1 //===-------------- PPCMIPeephole.cpp - MI Peephole Cleanups -------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===---------------------------------------------------------------------===// 8 // 9 // This pass performs peephole optimizations to clean up ugly code 10 // sequences at the MachineInstruction layer. It runs at the end of 11 // the SSA phases, following VSX swap removal. A pass of dead code 12 // elimination follows this one for quick clean-up of any dead 13 // instructions introduced here. Although we could do this as callbacks 14 // from the generic peephole pass, this would have a couple of bad 15 // effects: it might remove optimization opportunities for VSX swap 16 // removal, and it would miss cleanups made possible following VSX 17 // swap removal. 18 // 19 //===---------------------------------------------------------------------===// 20 21 #include "MCTargetDesc/PPCMCTargetDesc.h" 22 #include "MCTargetDesc/PPCPredicates.h" 23 #include "PPC.h" 24 #include "PPCInstrBuilder.h" 25 #include "PPCInstrInfo.h" 26 #include "PPCMachineFunctionInfo.h" 27 #include "PPCTargetMachine.h" 28 #include "llvm/ADT/Statistic.h" 29 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" 30 #include "llvm/CodeGen/MachineDominators.h" 31 #include "llvm/CodeGen/MachineFunctionPass.h" 32 #include "llvm/CodeGen/MachineInstrBuilder.h" 33 #include "llvm/CodeGen/MachinePostDominators.h" 34 #include "llvm/CodeGen/MachineRegisterInfo.h" 35 #include "llvm/InitializePasses.h" 36 #include "llvm/Support/Debug.h" 37 38 using namespace llvm; 39 40 #define DEBUG_TYPE "ppc-mi-peepholes" 41 42 STATISTIC(RemoveTOCSave, "Number of TOC saves removed"); 43 STATISTIC(MultiTOCSaves, 44 "Number of functions with multiple TOC saves that must be kept"); 45 STATISTIC(NumTOCSavesInPrologue, "Number of TOC saves placed in the prologue"); 46 STATISTIC(NumEliminatedSExt, "Number of eliminated sign-extensions"); 47 STATISTIC(NumEliminatedZExt, "Number of eliminated zero-extensions"); 48 STATISTIC(NumOptADDLIs, "Number of optimized ADD instruction fed by LI"); 49 STATISTIC(NumConvertedToImmediateForm, 50 "Number of instructions converted to their immediate form"); 51 STATISTIC(NumFunctionsEnteredInMIPeephole, 52 "Number of functions entered in PPC MI Peepholes"); 53 STATISTIC(NumFixedPointIterations, 54 "Number of fixed-point iterations converting reg-reg instructions " 55 "to reg-imm ones"); 56 STATISTIC(NumRotatesCollapsed, 57 "Number of pairs of rotate left, clear left/right collapsed"); 58 STATISTIC(NumEXTSWAndSLDICombined, 59 "Number of pairs of EXTSW and SLDI combined as EXTSWSLI"); 60 STATISTIC(NumLoadImmZeroFoldedAndRemoved, 61 "Number of LI(8) reg, 0 that are folded to r0 and removed"); 62 63 static cl::opt<bool> 64 FixedPointRegToImm("ppc-reg-to-imm-fixed-point", cl::Hidden, cl::init(true), 65 cl::desc("Iterate to a fixed point when attempting to " 66 "convert reg-reg instructions to reg-imm")); 67 68 static cl::opt<bool> 69 ConvertRegReg("ppc-convert-rr-to-ri", cl::Hidden, cl::init(true), 70 cl::desc("Convert eligible reg+reg instructions to reg+imm")); 71 72 static cl::opt<bool> 73 EnableSExtElimination("ppc-eliminate-signext", 74 cl::desc("enable elimination of sign-extensions"), 75 cl::init(false), cl::Hidden); 76 77 static cl::opt<bool> 78 EnableZExtElimination("ppc-eliminate-zeroext", 79 cl::desc("enable elimination of zero-extensions"), 80 cl::init(false), cl::Hidden); 81 82 namespace { 83 84 struct PPCMIPeephole : public MachineFunctionPass { 85 86 static char ID; 87 const PPCInstrInfo *TII; 88 MachineFunction *MF; 89 MachineRegisterInfo *MRI; 90 91 PPCMIPeephole() : MachineFunctionPass(ID) { 92 initializePPCMIPeepholePass(*PassRegistry::getPassRegistry()); 93 } 94 95 private: 96 MachineDominatorTree *MDT; 97 MachinePostDominatorTree *MPDT; 98 MachineBlockFrequencyInfo *MBFI; 99 uint64_t EntryFreq; 100 101 // Initialize class variables. 102 void initialize(MachineFunction &MFParm); 103 104 // Perform peepholes. 105 bool simplifyCode(void); 106 107 // Perform peepholes. 108 bool eliminateRedundantCompare(void); 109 bool eliminateRedundantTOCSaves(std::map<MachineInstr *, bool> &TOCSaves); 110 bool combineSEXTAndSHL(MachineInstr &MI, MachineInstr *&ToErase); 111 bool emitRLDICWhenLoweringJumpTables(MachineInstr &MI); 112 void UpdateTOCSaves(std::map<MachineInstr *, bool> &TOCSaves, 113 MachineInstr *MI); 114 115 public: 116 117 void getAnalysisUsage(AnalysisUsage &AU) const override { 118 AU.addRequired<MachineDominatorTree>(); 119 AU.addRequired<MachinePostDominatorTree>(); 120 AU.addRequired<MachineBlockFrequencyInfo>(); 121 AU.addPreserved<MachineDominatorTree>(); 122 AU.addPreserved<MachinePostDominatorTree>(); 123 AU.addPreserved<MachineBlockFrequencyInfo>(); 124 MachineFunctionPass::getAnalysisUsage(AU); 125 } 126 127 // Main entry point for this pass. 128 bool runOnMachineFunction(MachineFunction &MF) override { 129 initialize(MF); 130 // At this point, TOC pointer should not be used in a function that uses 131 // PC-Relative addressing. 132 assert((MF.getRegInfo().use_empty(PPC::X2) || 133 !MF.getSubtarget<PPCSubtarget>().isUsingPCRelativeCalls()) && 134 "TOC pointer used in a function using PC-Relative addressing!"); 135 if (skipFunction(MF.getFunction())) 136 return false; 137 return simplifyCode(); 138 } 139 }; 140 141 // Initialize class variables. 142 void PPCMIPeephole::initialize(MachineFunction &MFParm) { 143 MF = &MFParm; 144 MRI = &MF->getRegInfo(); 145 MDT = &getAnalysis<MachineDominatorTree>(); 146 MPDT = &getAnalysis<MachinePostDominatorTree>(); 147 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 148 EntryFreq = MBFI->getEntryFreq(); 149 TII = MF->getSubtarget<PPCSubtarget>().getInstrInfo(); 150 LLVM_DEBUG(dbgs() << "*** PowerPC MI peephole pass ***\n\n"); 151 LLVM_DEBUG(MF->dump()); 152 } 153 154 static MachineInstr *getVRegDefOrNull(MachineOperand *Op, 155 MachineRegisterInfo *MRI) { 156 assert(Op && "Invalid Operand!"); 157 if (!Op->isReg()) 158 return nullptr; 159 160 Register Reg = Op->getReg(); 161 if (!Register::isVirtualRegister(Reg)) 162 return nullptr; 163 164 return MRI->getVRegDef(Reg); 165 } 166 167 // This function returns number of known zero bits in output of MI 168 // starting from the most significant bit. 169 static unsigned 170 getKnownLeadingZeroCount(MachineInstr *MI, const PPCInstrInfo *TII) { 171 unsigned Opcode = MI->getOpcode(); 172 if (Opcode == PPC::RLDICL || Opcode == PPC::RLDICL_rec || 173 Opcode == PPC::RLDCL || Opcode == PPC::RLDCL_rec) 174 return MI->getOperand(3).getImm(); 175 176 if ((Opcode == PPC::RLDIC || Opcode == PPC::RLDIC_rec) && 177 MI->getOperand(3).getImm() <= 63 - MI->getOperand(2).getImm()) 178 return MI->getOperand(3).getImm(); 179 180 if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINM_rec || 181 Opcode == PPC::RLWNM || Opcode == PPC::RLWNM_rec || 182 Opcode == PPC::RLWINM8 || Opcode == PPC::RLWNM8) && 183 MI->getOperand(3).getImm() <= MI->getOperand(4).getImm()) 184 return 32 + MI->getOperand(3).getImm(); 185 186 if (Opcode == PPC::ANDI_rec) { 187 uint16_t Imm = MI->getOperand(2).getImm(); 188 return 48 + countLeadingZeros(Imm); 189 } 190 191 if (Opcode == PPC::CNTLZW || Opcode == PPC::CNTLZW_rec || 192 Opcode == PPC::CNTTZW || Opcode == PPC::CNTTZW_rec || 193 Opcode == PPC::CNTLZW8 || Opcode == PPC::CNTTZW8) 194 // The result ranges from 0 to 32. 195 return 58; 196 197 if (Opcode == PPC::CNTLZD || Opcode == PPC::CNTLZD_rec || 198 Opcode == PPC::CNTTZD || Opcode == PPC::CNTTZD_rec) 199 // The result ranges from 0 to 64. 200 return 57; 201 202 if (Opcode == PPC::LHZ || Opcode == PPC::LHZX || 203 Opcode == PPC::LHZ8 || Opcode == PPC::LHZX8 || 204 Opcode == PPC::LHZU || Opcode == PPC::LHZUX || 205 Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8) 206 return 48; 207 208 if (Opcode == PPC::LBZ || Opcode == PPC::LBZX || 209 Opcode == PPC::LBZ8 || Opcode == PPC::LBZX8 || 210 Opcode == PPC::LBZU || Opcode == PPC::LBZUX || 211 Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8) 212 return 56; 213 214 if (TII->isZeroExtended(*MI)) 215 return 32; 216 217 return 0; 218 } 219 220 // This function maintains a map for the pairs <TOC Save Instr, Keep> 221 // Each time a new TOC save is encountered, it checks if any of the existing 222 // ones are dominated by the new one. If so, it marks the existing one as 223 // redundant by setting it's entry in the map as false. It then adds the new 224 // instruction to the map with either true or false depending on if any 225 // existing instructions dominated the new one. 226 void PPCMIPeephole::UpdateTOCSaves( 227 std::map<MachineInstr *, bool> &TOCSaves, MachineInstr *MI) { 228 assert(TII->isTOCSaveMI(*MI) && "Expecting a TOC save instruction here"); 229 // FIXME: Saving TOC in prologue hasn't been implemented well in AIX ABI part, 230 // here only support it under ELFv2. 231 if (MF->getSubtarget<PPCSubtarget>().isELFv2ABI()) { 232 PPCFunctionInfo *FI = MF->getInfo<PPCFunctionInfo>(); 233 234 MachineBasicBlock *Entry = &MF->front(); 235 uint64_t CurrBlockFreq = MBFI->getBlockFreq(MI->getParent()).getFrequency(); 236 237 // If the block in which the TOC save resides is in a block that 238 // post-dominates Entry, or a block that is hotter than entry (keep in mind 239 // that early MachineLICM has already run so the TOC save won't be hoisted) 240 // we can just do the save in the prologue. 241 if (CurrBlockFreq > EntryFreq || MPDT->dominates(MI->getParent(), Entry)) 242 FI->setMustSaveTOC(true); 243 244 // If we are saving the TOC in the prologue, all the TOC saves can be 245 // removed from the code. 246 if (FI->mustSaveTOC()) { 247 for (auto &TOCSave : TOCSaves) 248 TOCSave.second = false; 249 // Add new instruction to map. 250 TOCSaves[MI] = false; 251 return; 252 } 253 } 254 255 bool Keep = true; 256 for (auto It = TOCSaves.begin(); It != TOCSaves.end(); It++ ) { 257 MachineInstr *CurrInst = It->first; 258 // If new instruction dominates an existing one, mark existing one as 259 // redundant. 260 if (It->second && MDT->dominates(MI, CurrInst)) 261 It->second = false; 262 // Check if the new instruction is redundant. 263 if (MDT->dominates(CurrInst, MI)) { 264 Keep = false; 265 break; 266 } 267 } 268 // Add new instruction to map. 269 TOCSaves[MI] = Keep; 270 } 271 272 // This function returns a list of all PHI nodes in the tree starting from 273 // the RootPHI node. We perform a BFS traversal to get an ordered list of nodes. 274 // The list initially only contains the root PHI. When we visit a PHI node, we 275 // add it to the list. We continue to look for other PHI node operands while 276 // there are nodes to visit in the list. The function returns false if the 277 // optimization cannot be applied on this tree. 278 static bool collectUnprimedAccPHIs(MachineRegisterInfo *MRI, 279 MachineInstr *RootPHI, 280 SmallVectorImpl<MachineInstr *> &PHIs) { 281 PHIs.push_back(RootPHI); 282 unsigned VisitedIndex = 0; 283 while (VisitedIndex < PHIs.size()) { 284 MachineInstr *VisitedPHI = PHIs[VisitedIndex]; 285 for (unsigned PHIOp = 1, NumOps = VisitedPHI->getNumOperands(); 286 PHIOp != NumOps; PHIOp += 2) { 287 Register RegOp = VisitedPHI->getOperand(PHIOp).getReg(); 288 if (!Register::isVirtualRegister(RegOp)) 289 return false; 290 MachineInstr *Instr = MRI->getVRegDef(RegOp); 291 // While collecting the PHI nodes, we check if they can be converted (i.e. 292 // all the operands are either copies, implicit defs or PHI nodes). 293 unsigned Opcode = Instr->getOpcode(); 294 if (Opcode == PPC::COPY) { 295 Register Reg = Instr->getOperand(1).getReg(); 296 if (!Register::isVirtualRegister(Reg) || 297 MRI->getRegClass(Reg) != &PPC::ACCRCRegClass) 298 return false; 299 } else if (Opcode != PPC::IMPLICIT_DEF && Opcode != PPC::PHI) 300 return false; 301 // If we detect a cycle in the PHI nodes, we exit. It would be 302 // possible to change cycles as well, but that would add a lot 303 // of complexity for a case that is unlikely to occur with MMA 304 // code. 305 if (Opcode != PPC::PHI) 306 continue; 307 if (llvm::is_contained(PHIs, Instr)) 308 return false; 309 PHIs.push_back(Instr); 310 } 311 VisitedIndex++; 312 } 313 return true; 314 } 315 316 // This function changes the unprimed accumulator PHI nodes in the PHIs list to 317 // primed accumulator PHI nodes. The list is traversed in reverse order to 318 // change all the PHI operands of a PHI node before changing the node itself. 319 // We keep a map to associate each changed PHI node to its non-changed form. 320 static void convertUnprimedAccPHIs(const PPCInstrInfo *TII, 321 MachineRegisterInfo *MRI, 322 SmallVectorImpl<MachineInstr *> &PHIs, 323 Register Dst) { 324 DenseMap<MachineInstr *, MachineInstr *> ChangedPHIMap; 325 for (auto It = PHIs.rbegin(), End = PHIs.rend(); It != End; ++It) { 326 MachineInstr *PHI = *It; 327 SmallVector<std::pair<MachineOperand, MachineOperand>, 4> PHIOps; 328 // We check if the current PHI node can be changed by looking at its 329 // operands. If all the operands are either copies from primed 330 // accumulators, implicit definitions or other unprimed accumulator 331 // PHI nodes, we change it. 332 for (unsigned PHIOp = 1, NumOps = PHI->getNumOperands(); PHIOp != NumOps; 333 PHIOp += 2) { 334 Register RegOp = PHI->getOperand(PHIOp).getReg(); 335 MachineInstr *PHIInput = MRI->getVRegDef(RegOp); 336 unsigned Opcode = PHIInput->getOpcode(); 337 assert((Opcode == PPC::COPY || Opcode == PPC::IMPLICIT_DEF || 338 Opcode == PPC::PHI) && 339 "Unexpected instruction"); 340 if (Opcode == PPC::COPY) { 341 assert(MRI->getRegClass(PHIInput->getOperand(1).getReg()) == 342 &PPC::ACCRCRegClass && 343 "Unexpected register class"); 344 PHIOps.push_back({PHIInput->getOperand(1), PHI->getOperand(PHIOp + 1)}); 345 } else if (Opcode == PPC::IMPLICIT_DEF) { 346 Register AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass); 347 BuildMI(*PHIInput->getParent(), PHIInput, PHIInput->getDebugLoc(), 348 TII->get(PPC::IMPLICIT_DEF), AccReg); 349 PHIOps.push_back({MachineOperand::CreateReg(AccReg, false), 350 PHI->getOperand(PHIOp + 1)}); 351 } else if (Opcode == PPC::PHI) { 352 // We found a PHI operand. At this point we know this operand 353 // has already been changed so we get its associated changed form 354 // from the map. 355 assert(ChangedPHIMap.count(PHIInput) == 1 && 356 "This PHI node should have already been changed."); 357 MachineInstr *PrimedAccPHI = ChangedPHIMap.lookup(PHIInput); 358 PHIOps.push_back({MachineOperand::CreateReg( 359 PrimedAccPHI->getOperand(0).getReg(), false), 360 PHI->getOperand(PHIOp + 1)}); 361 } 362 } 363 Register AccReg = Dst; 364 // If the PHI node we are changing is the root node, the register it defines 365 // will be the destination register of the original copy (of the PHI def). 366 // For all other PHI's in the list, we need to create another primed 367 // accumulator virtual register as the PHI will no longer define the 368 // unprimed accumulator. 369 if (PHI != PHIs[0]) 370 AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass); 371 MachineInstrBuilder NewPHI = BuildMI( 372 *PHI->getParent(), PHI, PHI->getDebugLoc(), TII->get(PPC::PHI), AccReg); 373 for (auto RegMBB : PHIOps) 374 NewPHI.add(RegMBB.first).add(RegMBB.second); 375 ChangedPHIMap[PHI] = NewPHI.getInstr(); 376 } 377 } 378 379 // Perform peephole optimizations. 380 bool PPCMIPeephole::simplifyCode(void) { 381 bool Simplified = false; 382 MachineInstr* ToErase = nullptr; 383 std::map<MachineInstr *, bool> TOCSaves; 384 const TargetRegisterInfo *TRI = &TII->getRegisterInfo(); 385 NumFunctionsEnteredInMIPeephole++; 386 if (ConvertRegReg) { 387 // Fixed-point conversion of reg/reg instructions fed by load-immediate 388 // into reg/imm instructions. FIXME: This is expensive, control it with 389 // an option. 390 bool SomethingChanged = false; 391 do { 392 NumFixedPointIterations++; 393 SomethingChanged = false; 394 for (MachineBasicBlock &MBB : *MF) { 395 for (MachineInstr &MI : MBB) { 396 if (MI.isDebugInstr()) 397 continue; 398 399 if (TII->convertToImmediateForm(MI)) { 400 // We don't erase anything in case the def has other uses. Let DCE 401 // remove it if it can be removed. 402 LLVM_DEBUG(dbgs() << "Converted instruction to imm form: "); 403 LLVM_DEBUG(MI.dump()); 404 NumConvertedToImmediateForm++; 405 SomethingChanged = true; 406 Simplified = true; 407 continue; 408 } 409 } 410 } 411 } while (SomethingChanged && FixedPointRegToImm); 412 } 413 414 for (MachineBasicBlock &MBB : *MF) { 415 for (MachineInstr &MI : MBB) { 416 417 // If the previous instruction was marked for elimination, 418 // remove it now. 419 if (ToErase) { 420 ToErase->eraseFromParent(); 421 ToErase = nullptr; 422 } 423 424 // Ignore debug instructions. 425 if (MI.isDebugInstr()) 426 continue; 427 428 // Per-opcode peepholes. 429 switch (MI.getOpcode()) { 430 431 default: 432 break; 433 case PPC::COPY: { 434 Register Src = MI.getOperand(1).getReg(); 435 Register Dst = MI.getOperand(0).getReg(); 436 if (!Register::isVirtualRegister(Src) || 437 !Register::isVirtualRegister(Dst)) 438 break; 439 if (MRI->getRegClass(Src) != &PPC::UACCRCRegClass || 440 MRI->getRegClass(Dst) != &PPC::ACCRCRegClass) 441 break; 442 443 // We are copying an unprimed accumulator to a primed accumulator. 444 // If the input to the copy is a PHI that is fed only by (i) copies in 445 // the other direction (ii) implicitly defined unprimed accumulators or 446 // (iii) other PHI nodes satisfying (i) and (ii), we can change 447 // the PHI to a PHI on primed accumulators (as long as we also change 448 // its operands). To detect and change such copies, we first get a list 449 // of all the PHI nodes starting from the root PHI node in BFS order. 450 // We then visit all these PHI nodes to check if they can be changed to 451 // primed accumulator PHI nodes and if so, we change them. 452 MachineInstr *RootPHI = MRI->getVRegDef(Src); 453 if (RootPHI->getOpcode() != PPC::PHI) 454 break; 455 456 SmallVector<MachineInstr *, 4> PHIs; 457 if (!collectUnprimedAccPHIs(MRI, RootPHI, PHIs)) 458 break; 459 460 convertUnprimedAccPHIs(TII, MRI, PHIs, Dst); 461 462 ToErase = &MI; 463 break; 464 } 465 case PPC::LI: 466 case PPC::LI8: { 467 // If we are materializing a zero, look for any use operands for which 468 // zero means immediate zero. All such operands can be replaced with 469 // PPC::ZERO. 470 if (!MI.getOperand(1).isImm() || MI.getOperand(1).getImm() != 0) 471 break; 472 unsigned MIDestReg = MI.getOperand(0).getReg(); 473 for (MachineInstr& UseMI : MRI->use_instructions(MIDestReg)) 474 Simplified |= TII->onlyFoldImmediate(UseMI, MI, MIDestReg); 475 if (MRI->use_nodbg_empty(MIDestReg)) { 476 ++NumLoadImmZeroFoldedAndRemoved; 477 ToErase = &MI; 478 } 479 break; 480 } 481 case PPC::STW: 482 case PPC::STD: { 483 MachineFrameInfo &MFI = MF->getFrameInfo(); 484 if (MFI.hasVarSizedObjects() || 485 (!MF->getSubtarget<PPCSubtarget>().isELFv2ABI() && 486 !MF->getSubtarget<PPCSubtarget>().isAIXABI())) 487 break; 488 // When encountering a TOC save instruction, call UpdateTOCSaves 489 // to add it to the TOCSaves map and mark any existing TOC saves 490 // it dominates as redundant. 491 if (TII->isTOCSaveMI(MI)) 492 UpdateTOCSaves(TOCSaves, &MI); 493 break; 494 } 495 case PPC::XXPERMDI: { 496 // Perform simplifications of 2x64 vector swaps and splats. 497 // A swap is identified by an immediate value of 2, and a splat 498 // is identified by an immediate value of 0 or 3. 499 int Immed = MI.getOperand(3).getImm(); 500 501 if (Immed == 1) 502 break; 503 504 // For each of these simplifications, we need the two source 505 // regs to match. Unfortunately, MachineCSE ignores COPY and 506 // SUBREG_TO_REG, so for example we can see 507 // XXPERMDI t, SUBREG_TO_REG(s), SUBREG_TO_REG(s), immed. 508 // We have to look through chains of COPY and SUBREG_TO_REG 509 // to find the real source values for comparison. 510 unsigned TrueReg1 = 511 TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI); 512 unsigned TrueReg2 = 513 TRI->lookThruCopyLike(MI.getOperand(2).getReg(), MRI); 514 515 if (!(TrueReg1 == TrueReg2 && Register::isVirtualRegister(TrueReg1))) 516 break; 517 518 MachineInstr *DefMI = MRI->getVRegDef(TrueReg1); 519 520 if (!DefMI) 521 break; 522 523 unsigned DefOpc = DefMI->getOpcode(); 524 525 // If this is a splat fed by a splatting load, the splat is 526 // redundant. Replace with a copy. This doesn't happen directly due 527 // to code in PPCDAGToDAGISel.cpp, but it can happen when converting 528 // a load of a double to a vector of 64-bit integers. 529 auto isConversionOfLoadAndSplat = [=]() -> bool { 530 if (DefOpc != PPC::XVCVDPSXDS && DefOpc != PPC::XVCVDPUXDS) 531 return false; 532 unsigned FeedReg1 = 533 TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI); 534 if (Register::isVirtualRegister(FeedReg1)) { 535 MachineInstr *LoadMI = MRI->getVRegDef(FeedReg1); 536 if (LoadMI && LoadMI->getOpcode() == PPC::LXVDSX) 537 return true; 538 } 539 return false; 540 }; 541 if ((Immed == 0 || Immed == 3) && 542 (DefOpc == PPC::LXVDSX || isConversionOfLoadAndSplat())) { 543 LLVM_DEBUG(dbgs() << "Optimizing load-and-splat/splat " 544 "to load-and-splat/copy: "); 545 LLVM_DEBUG(MI.dump()); 546 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 547 MI.getOperand(0).getReg()) 548 .add(MI.getOperand(1)); 549 ToErase = &MI; 550 Simplified = true; 551 } 552 553 // If this is a splat or a swap fed by another splat, we 554 // can replace it with a copy. 555 if (DefOpc == PPC::XXPERMDI) { 556 unsigned DefReg1 = DefMI->getOperand(1).getReg(); 557 unsigned DefReg2 = DefMI->getOperand(2).getReg(); 558 unsigned DefImmed = DefMI->getOperand(3).getImm(); 559 560 // If the two inputs are not the same register, check to see if 561 // they originate from the same virtual register after only 562 // copy-like instructions. 563 if (DefReg1 != DefReg2) { 564 unsigned FeedReg1 = TRI->lookThruCopyLike(DefReg1, MRI); 565 unsigned FeedReg2 = TRI->lookThruCopyLike(DefReg2, MRI); 566 567 if (!(FeedReg1 == FeedReg2 && 568 Register::isVirtualRegister(FeedReg1))) 569 break; 570 } 571 572 if (DefImmed == 0 || DefImmed == 3) { 573 LLVM_DEBUG(dbgs() << "Optimizing splat/swap or splat/splat " 574 "to splat/copy: "); 575 LLVM_DEBUG(MI.dump()); 576 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 577 MI.getOperand(0).getReg()) 578 .add(MI.getOperand(1)); 579 ToErase = &MI; 580 Simplified = true; 581 } 582 583 // If this is a splat fed by a swap, we can simplify modify 584 // the splat to splat the other value from the swap's input 585 // parameter. 586 else if ((Immed == 0 || Immed == 3) && DefImmed == 2) { 587 LLVM_DEBUG(dbgs() << "Optimizing swap/splat => splat: "); 588 LLVM_DEBUG(MI.dump()); 589 MI.getOperand(1).setReg(DefReg1); 590 MI.getOperand(2).setReg(DefReg2); 591 MI.getOperand(3).setImm(3 - Immed); 592 Simplified = true; 593 } 594 595 // If this is a swap fed by a swap, we can replace it 596 // with a copy from the first swap's input. 597 else if (Immed == 2 && DefImmed == 2) { 598 LLVM_DEBUG(dbgs() << "Optimizing swap/swap => copy: "); 599 LLVM_DEBUG(MI.dump()); 600 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 601 MI.getOperand(0).getReg()) 602 .add(DefMI->getOperand(1)); 603 ToErase = &MI; 604 Simplified = true; 605 } 606 } else if ((Immed == 0 || Immed == 3) && DefOpc == PPC::XXPERMDIs && 607 (DefMI->getOperand(2).getImm() == 0 || 608 DefMI->getOperand(2).getImm() == 3)) { 609 // Splat fed by another splat - switch the output of the first 610 // and remove the second. 611 DefMI->getOperand(0).setReg(MI.getOperand(0).getReg()); 612 ToErase = &MI; 613 Simplified = true; 614 LLVM_DEBUG(dbgs() << "Removing redundant splat: "); 615 LLVM_DEBUG(MI.dump()); 616 } 617 break; 618 } 619 case PPC::VSPLTB: 620 case PPC::VSPLTH: 621 case PPC::XXSPLTW: { 622 unsigned MyOpcode = MI.getOpcode(); 623 unsigned OpNo = MyOpcode == PPC::XXSPLTW ? 1 : 2; 624 unsigned TrueReg = 625 TRI->lookThruCopyLike(MI.getOperand(OpNo).getReg(), MRI); 626 if (!Register::isVirtualRegister(TrueReg)) 627 break; 628 MachineInstr *DefMI = MRI->getVRegDef(TrueReg); 629 if (!DefMI) 630 break; 631 unsigned DefOpcode = DefMI->getOpcode(); 632 auto isConvertOfSplat = [=]() -> bool { 633 if (DefOpcode != PPC::XVCVSPSXWS && DefOpcode != PPC::XVCVSPUXWS) 634 return false; 635 Register ConvReg = DefMI->getOperand(1).getReg(); 636 if (!Register::isVirtualRegister(ConvReg)) 637 return false; 638 MachineInstr *Splt = MRI->getVRegDef(ConvReg); 639 return Splt && (Splt->getOpcode() == PPC::LXVWSX || 640 Splt->getOpcode() == PPC::XXSPLTW); 641 }; 642 bool AlreadySplat = (MyOpcode == DefOpcode) || 643 (MyOpcode == PPC::VSPLTB && DefOpcode == PPC::VSPLTBs) || 644 (MyOpcode == PPC::VSPLTH && DefOpcode == PPC::VSPLTHs) || 645 (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::XXSPLTWs) || 646 (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::LXVWSX) || 647 (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::MTVSRWS)|| 648 (MyOpcode == PPC::XXSPLTW && isConvertOfSplat()); 649 // If the instruction[s] that feed this splat have already splat 650 // the value, this splat is redundant. 651 if (AlreadySplat) { 652 LLVM_DEBUG(dbgs() << "Changing redundant splat to a copy: "); 653 LLVM_DEBUG(MI.dump()); 654 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 655 MI.getOperand(0).getReg()) 656 .add(MI.getOperand(OpNo)); 657 ToErase = &MI; 658 Simplified = true; 659 } 660 // Splat fed by a shift. Usually when we align value to splat into 661 // vector element zero. 662 if (DefOpcode == PPC::XXSLDWI) { 663 Register ShiftRes = DefMI->getOperand(0).getReg(); 664 Register ShiftOp1 = DefMI->getOperand(1).getReg(); 665 Register ShiftOp2 = DefMI->getOperand(2).getReg(); 666 unsigned ShiftImm = DefMI->getOperand(3).getImm(); 667 unsigned SplatImm = 668 MI.getOperand(MyOpcode == PPC::XXSPLTW ? 2 : 1).getImm(); 669 if (ShiftOp1 == ShiftOp2) { 670 unsigned NewElem = (SplatImm + ShiftImm) & 0x3; 671 if (MRI->hasOneNonDBGUse(ShiftRes)) { 672 LLVM_DEBUG(dbgs() << "Removing redundant shift: "); 673 LLVM_DEBUG(DefMI->dump()); 674 ToErase = DefMI; 675 } 676 Simplified = true; 677 LLVM_DEBUG(dbgs() << "Changing splat immediate from " << SplatImm 678 << " to " << NewElem << " in instruction: "); 679 LLVM_DEBUG(MI.dump()); 680 MI.getOperand(1).setReg(ShiftOp1); 681 MI.getOperand(2).setImm(NewElem); 682 } 683 } 684 break; 685 } 686 case PPC::XVCVDPSP: { 687 // If this is a DP->SP conversion fed by an FRSP, the FRSP is redundant. 688 unsigned TrueReg = 689 TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI); 690 if (!Register::isVirtualRegister(TrueReg)) 691 break; 692 MachineInstr *DefMI = MRI->getVRegDef(TrueReg); 693 694 // This can occur when building a vector of single precision or integer 695 // values. 696 if (DefMI && DefMI->getOpcode() == PPC::XXPERMDI) { 697 unsigned DefsReg1 = 698 TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI); 699 unsigned DefsReg2 = 700 TRI->lookThruCopyLike(DefMI->getOperand(2).getReg(), MRI); 701 if (!Register::isVirtualRegister(DefsReg1) || 702 !Register::isVirtualRegister(DefsReg2)) 703 break; 704 MachineInstr *P1 = MRI->getVRegDef(DefsReg1); 705 MachineInstr *P2 = MRI->getVRegDef(DefsReg2); 706 707 if (!P1 || !P2) 708 break; 709 710 // Remove the passed FRSP/XSRSP instruction if it only feeds this MI 711 // and set any uses of that FRSP/XSRSP (in this MI) to the source of 712 // the FRSP/XSRSP. 713 auto removeFRSPIfPossible = [&](MachineInstr *RoundInstr) { 714 unsigned Opc = RoundInstr->getOpcode(); 715 if ((Opc == PPC::FRSP || Opc == PPC::XSRSP) && 716 MRI->hasOneNonDBGUse(RoundInstr->getOperand(0).getReg())) { 717 Simplified = true; 718 Register ConvReg1 = RoundInstr->getOperand(1).getReg(); 719 Register FRSPDefines = RoundInstr->getOperand(0).getReg(); 720 MachineInstr &Use = *(MRI->use_instr_nodbg_begin(FRSPDefines)); 721 for (int i = 0, e = Use.getNumOperands(); i < e; ++i) 722 if (Use.getOperand(i).isReg() && 723 Use.getOperand(i).getReg() == FRSPDefines) 724 Use.getOperand(i).setReg(ConvReg1); 725 LLVM_DEBUG(dbgs() << "Removing redundant FRSP/XSRSP:\n"); 726 LLVM_DEBUG(RoundInstr->dump()); 727 LLVM_DEBUG(dbgs() << "As it feeds instruction:\n"); 728 LLVM_DEBUG(MI.dump()); 729 LLVM_DEBUG(dbgs() << "Through instruction:\n"); 730 LLVM_DEBUG(DefMI->dump()); 731 RoundInstr->eraseFromParent(); 732 } 733 }; 734 735 // If the input to XVCVDPSP is a vector that was built (even 736 // partially) out of FRSP's, the FRSP(s) can safely be removed 737 // since this instruction performs the same operation. 738 if (P1 != P2) { 739 removeFRSPIfPossible(P1); 740 removeFRSPIfPossible(P2); 741 break; 742 } 743 removeFRSPIfPossible(P1); 744 } 745 break; 746 } 747 case PPC::EXTSH: 748 case PPC::EXTSH8: 749 case PPC::EXTSH8_32_64: { 750 if (!EnableSExtElimination) break; 751 Register NarrowReg = MI.getOperand(1).getReg(); 752 if (!Register::isVirtualRegister(NarrowReg)) 753 break; 754 755 MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg); 756 // If we've used a zero-extending load that we will sign-extend, 757 // just do a sign-extending load. 758 if (SrcMI->getOpcode() == PPC::LHZ || 759 SrcMI->getOpcode() == PPC::LHZX) { 760 if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg())) 761 break; 762 auto is64Bit = [] (unsigned Opcode) { 763 return Opcode == PPC::EXTSH8; 764 }; 765 auto isXForm = [] (unsigned Opcode) { 766 return Opcode == PPC::LHZX; 767 }; 768 auto getSextLoadOp = [] (bool is64Bit, bool isXForm) { 769 if (is64Bit) 770 if (isXForm) return PPC::LHAX8; 771 else return PPC::LHA8; 772 else 773 if (isXForm) return PPC::LHAX; 774 else return PPC::LHA; 775 }; 776 unsigned Opc = getSextLoadOp(is64Bit(MI.getOpcode()), 777 isXForm(SrcMI->getOpcode())); 778 LLVM_DEBUG(dbgs() << "Zero-extending load\n"); 779 LLVM_DEBUG(SrcMI->dump()); 780 LLVM_DEBUG(dbgs() << "and sign-extension\n"); 781 LLVM_DEBUG(MI.dump()); 782 LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n"); 783 SrcMI->setDesc(TII->get(Opc)); 784 SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg()); 785 ToErase = &MI; 786 Simplified = true; 787 NumEliminatedSExt++; 788 } 789 break; 790 } 791 case PPC::EXTSW: 792 case PPC::EXTSW_32: 793 case PPC::EXTSW_32_64: { 794 if (!EnableSExtElimination) break; 795 Register NarrowReg = MI.getOperand(1).getReg(); 796 if (!Register::isVirtualRegister(NarrowReg)) 797 break; 798 799 MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg); 800 // If we've used a zero-extending load that we will sign-extend, 801 // just do a sign-extending load. 802 if (SrcMI->getOpcode() == PPC::LWZ || 803 SrcMI->getOpcode() == PPC::LWZX) { 804 if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg())) 805 break; 806 auto is64Bit = [] (unsigned Opcode) { 807 return Opcode == PPC::EXTSW || Opcode == PPC::EXTSW_32_64; 808 }; 809 auto isXForm = [] (unsigned Opcode) { 810 return Opcode == PPC::LWZX; 811 }; 812 auto getSextLoadOp = [] (bool is64Bit, bool isXForm) { 813 if (is64Bit) 814 if (isXForm) return PPC::LWAX; 815 else return PPC::LWA; 816 else 817 if (isXForm) return PPC::LWAX_32; 818 else return PPC::LWA_32; 819 }; 820 unsigned Opc = getSextLoadOp(is64Bit(MI.getOpcode()), 821 isXForm(SrcMI->getOpcode())); 822 LLVM_DEBUG(dbgs() << "Zero-extending load\n"); 823 LLVM_DEBUG(SrcMI->dump()); 824 LLVM_DEBUG(dbgs() << "and sign-extension\n"); 825 LLVM_DEBUG(MI.dump()); 826 LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n"); 827 SrcMI->setDesc(TII->get(Opc)); 828 SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg()); 829 ToErase = &MI; 830 Simplified = true; 831 NumEliminatedSExt++; 832 } else if (MI.getOpcode() == PPC::EXTSW_32_64 && 833 TII->isSignExtended(*SrcMI)) { 834 // We can eliminate EXTSW if the input is known to be already 835 // sign-extended. 836 LLVM_DEBUG(dbgs() << "Removing redundant sign-extension\n"); 837 Register TmpReg = 838 MF->getRegInfo().createVirtualRegister(&PPC::G8RCRegClass); 839 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::IMPLICIT_DEF), 840 TmpReg); 841 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::INSERT_SUBREG), 842 MI.getOperand(0).getReg()) 843 .addReg(TmpReg) 844 .addReg(NarrowReg) 845 .addImm(PPC::sub_32); 846 ToErase = &MI; 847 Simplified = true; 848 NumEliminatedSExt++; 849 } 850 break; 851 } 852 case PPC::RLDICL: { 853 // We can eliminate RLDICL (e.g. for zero-extension) 854 // if all bits to clear are already zero in the input. 855 // This code assume following code sequence for zero-extension. 856 // %6 = COPY %5:sub_32; (optional) 857 // %8 = IMPLICIT_DEF; 858 // %7<def,tied1> = INSERT_SUBREG %8<tied0>, %6, sub_32; 859 if (!EnableZExtElimination) break; 860 861 if (MI.getOperand(2).getImm() != 0) 862 break; 863 864 Register SrcReg = MI.getOperand(1).getReg(); 865 if (!Register::isVirtualRegister(SrcReg)) 866 break; 867 868 MachineInstr *SrcMI = MRI->getVRegDef(SrcReg); 869 if (!(SrcMI && SrcMI->getOpcode() == PPC::INSERT_SUBREG && 870 SrcMI->getOperand(0).isReg() && SrcMI->getOperand(1).isReg())) 871 break; 872 873 MachineInstr *ImpDefMI, *SubRegMI; 874 ImpDefMI = MRI->getVRegDef(SrcMI->getOperand(1).getReg()); 875 SubRegMI = MRI->getVRegDef(SrcMI->getOperand(2).getReg()); 876 if (ImpDefMI->getOpcode() != PPC::IMPLICIT_DEF) break; 877 878 SrcMI = SubRegMI; 879 if (SubRegMI->getOpcode() == PPC::COPY) { 880 Register CopyReg = SubRegMI->getOperand(1).getReg(); 881 if (Register::isVirtualRegister(CopyReg)) 882 SrcMI = MRI->getVRegDef(CopyReg); 883 } 884 885 unsigned KnownZeroCount = getKnownLeadingZeroCount(SrcMI, TII); 886 if (MI.getOperand(3).getImm() <= KnownZeroCount) { 887 LLVM_DEBUG(dbgs() << "Removing redundant zero-extension\n"); 888 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 889 MI.getOperand(0).getReg()) 890 .addReg(SrcReg); 891 ToErase = &MI; 892 Simplified = true; 893 NumEliminatedZExt++; 894 } 895 break; 896 } 897 898 // TODO: Any instruction that has an immediate form fed only by a PHI 899 // whose operands are all load immediate can be folded away. We currently 900 // do this for ADD instructions, but should expand it to arithmetic and 901 // binary instructions with immediate forms in the future. 902 case PPC::ADD4: 903 case PPC::ADD8: { 904 auto isSingleUsePHI = [&](MachineOperand *PhiOp) { 905 assert(PhiOp && "Invalid Operand!"); 906 MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI); 907 908 return DefPhiMI && (DefPhiMI->getOpcode() == PPC::PHI) && 909 MRI->hasOneNonDBGUse(DefPhiMI->getOperand(0).getReg()); 910 }; 911 912 auto dominatesAllSingleUseLIs = [&](MachineOperand *DominatorOp, 913 MachineOperand *PhiOp) { 914 assert(PhiOp && "Invalid Operand!"); 915 assert(DominatorOp && "Invalid Operand!"); 916 MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI); 917 MachineInstr *DefDomMI = getVRegDefOrNull(DominatorOp, MRI); 918 919 // Note: the vregs only show up at odd indices position of PHI Node, 920 // the even indices position save the BB info. 921 for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) { 922 MachineInstr *LiMI = 923 getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI); 924 if (!LiMI || 925 (LiMI->getOpcode() != PPC::LI && LiMI->getOpcode() != PPC::LI8) 926 || !MRI->hasOneNonDBGUse(LiMI->getOperand(0).getReg()) || 927 !MDT->dominates(DefDomMI, LiMI)) 928 return false; 929 } 930 931 return true; 932 }; 933 934 MachineOperand Op1 = MI.getOperand(1); 935 MachineOperand Op2 = MI.getOperand(2); 936 if (isSingleUsePHI(&Op2) && dominatesAllSingleUseLIs(&Op1, &Op2)) 937 std::swap(Op1, Op2); 938 else if (!isSingleUsePHI(&Op1) || !dominatesAllSingleUseLIs(&Op2, &Op1)) 939 break; // We don't have an ADD fed by LI's that can be transformed 940 941 // Now we know that Op1 is the PHI node and Op2 is the dominator 942 Register DominatorReg = Op2.getReg(); 943 944 const TargetRegisterClass *TRC = MI.getOpcode() == PPC::ADD8 945 ? &PPC::G8RC_and_G8RC_NOX0RegClass 946 : &PPC::GPRC_and_GPRC_NOR0RegClass; 947 MRI->setRegClass(DominatorReg, TRC); 948 949 // replace LIs with ADDIs 950 MachineInstr *DefPhiMI = getVRegDefOrNull(&Op1, MRI); 951 for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) { 952 MachineInstr *LiMI = getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI); 953 LLVM_DEBUG(dbgs() << "Optimizing LI to ADDI: "); 954 LLVM_DEBUG(LiMI->dump()); 955 956 // There could be repeated registers in the PHI, e.g: %1 = 957 // PHI %6, <%bb.2>, %8, <%bb.3>, %8, <%bb.6>; So if we've 958 // already replaced the def instruction, skip. 959 if (LiMI->getOpcode() == PPC::ADDI || LiMI->getOpcode() == PPC::ADDI8) 960 continue; 961 962 assert((LiMI->getOpcode() == PPC::LI || 963 LiMI->getOpcode() == PPC::LI8) && 964 "Invalid Opcode!"); 965 auto LiImm = LiMI->getOperand(1).getImm(); // save the imm of LI 966 LiMI->RemoveOperand(1); // remove the imm of LI 967 LiMI->setDesc(TII->get(LiMI->getOpcode() == PPC::LI ? PPC::ADDI 968 : PPC::ADDI8)); 969 MachineInstrBuilder(*LiMI->getParent()->getParent(), *LiMI) 970 .addReg(DominatorReg) 971 .addImm(LiImm); // restore the imm of LI 972 LLVM_DEBUG(LiMI->dump()); 973 } 974 975 // Replace ADD with COPY 976 LLVM_DEBUG(dbgs() << "Optimizing ADD to COPY: "); 977 LLVM_DEBUG(MI.dump()); 978 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 979 MI.getOperand(0).getReg()) 980 .add(Op1); 981 ToErase = &MI; 982 Simplified = true; 983 NumOptADDLIs++; 984 break; 985 } 986 case PPC::RLDICR: { 987 Simplified |= emitRLDICWhenLoweringJumpTables(MI) || 988 combineSEXTAndSHL(MI, ToErase); 989 break; 990 } 991 case PPC::RLWINM: 992 case PPC::RLWINM_rec: 993 case PPC::RLWINM8: 994 case PPC::RLWINM8_rec: { 995 Simplified = TII->combineRLWINM(MI, &ToErase); 996 if (Simplified) 997 ++NumRotatesCollapsed; 998 break; 999 } 1000 } 1001 } 1002 1003 // If the last instruction was marked for elimination, 1004 // remove it now. 1005 if (ToErase) { 1006 ToErase->eraseFromParent(); 1007 ToErase = nullptr; 1008 } 1009 } 1010 1011 // Eliminate all the TOC save instructions which are redundant. 1012 Simplified |= eliminateRedundantTOCSaves(TOCSaves); 1013 PPCFunctionInfo *FI = MF->getInfo<PPCFunctionInfo>(); 1014 if (FI->mustSaveTOC()) 1015 NumTOCSavesInPrologue++; 1016 1017 // We try to eliminate redundant compare instruction. 1018 Simplified |= eliminateRedundantCompare(); 1019 1020 return Simplified; 1021 } 1022 1023 // helper functions for eliminateRedundantCompare 1024 static bool isEqOrNe(MachineInstr *BI) { 1025 PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); 1026 unsigned PredCond = PPC::getPredicateCondition(Pred); 1027 return (PredCond == PPC::PRED_EQ || PredCond == PPC::PRED_NE); 1028 } 1029 1030 static bool isSupportedCmpOp(unsigned opCode) { 1031 return (opCode == PPC::CMPLD || opCode == PPC::CMPD || 1032 opCode == PPC::CMPLW || opCode == PPC::CMPW || 1033 opCode == PPC::CMPLDI || opCode == PPC::CMPDI || 1034 opCode == PPC::CMPLWI || opCode == PPC::CMPWI); 1035 } 1036 1037 static bool is64bitCmpOp(unsigned opCode) { 1038 return (opCode == PPC::CMPLD || opCode == PPC::CMPD || 1039 opCode == PPC::CMPLDI || opCode == PPC::CMPDI); 1040 } 1041 1042 static bool isSignedCmpOp(unsigned opCode) { 1043 return (opCode == PPC::CMPD || opCode == PPC::CMPW || 1044 opCode == PPC::CMPDI || opCode == PPC::CMPWI); 1045 } 1046 1047 static unsigned getSignedCmpOpCode(unsigned opCode) { 1048 if (opCode == PPC::CMPLD) return PPC::CMPD; 1049 if (opCode == PPC::CMPLW) return PPC::CMPW; 1050 if (opCode == PPC::CMPLDI) return PPC::CMPDI; 1051 if (opCode == PPC::CMPLWI) return PPC::CMPWI; 1052 return opCode; 1053 } 1054 1055 // We can decrement immediate x in (GE x) by changing it to (GT x-1) or 1056 // (LT x) to (LE x-1) 1057 static unsigned getPredicateToDecImm(MachineInstr *BI, MachineInstr *CMPI) { 1058 uint64_t Imm = CMPI->getOperand(2).getImm(); 1059 bool SignedCmp = isSignedCmpOp(CMPI->getOpcode()); 1060 if ((!SignedCmp && Imm == 0) || (SignedCmp && Imm == 0x8000)) 1061 return 0; 1062 1063 PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); 1064 unsigned PredCond = PPC::getPredicateCondition(Pred); 1065 unsigned PredHint = PPC::getPredicateHint(Pred); 1066 if (PredCond == PPC::PRED_GE) 1067 return PPC::getPredicate(PPC::PRED_GT, PredHint); 1068 if (PredCond == PPC::PRED_LT) 1069 return PPC::getPredicate(PPC::PRED_LE, PredHint); 1070 1071 return 0; 1072 } 1073 1074 // We can increment immediate x in (GT x) by changing it to (GE x+1) or 1075 // (LE x) to (LT x+1) 1076 static unsigned getPredicateToIncImm(MachineInstr *BI, MachineInstr *CMPI) { 1077 uint64_t Imm = CMPI->getOperand(2).getImm(); 1078 bool SignedCmp = isSignedCmpOp(CMPI->getOpcode()); 1079 if ((!SignedCmp && Imm == 0xFFFF) || (SignedCmp && Imm == 0x7FFF)) 1080 return 0; 1081 1082 PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); 1083 unsigned PredCond = PPC::getPredicateCondition(Pred); 1084 unsigned PredHint = PPC::getPredicateHint(Pred); 1085 if (PredCond == PPC::PRED_GT) 1086 return PPC::getPredicate(PPC::PRED_GE, PredHint); 1087 if (PredCond == PPC::PRED_LE) 1088 return PPC::getPredicate(PPC::PRED_LT, PredHint); 1089 1090 return 0; 1091 } 1092 1093 // This takes a Phi node and returns a register value for the specified BB. 1094 static unsigned getIncomingRegForBlock(MachineInstr *Phi, 1095 MachineBasicBlock *MBB) { 1096 for (unsigned I = 2, E = Phi->getNumOperands() + 1; I != E; I += 2) { 1097 MachineOperand &MO = Phi->getOperand(I); 1098 if (MO.getMBB() == MBB) 1099 return Phi->getOperand(I-1).getReg(); 1100 } 1101 llvm_unreachable("invalid src basic block for this Phi node\n"); 1102 return 0; 1103 } 1104 1105 // This function tracks the source of the register through register copy. 1106 // If BB1 and BB2 are non-NULL, we also track PHI instruction in BB2 1107 // assuming that the control comes from BB1 into BB2. 1108 static unsigned getSrcVReg(unsigned Reg, MachineBasicBlock *BB1, 1109 MachineBasicBlock *BB2, MachineRegisterInfo *MRI) { 1110 unsigned SrcReg = Reg; 1111 while (1) { 1112 unsigned NextReg = SrcReg; 1113 MachineInstr *Inst = MRI->getVRegDef(SrcReg); 1114 if (BB1 && Inst->getOpcode() == PPC::PHI && Inst->getParent() == BB2) { 1115 NextReg = getIncomingRegForBlock(Inst, BB1); 1116 // We track through PHI only once to avoid infinite loop. 1117 BB1 = nullptr; 1118 } 1119 else if (Inst->isFullCopy()) 1120 NextReg = Inst->getOperand(1).getReg(); 1121 if (NextReg == SrcReg || !Register::isVirtualRegister(NextReg)) 1122 break; 1123 SrcReg = NextReg; 1124 } 1125 return SrcReg; 1126 } 1127 1128 static bool eligibleForCompareElimination(MachineBasicBlock &MBB, 1129 MachineBasicBlock *&PredMBB, 1130 MachineBasicBlock *&MBBtoMoveCmp, 1131 MachineRegisterInfo *MRI) { 1132 1133 auto isEligibleBB = [&](MachineBasicBlock &BB) { 1134 auto BII = BB.getFirstInstrTerminator(); 1135 // We optimize BBs ending with a conditional branch. 1136 // We check only for BCC here, not BCCLR, because BCCLR 1137 // will be formed only later in the pipeline. 1138 if (BB.succ_size() == 2 && 1139 BII != BB.instr_end() && 1140 (*BII).getOpcode() == PPC::BCC && 1141 (*BII).getOperand(1).isReg()) { 1142 // We optimize only if the condition code is used only by one BCC. 1143 Register CndReg = (*BII).getOperand(1).getReg(); 1144 if (!Register::isVirtualRegister(CndReg) || !MRI->hasOneNonDBGUse(CndReg)) 1145 return false; 1146 1147 MachineInstr *CMPI = MRI->getVRegDef(CndReg); 1148 // We assume compare and branch are in the same BB for ease of analysis. 1149 if (CMPI->getParent() != &BB) 1150 return false; 1151 1152 // We skip this BB if a physical register is used in comparison. 1153 for (MachineOperand &MO : CMPI->operands()) 1154 if (MO.isReg() && !Register::isVirtualRegister(MO.getReg())) 1155 return false; 1156 1157 return true; 1158 } 1159 return false; 1160 }; 1161 1162 // If this BB has more than one successor, we can create a new BB and 1163 // move the compare instruction in the new BB. 1164 // So far, we do not move compare instruction to a BB having multiple 1165 // successors to avoid potentially increasing code size. 1166 auto isEligibleForMoveCmp = [](MachineBasicBlock &BB) { 1167 return BB.succ_size() == 1; 1168 }; 1169 1170 if (!isEligibleBB(MBB)) 1171 return false; 1172 1173 unsigned NumPredBBs = MBB.pred_size(); 1174 if (NumPredBBs == 1) { 1175 MachineBasicBlock *TmpMBB = *MBB.pred_begin(); 1176 if (isEligibleBB(*TmpMBB)) { 1177 PredMBB = TmpMBB; 1178 MBBtoMoveCmp = nullptr; 1179 return true; 1180 } 1181 } 1182 else if (NumPredBBs == 2) { 1183 // We check for partially redundant case. 1184 // So far, we support cases with only two predecessors 1185 // to avoid increasing the number of instructions. 1186 MachineBasicBlock::pred_iterator PI = MBB.pred_begin(); 1187 MachineBasicBlock *Pred1MBB = *PI; 1188 MachineBasicBlock *Pred2MBB = *(PI+1); 1189 1190 if (isEligibleBB(*Pred1MBB) && isEligibleForMoveCmp(*Pred2MBB)) { 1191 // We assume Pred1MBB is the BB containing the compare to be merged and 1192 // Pred2MBB is the BB to which we will append a compare instruction. 1193 // Hence we can proceed as is. 1194 } 1195 else if (isEligibleBB(*Pred2MBB) && isEligibleForMoveCmp(*Pred1MBB)) { 1196 // We need to swap Pred1MBB and Pred2MBB to canonicalize. 1197 std::swap(Pred1MBB, Pred2MBB); 1198 } 1199 else return false; 1200 1201 // Here, Pred2MBB is the BB to which we need to append a compare inst. 1202 // We cannot move the compare instruction if operands are not available 1203 // in Pred2MBB (i.e. defined in MBB by an instruction other than PHI). 1204 MachineInstr *BI = &*MBB.getFirstInstrTerminator(); 1205 MachineInstr *CMPI = MRI->getVRegDef(BI->getOperand(1).getReg()); 1206 for (int I = 1; I <= 2; I++) 1207 if (CMPI->getOperand(I).isReg()) { 1208 MachineInstr *Inst = MRI->getVRegDef(CMPI->getOperand(I).getReg()); 1209 if (Inst->getParent() == &MBB && Inst->getOpcode() != PPC::PHI) 1210 return false; 1211 } 1212 1213 PredMBB = Pred1MBB; 1214 MBBtoMoveCmp = Pred2MBB; 1215 return true; 1216 } 1217 1218 return false; 1219 } 1220 1221 // This function will iterate over the input map containing a pair of TOC save 1222 // instruction and a flag. The flag will be set to false if the TOC save is 1223 // proven redundant. This function will erase from the basic block all the TOC 1224 // saves marked as redundant. 1225 bool PPCMIPeephole::eliminateRedundantTOCSaves( 1226 std::map<MachineInstr *, bool> &TOCSaves) { 1227 bool Simplified = false; 1228 int NumKept = 0; 1229 for (auto TOCSave : TOCSaves) { 1230 if (!TOCSave.second) { 1231 TOCSave.first->eraseFromParent(); 1232 RemoveTOCSave++; 1233 Simplified = true; 1234 } else { 1235 NumKept++; 1236 } 1237 } 1238 1239 if (NumKept > 1) 1240 MultiTOCSaves++; 1241 1242 return Simplified; 1243 } 1244 1245 // If multiple conditional branches are executed based on the (essentially) 1246 // same comparison, we merge compare instructions into one and make multiple 1247 // conditional branches on this comparison. 1248 // For example, 1249 // if (a == 0) { ... } 1250 // else if (a < 0) { ... } 1251 // can be executed by one compare and two conditional branches instead of 1252 // two pairs of a compare and a conditional branch. 1253 // 1254 // This method merges two compare instructions in two MBBs and modifies the 1255 // compare and conditional branch instructions if needed. 1256 // For the above example, the input for this pass looks like: 1257 // cmplwi r3, 0 1258 // beq 0, .LBB0_3 1259 // cmpwi r3, -1 1260 // bgt 0, .LBB0_4 1261 // So, before merging two compares, we need to modify these instructions as 1262 // cmpwi r3, 0 ; cmplwi and cmpwi yield same result for beq 1263 // beq 0, .LBB0_3 1264 // cmpwi r3, 0 ; greather than -1 means greater or equal to 0 1265 // bge 0, .LBB0_4 1266 1267 bool PPCMIPeephole::eliminateRedundantCompare(void) { 1268 bool Simplified = false; 1269 1270 for (MachineBasicBlock &MBB2 : *MF) { 1271 MachineBasicBlock *MBB1 = nullptr, *MBBtoMoveCmp = nullptr; 1272 1273 // For fully redundant case, we select two basic blocks MBB1 and MBB2 1274 // as an optimization target if 1275 // - both MBBs end with a conditional branch, 1276 // - MBB1 is the only predecessor of MBB2, and 1277 // - compare does not take a physical register as a operand in both MBBs. 1278 // In this case, eligibleForCompareElimination sets MBBtoMoveCmp nullptr. 1279 // 1280 // As partially redundant case, we additionally handle if MBB2 has one 1281 // additional predecessor, which has only one successor (MBB2). 1282 // In this case, we move the compare instruction originally in MBB2 into 1283 // MBBtoMoveCmp. This partially redundant case is typically appear by 1284 // compiling a while loop; here, MBBtoMoveCmp is the loop preheader. 1285 // 1286 // Overview of CFG of related basic blocks 1287 // Fully redundant case Partially redundant case 1288 // -------- ---------------- -------- 1289 // | MBB1 | (w/ 2 succ) | MBBtoMoveCmp | | MBB1 | (w/ 2 succ) 1290 // -------- ---------------- -------- 1291 // | \ (w/ 1 succ) \ | \ 1292 // | \ \ | \ 1293 // | \ | 1294 // -------- -------- 1295 // | MBB2 | (w/ 1 pred | MBB2 | (w/ 2 pred 1296 // -------- and 2 succ) -------- and 2 succ) 1297 // | \ | \ 1298 // | \ | \ 1299 // 1300 if (!eligibleForCompareElimination(MBB2, MBB1, MBBtoMoveCmp, MRI)) 1301 continue; 1302 1303 MachineInstr *BI1 = &*MBB1->getFirstInstrTerminator(); 1304 MachineInstr *CMPI1 = MRI->getVRegDef(BI1->getOperand(1).getReg()); 1305 1306 MachineInstr *BI2 = &*MBB2.getFirstInstrTerminator(); 1307 MachineInstr *CMPI2 = MRI->getVRegDef(BI2->getOperand(1).getReg()); 1308 bool IsPartiallyRedundant = (MBBtoMoveCmp != nullptr); 1309 1310 // We cannot optimize an unsupported compare opcode or 1311 // a mix of 32-bit and 64-bit comaprisons 1312 if (!isSupportedCmpOp(CMPI1->getOpcode()) || 1313 !isSupportedCmpOp(CMPI2->getOpcode()) || 1314 is64bitCmpOp(CMPI1->getOpcode()) != is64bitCmpOp(CMPI2->getOpcode())) 1315 continue; 1316 1317 unsigned NewOpCode = 0; 1318 unsigned NewPredicate1 = 0, NewPredicate2 = 0; 1319 int16_t Imm1 = 0, NewImm1 = 0, Imm2 = 0, NewImm2 = 0; 1320 bool SwapOperands = false; 1321 1322 if (CMPI1->getOpcode() != CMPI2->getOpcode()) { 1323 // Typically, unsigned comparison is used for equality check, but 1324 // we replace it with a signed comparison if the comparison 1325 // to be merged is a signed comparison. 1326 // In other cases of opcode mismatch, we cannot optimize this. 1327 1328 // We cannot change opcode when comparing against an immediate 1329 // if the most significant bit of the immediate is one 1330 // due to the difference in sign extension. 1331 auto CmpAgainstImmWithSignBit = [](MachineInstr *I) { 1332 if (!I->getOperand(2).isImm()) 1333 return false; 1334 int16_t Imm = (int16_t)I->getOperand(2).getImm(); 1335 return Imm < 0; 1336 }; 1337 1338 if (isEqOrNe(BI2) && !CmpAgainstImmWithSignBit(CMPI2) && 1339 CMPI1->getOpcode() == getSignedCmpOpCode(CMPI2->getOpcode())) 1340 NewOpCode = CMPI1->getOpcode(); 1341 else if (isEqOrNe(BI1) && !CmpAgainstImmWithSignBit(CMPI1) && 1342 getSignedCmpOpCode(CMPI1->getOpcode()) == CMPI2->getOpcode()) 1343 NewOpCode = CMPI2->getOpcode(); 1344 else continue; 1345 } 1346 1347 if (CMPI1->getOperand(2).isReg() && CMPI2->getOperand(2).isReg()) { 1348 // In case of comparisons between two registers, these two registers 1349 // must be same to merge two comparisons. 1350 unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(), 1351 nullptr, nullptr, MRI); 1352 unsigned Cmp1Operand2 = getSrcVReg(CMPI1->getOperand(2).getReg(), 1353 nullptr, nullptr, MRI); 1354 unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(), 1355 MBB1, &MBB2, MRI); 1356 unsigned Cmp2Operand2 = getSrcVReg(CMPI2->getOperand(2).getReg(), 1357 MBB1, &MBB2, MRI); 1358 1359 if (Cmp1Operand1 == Cmp2Operand1 && Cmp1Operand2 == Cmp2Operand2) { 1360 // Same pair of registers in the same order; ready to merge as is. 1361 } 1362 else if (Cmp1Operand1 == Cmp2Operand2 && Cmp1Operand2 == Cmp2Operand1) { 1363 // Same pair of registers in different order. 1364 // We reverse the predicate to merge compare instructions. 1365 PPC::Predicate Pred = (PPC::Predicate)BI2->getOperand(0).getImm(); 1366 NewPredicate2 = (unsigned)PPC::getSwappedPredicate(Pred); 1367 // In case of partial redundancy, we need to swap operands 1368 // in another compare instruction. 1369 SwapOperands = true; 1370 } 1371 else continue; 1372 } 1373 else if (CMPI1->getOperand(2).isImm() && CMPI2->getOperand(2).isImm()) { 1374 // In case of comparisons between a register and an immediate, 1375 // the operand register must be same for two compare instructions. 1376 unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(), 1377 nullptr, nullptr, MRI); 1378 unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(), 1379 MBB1, &MBB2, MRI); 1380 if (Cmp1Operand1 != Cmp2Operand1) 1381 continue; 1382 1383 NewImm1 = Imm1 = (int16_t)CMPI1->getOperand(2).getImm(); 1384 NewImm2 = Imm2 = (int16_t)CMPI2->getOperand(2).getImm(); 1385 1386 // If immediate are not same, we try to adjust by changing predicate; 1387 // e.g. GT imm means GE (imm+1). 1388 if (Imm1 != Imm2 && (!isEqOrNe(BI2) || !isEqOrNe(BI1))) { 1389 int Diff = Imm1 - Imm2; 1390 if (Diff < -2 || Diff > 2) 1391 continue; 1392 1393 unsigned PredToInc1 = getPredicateToIncImm(BI1, CMPI1); 1394 unsigned PredToDec1 = getPredicateToDecImm(BI1, CMPI1); 1395 unsigned PredToInc2 = getPredicateToIncImm(BI2, CMPI2); 1396 unsigned PredToDec2 = getPredicateToDecImm(BI2, CMPI2); 1397 if (Diff == 2) { 1398 if (PredToInc2 && PredToDec1) { 1399 NewPredicate2 = PredToInc2; 1400 NewPredicate1 = PredToDec1; 1401 NewImm2++; 1402 NewImm1--; 1403 } 1404 } 1405 else if (Diff == 1) { 1406 if (PredToInc2) { 1407 NewImm2++; 1408 NewPredicate2 = PredToInc2; 1409 } 1410 else if (PredToDec1) { 1411 NewImm1--; 1412 NewPredicate1 = PredToDec1; 1413 } 1414 } 1415 else if (Diff == -1) { 1416 if (PredToDec2) { 1417 NewImm2--; 1418 NewPredicate2 = PredToDec2; 1419 } 1420 else if (PredToInc1) { 1421 NewImm1++; 1422 NewPredicate1 = PredToInc1; 1423 } 1424 } 1425 else if (Diff == -2) { 1426 if (PredToDec2 && PredToInc1) { 1427 NewPredicate2 = PredToDec2; 1428 NewPredicate1 = PredToInc1; 1429 NewImm2--; 1430 NewImm1++; 1431 } 1432 } 1433 } 1434 1435 // We cannot merge two compares if the immediates are not same. 1436 if (NewImm2 != NewImm1) 1437 continue; 1438 } 1439 1440 LLVM_DEBUG(dbgs() << "Optimize two pairs of compare and branch:\n"); 1441 LLVM_DEBUG(CMPI1->dump()); 1442 LLVM_DEBUG(BI1->dump()); 1443 LLVM_DEBUG(CMPI2->dump()); 1444 LLVM_DEBUG(BI2->dump()); 1445 1446 // We adjust opcode, predicates and immediate as we determined above. 1447 if (NewOpCode != 0 && NewOpCode != CMPI1->getOpcode()) { 1448 CMPI1->setDesc(TII->get(NewOpCode)); 1449 } 1450 if (NewPredicate1) { 1451 BI1->getOperand(0).setImm(NewPredicate1); 1452 } 1453 if (NewPredicate2) { 1454 BI2->getOperand(0).setImm(NewPredicate2); 1455 } 1456 if (NewImm1 != Imm1) { 1457 CMPI1->getOperand(2).setImm(NewImm1); 1458 } 1459 1460 if (IsPartiallyRedundant) { 1461 // We touch up the compare instruction in MBB2 and move it to 1462 // a previous BB to handle partially redundant case. 1463 if (SwapOperands) { 1464 Register Op1 = CMPI2->getOperand(1).getReg(); 1465 Register Op2 = CMPI2->getOperand(2).getReg(); 1466 CMPI2->getOperand(1).setReg(Op2); 1467 CMPI2->getOperand(2).setReg(Op1); 1468 } 1469 if (NewImm2 != Imm2) 1470 CMPI2->getOperand(2).setImm(NewImm2); 1471 1472 for (int I = 1; I <= 2; I++) { 1473 if (CMPI2->getOperand(I).isReg()) { 1474 MachineInstr *Inst = MRI->getVRegDef(CMPI2->getOperand(I).getReg()); 1475 if (Inst->getParent() != &MBB2) 1476 continue; 1477 1478 assert(Inst->getOpcode() == PPC::PHI && 1479 "We cannot support if an operand comes from this BB."); 1480 unsigned SrcReg = getIncomingRegForBlock(Inst, MBBtoMoveCmp); 1481 CMPI2->getOperand(I).setReg(SrcReg); 1482 } 1483 } 1484 auto I = MachineBasicBlock::iterator(MBBtoMoveCmp->getFirstTerminator()); 1485 MBBtoMoveCmp->splice(I, &MBB2, MachineBasicBlock::iterator(CMPI2)); 1486 1487 DebugLoc DL = CMPI2->getDebugLoc(); 1488 Register NewVReg = MRI->createVirtualRegister(&PPC::CRRCRegClass); 1489 BuildMI(MBB2, MBB2.begin(), DL, 1490 TII->get(PPC::PHI), NewVReg) 1491 .addReg(BI1->getOperand(1).getReg()).addMBB(MBB1) 1492 .addReg(BI2->getOperand(1).getReg()).addMBB(MBBtoMoveCmp); 1493 BI2->getOperand(1).setReg(NewVReg); 1494 } 1495 else { 1496 // We finally eliminate compare instruction in MBB2. 1497 BI2->getOperand(1).setReg(BI1->getOperand(1).getReg()); 1498 CMPI2->eraseFromParent(); 1499 } 1500 BI2->getOperand(1).setIsKill(true); 1501 BI1->getOperand(1).setIsKill(false); 1502 1503 LLVM_DEBUG(dbgs() << "into a compare and two branches:\n"); 1504 LLVM_DEBUG(CMPI1->dump()); 1505 LLVM_DEBUG(BI1->dump()); 1506 LLVM_DEBUG(BI2->dump()); 1507 if (IsPartiallyRedundant) { 1508 LLVM_DEBUG(dbgs() << "The following compare is moved into " 1509 << printMBBReference(*MBBtoMoveCmp) 1510 << " to handle partial redundancy.\n"); 1511 LLVM_DEBUG(CMPI2->dump()); 1512 } 1513 1514 Simplified = true; 1515 } 1516 1517 return Simplified; 1518 } 1519 1520 // We miss the opportunity to emit an RLDIC when lowering jump tables 1521 // since ISEL sees only a single basic block. When selecting, the clear 1522 // and shift left will be in different blocks. 1523 bool PPCMIPeephole::emitRLDICWhenLoweringJumpTables(MachineInstr &MI) { 1524 if (MI.getOpcode() != PPC::RLDICR) 1525 return false; 1526 1527 Register SrcReg = MI.getOperand(1).getReg(); 1528 if (!Register::isVirtualRegister(SrcReg)) 1529 return false; 1530 1531 MachineInstr *SrcMI = MRI->getVRegDef(SrcReg); 1532 if (SrcMI->getOpcode() != PPC::RLDICL) 1533 return false; 1534 1535 MachineOperand MOpSHSrc = SrcMI->getOperand(2); 1536 MachineOperand MOpMBSrc = SrcMI->getOperand(3); 1537 MachineOperand MOpSHMI = MI.getOperand(2); 1538 MachineOperand MOpMEMI = MI.getOperand(3); 1539 if (!(MOpSHSrc.isImm() && MOpMBSrc.isImm() && MOpSHMI.isImm() && 1540 MOpMEMI.isImm())) 1541 return false; 1542 1543 uint64_t SHSrc = MOpSHSrc.getImm(); 1544 uint64_t MBSrc = MOpMBSrc.getImm(); 1545 uint64_t SHMI = MOpSHMI.getImm(); 1546 uint64_t MEMI = MOpMEMI.getImm(); 1547 uint64_t NewSH = SHSrc + SHMI; 1548 uint64_t NewMB = MBSrc - SHMI; 1549 if (NewMB > 63 || NewSH > 63) 1550 return false; 1551 1552 // The bits cleared with RLDICL are [0, MBSrc). 1553 // The bits cleared with RLDICR are (MEMI, 63]. 1554 // After the sequence, the bits cleared are: 1555 // [0, MBSrc-SHMI) and (MEMI, 63). 1556 // 1557 // The bits cleared with RLDIC are [0, NewMB) and (63-NewSH, 63]. 1558 if ((63 - NewSH) != MEMI) 1559 return false; 1560 1561 LLVM_DEBUG(dbgs() << "Converting pair: "); 1562 LLVM_DEBUG(SrcMI->dump()); 1563 LLVM_DEBUG(MI.dump()); 1564 1565 MI.setDesc(TII->get(PPC::RLDIC)); 1566 MI.getOperand(1).setReg(SrcMI->getOperand(1).getReg()); 1567 MI.getOperand(2).setImm(NewSH); 1568 MI.getOperand(3).setImm(NewMB); 1569 MI.getOperand(1).setIsKill(SrcMI->getOperand(1).isKill()); 1570 SrcMI->getOperand(1).setIsKill(false); 1571 1572 LLVM_DEBUG(dbgs() << "To: "); 1573 LLVM_DEBUG(MI.dump()); 1574 NumRotatesCollapsed++; 1575 // If SrcReg has no non-debug use it's safe to delete its def SrcMI. 1576 if (MRI->use_nodbg_empty(SrcReg)) { 1577 assert(!SrcMI->hasImplicitDef() && 1578 "Not expecting an implicit def with this instr."); 1579 SrcMI->eraseFromParent(); 1580 } 1581 return true; 1582 } 1583 1584 // For case in LLVM IR 1585 // entry: 1586 // %iconv = sext i32 %index to i64 1587 // br i1 undef label %true, label %false 1588 // true: 1589 // %ptr = getelementptr inbounds i32, i32* null, i64 %iconv 1590 // ... 1591 // PPCISelLowering::combineSHL fails to combine, because sext and shl are in 1592 // different BBs when conducting instruction selection. We can do a peephole 1593 // optimization to combine these two instructions into extswsli after 1594 // instruction selection. 1595 bool PPCMIPeephole::combineSEXTAndSHL(MachineInstr &MI, 1596 MachineInstr *&ToErase) { 1597 if (MI.getOpcode() != PPC::RLDICR) 1598 return false; 1599 1600 if (!MF->getSubtarget<PPCSubtarget>().isISA3_0()) 1601 return false; 1602 1603 assert(MI.getNumOperands() == 4 && "RLDICR should have 4 operands"); 1604 1605 MachineOperand MOpSHMI = MI.getOperand(2); 1606 MachineOperand MOpMEMI = MI.getOperand(3); 1607 if (!(MOpSHMI.isImm() && MOpMEMI.isImm())) 1608 return false; 1609 1610 uint64_t SHMI = MOpSHMI.getImm(); 1611 uint64_t MEMI = MOpMEMI.getImm(); 1612 if (SHMI + MEMI != 63) 1613 return false; 1614 1615 Register SrcReg = MI.getOperand(1).getReg(); 1616 if (!Register::isVirtualRegister(SrcReg)) 1617 return false; 1618 1619 MachineInstr *SrcMI = MRI->getVRegDef(SrcReg); 1620 if (SrcMI->getOpcode() != PPC::EXTSW && 1621 SrcMI->getOpcode() != PPC::EXTSW_32_64) 1622 return false; 1623 1624 // If the register defined by extsw has more than one use, combination is not 1625 // needed. 1626 if (!MRI->hasOneNonDBGUse(SrcReg)) 1627 return false; 1628 1629 assert(SrcMI->getNumOperands() == 2 && "EXTSW should have 2 operands"); 1630 assert(SrcMI->getOperand(1).isReg() && 1631 "EXTSW's second operand should be a register"); 1632 if (!Register::isVirtualRegister(SrcMI->getOperand(1).getReg())) 1633 return false; 1634 1635 LLVM_DEBUG(dbgs() << "Combining pair: "); 1636 LLVM_DEBUG(SrcMI->dump()); 1637 LLVM_DEBUG(MI.dump()); 1638 1639 MachineInstr *NewInstr = 1640 BuildMI(*MI.getParent(), &MI, MI.getDebugLoc(), 1641 SrcMI->getOpcode() == PPC::EXTSW ? TII->get(PPC::EXTSWSLI) 1642 : TII->get(PPC::EXTSWSLI_32_64), 1643 MI.getOperand(0).getReg()) 1644 .add(SrcMI->getOperand(1)) 1645 .add(MOpSHMI); 1646 (void)NewInstr; 1647 1648 LLVM_DEBUG(dbgs() << "TO: "); 1649 LLVM_DEBUG(NewInstr->dump()); 1650 ++NumEXTSWAndSLDICombined; 1651 ToErase = &MI; 1652 // SrcMI, which is extsw, is of no use now, erase it. 1653 SrcMI->eraseFromParent(); 1654 return true; 1655 } 1656 1657 } // end default namespace 1658 1659 INITIALIZE_PASS_BEGIN(PPCMIPeephole, DEBUG_TYPE, 1660 "PowerPC MI Peephole Optimization", false, false) 1661 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) 1662 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 1663 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree) 1664 INITIALIZE_PASS_END(PPCMIPeephole, DEBUG_TYPE, 1665 "PowerPC MI Peephole Optimization", false, false) 1666 1667 char PPCMIPeephole::ID = 0; 1668 FunctionPass* 1669 llvm::createPPCMIPeepholePass() { return new PPCMIPeephole(); } 1670 1671