PPCInstrInfo.cpp revision ebe69fe11e48d322045d5949c83283927a0d790b
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/Support/CommandLine.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/ErrorHandling.h" 37#include "llvm/Support/TargetRegistry.h" 38#include "llvm/Support/raw_ostream.h" 39 40using namespace llvm; 41 42#define DEBUG_TYPE "ppc-instr-info" 43 44#define GET_INSTRMAP_INFO 45#define GET_INSTRINFO_CTOR_DTOR 46#include "PPCGenInstrInfo.inc" 47 48static cl:: 49opt<bool> DisableCTRLoopAnal("disable-ppc-ctrloop-analysis", cl::Hidden, 50 cl::desc("Disable analysis for CTR loops")); 51 52static cl::opt<bool> DisableCmpOpt("disable-ppc-cmp-opt", 53cl::desc("Disable compare instruction optimization"), cl::Hidden); 54 55static cl::opt<bool> VSXSelfCopyCrash("crash-on-ppc-vsx-self-copy", 56cl::desc("Causes the backend to crash instead of generating a nop VSX copy"), 57cl::Hidden); 58 59// Pin the vtable to this file. 60void PPCInstrInfo::anchor() {} 61 62PPCInstrInfo::PPCInstrInfo(PPCSubtarget &STI) 63 : PPCGenInstrInfo(PPC::ADJCALLSTACKDOWN, PPC::ADJCALLSTACKUP), 64 Subtarget(STI), RI(STI) {} 65 66/// CreateTargetHazardRecognizer - Return the hazard recognizer to use for 67/// this target when scheduling the DAG. 68ScheduleHazardRecognizer * 69PPCInstrInfo::CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, 70 const ScheduleDAG *DAG) const { 71 unsigned Directive = 72 static_cast<const PPCSubtarget *>(STI)->getDarwinDirective(); 73 if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2 || 74 Directive == PPC::DIR_E500mc || Directive == PPC::DIR_E5500) { 75 const InstrItineraryData *II = 76 static_cast<const PPCSubtarget *>(STI)->getInstrItineraryData(); 77 return new ScoreboardHazardRecognizer(II, DAG); 78 } 79 80 return TargetInstrInfo::CreateTargetHazardRecognizer(STI, DAG); 81} 82 83/// CreateTargetPostRAHazardRecognizer - Return the postRA hazard recognizer 84/// to use for this target when scheduling the DAG. 85ScheduleHazardRecognizer * 86PPCInstrInfo::CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, 87 const ScheduleDAG *DAG) const { 88 unsigned Directive = 89 DAG->MF.getSubtarget<PPCSubtarget>().getDarwinDirective(); 90 91 if (Directive == PPC::DIR_PWR7 || Directive == PPC::DIR_PWR8) 92 return new PPCDispatchGroupSBHazardRecognizer(II, DAG); 93 94 // Most subtargets use a PPC970 recognizer. 95 if (Directive != PPC::DIR_440 && Directive != PPC::DIR_A2 && 96 Directive != PPC::DIR_E500mc && Directive != PPC::DIR_E5500) { 97 assert(DAG->TII && "No InstrInfo?"); 98 99 return new PPCHazardRecognizer970(*DAG); 100 } 101 102 return new ScoreboardHazardRecognizer(II, DAG); 103} 104 105 106int PPCInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, 107 const MachineInstr *DefMI, unsigned DefIdx, 108 const MachineInstr *UseMI, 109 unsigned UseIdx) const { 110 int Latency = PPCGenInstrInfo::getOperandLatency(ItinData, DefMI, DefIdx, 111 UseMI, UseIdx); 112 113 const MachineOperand &DefMO = DefMI->getOperand(DefIdx); 114 unsigned Reg = DefMO.getReg(); 115 116 const TargetRegisterInfo *TRI = &getRegisterInfo(); 117 bool IsRegCR; 118 if (TRI->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. 157bool 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 171unsigned 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 200unsigned 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. 231MachineInstr * 232PPCInstrInfo::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 308bool 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 323void 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. 341void 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. 348bool 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 515unsigned 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 548unsigned 549PPCInstrInfo::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. 595bool 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 635void 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 700void PPCInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 701 MachineBasicBlock::iterator I, DebugLoc DL, 702 unsigned DestReg, unsigned SrcReg, 703 bool KillSrc) const { 704 // We can end up with self copies and similar things as a result of VSX copy 705 // legalization. Promote them here. 706 const TargetRegisterInfo *TRI = &getRegisterInfo(); 707 if (PPC::F8RCRegClass.contains(DestReg) && 708 PPC::VSRCRegClass.contains(SrcReg)) { 709 unsigned SuperReg = 710 TRI->getMatchingSuperReg(DestReg, PPC::sub_64, &PPC::VSRCRegClass); 711 712 if (VSXSelfCopyCrash && SrcReg == SuperReg) 713 llvm_unreachable("nop VSX copy"); 714 715 DestReg = SuperReg; 716 } else if (PPC::VRRCRegClass.contains(DestReg) && 717 PPC::VSRCRegClass.contains(SrcReg)) { 718 unsigned SuperReg = 719 TRI->getMatchingSuperReg(DestReg, PPC::sub_128, &PPC::VSRCRegClass); 720 721 if (VSXSelfCopyCrash && SrcReg == SuperReg) 722 llvm_unreachable("nop VSX copy"); 723 724 DestReg = SuperReg; 725 } else if (PPC::F8RCRegClass.contains(SrcReg) && 726 PPC::VSRCRegClass.contains(DestReg)) { 727 unsigned SuperReg = 728 TRI->getMatchingSuperReg(SrcReg, PPC::sub_64, &PPC::VSRCRegClass); 729 730 if (VSXSelfCopyCrash && DestReg == SuperReg) 731 llvm_unreachable("nop VSX copy"); 732 733 SrcReg = SuperReg; 734 } else if (PPC::VRRCRegClass.contains(SrcReg) && 735 PPC::VSRCRegClass.contains(DestReg)) { 736 unsigned SuperReg = 737 TRI->getMatchingSuperReg(SrcReg, PPC::sub_128, &PPC::VSRCRegClass); 738 739 if (VSXSelfCopyCrash && DestReg == SuperReg) 740 llvm_unreachable("nop VSX copy"); 741 742 SrcReg = SuperReg; 743 } 744 745 unsigned Opc; 746 if (PPC::GPRCRegClass.contains(DestReg, SrcReg)) 747 Opc = PPC::OR; 748 else if (PPC::G8RCRegClass.contains(DestReg, SrcReg)) 749 Opc = PPC::OR8; 750 else if (PPC::F4RCRegClass.contains(DestReg, SrcReg)) 751 Opc = PPC::FMR; 752 else if (PPC::CRRCRegClass.contains(DestReg, SrcReg)) 753 Opc = PPC::MCRF; 754 else if (PPC::VRRCRegClass.contains(DestReg, SrcReg)) 755 Opc = PPC::VOR; 756 else if (PPC::VSRCRegClass.contains(DestReg, SrcReg)) 757 // There are two different ways this can be done: 758 // 1. xxlor : This has lower latency (on the P7), 2 cycles, but can only 759 // issue in VSU pipeline 0. 760 // 2. xmovdp/xmovsp: This has higher latency (on the P7), 6 cycles, but 761 // can go to either pipeline. 762 // We'll always use xxlor here, because in practically all cases where 763 // copies are generated, they are close enough to some use that the 764 // lower-latency form is preferable. 765 Opc = PPC::XXLOR; 766 else if (PPC::VSFRCRegClass.contains(DestReg, SrcReg)) 767 Opc = PPC::XXLORf; 768 else if (PPC::QFRCRegClass.contains(DestReg, SrcReg)) 769 Opc = PPC::QVFMR; 770 else if (PPC::QSRCRegClass.contains(DestReg, SrcReg)) 771 Opc = PPC::QVFMRs; 772 else if (PPC::QBRCRegClass.contains(DestReg, SrcReg)) 773 Opc = PPC::QVFMRb; 774 else if (PPC::CRBITRCRegClass.contains(DestReg, SrcReg)) 775 Opc = PPC::CROR; 776 else 777 llvm_unreachable("Impossible reg-to-reg copy"); 778 779 const MCInstrDesc &MCID = get(Opc); 780 if (MCID.getNumOperands() == 3) 781 BuildMI(MBB, I, DL, MCID, DestReg) 782 .addReg(SrcReg).addReg(SrcReg, getKillRegState(KillSrc)); 783 else 784 BuildMI(MBB, I, DL, MCID, DestReg).addReg(SrcReg, getKillRegState(KillSrc)); 785} 786 787// This function returns true if a CR spill is necessary and false otherwise. 788bool 789PPCInstrInfo::StoreRegToStackSlot(MachineFunction &MF, 790 unsigned SrcReg, bool isKill, 791 int FrameIdx, 792 const TargetRegisterClass *RC, 793 SmallVectorImpl<MachineInstr*> &NewMIs, 794 bool &NonRI, bool &SpillsVRS) const{ 795 // Note: If additional store instructions are added here, 796 // update isStoreToStackSlot. 797 798 DebugLoc DL; 799 if (PPC::GPRCRegClass.hasSubClassEq(RC) || 800 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) { 801 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STW)) 802 .addReg(SrcReg, 803 getKillRegState(isKill)), 804 FrameIdx)); 805 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) || 806 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) { 807 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STD)) 808 .addReg(SrcReg, 809 getKillRegState(isKill)), 810 FrameIdx)); 811 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) { 812 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFD)) 813 .addReg(SrcReg, 814 getKillRegState(isKill)), 815 FrameIdx)); 816 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) { 817 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STFS)) 818 .addReg(SrcReg, 819 getKillRegState(isKill)), 820 FrameIdx)); 821 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) { 822 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CR)) 823 .addReg(SrcReg, 824 getKillRegState(isKill)), 825 FrameIdx)); 826 return true; 827 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) { 828 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_CRBIT)) 829 .addReg(SrcReg, 830 getKillRegState(isKill)), 831 FrameIdx)); 832 return true; 833 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) { 834 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STVX)) 835 .addReg(SrcReg, 836 getKillRegState(isKill)), 837 FrameIdx)); 838 NonRI = true; 839 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) { 840 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXVD2X)) 841 .addReg(SrcReg, 842 getKillRegState(isKill)), 843 FrameIdx)); 844 NonRI = true; 845 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) { 846 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::STXSDX)) 847 .addReg(SrcReg, 848 getKillRegState(isKill)), 849 FrameIdx)); 850 NonRI = true; 851 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) { 852 assert(Subtarget.isDarwin() && 853 "VRSAVE only needs spill/restore on Darwin"); 854 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::SPILL_VRSAVE)) 855 .addReg(SrcReg, 856 getKillRegState(isKill)), 857 FrameIdx)); 858 SpillsVRS = true; 859 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) { 860 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFDX)) 861 .addReg(SrcReg, 862 getKillRegState(isKill)), 863 FrameIdx)); 864 NonRI = true; 865 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) { 866 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFSXs)) 867 .addReg(SrcReg, 868 getKillRegState(isKill)), 869 FrameIdx)); 870 NonRI = true; 871 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) { 872 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVSTFDXb)) 873 .addReg(SrcReg, 874 getKillRegState(isKill)), 875 FrameIdx)); 876 NonRI = true; 877 } else { 878 llvm_unreachable("Unknown regclass!"); 879 } 880 881 return false; 882} 883 884void 885PPCInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 886 MachineBasicBlock::iterator MI, 887 unsigned SrcReg, bool isKill, int FrameIdx, 888 const TargetRegisterClass *RC, 889 const TargetRegisterInfo *TRI) const { 890 MachineFunction &MF = *MBB.getParent(); 891 SmallVector<MachineInstr*, 4> NewMIs; 892 893 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); 894 FuncInfo->setHasSpills(); 895 896 bool NonRI = false, SpillsVRS = false; 897 if (StoreRegToStackSlot(MF, SrcReg, isKill, FrameIdx, RC, NewMIs, 898 NonRI, SpillsVRS)) 899 FuncInfo->setSpillsCR(); 900 901 if (SpillsVRS) 902 FuncInfo->setSpillsVRSAVE(); 903 904 if (NonRI) 905 FuncInfo->setHasNonRISpills(); 906 907 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i) 908 MBB.insert(MI, NewMIs[i]); 909 910 const MachineFrameInfo &MFI = *MF.getFrameInfo(); 911 MachineMemOperand *MMO = 912 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx), 913 MachineMemOperand::MOStore, 914 MFI.getObjectSize(FrameIdx), 915 MFI.getObjectAlignment(FrameIdx)); 916 NewMIs.back()->addMemOperand(MF, MMO); 917} 918 919bool 920PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL, 921 unsigned DestReg, int FrameIdx, 922 const TargetRegisterClass *RC, 923 SmallVectorImpl<MachineInstr*> &NewMIs, 924 bool &NonRI, bool &SpillsVRS) const{ 925 // Note: If additional load instructions are added here, 926 // update isLoadFromStackSlot. 927 928 if (PPC::GPRCRegClass.hasSubClassEq(RC) || 929 PPC::GPRC_NOR0RegClass.hasSubClassEq(RC)) { 930 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LWZ), 931 DestReg), FrameIdx)); 932 } else if (PPC::G8RCRegClass.hasSubClassEq(RC) || 933 PPC::G8RC_NOX0RegClass.hasSubClassEq(RC)) { 934 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LD), DestReg), 935 FrameIdx)); 936 } else if (PPC::F8RCRegClass.hasSubClassEq(RC)) { 937 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFD), DestReg), 938 FrameIdx)); 939 } else if (PPC::F4RCRegClass.hasSubClassEq(RC)) { 940 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFS), DestReg), 941 FrameIdx)); 942 } else if (PPC::CRRCRegClass.hasSubClassEq(RC)) { 943 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, 944 get(PPC::RESTORE_CR), DestReg), 945 FrameIdx)); 946 return true; 947 } else if (PPC::CRBITRCRegClass.hasSubClassEq(RC)) { 948 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, 949 get(PPC::RESTORE_CRBIT), DestReg), 950 FrameIdx)); 951 return true; 952 } else if (PPC::VRRCRegClass.hasSubClassEq(RC)) { 953 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LVX), DestReg), 954 FrameIdx)); 955 NonRI = true; 956 } else if (PPC::VSRCRegClass.hasSubClassEq(RC)) { 957 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXVD2X), DestReg), 958 FrameIdx)); 959 NonRI = true; 960 } else if (PPC::VSFRCRegClass.hasSubClassEq(RC)) { 961 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LXSDX), DestReg), 962 FrameIdx)); 963 NonRI = true; 964 } else if (PPC::VRSAVERCRegClass.hasSubClassEq(RC)) { 965 assert(Subtarget.isDarwin() && 966 "VRSAVE only needs spill/restore on Darwin"); 967 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, 968 get(PPC::RESTORE_VRSAVE), 969 DestReg), 970 FrameIdx)); 971 SpillsVRS = true; 972 } else if (PPC::QFRCRegClass.hasSubClassEq(RC)) { 973 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFDX), DestReg), 974 FrameIdx)); 975 NonRI = true; 976 } else if (PPC::QSRCRegClass.hasSubClassEq(RC)) { 977 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFSXs), DestReg), 978 FrameIdx)); 979 NonRI = true; 980 } else if (PPC::QBRCRegClass.hasSubClassEq(RC)) { 981 NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::QVLFDXb), DestReg), 982 FrameIdx)); 983 NonRI = true; 984 } else { 985 llvm_unreachable("Unknown regclass!"); 986 } 987 988 return false; 989} 990 991void 992PPCInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, 993 MachineBasicBlock::iterator MI, 994 unsigned DestReg, int FrameIdx, 995 const TargetRegisterClass *RC, 996 const TargetRegisterInfo *TRI) const { 997 MachineFunction &MF = *MBB.getParent(); 998 SmallVector<MachineInstr*, 4> NewMIs; 999 DebugLoc DL; 1000 if (MI != MBB.end()) DL = MI->getDebugLoc(); 1001 1002 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); 1003 FuncInfo->setHasSpills(); 1004 1005 bool NonRI = false, SpillsVRS = false; 1006 if (LoadRegFromStackSlot(MF, DL, DestReg, FrameIdx, RC, NewMIs, 1007 NonRI, SpillsVRS)) 1008 FuncInfo->setSpillsCR(); 1009 1010 if (SpillsVRS) 1011 FuncInfo->setSpillsVRSAVE(); 1012 1013 if (NonRI) 1014 FuncInfo->setHasNonRISpills(); 1015 1016 for (unsigned i = 0, e = NewMIs.size(); i != e; ++i) 1017 MBB.insert(MI, NewMIs[i]); 1018 1019 const MachineFrameInfo &MFI = *MF.getFrameInfo(); 1020 MachineMemOperand *MMO = 1021 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx), 1022 MachineMemOperand::MOLoad, 1023 MFI.getObjectSize(FrameIdx), 1024 MFI.getObjectAlignment(FrameIdx)); 1025 NewMIs.back()->addMemOperand(MF, MMO); 1026} 1027 1028bool PPCInstrInfo:: 1029ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 1030 assert(Cond.size() == 2 && "Invalid PPC branch opcode!"); 1031 if (Cond[1].getReg() == PPC::CTR8 || Cond[1].getReg() == PPC::CTR) 1032 Cond[0].setImm(Cond[0].getImm() == 0 ? 1 : 0); 1033 else 1034 // Leave the CR# the same, but invert the condition. 1035 Cond[0].setImm(PPC::InvertPredicate((PPC::Predicate)Cond[0].getImm())); 1036 return false; 1037} 1038 1039bool PPCInstrInfo::FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI, 1040 unsigned Reg, MachineRegisterInfo *MRI) const { 1041 // For some instructions, it is legal to fold ZERO into the RA register field. 1042 // A zero immediate should always be loaded with a single li. 1043 unsigned DefOpc = DefMI->getOpcode(); 1044 if (DefOpc != PPC::LI && DefOpc != PPC::LI8) 1045 return false; 1046 if (!DefMI->getOperand(1).isImm()) 1047 return false; 1048 if (DefMI->getOperand(1).getImm() != 0) 1049 return false; 1050 1051 // Note that we cannot here invert the arguments of an isel in order to fold 1052 // a ZERO into what is presented as the second argument. All we have here 1053 // is the condition bit, and that might come from a CR-logical bit operation. 1054 1055 const MCInstrDesc &UseMCID = UseMI->getDesc(); 1056 1057 // Only fold into real machine instructions. 1058 if (UseMCID.isPseudo()) 1059 return false; 1060 1061 unsigned UseIdx; 1062 for (UseIdx = 0; UseIdx < UseMI->getNumOperands(); ++UseIdx) 1063 if (UseMI->getOperand(UseIdx).isReg() && 1064 UseMI->getOperand(UseIdx).getReg() == Reg) 1065 break; 1066 1067 assert(UseIdx < UseMI->getNumOperands() && "Cannot find Reg in UseMI"); 1068 assert(UseIdx < UseMCID.getNumOperands() && "No operand description for Reg"); 1069 1070 const MCOperandInfo *UseInfo = &UseMCID.OpInfo[UseIdx]; 1071 1072 // We can fold the zero if this register requires a GPRC_NOR0/G8RC_NOX0 1073 // register (which might also be specified as a pointer class kind). 1074 if (UseInfo->isLookupPtrRegClass()) { 1075 if (UseInfo->RegClass /* Kind */ != 1) 1076 return false; 1077 } else { 1078 if (UseInfo->RegClass != PPC::GPRC_NOR0RegClassID && 1079 UseInfo->RegClass != PPC::G8RC_NOX0RegClassID) 1080 return false; 1081 } 1082 1083 // Make sure this is not tied to an output register (or otherwise 1084 // constrained). This is true for ST?UX registers, for example, which 1085 // are tied to their output registers. 1086 if (UseInfo->Constraints != 0) 1087 return false; 1088 1089 unsigned ZeroReg; 1090 if (UseInfo->isLookupPtrRegClass()) { 1091 bool isPPC64 = Subtarget.isPPC64(); 1092 ZeroReg = isPPC64 ? PPC::ZERO8 : PPC::ZERO; 1093 } else { 1094 ZeroReg = UseInfo->RegClass == PPC::G8RC_NOX0RegClassID ? 1095 PPC::ZERO8 : PPC::ZERO; 1096 } 1097 1098 bool DeleteDef = MRI->hasOneNonDBGUse(Reg); 1099 UseMI->getOperand(UseIdx).setReg(ZeroReg); 1100 1101 if (DeleteDef) 1102 DefMI->eraseFromParent(); 1103 1104 return true; 1105} 1106 1107static bool MBBDefinesCTR(MachineBasicBlock &MBB) { 1108 for (MachineBasicBlock::iterator I = MBB.begin(), IE = MBB.end(); 1109 I != IE; ++I) 1110 if (I->definesRegister(PPC::CTR) || I->definesRegister(PPC::CTR8)) 1111 return true; 1112 return false; 1113} 1114 1115// We should make sure that, if we're going to predicate both sides of a 1116// condition (a diamond), that both sides don't define the counter register. We 1117// can predicate counter-decrement-based branches, but while that predicates 1118// the branching, it does not predicate the counter decrement. If we tried to 1119// merge the triangle into one predicated block, we'd decrement the counter 1120// twice. 1121bool PPCInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB, 1122 unsigned NumT, unsigned ExtraT, 1123 MachineBasicBlock &FMBB, 1124 unsigned NumF, unsigned ExtraF, 1125 const BranchProbability &Probability) const { 1126 return !(MBBDefinesCTR(TMBB) && MBBDefinesCTR(FMBB)); 1127} 1128 1129 1130bool PPCInstrInfo::isPredicated(const MachineInstr *MI) const { 1131 // The predicated branches are identified by their type, not really by the 1132 // explicit presence of a predicate. Furthermore, some of them can be 1133 // predicated more than once. Because if conversion won't try to predicate 1134 // any instruction which already claims to be predicated (by returning true 1135 // here), always return false. In doing so, we let isPredicable() be the 1136 // final word on whether not the instruction can be (further) predicated. 1137 1138 return false; 1139} 1140 1141bool PPCInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const { 1142 if (!MI->isTerminator()) 1143 return false; 1144 1145 // Conditional branch is a special case. 1146 if (MI->isBranch() && !MI->isBarrier()) 1147 return true; 1148 1149 return !isPredicated(MI); 1150} 1151 1152bool PPCInstrInfo::PredicateInstruction( 1153 MachineInstr *MI, 1154 const SmallVectorImpl<MachineOperand> &Pred) const { 1155 unsigned OpC = MI->getOpcode(); 1156 if (OpC == PPC::BLR || OpC == PPC::BLR8) { 1157 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) { 1158 bool isPPC64 = Subtarget.isPPC64(); 1159 MI->setDesc(get(Pred[0].getImm() ? 1160 (isPPC64 ? PPC::BDNZLR8 : PPC::BDNZLR) : 1161 (isPPC64 ? PPC::BDZLR8 : PPC::BDZLR))); 1162 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) { 1163 MI->setDesc(get(PPC::BCLR)); 1164 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1165 .addReg(Pred[1].getReg()); 1166 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { 1167 MI->setDesc(get(PPC::BCLRn)); 1168 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1169 .addReg(Pred[1].getReg()); 1170 } else { 1171 MI->setDesc(get(PPC::BCCLR)); 1172 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1173 .addImm(Pred[0].getImm()) 1174 .addReg(Pred[1].getReg()); 1175 } 1176 1177 return true; 1178 } else if (OpC == PPC::B) { 1179 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) { 1180 bool isPPC64 = Subtarget.isPPC64(); 1181 MI->setDesc(get(Pred[0].getImm() ? 1182 (isPPC64 ? PPC::BDNZ8 : PPC::BDNZ) : 1183 (isPPC64 ? PPC::BDZ8 : PPC::BDZ))); 1184 } else if (Pred[0].getImm() == PPC::PRED_BIT_SET) { 1185 MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); 1186 MI->RemoveOperand(0); 1187 1188 MI->setDesc(get(PPC::BC)); 1189 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1190 .addReg(Pred[1].getReg()) 1191 .addMBB(MBB); 1192 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { 1193 MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); 1194 MI->RemoveOperand(0); 1195 1196 MI->setDesc(get(PPC::BCn)); 1197 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1198 .addReg(Pred[1].getReg()) 1199 .addMBB(MBB); 1200 } else { 1201 MachineBasicBlock *MBB = MI->getOperand(0).getMBB(); 1202 MI->RemoveOperand(0); 1203 1204 MI->setDesc(get(PPC::BCC)); 1205 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1206 .addImm(Pred[0].getImm()) 1207 .addReg(Pred[1].getReg()) 1208 .addMBB(MBB); 1209 } 1210 1211 return true; 1212 } else if (OpC == PPC::BCTR || OpC == PPC::BCTR8 || 1213 OpC == PPC::BCTRL || OpC == PPC::BCTRL8) { 1214 if (Pred[1].getReg() == PPC::CTR8 || Pred[1].getReg() == PPC::CTR) 1215 llvm_unreachable("Cannot predicate bctr[l] on the ctr register"); 1216 1217 bool setLR = OpC == PPC::BCTRL || OpC == PPC::BCTRL8; 1218 bool isPPC64 = Subtarget.isPPC64(); 1219 1220 if (Pred[0].getImm() == PPC::PRED_BIT_SET) { 1221 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8 : PPC::BCCTR8) : 1222 (setLR ? PPC::BCCTRL : PPC::BCCTR))); 1223 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1224 .addReg(Pred[1].getReg()); 1225 return true; 1226 } else if (Pred[0].getImm() == PPC::PRED_BIT_UNSET) { 1227 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCTRL8n : PPC::BCCTR8n) : 1228 (setLR ? PPC::BCCTRLn : PPC::BCCTRn))); 1229 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1230 .addReg(Pred[1].getReg()); 1231 return true; 1232 } 1233 1234 MI->setDesc(get(isPPC64 ? (setLR ? PPC::BCCCTRL8 : PPC::BCCCTR8) : 1235 (setLR ? PPC::BCCCTRL : PPC::BCCCTR))); 1236 MachineInstrBuilder(*MI->getParent()->getParent(), MI) 1237 .addImm(Pred[0].getImm()) 1238 .addReg(Pred[1].getReg()); 1239 return true; 1240 } 1241 1242 return false; 1243} 1244 1245bool PPCInstrInfo::SubsumesPredicate( 1246 const SmallVectorImpl<MachineOperand> &Pred1, 1247 const SmallVectorImpl<MachineOperand> &Pred2) const { 1248 assert(Pred1.size() == 2 && "Invalid PPC first predicate"); 1249 assert(Pred2.size() == 2 && "Invalid PPC second predicate"); 1250 1251 if (Pred1[1].getReg() == PPC::CTR8 || Pred1[1].getReg() == PPC::CTR) 1252 return false; 1253 if (Pred2[1].getReg() == PPC::CTR8 || Pred2[1].getReg() == PPC::CTR) 1254 return false; 1255 1256 // P1 can only subsume P2 if they test the same condition register. 1257 if (Pred1[1].getReg() != Pred2[1].getReg()) 1258 return false; 1259 1260 PPC::Predicate P1 = (PPC::Predicate) Pred1[0].getImm(); 1261 PPC::Predicate P2 = (PPC::Predicate) Pred2[0].getImm(); 1262 1263 if (P1 == P2) 1264 return true; 1265 1266 // Does P1 subsume P2, e.g. GE subsumes GT. 1267 if (P1 == PPC::PRED_LE && 1268 (P2 == PPC::PRED_LT || P2 == PPC::PRED_EQ)) 1269 return true; 1270 if (P1 == PPC::PRED_GE && 1271 (P2 == PPC::PRED_GT || P2 == PPC::PRED_EQ)) 1272 return true; 1273 1274 return false; 1275} 1276 1277bool PPCInstrInfo::DefinesPredicate(MachineInstr *MI, 1278 std::vector<MachineOperand> &Pred) const { 1279 // Note: At the present time, the contents of Pred from this function is 1280 // unused by IfConversion. This implementation follows ARM by pushing the 1281 // CR-defining operand. Because the 'DZ' and 'DNZ' count as types of 1282 // predicate, instructions defining CTR or CTR8 are also included as 1283 // predicate-defining instructions. 1284 1285 const TargetRegisterClass *RCs[] = 1286 { &PPC::CRRCRegClass, &PPC::CRBITRCRegClass, 1287 &PPC::CTRRCRegClass, &PPC::CTRRC8RegClass }; 1288 1289 bool Found = false; 1290 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 1291 const MachineOperand &MO = MI->getOperand(i); 1292 for (unsigned c = 0; c < array_lengthof(RCs) && !Found; ++c) { 1293 const TargetRegisterClass *RC = RCs[c]; 1294 if (MO.isReg()) { 1295 if (MO.isDef() && RC->contains(MO.getReg())) { 1296 Pred.push_back(MO); 1297 Found = true; 1298 } 1299 } else if (MO.isRegMask()) { 1300 for (TargetRegisterClass::iterator I = RC->begin(), 1301 IE = RC->end(); I != IE; ++I) 1302 if (MO.clobbersPhysReg(*I)) { 1303 Pred.push_back(MO); 1304 Found = true; 1305 } 1306 } 1307 } 1308 } 1309 1310 return Found; 1311} 1312 1313bool PPCInstrInfo::isPredicable(MachineInstr *MI) const { 1314 unsigned OpC = MI->getOpcode(); 1315 switch (OpC) { 1316 default: 1317 return false; 1318 case PPC::B: 1319 case PPC::BLR: 1320 case PPC::BLR8: 1321 case PPC::BCTR: 1322 case PPC::BCTR8: 1323 case PPC::BCTRL: 1324 case PPC::BCTRL8: 1325 return true; 1326 } 1327} 1328 1329bool PPCInstrInfo::analyzeCompare(const MachineInstr *MI, 1330 unsigned &SrcReg, unsigned &SrcReg2, 1331 int &Mask, int &Value) const { 1332 unsigned Opc = MI->getOpcode(); 1333 1334 switch (Opc) { 1335 default: return false; 1336 case PPC::CMPWI: 1337 case PPC::CMPLWI: 1338 case PPC::CMPDI: 1339 case PPC::CMPLDI: 1340 SrcReg = MI->getOperand(1).getReg(); 1341 SrcReg2 = 0; 1342 Value = MI->getOperand(2).getImm(); 1343 Mask = 0xFFFF; 1344 return true; 1345 case PPC::CMPW: 1346 case PPC::CMPLW: 1347 case PPC::CMPD: 1348 case PPC::CMPLD: 1349 case PPC::FCMPUS: 1350 case PPC::FCMPUD: 1351 SrcReg = MI->getOperand(1).getReg(); 1352 SrcReg2 = MI->getOperand(2).getReg(); 1353 return true; 1354 } 1355} 1356 1357bool PPCInstrInfo::optimizeCompareInstr(MachineInstr *CmpInstr, 1358 unsigned SrcReg, unsigned SrcReg2, 1359 int Mask, int Value, 1360 const MachineRegisterInfo *MRI) const { 1361 if (DisableCmpOpt) 1362 return false; 1363 1364 int OpC = CmpInstr->getOpcode(); 1365 unsigned CRReg = CmpInstr->getOperand(0).getReg(); 1366 1367 // FP record forms set CR1 based on the execption status bits, not a 1368 // comparison with zero. 1369 if (OpC == PPC::FCMPUS || OpC == PPC::FCMPUD) 1370 return false; 1371 1372 // The record forms set the condition register based on a signed comparison 1373 // with zero (so says the ISA manual). This is not as straightforward as it 1374 // seems, however, because this is always a 64-bit comparison on PPC64, even 1375 // for instructions that are 32-bit in nature (like slw for example). 1376 // So, on PPC32, for unsigned comparisons, we can use the record forms only 1377 // for equality checks (as those don't depend on the sign). On PPC64, 1378 // we are restricted to equality for unsigned 64-bit comparisons and for 1379 // signed 32-bit comparisons the applicability is more restricted. 1380 bool isPPC64 = Subtarget.isPPC64(); 1381 bool is32BitSignedCompare = OpC == PPC::CMPWI || OpC == PPC::CMPW; 1382 bool is32BitUnsignedCompare = OpC == PPC::CMPLWI || OpC == PPC::CMPLW; 1383 bool is64BitUnsignedCompare = OpC == PPC::CMPLDI || OpC == PPC::CMPLD; 1384 1385 // Get the unique definition of SrcReg. 1386 MachineInstr *MI = MRI->getUniqueVRegDef(SrcReg); 1387 if (!MI) return false; 1388 int MIOpC = MI->getOpcode(); 1389 1390 bool equalityOnly = false; 1391 bool noSub = false; 1392 if (isPPC64) { 1393 if (is32BitSignedCompare) { 1394 // We can perform this optimization only if MI is sign-extending. 1395 if (MIOpC == PPC::SRAW || MIOpC == PPC::SRAWo || 1396 MIOpC == PPC::SRAWI || MIOpC == PPC::SRAWIo || 1397 MIOpC == PPC::EXTSB || MIOpC == PPC::EXTSBo || 1398 MIOpC == PPC::EXTSH || MIOpC == PPC::EXTSHo || 1399 MIOpC == PPC::EXTSW || MIOpC == PPC::EXTSWo) { 1400 noSub = true; 1401 } else 1402 return false; 1403 } else if (is32BitUnsignedCompare) { 1404 // We can perform this optimization, equality only, if MI is 1405 // zero-extending. 1406 if (MIOpC == PPC::CNTLZW || MIOpC == PPC::CNTLZWo || 1407 MIOpC == PPC::SLW || MIOpC == PPC::SLWo || 1408 MIOpC == PPC::SRW || MIOpC == PPC::SRWo) { 1409 noSub = true; 1410 equalityOnly = true; 1411 } else 1412 return false; 1413 } else 1414 equalityOnly = is64BitUnsignedCompare; 1415 } else 1416 equalityOnly = is32BitUnsignedCompare; 1417 1418 if (equalityOnly) { 1419 // We need to check the uses of the condition register in order to reject 1420 // non-equality comparisons. 1421 for (MachineRegisterInfo::use_instr_iterator I =MRI->use_instr_begin(CRReg), 1422 IE = MRI->use_instr_end(); I != IE; ++I) { 1423 MachineInstr *UseMI = &*I; 1424 if (UseMI->getOpcode() == PPC::BCC) { 1425 unsigned Pred = UseMI->getOperand(0).getImm(); 1426 if (Pred != PPC::PRED_EQ && Pred != PPC::PRED_NE) 1427 return false; 1428 } else if (UseMI->getOpcode() == PPC::ISEL || 1429 UseMI->getOpcode() == PPC::ISEL8) { 1430 unsigned SubIdx = UseMI->getOperand(3).getSubReg(); 1431 if (SubIdx != PPC::sub_eq) 1432 return false; 1433 } else 1434 return false; 1435 } 1436 } 1437 1438 MachineBasicBlock::iterator I = CmpInstr; 1439 1440 // Scan forward to find the first use of the compare. 1441 for (MachineBasicBlock::iterator EL = CmpInstr->getParent()->end(); 1442 I != EL; ++I) { 1443 bool FoundUse = false; 1444 for (MachineRegisterInfo::use_instr_iterator J =MRI->use_instr_begin(CRReg), 1445 JE = MRI->use_instr_end(); J != JE; ++J) 1446 if (&*J == &*I) { 1447 FoundUse = true; 1448 break; 1449 } 1450 1451 if (FoundUse) 1452 break; 1453 } 1454 1455 // There are two possible candidates which can be changed to set CR[01]. 1456 // One is MI, the other is a SUB instruction. 1457 // For CMPrr(r1,r2), we are looking for SUB(r1,r2) or SUB(r2,r1). 1458 MachineInstr *Sub = nullptr; 1459 if (SrcReg2 != 0) 1460 // MI is not a candidate for CMPrr. 1461 MI = nullptr; 1462 // FIXME: Conservatively refuse to convert an instruction which isn't in the 1463 // same BB as the comparison. This is to allow the check below to avoid calls 1464 // (and other explicit clobbers); instead we should really check for these 1465 // more explicitly (in at least a few predecessors). 1466 else if (MI->getParent() != CmpInstr->getParent() || Value != 0) { 1467 // PPC does not have a record-form SUBri. 1468 return false; 1469 } 1470 1471 // Search for Sub. 1472 const TargetRegisterInfo *TRI = &getRegisterInfo(); 1473 --I; 1474 1475 // Get ready to iterate backward from CmpInstr. 1476 MachineBasicBlock::iterator E = MI, 1477 B = CmpInstr->getParent()->begin(); 1478 1479 for (; I != E && !noSub; --I) { 1480 const MachineInstr &Instr = *I; 1481 unsigned IOpC = Instr.getOpcode(); 1482 1483 if (&*I != CmpInstr && ( 1484 Instr.modifiesRegister(PPC::CR0, TRI) || 1485 Instr.readsRegister(PPC::CR0, TRI))) 1486 // This instruction modifies or uses the record condition register after 1487 // the one we want to change. While we could do this transformation, it 1488 // would likely not be profitable. This transformation removes one 1489 // instruction, and so even forcing RA to generate one move probably 1490 // makes it unprofitable. 1491 return false; 1492 1493 // Check whether CmpInstr can be made redundant by the current instruction. 1494 if ((OpC == PPC::CMPW || OpC == PPC::CMPLW || 1495 OpC == PPC::CMPD || OpC == PPC::CMPLD) && 1496 (IOpC == PPC::SUBF || IOpC == PPC::SUBF8) && 1497 ((Instr.getOperand(1).getReg() == SrcReg && 1498 Instr.getOperand(2).getReg() == SrcReg2) || 1499 (Instr.getOperand(1).getReg() == SrcReg2 && 1500 Instr.getOperand(2).getReg() == SrcReg))) { 1501 Sub = &*I; 1502 break; 1503 } 1504 1505 if (I == B) 1506 // The 'and' is below the comparison instruction. 1507 return false; 1508 } 1509 1510 // Return false if no candidates exist. 1511 if (!MI && !Sub) 1512 return false; 1513 1514 // The single candidate is called MI. 1515 if (!MI) MI = Sub; 1516 1517 int NewOpC = -1; 1518 MIOpC = MI->getOpcode(); 1519 if (MIOpC == PPC::ANDIo || MIOpC == PPC::ANDIo8) 1520 NewOpC = MIOpC; 1521 else { 1522 NewOpC = PPC::getRecordFormOpcode(MIOpC); 1523 if (NewOpC == -1 && PPC::getNonRecordFormOpcode(MIOpC) != -1) 1524 NewOpC = MIOpC; 1525 } 1526 1527 // FIXME: On the non-embedded POWER architectures, only some of the record 1528 // forms are fast, and we should use only the fast ones. 1529 1530 // The defining instruction has a record form (or is already a record 1531 // form). It is possible, however, that we'll need to reverse the condition 1532 // code of the users. 1533 if (NewOpC == -1) 1534 return false; 1535 1536 SmallVector<std::pair<MachineOperand*, PPC::Predicate>, 4> PredsToUpdate; 1537 SmallVector<std::pair<MachineOperand*, unsigned>, 4> SubRegsToUpdate; 1538 1539 // If we have SUB(r1, r2) and CMP(r2, r1), the condition code based on CMP 1540 // needs to be updated to be based on SUB. Push the condition code 1541 // operands to OperandsToUpdate. If it is safe to remove CmpInstr, the 1542 // condition code of these operands will be modified. 1543 bool ShouldSwap = false; 1544 if (Sub) { 1545 ShouldSwap = SrcReg2 != 0 && Sub->getOperand(1).getReg() == SrcReg2 && 1546 Sub->getOperand(2).getReg() == SrcReg; 1547 1548 // The operands to subf are the opposite of sub, so only in the fixed-point 1549 // case, invert the order. 1550 ShouldSwap = !ShouldSwap; 1551 } 1552 1553 if (ShouldSwap) 1554 for (MachineRegisterInfo::use_instr_iterator 1555 I = MRI->use_instr_begin(CRReg), IE = MRI->use_instr_end(); 1556 I != IE; ++I) { 1557 MachineInstr *UseMI = &*I; 1558 if (UseMI->getOpcode() == PPC::BCC) { 1559 PPC::Predicate Pred = (PPC::Predicate) UseMI->getOperand(0).getImm(); 1560 assert((!equalityOnly || 1561 Pred == PPC::PRED_EQ || Pred == PPC::PRED_NE) && 1562 "Invalid predicate for equality-only optimization"); 1563 PredsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(0)), 1564 PPC::getSwappedPredicate(Pred))); 1565 } else if (UseMI->getOpcode() == PPC::ISEL || 1566 UseMI->getOpcode() == PPC::ISEL8) { 1567 unsigned NewSubReg = UseMI->getOperand(3).getSubReg(); 1568 assert((!equalityOnly || NewSubReg == PPC::sub_eq) && 1569 "Invalid CR bit for equality-only optimization"); 1570 1571 if (NewSubReg == PPC::sub_lt) 1572 NewSubReg = PPC::sub_gt; 1573 else if (NewSubReg == PPC::sub_gt) 1574 NewSubReg = PPC::sub_lt; 1575 1576 SubRegsToUpdate.push_back(std::make_pair(&(UseMI->getOperand(3)), 1577 NewSubReg)); 1578 } else // We need to abort on a user we don't understand. 1579 return false; 1580 } 1581 1582 // Create a new virtual register to hold the value of the CR set by the 1583 // record-form instruction. If the instruction was not previously in 1584 // record form, then set the kill flag on the CR. 1585 CmpInstr->eraseFromParent(); 1586 1587 MachineBasicBlock::iterator MII = MI; 1588 BuildMI(*MI->getParent(), std::next(MII), MI->getDebugLoc(), 1589 get(TargetOpcode::COPY), CRReg) 1590 .addReg(PPC::CR0, MIOpC != NewOpC ? RegState::Kill : 0); 1591 1592 if (MIOpC != NewOpC) { 1593 // We need to be careful here: we're replacing one instruction with 1594 // another, and we need to make sure that we get all of the right 1595 // implicit uses and defs. On the other hand, the caller may be holding 1596 // an iterator to this instruction, and so we can't delete it (this is 1597 // specifically the case if this is the instruction directly after the 1598 // compare). 1599 1600 const MCInstrDesc &NewDesc = get(NewOpC); 1601 MI->setDesc(NewDesc); 1602 1603 if (NewDesc.ImplicitDefs) 1604 for (const uint16_t *ImpDefs = NewDesc.getImplicitDefs(); 1605 *ImpDefs; ++ImpDefs) 1606 if (!MI->definesRegister(*ImpDefs)) 1607 MI->addOperand(*MI->getParent()->getParent(), 1608 MachineOperand::CreateReg(*ImpDefs, true, true)); 1609 if (NewDesc.ImplicitUses) 1610 for (const uint16_t *ImpUses = NewDesc.getImplicitUses(); 1611 *ImpUses; ++ImpUses) 1612 if (!MI->readsRegister(*ImpUses)) 1613 MI->addOperand(*MI->getParent()->getParent(), 1614 MachineOperand::CreateReg(*ImpUses, false, true)); 1615 } 1616 1617 // Modify the condition code of operands in OperandsToUpdate. 1618 // Since we have SUB(r1, r2) and CMP(r2, r1), the condition code needs to 1619 // be changed from r2 > r1 to r1 < r2, from r2 < r1 to r1 > r2, etc. 1620 for (unsigned i = 0, e = PredsToUpdate.size(); i < e; i++) 1621 PredsToUpdate[i].first->setImm(PredsToUpdate[i].second); 1622 1623 for (unsigned i = 0, e = SubRegsToUpdate.size(); i < e; i++) 1624 SubRegsToUpdate[i].first->setSubReg(SubRegsToUpdate[i].second); 1625 1626 return true; 1627} 1628 1629/// GetInstSize - Return the number of bytes of code the specified 1630/// instruction may be. This returns the maximum number of bytes. 1631/// 1632unsigned PPCInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const { 1633 unsigned Opcode = MI->getOpcode(); 1634 1635 if (Opcode == PPC::INLINEASM) { 1636 const MachineFunction *MF = MI->getParent()->getParent(); 1637 const char *AsmStr = MI->getOperand(0).getSymbolName(); 1638 return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo()); 1639 } else if (Opcode == TargetOpcode::STACKMAP) { 1640 return MI->getOperand(1).getImm(); 1641 } else if (Opcode == TargetOpcode::PATCHPOINT) { 1642 PatchPointOpers Opers(MI); 1643 return Opers.getMetaOper(PatchPointOpers::NBytesPos).getImm(); 1644 } else { 1645 const MCInstrDesc &Desc = get(Opcode); 1646 return Desc.getSize(); 1647 } 1648} 1649 1650