(MachineOperand::CreateImm(Branch.CCValid)); 420 Cond.push_back(MachineOperand::CreateImm(Branch.CCMask)); 421 continue; 422 } 423 424 // Handle subsequent conditional branches. 425 assert(Cond.size() == 2 && TBB && "Should have seen a conditional branch"); 426 427 // Only handle the case where all conditional branches branch to the same 428 // destination. 429 if (TBB != Branch.getMBBTarget()) 430 return true; 431 432 // If the conditions are the same, we can leave them alone. 433 unsigned OldCCValid = Cond[0].getImm(); 434 unsigned OldCCMask = Cond[1].getImm(); 435 if (OldCCValid == Branch.CCValid && OldCCMask == Branch.CCMask) 436 continue; 437 438 // FIXME: Try combining conditions like X86 does. Should be easy on Z! 439 return false; 440 } 441 442 return false; 443 } 444 445 unsigned SystemZInstrInfo::removeBranch(MachineBasicBlock &MBB, 446 int *BytesRemoved) const { 447 assert(!BytesRemoved && "code size not handled"); 448 449 // Most of the code and comments here are boilerplate. 450 MachineBasicBlock::iterator I = MBB.end(); 451 unsigned Count = 0; 452 453 while (I != MBB.begin()) { 454 --I; 455 if (I->isDebugInstr()) 456 continue; 457 if (!I->isBranch()) 458 break; 459 if (!getBranchInfo(*I).hasMBBTarget()) 460 break; 461 // Remove the branch. 462 I->eraseFromParent(); 463 I = MBB.end(); 464 ++Count; 465 } 466 467 return Count; 468 } 469 470 bool SystemZInstrInfo:: 471 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 472 assert(Cond.size() == 2 && "Invalid condition"); 473 Cond[1].setImm(Cond[1].getImm() ^ Cond[0].getImm()); 474 return false; 475 } 476 477 unsigned SystemZInstrInfo::insertBranch(MachineBasicBlock &MBB, 478 MachineBasicBlock *TBB, 479 MachineBasicBlock *FBB, 480 ArrayRef<MachineOperand> Cond, 481 const DebugLoc &DL, 482 int *BytesAdded) const { 483 // In this function we output 32-bit branches, which should always 484 // have enough range. They can be shortened and relaxed by later code 485 // in the pipeline, if desired. 486 487 // Shouldn't be a fall through. 488 assert(TBB && "insertBranch must not be told to insert a fallthrough"); 489 assert((Cond.size() == 2 || Cond.size() == 0) && 490 "SystemZ branch conditions have one component!"); 491 assert(!BytesAdded && "code size not handled"); 492 493 if (Cond.empty()) { 494 // Unconditional branch? 495 assert(!FBB && "Unconditional branch with multiple successors!"); 496 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(TBB); 497 return 1; 498 } 499 500 // Conditional branch. 501 unsigned Count = 0; 502 unsigned CCValid = Cond[0].getImm(); 503 unsigned CCMask = Cond[1].getImm(); 504 BuildMI(&MBB, DL, get(SystemZ::BRC)) 505 .addImm(CCValid).addImm(CCMask).addMBB(TBB); 506 ++Count; 507 508 if (FBB) { 509 // Two-way Conditional branch. Insert the second branch. 510 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(FBB); 511 ++Count; 512 } 513 return Count; 514 } 515 516 bool SystemZInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg, 517 unsigned &SrcReg2, int &Mask, 518 int &Value) const { 519 assert(MI.isCompare() && "Caller should have checked for a comparison"); 520 521 if (MI.getNumExplicitOperands() == 2 && MI.getOperand(0).isReg() && 522 MI.getOperand(1).isImm()) { 523 SrcReg = MI.getOperand(0).getReg(); 524 SrcReg2 = 0; 525 Value = MI.getOperand(1).getImm(); 526 Mask = ~0; 527 return true; 528 } 529 530 return false; 531 } 532 533 bool SystemZInstrInfo::canInsertSelect(const MachineBasicBlock &MBB, 534 ArrayRef<MachineOperand> Pred, 535 unsigned DstReg, unsigned TrueReg, 536 unsigned FalseReg, int &CondCycles, 537 int &TrueCycles, 538 int &FalseCycles) const { 539 // Not all subtargets have LOCR instructions. 540 if (!STI.hasLoadStoreOnCond()) 541 return false; 542 if (Pred.size() != 2) 543 return false; 544 545 // Check register classes. 546 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 547 const TargetRegisterClass *RC = 548 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 549 if (!RC) 550 return false; 551 552 // We have LOCR instructions for 32 and 64 bit general purpose registers. 553 if ((STI.hasLoadStoreOnCond2() && 554 SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) || 555 SystemZ::GR32BitRegClass.hasSubClassEq(RC) || 556 SystemZ::GR64BitRegClass.hasSubClassEq(RC)) { 557 CondCycles = 2; 558 TrueCycles = 2; 559 FalseCycles = 2; 560 return true; 561 } 562 563 // Can't do anything else. 564 return false; 565 } 566 567 void SystemZInstrInfo::insertSelect(MachineBasicBlock &MBB, 568 MachineBasicBlock::iterator I, 569 const DebugLoc &DL, unsigned DstReg, 570 ArrayRef<MachineOperand> Pred, 571 unsigned TrueReg, 572 unsigned FalseReg) const { 573 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 574 const TargetRegisterClass *RC = MRI.getRegClass(DstReg); 575 576 assert(Pred.size() == 2 && "Invalid condition"); 577 unsigned CCValid = Pred[0].getImm(); 578 unsigned CCMask = Pred[1].getImm(); 579 580 unsigned Opc; 581 if (SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) { 582 if (STI.hasMiscellaneousExtensions3()) 583 Opc = SystemZ::SELRMux; 584 else if (STI.hasLoadStoreOnCond2()) 585 Opc = SystemZ::LOCRMux; 586 else { 587 Opc = SystemZ::LOCR; 588 MRI.constrainRegClass(DstReg, &SystemZ::GR32BitRegClass); 589 Register TReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass); 590 Register FReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass); 591 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), TReg).addReg(TrueReg); 592 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), FReg).addReg(FalseReg); 593 TrueReg = TReg; 594 FalseReg = FReg; 595 } 596 } else if (SystemZ::GR64BitRegClass.hasSubClassEq(RC)) { 597 if (STI.hasMiscellaneousExtensions3()) 598 Opc = SystemZ::SELGR; 599 else 600 Opc = SystemZ::LOCGR; 601 } else 602 llvm_unreachable("Invalid register class"); 603 604 BuildMI(MBB, I, DL, get(Opc), DstReg) 605 .addReg(FalseReg).addReg(TrueReg) 606 .addImm(CCValid).addImm(CCMask); 607 } 608 609 bool SystemZInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, 610 unsigned Reg, 611 MachineRegisterInfo *MRI) const { 612 unsigned DefOpc = DefMI.getOpcode(); 613 if (DefOpc != SystemZ::LHIMux && DefOpc != SystemZ::LHI && 614 DefOpc != SystemZ::LGHI) 615 return false; 616 if (DefMI.getOperand(0).getReg() != Reg) 617 return false; 618 int32_t ImmVal = (int32_t)DefMI.getOperand(1).getImm(); 619 620 unsigned UseOpc = UseMI.getOpcode(); 621 unsigned NewUseOpc; 622 unsigned UseIdx; 623 int CommuteIdx = -1; 624 bool TieOps = false; 625 switch (UseOpc) { 626 case SystemZ::SELRMux: 627 TieOps = true; 628 LLVM_FALLTHROUGH; 629 case SystemZ::LOCRMux: 630 if (!STI.hasLoadStoreOnCond2()) 631 return false; 632 NewUseOpc = SystemZ::LOCHIMux; 633 if (UseMI.getOperand(2).getReg() == Reg) 634 UseIdx = 2; 635 else if (UseMI.getOperand(1).getReg() == Reg) 636 UseIdx = 2, CommuteIdx = 1; 637 else 638 return false; 639 break; 640 case SystemZ::SELGR: 641 TieOps = true; 642 LLVM_FALLTHROUGH; 643 case SystemZ::LOCGR: 644 if (!STI.hasLoadStoreOnCond2()) 645 return false; 646 NewUseOpc = SystemZ::LOCGHI; 647 if (UseMI.getOperand(2).getReg() == Reg) 648 UseIdx = 2; 649 else if (UseMI.getOperand(1).getReg() == Reg) 650 UseIdx = 2, CommuteIdx = 1; 651 else 652 return false; 653 break; 654 default: 655 return false; 656 } 657 658 if (CommuteIdx != -1) 659 if (!commuteInstruction(UseMI, false, CommuteIdx, UseIdx)) 660 return false; 661 662 bool DeleteDef = MRI->hasOneNonDBGUse(Reg); 663 UseMI.setDesc(get(NewUseOpc)); 664 if (TieOps) 665 UseMI.tieOperands(0, 1); 666 UseMI.getOperand(UseIdx).ChangeToImmediate(ImmVal); 667 if (DeleteDef) 668 DefMI.eraseFromParent(); 669 670 return true; 671 } 672 673 bool SystemZInstrInfo::isPredicable(const MachineInstr &MI) const { 674 unsigned Opcode = MI.getOpcode(); 675 if (Opcode == SystemZ::Return || 676 Opcode == SystemZ::Trap || 677 Opcode == SystemZ::CallJG || 678 Opcode == SystemZ::CallBR) 679 return true; 680 return false; 681 } 682 683 bool SystemZInstrInfo:: 684 isProfitableToIfCvt(MachineBasicBlock &MBB, 685 unsigned NumCycles, unsigned ExtraPredCycles, 686 BranchProbability Probability) const { 687 // Avoid using conditional returns at the end of a loop (since then 688 // we'd need to emit an unconditional branch to the beginning anyway, 689 // making the loop body longer). This doesn't apply for low-probability 690 // loops (eg. compare-and-swap retry), so just decide based on branch 691 // probability instead of looping structure. 692 // However, since Compare and Trap instructions cost the same as a regular 693 // Compare instruction, we should allow the if conversion to convert this 694 // into a Conditional Compare regardless of the branch probability. 695 if (MBB.getLastNonDebugInstr()->getOpcode() != SystemZ::Trap && 696 MBB.succ_empty() && Probability < BranchProbability(1, 8)) 697 return false; 698 // For now only convert single instructions. 699 return NumCycles == 1; 700 } 701 702 bool SystemZInstrInfo:: 703 isProfitableToIfCvt(MachineBasicBlock &TMBB, 704 unsigned NumCyclesT, unsigned ExtraPredCyclesT, 705 MachineBasicBlock &FMBB, 706 unsigned NumCyclesF, unsigned ExtraPredCyclesF, 707 BranchProbability Probability) const { 708 // For now avoid converting mutually-exclusive cases. 709 return false; 710 } 711 712 bool SystemZInstrInfo:: 713 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles, 714 BranchProbability Probability) const { 715 // For now only duplicate single instructions. 716 return NumCycles == 1; 717 } 718 719 bool SystemZInstrInfo::PredicateInstruction( 720 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const { 721 assert(Pred.size() == 2 && "Invalid condition"); 722 unsigned CCValid = Pred[0].getImm(); 723 unsigned CCMask = Pred[1].getImm(); 724 assert(CCMask > 0 && CCMask < 15 && "Invalid predicate"); 725 unsigned Opcode = MI.getOpcode(); 726 if (Opcode == SystemZ::Trap) { 727 MI.setDesc(get(SystemZ::CondTrap)); 728 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 729 .addImm(CCValid).addImm(CCMask) 730 .addReg(SystemZ::CC, RegState::Implicit); 731 return true; 732 } 733 if (Opcode == SystemZ::Return) { 734 MI.setDesc(get(SystemZ::CondReturn)); 735 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 736 .addImm(CCValid).addImm(CCMask) 737 .addReg(SystemZ::CC, RegState::Implicit); 738 return true; 739 } 740 if (Opcode == SystemZ::CallJG) { 741 MachineOperand FirstOp = MI.getOperand(0); 742 const uint32_t *RegMask = MI.getOperand(1).getRegMask(); 743 MI.RemoveOperand(1); 744 MI.RemoveOperand(0); 745 MI.setDesc(get(SystemZ::CallBRCL)); 746 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 747 .addImm(CCValid) 748 .addImm(CCMask) 749 .add(FirstOp) 750 .addRegMask(RegMask) 751 .addReg(SystemZ::CC, RegState::Implicit); 752 return true; 753 } 754 if (Opcode == SystemZ::CallBR) { 755 const uint32_t *RegMask = MI.getOperand(0).getRegMask(); 756 MI.RemoveOperand(0); 757 MI.setDesc(get(SystemZ::CallBCR)); 758 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 759 .addImm(CCValid).addImm(CCMask) 760 .addRegMask(RegMask) 761 .addReg(SystemZ::CC, RegState::Implicit); 762 return true; 763 } 764 return false; 765 } 766 767 void SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 768 MachineBasicBlock::iterator MBBI, 769 const DebugLoc &DL, MCRegister DestReg, 770 MCRegister SrcReg, bool KillSrc) const { 771 // Split 128-bit GPR moves into two 64-bit moves. Add implicit uses of the 772 // super register in case one of the subregs is undefined. 773 // This handles ADDR128 too. 774 if (SystemZ::GR128BitRegClass.contains(DestReg, SrcReg)) { 775 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_h64), 776 RI.getSubReg(SrcReg, SystemZ::subreg_h64), KillSrc); 777 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI)) 778 .addReg(SrcReg, RegState::Implicit); 779 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_l64), 780 RI.getSubReg(SrcReg, SystemZ::subreg_l64), KillSrc); 781 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI)) 782 .addReg(SrcReg, (getKillRegState(KillSrc) | RegState::Implicit)); 783 return; 784 } 785 786 if (SystemZ::GRX32BitRegClass.contains(DestReg, SrcReg)) { 787 emitGRX32Move(MBB, MBBI, DL, DestReg, SrcReg, SystemZ::LR, 32, KillSrc, 788 false); 789 return; 790 } 791 792 // Move 128-bit floating-point values between VR128 and FP128. 793 if (SystemZ::VR128BitRegClass.contains(DestReg) && 794 SystemZ::FP128BitRegClass.contains(SrcReg)) { 795 MCRegister SrcRegHi = 796 RI.getMatchingSuperReg(RI.getSubReg(SrcReg, SystemZ::subreg_h64), 797 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass); 798 MCRegister SrcRegLo = 799 RI.getMatchingSuperReg(RI.getSubReg(SrcReg, SystemZ::subreg_l64), 800 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass); 801 802 BuildMI(MBB, MBBI, DL, get(SystemZ::VMRHG), DestReg) 803 .addReg(SrcRegHi, getKillRegState(KillSrc)) 804 .addReg(SrcRegLo, getKillRegState(KillSrc)); 805 return; 806 } 807 if (SystemZ::FP128BitRegClass.contains(DestReg) && 808 SystemZ::VR128BitRegClass.contains(SrcReg)) { 809 MCRegister DestRegHi = 810 RI.getMatchingSuperReg(RI.getSubReg(DestReg, SystemZ::subreg_h64), 811 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass); 812 MCRegister DestRegLo = 813 RI.getMatchingSuperReg(RI.getSubReg(DestReg, SystemZ::subreg_l64), 814 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass); 815 816 if (DestRegHi != SrcReg) 817 copyPhysReg(MBB, MBBI, DL, DestRegHi, SrcReg, false); 818 BuildMI(MBB, MBBI, DL, get(SystemZ::VREPG), DestRegLo) 819 .addReg(SrcReg, getKillRegState(KillSrc)).addImm(1); 820 return; 821 } 822 823 // Move CC value from/to a GR32. 824 if (SrcReg == SystemZ::CC) { 825 auto MIB = BuildMI(MBB, MBBI, DL, get(SystemZ::IPM), DestReg); 826 if (KillSrc) { 827 const MachineFunction *MF = MBB.getParent(); 828 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); 829 MIB->addRegisterKilled(SrcReg, TRI); 830 } 831 return; 832 } 833 if (DestReg == SystemZ::CC) { 834 BuildMI(MBB, MBBI, DL, get(SystemZ::TMLH)) 835 .addReg(SrcReg, getKillRegState(KillSrc)) 836 .addImm(3 << (SystemZ::IPM_CC - 16)); 837 return; 838 } 839 840 // Everything else needs only one instruction. 841 unsigned Opcode; 842 if (SystemZ::GR64BitRegClass.contains(DestReg, SrcReg)) 843 Opcode = SystemZ::LGR; 844 else if (SystemZ::FP32BitRegClass.contains(DestReg, SrcReg)) 845 // For z13 we prefer LDR over LER to avoid partial register dependencies. 846 Opcode = STI.hasVector() ? SystemZ::LDR32 : SystemZ::LER; 847 else if (SystemZ::FP64BitRegClass.contains(DestReg, SrcReg)) 848 Opcode = SystemZ::LDR; 849 else if (SystemZ::FP128BitRegClass.contains(DestReg, SrcReg)) 850 Opcode = SystemZ::LXR; 851 else if (SystemZ::VR32BitRegClass.contains(DestReg, SrcReg)) 852 Opcode = SystemZ::VLR32; 853 else if (SystemZ::VR64BitRegClass.contains(DestReg, SrcReg)) 854 Opcode = SystemZ::VLR64; 855 else if (SystemZ::VR128BitRegClass.contains(DestReg, SrcReg)) 856 Opcode = SystemZ::VLR; 857 else if (SystemZ::AR32BitRegClass.contains(DestReg, SrcReg)) 858 Opcode = SystemZ::CPYA; 859 else if (SystemZ::AR32BitRegClass.contains(DestReg) && 860 SystemZ::GR32BitRegClass.contains(SrcReg)) 861 Opcode = SystemZ::SAR; 862 else if (SystemZ::GR32BitRegClass.contains(DestReg) && 863 SystemZ::AR32BitRegClass.contains(SrcReg)) 864 Opcode = SystemZ::EAR; 865 else 866 llvm_unreachable("Impossible reg-to-reg copy"); 867 868 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 869 .addReg(SrcReg, getKillRegState(KillSrc)); 870 } 871 872 void SystemZInstrInfo::storeRegToStackSlot( 873 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, Register SrcReg, 874 bool isKill, int FrameIdx, const TargetRegisterClass *RC, 875 const TargetRegisterInfo *TRI) const { 876 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 877 878 // Callers may expect a single instruction, so keep 128-bit moves 879 // together for now and lower them after register allocation. 880 unsigned LoadOpcode, StoreOpcode; 881 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 882 addFrameReference(BuildMI(MBB, MBBI, DL, get(StoreOpcode)) 883 .addReg(SrcReg, getKillRegState(isKill)), 884 FrameIdx); 885 } 886 887 void SystemZInstrInfo::loadRegFromStackSlot( 888 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, Register DestReg, 889 int FrameIdx, const TargetRegisterClass *RC, 890 const TargetRegisterInfo *TRI) const { 891 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 892 893 // Callers may expect a single instruction, so keep 128-bit moves 894 // together for now and lower them after register allocation. 895 unsigned LoadOpcode, StoreOpcode; 896 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 897 addFrameReference(BuildMI(MBB, MBBI, DL, get(LoadOpcode), DestReg), 898 FrameIdx); 899 } 900 901 // Return true if MI is a simple load or store with a 12-bit displacement 902 // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 903 static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) { 904 const MCInstrDesc &MCID = MI->getDesc(); 905 return ((MCID.TSFlags & Flag) && 906 isUInt<12>(MI->getOperand(2).getImm()) && 907 MI->getOperand(3).getReg() == 0); 908 } 909 910 namespace { 911 912 struct LogicOp { 913 LogicOp() = default; 914 LogicOp(unsigned regSize, unsigned immLSB, unsigned immSize) 915 : RegSize(regSize), ImmLSB(immLSB), ImmSize(immSize) {} 916 917 explicit operator bool() const { return RegSize; } 918 919 unsigned RegSize = 0; 920 unsigned ImmLSB = 0; 921 unsigned ImmSize = 0; 922 }; 923 924 } // end anonymous namespace 925 926 static LogicOp interpretAndImmediate(unsigned Opcode) { 927 switch (Opcode) { 928 case SystemZ::NILMux: return LogicOp(32, 0, 16); 929 case SystemZ::NIHMux: return LogicOp(32, 16, 16); 930 case SystemZ::NILL64: return LogicOp(64, 0, 16); 931 case SystemZ::NILH64: return LogicOp(64, 16, 16); 932 case SystemZ::NIHL64: return LogicOp(64, 32, 16); 933 case SystemZ::NIHH64: return LogicOp(64, 48, 16); 934 case SystemZ::NIFMux: return LogicOp(32, 0, 32); 935 case SystemZ::NILF64: return LogicOp(64, 0, 32); 936 case SystemZ::NIHF64: return LogicOp(64, 32, 32); 937 default: return LogicOp(); 938 } 939 } 940 941 static void transferDeadCC(MachineInstr *OldMI, MachineInstr *NewMI) { 942 if (OldMI->registerDefIsDead(SystemZ::CC)) { 943 MachineOperand *CCDef = NewMI->findRegisterDefOperand(SystemZ::CC); 944 if (CCDef != nullptr) 945 CCDef->setIsDead(true); 946 } 947 } 948 949 static void transferMIFlag(MachineInstr *OldMI, MachineInstr *NewMI, 950 MachineInstr::MIFlag Flag) { 951 if (OldMI->getFlag(Flag)) 952 NewMI->setFlag(Flag); 953 } 954 955 MachineInstr *SystemZInstrInfo::convertToThreeAddress( 956 MachineFunction::iterator &MFI, MachineInstr &MI, LiveVariables *LV) const { 957 MachineBasicBlock *MBB = MI.getParent(); 958 959 // Try to convert an AND into an RISBG-type instruction. 960 // TODO: It might be beneficial to select RISBG and shorten to AND instead. 961 if (LogicOp And = interpretAndImmediate(MI.getOpcode())) { 962 uint64_t Imm = MI.getOperand(2).getImm() << And.ImmLSB; 963 // AND IMMEDIATE leaves the other bits of the register unchanged. 964 Imm |= allOnes(And.RegSize) & ~(allOnes(And.ImmSize) << And.ImmLSB); 965 unsigned Start, End; 966 if (isRxSBGMask(Imm, And.RegSize, Start, End)) { 967 unsigned NewOpcode; 968 if (And.RegSize == 64) { 969 NewOpcode = SystemZ::RISBG; 970 // Prefer RISBGN if available, since it does not clobber CC. 971 if (STI.hasMiscellaneousExtensions()) 972 NewOpcode = SystemZ::RISBGN; 973 } else { 974 NewOpcode = SystemZ::RISBMux; 975 Start &= 31; 976 End &= 31; 977 } 978 MachineOperand &Dest = MI.getOperand(0); 979 MachineOperand &Src = MI.getOperand(1); 980 MachineInstrBuilder MIB = 981 BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpcode)) 982 .add(Dest) 983 .addReg(0) 984 .addReg(Src.getReg(), getKillRegState(Src.isKill()), 985 Src.getSubReg()) 986 .addImm(Start) 987 .addImm(End + 128) 988 .addImm(0); 989 if (LV) { 990 unsigned NumOps = MI.getNumOperands(); 991 for (unsigned I = 1; I < NumOps; ++I) { 992 MachineOperand &Op = MI.getOperand(I); 993 if (Op.isReg() && Op.isKill()) 994 LV->replaceKillInstruction(Op.getReg(), MI, *MIB); 995 } 996 } 997 transferDeadCC(&MI, MIB); 998 return MIB; 999 } 1000 } 1001 return nullptr; 1002 } 1003 1004 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1005 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1006 MachineBasicBlock::iterator InsertPt, int FrameIndex, 1007 LiveIntervals *LIS, VirtRegMap *VRM) const { 1008 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 1009 const MachineFrameInfo &MFI = MF.getFrameInfo(); 1010 unsigned Size = MFI.getObjectSize(FrameIndex); 1011 unsigned Opcode = MI.getOpcode(); 1012 1013 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 1014 if (LIS != nullptr && (Opcode == SystemZ::LA || Opcode == SystemZ::LAY) && 1015 isInt<8>(MI.getOperand(2).getImm()) && !MI.getOperand(3).getReg()) { 1016 1017 // Check CC liveness, since new instruction introduces a dead 1018 // def of CC. 1019 MCRegUnitIterator CCUnit(SystemZ::CC, TRI); 1020 LiveRange &CCLiveRange = LIS->getRegUnit(*CCUnit); 1021 ++CCUnit; 1022 assert(!CCUnit.isValid() && "CC only has one reg unit."); 1023 SlotIndex MISlot = 1024 LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot(); 1025 if (!CCLiveRange.liveAt(MISlot)) { 1026 // LA(Y) %reg, CONST(%reg) -> AGSI %mem, CONST 1027 MachineInstr *BuiltMI = BuildMI(*InsertPt->getParent(), InsertPt, 1028 MI.getDebugLoc(), get(SystemZ::AGSI)) 1029 .addFrameIndex(FrameIndex) 1030 .addImm(0) 1031 .addImm(MI.getOperand(2).getImm()); 1032 BuiltMI->findRegisterDefOperand(SystemZ::CC)->setIsDead(true); 1033 CCLiveRange.createDeadDef(MISlot, LIS->getVNInfoAllocator()); 1034 return BuiltMI; 1035 } 1036 } 1037 return nullptr; 1038 } 1039 1040 // All other cases require a single operand. 1041 if (Ops.size() != 1) 1042 return nullptr; 1043 1044 unsigned OpNum = Ops[0]; 1045 assert(Size * 8 == 1046 TRI->getRegSizeInBits(*MF.getRegInfo() 1047 .getRegClass(MI.getOperand(OpNum).getReg())) && 1048 "Invalid size combination"); 1049 1050 if ((Opcode == SystemZ::AHI || Opcode == SystemZ::AGHI) && OpNum == 0 && 1051 isInt<8>(MI.getOperand(2).getImm())) { 1052 // A(G)HI %reg, CONST -> A(G)SI %mem, CONST 1053 Opcode = (Opcode == SystemZ::AHI ? SystemZ::ASI : SystemZ::AGSI); 1054 MachineInstr *BuiltMI = 1055 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode)) 1056 .addFrameIndex(FrameIndex) 1057 .addImm(0) 1058 .addImm(MI.getOperand(2).getImm()); 1059 transferDeadCC(&MI, BuiltMI); 1060 transferMIFlag(&MI, BuiltMI, MachineInstr::NoSWrap); 1061 return BuiltMI; 1062 } 1063 1064 if ((Opcode == SystemZ::ALFI && OpNum == 0 && 1065 isInt<8>((int32_t)MI.getOperand(2).getImm())) || 1066 (Opcode == SystemZ::ALGFI && OpNum == 0 && 1067 isInt<8>((int64_t)MI.getOperand(2).getImm()))) { 1068 // AL(G)FI %reg, CONST -> AL(G)SI %mem, CONST 1069 Opcode = (Opcode == SystemZ::ALFI ? SystemZ::ALSI : SystemZ::ALGSI); 1070 MachineInstr *BuiltMI = 1071 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode)) 1072 .addFrameIndex(FrameIndex) 1073 .addImm(0) 1074 .addImm((int8_t)MI.getOperand(2).getImm()); 1075 transferDeadCC(&MI, BuiltMI); 1076 return BuiltMI; 1077 } 1078 1079 if ((Opcode == SystemZ::SLFI && OpNum == 0 && 1080 isInt<8>((int32_t)-MI.getOperand(2).getImm())) || 1081 (Opcode == SystemZ::SLGFI && OpNum == 0 && 1082 isInt<8>((int64_t)-MI.getOperand(2).getImm()))) { 1083 // SL(G)FI %reg, CONST -> AL(G)SI %mem, -CONST 1084 Opcode = (Opcode == SystemZ::SLFI ? SystemZ::ALSI : SystemZ::ALGSI); 1085 MachineInstr *BuiltMI = 1086 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode)) 1087 .addFrameIndex(FrameIndex) 1088 .addImm(0) 1089 .addImm((int8_t)-MI.getOperand(2).getImm()); 1090 transferDeadCC(&MI, BuiltMI); 1091 return BuiltMI; 1092 } 1093 1094 if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) { 1095 bool Op0IsGPR = (Opcode == SystemZ::LGDR); 1096 bool Op1IsGPR = (Opcode == SystemZ::LDGR); 1097 // If we're spilling the destination of an LDGR or LGDR, store the 1098 // source register instead. 1099 if (OpNum == 0) { 1100 unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD; 1101 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1102 get(StoreOpcode)) 1103 .add(MI.getOperand(1)) 1104 .addFrameIndex(FrameIndex) 1105 .addImm(0) 1106 .addReg(0); 1107 } 1108 // If we're spilling the source of an LDGR or LGDR, load the 1109 // destination register instead. 1110 if (OpNum == 1) { 1111 unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD; 1112 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1113 get(LoadOpcode)) 1114 .add(MI.getOperand(0)) 1115 .addFrameIndex(FrameIndex) 1116 .addImm(0) 1117 .addReg(0); 1118 } 1119 } 1120 1121 // Look for cases where the source of a simple store or the destination 1122 // of a simple load is being spilled. Try to use MVC instead. 1123 // 1124 // Although MVC is in practice a fast choice in these cases, it is still 1125 // logically a bytewise copy. This means that we cannot use it if the 1126 // load or store is volatile. We also wouldn't be able to use MVC if 1127 // the two memories partially overlap, but that case cannot occur here, 1128 // because we know that one of the memories is a full frame index. 1129 // 1130 // For performance reasons, we also want to avoid using MVC if the addresses 1131 // might be equal. We don't worry about that case here, because spill slot 1132 // coloring happens later, and because we have special code to remove 1133 // MVCs that turn out to be redundant. 1134 if (OpNum == 0 && MI.hasOneMemOperand()) { 1135 MachineMemOperand *MMO = *MI.memoperands_begin(); 1136 if (MMO->getSize() == Size && !MMO->isVolatile() && !MMO->isAtomic()) { 1137 // Handle conversion of loads. 1138 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXLoad)) { 1139 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1140 get(SystemZ::MVC)) 1141 .addFrameIndex(FrameIndex) 1142 .addImm(0) 1143 .addImm(Size) 1144 .add(MI.getOperand(1)) 1145 .addImm(MI.getOperand(2).getImm()) 1146 .addMemOperand(MMO); 1147 } 1148 // Handle conversion of stores. 1149 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXStore)) { 1150 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1151 get(SystemZ::MVC)) 1152 .add(MI.getOperand(1)) 1153 .addImm(MI.getOperand(2).getImm()) 1154 .addImm(Size) 1155 .addFrameIndex(FrameIndex) 1156 .addImm(0) 1157 .addMemOperand(MMO); 1158 } 1159 } 1160 } 1161 1162 // If the spilled operand is the final one or the instruction is 1163 // commutable, try to change <INSN>R into <INSN>. 1164 unsigned NumOps = MI.getNumExplicitOperands(); 1165 int MemOpcode = SystemZ::getMemOpcode(Opcode); 1166 1167 // See if this is a 3-address instruction that is convertible to 2-address 1168 // and suitable for folding below. Only try this with virtual registers 1169 // and a provided VRM (during regalloc). 1170 bool NeedsCommute = false; 1171 if (SystemZ::getTwoOperandOpcode(Opcode) != -1 && MemOpcode != -1) { 1172 if (VRM == nullptr) 1173 MemOpcode = -1; 1174 else { 1175 assert(NumOps == 3 && "Expected two source registers."); 1176 Register DstReg = MI.getOperand(0).getReg(); 1177 Register DstPhys = 1178 (Register::isVirtualRegister(DstReg) ? VRM->getPhys(DstReg) : DstReg); 1179 Register SrcReg = (OpNum == 2 ? MI.getOperand(1).getReg() 1180 : ((OpNum == 1 && MI.isCommutable()) 1181 ? MI.getOperand(2).getReg() 1182 : Register())); 1183 if (DstPhys && !SystemZ::GRH32BitRegClass.contains(DstPhys) && SrcReg && 1184 Register::isVirtualRegister(SrcReg) && 1185 DstPhys == VRM->getPhys(SrcReg)) 1186 NeedsCommute = (OpNum == 1); 1187 else 1188 MemOpcode = -1; 1189 } 1190 } 1191 1192 if (MemOpcode >= 0) { 1193 if ((OpNum == NumOps - 1) || NeedsCommute) { 1194 const MCInstrDesc &MemDesc = get(MemOpcode); 1195 uint64_t AccessBytes = SystemZII::getAccessSize(MemDesc.TSFlags); 1196 assert(AccessBytes != 0 && "Size of access should be known"); 1197 assert(AccessBytes <= Size && "Access outside the frame index"); 1198 uint64_t Offset = Size - AccessBytes; 1199 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt, 1200 MI.getDebugLoc(), get(MemOpcode)); 1201 MIB.add(MI.getOperand(0)); 1202 if (NeedsCommute) 1203 MIB.add(MI.getOperand(2)); 1204 else 1205 for (unsigned I = 1; I < OpNum; ++I) 1206 MIB.add(MI.getOperand(I)); 1207 MIB.addFrameIndex(FrameIndex).addImm(Offset); 1208 if (MemDesc.TSFlags & SystemZII::HasIndex) 1209 MIB.addReg(0); 1210 transferDeadCC(&MI, MIB); 1211 transferMIFlag(&MI, MIB, MachineInstr::NoSWrap); 1212 return MIB; 1213 } 1214 } 1215 1216 return nullptr; 1217 } 1218 1219 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1220 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1221 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, 1222 LiveIntervals *LIS) const { 1223 return nullptr; 1224 } 1225 1226 bool SystemZInstrInfo::expandPostRAPseudo(MachineInstr &MI) const { 1227 switch (MI.getOpcode()) { 1228 case SystemZ::L128: 1229 splitMove(MI, SystemZ::LG); 1230 return true; 1231 1232 case SystemZ::ST128: 1233 splitMove(MI, SystemZ::STG); 1234 return true; 1235 1236 case SystemZ::LX: 1237 splitMove(MI, SystemZ::LD); 1238 return true; 1239 1240 case SystemZ::STX: 1241 splitMove(MI, SystemZ::STD); 1242 return true; 1243 1244 case SystemZ::LBMux: 1245 expandRXYPseudo(MI, SystemZ::LB, SystemZ::LBH); 1246 return true; 1247 1248 case SystemZ::LHMux: 1249 expandRXYPseudo(MI, SystemZ::LH, SystemZ::LHH); 1250 return true; 1251 1252 case SystemZ::LLCRMux: 1253 expandZExtPseudo(MI, SystemZ::LLCR, 8); 1254 return true; 1255 1256 case SystemZ::LLHRMux: 1257 expandZExtPseudo(MI, SystemZ::LLHR, 16); 1258 return true; 1259 1260 case SystemZ::LLCMux: 1261 expandRXYPseudo(MI, SystemZ::LLC, SystemZ::LLCH); 1262 return true; 1263 1264 case SystemZ::LLHMux: 1265 expandRXYPseudo(MI, SystemZ::LLH, SystemZ::LLHH); 1266 return true; 1267 1268 case SystemZ::LMux: 1269 expandRXYPseudo(MI, SystemZ::L, SystemZ::LFH); 1270 return true; 1271 1272 case SystemZ::LOCMux: 1273 expandLOCPseudo(MI, SystemZ::LOC, SystemZ::LOCFH); 1274 return true; 1275 1276 case SystemZ::LOCHIMux: 1277 expandLOCPseudo(MI, SystemZ::LOCHI, SystemZ::LOCHHI); 1278 return true; 1279 1280 case SystemZ::STCMux: 1281 expandRXYPseudo(MI, SystemZ::STC, SystemZ::STCH); 1282 return true; 1283 1284 case SystemZ::STHMux: 1285 expandRXYPseudo(MI, SystemZ::STH, SystemZ::STHH); 1286 return true; 1287 1288 case SystemZ::STMux: 1289 expandRXYPseudo(MI, SystemZ::ST, SystemZ::STFH); 1290 return true; 1291 1292 case SystemZ::STOCMux: 1293 expandLOCPseudo(MI, SystemZ::STOC, SystemZ::STOCFH); 1294 return true; 1295 1296 case SystemZ::LHIMux: 1297 expandRIPseudo(MI, SystemZ::LHI, SystemZ::IIHF, true); 1298 return true; 1299 1300 case SystemZ::IIFMux: 1301 expandRIPseudo(MI, SystemZ::IILF, SystemZ::IIHF, false); 1302 return true; 1303 1304 case SystemZ::IILMux: 1305 expandRIPseudo(MI, SystemZ::IILL, SystemZ::IIHL, false); 1306 return true; 1307 1308 case SystemZ::IIHMux: 1309 expandRIPseudo(MI, SystemZ::IILH, SystemZ::IIHH, false); 1310 return true; 1311 1312 case SystemZ::NIFMux: 1313 expandRIPseudo(MI, SystemZ::NILF, SystemZ::NIHF, false); 1314 return true; 1315 1316 case SystemZ::NILMux: 1317 expandRIPseudo(MI, SystemZ::NILL, SystemZ::NIHL, false); 1318 return true; 1319 1320 case SystemZ::NIHMux: 1321 expandRIPseudo(MI, SystemZ::NILH, SystemZ::NIHH, false); 1322 return true; 1323 1324 case SystemZ::OIFMux: 1325 expandRIPseudo(MI, SystemZ::OILF, SystemZ::OIHF, false); 1326 return true; 1327 1328 case SystemZ::OILMux: 1329 expandRIPseudo(MI, SystemZ::OILL, SystemZ::OIHL, false); 1330 return true; 1331 1332 case SystemZ::OIHMux: 1333 expandRIPseudo(MI, SystemZ::OILH, SystemZ::OIHH, false); 1334 return true; 1335 1336 case SystemZ::XIFMux: 1337 expandRIPseudo(MI, SystemZ::XILF, SystemZ::XIHF, false); 1338 return true; 1339 1340 case SystemZ::TMLMux: 1341 expandRIPseudo(MI, SystemZ::TMLL, SystemZ::TMHL, false); 1342 return true; 1343 1344 case SystemZ::TMHMux: 1345 expandRIPseudo(MI, SystemZ::TMLH, SystemZ::TMHH, false); 1346 return true; 1347 1348 case SystemZ::AHIMux: 1349 expandRIPseudo(MI, SystemZ::AHI, SystemZ::AIH, false); 1350 return true; 1351 1352 case SystemZ::AHIMuxK: 1353 expandRIEPseudo(MI, SystemZ::AHI, SystemZ::AHIK, SystemZ::AIH); 1354 return true; 1355 1356 case SystemZ::AFIMux: 1357 expandRIPseudo(MI, SystemZ::AFI, SystemZ::AIH, false); 1358 return true; 1359 1360 case SystemZ::CHIMux: 1361 expandRIPseudo(MI, SystemZ::CHI, SystemZ::CIH, false); 1362 return true; 1363 1364 case SystemZ::CFIMux: 1365 expandRIPseudo(MI, SystemZ::CFI, SystemZ::CIH, false); 1366 return true; 1367 1368 case SystemZ::CLFIMux: 1369 expandRIPseudo(MI, SystemZ::CLFI, SystemZ::CLIH, false); 1370 return true; 1371 1372 case SystemZ::CMux: 1373 expandRXYPseudo(MI, SystemZ::C, SystemZ::CHF); 1374 return true; 1375 1376 case SystemZ::CLMux: 1377 expandRXYPseudo(MI, SystemZ::CL, SystemZ::CLHF); 1378 return true; 1379 1380 case SystemZ::RISBMux: { 1381 bool DestIsHigh = SystemZ::isHighReg(MI.getOperand(0).getReg()); 1382 bool SrcIsHigh = SystemZ::isHighReg(MI.getOperand(2).getReg()); 1383 if (SrcIsHigh == DestIsHigh) 1384 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHH : SystemZ::RISBLL)); 1385 else { 1386 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHL : SystemZ::RISBLH)); 1387 MI.getOperand(5).setImm(MI.getOperand(5).getImm() ^ 32); 1388 } 1389 return true; 1390 } 1391 1392 case SystemZ::ADJDYNALLOC: 1393 splitAdjDynAlloc(MI); 1394 return true; 1395 1396 case TargetOpcode::LOAD_STACK_GUARD: 1397 expandLoadStackGuard(&MI); 1398 return true; 1399 1400 default: 1401 return false; 1402 } 1403 } 1404 1405 unsigned SystemZInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { 1406 if (MI.isInlineAsm()) { 1407 const MachineFunction *MF = MI.getParent()->getParent(); 1408 const char *AsmStr = MI.getOperand(0).getSymbolName(); 1409 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo()); 1410 } 1411 return MI.getDesc().getSize(); 1412 } 1413 1414 SystemZII::Branch 1415 SystemZInstrInfo::getBranchInfo(const MachineInstr &MI) const { 1416 switch (MI.getOpcode()) { 1417 case SystemZ::BR: 1418 case SystemZ::BI: 1419 case SystemZ::J: 1420 case SystemZ::JG: 1421 return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY, 1422 SystemZ::CCMASK_ANY, &MI.getOperand(0)); 1423 1424 case SystemZ::BRC: 1425 case SystemZ::BRCL: 1426 return SystemZII::Branch(SystemZII::BranchNormal, MI.getOperand(0).getImm(), 1427 MI.getOperand(1).getImm(), &MI.getOperand(2)); 1428 1429 case SystemZ::BRCT: 1430 case SystemZ::BRCTH: 1431 return SystemZII::Branch(SystemZII::BranchCT, SystemZ::CCMASK_ICMP, 1432 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1433 1434 case SystemZ::BRCTG: 1435 return SystemZII::Branch(SystemZII::BranchCTG, SystemZ::CCMASK_ICMP, 1436 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1437 1438 case SystemZ::CIJ: 1439 case SystemZ::CRJ: 1440 return SystemZII::Branch(SystemZII::BranchC, SystemZ::CCMASK_ICMP, 1441 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1442 1443 case SystemZ::CLIJ: 1444 case SystemZ::CLRJ: 1445 return SystemZII::Branch(SystemZII::BranchCL, SystemZ::CCMASK_ICMP, 1446 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1447 1448 case SystemZ::CGIJ: 1449 case SystemZ::CGRJ: 1450 return SystemZII::Branch(SystemZII::BranchCG, SystemZ::CCMASK_ICMP, 1451 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1452 1453 case SystemZ::CLGIJ: 1454 case SystemZ::CLGRJ: 1455 return SystemZII::Branch(SystemZII::BranchCLG, SystemZ::CCMASK_ICMP, 1456 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1457 1458 case SystemZ::INLINEASM_BR: 1459 // Don't try to analyze asm goto, so pass nullptr as branch target argument. 1460 return SystemZII::Branch(SystemZII::AsmGoto, 0, 0, nullptr); 1461 1462 default: 1463 llvm_unreachable("Unrecognized branch opcode"); 1464 } 1465 } 1466 1467 void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC, 1468 unsigned &LoadOpcode, 1469 unsigned &StoreOpcode) const { 1470 if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) { 1471 LoadOpcode = SystemZ::L; 1472 StoreOpcode = SystemZ::ST; 1473 } else if (RC == &SystemZ::GRH32BitRegClass) { 1474 LoadOpcode = SystemZ::LFH; 1475 StoreOpcode = SystemZ::STFH; 1476 } else if (RC == &SystemZ::GRX32BitRegClass) { 1477 LoadOpcode = SystemZ::LMux; 1478 StoreOpcode = SystemZ::STMux; 1479 } else if (RC == &SystemZ::GR64BitRegClass || 1480 RC == &SystemZ::ADDR64BitRegClass) { 1481 LoadOpcode = SystemZ::LG; 1482 StoreOpcode = SystemZ::STG; 1483 } else if (RC == &SystemZ::GR128BitRegClass || 1484 RC == &SystemZ::ADDR128BitRegClass) { 1485 LoadOpcode = SystemZ::L128; 1486 StoreOpcode = SystemZ::ST128; 1487 } else if (RC == &SystemZ::FP32BitRegClass) { 1488 LoadOpcode = SystemZ::LE; 1489 StoreOpcode = SystemZ::STE; 1490 } else if (RC == &SystemZ::FP64BitRegClass) { 1491 LoadOpcode = SystemZ::LD; 1492 StoreOpcode = SystemZ::STD; 1493 } else if (RC == &SystemZ::FP128BitRegClass) { 1494 LoadOpcode = SystemZ::LX; 1495 StoreOpcode = SystemZ::STX; 1496 } else if (RC == &SystemZ::VR32BitRegClass) { 1497 LoadOpcode = SystemZ::VL32; 1498 StoreOpcode = SystemZ::VST32; 1499 } else if (RC == &SystemZ::VR64BitRegClass) { 1500 LoadOpcode = SystemZ::VL64; 1501 StoreOpcode = SystemZ::VST64; 1502 } else if (RC == &SystemZ::VF128BitRegClass || 1503 RC == &SystemZ::VR128BitRegClass) { 1504 LoadOpcode = SystemZ::VL; 1505 StoreOpcode = SystemZ::VST; 1506 } else 1507 llvm_unreachable("Unsupported regclass to load or store"); 1508 } 1509 1510 unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode, 1511 int64_t Offset) const { 1512 const MCInstrDesc &MCID = get(Opcode); 1513 int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset); 1514 if (isUInt<12>(Offset) && isUInt<12>(Offset2)) { 1515 // Get the instruction to use for unsigned 12-bit displacements. 1516 int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode); 1517 if (Disp12Opcode >= 0) 1518 return Disp12Opcode; 1519 1520 // All address-related instructions can use unsigned 12-bit 1521 // displacements. 1522 return Opcode; 1523 } 1524 if (isInt<20>(Offset) && isInt<20>(Offset2)) { 1525 // Get the instruction to use for signed 20-bit displacements. 1526 int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode); 1527 if (Disp20Opcode >= 0) 1528 return Disp20Opcode; 1529 1530 // Check whether Opcode allows signed 20-bit displacements. 1531 if (MCID.TSFlags & SystemZII::Has20BitOffset) 1532 return Opcode; 1533 } 1534 return 0; 1535 } 1536 1537 unsigned SystemZInstrInfo::getLoadAndTest(unsigned Opcode) const { 1538 switch (Opcode) { 1539 case SystemZ::L: return SystemZ::LT; 1540 case SystemZ::LY: return SystemZ::LT; 1541 case SystemZ::LG: return SystemZ::LTG; 1542 case SystemZ::LGF: return SystemZ::LTGF; 1543 case SystemZ::LR: return SystemZ::LTR; 1544 case SystemZ::LGFR: return SystemZ::LTGFR; 1545 case SystemZ::LGR: return SystemZ::LTGR; 1546 case SystemZ::LER: return SystemZ::LTEBR; 1547 case SystemZ::LDR: return SystemZ::LTDBR; 1548 case SystemZ::LXR: return SystemZ::LTXBR; 1549 case SystemZ::LCDFR: return SystemZ::LCDBR; 1550 case SystemZ::LPDFR: return SystemZ::LPDBR; 1551 case SystemZ::LNDFR: return SystemZ::LNDBR; 1552 case SystemZ::LCDFR_32: return SystemZ::LCEBR; 1553 case SystemZ::LPDFR_32: return SystemZ::LPEBR; 1554 case SystemZ::LNDFR_32: return SystemZ::LNEBR; 1555 // On zEC12 we prefer to use RISBGN. But if there is a chance to 1556 // actually use the condition code, we may turn it back into RISGB. 1557 // Note that RISBG is not really a "load-and-test" instruction, 1558 // but sets the same condition code values, so is OK to use here. 1559 case SystemZ::RISBGN: return SystemZ::RISBG; 1560 default: return 0; 1561 } 1562 } 1563 1564 // Return true if Mask matches the regexp 0*1+0*, given that zero masks 1565 // have already been filtered out. Store the first set bit in LSB and 1566 // the number of set bits in Length if so. 1567 static bool isStringOfOnes(uint64_t Mask, unsigned &LSB, unsigned &Length) { 1568 unsigned First = findFirstSet(Mask); 1569 uint64_t Top = (Mask >> First) + 1; 1570 if ((Top & -Top) == Top) { 1571 LSB = First; 1572 Length = findFirstSet(Top); 1573 return true; 1574 } 1575 return false; 1576 } 1577 1578 bool SystemZInstrInfo::isRxSBGMask(uint64_t Mask, unsigned BitSize, 1579 unsigned &Start, unsigned &End) const { 1580 // Reject trivial all-zero masks. 1581 Mask &= allOnes(BitSize); 1582 if (Mask == 0) 1583 return false; 1584 1585 // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of 1586 // the msb and End specifies the index of the lsb. 1587 unsigned LSB, Length; 1588 if (isStringOfOnes(Mask, LSB, Length)) { 1589 Start = 63 - (LSB + Length - 1); 1590 End = 63 - LSB; 1591 return true; 1592 } 1593 1594 // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb 1595 // of the low 1s and End specifies the lsb of the high 1s. 1596 if (isStringOfOnes(Mask ^ allOnes(BitSize), LSB, Length)) { 1597 assert(LSB > 0 && "Bottom bit must be set"); 1598 assert(LSB + Length < BitSize && "Top bit must be set"); 1599 Start = 63 - (LSB - 1); 1600 End = 63 - (LSB + Length); 1601 return true; 1602 } 1603 1604 return false; 1605 } 1606 1607 unsigned SystemZInstrInfo::getFusedCompare(unsigned Opcode, 1608 SystemZII::FusedCompareType Type, 1609 const MachineInstr *MI) const { 1610 switch (Opcode) { 1611 case SystemZ::CHI: 1612 case SystemZ::CGHI: 1613 if (!(MI && isInt<8>(MI->getOperand(1).getImm()))) 1614 return 0; 1615 break; 1616 case SystemZ::CLFI: 1617 case SystemZ::CLGFI: 1618 if (!(MI && isUInt<8>(MI->getOperand(1).getImm()))) 1619 return 0; 1620 break; 1621 case SystemZ::CL: 1622 case SystemZ::CLG: 1623 if (!STI.hasMiscellaneousExtensions()) 1624 return 0; 1625 if (!(MI && MI->getOperand(3).getReg() == 0)) 1626 return 0; 1627 break; 1628 } 1629 switch (Type) { 1630 case SystemZII::CompareAndBranch: 1631 switch (Opcode) { 1632 case SystemZ::CR: 1633 return SystemZ::CRJ; 1634 case SystemZ::CGR: 1635 return SystemZ::CGRJ; 1636 case SystemZ::CHI: 1637 return SystemZ::CIJ; 1638 case SystemZ::CGHI: 1639 return SystemZ::CGIJ; 1640 case SystemZ::CLR: 1641 return SystemZ::CLRJ; 1642 case SystemZ::CLGR: 1643 return SystemZ::CLGRJ; 1644 case SystemZ::CLFI: 1645 return SystemZ::CLIJ; 1646 case SystemZ::CLGFI: 1647 return SystemZ::CLGIJ; 1648 default: 1649 return 0; 1650 } 1651 case SystemZII::CompareAndReturn: 1652 switch (Opcode) { 1653 case SystemZ::CR: 1654 return SystemZ::CRBReturn; 1655 case SystemZ::CGR: 1656 return SystemZ::CGRBReturn; 1657 case SystemZ::CHI: 1658 return SystemZ::CIBReturn; 1659 case SystemZ::CGHI: 1660 return SystemZ::CGIBReturn; 1661 case SystemZ::CLR: 1662 return SystemZ::CLRBReturn; 1663 case SystemZ::CLGR: 1664 return SystemZ::CLGRBReturn; 1665 case SystemZ::CLFI: 1666 return SystemZ::CLIBReturn; 1667 case SystemZ::CLGFI: 1668 return SystemZ::CLGIBReturn; 1669 default: 1670 return 0; 1671 } 1672 case SystemZII::CompareAndSibcall: 1673 switch (Opcode) { 1674 case SystemZ::CR: 1675 return SystemZ::CRBCall; 1676 case SystemZ::CGR: 1677 return SystemZ::CGRBCall; 1678 case SystemZ::CHI: 1679 return SystemZ::CIBCall; 1680 case SystemZ::CGHI: 1681 return SystemZ::CGIBCall; 1682 case SystemZ::CLR: 1683 return SystemZ::CLRBCall; 1684 case SystemZ::CLGR: 1685 return SystemZ::CLGRBCall; 1686 case SystemZ::CLFI: 1687 return SystemZ::CLIBCall; 1688 case SystemZ::CLGFI: 1689 return SystemZ::CLGIBCall; 1690 default: 1691 return 0; 1692 } 1693 case SystemZII::CompareAndTrap: 1694 switch (Opcode) { 1695 case SystemZ::CR: 1696 return SystemZ::CRT; 1697 case SystemZ::CGR: 1698 return SystemZ::CGRT; 1699 case SystemZ::CHI: 1700 return SystemZ::CIT; 1701 case SystemZ::CGHI: 1702 return SystemZ::CGIT; 1703 case SystemZ::CLR: 1704 return SystemZ::CLRT; 1705 case SystemZ::CLGR: 1706 return SystemZ::CLGRT; 1707 case SystemZ::CLFI: 1708 return SystemZ::CLFIT; 1709 case SystemZ::CLGFI: 1710 return SystemZ::CLGIT; 1711 case SystemZ::CL: 1712 return SystemZ::CLT; 1713 case SystemZ::CLG: 1714 return SystemZ::CLGT; 1715 default: 1716 return 0; 1717 } 1718 } 1719 return 0; 1720 } 1721 1722 unsigned SystemZInstrInfo::getLoadAndTrap(unsigned Opcode) const { 1723 if (!STI.hasLoadAndTrap()) 1724 return 0; 1725 switch (Opcode) { 1726 case SystemZ::L: 1727 case SystemZ::LY: 1728 return SystemZ::LAT; 1729 case SystemZ::LG: 1730 return SystemZ::LGAT; 1731 case SystemZ::LFH: 1732 return SystemZ::LFHAT; 1733 case SystemZ::LLGF: 1734 return SystemZ::LLGFAT; 1735 case SystemZ::LLGT: 1736 return SystemZ::LLGTAT; 1737 } 1738 return 0; 1739 } 1740 1741 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB, 1742 MachineBasicBlock::iterator MBBI, 1743 unsigned Reg, uint64_t Value) const { 1744 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 1745 unsigned Opcode; 1746 if (isInt<16>(Value)) 1747 Opcode = SystemZ::LGHI; 1748 else if (SystemZ::isImmLL(Value)) 1749 Opcode = SystemZ::LLILL; 1750 else if (SystemZ::isImmLH(Value)) { 1751 Opcode = SystemZ::LLILH; 1752 Value >>= 16; 1753 } else { 1754 assert(isInt<32>(Value) && "Huge values not handled yet"); 1755 Opcode = SystemZ::LGFI; 1756 } 1757 BuildMI(MBB, MBBI, DL, get(Opcode), Reg).addImm(Value); 1758 } 1759 1760 bool SystemZInstrInfo::verifyInstruction(const MachineInstr &MI, 1761 StringRef &ErrInfo) const { 1762 const MCInstrDesc &MCID = MI.getDesc(); 1763 for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) { 1764 if (I >= MCID.getNumOperands()) 1765 break; 1766 const MachineOperand &Op = MI.getOperand(I); 1767 const MCOperandInfo &MCOI = MCID.OpInfo[I]; 1768 // Addressing modes have register and immediate operands. Op should be a 1769 // register (or frame index) operand if MCOI.RegClass contains a valid 1770 // register class, or an immediate otherwise. 1771 if (MCOI.OperandType == MCOI::OPERAND_MEMORY && 1772 ((MCOI.RegClass != -1 && !Op.isReg() && !Op.isFI()) || 1773 (MCOI.RegClass == -1 && !Op.isImm()))) { 1774 ErrInfo = "Addressing mode operands corrupt!"; 1775 return false; 1776 } 1777 } 1778 1779 return true; 1780 } 1781 1782 bool SystemZInstrInfo:: 1783 areMemAccessesTriviallyDisjoint(const MachineInstr &MIa, 1784 const MachineInstr &MIb) const { 1785 1786 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand()) 1787 return false; 1788 1789 // If mem-operands show that the same address Value is used by both 1790 // instructions, check for non-overlapping offsets and widths. Not 1791 // sure if a register based analysis would be an improvement... 1792 1793 MachineMemOperand *MMOa = *MIa.memoperands_begin(); 1794 MachineMemOperand *MMOb = *MIb.memoperands_begin(); 1795 const Value *VALa = MMOa->getValue(); 1796 const Value *VALb = MMOb->getValue(); 1797 bool SameVal = (VALa && VALb && (VALa == VALb)); 1798 if (!SameVal) { 1799 const PseudoSourceValue *PSVa = MMOa->getPseudoValue(); 1800 const PseudoSourceValue *PSVb = MMOb->getPseudoValue(); 1801 if (PSVa && PSVb && (PSVa == PSVb)) 1802 SameVal = true; 1803 } 1804 if (SameVal) { 1805 int OffsetA = MMOa->getOffset(), OffsetB = MMOb->getOffset(); 1806 int WidthA = MMOa->getSize(), WidthB = MMOb->getSize(); 1807 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB; 1808 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA; 1809 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; 1810 if (LowOffset + LowWidth <= HighOffset) 1811 return true; 1812 } 1813 1814 return false; 1815 } 1816