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