1 //===-- PPCInstrInfo.cpp - PowerPC Instruction Information ----------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains the PowerPC implementation of the TargetInstrInfo class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "PPCInstrInfo.h" 15 #include "MCTargetDesc/PPCPredicates.h" 16 #include "PPC.h" 17 #include "PPCHazardRecognizers.h" 18 #include "PPCInstrBuilder.h" 19 #include "PPCMachineFunctionInfo.h" 20 #include "PPCTargetMachine.h" 21 #include "llvm/ADT/STLExtras.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/CodeGen/LiveIntervalAnalysis.h" 24 #include "llvm/CodeGen/MachineFrameInfo.h" 25 #include "llvm/CodeGen/MachineFunctionPass.h" 26 #include "llvm/CodeGen/MachineInstrBuilder.h" 27 #include "llvm/CodeGen/MachineMemOperand.h" 28 #include "llvm/CodeGen/MachineRegisterInfo.h" 29 #include "llvm/CodeGen/PseudoSourceValue.h" 30 #include "llvm/CodeGen/ScheduleDAG.h" 31 #include "llvm/CodeGen/SlotIndexes.h" 32 #include "llvm/CodeGen/StackMaps.h" 33 #include "llvm/MC/MCAsmInfo.h" 34 #include "llvm/MC/MCInst.h" 35 #include "llvm/Support/CommandLine.h" 36 #include "llvm/Support/Debug.h" 37 #include "llvm/Support/ErrorHandling.h" 38 #include "llvm/Support/TargetRegistry.h" 39 #include "llvm/Support/raw_ostream.h" 40 41 using namespace llvm; 42 43 #define DEBUG_TYPE "ppc-instr-info" 44 45 #define GET_INSTRMAP_INFO 46 #define GET_INSTRINFO_CTOR_DTOR 47 #include "PPCGenInstrInfo.inc" 48 49 static cl:: 50 opt<bool> DisableCTRLoopAnal("disable-ppc-ctrloop-analysis", cl::Hidden, 51 cl::desc("Disable analysis for CTR loops")); 52 53 static cl::opt<bool> DisableCmpOpt("disable-ppc-cmp-opt", 54 cl::desc("Disable compare instruction optimization"), cl::Hidden); 55 56 static cl::opt<bool> VSXSelfCopyCrash("crash-on-ppc-vsx-self-copy", 57 cl::desc("Causes the backend to crash instead of generating a nop VSX copy"), 58 cl::Hidden); 59 60 // Pin the vtable to this file. 61 void PPCInstrInfo::anchor() {} 62 63 PPCInstrInfo::PPCInstrInfo(PPCSubtarget &STI) 64 : PPCGenInstrInfo(PPC::ADJCALLSTACKDOWN, PPC::ADJCALLSTACKUP), 65 Subtarget(STI), RI(STI.getTargetMachine()) {} 66 67 /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for 68 /// this target when scheduling the DAG. 69 ScheduleHazardRecognizer * 70 PPCInstrInfo::CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, 71 const ScheduleDAG *DAG) const { 72 unsigned Directive = 73 static_cast<const PPCSubtarget *>(STI)->getDarwinDirective(); 74 if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2 || 75 Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500) { 76 const InstrItineraryData *II = 77 static_cast<const PPCSubtarget *>(STI)->getInstrItineraryData(); 78 return new ScoreboardHazardRecognizer(II, DAG); 79 } 80 81 return TargetInstrInfo::CreateTargetHazardRecognizer(STI, DAG); 82 } 83 84 /// CreateTargetPostRAHazardRecognizer - Return the postRA hazard recognizer 85 /// to use for this target when scheduling the DAG. 86 ScheduleHazardRecognizer * 87 PPCInstrInfo::CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, 88 const ScheduleDAG *DAG) const { 89 unsigned Directive = 90 DAG->MF.getSubtarget<PPCSubtarget>().getDarwinDirective(); 91 92 if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8) 93 return new PPCDispatchGroupSBHazardRecognizer(II, DAG); 94 95 // Most subtargets use a PPC970 recognizer. 96 if (Directive != PPC::DIR_440 && Directive != PPC::DIR_A2 && 97 Directive != PPC::DIR_E500mc && Directive != PPC::DIR_E5500) { 98 assert(DAG->TII && "No InstrInfo?"); 99 100 return new PPCHazardRecognizer970(*DAG); 101 } 102 103 return new ScoreboardHazardRecognizer(II, DAG); 104 } 105 106 107 int PPCInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 108 const MachineInstr *DefMI, unsigned DefIdx, 109 const MachineInstr *UseMI, 110 unsigned UseIdx) const { 111 int Latency = PPCGenInstrInfo::getOperandLatency(ItinData, DefMI, DefIdx, 112 UseMI, UseIdx); 113 114 const MachineOperand &DefMO = DefMI->getOperand(DefIdx); 115 unsigned Reg = DefMO.getReg(); 116 117 bool IsRegCR; 118 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 119 const MachineRegisterInfo *MRI = 120 &DefMI->getParent()->getParent()->getRegInfo(); 121 IsRegCR = MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRRCRegClass) || 122 MRI->getRegClass(Reg)->hasSuperClassEq(&PPC::CRBITRCRegClass); 123 } else { 124 IsRegCR = PPC::CRRCRegClass.contains(Reg) || 125 PPC::CRBITRCRegClass.contains(Reg); 126 } 127 128 if (UseMI->isBranch() && IsRegCR) { 129 if (Latency < 0) 130 Latency = getInstrLatency(ItinData, DefMI); 131 132 // On some cores, there is an additional delay between writing to a condition 133 // register, and using it from a branch. 134 unsigned Directive = Subtarget.getDarwinDirective(); 135 switch (Directive) { 136 default: break; 137 case PPC::DIR_7400: 138 case PPC::DIR_750: 139 case PPC::DIR_970: 140 case PPC::DIR_E5500: 141 case PPC::DIR_PWR4: 142 case PPC::DIR_PWR5: 143 case PPC::DIR_PWR5X: 144 case PPC::DIR_PWR6: 145 case PPC::DIR_PWR6X: 146 case PPC::DIR_PWR7: 147 case PPC::DIR_PWR8: 148 Latency += 2; 149 break; 150 } 151 } 152 153 return Latency; 154 } 155 156 // Detect 32 -> 64-bit extensions where we may reuse the low sub-register. 157 bool PPCInstrInfo::isCoalescableExtInstr(const MachineInstr &MI, 158 unsigned &SrcReg, unsigned &DstReg, 159 unsigned &SubIdx) const { 160 switch (MI.getOpcode()) { 161 default: return false; 162 case PPC::EXTSW: 163 case PPC::EXTSW_32_64: 164 SrcReg = MI.getOperand(1).getReg(); 165 DstReg = MI.getOperand(0).getReg(); 166 SubIdx = PPC::sub_32; 167 return true; 168 } 169 } 170 171 unsigned PPCInstrInfo::isLoadFromStackSlot(const MachineInstr *MI, 172 int &FrameIndex) const { 173 // Note: This list must be kept consistent with LoadRegFromStackSlot. 174 switch (MI->getOpcode()) { 175 default: break; 176 case PPC::LD: 177 case PPC::LWZ: 178 case PPC::LFS: 179 case PPC::LFD: 180 case PPC::RESTORE_CR: 181 case PPC::RESTORE_CRBIT: 182 case PPC::LVX: 183 case PPC::LXVD2X: 184 case PPC::QVLFDX: 185 case PPC::QVLFSXs: 186 case PPC::QVLFDXb: 187 case PPC::RESTORE_VRSAVE: 188 // Check for the operands added by addFrameReference (the immediate is the 189 // offset which defaults to 0). 190 if (MI->getOperand(1).isImm() && !MI->getOperand(1).getImm() && 191 MI->getOperand(2).isFI()) { 192 FrameIndex = MI->getOperand(2).getIndex(); 193 return MI->getOperand(0).getReg(); 194 } 195 break; 196 } 197 return 0; 198 } 199 200 unsigned PPCInstrInfo::isStoreToStackSlot(const MachineInstr *MI, 201 int &FrameIndex) const { 202 // Note: This list must be kept consistent with StoreRegToStackSlot. 203 switch (MI->getOpcode()) { 204 default: break; 205 case PPC::STD: 206 case PPC::STW: 207 case PPC::STFS: 208 case PPC::STFD: 209 case PPC::SPILL_CR: 210 case PPC::SPILL_CRBIT: 211 case PPC::STVX: 212 case PPC::STXVD2X: 213 case PPC::QVSTFDX: 214 case PPC::QVSTFSXs: 215 case PPC::QVSTFDXb: 216 case PPC::SPILL_VRSAVE: 217 // Check for the operands added by addFrameReference (the immediate is the 218 // offset which defaults to 0). 219 if (MI->getOperand(1).isImm() && !MI->getOperand(1).getImm() && 220 MI->getOperand(2).isFI()) { 221 FrameIndex = MI->getOperand(2).getIndex(); 222 return MI->getOperand(0).getReg(); 223 } 224 break; 225 } 226 return 0; 227 } 228 229 // commuteInstruction - We can commute rlwimi instructions, but only if the 230 // rotate amt is zero. We also have to munge the immediates a bit. 231 MachineInstr * 232 PPCInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const { 233 MachineFunction &MF = *MI->getParent()->getParent(); 234 235 // Normal instructions can be commuted the obvious way. 236 if (MI->getOpcode() != PPC::RLWIMI && 237 MI->getOpcode() != PPC::RLWIMIo) 238 return TargetInstrInfo::commuteInstruction(MI, NewMI); 239 // Note that RLWIMI can be commuted as a 32-bit instruction, but not as a 240 // 64-bit instruction (so we don't handle PPC::RLWIMI8 here), because 241 // changing the relative order of the mask operands might change what happens 242 // to the high-bits of the mask (and, thus, the result). 243 244 // Cannot commute if it has a non-zero rotate count. 245 if (MI->getOperand(3).getImm() != 0) 246 return nullptr; 247 248 // If we have a zero rotate count, we have: 249 // M = mask(MB,ME) 250 // Op0 = (Op1 & ~M) | (Op2 & M) 251 // Change this to: 252 // M = mask((ME+1)&31, (MB-1)&31) 253 // Op0 = (Op2 & ~M) | (Op1 & M) 254 255 // Swap op1/op2 256 unsigned Reg0 = MI->getOperand(0).getReg(); 257 unsigned Reg1 = MI->getOperand(1).getReg(); 258 unsigned Reg2 = MI->getOperand(2).getReg(); 259 unsigned SubReg1 = MI->getOperand(1).getSubReg(); 260 unsigned SubReg2 = MI->getOperand(2).getSubReg(); 261 bool Reg1IsKill = MI->getOperand(1).isKill(); 262 bool Reg2IsKill = MI->getOperand(2).isKill(); 263 bool ChangeReg0 = false; 264 // If machine instrs are no longer in two-address forms, update 265 // destination register as well. 266 if (Reg0 == Reg1) { 267 // Must be two address instruction! 268 assert(MI->getDesc().getOperandConstraint(0, MCOI::TIED_TO) && 269 "Expecting a two-address instruction!"); 270 assert(MI->getOperand(0).getSubReg() == SubReg1 && "Tied subreg mismatch"); 271 Reg2IsKill = false; 272 ChangeReg0 = true; 273 } 274 275 // Masks. 276 unsigned MB = MI->getOperand(4).getImm(); 277 unsigned ME = MI->getOperand(5).getImm(); 278 279 if (NewMI) { 280 // Create a new instruction. 281 unsigned Reg0 = ChangeReg0 ? Reg2 : MI->getOperand(0).getReg(); 282 bool Reg0IsDead = MI->getOperand(0).isDead(); 283 return BuildMI(MF, MI->getDebugLoc(), MI->getDesc()) 284 .addReg(Reg0, RegState::Define | getDeadRegState(Reg0IsDead)) 285 .addReg(Reg2, getKillRegState(Reg2IsKill)) 286 .addReg(Reg1, getKillRegState(Reg1IsKill)) 287 .addImm((ME+1) & 31) 288 .addImm((MB-1) & 31); 289 } 290 291 if (ChangeReg0) { 292 MI->getOperand(0).setReg(Reg2); 293 MI->getOperand(0).setSubReg(SubReg2); 294 } 295 MI->getOperand(2).setReg(Reg1); 296 MI->getOperand(1).setReg(Reg2); 297 MI->getOperand(2).setSubReg(SubReg1); 298 MI->getOperand(1).setSubReg(SubReg2); 299 MI->getOperand(2).setIsKill(Reg1IsKill); 300 MI->getOperand(1).setIsKill(Reg2IsKill); 301 302 // Swap the mask around. 303 MI->getOperand(4).setImm((ME+1) & 31); 304 MI->getOperand(5).setImm((MB-1) & 31); 305 return MI; 306 } 307 308 bool PPCInstrInfo::findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1, 309 unsigned &SrcOpIdx2) const { 310 // For VSX A-Type FMA instructions, it is the first two operands that can be 311 // commuted, however, because the non-encoded tied input operand is listed 312 // first, the operands to swap are actually the second and third. 313 314 int AltOpc = PPC::getAltVSXFMAOpcode(MI->getOpcode()); 315 if (AltOpc == -1) 316 return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2); 317 318 SrcOpIdx1 = 2; 319 SrcOpIdx2 = 3; 320 return true; 321 } 322 323 void PPCInstrInfo::insertNoop(MachineBasicBlock &MBB, 324 MachineBasicBlock::iterator MI) const { 325 // This function is used for scheduling, and the nop wanted here is the type 326 // that terminates dispatch groups on the POWER cores. 327 unsigned Directive = Subtarget.getDarwinDirective(); 328 unsigned Opcode; 329 switch (Directive) { 330 default: Opcode = PPC::NOP; break; 331 case PPC::DIR_PWR6: Opcode = PPC::NOP_GT_PWR6; break; 332 case PPC::DIR_PWR7: Opcode = PPC::NOP_GT_PWR7; break; 333 case PPC::DIR_PWR8: Opcode = PPC::NOP_GT_PWR7; break; /* FIXME: Update when P8 InstrScheduling model is ready */ 334 } 335 336 DebugLoc DL; 337 BuildMI(MBB, MI, DL, get(Opcode)); 338 } 339 340 /// getNoopForMachoTarget - Return the noop instruction to use for a noop. 341 void PPCInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const { 342 NopInst.setOpcode(PPC::NOP); 343 } 344 345 // Branch analysis. 346 // Note: If the condition register is set to CTR or CTR8 then this is a 347 // BDNZ (imm == 1) or BDZ (imm == 0) branch. 348 bool PPCInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB, 349 MachineBasicBlock *&FBB, 350 SmallVectorImpl<MachineOperand> &Cond, 351 bool AllowModify) const { 352 bool isPPC64 = Subtarget.isPPC64(); 353 354 // If the block has no terminators, it just falls into the block after it. 355 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); 356 if (I == MBB.end()) 357 return false; 358 359 if (!isUnpredicatedTerminator(I)) 360 return false; 361 362 // Get the last instruction in the block. 363 MachineInstr *LastInst = I; 364 365 // If there is only one terminator instruction, process it. 366 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) { 367 if (LastInst->getOpcode() == PPC::B) { 368 if (!LastInst->getOperand(0).isMBB()) 369 return true; 370 TBB = LastInst->getOperand(0).getMBB(); 371 return false; 372 } else if (LastInst->getOpcode() == PPC::BCC) { 373 if (!LastInst->getOperand(2).isMBB()) 374 return true; 375 // Block ends with fall-through condbranch. 376 TBB = LastInst->getOperand(2).getMBB(); 377 Cond.push_back(LastInst->getOperand(0)); 378 Cond.push_back(LastInst->getOperand(1)); 379 return false; 380 } else if (LastInst->getOpcode() == PPC::BC) { 381 if (!LastInst->getOperand(1).isMBB()) 382 return true; 383 // Block ends with fall-through condbranch. 384 TBB = LastInst->getOperand(1).getMBB(); 385 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET)); 386 Cond.push_back(LastInst->getOperand(0)); 387 return false; 388 } else if (LastInst->getOpcode() == PPC::BCn) { 389 if (!LastInst->getOperand(1).isMBB()) 390 return true; 391 // Block ends with fall-through condbranch. 392 TBB = LastInst->getOperand(1).getMBB(); 393 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET)); 394 Cond.push_back(LastInst->getOperand(0)); 395 return false; 396 } else if (LastInst->getOpcode() == PPC::BDNZ8 || 397 LastInst->getOpcode() == PPC::BDNZ) { 398 if (!LastInst->getOperand(0).isMBB()) 399 return true; 400 if (DisableCTRLoopAnal) 401 return true; 402 TBB = LastInst->getOperand(0).getMBB(); 403 Cond.push_back(MachineOperand::CreateImm(1)); 404 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, 405 true)); 406 return false; 407 } else if (LastInst->getOpcode() == PPC::BDZ8 || 408 LastInst->getOpcode() == PPC::BDZ) { 409 if (!LastInst->getOperand(0).isMBB()) 410 return true; 411 if (DisableCTRLoopAnal) 412 return true; 413 TBB = LastInst->getOperand(0).getMBB(); 414 Cond.push_back(MachineOperand::CreateImm(0)); 415 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, 416 true)); 417 return false; 418 } 419 420 // Otherwise, don't know what this is. 421 return true; 422 } 423 424 // Get the instruction before it if it's a terminator. 425 MachineInstr *SecondLastInst = I; 426 427 // If there are three terminators, we don't know what sort of block this is. 428 if (SecondLastInst && I != MBB.begin() && 429 isUnpredicatedTerminator(--I)) 430 return true; 431 432 // If the block ends with PPC::B and PPC:BCC, handle it. 433 if (SecondLastInst->getOpcode() == PPC::BCC && 434 LastInst->getOpcode() == PPC::B) { 435 if (!SecondLastInst->getOperand(2).isMBB() || 436 !LastInst->getOperand(0).isMBB()) 437 return true; 438 TBB = SecondLastInst->getOperand(2).getMBB(); 439 Cond.push_back(SecondLastInst->getOperand(0)); 440 Cond.push_back(SecondLastInst->getOperand(1)); 441 FBB = LastInst->getOperand(0).getMBB(); 442 return false; 443 } else if (SecondLastInst->getOpcode() == PPC::BC && 444 LastInst->getOpcode() == PPC::B) { 445 if (!SecondLastInst->getOperand(1).isMBB() || 446 !LastInst->getOperand(0).isMBB()) 447 return true; 448 TBB = SecondLastInst->getOperand(1).getMBB(); 449 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET)); 450 Cond.push_back(SecondLastInst->getOperand(0)); 451 FBB = LastInst->getOperand(0).getMBB(); 452 return false; 453 } else if (SecondLastInst->getOpcode() == PPC::BCn && 454 LastInst->getOpcode() == PPC::B) { 455 if (!SecondLastInst->getOperand(1).isMBB() || 456 !LastInst->getOperand(0).isMBB()) 457 return true; 458 TBB = SecondLastInst->getOperand(1).getMBB(); 459 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET)); 460 Cond.push_back(SecondLastInst->getOperand(0)); 461 FBB = LastInst->getOperand(0).getMBB(); 462 return false; 463 } else if ((SecondLastInst->getOpcode() == PPC::BDNZ8 || 464 SecondLastInst->getOpcode() == PPC::BDNZ) && 465 LastInst->getOpcode() == PPC::B) { 466 if (!SecondLastInst->getOperand(0).isMBB() || 467 !LastInst->getOperand(0).isMBB()) 468 return true; 469 if (DisableCTRLoopAnal) 470 return true; 471 TBB = SecondLastInst->getOperand(0).getMBB(); 472 Cond.push_back(MachineOperand::CreateImm(1)); 473 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, 474 true)); 475 FBB = LastInst->getOperand(0).getMBB(); 476 return false; 477 } else if ((SecondLastInst->getOpcode() == PPC::BDZ8 || 478 SecondLastInst->getOpcode() == PPC::BDZ) && 479 LastInst->getOpcode() == PPC::B) { 480 if (!SecondLastInst->getOperand(0).isMBB() || 481 !LastInst->getOperand(0).isMBB()) 482 return true; 483 if (DisableCTRLoopAnal) 484 return true; 485 TBB = SecondLastInst->getOperand(0).getMBB(); 486 Cond.push_back(MachineOperand::CreateImm(0)); 487 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, 488 true)); 489 FBB = LastInst->getOperand(0).getMBB(); 490 return false; 491 } 492 493 // If the block ends with two PPC:Bs, handle it. The second one is not 494 // executed, so remove it. 495 if (SecondLastInst->getOpcode() == PPC::B && 496 LastInst->getOpcode() == PPC::B) { 497 if (!SecondLastInst->getOperand(0).isMBB()) 498 return true; 499 TBB = SecondLastInst->getOperand(0).getMBB(); 500 I = LastInst; 501 if (AllowModify) 502 I->eraseFromParent(); 503 return false; 504 } 505 506 // Otherwise, can't handle this. 507 return true; 508 } 509 510 unsigned PPCInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { 511 MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); 512 if (I == MBB.end()) 513 return 0; 514 515 if (I->getOpcode() != PPC::B && I->getOpcode() != PPC::BCC && 516 I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn && 517 I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ && 518 I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ) 519 return 0; 520 521 // Remove the branch. 522 I->eraseFromParent(); 523 524 I = MBB.end(); 525 526 if (I == MBB.begin()) return 1; 527 --I; 528 if (I->getOpcode() != PPC::BCC && 529 I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn && 530 I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ && 531 I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ) 532 return 1; 533 534 // Remove the branch. 535 I->eraseFromParent(); 536 return 2; 537 } 538 539 unsigned 540 PPCInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, 541 MachineBasicBlock *FBB, 542 ArrayRef<MachineOperand> Cond, 543 DebugLoc DL) const { 544 // Shouldn't be a fall through. 545 assert(TBB && "InsertBranch must not be told to insert a fallthrough"); 546 assert((Cond.size() == 2 || Cond.size() == 0) && 547 "PPC branch conditions have two components!"); 548 549 bool isPPC64 = Subtarget.isPPC64(); 550 551 // One-way branch. 552 if (!FBB) { 553 if (Cond.empty()) // Unconditional branch 554 BuildMI(&MBB, DL, get(PPC::B)).addMBB(TBB); 555 else if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) 556 BuildMI(&MBB, DL, get(Cond[0].getImm() ? 557 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : 558 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB); 559 else if (Cond[0].getImm() == PPC::PRED_BIT_SET) 560 BuildMI(&MBB, DL, get(PPC::BC)).addOperand(Cond[1]).addMBB(TBB); 561 else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET) 562 BuildMI(&MBB, DL, get(PPC::BCn)).addOperand(Cond[1]).addMBB(TBB); 563 else // Conditional branch 564 BuildMI(&MBB, DL, get(PPC::BCC)) 565 .addImm(Cond[0].getImm()).addOperand(Cond[1]).addMBB(TBB); 566 return 1; 567 } 568 569 // Two-way Conditional Branch. 570 if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) 571 BuildMI(&MBB, DL, get(Cond[0].getImm() ? 572 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : 573 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB); 574 else if (Cond[0].getImm() == PPC::PRED_BIT_SET) 575 BuildMI(&MBB, DL, get(PPC::BC)).addOperand(Cond[1]).addMBB(TBB); 576 else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET) 577 BuildMI(&MBB, DL, get(PPC::BCn)).addOperand(Cond[1]).addMBB(TBB); 578 else 579 BuildMI(&MBB, DL, get(PPC::BCC)) 580 .addImm(Cond[0].getImm()).addOperand(Cond[1]).addMBB(TBB); 581 BuildMI(&MBB, DL, get(PPC::B)).addMBB(FBB); 582 return 2; 583 } 584 585 // Select analysis. 586 bool PPCInstrInfo::canInsertSelect(const MachineBasicBlock &MBB, 587 ArrayRef<MachineOperand> Cond, 588 unsigned TrueReg, unsigned FalseReg, 589 int &CondCycles, int &TrueCycles, int &FalseCycles) const { 590 if (!Subtarget.hasISEL()) 591 return false; 592 593 if (Cond.size() != 2) 594 return false; 595 596 // If this is really a bdnz-like condition, then it cannot be turned into a 597 // select. 598 if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) 599 return false; 600 601 // Check register classes. 602 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 603 const TargetRegisterClass *RC = 604 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 605 if (!RC) 606 return false; 607 608 // isel is for regular integer GPRs only. 609 if (!PPC::GPRCRegClass.hasSubClassEq(RC) && 610 !PPC::GPRC_NOR0RegClass.hasSubClassEq(RC) && 611 !PPC::G8RCRegClass.hasSubClassEq(RC) && 612 !PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) 613 return false; 614 615 // FIXME: These numbers are for the A2, how well they work for other cores is 616 // an open question. On the A2, the isel instruction has a 2-cycle latency 617 // but single-cycle throughput. These numbers are used in combination with 618 // the MispredictPenalty setting from the active SchedMachineModel. 619 CondCycles = 1; 620 TrueCycles = 1; 621 FalseCycles = 1; 622 623 return true; 624 } 625 626 void PPCInstrInfo::insertSelect(MachineBasicBlock &MBB, 627 MachineBasicBlock::iterator MI, DebugLoc dl, 628 unsigned DestReg, ArrayRef<MachineOperand> Cond, 629 unsigned TrueReg, unsigned FalseReg) const { 630 assert(Cond.size() == 2 && 631 "PPC branch conditions have two components!"); 632 633 assert(Subtarget.hasISEL() && 634 "Cannot insert select on target without ISEL support"); 635 636 // Get the register classes. 637 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 638 const TargetRegisterClass *RC = 639 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 640 assert(RC && "TrueReg and FalseReg must have overlapping register classes"); 641 642 bool Is64Bit = PPC::G8RCRegClass.hasSubClassEq(RC) || 643 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC); 644 assert((Is64Bit || 645 PPC::GPRCRegClass.hasSubClassEq(RC) || 646 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) && 647 "isel is for regular integer GPRs only"); 648 649 unsigned OpCode = Is64Bit ? PPC::ISEL8 : PPC::ISEL; 650 unsigned SelectPred = Cond[0].getImm(); 651 652 unsigned SubIdx; 653 bool SwapOps; 654 switch (SelectPred) { 655 default: llvm_unreachable("invalid predicate for isel"); 656 case PPC::PRED_EQ: SubIdx = PPC::sub_eq; SwapOps = false; break; 657 case PPC::PRED_NE: SubIdx = PPC::sub_eq; SwapOps = true; break; 658 case PPC::PRED_LT: SubIdx = PPC::sub_lt; SwapOps = false; break; 659 case PPC::PRED_GE: SubIdx = PPC::sub_lt; SwapOps = true; break; 660 case PPC::PRED_GT: SubIdx = PPC::sub_gt; SwapOps = false; break; 661 case PPC::PRED_LE: SubIdx = PPC::sub_gt; SwapOps = true; break; 662 case PPC::PRED_UN: SubIdx = PPC::sub_un; SwapOps = false; break; 663 case PPC::PRED_NU: SubIdx = PPC::sub_un; SwapOps = true; break; 664 case PPC::PRED_BIT_SET: SubIdx = 0; SwapOps = false; break; 665 case PPC::PRED_BIT_UNSET: SubIdx = 0; SwapOps = true; break; 666 } 667 668 unsigned FirstReg = SwapOps ? FalseReg : TrueReg, 669 SecondReg = SwapOps ? TrueReg : FalseReg; 670 671 // The first input register of isel cannot be r0. If it is a member 672 // of a register class that can be r0, then copy it first (the 673 // register allocator should eliminate the copy). 674 if (MRI.getRegClass(FirstReg)->contains(PPC::R0) || 675 MRI.getRegClass(FirstReg)->contains(PPC::X0)) { 676 const TargetRegisterClass *FirstRC = 677 MRI.getRegClass(FirstReg)->contains(PPC::X0) ? 678 &PPC::G8RC_NOX0RegClass : &PPC::GPRC_NOR0RegClass; 679 unsigned OldFirstReg = FirstReg; 680 FirstReg = MRI.createVirtualRegister(FirstRC); 681 BuildMI(MBB, MI, dl, get(TargetOpcode::COPY), FirstReg) 682 .addReg(OldFirstReg); 683 } 684 685 BuildMI(MBB, MI, dl, get(OpCode), DestReg) 686 .addReg(FirstReg).addReg(SecondReg) 687 .addReg(Cond[1].getReg(), 0, SubIdx); 688 } 689 690 static unsigned getCRBitValue(unsigned CRBit) { 691 unsigned Ret = 4; 692 if (CRBit == PPC::CR0LT || CRBit == PPC::CR1LT || 693 CRBit == PPC::CR2LT || CRBit == PPC::CR3LT || 694 CRBit == PPC::CR4LT || CRBit == PPC::CR5LT || 695 CRBit == PPC::CR6LT || CRBit == PPC::CR7LT) 696 Ret = 3; 697 if (CRBit == PPC::CR0GT || CRBit == PPC::CR1GT || 698 CRBit == PPC::CR2GT || CRBit == PPC::CR3GT || 699 CRBit == PPC::CR4GT || CRBit == PPC::CR5GT || 700 CRBit == PPC::CR6GT || CRBit == PPC::CR7GT) 701 Ret = 2; 702 if (CRBit == PPC::CR0EQ || CRBit == PPC::CR1EQ || 703 CRBit == PPC::CR2EQ || CRBit == PPC::CR3EQ || 704 CRBit == PPC::CR4EQ || CRBit == PPC::CR5EQ || 705 CRBit == PPC::CR6EQ || CRBit == PPC::CR7EQ) 706 Ret = 1; 707 if (CRBit == PPC::CR0UN || CRBit == PPC::CR1UN || 708 CRBit == PPC::CR2UN || CRBit == PPC::CR3UN || 709 CRBit == PPC::CR4UN || CRBit == PPC::CR5UN || 710 CRBit == PPC::CR6UN || CRBit == PPC::CR7UN) 711 Ret = 0; 712 713 assert(Ret != 4 && "Invalid CR bit register"); 714 return Ret; 715 } 716 717 void PPCInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 718 MachineBasicBlock::iterator I, DebugLoc DL, 719 unsigned DestReg, unsigned SrcReg, 720 bool KillSrc) const { 721 // We can end up with self copies and similar things as a result of VSX copy 722 // legalization. Promote them here. 723 const TargetRegisterInfo *TRI = &getRegisterInfo(); 724 if (PPC::F8RCRegClass.contains(DestReg) && 725 PPC::VSRCRegClass.contains(SrcReg)) { 726 unsigned SuperReg = 727 TRI->getMatchingSuperReg(DestReg, PPC::sub_64, &PPC::VSRCRegClass); 728 729 if (VSXSelfCopyCrash && SrcReg == SuperReg) 730 llvm_unreachable("nop VSX copy"); 731 732 DestReg = SuperReg; 733 } else if (PPC::VRRCRegClass.contains(DestReg) && 734 PPC::VSRCRegClass.contains(SrcReg)) { 735 unsigned SuperReg = 736 TRI->getMatchingSuperReg(DestReg, PPC::sub_128, &PPC::VSRCRegClass); 737 738 if (VSXSelfCopyCrash && SrcReg == SuperReg) 739 llvm_unreachable("nop VSX copy"); 740 741 DestReg = SuperReg; 742 } else if (PPC::F8RCRegClass.contains(SrcReg) && 743 PPC::VSRCRegClass.contains(DestReg)) { 744 unsigned SuperReg = 745 TRI->getMatchingSuperReg(SrcReg, PPC::sub_64, &PPC::VSRCRegClass); 746 747 if (VSXSelfCopyCrash && DestReg == SuperReg) 748 llvm_unreachable("nop VSX copy"); 749 750 SrcReg = SuperReg; 751 } else if (PPC::VRRCRegClass.contains(SrcReg) && 752 PPC::VSRCRegClass.contains(DestReg)) { 753 unsigned SuperReg = 754 TRI->getMatchingSuperReg(SrcReg, PPC::sub_128, &PPC::VSRCRegClass); 755 756 if (VSXSelfCopyCrash && DestReg == SuperReg) 757 llvm_unreachable("nop VSX copy"); 758 759 SrcReg = SuperReg; 760 } 761 762 // Different class register copy 763 if (PPC::CRBITRCRegClass.contains(SrcReg) && 764 PPC::GPRCRegClass.contains(DestReg)) { 765 unsigned CRReg = getCRFromCRBit(SrcReg); 766 BuildMI(MBB, I, DL, get(PPC::MFOCRF), DestReg) 767 .addReg(CRReg), getKillRegState(KillSrc); 768 // Rotate the CR bit in the CR fields to be the least significant bit and 769 // then mask with 0x1 (MB = ME = 31). 770 BuildMI(MBB, I, DL, get(PPC::RLWINM), DestReg) 771 .addReg(DestReg, RegState::Kill) 772 .addImm(TRI->getEncodingValue(CRReg) * 4 + (4 - getCRBitValue(SrcReg))) 773 .addImm(31) 774 .addImm(31); 775 return; 776 } else if (PPC::CRRCRegClass.contains(SrcReg) && 777 PPC::G8RCRegClass.contains(DestReg)) { 778 BuildMI(MBB, I, DL, get(PPC::MFOCRF8), DestReg) 779 .addReg(SrcReg), getKillRegState(KillSrc); 780 return; 781 } else if (PPC::CRRCRegClass.contains(SrcReg) && 782 PPC::GPRCRegClass.contains(DestReg)) { 783 BuildMI(MBB, I, DL, get(PPC::MFOCRF), DestReg) 784 .addReg(SrcReg), getKillRegState(KillSrc); 785 return; 786 } 787 788 unsigned Opc; 789 if (PPC::GPRCRegClass.contains(DestReg, SrcReg)) 790 Opc = PPC::OR; 791 else if (PPC::G8RCRegClass.contains(DestReg, SrcReg)) 792 Opc = PPC::OR8; 793 else if (PPC::F4RCRegClass.contains(DestReg, SrcReg)) 794 Opc = PPC::FMR; 795 else if (PPC::CRRCRegClass.contains(DestReg, SrcReg)) 796 Opc = PPC::MCRF; 797 else if (PPC::VRRCRegClass.contains(DestReg, SrcReg)) 798 Opc = PPC::VOR; 799 else if (PPC::VSRCRegClass.contains(DestReg, SrcReg)) 800 // There are two different ways this can be done: 801 // 1. xxlor : This has lower latency (on the P7), 2 cycles, but can only 802 // issue in VSU pipeline 0. 803 // 2. xmovdp/xmovsp: This has higher latency (on the P7), 6 cycles, but 804 // can go to either pipeline. 805 // We'll always use xxlor here, because in practically all cases where 806 // copies are generated, they are close enough to some use that the 807 // lower-latency form is preferable. 808 Opc = PPC::XXLOR; 809 else if (PPC::VSFRCRegClass.contains(DestReg, SrcReg) || 810 PPC::VSSRCRegClass.contains(DestReg, SrcReg)) 811 Opc = PPC::XXLORf; 812 else if (PPC::QFRCRegClass.contains(DestReg, SrcReg)) 813 Opc = PPC::QVFMR; 814 else if (PPC::QSRCRegClass.contains(DestReg, SrcReg)) 815 Opc = PPC::QVFMRs; 816 else if (PPC::QBRCRegClass.contains(DestReg, SrcReg)) 817 Opc = PPC::QVFMRb; 818 else if (PPC::CRBITRCRegClass.contains(DestReg, SrcReg)) 819 Opc = PPC::CROR; 820 else 821 llvm_unreachable("Impossible reg-to-reg copy"); 822 823 const MCInstrDesc &MCID = get(Opc); 824 if (MCID.getNumOperands() == 3) 825 BuildMI(MBB, I, DL, MCID, DestReg) 826 .addReg(SrcReg).addReg(SrcReg, getKillRegState(KillSrc)); 827 else 828 BuildMI(MBB, I, DL, MCID, DestReg).addReg(SrcReg, getKillRegState(KillSrc)); 829 } 830 831 // This function returns true if a CR spill is necessary and false otherwise. 832 bool 833 PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF, 834 unsigned SrcReg, bool isKill, 835 int FrameIdx, 836 const TargetRegisterClass *RC, 837 SmallVectorImpl<MachineInstr*> &NewMIs, 838 bool &NonRI, bool &SpillsVRS) const{ 839 // Note: If additional store instructions are added here, 840 // update isStoreToStackSlot. 841 842 DebugLoc DL; 843 if (PPC::GPRCRegClass.hasSubClassEq(RC) || 844 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) { 845 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STW)) 846 .addReg(SrcReg, 847 getKillRegState(isKill)), 848 FrameIdx)); 849 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) || 850 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) { 851 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STD)) 852 .addReg(SrcReg, 853 getKillRegState(isKill)), 854 FrameIdx)); 855 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) { 856 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFD)) 857 .addReg(SrcReg, 858 getKillRegState(isKill)), 859 FrameIdx)); 860 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) { 861 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFS)) 862 .addReg(SrcReg, 863 getKillRegState(isKill)), 864 FrameIdx)); 865 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) { 866 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CR)) 867 .addReg(SrcReg, 868 getKillRegState(isKill)), 869 FrameIdx)); 870 return true; 871 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) { 872 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CRBIT)) 873 .addReg(SrcReg, 874 getKillRegState(isKill)), 875 FrameIdx)); 876 return true; 877 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) { 878 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STVX)) 879 .addReg(SrcReg, 880 getKillRegState(isKill)), 881 FrameIdx)); 882 NonRI = true; 883 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) { 884 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXVD2X)) 885 .addReg(SrcReg, 886 getKillRegState(isKill)), 887 FrameIdx)); 888 NonRI = true; 889 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) { 890 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXSDX)) 891 .addReg(SrcReg, 892 getKillRegState(isKill)), 893 FrameIdx)); 894 NonRI = true; 895 } else if (PPC::VSSRCRegClass.hasSubClassEq(RC)) { 896 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXSSPX)) 897 .addReg(SrcReg, 898 getKillRegState(isKill)), 899 FrameIdx)); 900 NonRI = true; 901 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) { 902 assert(Subtarget.isDarwin() && 903 "VRSAVE only needs spill/restore on Darwin"); 904 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_VRSAVE)) 905 .addReg(SrcReg, 906 getKillRegState(isKill)), 907 FrameIdx)); 908 SpillsVRS = true; 909 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) { 910 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFDX)) 911 .addReg(SrcReg, 912 getKillRegState(isKill)), 913 FrameIdx)); 914 NonRI = true; 915 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) { 916 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFSXs)) 917 .addReg(SrcReg, 918 getKillRegState(isKill)), 919 FrameIdx)); 920 NonRI = true; 921 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) { 922 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFDXb)) 923 .addReg(SrcReg, 924 getKillRegState(isKill)), 925 FrameIdx)); 926 NonRI = true; 927 } else { 928 llvm_unreachable("Unknown regclass!"); 929 } 930 931 return false; 932 } 933 934 void 935 PPCInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 936 MachineBasicBlock::iterator MI, 937 unsigned SrcReg, bool isKill, int FrameIdx, 938 const TargetRegisterClass *RC, 939 const TargetRegisterInfo *TRI) const { 940 MachineFunction &MF = *MBB.getParent(); 941 SmallVector<MachineInstr*, 4> NewMIs; 942 943 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); 944 FuncInfo->setHasSpills(); 945 946 bool NonRI = false, SpillsVRS = false; 947 if (StoreRegToStackSlot(MF, SrcReg, isKill, FrameIdx, RC, NewMIs, 948 NonRI, SpillsVRS)) 949 FuncInfo->setSpillsCR(); 950 951 if (SpillsVRS) 952 FuncInfo->setSpillsVRSAVE(); 953 954 if (NonRI) 955 FuncInfo->setHasNonRISpills(); 956 957 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i) 958 MBB.insert(MI, NewMIs[i]); 959 960 const MachineFrameInfo &MFI = *MF.getFrameInfo(); 961 MachineMemOperand *MMO = 962 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx), 963 MachineMemOperand::MOStore, 964 MFI.getObjectSize(FrameIdx), 965 MFI.getObjectAlignment(FrameIdx)); 966 NewMIs.back()->addMemOperand(MF, MMO); 967 } 968 969 bool 970 PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL, 971 unsigned DestReg, int FrameIdx, 972 const TargetRegisterClass *RC, 973 SmallVectorImpl<MachineInstr*> &NewMIs, 974 bool &NonRI, bool &SpillsVRS) const{ 975 // Note: If additional load instructions are added here, 976 // update isLoadFromStackSlot. 977 978 if (PPC::GPRCRegClass.hasSubClassEq(RC) || 979 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) { 980 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LWZ), 981 DestReg), FrameIdx)); 982 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) || 983 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) { 984 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LD), DestReg), 985 FrameIdx)); 986 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) { 987 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFD), DestReg), 988 FrameIdx)); 989 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) { 990 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFS), DestReg), 991 FrameIdx)); 992 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) { 993 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, 994 get(PPC::RESTORE_CR), DestReg), 995 FrameIdx)); 996 return true; 997 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) { 998 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, 999 get(PPC::RESTORE_CRBIT), DestReg), 1000 FrameIdx)); 1001 return true; 1002 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) { 1003 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LVX), DestReg), 1004 FrameIdx)); 1005 NonRI = true; 1006 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) { 1007 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXVD2X), DestReg), 1008 FrameIdx)); 1009 NonRI = true; 1010 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) { 1011 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXSDX), DestReg), 1012 FrameIdx)); 1013 NonRI = true; 1014 } else if (PPC::VSSRCRegClass.hasSubClassEq(RC)) { 1015 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXSSPX), DestReg), 1016 FrameIdx)); 1017 NonRI = true; 1018 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) { 1019 assert(Subtarget.isDarwin() && 1020 "VRSAVE only needs spill/restore on Darwin"); 1021 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, 1022 get(PPC::RESTORE_VRSAVE), 1023 DestReg), 1024 FrameIdx)); 1025 SpillsVRS = true; 1026 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) { 1027 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFDX), DestReg), 1028 FrameIdx)); 1029 NonRI = true; 1030 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) { 1031 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFSXs), DestReg), 1032 FrameIdx)); 1033 NonRI = true; 1034 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) { 1035 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFDXb), DestReg), 1036 FrameIdx)); 1037 NonRI = true; 1038 } else { 1039 llvm_unreachable("Unknown regclass!"); 1040 } 1041 1042 return false; 1043 } 1044 1045 void 1046 PPCInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, 1047 MachineBasicBlock::iterator MI, 1048 unsigned DestReg, int FrameIdx, 1049 const TargetRegisterClass *RC, 1050 const TargetRegisterInfo *TRI) const { 1051 MachineFunction &MF = *MBB.getParent(); 1052 SmallVector<MachineInstr*, 4> NewMIs; 1053 DebugLoc DL; 1054 if (MI != MBB.end()) DL = MI->getDebugLoc(); 1055 1056 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); 1057 FuncInfo->setHasSpills(); 1058 1059 bool NonRI = false, SpillsVRS = false; 1060 if (LoadRegFromStackSlot(MF, DL, DestReg, FrameIdx, RC, NewMIs, 1061 NonRI, SpillsVRS)) 1062 FuncInfo->setSpillsCR(); 1063 1064 if (SpillsVRS) 1065 FuncInfo->setSpillsVRSAVE(); 1066 1067 if (NonRI) 1068 FuncInfo->setHasNonRISpills(); 1069 1070 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i) 1071 MBB.insert(MI, NewMIs[i]); 1072 1073 const MachineFrameInfo &MFI = *MF.getFrameInfo(); 1074 MachineMemOperand *MMO = 1075 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx), 1076 MachineMemOperand::MOLoad, 1077 MFI.getObjectSize(FrameIdx), 1078 MFI.getObjectAlignment(FrameIdx)); 1079 NewMIs.back()->addMemOperand(MF, MMO); 1080 } 1081 1082 bool PPCInstrInfo:: 1083 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 1084 assert(Cond.size() == 2 && "Invalid PPC branch opcode!"); 1085 if (Cond[1].getReg() == PPC::CTR8 || Cond[1].getReg() == PPC::CTR) 1086 Cond[0].setImm(Cond[0].getImm() == 0 ? 1 : 0); 1087 else 1088 // Leave the CR# the same, but invert the condition. 1089 Cond[0].setImm(PPC::InvertPredicate((PPC::Predicate)Cond[0].getImm())); 1090 return false; 1091 } 1092 1093 bool PPCInstrInfo::FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI, 1094 unsigned Reg, MachineRegisterInfo *MRI) const { 1095 // For some instructions, it is legal to fold ZERO into the RA register field. 1096 // A zero immediate should always be loaded with a single li. 1097 unsigned DefOpc = DefMI->getOpcode(); 1098 if (DefOpc != PPC::LI && DefOpc != PPC::LI8) 1099 return false; 1100 if (!DefMI->getOperand(1).isImm()) 1101 return false; 1102 if (DefMI->getOperand(1).getImm() != 0) 1103 return false; 1104 1105 // Note that we cannot here invert the arguments of an isel in order to fold 1106 // a ZERO into what is presented as the second argument. All we have here 1107 // is the condition bit, and that might come from a CR-logical bit operation. 1108 1109 const MCInstrDesc &UseMCID = UseMI->getDesc(); 1110 1111 // Only fold into real machine instructions. 1112 if (UseMCID.isPseudo()) 1113 return false; 1114 1115 unsigned UseIdx; 1116 for (UseIdx = 0; UseIdx < UseMI->getNumOperands(); ++UseIdx) 1117 if (UseMI->getOperand(UseIdx).isReg() && 1118 UseMI->getOperand(UseIdx).getReg() == Reg) 1119 break; 1120 1121 assert(UseIdx < UseMI->getNumOperands() && "Cannot find Reg in UseMI"); 1122 assert(UseIdx < UseMCID.getNumOperands() && "No operand description for Reg"); 1123 1124 const MCOperandInfo *UseInfo = &UseMCID.OpInfo[UseIdx]; 1125 1126 // We can fold the zero if this register requires a GPRC_NOR0/G8RC_NOX0 1127 // register (which might also be specified as a pointer class kind). 1128 if (UseInfo->isLookupPtrRegClass()) { 1129 if (UseInfo->RegClass /* Kind */ != 1) 1130 return false; 1131 } else { 1132 if (UseInfo->RegClass != PPC::GPRC_NOR0RegClassID && 1133 UseInfo->RegClass != PPC::G8RC_NOX0RegClassID) 1134 return false; 1135 } 1136 1137 // Make sure this is not tied to an output register (or otherwise 1138 // constrained). This is true for ST?UX registers, for example, which 1139 // are tied to their output registers. 1140 if (UseInfo->Constraints != 0) 1141 return false; 1142 1143 unsigned ZeroReg; 1144 if (UseInfo->isLookupPtrRegClass()) { 1145 bool isPPC64 = Subtarget.isPPC64(); 1146 ZeroReg = isPPC64 ? PPC::ZERO8 : PPC::ZERO; 1147 } else { 1148 ZeroReg = UseInfo->RegClass == PPC::G8RC_NOX0RegClassID ? 1149 PPC::ZERO8 : PPC::ZERO; 1150 } 1151 1152 bool DeleteDef = MRI->hasOneNonDBGUse(Reg); 1153 UseMI->getOperand(UseIdx).setReg(ZeroReg); 1154 1155 if (DeleteDef) 1156 DefMI->eraseFromParent(); 1157 1158 return true; 1159 } 1160 1161 static bool MBBDefinesCTR(MachineBasicBlock &MBB) { 1162 for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end(); 1163 I != IE; ++I) 1164 if (I->definesRegister(PPC::CTR) || I->definesRegister(PPC::CTR8)) 1165 return true; 1166 return false; 1167 } 1168 1169 // We should make sure that, if we're going to predicate both sides of a 1170 // condition (a diamond), that both sides don't define the counter register. We 1171 // can predicate counter-decrement-based branches, but while that predicates 1172 // the branching, it does not predicate the counter decrement. If we tried to 1173 // merge the triangle into one predicated block, we'd decrement the counter 1174 // twice. 1175 bool PPCInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB, 1176 unsigned NumT, unsigned ExtraT, 1177 MachineBasicBlock &FMBB, 1178 unsigned NumF, unsigned ExtraF, 1179 const BranchProbability &Probability) const { 1180 return !(MBBDefinesCTR(TMBB) && MBBDefinesCTR(FMBB)); 1181 } 1182 1183 1184 bool PPCInstrInfo::isPredicated(const MachineInstr *MI) const { 1185 // The predicated branches are identified by their type, not really by the 1186 // explicit presence of a predicate. Furthermore, some of them can be 1187 // predicated more than once. Because if conversion won't try to predicate 1188 // any instruction which already claims to be predicated (by returning true 1189 // here), always return false. In doing so, we let isPredicable() be the 1190 // final word on whether not the instruction can be (further) predicated. 1191 1192 return false; 1193 } 1194 1195 bool PPCInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const { 1196 if (!MI->isTerminator()) 1197 return false; 1198 1199 // Conditional branch is a special case. 1200 if (MI->isBranch() && !MI->isBarrier()) 1201 return true; 1202 1203 return !isPredicated(MI); 1204 } 1205 1206 bool PPCInstrInfo::PredicateInstruction(MachineInstr *MI, 1207 ArrayRef<MachineOperand> Pred) const { 1208 unsigned OpC = MI->getOpcode(); 1209 if (OpC == PPC::BLR || OpC == PPC::BLR8) { 1210 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) { 1211 bool isPPC64 = Subtarget.isPPC64(); 1212 MI->setDesc(get(Pred[0].getImm() ? 1213 (isPPC64 ? PPC::BDNZLR8 : PPC::BDNZLR) : 1214 (isPPC64 ? PPC::BDZLR8 : PPC::BDZLR))); 1215 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) { 1216 MI->setDesc(get(PPC::BCLR)); 1217 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1218 .addReg(Pred[1].getReg()); 1219 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { 1220 MI->setDesc(get(PPC::BCLRn)); 1221 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1222 .addReg(Pred[1].getReg()); 1223 } else { 1224 MI->setDesc(get(PPC::BCCLR)); 1225 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1226 .addImm(Pred[0].getImm()) 1227 .addReg(Pred[1].getReg()); 1228 } 1229 1230 return true; 1231 } else if (OpC == PPC::B) { 1232 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) { 1233 bool isPPC64 = Subtarget.isPPC64(); 1234 MI->setDesc(get(Pred[0].getImm() ? 1235 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : 1236 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))); 1237 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) { 1238 MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); 1239 MI->RemoveOperand(0); 1240 1241 MI->setDesc(get(PPC::BC)); 1242 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1243 .addReg(Pred[1].getReg()) 1244 .addMBB(MBB); 1245 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { 1246 MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); 1247 MI->RemoveOperand(0); 1248 1249 MI->setDesc(get(PPC::BCn)); 1250 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1251 .addReg(Pred[1].getReg()) 1252 .addMBB(MBB); 1253 } else { 1254 MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); 1255 MI->RemoveOperand(0); 1256 1257 MI->setDesc(get(PPC::BCC)); 1258 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1259 .addImm(Pred[0].getImm()) 1260 .addReg(Pred[1].getReg()) 1261 .addMBB(MBB); 1262 } 1263 1264 return true; 1265 } else if (OpC == PPC::BCTR || OpC == PPC::BCTR8 || 1266 OpC == PPC::BCTRL || OpC == PPC::BCTRL8) { 1267 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) 1268 llvm_unreachable("Cannot predicate bctr[l] on the ctr register"); 1269 1270 bool setLR = OpC == PPC::BCTRL || OpC == PPC::BCTRL8; 1271 bool isPPC64 = Subtarget.isPPC64(); 1272 1273 if (Pred[0].getImm() == PPC::PRED_BIT_SET) { 1274 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8) : 1275 (setLR ? PPC::BCCTRL : PPC::BCCTR))); 1276 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1277 .addReg(Pred[1].getReg()); 1278 return true; 1279 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { 1280 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8n : PPC::BCCTR8n) : 1281 (setLR ? PPC::BCCTRLn : PPC::BCCTRn))); 1282 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1283 .addReg(Pred[1].getReg()); 1284 return true; 1285 } 1286 1287 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCCTRL8 : PPC::BCCCTR8) : 1288 (setLR ? PPC::BCCCTRL : PPC::BCCCTR))); 1289 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1290 .addImm(Pred[0].getImm()) 1291 .addReg(Pred[1].getReg()); 1292 return true; 1293 } 1294 1295 return false; 1296 } 1297 1298 bool PPCInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1, 1299 ArrayRef<MachineOperand> Pred2) const { 1300 assert(Pred1.size() == 2 && "Invalid PPC first predicate"); 1301 assert(Pred2.size() == 2 && "Invalid PPC second predicate"); 1302 1303 if (Pred1[1].getReg() == PPC::CTR8 || Pred1[1].getReg() == PPC::CTR) 1304 return false; 1305 if (Pred2[1].getReg() == PPC::CTR8 || Pred2[1].getReg() == PPC::CTR) 1306 return false; 1307 1308 // P1 can only subsume P2 if they test the same condition register. 1309 if (Pred1[1].getReg() != Pred2[1].getReg()) 1310 return false; 1311 1312 PPC::Predicate P1 = (PPC::Predicate) Pred1[0].getImm(); 1313 PPC::Predicate P2 = (PPC::Predicate) Pred2[0].getImm(); 1314 1315 if (P1 == P2) 1316 return true; 1317 1318 // Does P1 subsume P2, e.g. GE subsumes GT. 1319 if (P1 == PPC::PRED_LE && 1320 (P2 == PPC::PRED_LT || P2 == PPC::PRED_EQ)) 1321 return true; 1322 if (P1 == PPC::PRED_GE && 1323 (P2 == PPC::PRED_GT || P2 == PPC::PRED_EQ)) 1324 return true; 1325 1326 return false; 1327 } 1328 1329 bool PPCInstrInfo::DefinesPredicate(MachineInstr *MI, 1330 std::vector<MachineOperand> &Pred) const { 1331 // Note: At the present time, the contents of Pred from this function is 1332 // unused by IfConversion. This implementation follows ARM by pushing the 1333 // CR-defining operand. Because the 'DZ' and 'DNZ' count as types of 1334 // predicate, instructions defining CTR or CTR8 are also included as 1335 // predicate-defining instructions. 1336 1337 const TargetRegisterClass *RCs[] = 1338 { &PPC::CRRCRegClass, &PPC::CRBITRCRegClass, 1339 &PPC::CTRRCRegClass, &PPC::CTRRC8RegClass }; 1340 1341 bool Found = false; 1342 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 1343 const MachineOperand &MO = MI->getOperand(i); 1344 for (unsigned c = 0; c < array_lengthof(RCs) && !Found; ++c) { 1345 const TargetRegisterClass *RC = RCs[c]; 1346 if (MO.isReg()) { 1347 if (MO.isDef() && RC->contains(MO.getReg())) { 1348 Pred.push_back(MO); 1349 Found = true; 1350 } 1351 } else if (MO.isRegMask()) { 1352 for (TargetRegisterClass::iterator I = RC->begin(), 1353 IE = RC->end(); I != IE; ++I) 1354 if (MO.clobbersPhysReg(*I)) { 1355 Pred.push_back(MO); 1356 Found = true; 1357 } 1358 } 1359 } 1360 } 1361 1362 return Found; 1363 } 1364 1365 bool PPCInstrInfo::isPredicable(MachineInstr *MI) const { 1366 unsigned OpC = MI->getOpcode(); 1367 switch (OpC) { 1368 default: 1369 return false; 1370 case PPC::B: 1371 case PPC::BLR: 1372 case PPC::BLR8: 1373 case PPC::BCTR: 1374 case PPC::BCTR8: 1375 case PPC::BCTRL: 1376 case PPC::BCTRL8: 1377 return true; 1378 } 1379 } 1380 1381 bool PPCInstrInfo::analyzeCompare(const MachineInstr *MI, 1382 unsigned &SrcReg, unsigned &SrcReg2, 1383 int &Mask, int &Value) const { 1384 unsigned Opc = MI->getOpcode(); 1385 1386 switch (Opc) { 1387 default: return false; 1388 case PPC::CMPWI: 1389 case PPC::CMPLWI: 1390 case PPC::CMPDI: 1391 case PPC::CMPLDI: 1392 SrcReg = MI->getOperand(1).getReg(); 1393 SrcReg2 = 0; 1394 Value = MI->getOperand(2).getImm(); 1395 Mask = 0xFFFF; 1396 return true; 1397 case PPC::CMPW: 1398 case PPC::CMPLW: 1399 case PPC::CMPD: 1400 case PPC::CMPLD: 1401 case PPC::FCMPUS: 1402 case PPC::FCMPUD: 1403 SrcReg = MI->getOperand(1).getReg(); 1404 SrcReg2 = MI->getOperand(2).getReg(); 1405 return true; 1406 } 1407 } 1408 1409 bool PPCInstrInfo::optimizeCompareInstr(MachineInstr *CmpInstr, 1410 unsigned SrcReg, unsigned SrcReg2, 1411 int Mask, int Value, 1412 const MachineRegisterInfo *MRI) const { 1413 if (DisableCmpOpt) 1414 return false; 1415 1416 int OpC = CmpInstr->getOpcode(); 1417 unsigned CRReg = CmpInstr->getOperand(0).getReg(); 1418 1419 // FP record forms set CR1 based on the execption status bits, not a 1420 // comparison with zero. 1421 if (OpC == PPC::FCMPUS || OpC == PPC::FCMPUD) 1422 return false; 1423 1424 // The record forms set the condition register based on a signed comparison 1425 // with zero (so says the ISA manual). This is not as straightforward as it 1426 // seems, however, because this is always a 64-bit comparison on PPC64, even 1427 // for instructions that are 32-bit in nature (like slw for example). 1428 // So, on PPC32, for unsigned comparisons, we can use the record forms only 1429 // for equality checks (as those don't depend on the sign). On PPC64, 1430 // we are restricted to equality for unsigned 64-bit comparisons and for 1431 // signed 32-bit comparisons the applicability is more restricted. 1432 bool isPPC64 = Subtarget.isPPC64(); 1433 bool is32BitSignedCompare = OpC == PPC::CMPWI || OpC == PPC::CMPW; 1434 bool is32BitUnsignedCompare = OpC == PPC::CMPLWI || OpC == PPC::CMPLW; 1435 bool is64BitUnsignedCompare = OpC == PPC::CMPLDI || OpC == PPC::CMPLD; 1436 1437 // Get the unique definition of SrcReg. 1438 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg); 1439 if (!MI) return false; 1440 int MIOpC = MI->getOpcode(); 1441 1442 bool equalityOnly = false; 1443 bool noSub = false; 1444 if (isPPC64) { 1445 if (is32BitSignedCompare) { 1446 // We can perform this optimization only if MI is sign-extending. 1447 if (MIOpC == PPC::SRAW || MIOpC == PPC::SRAWo || 1448 MIOpC == PPC::SRAWI || MIOpC == PPC::SRAWIo || 1449 MIOpC == PPC::EXTSB || MIOpC == PPC::EXTSBo || 1450 MIOpC == PPC::EXTSH || MIOpC == PPC::EXTSHo || 1451 MIOpC == PPC::EXTSW || MIOpC == PPC::EXTSWo) { 1452 noSub = true; 1453 } else 1454 return false; 1455 } else if (is32BitUnsignedCompare) { 1456 // We can perform this optimization, equality only, if MI is 1457 // zero-extending. 1458 if (MIOpC == PPC::CNTLZW || MIOpC == PPC::CNTLZWo || 1459 MIOpC == PPC::SLW || MIOpC == PPC::SLWo || 1460 MIOpC == PPC::SRW || MIOpC == PPC::SRWo) { 1461 noSub = true; 1462 equalityOnly = true; 1463 } else 1464 return false; 1465 } else 1466 equalityOnly = is64BitUnsignedCompare; 1467 } else 1468 equalityOnly = is32BitUnsignedCompare; 1469 1470 if (equalityOnly) { 1471 // We need to check the uses of the condition register in order to reject 1472 // non-equality comparisons. 1473 for (MachineRegisterInfo::use_instr_iterator I =MRI->use_instr_begin(CRReg), 1474 IE = MRI->use_instr_end(); I != IE; ++I) { 1475 MachineInstr *UseMI = &*I; 1476 if (UseMI->getOpcode() == PPC::BCC) { 1477 unsigned Pred = UseMI->getOperand(0).getImm(); 1478 if (Pred != PPC::PRED_EQ && Pred != PPC::PRED_NE) 1479 return false; 1480 } else if (UseMI->getOpcode() == PPC::ISEL || 1481 UseMI->getOpcode() == PPC::ISEL8) { 1482 unsigned SubIdx = UseMI->getOperand(3).getSubReg(); 1483 if (SubIdx != PPC::sub_eq) 1484 return false; 1485 } else 1486 return false; 1487 } 1488 } 1489 1490 MachineBasicBlock::iterator I = CmpInstr; 1491 1492 // Scan forward to find the first use of the compare. 1493 for (MachineBasicBlock::iterator EL = CmpInstr->getParent()->end(); 1494 I != EL; ++I) { 1495 bool FoundUse = false; 1496 for (MachineRegisterInfo::use_instr_iterator J =MRI->use_instr_begin(CRReg), 1497 JE = MRI->use_instr_end(); J != JE; ++J) 1498 if (&*J == &*I) { 1499 FoundUse = true; 1500 break; 1501 } 1502 1503 if (FoundUse) 1504 break; 1505 } 1506 1507 // There are two possible candidates which can be changed to set CR[01]. 1508 // One is MI, the other is a SUB instruction. 1509 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1). 1510 MachineInstr *Sub = nullptr; 1511 if (SrcReg2 != 0) 1512 // MI is not a candidate for CMPrr. 1513 MI = nullptr; 1514 // FIXME: Conservatively refuse to convert an instruction which isn't in the 1515 // same BB as the comparison. This is to allow the check below to avoid calls 1516 // (and other explicit clobbers); instead we should really check for these 1517 // more explicitly (in at least a few predecessors). 1518 else if (MI->getParent() != CmpInstr->getParent() || Value != 0) { 1519 // PPC does not have a record-form SUBri. 1520 return false; 1521 } 1522 1523 // Search for Sub. 1524 const TargetRegisterInfo *TRI = &getRegisterInfo(); 1525 --I; 1526 1527 // Get ready to iterate backward from CmpInstr. 1528 MachineBasicBlock::iterator E = MI, 1529 B = CmpInstr->getParent()->begin(); 1530 1531 for (; I != E && !noSub; --I) { 1532 const MachineInstr &Instr = *I; 1533 unsigned IOpC = Instr.getOpcode(); 1534 1535 if (&*I != CmpInstr && ( 1536 Instr.modifiesRegister(PPC::CR0, TRI) || 1537 Instr.readsRegister(PPC::CR0, TRI))) 1538 // This instruction modifies or uses the record condition register after 1539 // the one we want to change. While we could do this transformation, it 1540 // would likely not be profitable. This transformation removes one 1541 // instruction, and so even forcing RA to generate one move probably 1542 // makes it unprofitable. 1543 return false; 1544 1545 // Check whether CmpInstr can be made redundant by the current instruction. 1546 if ((OpC == PPC::CMPW || OpC == PPC::CMPLW || 1547 OpC == PPC::CMPD || OpC == PPC::CMPLD) && 1548 (IOpC == PPC::SUBF || IOpC == PPC::SUBF8) && 1549 ((Instr.getOperand(1).getReg() == SrcReg && 1550 Instr.getOperand(2).getReg() == SrcReg2) || 1551 (Instr.getOperand(1).getReg() == SrcReg2 && 1552 Instr.getOperand(2).getReg() == SrcReg))) { 1553 Sub = &*I; 1554 break; 1555 } 1556 1557 if (I == B) 1558 // The 'and' is below the comparison instruction. 1559 return false; 1560 } 1561 1562 // Return false if no candidates exist. 1563 if (!MI && !Sub) 1564 return false; 1565 1566 // The single candidate is called MI. 1567 if (!MI) MI = Sub; 1568 1569 int NewOpC = -1; 1570 MIOpC = MI->getOpcode(); 1571 if (MIOpC == PPC::ANDIo || MIOpC == PPC::ANDIo8) 1572 NewOpC = MIOpC; 1573 else { 1574 NewOpC = PPC::getRecordFormOpcode(MIOpC); 1575 if (NewOpC == -1 && PPC::getNonRecordFormOpcode(MIOpC) != -1) 1576 NewOpC = MIOpC; 1577 } 1578 1579 // FIXME: On the non-embedded POWER architectures, only some of the record 1580 // forms are fast, and we should use only the fast ones. 1581 1582 // The defining instruction has a record form (or is already a record 1583 // form). It is possible, however, that we'll need to reverse the condition 1584 // code of the users. 1585 if (NewOpC == -1) 1586 return false; 1587 1588 SmallVector<std::pair<MachineOperand*, PPC::Predicate>, 4> PredsToUpdate; 1589 SmallVector<std::pair<MachineOperand*, unsigned>, 4> SubRegsToUpdate; 1590 1591 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based on CMP 1592 // needs to be updated to be based on SUB. Push the condition code 1593 // operands to OperandsToUpdate. If it is safe to remove CmpInstr, the 1594 // condition code of these operands will be modified. 1595 bool ShouldSwap = false; 1596 if (Sub) { 1597 ShouldSwap = SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 && 1598 Sub->getOperand(2).getReg() == SrcReg; 1599 1600 // The operands to subf are the opposite of sub, so only in the fixed-point 1601 // case, invert the order. 1602 ShouldSwap = !ShouldSwap; 1603 } 1604 1605 if (ShouldSwap) 1606 for (MachineRegisterInfo::use_instr_iterator 1607 I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end(); 1608 I != IE; ++I) { 1609 MachineInstr *UseMI = &*I; 1610 if (UseMI->getOpcode() == PPC::BCC) { 1611 PPC::Predicate Pred = (PPC::Predicate) UseMI->getOperand(0).getImm(); 1612 assert((!equalityOnly || 1613 Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE) && 1614 "Invalid predicate for equality-only optimization"); 1615 PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)), 1616 PPC::getSwappedPredicate(Pred))); 1617 } else if (UseMI->getOpcode() == PPC::ISEL || 1618 UseMI->getOpcode() == PPC::ISEL8) { 1619 unsigned NewSubReg = UseMI->getOperand(3).getSubReg(); 1620 assert((!equalityOnly || NewSubReg == PPC::sub_eq) && 1621 "Invalid CR bit for equality-only optimization"); 1622 1623 if (NewSubReg == PPC::sub_lt) 1624 NewSubReg = PPC::sub_gt; 1625 else if (NewSubReg == PPC::sub_gt) 1626 NewSubReg = PPC::sub_lt; 1627 1628 SubRegsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(3)), 1629 NewSubReg)); 1630 } else // We need to abort on a user we don't understand. 1631 return false; 1632 } 1633 1634 // Create a new virtual register to hold the value of the CR set by the 1635 // record-form instruction. If the instruction was not previously in 1636 // record form, then set the kill flag on the CR. 1637 CmpInstr->eraseFromParent(); 1638 1639 MachineBasicBlock::iterator MII = MI; 1640 BuildMI(*MI->getParent(), std::next(MII), MI->getDebugLoc(), 1641 get(TargetOpcode::COPY), CRReg) 1642 .addReg(PPC::CR0, MIOpC != NewOpC ? RegState::Kill : 0); 1643 1644 if (MIOpC != NewOpC) { 1645 // We need to be careful here: we're replacing one instruction with 1646 // another, and we need to make sure that we get all of the right 1647 // implicit uses and defs. On the other hand, the caller may be holding 1648 // an iterator to this instruction, and so we can't delete it (this is 1649 // specifically the case if this is the instruction directly after the 1650 // compare). 1651 1652 const MCInstrDesc &NewDesc = get(NewOpC); 1653 MI->setDesc(NewDesc); 1654 1655 if (NewDesc.ImplicitDefs) 1656 for (const uint16_t *ImpDefs = NewDesc.getImplicitDefs(); 1657 *ImpDefs; ++ImpDefs) 1658 if (!MI->definesRegister(*ImpDefs)) 1659 MI->addOperand(*MI->getParent()->getParent(), 1660 MachineOperand::CreateReg(*ImpDefs, true, true)); 1661 if (NewDesc.ImplicitUses) 1662 for (const uint16_t *ImpUses = NewDesc.getImplicitUses(); 1663 *ImpUses; ++ImpUses) 1664 if (!MI->readsRegister(*ImpUses)) 1665 MI->addOperand(*MI->getParent()->getParent(), 1666 MachineOperand::CreateReg(*ImpUses, false, true)); 1667 } 1668 1669 // Modify the condition code of operands in OperandsToUpdate. 1670 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to 1671 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc. 1672 for (unsigned i = 0, e = PredsToUpdate.size(); i < e; i++) 1673 PredsToUpdate[i].first->setImm(PredsToUpdate[i].second); 1674 1675 for (unsigned i = 0, e = SubRegsToUpdate.size(); i < e; i++) 1676 SubRegsToUpdate[i].first->setSubReg(SubRegsToUpdate[i].second); 1677 1678 return true; 1679 } 1680 1681 /// GetInstSize - Return the number of bytes of code the specified 1682 /// instruction may be. This returns the maximum number of bytes. 1683 /// 1684 unsigned PPCInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const { 1685 unsigned Opcode = MI->getOpcode(); 1686 1687 if (Opcode == PPC::INLINEASM) { 1688 const MachineFunction *MF = MI->getParent()->getParent(); 1689 const char *AsmStr = MI->getOperand(0).getSymbolName(); 1690 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo()); 1691 } else if (Opcode == TargetOpcode::STACKMAP) { 1692 return MI->getOperand(1).getImm(); 1693 } else if (Opcode == TargetOpcode::PATCHPOINT) { 1694 PatchPointOpers Opers(MI); 1695 return Opers.getMetaOper(PatchPointOpers::NBytesPos).getImm(); 1696 } else { 1697 const MCInstrDesc &Desc = get(Opcode); 1698 return Desc.getSize(); 1699 } 1700 } 1701 1702