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.end(); 356 if (I == MBB.begin()) 357 return false; 358 --I; 359 while (I->isDebugValue()) { 360 if (I == MBB.begin()) 361 return false; 362 --I; 363 } 364 if (!isUnpredicatedTerminator(I)) 365 return false; 366 367 // Get the last instruction in the block. 368 MachineInstr *LastInst = I; 369 370 // If there is only one terminator instruction, process it. 371 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) { 372 if (LastInst->getOpcode() == PPC::B) { 373 if (!LastInst->getOperand(0).isMBB()) 374 return true; 375 TBB = LastInst->getOperand(0).getMBB(); 376 return false; 377 } else if (LastInst->getOpcode() == PPC::BCC) { 378 if (!LastInst->getOperand(2).isMBB()) 379 return true; 380 // Block ends with fall-through condbranch. 381 TBB = LastInst->getOperand(2).getMBB(); 382 Cond.push_back(LastInst->getOperand(0)); 383 Cond.push_back(LastInst->getOperand(1)); 384 return false; 385 } else if (LastInst->getOpcode() == PPC::BC) { 386 if (!LastInst->getOperand(1).isMBB()) 387 return true; 388 // Block ends with fall-through condbranch. 389 TBB = LastInst->getOperand(1).getMBB(); 390 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET)); 391 Cond.push_back(LastInst->getOperand(0)); 392 return false; 393 } else if (LastInst->getOpcode() == PPC::BCn) { 394 if (!LastInst->getOperand(1).isMBB()) 395 return true; 396 // Block ends with fall-through condbranch. 397 TBB = LastInst->getOperand(1).getMBB(); 398 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET)); 399 Cond.push_back(LastInst->getOperand(0)); 400 return false; 401 } else if (LastInst->getOpcode() == PPC::BDNZ8 || 402 LastInst->getOpcode() == PPC::BDNZ) { 403 if (!LastInst->getOperand(0).isMBB()) 404 return true; 405 if (DisableCTRLoopAnal) 406 return true; 407 TBB = LastInst->getOperand(0).getMBB(); 408 Cond.push_back(MachineOperand::CreateImm(1)); 409 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, 410 true)); 411 return false; 412 } else if (LastInst->getOpcode() == PPC::BDZ8 || 413 LastInst->getOpcode() == PPC::BDZ) { 414 if (!LastInst->getOperand(0).isMBB()) 415 return true; 416 if (DisableCTRLoopAnal) 417 return true; 418 TBB = LastInst->getOperand(0).getMBB(); 419 Cond.push_back(MachineOperand::CreateImm(0)); 420 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, 421 true)); 422 return false; 423 } 424 425 // Otherwise, don't know what this is. 426 return true; 427 } 428 429 // Get the instruction before it if it's a terminator. 430 MachineInstr *SecondLastInst = I; 431 432 // If there are three terminators, we don't know what sort of block this is. 433 if (SecondLastInst && I != MBB.begin() && 434 isUnpredicatedTerminator(--I)) 435 return true; 436 437 // If the block ends with PPC::B and PPC:BCC, handle it. 438 if (SecondLastInst->getOpcode() == PPC::BCC && 439 LastInst->getOpcode() == PPC::B) { 440 if (!SecondLastInst->getOperand(2).isMBB() || 441 !LastInst->getOperand(0).isMBB()) 442 return true; 443 TBB = SecondLastInst->getOperand(2).getMBB(); 444 Cond.push_back(SecondLastInst->getOperand(0)); 445 Cond.push_back(SecondLastInst->getOperand(1)); 446 FBB = LastInst->getOperand(0).getMBB(); 447 return false; 448 } else if (SecondLastInst->getOpcode() == PPC::BC && 449 LastInst->getOpcode() == PPC::B) { 450 if (!SecondLastInst->getOperand(1).isMBB() || 451 !LastInst->getOperand(0).isMBB()) 452 return true; 453 TBB = SecondLastInst->getOperand(1).getMBB(); 454 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET)); 455 Cond.push_back(SecondLastInst->getOperand(0)); 456 FBB = LastInst->getOperand(0).getMBB(); 457 return false; 458 } else if (SecondLastInst->getOpcode() == PPC::BCn && 459 LastInst->getOpcode() == PPC::B) { 460 if (!SecondLastInst->getOperand(1).isMBB() || 461 !LastInst->getOperand(0).isMBB()) 462 return true; 463 TBB = SecondLastInst->getOperand(1).getMBB(); 464 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_UNSET)); 465 Cond.push_back(SecondLastInst->getOperand(0)); 466 FBB = LastInst->getOperand(0).getMBB(); 467 return false; 468 } else if ((SecondLastInst->getOpcode() == PPC::BDNZ8 || 469 SecondLastInst->getOpcode() == PPC::BDNZ) && 470 LastInst->getOpcode() == PPC::B) { 471 if (!SecondLastInst->getOperand(0).isMBB() || 472 !LastInst->getOperand(0).isMBB()) 473 return true; 474 if (DisableCTRLoopAnal) 475 return true; 476 TBB = SecondLastInst->getOperand(0).getMBB(); 477 Cond.push_back(MachineOperand::CreateImm(1)); 478 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, 479 true)); 480 FBB = LastInst->getOperand(0).getMBB(); 481 return false; 482 } else if ((SecondLastInst->getOpcode() == PPC::BDZ8 || 483 SecondLastInst->getOpcode() == PPC::BDZ) && 484 LastInst->getOpcode() == PPC::B) { 485 if (!SecondLastInst->getOperand(0).isMBB() || 486 !LastInst->getOperand(0).isMBB()) 487 return true; 488 if (DisableCTRLoopAnal) 489 return true; 490 TBB = SecondLastInst->getOperand(0).getMBB(); 491 Cond.push_back(MachineOperand::CreateImm(0)); 492 Cond.push_back(MachineOperand::CreateReg(isPPC64 ? PPC::CTR8 : PPC::CTR, 493 true)); 494 FBB = LastInst->getOperand(0).getMBB(); 495 return false; 496 } 497 498 // If the block ends with two PPC:Bs, handle it. The second one is not 499 // executed, so remove it. 500 if (SecondLastInst->getOpcode() == PPC::B && 501 LastInst->getOpcode() == PPC::B) { 502 if (!SecondLastInst->getOperand(0).isMBB()) 503 return true; 504 TBB = SecondLastInst->getOperand(0).getMBB(); 505 I = LastInst; 506 if (AllowModify) 507 I->eraseFromParent(); 508 return false; 509 } 510 511 // Otherwise, can't handle this. 512 return true; 513 } 514 515 unsigned PPCInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { 516 MachineBasicBlock::iterator I = MBB.end(); 517 if (I == MBB.begin()) return 0; 518 --I; 519 while (I->isDebugValue()) { 520 if (I == MBB.begin()) 521 return 0; 522 --I; 523 } 524 if (I->getOpcode() != PPC::B && I->getOpcode() != PPC::BCC && 525 I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn && 526 I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ && 527 I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ) 528 return 0; 529 530 // Remove the branch. 531 I->eraseFromParent(); 532 533 I = MBB.end(); 534 535 if (I == MBB.begin()) return 1; 536 --I; 537 if (I->getOpcode() != PPC::BCC && 538 I->getOpcode() != PPC::BC && I->getOpcode() != PPC::BCn && 539 I->getOpcode() != PPC::BDNZ8 && I->getOpcode() != PPC::BDNZ && 540 I->getOpcode() != PPC::BDZ8 && I->getOpcode() != PPC::BDZ) 541 return 1; 542 543 // Remove the branch. 544 I->eraseFromParent(); 545 return 2; 546 } 547 548 unsigned 549 PPCInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, 550 MachineBasicBlock *FBB, 551 const SmallVectorImpl<MachineOperand> &Cond, 552 DebugLoc DL) const { 553 // Shouldn't be a fall through. 554 assert(TBB && "InsertBranch must not be told to insert a fallthrough"); 555 assert((Cond.size() == 2 || Cond.size() == 0) && 556 "PPC branch conditions have two components!"); 557 558 bool isPPC64 = Subtarget.isPPC64(); 559 560 // One-way branch. 561 if (!FBB) { 562 if (Cond.empty()) // Unconditional branch 563 BuildMI(&MBB, DL, get(PPC::B)).addMBB(TBB); 564 else if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) 565 BuildMI(&MBB, DL, get(Cond[0].getImm() ? 566 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : 567 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB); 568 else if (Cond[0].getImm() == PPC::PRED_BIT_SET) 569 BuildMI(&MBB, DL, get(PPC::BC)).addOperand(Cond[1]).addMBB(TBB); 570 else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET) 571 BuildMI(&MBB, DL, get(PPC::BCn)).addOperand(Cond[1]).addMBB(TBB); 572 else // Conditional branch 573 BuildMI(&MBB, DL, get(PPC::BCC)) 574 .addImm(Cond[0].getImm()).addOperand(Cond[1]).addMBB(TBB); 575 return 1; 576 } 577 578 // Two-way Conditional Branch. 579 if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) 580 BuildMI(&MBB, DL, get(Cond[0].getImm() ? 581 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : 582 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))).addMBB(TBB); 583 else if (Cond[0].getImm() == PPC::PRED_BIT_SET) 584 BuildMI(&MBB, DL, get(PPC::BC)).addOperand(Cond[1]).addMBB(TBB); 585 else if (Cond[0].getImm() == PPC::PRED_BIT_UNSET) 586 BuildMI(&MBB, DL, get(PPC::BCn)).addOperand(Cond[1]).addMBB(TBB); 587 else 588 BuildMI(&MBB, DL, get(PPC::BCC)) 589 .addImm(Cond[0].getImm()).addOperand(Cond[1]).addMBB(TBB); 590 BuildMI(&MBB, DL, get(PPC::B)).addMBB(FBB); 591 return 2; 592 } 593 594 // Select analysis. 595 bool PPCInstrInfo::canInsertSelect(const MachineBasicBlock &MBB, 596 const SmallVectorImpl<MachineOperand> &Cond, 597 unsigned TrueReg, unsigned FalseReg, 598 int &CondCycles, int &TrueCycles, int &FalseCycles) const { 599 if (!Subtarget.hasISEL()) 600 return false; 601 602 if (Cond.size() != 2) 603 return false; 604 605 // If this is really a bdnz-like condition, then it cannot be turned into a 606 // select. 607 if (Cond[1].getReg() == PPC::CTR || Cond[1].getReg() == PPC::CTR8) 608 return false; 609 610 // Check register classes. 611 const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 612 const TargetRegisterClass *RC = 613 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 614 if (!RC) 615 return false; 616 617 // isel is for regular integer GPRs only. 618 if (!PPC::GPRCRegClass.hasSubClassEq(RC) && 619 !PPC::GPRC_NOR0RegClass.hasSubClassEq(RC) && 620 !PPC::G8RCRegClass.hasSubClassEq(RC) && 621 !PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) 622 return false; 623 624 // FIXME: These numbers are for the A2, how well they work for other cores is 625 // an open question. On the A2, the isel instruction has a 2-cycle latency 626 // but single-cycle throughput. These numbers are used in combination with 627 // the MispredictPenalty setting from the active SchedMachineModel. 628 CondCycles = 1; 629 TrueCycles = 1; 630 FalseCycles = 1; 631 632 return true; 633 } 634 635 void PPCInstrInfo::insertSelect(MachineBasicBlock &MBB, 636 MachineBasicBlock::iterator MI, DebugLoc dl, 637 unsigned DestReg, 638 const SmallVectorImpl<MachineOperand> &Cond, 639 unsigned TrueReg, unsigned FalseReg) const { 640 assert(Cond.size() == 2 && 641 "PPC branch conditions have two components!"); 642 643 assert(Subtarget.hasISEL() && 644 "Cannot insert select on target without ISEL support"); 645 646 // Get the register classes. 647 MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); 648 const TargetRegisterClass *RC = 649 RI.getCommonSubClass(MRI.getRegClass(TrueReg), MRI.getRegClass(FalseReg)); 650 assert(RC && "TrueReg and FalseReg must have overlapping register classes"); 651 652 bool Is64Bit = PPC::G8RCRegClass.hasSubClassEq(RC) || 653 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC); 654 assert((Is64Bit || 655 PPC::GPRCRegClass.hasSubClassEq(RC) || 656 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) && 657 "isel is for regular integer GPRs only"); 658 659 unsigned OpCode = Is64Bit ? PPC::ISEL8 : PPC::ISEL; 660 unsigned SelectPred = Cond[0].getImm(); 661 662 unsigned SubIdx; 663 bool SwapOps; 664 switch (SelectPred) { 665 default: llvm_unreachable("invalid predicate for isel"); 666 case PPC::PRED_EQ: SubIdx = PPC::sub_eq; SwapOps = false; break; 667 case PPC::PRED_NE: SubIdx = PPC::sub_eq; SwapOps = true; break; 668 case PPC::PRED_LT: SubIdx = PPC::sub_lt; SwapOps = false; break; 669 case PPC::PRED_GE: SubIdx = PPC::sub_lt; SwapOps = true; break; 670 case PPC::PRED_GT: SubIdx = PPC::sub_gt; SwapOps = false; break; 671 case PPC::PRED_LE: SubIdx = PPC::sub_gt; SwapOps = true; break; 672 case PPC::PRED_UN: SubIdx = PPC::sub_un; SwapOps = false; break; 673 case PPC::PRED_NU: SubIdx = PPC::sub_un; SwapOps = true; break; 674 case PPC::PRED_BIT_SET: SubIdx = 0; SwapOps = false; break; 675 case PPC::PRED_BIT_UNSET: SubIdx = 0; SwapOps = true; break; 676 } 677 678 unsigned FirstReg = SwapOps ? FalseReg : TrueReg, 679 SecondReg = SwapOps ? TrueReg : FalseReg; 680 681 // The first input register of isel cannot be r0. If it is a member 682 // of a register class that can be r0, then copy it first (the 683 // register allocator should eliminate the copy). 684 if (MRI.getRegClass(FirstReg)->contains(PPC::R0) || 685 MRI.getRegClass(FirstReg)->contains(PPC::X0)) { 686 const TargetRegisterClass *FirstRC = 687 MRI.getRegClass(FirstReg)->contains(PPC::X0) ? 688 &PPC::G8RC_NOX0RegClass : &PPC::GPRC_NOR0RegClass; 689 unsigned OldFirstReg = FirstReg; 690 FirstReg = MRI.createVirtualRegister(FirstRC); 691 BuildMI(MBB, MI, dl, get(TargetOpcode::COPY), FirstReg) 692 .addReg(OldFirstReg); 693 } 694 695 BuildMI(MBB, MI, dl, get(OpCode), DestReg) 696 .addReg(FirstReg).addReg(SecondReg) 697 .addReg(Cond[1].getReg(), 0, SubIdx); 698 } 699 700 static unsigned getCRBitValue(unsigned CRBit) { 701 unsigned Ret = 4; 702 if (CRBit == PPC::CR0LT || CRBit == PPC::CR1LT || 703 CRBit == PPC::CR2LT || CRBit == PPC::CR3LT || 704 CRBit == PPC::CR4LT || CRBit == PPC::CR5LT || 705 CRBit == PPC::CR6LT || CRBit == PPC::CR7LT) 706 Ret = 3; 707 if (CRBit == PPC::CR0GT || CRBit == PPC::CR1GT || 708 CRBit == PPC::CR2GT || CRBit == PPC::CR3GT || 709 CRBit == PPC::CR4GT || CRBit == PPC::CR5GT || 710 CRBit == PPC::CR6GT || CRBit == PPC::CR7GT) 711 Ret = 2; 712 if (CRBit == PPC::CR0EQ || CRBit == PPC::CR1EQ || 713 CRBit == PPC::CR2EQ || CRBit == PPC::CR3EQ || 714 CRBit == PPC::CR4EQ || CRBit == PPC::CR5EQ || 715 CRBit == PPC::CR6EQ || CRBit == PPC::CR7EQ) 716 Ret = 1; 717 if (CRBit == PPC::CR0UN || CRBit == PPC::CR1UN || 718 CRBit == PPC::CR2UN || CRBit == PPC::CR3UN || 719 CRBit == PPC::CR4UN || CRBit == PPC::CR5UN || 720 CRBit == PPC::CR6UN || CRBit == PPC::CR7UN) 721 Ret = 0; 722 723 assert(Ret != 4 && "Invalid CR bit register"); 724 return Ret; 725 } 726 727 void PPCInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 728 MachineBasicBlock::iterator I, DebugLoc DL, 729 unsigned DestReg, unsigned SrcReg, 730 bool KillSrc) const { 731 // We can end up with self copies and similar things as a result of VSX copy 732 // legalization. Promote them here. 733 const TargetRegisterInfo *TRI = &getRegisterInfo(); 734 if (PPC::F8RCRegClass.contains(DestReg) && 735 PPC::VSRCRegClass.contains(SrcReg)) { 736 unsigned SuperReg = 737 TRI->getMatchingSuperReg(DestReg, PPC::sub_64, &PPC::VSRCRegClass); 738 739 if (VSXSelfCopyCrash && SrcReg == SuperReg) 740 llvm_unreachable("nop VSX copy"); 741 742 DestReg = SuperReg; 743 } else if (PPC::VRRCRegClass.contains(DestReg) && 744 PPC::VSRCRegClass.contains(SrcReg)) { 745 unsigned SuperReg = 746 TRI->getMatchingSuperReg(DestReg, PPC::sub_128, &PPC::VSRCRegClass); 747 748 if (VSXSelfCopyCrash && SrcReg == SuperReg) 749 llvm_unreachable("nop VSX copy"); 750 751 DestReg = SuperReg; 752 } else if (PPC::F8RCRegClass.contains(SrcReg) && 753 PPC::VSRCRegClass.contains(DestReg)) { 754 unsigned SuperReg = 755 TRI->getMatchingSuperReg(SrcReg, PPC::sub_64, &PPC::VSRCRegClass); 756 757 if (VSXSelfCopyCrash && DestReg == SuperReg) 758 llvm_unreachable("nop VSX copy"); 759 760 SrcReg = SuperReg; 761 } else if (PPC::VRRCRegClass.contains(SrcReg) && 762 PPC::VSRCRegClass.contains(DestReg)) { 763 unsigned SuperReg = 764 TRI->getMatchingSuperReg(SrcReg, PPC::sub_128, &PPC::VSRCRegClass); 765 766 if (VSXSelfCopyCrash && DestReg == SuperReg) 767 llvm_unreachable("nop VSX copy"); 768 769 SrcReg = SuperReg; 770 } 771 772 // Different class register copy 773 if (PPC::CRBITRCRegClass.contains(SrcReg) && 774 PPC::GPRCRegClass.contains(DestReg)) { 775 unsigned CRReg = getCRFromCRBit(SrcReg); 776 BuildMI(MBB, I, DL, get(PPC::MFOCRF), DestReg) 777 .addReg(CRReg), getKillRegState(KillSrc); 778 // Rotate the CR bit in the CR fields to be the least significant bit and 779 // then mask with 0x1 (MB = ME = 31). 780 BuildMI(MBB, I, DL, get(PPC::RLWINM), DestReg) 781 .addReg(DestReg, RegState::Kill) 782 .addImm(TRI->getEncodingValue(CRReg) * 4 + (4 - getCRBitValue(SrcReg))) 783 .addImm(31) 784 .addImm(31); 785 return; 786 } else if (PPC::CRRCRegClass.contains(SrcReg) && 787 PPC::G8RCRegClass.contains(DestReg)) { 788 BuildMI(MBB, I, DL, get(PPC::MFOCRF8), DestReg) 789 .addReg(SrcReg), getKillRegState(KillSrc); 790 return; 791 } else if (PPC::CRRCRegClass.contains(SrcReg) && 792 PPC::GPRCRegClass.contains(DestReg)) { 793 BuildMI(MBB, I, DL, get(PPC::MFOCRF), DestReg) 794 .addReg(SrcReg), getKillRegState(KillSrc); 795 return; 796 } 797 798 unsigned Opc; 799 if (PPC::GPRCRegClass.contains(DestReg, SrcReg)) 800 Opc = PPC::OR; 801 else if (PPC::G8RCRegClass.contains(DestReg, SrcReg)) 802 Opc = PPC::OR8; 803 else if (PPC::F4RCRegClass.contains(DestReg, SrcReg)) 804 Opc = PPC::FMR; 805 else if (PPC::CRRCRegClass.contains(DestReg, SrcReg)) 806 Opc = PPC::MCRF; 807 else if (PPC::VRRCRegClass.contains(DestReg, SrcReg)) 808 Opc = PPC::VOR; 809 else if (PPC::VSRCRegClass.contains(DestReg, SrcReg)) 810 // There are two different ways this can be done: 811 // 1. xxlor : This has lower latency (on the P7), 2 cycles, but can only 812 // issue in VSU pipeline 0. 813 // 2. xmovdp/xmovsp: This has higher latency (on the P7), 6 cycles, but 814 // can go to either pipeline. 815 // We'll always use xxlor here, because in practically all cases where 816 // copies are generated, they are close enough to some use that the 817 // lower-latency form is preferable. 818 Opc = PPC::XXLOR; 819 else if (PPC::VSFRCRegClass.contains(DestReg, SrcReg) || 820 PPC::VSSRCRegClass.contains(DestReg, SrcReg)) 821 Opc = PPC::XXLORf; 822 else if (PPC::QFRCRegClass.contains(DestReg, SrcReg)) 823 Opc = PPC::QVFMR; 824 else if (PPC::QSRCRegClass.contains(DestReg, SrcReg)) 825 Opc = PPC::QVFMRs; 826 else if (PPC::QBRCRegClass.contains(DestReg, SrcReg)) 827 Opc = PPC::QVFMRb; 828 else if (PPC::CRBITRCRegClass.contains(DestReg, SrcReg)) 829 Opc = PPC::CROR; 830 else 831 llvm_unreachable("Impossible reg-to-reg copy"); 832 833 const MCInstrDesc &MCID = get(Opc); 834 if (MCID.getNumOperands() == 3) 835 BuildMI(MBB, I, DL, MCID, DestReg) 836 .addReg(SrcReg).addReg(SrcReg, getKillRegState(KillSrc)); 837 else 838 BuildMI(MBB, I, DL, MCID, DestReg).addReg(SrcReg, getKillRegState(KillSrc)); 839 } 840 841 // This function returns true if a CR spill is necessary and false otherwise. 842 bool 843 PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF, 844 unsigned SrcReg, bool isKill, 845 int FrameIdx, 846 const TargetRegisterClass *RC, 847 SmallVectorImpl<MachineInstr*> &NewMIs, 848 bool &NonRI, bool &SpillsVRS) const{ 849 // Note: If additional store instructions are added here, 850 // update isStoreToStackSlot. 851 852 DebugLoc DL; 853 if (PPC::GPRCRegClass.hasSubClassEq(RC) || 854 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) { 855 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STW)) 856 .addReg(SrcReg, 857 getKillRegState(isKill)), 858 FrameIdx)); 859 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) || 860 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) { 861 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STD)) 862 .addReg(SrcReg, 863 getKillRegState(isKill)), 864 FrameIdx)); 865 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) { 866 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFD)) 867 .addReg(SrcReg, 868 getKillRegState(isKill)), 869 FrameIdx)); 870 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) { 871 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFS)) 872 .addReg(SrcReg, 873 getKillRegState(isKill)), 874 FrameIdx)); 875 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) { 876 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CR)) 877 .addReg(SrcReg, 878 getKillRegState(isKill)), 879 FrameIdx)); 880 return true; 881 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) { 882 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CRBIT)) 883 .addReg(SrcReg, 884 getKillRegState(isKill)), 885 FrameIdx)); 886 return true; 887 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) { 888 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STVX)) 889 .addReg(SrcReg, 890 getKillRegState(isKill)), 891 FrameIdx)); 892 NonRI = true; 893 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) { 894 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXVD2X)) 895 .addReg(SrcReg, 896 getKillRegState(isKill)), 897 FrameIdx)); 898 NonRI = true; 899 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) { 900 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXSDX)) 901 .addReg(SrcReg, 902 getKillRegState(isKill)), 903 FrameIdx)); 904 NonRI = true; 905 } else if (PPC::VSSRCRegClass.hasSubClassEq(RC)) { 906 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXSSPX)) 907 .addReg(SrcReg, 908 getKillRegState(isKill)), 909 FrameIdx)); 910 NonRI = true; 911 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) { 912 assert(Subtarget.isDarwin() && 913 "VRSAVE only needs spill/restore on Darwin"); 914 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_VRSAVE)) 915 .addReg(SrcReg, 916 getKillRegState(isKill)), 917 FrameIdx)); 918 SpillsVRS = true; 919 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) { 920 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFDX)) 921 .addReg(SrcReg, 922 getKillRegState(isKill)), 923 FrameIdx)); 924 NonRI = true; 925 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) { 926 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFSXs)) 927 .addReg(SrcReg, 928 getKillRegState(isKill)), 929 FrameIdx)); 930 NonRI = true; 931 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) { 932 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFDXb)) 933 .addReg(SrcReg, 934 getKillRegState(isKill)), 935 FrameIdx)); 936 NonRI = true; 937 } else { 938 llvm_unreachable("Unknown regclass!"); 939 } 940 941 return false; 942 } 943 944 void 945 PPCInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 946 MachineBasicBlock::iterator MI, 947 unsigned SrcReg, bool isKill, int FrameIdx, 948 const TargetRegisterClass *RC, 949 const TargetRegisterInfo *TRI) const { 950 MachineFunction &MF = *MBB.getParent(); 951 SmallVector<MachineInstr*, 4> NewMIs; 952 953 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); 954 FuncInfo->setHasSpills(); 955 956 bool NonRI = false, SpillsVRS = false; 957 if (StoreRegToStackSlot(MF, SrcReg, isKill, FrameIdx, RC, NewMIs, 958 NonRI, SpillsVRS)) 959 FuncInfo->setSpillsCR(); 960 961 if (SpillsVRS) 962 FuncInfo->setSpillsVRSAVE(); 963 964 if (NonRI) 965 FuncInfo->setHasNonRISpills(); 966 967 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i) 968 MBB.insert(MI, NewMIs[i]); 969 970 const MachineFrameInfo &MFI = *MF.getFrameInfo(); 971 MachineMemOperand *MMO = 972 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx), 973 MachineMemOperand::MOStore, 974 MFI.getObjectSize(FrameIdx), 975 MFI.getObjectAlignment(FrameIdx)); 976 NewMIs.back()->addMemOperand(MF, MMO); 977 } 978 979 bool 980 PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL, 981 unsigned DestReg, int FrameIdx, 982 const TargetRegisterClass *RC, 983 SmallVectorImpl<MachineInstr*> &NewMIs, 984 bool &NonRI, bool &SpillsVRS) const{ 985 // Note: If additional load instructions are added here, 986 // update isLoadFromStackSlot. 987 988 if (PPC::GPRCRegClass.hasSubClassEq(RC) || 989 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) { 990 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LWZ), 991 DestReg), FrameIdx)); 992 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) || 993 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) { 994 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LD), DestReg), 995 FrameIdx)); 996 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) { 997 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFD), DestReg), 998 FrameIdx)); 999 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) { 1000 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFS), DestReg), 1001 FrameIdx)); 1002 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) { 1003 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, 1004 get(PPC::RESTORE_CR), DestReg), 1005 FrameIdx)); 1006 return true; 1007 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) { 1008 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, 1009 get(PPC::RESTORE_CRBIT), DestReg), 1010 FrameIdx)); 1011 return true; 1012 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) { 1013 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LVX), DestReg), 1014 FrameIdx)); 1015 NonRI = true; 1016 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) { 1017 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXVD2X), DestReg), 1018 FrameIdx)); 1019 NonRI = true; 1020 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) { 1021 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXSDX), DestReg), 1022 FrameIdx)); 1023 NonRI = true; 1024 } else if (PPC::VSSRCRegClass.hasSubClassEq(RC)) { 1025 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXSSPX), DestReg), 1026 FrameIdx)); 1027 NonRI = true; 1028 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) { 1029 assert(Subtarget.isDarwin() && 1030 "VRSAVE only needs spill/restore on Darwin"); 1031 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, 1032 get(PPC::RESTORE_VRSAVE), 1033 DestReg), 1034 FrameIdx)); 1035 SpillsVRS = true; 1036 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) { 1037 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFDX), DestReg), 1038 FrameIdx)); 1039 NonRI = true; 1040 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) { 1041 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFSXs), DestReg), 1042 FrameIdx)); 1043 NonRI = true; 1044 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) { 1045 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFDXb), DestReg), 1046 FrameIdx)); 1047 NonRI = true; 1048 } else { 1049 llvm_unreachable("Unknown regclass!"); 1050 } 1051 1052 return false; 1053 } 1054 1055 void 1056 PPCInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, 1057 MachineBasicBlock::iterator MI, 1058 unsigned DestReg, int FrameIdx, 1059 const TargetRegisterClass *RC, 1060 const TargetRegisterInfo *TRI) const { 1061 MachineFunction &MF = *MBB.getParent(); 1062 SmallVector<MachineInstr*, 4> NewMIs; 1063 DebugLoc DL; 1064 if (MI != MBB.end()) DL = MI->getDebugLoc(); 1065 1066 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); 1067 FuncInfo->setHasSpills(); 1068 1069 bool NonRI = false, SpillsVRS = false; 1070 if (LoadRegFromStackSlot(MF, DL, DestReg, FrameIdx, RC, NewMIs, 1071 NonRI, SpillsVRS)) 1072 FuncInfo->setSpillsCR(); 1073 1074 if (SpillsVRS) 1075 FuncInfo->setSpillsVRSAVE(); 1076 1077 if (NonRI) 1078 FuncInfo->setHasNonRISpills(); 1079 1080 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i) 1081 MBB.insert(MI, NewMIs[i]); 1082 1083 const MachineFrameInfo &MFI = *MF.getFrameInfo(); 1084 MachineMemOperand *MMO = 1085 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx), 1086 MachineMemOperand::MOLoad, 1087 MFI.getObjectSize(FrameIdx), 1088 MFI.getObjectAlignment(FrameIdx)); 1089 NewMIs.back()->addMemOperand(MF, MMO); 1090 } 1091 1092 bool PPCInstrInfo:: 1093 ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 1094 assert(Cond.size() == 2 && "Invalid PPC branch opcode!"); 1095 if (Cond[1].getReg() == PPC::CTR8 || Cond[1].getReg() == PPC::CTR) 1096 Cond[0].setImm(Cond[0].getImm() == 0 ? 1 : 0); 1097 else 1098 // Leave the CR# the same, but invert the condition. 1099 Cond[0].setImm(PPC::InvertPredicate((PPC::Predicate)Cond[0].getImm())); 1100 return false; 1101 } 1102 1103 bool PPCInstrInfo::FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI, 1104 unsigned Reg, MachineRegisterInfo *MRI) const { 1105 // For some instructions, it is legal to fold ZERO into the RA register field. 1106 // A zero immediate should always be loaded with a single li. 1107 unsigned DefOpc = DefMI->getOpcode(); 1108 if (DefOpc != PPC::LI && DefOpc != PPC::LI8) 1109 return false; 1110 if (!DefMI->getOperand(1).isImm()) 1111 return false; 1112 if (DefMI->getOperand(1).getImm() != 0) 1113 return false; 1114 1115 // Note that we cannot here invert the arguments of an isel in order to fold 1116 // a ZERO into what is presented as the second argument. All we have here 1117 // is the condition bit, and that might come from a CR-logical bit operation. 1118 1119 const MCInstrDesc &UseMCID = UseMI->getDesc(); 1120 1121 // Only fold into real machine instructions. 1122 if (UseMCID.isPseudo()) 1123 return false; 1124 1125 unsigned UseIdx; 1126 for (UseIdx = 0; UseIdx < UseMI->getNumOperands(); ++UseIdx) 1127 if (UseMI->getOperand(UseIdx).isReg() && 1128 UseMI->getOperand(UseIdx).getReg() == Reg) 1129 break; 1130 1131 assert(UseIdx < UseMI->getNumOperands() && "Cannot find Reg in UseMI"); 1132 assert(UseIdx < UseMCID.getNumOperands() && "No operand description for Reg"); 1133 1134 const MCOperandInfo *UseInfo = &UseMCID.OpInfo[UseIdx]; 1135 1136 // We can fold the zero if this register requires a GPRC_NOR0/G8RC_NOX0 1137 // register (which might also be specified as a pointer class kind). 1138 if (UseInfo->isLookupPtrRegClass()) { 1139 if (UseInfo->RegClass /* Kind */ != 1) 1140 return false; 1141 } else { 1142 if (UseInfo->RegClass != PPC::GPRC_NOR0RegClassID && 1143 UseInfo->RegClass != PPC::G8RC_NOX0RegClassID) 1144 return false; 1145 } 1146 1147 // Make sure this is not tied to an output register (or otherwise 1148 // constrained). This is true for ST?UX registers, for example, which 1149 // are tied to their output registers. 1150 if (UseInfo->Constraints != 0) 1151 return false; 1152 1153 unsigned ZeroReg; 1154 if (UseInfo->isLookupPtrRegClass()) { 1155 bool isPPC64 = Subtarget.isPPC64(); 1156 ZeroReg = isPPC64 ? PPC::ZERO8 : PPC::ZERO; 1157 } else { 1158 ZeroReg = UseInfo->RegClass == PPC::G8RC_NOX0RegClassID ? 1159 PPC::ZERO8 : PPC::ZERO; 1160 } 1161 1162 bool DeleteDef = MRI->hasOneNonDBGUse(Reg); 1163 UseMI->getOperand(UseIdx).setReg(ZeroReg); 1164 1165 if (DeleteDef) 1166 DefMI->eraseFromParent(); 1167 1168 return true; 1169 } 1170 1171 static bool MBBDefinesCTR(MachineBasicBlock &MBB) { 1172 for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end(); 1173 I != IE; ++I) 1174 if (I->definesRegister(PPC::CTR) || I->definesRegister(PPC::CTR8)) 1175 return true; 1176 return false; 1177 } 1178 1179 // We should make sure that, if we're going to predicate both sides of a 1180 // condition (a diamond), that both sides don't define the counter register. We 1181 // can predicate counter-decrement-based branches, but while that predicates 1182 // the branching, it does not predicate the counter decrement. If we tried to 1183 // merge the triangle into one predicated block, we'd decrement the counter 1184 // twice. 1185 bool PPCInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB, 1186 unsigned NumT, unsigned ExtraT, 1187 MachineBasicBlock &FMBB, 1188 unsigned NumF, unsigned ExtraF, 1189 const BranchProbability &Probability) const { 1190 return !(MBBDefinesCTR(TMBB) && MBBDefinesCTR(FMBB)); 1191 } 1192 1193 1194 bool PPCInstrInfo::isPredicated(const MachineInstr *MI) const { 1195 // The predicated branches are identified by their type, not really by the 1196 // explicit presence of a predicate. Furthermore, some of them can be 1197 // predicated more than once. Because if conversion won't try to predicate 1198 // any instruction which already claims to be predicated (by returning true 1199 // here), always return false. In doing so, we let isPredicable() be the 1200 // final word on whether not the instruction can be (further) predicated. 1201 1202 return false; 1203 } 1204 1205 bool PPCInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const { 1206 if (!MI->isTerminator()) 1207 return false; 1208 1209 // Conditional branch is a special case. 1210 if (MI->isBranch() && !MI->isBarrier()) 1211 return true; 1212 1213 return !isPredicated(MI); 1214 } 1215 1216 bool PPCInstrInfo::PredicateInstruction( 1217 MachineInstr *MI, 1218 const SmallVectorImpl<MachineOperand> &Pred) const { 1219 unsigned OpC = MI->getOpcode(); 1220 if (OpC == PPC::BLR || OpC == PPC::BLR8) { 1221 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) { 1222 bool isPPC64 = Subtarget.isPPC64(); 1223 MI->setDesc(get(Pred[0].getImm() ? 1224 (isPPC64 ? PPC::BDNZLR8 : PPC::BDNZLR) : 1225 (isPPC64 ? PPC::BDZLR8 : PPC::BDZLR))); 1226 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) { 1227 MI->setDesc(get(PPC::BCLR)); 1228 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1229 .addReg(Pred[1].getReg()); 1230 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { 1231 MI->setDesc(get(PPC::BCLRn)); 1232 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1233 .addReg(Pred[1].getReg()); 1234 } else { 1235 MI->setDesc(get(PPC::BCCLR)); 1236 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1237 .addImm(Pred[0].getImm()) 1238 .addReg(Pred[1].getReg()); 1239 } 1240 1241 return true; 1242 } else if (OpC == PPC::B) { 1243 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) { 1244 bool isPPC64 = Subtarget.isPPC64(); 1245 MI->setDesc(get(Pred[0].getImm() ? 1246 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : 1247 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))); 1248 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) { 1249 MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); 1250 MI->RemoveOperand(0); 1251 1252 MI->setDesc(get(PPC::BC)); 1253 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1254 .addReg(Pred[1].getReg()) 1255 .addMBB(MBB); 1256 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { 1257 MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); 1258 MI->RemoveOperand(0); 1259 1260 MI->setDesc(get(PPC::BCn)); 1261 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1262 .addReg(Pred[1].getReg()) 1263 .addMBB(MBB); 1264 } else { 1265 MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); 1266 MI->RemoveOperand(0); 1267 1268 MI->setDesc(get(PPC::BCC)); 1269 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1270 .addImm(Pred[0].getImm()) 1271 .addReg(Pred[1].getReg()) 1272 .addMBB(MBB); 1273 } 1274 1275 return true; 1276 } else if (OpC == PPC::BCTR || OpC == PPC::BCTR8 || 1277 OpC == PPC::BCTRL || OpC == PPC::BCTRL8) { 1278 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) 1279 llvm_unreachable("Cannot predicate bctr[l] on the ctr register"); 1280 1281 bool setLR = OpC == PPC::BCTRL || OpC == PPC::BCTRL8; 1282 bool isPPC64 = Subtarget.isPPC64(); 1283 1284 if (Pred[0].getImm() == PPC::PRED_BIT_SET) { 1285 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8) : 1286 (setLR ? PPC::BCCTRL : PPC::BCCTR))); 1287 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1288 .addReg(Pred[1].getReg()); 1289 return true; 1290 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { 1291 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8n : PPC::BCCTR8n) : 1292 (setLR ? PPC::BCCTRLn : PPC::BCCTRn))); 1293 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1294 .addReg(Pred[1].getReg()); 1295 return true; 1296 } 1297 1298 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCCTRL8 : PPC::BCCCTR8) : 1299 (setLR ? PPC::BCCCTRL : PPC::BCCCTR))); 1300 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1301 .addImm(Pred[0].getImm()) 1302 .addReg(Pred[1].getReg()); 1303 return true; 1304 } 1305 1306 return false; 1307 } 1308 1309 bool PPCInstrInfo::SubsumesPredicate( 1310 const SmallVectorImpl<MachineOperand> &Pred1, 1311 const SmallVectorImpl<MachineOperand> &Pred2) const { 1312 assert(Pred1.size() == 2 && "Invalid PPC first predicate"); 1313 assert(Pred2.size() == 2 && "Invalid PPC second predicate"); 1314 1315 if (Pred1[1].getReg() == PPC::CTR8 || Pred1[1].getReg() == PPC::CTR) 1316 return false; 1317 if (Pred2[1].getReg() == PPC::CTR8 || Pred2[1].getReg() == PPC::CTR) 1318 return false; 1319 1320 // P1 can only subsume P2 if they test the same condition register. 1321 if (Pred1[1].getReg() != Pred2[1].getReg()) 1322 return false; 1323 1324 PPC::Predicate P1 = (PPC::Predicate) Pred1[0].getImm(); 1325 PPC::Predicate P2 = (PPC::Predicate) Pred2[0].getImm(); 1326 1327 if (P1 == P2) 1328 return true; 1329 1330 // Does P1 subsume P2, e.g. GE subsumes GT. 1331 if (P1 == PPC::PRED_LE && 1332 (P2 == PPC::PRED_LT || P2 == PPC::PRED_EQ)) 1333 return true; 1334 if (P1 == PPC::PRED_GE && 1335 (P2 == PPC::PRED_GT || P2 == PPC::PRED_EQ)) 1336 return true; 1337 1338 return false; 1339 } 1340 1341 bool PPCInstrInfo::DefinesPredicate(MachineInstr *MI, 1342 std::vector<MachineOperand> &Pred) const { 1343 // Note: At the present time, the contents of Pred from this function is 1344 // unused by IfConversion. This implementation follows ARM by pushing the 1345 // CR-defining operand. Because the 'DZ' and 'DNZ' count as types of 1346 // predicate, instructions defining CTR or CTR8 are also included as 1347 // predicate-defining instructions. 1348 1349 const TargetRegisterClass *RCs[] = 1350 { &PPC::CRRCRegClass, &PPC::CRBITRCRegClass, 1351 &PPC::CTRRCRegClass, &PPC::CTRRC8RegClass }; 1352 1353 bool Found = false; 1354 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 1355 const MachineOperand &MO = MI->getOperand(i); 1356 for (unsigned c = 0; c < array_lengthof(RCs) && !Found; ++c) { 1357 const TargetRegisterClass *RC = RCs[c]; 1358 if (MO.isReg()) { 1359 if (MO.isDef() && RC->contains(MO.getReg())) { 1360 Pred.push_back(MO); 1361 Found = true; 1362 } 1363 } else if (MO.isRegMask()) { 1364 for (TargetRegisterClass::iterator I = RC->begin(), 1365 IE = RC->end(); I != IE; ++I) 1366 if (MO.clobbersPhysReg(*I)) { 1367 Pred.push_back(MO); 1368 Found = true; 1369 } 1370 } 1371 } 1372 } 1373 1374 return Found; 1375 } 1376 1377 bool PPCInstrInfo::isPredicable(MachineInstr *MI) const { 1378 unsigned OpC = MI->getOpcode(); 1379 switch (OpC) { 1380 default: 1381 return false; 1382 case PPC::B: 1383 case PPC::BLR: 1384 case PPC::BLR8: 1385 case PPC::BCTR: 1386 case PPC::BCTR8: 1387 case PPC::BCTRL: 1388 case PPC::BCTRL8: 1389 return true; 1390 } 1391 } 1392 1393 bool PPCInstrInfo::analyzeCompare(const MachineInstr *MI, 1394 unsigned &SrcReg, unsigned &SrcReg2, 1395 int &Mask, int &Value) const { 1396 unsigned Opc = MI->getOpcode(); 1397 1398 switch (Opc) { 1399 default: return false; 1400 case PPC::CMPWI: 1401 case PPC::CMPLWI: 1402 case PPC::CMPDI: 1403 case PPC::CMPLDI: 1404 SrcReg = MI->getOperand(1).getReg(); 1405 SrcReg2 = 0; 1406 Value = MI->getOperand(2).getImm(); 1407 Mask = 0xFFFF; 1408 return true; 1409 case PPC::CMPW: 1410 case PPC::CMPLW: 1411 case PPC::CMPD: 1412 case PPC::CMPLD: 1413 case PPC::FCMPUS: 1414 case PPC::FCMPUD: 1415 SrcReg = MI->getOperand(1).getReg(); 1416 SrcReg2 = MI->getOperand(2).getReg(); 1417 return true; 1418 } 1419 } 1420 1421 bool PPCInstrInfo::optimizeCompareInstr(MachineInstr *CmpInstr, 1422 unsigned SrcReg, unsigned SrcReg2, 1423 int Mask, int Value, 1424 const MachineRegisterInfo *MRI) const { 1425 if (DisableCmpOpt) 1426 return false; 1427 1428 int OpC = CmpInstr->getOpcode(); 1429 unsigned CRReg = CmpInstr->getOperand(0).getReg(); 1430 1431 // FP record forms set CR1 based on the execption status bits, not a 1432 // comparison with zero. 1433 if (OpC == PPC::FCMPUS || OpC == PPC::FCMPUD) 1434 return false; 1435 1436 // The record forms set the condition register based on a signed comparison 1437 // with zero (so says the ISA manual). This is not as straightforward as it 1438 // seems, however, because this is always a 64-bit comparison on PPC64, even 1439 // for instructions that are 32-bit in nature (like slw for example). 1440 // So, on PPC32, for unsigned comparisons, we can use the record forms only 1441 // for equality checks (as those don't depend on the sign). On PPC64, 1442 // we are restricted to equality for unsigned 64-bit comparisons and for 1443 // signed 32-bit comparisons the applicability is more restricted. 1444 bool isPPC64 = Subtarget.isPPC64(); 1445 bool is32BitSignedCompare = OpC == PPC::CMPWI || OpC == PPC::CMPW; 1446 bool is32BitUnsignedCompare = OpC == PPC::CMPLWI || OpC == PPC::CMPLW; 1447 bool is64BitUnsignedCompare = OpC == PPC::CMPLDI || OpC == PPC::CMPLD; 1448 1449 // Get the unique definition of SrcReg. 1450 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg); 1451 if (!MI) return false; 1452 int MIOpC = MI->getOpcode(); 1453 1454 bool equalityOnly = false; 1455 bool noSub = false; 1456 if (isPPC64) { 1457 if (is32BitSignedCompare) { 1458 // We can perform this optimization only if MI is sign-extending. 1459 if (MIOpC == PPC::SRAW || MIOpC == PPC::SRAWo || 1460 MIOpC == PPC::SRAWI || MIOpC == PPC::SRAWIo || 1461 MIOpC == PPC::EXTSB || MIOpC == PPC::EXTSBo || 1462 MIOpC == PPC::EXTSH || MIOpC == PPC::EXTSHo || 1463 MIOpC == PPC::EXTSW || MIOpC == PPC::EXTSWo) { 1464 noSub = true; 1465 } else 1466 return false; 1467 } else if (is32BitUnsignedCompare) { 1468 // We can perform this optimization, equality only, if MI is 1469 // zero-extending. 1470 if (MIOpC == PPC::CNTLZW || MIOpC == PPC::CNTLZWo || 1471 MIOpC == PPC::SLW || MIOpC == PPC::SLWo || 1472 MIOpC == PPC::SRW || MIOpC == PPC::SRWo) { 1473 noSub = true; 1474 equalityOnly = true; 1475 } else 1476 return false; 1477 } else 1478 equalityOnly = is64BitUnsignedCompare; 1479 } else 1480 equalityOnly = is32BitUnsignedCompare; 1481 1482 if (equalityOnly) { 1483 // We need to check the uses of the condition register in order to reject 1484 // non-equality comparisons. 1485 for (MachineRegisterInfo::use_instr_iterator I =MRI->use_instr_begin(CRReg), 1486 IE = MRI->use_instr_end(); I != IE; ++I) { 1487 MachineInstr *UseMI = &*I; 1488 if (UseMI->getOpcode() == PPC::BCC) { 1489 unsigned Pred = UseMI->getOperand(0).getImm(); 1490 if (Pred != PPC::PRED_EQ && Pred != PPC::PRED_NE) 1491 return false; 1492 } else if (UseMI->getOpcode() == PPC::ISEL || 1493 UseMI->getOpcode() == PPC::ISEL8) { 1494 unsigned SubIdx = UseMI->getOperand(3).getSubReg(); 1495 if (SubIdx != PPC::sub_eq) 1496 return false; 1497 } else 1498 return false; 1499 } 1500 } 1501 1502 MachineBasicBlock::iterator I = CmpInstr; 1503 1504 // Scan forward to find the first use of the compare. 1505 for (MachineBasicBlock::iterator EL = CmpInstr->getParent()->end(); 1506 I != EL; ++I) { 1507 bool FoundUse = false; 1508 for (MachineRegisterInfo::use_instr_iterator J =MRI->use_instr_begin(CRReg), 1509 JE = MRI->use_instr_end(); J != JE; ++J) 1510 if (&*J == &*I) { 1511 FoundUse = true; 1512 break; 1513 } 1514 1515 if (FoundUse) 1516 break; 1517 } 1518 1519 // There are two possible candidates which can be changed to set CR[01]. 1520 // One is MI, the other is a SUB instruction. 1521 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1). 1522 MachineInstr *Sub = nullptr; 1523 if (SrcReg2 != 0) 1524 // MI is not a candidate for CMPrr. 1525 MI = nullptr; 1526 // FIXME: Conservatively refuse to convert an instruction which isn't in the 1527 // same BB as the comparison. This is to allow the check below to avoid calls 1528 // (and other explicit clobbers); instead we should really check for these 1529 // more explicitly (in at least a few predecessors). 1530 else if (MI->getParent() != CmpInstr->getParent() || Value != 0) { 1531 // PPC does not have a record-form SUBri. 1532 return false; 1533 } 1534 1535 // Search for Sub. 1536 const TargetRegisterInfo *TRI = &getRegisterInfo(); 1537 --I; 1538 1539 // Get ready to iterate backward from CmpInstr. 1540 MachineBasicBlock::iterator E = MI, 1541 B = CmpInstr->getParent()->begin(); 1542 1543 for (; I != E && !noSub; --I) { 1544 const MachineInstr &Instr = *I; 1545 unsigned IOpC = Instr.getOpcode(); 1546 1547 if (&*I != CmpInstr && ( 1548 Instr.modifiesRegister(PPC::CR0, TRI) || 1549 Instr.readsRegister(PPC::CR0, TRI))) 1550 // This instruction modifies or uses the record condition register after 1551 // the one we want to change. While we could do this transformation, it 1552 // would likely not be profitable. This transformation removes one 1553 // instruction, and so even forcing RA to generate one move probably 1554 // makes it unprofitable. 1555 return false; 1556 1557 // Check whether CmpInstr can be made redundant by the current instruction. 1558 if ((OpC == PPC::CMPW || OpC == PPC::CMPLW || 1559 OpC == PPC::CMPD || OpC == PPC::CMPLD) && 1560 (IOpC == PPC::SUBF || IOpC == PPC::SUBF8) && 1561 ((Instr.getOperand(1).getReg() == SrcReg && 1562 Instr.getOperand(2).getReg() == SrcReg2) || 1563 (Instr.getOperand(1).getReg() == SrcReg2 && 1564 Instr.getOperand(2).getReg() == SrcReg))) { 1565 Sub = &*I; 1566 break; 1567 } 1568 1569 if (I == B) 1570 // The 'and' is below the comparison instruction. 1571 return false; 1572 } 1573 1574 // Return false if no candidates exist. 1575 if (!MI && !Sub) 1576 return false; 1577 1578 // The single candidate is called MI. 1579 if (!MI) MI = Sub; 1580 1581 int NewOpC = -1; 1582 MIOpC = MI->getOpcode(); 1583 if (MIOpC == PPC::ANDIo || MIOpC == PPC::ANDIo8) 1584 NewOpC = MIOpC; 1585 else { 1586 NewOpC = PPC::getRecordFormOpcode(MIOpC); 1587 if (NewOpC == -1 && PPC::getNonRecordFormOpcode(MIOpC) != -1) 1588 NewOpC = MIOpC; 1589 } 1590 1591 // FIXME: On the non-embedded POWER architectures, only some of the record 1592 // forms are fast, and we should use only the fast ones. 1593 1594 // The defining instruction has a record form (or is already a record 1595 // form). It is possible, however, that we'll need to reverse the condition 1596 // code of the users. 1597 if (NewOpC == -1) 1598 return false; 1599 1600 SmallVector<std::pair<MachineOperand*, PPC::Predicate>, 4> PredsToUpdate; 1601 SmallVector<std::pair<MachineOperand*, unsigned>, 4> SubRegsToUpdate; 1602 1603 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based on CMP 1604 // needs to be updated to be based on SUB. Push the condition code 1605 // operands to OperandsToUpdate. If it is safe to remove CmpInstr, the 1606 // condition code of these operands will be modified. 1607 bool ShouldSwap = false; 1608 if (Sub) { 1609 ShouldSwap = SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 && 1610 Sub->getOperand(2).getReg() == SrcReg; 1611 1612 // The operands to subf are the opposite of sub, so only in the fixed-point 1613 // case, invert the order. 1614 ShouldSwap = !ShouldSwap; 1615 } 1616 1617 if (ShouldSwap) 1618 for (MachineRegisterInfo::use_instr_iterator 1619 I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end(); 1620 I != IE; ++I) { 1621 MachineInstr *UseMI = &*I; 1622 if (UseMI->getOpcode() == PPC::BCC) { 1623 PPC::Predicate Pred = (PPC::Predicate) UseMI->getOperand(0).getImm(); 1624 assert((!equalityOnly || 1625 Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE) && 1626 "Invalid predicate for equality-only optimization"); 1627 PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)), 1628 PPC::getSwappedPredicate(Pred))); 1629 } else if (UseMI->getOpcode() == PPC::ISEL || 1630 UseMI->getOpcode() == PPC::ISEL8) { 1631 unsigned NewSubReg = UseMI->getOperand(3).getSubReg(); 1632 assert((!equalityOnly || NewSubReg == PPC::sub_eq) && 1633 "Invalid CR bit for equality-only optimization"); 1634 1635 if (NewSubReg == PPC::sub_lt) 1636 NewSubReg = PPC::sub_gt; 1637 else if (NewSubReg == PPC::sub_gt) 1638 NewSubReg = PPC::sub_lt; 1639 1640 SubRegsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(3)), 1641 NewSubReg)); 1642 } else // We need to abort on a user we don't understand. 1643 return false; 1644 } 1645 1646 // Create a new virtual register to hold the value of the CR set by the 1647 // record-form instruction. If the instruction was not previously in 1648 // record form, then set the kill flag on the CR. 1649 CmpInstr->eraseFromParent(); 1650 1651 MachineBasicBlock::iterator MII = MI; 1652 BuildMI(*MI->getParent(), std::next(MII), MI->getDebugLoc(), 1653 get(TargetOpcode::COPY), CRReg) 1654 .addReg(PPC::CR0, MIOpC != NewOpC ? RegState::Kill : 0); 1655 1656 if (MIOpC != NewOpC) { 1657 // We need to be careful here: we're replacing one instruction with 1658 // another, and we need to make sure that we get all of the right 1659 // implicit uses and defs. On the other hand, the caller may be holding 1660 // an iterator to this instruction, and so we can't delete it (this is 1661 // specifically the case if this is the instruction directly after the 1662 // compare). 1663 1664 const MCInstrDesc &NewDesc = get(NewOpC); 1665 MI->setDesc(NewDesc); 1666 1667 if (NewDesc.ImplicitDefs) 1668 for (const uint16_t *ImpDefs = NewDesc.getImplicitDefs(); 1669 *ImpDefs; ++ImpDefs) 1670 if (!MI->definesRegister(*ImpDefs)) 1671 MI->addOperand(*MI->getParent()->getParent(), 1672 MachineOperand::CreateReg(*ImpDefs, true, true)); 1673 if (NewDesc.ImplicitUses) 1674 for (const uint16_t *ImpUses = NewDesc.getImplicitUses(); 1675 *ImpUses; ++ImpUses) 1676 if (!MI->readsRegister(*ImpUses)) 1677 MI->addOperand(*MI->getParent()->getParent(), 1678 MachineOperand::CreateReg(*ImpUses, false, true)); 1679 } 1680 1681 // Modify the condition code of operands in OperandsToUpdate. 1682 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to 1683 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc. 1684 for (unsigned i = 0, e = PredsToUpdate.size(); i < e; i++) 1685 PredsToUpdate[i].first->setImm(PredsToUpdate[i].second); 1686 1687 for (unsigned i = 0, e = SubRegsToUpdate.size(); i < e; i++) 1688 SubRegsToUpdate[i].first->setSubReg(SubRegsToUpdate[i].second); 1689 1690 return true; 1691 } 1692 1693 /// GetInstSize - Return the number of bytes of code the specified 1694 /// instruction may be. This returns the maximum number of bytes. 1695 /// 1696 unsigned PPCInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const { 1697 unsigned Opcode = MI->getOpcode(); 1698 1699 if (Opcode == PPC::INLINEASM) { 1700 const MachineFunction *MF = MI->getParent()->getParent(); 1701 const char *AsmStr = MI->getOperand(0).getSymbolName(); 1702 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo()); 1703 } else if (Opcode == TargetOpcode::STACKMAP) { 1704 return MI->getOperand(1).getImm(); 1705 } else if (Opcode == TargetOpcode::PATCHPOINT) { 1706 PatchPointOpers Opers(MI); 1707 return Opers.getMetaOper(PatchPointOpers::NBytesPos).getImm(); 1708 } else { 1709 const MCInstrDesc &Desc = get(Opcode); 1710 return Desc.getSize(); 1711 } 1712 } 1713 1714