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