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 a select pseudo instruction. Replace it with LowOpcode if source 227 // and destination are all low GR32s and HighOpcode if source and destination 228 // are all high GR32s. Otherwise, use the two-operand MixedOpcode. 229 void SystemZInstrInfo::expandSELRPseudo(MachineInstr &MI, unsigned LowOpcode, 230 unsigned HighOpcode, 231 unsigned MixedOpcode) const { 232 unsigned DestReg = MI.getOperand(0).getReg(); 233 unsigned Src1Reg = MI.getOperand(1).getReg(); 234 unsigned Src2Reg = MI.getOperand(2).getReg(); 235 bool DestIsHigh = isHighReg(DestReg); 236 bool Src1IsHigh = isHighReg(Src1Reg); 237 bool Src2IsHigh = isHighReg(Src2Reg); 238 239 // If sources and destination aren't all high or all low, we may be able to 240 // simplify the operation by moving one of the sources to the destination 241 // first. But only if this doesn't clobber the other source. 242 if (DestReg != Src1Reg && DestReg != Src2Reg) { 243 if (DestIsHigh != Src1IsHigh) { 244 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), DestReg, Src1Reg, 245 SystemZ::LR, 32, MI.getOperand(1).isKill(), 246 MI.getOperand(1).isUndef()); 247 MI.getOperand(1).setReg(DestReg); 248 Src1Reg = DestReg; 249 Src1IsHigh = DestIsHigh; 250 } else if (DestIsHigh != Src2IsHigh) { 251 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), DestReg, Src2Reg, 252 SystemZ::LR, 32, MI.getOperand(2).isKill(), 253 MI.getOperand(2).isUndef()); 254 MI.getOperand(2).setReg(DestReg); 255 Src2Reg = DestReg; 256 Src2IsHigh = DestIsHigh; 257 } 258 } 259 260 // If the destination (now) matches one source, prefer this to be first. 261 if (DestReg != Src1Reg && DestReg == Src2Reg) { 262 commuteInstruction(MI, false, 1, 2); 263 std::swap(Src1Reg, Src2Reg); 264 std::swap(Src1IsHigh, Src2IsHigh); 265 } 266 267 if (!DestIsHigh && !Src1IsHigh && !Src2IsHigh) 268 MI.setDesc(get(LowOpcode)); 269 else if (DestIsHigh && Src1IsHigh && Src2IsHigh) 270 MI.setDesc(get(HighOpcode)); 271 else { 272 // Given the simplifcation above, we must already have a two-operand case. 273 assert (DestReg == Src1Reg); 274 MI.setDesc(get(MixedOpcode)); 275 MI.tieOperands(0, 1); 276 LOCRMuxJumps++; 277 } 278 279 // If we were unable to implement the pseudo with a single instruction, we 280 // need to convert it back into a branch sequence. This cannot be done here 281 // since the caller of expandPostRAPseudo does not handle changes to the CFG 282 // correctly. This change is defered to the SystemZExpandPseudo pass. 283 } 284 285 // MI is an RR-style pseudo instruction that zero-extends the low Size bits 286 // of one GRX32 into another. Replace it with LowOpcode if both operands 287 // are low registers, otherwise use RISB[LH]G. 288 void SystemZInstrInfo::expandZExtPseudo(MachineInstr &MI, unsigned LowOpcode, 289 unsigned Size) const { 290 MachineInstrBuilder MIB = 291 emitGRX32Move(*MI.getParent(), MI, MI.getDebugLoc(), 292 MI.getOperand(0).getReg(), MI.getOperand(1).getReg(), LowOpcode, 293 Size, MI.getOperand(1).isKill(), MI.getOperand(1).isUndef()); 294 295 // Keep the remaining operands as-is. 296 for (unsigned I = 2; I < MI.getNumOperands(); ++I) 297 MIB.add(MI.getOperand(I)); 298 299 MI.eraseFromParent(); 300 } 301 302 void SystemZInstrInfo::expandLoadStackGuard(MachineInstr *MI) const { 303 MachineBasicBlock *MBB = MI->getParent(); 304 MachineFunction &MF = *MBB->getParent(); 305 const unsigned Reg64 = MI->getOperand(0).getReg(); 306 const unsigned Reg32 = RI.getSubReg(Reg64, SystemZ::subreg_l32); 307 308 // EAR can only load the low subregister so us a shift for %a0 to produce 309 // the GR containing %a0 and %a1. 310 311 // ear <reg>, %a0 312 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::EAR), Reg32) 313 .addReg(SystemZ::A0) 314 .addReg(Reg64, RegState::ImplicitDefine); 315 316 // sllg <reg>, <reg>, 32 317 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::SLLG), Reg64) 318 .addReg(Reg64) 319 .addReg(0) 320 .addImm(32); 321 322 // ear <reg>, %a1 323 BuildMI(*MBB, MI, MI->getDebugLoc(), get(SystemZ::EAR), Reg32) 324 .addReg(SystemZ::A1); 325 326 // lg <reg>, 40(<reg>) 327 MI->setDesc(get(SystemZ::LG)); 328 MachineInstrBuilder(MF, MI).addReg(Reg64).addImm(40).addReg(0); 329 } 330 331 // Emit a zero-extending move from 32-bit GPR SrcReg to 32-bit GPR 332 // DestReg before MBBI in MBB. Use LowLowOpcode when both DestReg and SrcReg 333 // are low registers, otherwise use RISB[LH]G. Size is the number of bits 334 // taken from the low end of SrcReg (8 for LLCR, 16 for LLHR and 32 for LR). 335 // KillSrc is true if this move is the last use of SrcReg. 336 MachineInstrBuilder 337 SystemZInstrInfo::emitGRX32Move(MachineBasicBlock &MBB, 338 MachineBasicBlock::iterator MBBI, 339 const DebugLoc &DL, unsigned DestReg, 340 unsigned SrcReg, unsigned LowLowOpcode, 341 unsigned Size, bool KillSrc, 342 bool UndefSrc) const { 343 unsigned Opcode; 344 bool DestIsHigh = isHighReg(DestReg); 345 bool SrcIsHigh = isHighReg(SrcReg); 346 if (DestIsHigh && SrcIsHigh) 347 Opcode = SystemZ::RISBHH; 348 else if (DestIsHigh && !SrcIsHigh) 349 Opcode = SystemZ::RISBHL; 350 else if (!DestIsHigh && SrcIsHigh) 351 Opcode = SystemZ::RISBLH; 352 else { 353 return BuildMI(MBB, MBBI, DL, get(LowLowOpcode), DestReg) 354 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)); 355 } 356 unsigned Rotate = (DestIsHigh != SrcIsHigh ? 32 : 0); 357 return BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 358 .addReg(DestReg, RegState::Undef) 359 .addReg(SrcReg, getKillRegState(KillSrc) | getUndefRegState(UndefSrc)) 360 .addImm(32 - Size).addImm(128 + 31).addImm(Rotate); 361 } 362 363 MachineInstr *SystemZInstrInfo::commuteInstructionImpl(MachineInstr &MI, 364 bool NewMI, 365 unsigned OpIdx1, 366 unsigned OpIdx2) const { 367 auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & { 368 if (NewMI) 369 return *MI.getParent()->getParent()->CloneMachineInstr(&MI); 370 return MI; 371 }; 372 373 switch (MI.getOpcode()) { 374 case SystemZ::SELRMux: 375 case SystemZ::SELFHR: 376 case SystemZ::SELR: 377 case SystemZ::SELGR: 378 case SystemZ::LOCRMux: 379 case SystemZ::LOCFHR: 380 case SystemZ::LOCR: 381 case SystemZ::LOCGR: { 382 auto &WorkingMI = cloneIfNew(MI); 383 // Invert condition. 384 unsigned CCValid = WorkingMI.getOperand(3).getImm(); 385 unsigned CCMask = WorkingMI.getOperand(4).getImm(); 386 WorkingMI.getOperand(4).setImm(CCMask ^ CCValid); 387 return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false, 388 OpIdx1, OpIdx2); 389 } 390 default: 391 return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 392 } 393 } 394 395 // If MI is a simple load or store for a frame object, return the register 396 // it loads or stores and set FrameIndex to the index of the frame object. 397 // Return 0 otherwise. 398 // 399 // Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 400 static int isSimpleMove(const MachineInstr &MI, int &FrameIndex, 401 unsigned Flag) { 402 const MCInstrDesc &MCID = MI.getDesc(); 403 if ((MCID.TSFlags & Flag) && MI.getOperand(1).isFI() && 404 MI.getOperand(2).getImm() == 0 && MI.getOperand(3).getReg() == 0) { 405 FrameIndex = MI.getOperand(1).getIndex(); 406 return MI.getOperand(0).getReg(); 407 } 408 return 0; 409 } 410 411 unsigned SystemZInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, 412 int &FrameIndex) const { 413 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXLoad); 414 } 415 416 unsigned SystemZInstrInfo::isStoreToStackSlot(const MachineInstr &MI, 417 int &FrameIndex) const { 418 return isSimpleMove(MI, FrameIndex, SystemZII::SimpleBDXStore); 419 } 420 421 bool SystemZInstrInfo::isStackSlotCopy(const MachineInstr &MI, 422 int &DestFrameIndex, 423 int &SrcFrameIndex) const { 424 // Check for MVC 0(Length,FI1),0(FI2) 425 const MachineFrameInfo &MFI = MI.getParent()->getParent()->getFrameInfo(); 426 if (MI.getOpcode() != SystemZ::MVC || !MI.getOperand(0).isFI() || 427 MI.getOperand(1).getImm() != 0 || !MI.getOperand(3).isFI() || 428 MI.getOperand(4).getImm() != 0) 429 return false; 430 431 // Check that Length covers the full slots. 432 int64_t Length = MI.getOperand(2).getImm(); 433 unsigned FI1 = MI.getOperand(0).getIndex(); 434 unsigned FI2 = MI.getOperand(3).getIndex(); 435 if (MFI.getObjectSize(FI1) != Length || 436 MFI.getObjectSize(FI2) != Length) 437 return false; 438 439 DestFrameIndex = FI1; 440 SrcFrameIndex = FI2; 441 return true; 442 } 443 444 bool SystemZInstrInfo::analyzeBranch(MachineBasicBlock &MBB, 445 MachineBasicBlock *&TBB, 446 MachineBasicBlock *&FBB, 447 SmallVectorImpl<MachineOperand> &Cond, 448 bool AllowModify) const { 449 // Most of the code and comments here are boilerplate. 450 451 // Start from the bottom of the block and work up, examining the 452 // terminator instructions. 453 MachineBasicBlock::iterator I = MBB.end(); 454 while (I != MBB.begin()) { 455 --I; 456 if (I->isDebugInstr()) 457 continue; 458 459 // Working from the bottom, when we see a non-terminator instruction, we're 460 // done. 461 if (!isUnpredicatedTerminator(*I)) 462 break; 463 464 // A terminator that isn't a branch can't easily be handled by this 465 // analysis. 466 if (!I->isBranch()) 467 return true; 468 469 // Can't handle indirect branches. 470 SystemZII::Branch Branch(getBranchInfo(*I)); 471 if (!Branch.Target->isMBB()) 472 return true; 473 474 // Punt on compound branches. 475 if (Branch.Type != SystemZII::BranchNormal) 476 return true; 477 478 if (Branch.CCMask == SystemZ::CCMASK_ANY) { 479 // Handle unconditional branches. 480 if (!AllowModify) { 481 TBB = Branch.Target->getMBB(); 482 continue; 483 } 484 485 // If the block has any instructions after a JMP, delete them. 486 while (std::next(I) != MBB.end()) 487 std::next(I)->eraseFromParent(); 488 489 Cond.clear(); 490 FBB = nullptr; 491 492 // Delete the JMP if it's equivalent to a fall-through. 493 if (MBB.isLayoutSuccessor(Branch.Target->getMBB())) { 494 TBB = nullptr; 495 I->eraseFromParent(); 496 I = MBB.end(); 497 continue; 498 } 499 500 // TBB is used to indicate the unconditinal destination. 501 TBB = Branch.Target->getMBB(); 502 continue; 503 } 504 505 // Working from the bottom, handle the first conditional branch. 506 if (Cond.empty()) { 507 // FIXME: add X86-style branch swap 508 FBB = TBB; 509 TBB = Branch.Target->getMBB(); 510 Cond.push_back(MachineOperand::CreateImm(Branch.CCValid)); 511 Cond.push_back(MachineOperand::CreateImm(Branch.CCMask)); 512 continue; 513 } 514 515 // Handle subsequent conditional branches. 516 assert(Cond.size() == 2 && TBB && "Should have seen a conditional branch"); 517 518 // Only handle the case where all conditional branches branch to the same 519 // destination. 520 if (TBB != Branch.Target->getMBB()) 521 return true; 522 523 // If the conditions are the same, we can leave them alone. 524 unsigned OldCCValid = Cond[0].getImm(); 525 unsigned OldCCMask = Cond[1].getImm(); 526 if (OldCCValid == Branch.CCValid && OldCCMask == Branch.CCMask) 527 continue; 528 529 // FIXME: Try combining conditions like X86 does. Should be easy on Z! 530 return false; 531 } 532 533 return false; 534 } 535 536 unsigned SystemZInstrInfo::removeBranch(MachineBasicBlock &MBB, 537 int *BytesRemoved) const { 538 assert(!BytesRemoved && "code size not handled"); 539 540 // Most of the code and comments here are boilerplate. 541 MachineBasicBlock::iterator I = MBB.end(); 542 unsigned Count = 0; 543 544 while (I != MBB.begin()) { 545 --I; 546 if (I->isDebugInstr()) 547 continue; 548 if (!I->isBranch()) 549 break; 550 if (!getBranchInfo(*I).Target->isMBB()) 551 break; 552 // Remove the branch. 553 I->eraseFromParent(); 554 I = MBB.end(); 555 ++Count; 556 } 557 558 return Count; 559 } 560 561 bool SystemZInstrInfo:: 562 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 563 assert(Cond.size() == 2 && "Invalid condition"); 564 Cond[1].setImm(Cond[1].getImm() ^ Cond[0].getImm()); 565 return false; 566 } 567 568 unsigned SystemZInstrInfo::insertBranch(MachineBasicBlock &MBB, 569 MachineBasicBlock *TBB, 570 MachineBasicBlock *FBB, 571 ArrayRef<MachineOperand> Cond, 572 const DebugLoc &DL, 573 int *BytesAdded) const { 574 // In this function we output 32-bit branches, which should always 575 // have enough range. They can be shortened and relaxed by later code 576 // in the pipeline, if desired. 577 578 // Shouldn't be a fall through. 579 assert(TBB && "insertBranch must not be told to insert a fallthrough"); 580 assert((Cond.size() == 2 || Cond.size() == 0) && 581 "SystemZ branch conditions have one component!"); 582 assert(!BytesAdded && "code size not handled"); 583 584 if (Cond.empty()) { 585 // Unconditional branch? 586 assert(!FBB && "Unconditional branch with multiple successors!"); 587 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(TBB); 588 return 1; 589 } 590 591 // Conditional branch. 592 unsigned Count = 0; 593 unsigned CCValid = Cond[0].getImm(); 594 unsigned CCMask = Cond[1].getImm(); 595 BuildMI(&MBB, DL, get(SystemZ::BRC)) 596 .addImm(CCValid).addImm(CCMask).addMBB(TBB); 597 ++Count; 598 599 if (FBB) { 600 // Two-way Conditional branch. Insert the second branch. 601 BuildMI(&MBB, DL, get(SystemZ::J)).addMBB(FBB); 602 ++Count; 603 } 604 return Count; 605 } 606 607 bool SystemZInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg, 608 unsigned &SrcReg2, int &Mask, 609 int &Value) const { 610 assert(MI.isCompare() && "Caller should have checked for a comparison"); 611 612 if (MI.getNumExplicitOperands() == 2 && MI.getOperand(0).isReg() && 613 MI.getOperand(1).isImm()) { 614 SrcReg = MI.getOperand(0).getReg(); 615 SrcReg2 = 0; 616 Value = MI.getOperand(1).getImm(); 617 Mask = ~0; 618 return true; 619 } 620 621 return false; 622 } 623 624 bool SystemZInstrInfo::canInsertSelect(const MachineBasicBlock &MBB, 625 ArrayRef<MachineOperand> Pred, 626 unsigned TrueReg, unsigned FalseReg, 627 int &CondCycles, int &TrueCycles, 628 int &FalseCycles) const { 629 // Not all subtargets have LOCR instructions. 630 if (!STI.hasLoadStoreOnCond()) 631 return false; 632 if (Pred.size() != 2) 633 return false; 634 635 // Check register classes. 636 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 637 const TargetRegisterClass *RC = 638 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 639 if (!RC) 640 return false; 641 642 // We have LOCR instructions for 32 and 64 bit general purpose registers. 643 if ((STI.hasLoadStoreOnCond2() && 644 SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) || 645 SystemZ::GR32BitRegClass.hasSubClassEq(RC) || 646 SystemZ::GR64BitRegClass.hasSubClassEq(RC)) { 647 CondCycles = 2; 648 TrueCycles = 2; 649 FalseCycles = 2; 650 return true; 651 } 652 653 // Can't do anything else. 654 return false; 655 } 656 657 void SystemZInstrInfo::insertSelect(MachineBasicBlock &MBB, 658 MachineBasicBlock::iterator I, 659 const DebugLoc &DL, unsigned DstReg, 660 ArrayRef<MachineOperand> Pred, 661 unsigned TrueReg, 662 unsigned FalseReg) const { 663 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 664 const TargetRegisterClass *RC = MRI.getRegClass(DstReg); 665 666 assert(Pred.size() == 2 && "Invalid condition"); 667 unsigned CCValid = Pred[0].getImm(); 668 unsigned CCMask = Pred[1].getImm(); 669 670 unsigned Opc; 671 if (SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) { 672 if (STI.hasMiscellaneousExtensions3()) 673 Opc = SystemZ::SELRMux; 674 else if (STI.hasLoadStoreOnCond2()) 675 Opc = SystemZ::LOCRMux; 676 else { 677 Opc = SystemZ::LOCR; 678 MRI.constrainRegClass(DstReg, &SystemZ::GR32BitRegClass); 679 unsigned TReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass); 680 unsigned FReg = MRI.createVirtualRegister(&SystemZ::GR32BitRegClass); 681 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), TReg).addReg(TrueReg); 682 BuildMI(MBB, I, DL, get(TargetOpcode::COPY), FReg).addReg(FalseReg); 683 TrueReg = TReg; 684 FalseReg = FReg; 685 } 686 } else if (SystemZ::GR64BitRegClass.hasSubClassEq(RC)) { 687 if (STI.hasMiscellaneousExtensions3()) 688 Opc = SystemZ::SELGR; 689 else 690 Opc = SystemZ::LOCGR; 691 } else 692 llvm_unreachable("Invalid register class"); 693 694 BuildMI(MBB, I, DL, get(Opc), DstReg) 695 .addReg(FalseReg).addReg(TrueReg) 696 .addImm(CCValid).addImm(CCMask); 697 } 698 699 bool SystemZInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, 700 unsigned Reg, 701 MachineRegisterInfo *MRI) const { 702 unsigned DefOpc = DefMI.getOpcode(); 703 if (DefOpc != SystemZ::LHIMux && DefOpc != SystemZ::LHI && 704 DefOpc != SystemZ::LGHI) 705 return false; 706 if (DefMI.getOperand(0).getReg() != Reg) 707 return false; 708 int32_t ImmVal = (int32_t)DefMI.getOperand(1).getImm(); 709 710 unsigned UseOpc = UseMI.getOpcode(); 711 unsigned NewUseOpc; 712 unsigned UseIdx; 713 int CommuteIdx = -1; 714 bool TieOps = false; 715 switch (UseOpc) { 716 case SystemZ::SELRMux: 717 TieOps = true; 718 LLVM_FALLTHROUGH; 719 case SystemZ::LOCRMux: 720 if (!STI.hasLoadStoreOnCond2()) 721 return false; 722 NewUseOpc = SystemZ::LOCHIMux; 723 if (UseMI.getOperand(2).getReg() == Reg) 724 UseIdx = 2; 725 else if (UseMI.getOperand(1).getReg() == Reg) 726 UseIdx = 2, CommuteIdx = 1; 727 else 728 return false; 729 break; 730 case SystemZ::SELGR: 731 TieOps = true; 732 LLVM_FALLTHROUGH; 733 case SystemZ::LOCGR: 734 if (!STI.hasLoadStoreOnCond2()) 735 return false; 736 NewUseOpc = SystemZ::LOCGHI; 737 if (UseMI.getOperand(2).getReg() == Reg) 738 UseIdx = 2; 739 else if (UseMI.getOperand(1).getReg() == Reg) 740 UseIdx = 2, CommuteIdx = 1; 741 else 742 return false; 743 break; 744 default: 745 return false; 746 } 747 748 if (CommuteIdx != -1) 749 if (!commuteInstruction(UseMI, false, CommuteIdx, UseIdx)) 750 return false; 751 752 bool DeleteDef = MRI->hasOneNonDBGUse(Reg); 753 UseMI.setDesc(get(NewUseOpc)); 754 if (TieOps) 755 UseMI.tieOperands(0, 1); 756 UseMI.getOperand(UseIdx).ChangeToImmediate(ImmVal); 757 if (DeleteDef) 758 DefMI.eraseFromParent(); 759 760 return true; 761 } 762 763 bool SystemZInstrInfo::isPredicable(const MachineInstr &MI) const { 764 unsigned Opcode = MI.getOpcode(); 765 if (Opcode == SystemZ::Return || 766 Opcode == SystemZ::Trap || 767 Opcode == SystemZ::CallJG || 768 Opcode == SystemZ::CallBR) 769 return true; 770 return false; 771 } 772 773 bool SystemZInstrInfo:: 774 isProfitableToIfCvt(MachineBasicBlock &MBB, 775 unsigned NumCycles, unsigned ExtraPredCycles, 776 BranchProbability Probability) const { 777 // Avoid using conditional returns at the end of a loop (since then 778 // we'd need to emit an unconditional branch to the beginning anyway, 779 // making the loop body longer). This doesn't apply for low-probability 780 // loops (eg. compare-and-swap retry), so just decide based on branch 781 // probability instead of looping structure. 782 // However, since Compare and Trap instructions cost the same as a regular 783 // Compare instruction, we should allow the if conversion to convert this 784 // into a Conditional Compare regardless of the branch probability. 785 if (MBB.getLastNonDebugInstr()->getOpcode() != SystemZ::Trap && 786 MBB.succ_empty() && Probability < BranchProbability(1, 8)) 787 return false; 788 // For now only convert single instructions. 789 return NumCycles == 1; 790 } 791 792 bool SystemZInstrInfo:: 793 isProfitableToIfCvt(MachineBasicBlock &TMBB, 794 unsigned NumCyclesT, unsigned ExtraPredCyclesT, 795 MachineBasicBlock &FMBB, 796 unsigned NumCyclesF, unsigned ExtraPredCyclesF, 797 BranchProbability Probability) const { 798 // For now avoid converting mutually-exclusive cases. 799 return false; 800 } 801 802 bool SystemZInstrInfo:: 803 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles, 804 BranchProbability Probability) const { 805 // For now only duplicate single instructions. 806 return NumCycles == 1; 807 } 808 809 bool SystemZInstrInfo::PredicateInstruction( 810 MachineInstr &MI, ArrayRef<MachineOperand> Pred) const { 811 assert(Pred.size() == 2 && "Invalid condition"); 812 unsigned CCValid = Pred[0].getImm(); 813 unsigned CCMask = Pred[1].getImm(); 814 assert(CCMask > 0 && CCMask < 15 && "Invalid predicate"); 815 unsigned Opcode = MI.getOpcode(); 816 if (Opcode == SystemZ::Trap) { 817 MI.setDesc(get(SystemZ::CondTrap)); 818 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 819 .addImm(CCValid).addImm(CCMask) 820 .addReg(SystemZ::CC, RegState::Implicit); 821 return true; 822 } 823 if (Opcode == SystemZ::Return) { 824 MI.setDesc(get(SystemZ::CondReturn)); 825 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 826 .addImm(CCValid).addImm(CCMask) 827 .addReg(SystemZ::CC, RegState::Implicit); 828 return true; 829 } 830 if (Opcode == SystemZ::CallJG) { 831 MachineOperand FirstOp = MI.getOperand(0); 832 const uint32_t *RegMask = MI.getOperand(1).getRegMask(); 833 MI.RemoveOperand(1); 834 MI.RemoveOperand(0); 835 MI.setDesc(get(SystemZ::CallBRCL)); 836 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 837 .addImm(CCValid) 838 .addImm(CCMask) 839 .add(FirstOp) 840 .addRegMask(RegMask) 841 .addReg(SystemZ::CC, RegState::Implicit); 842 return true; 843 } 844 if (Opcode == SystemZ::CallBR) { 845 const uint32_t *RegMask = MI.getOperand(0).getRegMask(); 846 MI.RemoveOperand(0); 847 MI.setDesc(get(SystemZ::CallBCR)); 848 MachineInstrBuilder(*MI.getParent()->getParent(), MI) 849 .addImm(CCValid).addImm(CCMask) 850 .addRegMask(RegMask) 851 .addReg(SystemZ::CC, RegState::Implicit); 852 return true; 853 } 854 return false; 855 } 856 857 void SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 858 MachineBasicBlock::iterator MBBI, 859 const DebugLoc &DL, unsigned DestReg, 860 unsigned SrcReg, bool KillSrc) const { 861 // Split 128-bit GPR moves into two 64-bit moves. Add implicit uses of the 862 // super register in case one of the subregs is undefined. 863 // This handles ADDR128 too. 864 if (SystemZ::GR128BitRegClass.contains(DestReg, SrcReg)) { 865 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_h64), 866 RI.getSubReg(SrcReg, SystemZ::subreg_h64), KillSrc); 867 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI)) 868 .addReg(SrcReg, RegState::Implicit); 869 copyPhysReg(MBB, MBBI, DL, RI.getSubReg(DestReg, SystemZ::subreg_l64), 870 RI.getSubReg(SrcReg, SystemZ::subreg_l64), KillSrc); 871 MachineInstrBuilder(*MBB.getParent(), std::prev(MBBI)) 872 .addReg(SrcReg, (getKillRegState(KillSrc) | RegState::Implicit)); 873 return; 874 } 875 876 if (SystemZ::GRX32BitRegClass.contains(DestReg, SrcReg)) { 877 emitGRX32Move(MBB, MBBI, DL, DestReg, SrcReg, SystemZ::LR, 32, KillSrc, 878 false); 879 return; 880 } 881 882 // Move 128-bit floating-point values between VR128 and FP128. 883 if (SystemZ::VR128BitRegClass.contains(DestReg) && 884 SystemZ::FP128BitRegClass.contains(SrcReg)) { 885 unsigned SrcRegHi = 886 RI.getMatchingSuperReg(RI.getSubReg(SrcReg, SystemZ::subreg_h64), 887 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass); 888 unsigned SrcRegLo = 889 RI.getMatchingSuperReg(RI.getSubReg(SrcReg, SystemZ::subreg_l64), 890 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass); 891 892 BuildMI(MBB, MBBI, DL, get(SystemZ::VMRHG), DestReg) 893 .addReg(SrcRegHi, getKillRegState(KillSrc)) 894 .addReg(SrcRegLo, getKillRegState(KillSrc)); 895 return; 896 } 897 if (SystemZ::FP128BitRegClass.contains(DestReg) && 898 SystemZ::VR128BitRegClass.contains(SrcReg)) { 899 unsigned DestRegHi = 900 RI.getMatchingSuperReg(RI.getSubReg(DestReg, SystemZ::subreg_h64), 901 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass); 902 unsigned DestRegLo = 903 RI.getMatchingSuperReg(RI.getSubReg(DestReg, SystemZ::subreg_l64), 904 SystemZ::subreg_h64, &SystemZ::VR128BitRegClass); 905 906 if (DestRegHi != SrcReg) 907 copyPhysReg(MBB, MBBI, DL, DestRegHi, SrcReg, false); 908 BuildMI(MBB, MBBI, DL, get(SystemZ::VREPG), DestRegLo) 909 .addReg(SrcReg, getKillRegState(KillSrc)).addImm(1); 910 return; 911 } 912 913 // Move CC value from/to a GR32. 914 if (SrcReg == SystemZ::CC) { 915 auto MIB = BuildMI(MBB, MBBI, DL, get(SystemZ::IPM), DestReg); 916 if (KillSrc) { 917 const MachineFunction *MF = MBB.getParent(); 918 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); 919 MIB->addRegisterKilled(SrcReg, TRI); 920 } 921 return; 922 } 923 if (DestReg == SystemZ::CC) { 924 BuildMI(MBB, MBBI, DL, get(SystemZ::TMLH)) 925 .addReg(SrcReg, getKillRegState(KillSrc)) 926 .addImm(3 << (SystemZ::IPM_CC - 16)); 927 return; 928 } 929 930 // Everything else needs only one instruction. 931 unsigned Opcode; 932 if (SystemZ::GR64BitRegClass.contains(DestReg, SrcReg)) 933 Opcode = SystemZ::LGR; 934 else if (SystemZ::FP32BitRegClass.contains(DestReg, SrcReg)) 935 // For z13 we prefer LDR over LER to avoid partial register dependencies. 936 Opcode = STI.hasVector() ? SystemZ::LDR32 : SystemZ::LER; 937 else if (SystemZ::FP64BitRegClass.contains(DestReg, SrcReg)) 938 Opcode = SystemZ::LDR; 939 else if (SystemZ::FP128BitRegClass.contains(DestReg, SrcReg)) 940 Opcode = SystemZ::LXR; 941 else if (SystemZ::VR32BitRegClass.contains(DestReg, SrcReg)) 942 Opcode = SystemZ::VLR32; 943 else if (SystemZ::VR64BitRegClass.contains(DestReg, SrcReg)) 944 Opcode = SystemZ::VLR64; 945 else if (SystemZ::VR128BitRegClass.contains(DestReg, SrcReg)) 946 Opcode = SystemZ::VLR; 947 else if (SystemZ::AR32BitRegClass.contains(DestReg, SrcReg)) 948 Opcode = SystemZ::CPYA; 949 else if (SystemZ::AR32BitRegClass.contains(DestReg) && 950 SystemZ::GR32BitRegClass.contains(SrcReg)) 951 Opcode = SystemZ::SAR; 952 else if (SystemZ::GR32BitRegClass.contains(DestReg) && 953 SystemZ::AR32BitRegClass.contains(SrcReg)) 954 Opcode = SystemZ::EAR; 955 else 956 llvm_unreachable("Impossible reg-to-reg copy"); 957 958 BuildMI(MBB, MBBI, DL, get(Opcode), DestReg) 959 .addReg(SrcReg, getKillRegState(KillSrc)); 960 } 961 962 void SystemZInstrInfo::storeRegToStackSlot( 963 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned SrcReg, 964 bool isKill, int FrameIdx, const TargetRegisterClass *RC, 965 const TargetRegisterInfo *TRI) const { 966 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 967 968 // Callers may expect a single instruction, so keep 128-bit moves 969 // together for now and lower them after register allocation. 970 unsigned LoadOpcode, StoreOpcode; 971 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 972 addFrameReference(BuildMI(MBB, MBBI, DL, get(StoreOpcode)) 973 .addReg(SrcReg, getKillRegState(isKill)), 974 FrameIdx); 975 } 976 977 void SystemZInstrInfo::loadRegFromStackSlot( 978 MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, unsigned DestReg, 979 int FrameIdx, const TargetRegisterClass *RC, 980 const TargetRegisterInfo *TRI) const { 981 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 982 983 // Callers may expect a single instruction, so keep 128-bit moves 984 // together for now and lower them after register allocation. 985 unsigned LoadOpcode, StoreOpcode; 986 getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); 987 addFrameReference(BuildMI(MBB, MBBI, DL, get(LoadOpcode), DestReg), 988 FrameIdx); 989 } 990 991 // Return true if MI is a simple load or store with a 12-bit displacement 992 // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. 993 static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) { 994 const MCInstrDesc &MCID = MI->getDesc(); 995 return ((MCID.TSFlags & Flag) && 996 isUInt<12>(MI->getOperand(2).getImm()) && 997 MI->getOperand(3).getReg() == 0); 998 } 999 1000 namespace { 1001 1002 struct LogicOp { 1003 LogicOp() = default; 1004 LogicOp(unsigned regSize, unsigned immLSB, unsigned immSize) 1005 : RegSize(regSize), ImmLSB(immLSB), ImmSize(immSize) {} 1006 1007 explicit operator bool() const { return RegSize; } 1008 1009 unsigned RegSize = 0; 1010 unsigned ImmLSB = 0; 1011 unsigned ImmSize = 0; 1012 }; 1013 1014 } // end anonymous namespace 1015 1016 static LogicOp interpretAndImmediate(unsigned Opcode) { 1017 switch (Opcode) { 1018 case SystemZ::NILMux: return LogicOp(32, 0, 16); 1019 case SystemZ::NIHMux: return LogicOp(32, 16, 16); 1020 case SystemZ::NILL64: return LogicOp(64, 0, 16); 1021 case SystemZ::NILH64: return LogicOp(64, 16, 16); 1022 case SystemZ::NIHL64: return LogicOp(64, 32, 16); 1023 case SystemZ::NIHH64: return LogicOp(64, 48, 16); 1024 case SystemZ::NIFMux: return LogicOp(32, 0, 32); 1025 case SystemZ::NILF64: return LogicOp(64, 0, 32); 1026 case SystemZ::NIHF64: return LogicOp(64, 32, 32); 1027 default: return LogicOp(); 1028 } 1029 } 1030 1031 static void transferDeadCC(MachineInstr *OldMI, MachineInstr *NewMI) { 1032 if (OldMI->registerDefIsDead(SystemZ::CC)) { 1033 MachineOperand *CCDef = NewMI->findRegisterDefOperand(SystemZ::CC); 1034 if (CCDef != nullptr) 1035 CCDef->setIsDead(true); 1036 } 1037 } 1038 1039 MachineInstr *SystemZInstrInfo::convertToThreeAddress( 1040 MachineFunction::iterator &MFI, MachineInstr &MI, LiveVariables *LV) const { 1041 MachineBasicBlock *MBB = MI.getParent(); 1042 1043 // Try to convert an AND into an RISBG-type instruction. 1044 // TODO: It might be beneficial to select RISBG and shorten to AND instead. 1045 if (LogicOp And = interpretAndImmediate(MI.getOpcode())) { 1046 uint64_t Imm = MI.getOperand(2).getImm() << And.ImmLSB; 1047 // AND IMMEDIATE leaves the other bits of the register unchanged. 1048 Imm |= allOnes(And.RegSize) & ~(allOnes(And.ImmSize) << And.ImmLSB); 1049 unsigned Start, End; 1050 if (isRxSBGMask(Imm, And.RegSize, Start, End)) { 1051 unsigned NewOpcode; 1052 if (And.RegSize == 64) { 1053 NewOpcode = SystemZ::RISBG; 1054 // Prefer RISBGN if available, since it does not clobber CC. 1055 if (STI.hasMiscellaneousExtensions()) 1056 NewOpcode = SystemZ::RISBGN; 1057 } else { 1058 NewOpcode = SystemZ::RISBMux; 1059 Start &= 31; 1060 End &= 31; 1061 } 1062 MachineOperand &Dest = MI.getOperand(0); 1063 MachineOperand &Src = MI.getOperand(1); 1064 MachineInstrBuilder MIB = 1065 BuildMI(*MBB, MI, MI.getDebugLoc(), get(NewOpcode)) 1066 .add(Dest) 1067 .addReg(0) 1068 .addReg(Src.getReg(), getKillRegState(Src.isKill()), 1069 Src.getSubReg()) 1070 .addImm(Start) 1071 .addImm(End + 128) 1072 .addImm(0); 1073 if (LV) { 1074 unsigned NumOps = MI.getNumOperands(); 1075 for (unsigned I = 1; I < NumOps; ++I) { 1076 MachineOperand &Op = MI.getOperand(I); 1077 if (Op.isReg() && Op.isKill()) 1078 LV->replaceKillInstruction(Op.getReg(), MI, *MIB); 1079 } 1080 } 1081 transferDeadCC(&MI, MIB); 1082 return MIB; 1083 } 1084 } 1085 return nullptr; 1086 } 1087 1088 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1089 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1090 MachineBasicBlock::iterator InsertPt, int FrameIndex, 1091 LiveIntervals *LIS, VirtRegMap *VRM) const { 1092 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 1093 const MachineFrameInfo &MFI = MF.getFrameInfo(); 1094 unsigned Size = MFI.getObjectSize(FrameIndex); 1095 unsigned Opcode = MI.getOpcode(); 1096 1097 if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { 1098 if (LIS != nullptr && (Opcode == SystemZ::LA || Opcode == SystemZ::LAY) && 1099 isInt<8>(MI.getOperand(2).getImm()) && !MI.getOperand(3).getReg()) { 1100 1101 // Check CC liveness, since new instruction introduces a dead 1102 // def of CC. 1103 MCRegUnitIterator CCUnit(SystemZ::CC, TRI); 1104 LiveRange &CCLiveRange = LIS->getRegUnit(*CCUnit); 1105 ++CCUnit; 1106 assert(!CCUnit.isValid() && "CC only has one reg unit."); 1107 SlotIndex MISlot = 1108 LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot(); 1109 if (!CCLiveRange.liveAt(MISlot)) { 1110 // LA(Y) %reg, CONST(%reg) -> AGSI %mem, CONST 1111 MachineInstr *BuiltMI = BuildMI(*InsertPt->getParent(), InsertPt, 1112 MI.getDebugLoc(), get(SystemZ::AGSI)) 1113 .addFrameIndex(FrameIndex) 1114 .addImm(0) 1115 .addImm(MI.getOperand(2).getImm()); 1116 BuiltMI->findRegisterDefOperand(SystemZ::CC)->setIsDead(true); 1117 CCLiveRange.createDeadDef(MISlot, LIS->getVNInfoAllocator()); 1118 return BuiltMI; 1119 } 1120 } 1121 return nullptr; 1122 } 1123 1124 // All other cases require a single operand. 1125 if (Ops.size() != 1) 1126 return nullptr; 1127 1128 unsigned OpNum = Ops[0]; 1129 assert(Size * 8 == 1130 TRI->getRegSizeInBits(*MF.getRegInfo() 1131 .getRegClass(MI.getOperand(OpNum).getReg())) && 1132 "Invalid size combination"); 1133 1134 if ((Opcode == SystemZ::AHI || Opcode == SystemZ::AGHI) && OpNum == 0 && 1135 isInt<8>(MI.getOperand(2).getImm())) { 1136 // A(G)HI %reg, CONST -> A(G)SI %mem, CONST 1137 Opcode = (Opcode == SystemZ::AHI ? SystemZ::ASI : SystemZ::AGSI); 1138 MachineInstr *BuiltMI = 1139 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode)) 1140 .addFrameIndex(FrameIndex) 1141 .addImm(0) 1142 .addImm(MI.getOperand(2).getImm()); 1143 transferDeadCC(&MI, BuiltMI); 1144 return BuiltMI; 1145 } 1146 1147 if ((Opcode == SystemZ::ALFI && OpNum == 0 && 1148 isInt<8>((int32_t)MI.getOperand(2).getImm())) || 1149 (Opcode == SystemZ::ALGFI && OpNum == 0 && 1150 isInt<8>((int64_t)MI.getOperand(2).getImm()))) { 1151 // AL(G)FI %reg, CONST -> AL(G)SI %mem, CONST 1152 Opcode = (Opcode == SystemZ::ALFI ? SystemZ::ALSI : SystemZ::ALGSI); 1153 MachineInstr *BuiltMI = 1154 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode)) 1155 .addFrameIndex(FrameIndex) 1156 .addImm(0) 1157 .addImm((int8_t)MI.getOperand(2).getImm()); 1158 transferDeadCC(&MI, BuiltMI); 1159 return BuiltMI; 1160 } 1161 1162 if ((Opcode == SystemZ::SLFI && OpNum == 0 && 1163 isInt<8>((int32_t)-MI.getOperand(2).getImm())) || 1164 (Opcode == SystemZ::SLGFI && OpNum == 0 && 1165 isInt<8>((int64_t)-MI.getOperand(2).getImm()))) { 1166 // SL(G)FI %reg, CONST -> AL(G)SI %mem, -CONST 1167 Opcode = (Opcode == SystemZ::SLFI ? SystemZ::ALSI : SystemZ::ALGSI); 1168 MachineInstr *BuiltMI = 1169 BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), get(Opcode)) 1170 .addFrameIndex(FrameIndex) 1171 .addImm(0) 1172 .addImm((int8_t)-MI.getOperand(2).getImm()); 1173 transferDeadCC(&MI, BuiltMI); 1174 return BuiltMI; 1175 } 1176 1177 if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) { 1178 bool Op0IsGPR = (Opcode == SystemZ::LGDR); 1179 bool Op1IsGPR = (Opcode == SystemZ::LDGR); 1180 // If we're spilling the destination of an LDGR or LGDR, store the 1181 // source register instead. 1182 if (OpNum == 0) { 1183 unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD; 1184 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1185 get(StoreOpcode)) 1186 .add(MI.getOperand(1)) 1187 .addFrameIndex(FrameIndex) 1188 .addImm(0) 1189 .addReg(0); 1190 } 1191 // If we're spilling the source of an LDGR or LGDR, load the 1192 // destination register instead. 1193 if (OpNum == 1) { 1194 unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD; 1195 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1196 get(LoadOpcode)) 1197 .add(MI.getOperand(0)) 1198 .addFrameIndex(FrameIndex) 1199 .addImm(0) 1200 .addReg(0); 1201 } 1202 } 1203 1204 // Look for cases where the source of a simple store or the destination 1205 // of a simple load is being spilled. Try to use MVC instead. 1206 // 1207 // Although MVC is in practice a fast choice in these cases, it is still 1208 // logically a bytewise copy. This means that we cannot use it if the 1209 // load or store is volatile. We also wouldn't be able to use MVC if 1210 // the two memories partially overlap, but that case cannot occur here, 1211 // because we know that one of the memories is a full frame index. 1212 // 1213 // For performance reasons, we also want to avoid using MVC if the addresses 1214 // might be equal. We don't worry about that case here, because spill slot 1215 // coloring happens later, and because we have special code to remove 1216 // MVCs that turn out to be redundant. 1217 if (OpNum == 0 && MI.hasOneMemOperand()) { 1218 MachineMemOperand *MMO = *MI.memoperands_begin(); 1219 if (MMO->getSize() == Size && !MMO->isVolatile() && !MMO->isAtomic()) { 1220 // Handle conversion of loads. 1221 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXLoad)) { 1222 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1223 get(SystemZ::MVC)) 1224 .addFrameIndex(FrameIndex) 1225 .addImm(0) 1226 .addImm(Size) 1227 .add(MI.getOperand(1)) 1228 .addImm(MI.getOperand(2).getImm()) 1229 .addMemOperand(MMO); 1230 } 1231 // Handle conversion of stores. 1232 if (isSimpleBD12Move(&MI, SystemZII::SimpleBDXStore)) { 1233 return BuildMI(*InsertPt->getParent(), InsertPt, MI.getDebugLoc(), 1234 get(SystemZ::MVC)) 1235 .add(MI.getOperand(1)) 1236 .addImm(MI.getOperand(2).getImm()) 1237 .addImm(Size) 1238 .addFrameIndex(FrameIndex) 1239 .addImm(0) 1240 .addMemOperand(MMO); 1241 } 1242 } 1243 } 1244 1245 // If the spilled operand is the final one or the instruction is 1246 // commutable, try to change <INSN>R into <INSN>. 1247 unsigned NumOps = MI.getNumExplicitOperands(); 1248 int MemOpcode = SystemZ::getMemOpcode(Opcode); 1249 1250 // See if this is a 3-address instruction that is convertible to 2-address 1251 // and suitable for folding below. Only try this with virtual registers 1252 // and a provided VRM (during regalloc). 1253 bool NeedsCommute = false; 1254 if (SystemZ::getTwoOperandOpcode(Opcode) != -1 && MemOpcode != -1) { 1255 if (VRM == nullptr) 1256 MemOpcode = -1; 1257 else { 1258 assert(NumOps == 3 && "Expected two source registers."); 1259 Register DstReg = MI.getOperand(0).getReg(); 1260 Register DstPhys = 1261 (Register::isVirtualRegister(DstReg) ? VRM->getPhys(DstReg) : DstReg); 1262 Register SrcReg = (OpNum == 2 ? MI.getOperand(1).getReg() 1263 : ((OpNum == 1 && MI.isCommutable()) 1264 ? MI.getOperand(2).getReg() 1265 : Register())); 1266 if (DstPhys && !SystemZ::GRH32BitRegClass.contains(DstPhys) && SrcReg && 1267 Register::isVirtualRegister(SrcReg) && 1268 DstPhys == VRM->getPhys(SrcReg)) 1269 NeedsCommute = (OpNum == 1); 1270 else 1271 MemOpcode = -1; 1272 } 1273 } 1274 1275 if (MemOpcode >= 0) { 1276 if ((OpNum == NumOps - 1) || NeedsCommute) { 1277 const MCInstrDesc &MemDesc = get(MemOpcode); 1278 uint64_t AccessBytes = SystemZII::getAccessSize(MemDesc.TSFlags); 1279 assert(AccessBytes != 0 && "Size of access should be known"); 1280 assert(AccessBytes <= Size && "Access outside the frame index"); 1281 uint64_t Offset = Size - AccessBytes; 1282 MachineInstrBuilder MIB = BuildMI(*InsertPt->getParent(), InsertPt, 1283 MI.getDebugLoc(), get(MemOpcode)); 1284 MIB.add(MI.getOperand(0)); 1285 if (NeedsCommute) 1286 MIB.add(MI.getOperand(2)); 1287 else 1288 for (unsigned I = 1; I < OpNum; ++I) 1289 MIB.add(MI.getOperand(I)); 1290 MIB.addFrameIndex(FrameIndex).addImm(Offset); 1291 if (MemDesc.TSFlags & SystemZII::HasIndex) 1292 MIB.addReg(0); 1293 transferDeadCC(&MI, MIB); 1294 return MIB; 1295 } 1296 } 1297 1298 return nullptr; 1299 } 1300 1301 MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( 1302 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1303 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, 1304 LiveIntervals *LIS) const { 1305 return nullptr; 1306 } 1307 1308 bool SystemZInstrInfo::expandPostRAPseudo(MachineInstr &MI) const { 1309 switch (MI.getOpcode()) { 1310 case SystemZ::L128: 1311 splitMove(MI, SystemZ::LG); 1312 return true; 1313 1314 case SystemZ::ST128: 1315 splitMove(MI, SystemZ::STG); 1316 return true; 1317 1318 case SystemZ::LX: 1319 splitMove(MI, SystemZ::LD); 1320 return true; 1321 1322 case SystemZ::STX: 1323 splitMove(MI, SystemZ::STD); 1324 return true; 1325 1326 case SystemZ::LBMux: 1327 expandRXYPseudo(MI, SystemZ::LB, SystemZ::LBH); 1328 return true; 1329 1330 case SystemZ::LHMux: 1331 expandRXYPseudo(MI, SystemZ::LH, SystemZ::LHH); 1332 return true; 1333 1334 case SystemZ::LLCRMux: 1335 expandZExtPseudo(MI, SystemZ::LLCR, 8); 1336 return true; 1337 1338 case SystemZ::LLHRMux: 1339 expandZExtPseudo(MI, SystemZ::LLHR, 16); 1340 return true; 1341 1342 case SystemZ::LLCMux: 1343 expandRXYPseudo(MI, SystemZ::LLC, SystemZ::LLCH); 1344 return true; 1345 1346 case SystemZ::LLHMux: 1347 expandRXYPseudo(MI, SystemZ::LLH, SystemZ::LLHH); 1348 return true; 1349 1350 case SystemZ::LMux: 1351 expandRXYPseudo(MI, SystemZ::L, SystemZ::LFH); 1352 return true; 1353 1354 case SystemZ::LOCMux: 1355 expandLOCPseudo(MI, SystemZ::LOC, SystemZ::LOCFH); 1356 return true; 1357 1358 case SystemZ::LOCHIMux: 1359 expandLOCPseudo(MI, SystemZ::LOCHI, SystemZ::LOCHHI); 1360 return true; 1361 1362 case SystemZ::LOCRMux: 1363 expandLOCRPseudo(MI, SystemZ::LOCR, SystemZ::LOCFHR); 1364 return true; 1365 1366 case SystemZ::SELRMux: 1367 expandSELRPseudo(MI, SystemZ::SELR, SystemZ::SELFHR, 1368 SystemZ::LOCRMux); 1369 return true; 1370 1371 case SystemZ::STCMux: 1372 expandRXYPseudo(MI, SystemZ::STC, SystemZ::STCH); 1373 return true; 1374 1375 case SystemZ::STHMux: 1376 expandRXYPseudo(MI, SystemZ::STH, SystemZ::STHH); 1377 return true; 1378 1379 case SystemZ::STMux: 1380 expandRXYPseudo(MI, SystemZ::ST, SystemZ::STFH); 1381 return true; 1382 1383 case SystemZ::STOCMux: 1384 expandLOCPseudo(MI, SystemZ::STOC, SystemZ::STOCFH); 1385 return true; 1386 1387 case SystemZ::LHIMux: 1388 expandRIPseudo(MI, SystemZ::LHI, SystemZ::IIHF, true); 1389 return true; 1390 1391 case SystemZ::IIFMux: 1392 expandRIPseudo(MI, SystemZ::IILF, SystemZ::IIHF, false); 1393 return true; 1394 1395 case SystemZ::IILMux: 1396 expandRIPseudo(MI, SystemZ::IILL, SystemZ::IIHL, false); 1397 return true; 1398 1399 case SystemZ::IIHMux: 1400 expandRIPseudo(MI, SystemZ::IILH, SystemZ::IIHH, false); 1401 return true; 1402 1403 case SystemZ::NIFMux: 1404 expandRIPseudo(MI, SystemZ::NILF, SystemZ::NIHF, false); 1405 return true; 1406 1407 case SystemZ::NILMux: 1408 expandRIPseudo(MI, SystemZ::NILL, SystemZ::NIHL, false); 1409 return true; 1410 1411 case SystemZ::NIHMux: 1412 expandRIPseudo(MI, SystemZ::NILH, SystemZ::NIHH, false); 1413 return true; 1414 1415 case SystemZ::OIFMux: 1416 expandRIPseudo(MI, SystemZ::OILF, SystemZ::OIHF, false); 1417 return true; 1418 1419 case SystemZ::OILMux: 1420 expandRIPseudo(MI, SystemZ::OILL, SystemZ::OIHL, false); 1421 return true; 1422 1423 case SystemZ::OIHMux: 1424 expandRIPseudo(MI, SystemZ::OILH, SystemZ::OIHH, false); 1425 return true; 1426 1427 case SystemZ::XIFMux: 1428 expandRIPseudo(MI, SystemZ::XILF, SystemZ::XIHF, false); 1429 return true; 1430 1431 case SystemZ::TMLMux: 1432 expandRIPseudo(MI, SystemZ::TMLL, SystemZ::TMHL, false); 1433 return true; 1434 1435 case SystemZ::TMHMux: 1436 expandRIPseudo(MI, SystemZ::TMLH, SystemZ::TMHH, false); 1437 return true; 1438 1439 case SystemZ::AHIMux: 1440 expandRIPseudo(MI, SystemZ::AHI, SystemZ::AIH, false); 1441 return true; 1442 1443 case SystemZ::AHIMuxK: 1444 expandRIEPseudo(MI, SystemZ::AHI, SystemZ::AHIK, SystemZ::AIH); 1445 return true; 1446 1447 case SystemZ::AFIMux: 1448 expandRIPseudo(MI, SystemZ::AFI, SystemZ::AIH, false); 1449 return true; 1450 1451 case SystemZ::CHIMux: 1452 expandRIPseudo(MI, SystemZ::CHI, SystemZ::CIH, false); 1453 return true; 1454 1455 case SystemZ::CFIMux: 1456 expandRIPseudo(MI, SystemZ::CFI, SystemZ::CIH, false); 1457 return true; 1458 1459 case SystemZ::CLFIMux: 1460 expandRIPseudo(MI, SystemZ::CLFI, SystemZ::CLIH, false); 1461 return true; 1462 1463 case SystemZ::CMux: 1464 expandRXYPseudo(MI, SystemZ::C, SystemZ::CHF); 1465 return true; 1466 1467 case SystemZ::CLMux: 1468 expandRXYPseudo(MI, SystemZ::CL, SystemZ::CLHF); 1469 return true; 1470 1471 case SystemZ::RISBMux: { 1472 bool DestIsHigh = isHighReg(MI.getOperand(0).getReg()); 1473 bool SrcIsHigh = isHighReg(MI.getOperand(2).getReg()); 1474 if (SrcIsHigh == DestIsHigh) 1475 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHH : SystemZ::RISBLL)); 1476 else { 1477 MI.setDesc(get(DestIsHigh ? SystemZ::RISBHL : SystemZ::RISBLH)); 1478 MI.getOperand(5).setImm(MI.getOperand(5).getImm() ^ 32); 1479 } 1480 return true; 1481 } 1482 1483 case SystemZ::ADJDYNALLOC: 1484 splitAdjDynAlloc(MI); 1485 return true; 1486 1487 case TargetOpcode::LOAD_STACK_GUARD: 1488 expandLoadStackGuard(&MI); 1489 return true; 1490 1491 default: 1492 return false; 1493 } 1494 } 1495 1496 unsigned SystemZInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { 1497 if (MI.isInlineAsm()) { 1498 const MachineFunction *MF = MI.getParent()->getParent(); 1499 const char *AsmStr = MI.getOperand(0).getSymbolName(); 1500 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo()); 1501 } 1502 return MI.getDesc().getSize(); 1503 } 1504 1505 SystemZII::Branch 1506 SystemZInstrInfo::getBranchInfo(const MachineInstr &MI) const { 1507 switch (MI.getOpcode()) { 1508 case SystemZ::BR: 1509 case SystemZ::BI: 1510 case SystemZ::J: 1511 case SystemZ::JG: 1512 return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY, 1513 SystemZ::CCMASK_ANY, &MI.getOperand(0)); 1514 1515 case SystemZ::BRC: 1516 case SystemZ::BRCL: 1517 return SystemZII::Branch(SystemZII::BranchNormal, MI.getOperand(0).getImm(), 1518 MI.getOperand(1).getImm(), &MI.getOperand(2)); 1519 1520 case SystemZ::BRCT: 1521 case SystemZ::BRCTH: 1522 return SystemZII::Branch(SystemZII::BranchCT, SystemZ::CCMASK_ICMP, 1523 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1524 1525 case SystemZ::BRCTG: 1526 return SystemZII::Branch(SystemZII::BranchCTG, SystemZ::CCMASK_ICMP, 1527 SystemZ::CCMASK_CMP_NE, &MI.getOperand(2)); 1528 1529 case SystemZ::CIJ: 1530 case SystemZ::CRJ: 1531 return SystemZII::Branch(SystemZII::BranchC, SystemZ::CCMASK_ICMP, 1532 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1533 1534 case SystemZ::CLIJ: 1535 case SystemZ::CLRJ: 1536 return SystemZII::Branch(SystemZII::BranchCL, SystemZ::CCMASK_ICMP, 1537 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1538 1539 case SystemZ::CGIJ: 1540 case SystemZ::CGRJ: 1541 return SystemZII::Branch(SystemZII::BranchCG, SystemZ::CCMASK_ICMP, 1542 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1543 1544 case SystemZ::CLGIJ: 1545 case SystemZ::CLGRJ: 1546 return SystemZII::Branch(SystemZII::BranchCLG, SystemZ::CCMASK_ICMP, 1547 MI.getOperand(2).getImm(), &MI.getOperand(3)); 1548 1549 default: 1550 llvm_unreachable("Unrecognized branch opcode"); 1551 } 1552 } 1553 1554 void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC, 1555 unsigned &LoadOpcode, 1556 unsigned &StoreOpcode) const { 1557 if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) { 1558 LoadOpcode = SystemZ::L; 1559 StoreOpcode = SystemZ::ST; 1560 } else if (RC == &SystemZ::GRH32BitRegClass) { 1561 LoadOpcode = SystemZ::LFH; 1562 StoreOpcode = SystemZ::STFH; 1563 } else if (RC == &SystemZ::GRX32BitRegClass) { 1564 LoadOpcode = SystemZ::LMux; 1565 StoreOpcode = SystemZ::STMux; 1566 } else if (RC == &SystemZ::GR64BitRegClass || 1567 RC == &SystemZ::ADDR64BitRegClass) { 1568 LoadOpcode = SystemZ::LG; 1569 StoreOpcode = SystemZ::STG; 1570 } else if (RC == &SystemZ::GR128BitRegClass || 1571 RC == &SystemZ::ADDR128BitRegClass) { 1572 LoadOpcode = SystemZ::L128; 1573 StoreOpcode = SystemZ::ST128; 1574 } else if (RC == &SystemZ::FP32BitRegClass) { 1575 LoadOpcode = SystemZ::LE; 1576 StoreOpcode = SystemZ::STE; 1577 } else if (RC == &SystemZ::FP64BitRegClass) { 1578 LoadOpcode = SystemZ::LD; 1579 StoreOpcode = SystemZ::STD; 1580 } else if (RC == &SystemZ::FP128BitRegClass) { 1581 LoadOpcode = SystemZ::LX; 1582 StoreOpcode = SystemZ::STX; 1583 } else if (RC == &SystemZ::VR32BitRegClass) { 1584 LoadOpcode = SystemZ::VL32; 1585 StoreOpcode = SystemZ::VST32; 1586 } else if (RC == &SystemZ::VR64BitRegClass) { 1587 LoadOpcode = SystemZ::VL64; 1588 StoreOpcode = SystemZ::VST64; 1589 } else if (RC == &SystemZ::VF128BitRegClass || 1590 RC == &SystemZ::VR128BitRegClass) { 1591 LoadOpcode = SystemZ::VL; 1592 StoreOpcode = SystemZ::VST; 1593 } else 1594 llvm_unreachable("Unsupported regclass to load or store"); 1595 } 1596 1597 unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode, 1598 int64_t Offset) const { 1599 const MCInstrDesc &MCID = get(Opcode); 1600 int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset); 1601 if (isUInt<12>(Offset) && isUInt<12>(Offset2)) { 1602 // Get the instruction to use for unsigned 12-bit displacements. 1603 int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode); 1604 if (Disp12Opcode >= 0) 1605 return Disp12Opcode; 1606 1607 // All address-related instructions can use unsigned 12-bit 1608 // displacements. 1609 return Opcode; 1610 } 1611 if (isInt<20>(Offset) && isInt<20>(Offset2)) { 1612 // Get the instruction to use for signed 20-bit displacements. 1613 int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode); 1614 if (Disp20Opcode >= 0) 1615 return Disp20Opcode; 1616 1617 // Check whether Opcode allows signed 20-bit displacements. 1618 if (MCID.TSFlags & SystemZII::Has20BitOffset) 1619 return Opcode; 1620 } 1621 return 0; 1622 } 1623 1624 unsigned SystemZInstrInfo::getLoadAndTest(unsigned Opcode) const { 1625 switch (Opcode) { 1626 case SystemZ::L: return SystemZ::LT; 1627 case SystemZ::LY: return SystemZ::LT; 1628 case SystemZ::LG: return SystemZ::LTG; 1629 case SystemZ::LGF: return SystemZ::LTGF; 1630 case SystemZ::LR: return SystemZ::LTR; 1631 case SystemZ::LGFR: return SystemZ::LTGFR; 1632 case SystemZ::LGR: return SystemZ::LTGR; 1633 case SystemZ::LER: return SystemZ::LTEBR; 1634 case SystemZ::LDR: return SystemZ::LTDBR; 1635 case SystemZ::LXR: return SystemZ::LTXBR; 1636 case SystemZ::LCDFR: return SystemZ::LCDBR; 1637 case SystemZ::LPDFR: return SystemZ::LPDBR; 1638 case SystemZ::LNDFR: return SystemZ::LNDBR; 1639 case SystemZ::LCDFR_32: return SystemZ::LCEBR; 1640 case SystemZ::LPDFR_32: return SystemZ::LPEBR; 1641 case SystemZ::LNDFR_32: return SystemZ::LNEBR; 1642 // On zEC12 we prefer to use RISBGN. But if there is a chance to 1643 // actually use the condition code, we may turn it back into RISGB. 1644 // Note that RISBG is not really a "load-and-test" instruction, 1645 // but sets the same condition code values, so is OK to use here. 1646 case SystemZ::RISBGN: return SystemZ::RISBG; 1647 default: return 0; 1648 } 1649 } 1650 1651 // Return true if Mask matches the regexp 0*1+0*, given that zero masks 1652 // have already been filtered out. Store the first set bit in LSB and 1653 // the number of set bits in Length if so. 1654 static bool isStringOfOnes(uint64_t Mask, unsigned &LSB, unsigned &Length) { 1655 unsigned First = findFirstSet(Mask); 1656 uint64_t Top = (Mask >> First) + 1; 1657 if ((Top & -Top) == Top) { 1658 LSB = First; 1659 Length = findFirstSet(Top); 1660 return true; 1661 } 1662 return false; 1663 } 1664 1665 bool SystemZInstrInfo::isRxSBGMask(uint64_t Mask, unsigned BitSize, 1666 unsigned &Start, unsigned &End) const { 1667 // Reject trivial all-zero masks. 1668 Mask &= allOnes(BitSize); 1669 if (Mask == 0) 1670 return false; 1671 1672 // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of 1673 // the msb and End specifies the index of the lsb. 1674 unsigned LSB, Length; 1675 if (isStringOfOnes(Mask, LSB, Length)) { 1676 Start = 63 - (LSB + Length - 1); 1677 End = 63 - LSB; 1678 return true; 1679 } 1680 1681 // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb 1682 // of the low 1s and End specifies the lsb of the high 1s. 1683 if (isStringOfOnes(Mask ^ allOnes(BitSize), LSB, Length)) { 1684 assert(LSB > 0 && "Bottom bit must be set"); 1685 assert(LSB + Length < BitSize && "Top bit must be set"); 1686 Start = 63 - (LSB - 1); 1687 End = 63 - (LSB + Length); 1688 return true; 1689 } 1690 1691 return false; 1692 } 1693 1694 unsigned SystemZInstrInfo::getFusedCompare(unsigned Opcode, 1695 SystemZII::FusedCompareType Type, 1696 const MachineInstr *MI) const { 1697 switch (Opcode) { 1698 case SystemZ::CHI: 1699 case SystemZ::CGHI: 1700 if (!(MI && isInt<8>(MI->getOperand(1).getImm()))) 1701 return 0; 1702 break; 1703 case SystemZ::CLFI: 1704 case SystemZ::CLGFI: 1705 if (!(MI && isUInt<8>(MI->getOperand(1).getImm()))) 1706 return 0; 1707 break; 1708 case SystemZ::CL: 1709 case SystemZ::CLG: 1710 if (!STI.hasMiscellaneousExtensions()) 1711 return 0; 1712 if (!(MI && MI->getOperand(3).getReg() == 0)) 1713 return 0; 1714 break; 1715 } 1716 switch (Type) { 1717 case SystemZII::CompareAndBranch: 1718 switch (Opcode) { 1719 case SystemZ::CR: 1720 return SystemZ::CRJ; 1721 case SystemZ::CGR: 1722 return SystemZ::CGRJ; 1723 case SystemZ::CHI: 1724 return SystemZ::CIJ; 1725 case SystemZ::CGHI: 1726 return SystemZ::CGIJ; 1727 case SystemZ::CLR: 1728 return SystemZ::CLRJ; 1729 case SystemZ::CLGR: 1730 return SystemZ::CLGRJ; 1731 case SystemZ::CLFI: 1732 return SystemZ::CLIJ; 1733 case SystemZ::CLGFI: 1734 return SystemZ::CLGIJ; 1735 default: 1736 return 0; 1737 } 1738 case SystemZII::CompareAndReturn: 1739 switch (Opcode) { 1740 case SystemZ::CR: 1741 return SystemZ::CRBReturn; 1742 case SystemZ::CGR: 1743 return SystemZ::CGRBReturn; 1744 case SystemZ::CHI: 1745 return SystemZ::CIBReturn; 1746 case SystemZ::CGHI: 1747 return SystemZ::CGIBReturn; 1748 case SystemZ::CLR: 1749 return SystemZ::CLRBReturn; 1750 case SystemZ::CLGR: 1751 return SystemZ::CLGRBReturn; 1752 case SystemZ::CLFI: 1753 return SystemZ::CLIBReturn; 1754 case SystemZ::CLGFI: 1755 return SystemZ::CLGIBReturn; 1756 default: 1757 return 0; 1758 } 1759 case SystemZII::CompareAndSibcall: 1760 switch (Opcode) { 1761 case SystemZ::CR: 1762 return SystemZ::CRBCall; 1763 case SystemZ::CGR: 1764 return SystemZ::CGRBCall; 1765 case SystemZ::CHI: 1766 return SystemZ::CIBCall; 1767 case SystemZ::CGHI: 1768 return SystemZ::CGIBCall; 1769 case SystemZ::CLR: 1770 return SystemZ::CLRBCall; 1771 case SystemZ::CLGR: 1772 return SystemZ::CLGRBCall; 1773 case SystemZ::CLFI: 1774 return SystemZ::CLIBCall; 1775 case SystemZ::CLGFI: 1776 return SystemZ::CLGIBCall; 1777 default: 1778 return 0; 1779 } 1780 case SystemZII::CompareAndTrap: 1781 switch (Opcode) { 1782 case SystemZ::CR: 1783 return SystemZ::CRT; 1784 case SystemZ::CGR: 1785 return SystemZ::CGRT; 1786 case SystemZ::CHI: 1787 return SystemZ::CIT; 1788 case SystemZ::CGHI: 1789 return SystemZ::CGIT; 1790 case SystemZ::CLR: 1791 return SystemZ::CLRT; 1792 case SystemZ::CLGR: 1793 return SystemZ::CLGRT; 1794 case SystemZ::CLFI: 1795 return SystemZ::CLFIT; 1796 case SystemZ::CLGFI: 1797 return SystemZ::CLGIT; 1798 case SystemZ::CL: 1799 return SystemZ::CLT; 1800 case SystemZ::CLG: 1801 return SystemZ::CLGT; 1802 default: 1803 return 0; 1804 } 1805 } 1806 return 0; 1807 } 1808 1809 unsigned SystemZInstrInfo::getLoadAndTrap(unsigned Opcode) const { 1810 if (!STI.hasLoadAndTrap()) 1811 return 0; 1812 switch (Opcode) { 1813 case SystemZ::L: 1814 case SystemZ::LY: 1815 return SystemZ::LAT; 1816 case SystemZ::LG: 1817 return SystemZ::LGAT; 1818 case SystemZ::LFH: 1819 return SystemZ::LFHAT; 1820 case SystemZ::LLGF: 1821 return SystemZ::LLGFAT; 1822 case SystemZ::LLGT: 1823 return SystemZ::LLGTAT; 1824 } 1825 return 0; 1826 } 1827 1828 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB, 1829 MachineBasicBlock::iterator MBBI, 1830 unsigned Reg, uint64_t Value) const { 1831 DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); 1832 unsigned Opcode; 1833 if (isInt<16>(Value)) 1834 Opcode = SystemZ::LGHI; 1835 else if (SystemZ::isImmLL(Value)) 1836 Opcode = SystemZ::LLILL; 1837 else if (SystemZ::isImmLH(Value)) { 1838 Opcode = SystemZ::LLILH; 1839 Value >>= 16; 1840 } else { 1841 assert(isInt<32>(Value) && "Huge values not handled yet"); 1842 Opcode = SystemZ::LGFI; 1843 } 1844 BuildMI(MBB, MBBI, DL, get(Opcode), Reg).addImm(Value); 1845 } 1846 1847 bool SystemZInstrInfo:: 1848 areMemAccessesTriviallyDisjoint(const MachineInstr &MIa, 1849 const MachineInstr &MIb, 1850 AliasAnalysis *AA) const { 1851 1852 if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand()) 1853 return false; 1854 1855 // If mem-operands show that the same address Value is used by both 1856 // instructions, check for non-overlapping offsets and widths. Not 1857 // sure if a register based analysis would be an improvement... 1858 1859 MachineMemOperand *MMOa = *MIa.memoperands_begin(); 1860 MachineMemOperand *MMOb = *MIb.memoperands_begin(); 1861 const Value *VALa = MMOa->getValue(); 1862 const Value *VALb = MMOb->getValue(); 1863 bool SameVal = (VALa && VALb && (VALa == VALb)); 1864 if (!SameVal) { 1865 const PseudoSourceValue *PSVa = MMOa->getPseudoValue(); 1866 const PseudoSourceValue *PSVb = MMOb->getPseudoValue(); 1867 if (PSVa && PSVb && (PSVa == PSVb)) 1868 SameVal = true; 1869 } 1870 if (SameVal) { 1871 int OffsetA = MMOa->getOffset(), OffsetB = MMOb->getOffset(); 1872 int WidthA = MMOa->getSize(), WidthB = MMOb->getSize(); 1873 int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB; 1874 int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA; 1875 int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; 1876 if (LowOffset + LowWidth <= HighOffset) 1877 return true; 1878 } 1879 1880 return false; 1881 } 1882