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/MachineFrameInfo.h" 32 #include "llvm/CodeGen/MachineFunctionPass.h" 33 #include "llvm/CodeGen/MachineInstrBuilder.h" 34 #include "llvm/CodeGen/MachinePostDominators.h" 35 #include "llvm/CodeGen/MachineRegisterInfo.h" 36 #include "llvm/InitializePasses.h" 37 #include "llvm/Support/Debug.h" 38 39 using namespace llvm; 40 41 #define DEBUG_TYPE "ppc-mi-peepholes" 42 43 STATISTIC(RemoveTOCSave, "Number of TOC saves removed"); 44 STATISTIC(MultiTOCSaves, 45 "Number of functions with multiple TOC saves that must be kept"); 46 STATISTIC(NumTOCSavesInPrologue, "Number of TOC saves placed in the prologue"); 47 STATISTIC(NumEliminatedSExt, "Number of eliminated sign-extensions"); 48 STATISTIC(NumEliminatedZExt, "Number of eliminated zero-extensions"); 49 STATISTIC(NumOptADDLIs, "Number of optimized ADD instruction fed by LI"); 50 STATISTIC(NumConvertedToImmediateForm, 51 "Number of instructions converted to their immediate form"); 52 STATISTIC(NumFunctionsEnteredInMIPeephole, 53 "Number of functions entered in PPC MI Peepholes"); 54 STATISTIC(NumFixedPointIterations, 55 "Number of fixed-point iterations converting reg-reg instructions " 56 "to reg-imm ones"); 57 STATISTIC(NumRotatesCollapsed, 58 "Number of pairs of rotate left, clear left/right collapsed"); 59 STATISTIC(NumEXTSWAndSLDICombined, 60 "Number of pairs of EXTSW and SLDI combined as EXTSWSLI"); 61 STATISTIC(NumLoadImmZeroFoldedAndRemoved, 62 "Number of LI(8) reg, 0 that are folded to r0 and removed"); 63 64 static cl::opt<bool> 65 FixedPointRegToImm("ppc-reg-to-imm-fixed-point", cl::Hidden, cl::init(true), 66 cl::desc("Iterate to a fixed point when attempting to " 67 "convert reg-reg instructions to reg-imm")); 68 69 static cl::opt<bool> 70 ConvertRegReg("ppc-convert-rr-to-ri", cl::Hidden, cl::init(true), 71 cl::desc("Convert eligible reg+reg instructions to reg+imm")); 72 73 static cl::opt<bool> 74 EnableSExtElimination("ppc-eliminate-signext", 75 cl::desc("enable elimination of sign-extensions"), 76 cl::init(true), cl::Hidden); 77 78 static cl::opt<bool> 79 EnableZExtElimination("ppc-eliminate-zeroext", 80 cl::desc("enable elimination of zero-extensions"), 81 cl::init(true), cl::Hidden); 82 83 static cl::opt<bool> 84 EnableTrapOptimization("ppc-opt-conditional-trap", 85 cl::desc("enable optimization of conditional traps"), 86 cl::init(false), cl::Hidden); 87 88 namespace { 89 90 struct PPCMIPeephole : public MachineFunctionPass { 91 92 static char ID; 93 const PPCInstrInfo *TII; 94 MachineFunction *MF; 95 MachineRegisterInfo *MRI; 96 97 PPCMIPeephole() : MachineFunctionPass(ID) { 98 initializePPCMIPeepholePass(*PassRegistry::getPassRegistry()); 99 } 100 101 private: 102 MachineDominatorTree *MDT; 103 MachinePostDominatorTree *MPDT; 104 MachineBlockFrequencyInfo *MBFI; 105 uint64_t EntryFreq; 106 107 // Initialize class variables. 108 void initialize(MachineFunction &MFParm); 109 110 // Perform peepholes. 111 bool simplifyCode(); 112 113 // Perform peepholes. 114 bool eliminateRedundantCompare(); 115 bool eliminateRedundantTOCSaves(std::map<MachineInstr *, bool> &TOCSaves); 116 bool combineSEXTAndSHL(MachineInstr &MI, MachineInstr *&ToErase); 117 bool emitRLDICWhenLoweringJumpTables(MachineInstr &MI); 118 void UpdateTOCSaves(std::map<MachineInstr *, bool> &TOCSaves, 119 MachineInstr *MI); 120 121 public: 122 123 void getAnalysisUsage(AnalysisUsage &AU) const override { 124 AU.addRequired<MachineDominatorTree>(); 125 AU.addRequired<MachinePostDominatorTree>(); 126 AU.addRequired<MachineBlockFrequencyInfo>(); 127 AU.addPreserved<MachineDominatorTree>(); 128 AU.addPreserved<MachinePostDominatorTree>(); 129 AU.addPreserved<MachineBlockFrequencyInfo>(); 130 MachineFunctionPass::getAnalysisUsage(AU); 131 } 132 133 // Main entry point for this pass. 134 bool runOnMachineFunction(MachineFunction &MF) override { 135 initialize(MF); 136 // At this point, TOC pointer should not be used in a function that uses 137 // PC-Relative addressing. 138 assert((MF.getRegInfo().use_empty(PPC::X2) || 139 !MF.getSubtarget<PPCSubtarget>().isUsingPCRelativeCalls()) && 140 "TOC pointer used in a function using PC-Relative addressing!"); 141 if (skipFunction(MF.getFunction())) 142 return false; 143 return simplifyCode(); 144 } 145 }; 146 147 // Initialize class variables. 148 void PPCMIPeephole::initialize(MachineFunction &MFParm) { 149 MF = &MFParm; 150 MRI = &MF->getRegInfo(); 151 MDT = &getAnalysis<MachineDominatorTree>(); 152 MPDT = &getAnalysis<MachinePostDominatorTree>(); 153 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 154 EntryFreq = MBFI->getEntryFreq(); 155 TII = MF->getSubtarget<PPCSubtarget>().getInstrInfo(); 156 LLVM_DEBUG(dbgs() << "*** PowerPC MI peephole pass ***\n\n"); 157 LLVM_DEBUG(MF->dump()); 158 } 159 160 static MachineInstr *getVRegDefOrNull(MachineOperand *Op, 161 MachineRegisterInfo *MRI) { 162 assert(Op && "Invalid Operand!"); 163 if (!Op->isReg()) 164 return nullptr; 165 166 Register Reg = Op->getReg(); 167 if (!Reg.isVirtual()) 168 return nullptr; 169 170 return MRI->getVRegDef(Reg); 171 } 172 173 // This function returns number of known zero bits in output of MI 174 // starting from the most significant bit. 175 static unsigned getKnownLeadingZeroCount(const unsigned Reg, 176 const PPCInstrInfo *TII, 177 const MachineRegisterInfo *MRI) { 178 MachineInstr *MI = MRI->getVRegDef(Reg); 179 unsigned Opcode = MI->getOpcode(); 180 if (Opcode == PPC::RLDICL || Opcode == PPC::RLDICL_rec || 181 Opcode == PPC::RLDCL || Opcode == PPC::RLDCL_rec) 182 return MI->getOperand(3).getImm(); 183 184 if ((Opcode == PPC::RLDIC || Opcode == PPC::RLDIC_rec) && 185 MI->getOperand(3).getImm() <= 63 - MI->getOperand(2).getImm()) 186 return MI->getOperand(3).getImm(); 187 188 if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINM_rec || 189 Opcode == PPC::RLWNM || Opcode == PPC::RLWNM_rec || 190 Opcode == PPC::RLWINM8 || Opcode == PPC::RLWNM8) && 191 MI->getOperand(3).getImm() <= MI->getOperand(4).getImm()) 192 return 32 + MI->getOperand(3).getImm(); 193 194 if (Opcode == PPC::ANDI_rec) { 195 uint16_t Imm = MI->getOperand(2).getImm(); 196 return 48 + llvm::countl_zero(Imm); 197 } 198 199 if (Opcode == PPC::CNTLZW || Opcode == PPC::CNTLZW_rec || 200 Opcode == PPC::CNTTZW || Opcode == PPC::CNTTZW_rec || 201 Opcode == PPC::CNTLZW8 || Opcode == PPC::CNTTZW8) 202 // The result ranges from 0 to 32. 203 return 58; 204 205 if (Opcode == PPC::CNTLZD || Opcode == PPC::CNTLZD_rec || 206 Opcode == PPC::CNTTZD || Opcode == PPC::CNTTZD_rec) 207 // The result ranges from 0 to 64. 208 return 57; 209 210 if (Opcode == PPC::LHZ || Opcode == PPC::LHZX || 211 Opcode == PPC::LHZ8 || Opcode == PPC::LHZX8 || 212 Opcode == PPC::LHZU || Opcode == PPC::LHZUX || 213 Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8) 214 return 48; 215 216 if (Opcode == PPC::LBZ || Opcode == PPC::LBZX || 217 Opcode == PPC::LBZ8 || Opcode == PPC::LBZX8 || 218 Opcode == PPC::LBZU || Opcode == PPC::LBZUX || 219 Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8) 220 return 56; 221 222 if (TII->isZeroExtended(Reg, MRI)) 223 return 32; 224 225 return 0; 226 } 227 228 // This function maintains a map for the pairs <TOC Save Instr, Keep> 229 // Each time a new TOC save is encountered, it checks if any of the existing 230 // ones are dominated by the new one. If so, it marks the existing one as 231 // redundant by setting it's entry in the map as false. It then adds the new 232 // instruction to the map with either true or false depending on if any 233 // existing instructions dominated the new one. 234 void PPCMIPeephole::UpdateTOCSaves( 235 std::map<MachineInstr *, bool> &TOCSaves, MachineInstr *MI) { 236 assert(TII->isTOCSaveMI(*MI) && "Expecting a TOC save instruction here"); 237 // FIXME: Saving TOC in prologue hasn't been implemented well in AIX ABI part, 238 // here only support it under ELFv2. 239 if (MF->getSubtarget<PPCSubtarget>().isELFv2ABI()) { 240 PPCFunctionInfo *FI = MF->getInfo<PPCFunctionInfo>(); 241 242 MachineBasicBlock *Entry = &MF->front(); 243 uint64_t CurrBlockFreq = MBFI->getBlockFreq(MI->getParent()).getFrequency(); 244 245 // If the block in which the TOC save resides is in a block that 246 // post-dominates Entry, or a block that is hotter than entry (keep in mind 247 // that early MachineLICM has already run so the TOC save won't be hoisted) 248 // we can just do the save in the prologue. 249 if (CurrBlockFreq > EntryFreq || MPDT->dominates(MI->getParent(), Entry)) 250 FI->setMustSaveTOC(true); 251 252 // If we are saving the TOC in the prologue, all the TOC saves can be 253 // removed from the code. 254 if (FI->mustSaveTOC()) { 255 for (auto &TOCSave : TOCSaves) 256 TOCSave.second = false; 257 // Add new instruction to map. 258 TOCSaves[MI] = false; 259 return; 260 } 261 } 262 263 bool Keep = true; 264 for (auto &I : TOCSaves) { 265 MachineInstr *CurrInst = I.first; 266 // If new instruction dominates an existing one, mark existing one as 267 // redundant. 268 if (I.second && MDT->dominates(MI, CurrInst)) 269 I.second = false; 270 // Check if the new instruction is redundant. 271 if (MDT->dominates(CurrInst, MI)) { 272 Keep = false; 273 break; 274 } 275 } 276 // Add new instruction to map. 277 TOCSaves[MI] = Keep; 278 } 279 280 // This function returns a list of all PHI nodes in the tree starting from 281 // the RootPHI node. We perform a BFS traversal to get an ordered list of nodes. 282 // The list initially only contains the root PHI. When we visit a PHI node, we 283 // add it to the list. We continue to look for other PHI node operands while 284 // there are nodes to visit in the list. The function returns false if the 285 // optimization cannot be applied on this tree. 286 static bool collectUnprimedAccPHIs(MachineRegisterInfo *MRI, 287 MachineInstr *RootPHI, 288 SmallVectorImpl<MachineInstr *> &PHIs) { 289 PHIs.push_back(RootPHI); 290 unsigned VisitedIndex = 0; 291 while (VisitedIndex < PHIs.size()) { 292 MachineInstr *VisitedPHI = PHIs[VisitedIndex]; 293 for (unsigned PHIOp = 1, NumOps = VisitedPHI->getNumOperands(); 294 PHIOp != NumOps; PHIOp += 2) { 295 Register RegOp = VisitedPHI->getOperand(PHIOp).getReg(); 296 if (!RegOp.isVirtual()) 297 return false; 298 MachineInstr *Instr = MRI->getVRegDef(RegOp); 299 // While collecting the PHI nodes, we check if they can be converted (i.e. 300 // all the operands are either copies, implicit defs or PHI nodes). 301 unsigned Opcode = Instr->getOpcode(); 302 if (Opcode == PPC::COPY) { 303 Register Reg = Instr->getOperand(1).getReg(); 304 if (!Reg.isVirtual() || MRI->getRegClass(Reg) != &PPC::ACCRCRegClass) 305 return false; 306 } else if (Opcode != PPC::IMPLICIT_DEF && Opcode != PPC::PHI) 307 return false; 308 // If we detect a cycle in the PHI nodes, we exit. It would be 309 // possible to change cycles as well, but that would add a lot 310 // of complexity for a case that is unlikely to occur with MMA 311 // code. 312 if (Opcode != PPC::PHI) 313 continue; 314 if (llvm::is_contained(PHIs, Instr)) 315 return false; 316 PHIs.push_back(Instr); 317 } 318 VisitedIndex++; 319 } 320 return true; 321 } 322 323 // This function changes the unprimed accumulator PHI nodes in the PHIs list to 324 // primed accumulator PHI nodes. The list is traversed in reverse order to 325 // change all the PHI operands of a PHI node before changing the node itself. 326 // We keep a map to associate each changed PHI node to its non-changed form. 327 static void convertUnprimedAccPHIs(const PPCInstrInfo *TII, 328 MachineRegisterInfo *MRI, 329 SmallVectorImpl<MachineInstr *> &PHIs, 330 Register Dst) { 331 DenseMap<MachineInstr *, MachineInstr *> ChangedPHIMap; 332 for (MachineInstr *PHI : llvm::reverse(PHIs)) { 333 SmallVector<std::pair<MachineOperand, MachineOperand>, 4> PHIOps; 334 // We check if the current PHI node can be changed by looking at its 335 // operands. If all the operands are either copies from primed 336 // accumulators, implicit definitions or other unprimed accumulator 337 // PHI nodes, we change it. 338 for (unsigned PHIOp = 1, NumOps = PHI->getNumOperands(); PHIOp != NumOps; 339 PHIOp += 2) { 340 Register RegOp = PHI->getOperand(PHIOp).getReg(); 341 MachineInstr *PHIInput = MRI->getVRegDef(RegOp); 342 unsigned Opcode = PHIInput->getOpcode(); 343 assert((Opcode == PPC::COPY || Opcode == PPC::IMPLICIT_DEF || 344 Opcode == PPC::PHI) && 345 "Unexpected instruction"); 346 if (Opcode == PPC::COPY) { 347 assert(MRI->getRegClass(PHIInput->getOperand(1).getReg()) == 348 &PPC::ACCRCRegClass && 349 "Unexpected register class"); 350 PHIOps.push_back({PHIInput->getOperand(1), PHI->getOperand(PHIOp + 1)}); 351 } else if (Opcode == PPC::IMPLICIT_DEF) { 352 Register AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass); 353 BuildMI(*PHIInput->getParent(), PHIInput, PHIInput->getDebugLoc(), 354 TII->get(PPC::IMPLICIT_DEF), AccReg); 355 PHIOps.push_back({MachineOperand::CreateReg(AccReg, false), 356 PHI->getOperand(PHIOp + 1)}); 357 } else if (Opcode == PPC::PHI) { 358 // We found a PHI operand. At this point we know this operand 359 // has already been changed so we get its associated changed form 360 // from the map. 361 assert(ChangedPHIMap.count(PHIInput) == 1 && 362 "This PHI node should have already been changed."); 363 MachineInstr *PrimedAccPHI = ChangedPHIMap.lookup(PHIInput); 364 PHIOps.push_back({MachineOperand::CreateReg( 365 PrimedAccPHI->getOperand(0).getReg(), false), 366 PHI->getOperand(PHIOp + 1)}); 367 } 368 } 369 Register AccReg = Dst; 370 // If the PHI node we are changing is the root node, the register it defines 371 // will be the destination register of the original copy (of the PHI def). 372 // For all other PHI's in the list, we need to create another primed 373 // accumulator virtual register as the PHI will no longer define the 374 // unprimed accumulator. 375 if (PHI != PHIs[0]) 376 AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass); 377 MachineInstrBuilder NewPHI = BuildMI( 378 *PHI->getParent(), PHI, PHI->getDebugLoc(), TII->get(PPC::PHI), AccReg); 379 for (auto RegMBB : PHIOps) 380 NewPHI.add(RegMBB.first).add(RegMBB.second); 381 ChangedPHIMap[PHI] = NewPHI.getInstr(); 382 LLVM_DEBUG(dbgs() << "Converting PHI: "); 383 LLVM_DEBUG(PHI->dump()); 384 LLVM_DEBUG(dbgs() << "To: "); 385 LLVM_DEBUG(NewPHI.getInstr()->dump()); 386 } 387 } 388 389 // Perform peephole optimizations. 390 bool PPCMIPeephole::simplifyCode() { 391 bool Simplified = false; 392 bool TrapOpt = false; 393 MachineInstr* ToErase = nullptr; 394 std::map<MachineInstr *, bool> TOCSaves; 395 const TargetRegisterInfo *TRI = &TII->getRegisterInfo(); 396 NumFunctionsEnteredInMIPeephole++; 397 if (ConvertRegReg) { 398 // Fixed-point conversion of reg/reg instructions fed by load-immediate 399 // into reg/imm instructions. FIXME: This is expensive, control it with 400 // an option. 401 bool SomethingChanged = false; 402 do { 403 NumFixedPointIterations++; 404 SomethingChanged = false; 405 for (MachineBasicBlock &MBB : *MF) { 406 for (MachineInstr &MI : MBB) { 407 if (MI.isDebugInstr()) 408 continue; 409 410 if (TII->convertToImmediateForm(MI)) { 411 // We don't erase anything in case the def has other uses. Let DCE 412 // remove it if it can be removed. 413 LLVM_DEBUG(dbgs() << "Converted instruction to imm form: "); 414 LLVM_DEBUG(MI.dump()); 415 NumConvertedToImmediateForm++; 416 SomethingChanged = true; 417 Simplified = true; 418 continue; 419 } 420 } 421 } 422 } while (SomethingChanged && FixedPointRegToImm); 423 } 424 425 for (MachineBasicBlock &MBB : *MF) { 426 for (MachineInstr &MI : MBB) { 427 428 // If the previous instruction was marked for elimination, 429 // remove it now. 430 if (ToErase) { 431 LLVM_DEBUG(dbgs() << "Deleting instruction: "); 432 LLVM_DEBUG(ToErase->dump()); 433 ToErase->eraseFromParent(); 434 ToErase = nullptr; 435 } 436 // If a conditional trap instruction got optimized to an 437 // unconditional trap, eliminate all the instructions after 438 // the trap. 439 if (EnableTrapOptimization && TrapOpt) { 440 ToErase = &MI; 441 continue; 442 } 443 444 // Ignore debug instructions. 445 if (MI.isDebugInstr()) 446 continue; 447 448 // Per-opcode peepholes. 449 switch (MI.getOpcode()) { 450 451 default: 452 break; 453 case PPC::COPY: { 454 Register Src = MI.getOperand(1).getReg(); 455 Register Dst = MI.getOperand(0).getReg(); 456 if (!Src.isVirtual() || !Dst.isVirtual()) 457 break; 458 if (MRI->getRegClass(Src) != &PPC::UACCRCRegClass || 459 MRI->getRegClass(Dst) != &PPC::ACCRCRegClass) 460 break; 461 462 // We are copying an unprimed accumulator to a primed accumulator. 463 // If the input to the copy is a PHI that is fed only by (i) copies in 464 // the other direction (ii) implicitly defined unprimed accumulators or 465 // (iii) other PHI nodes satisfying (i) and (ii), we can change 466 // the PHI to a PHI on primed accumulators (as long as we also change 467 // its operands). To detect and change such copies, we first get a list 468 // of all the PHI nodes starting from the root PHI node in BFS order. 469 // We then visit all these PHI nodes to check if they can be changed to 470 // primed accumulator PHI nodes and if so, we change them. 471 MachineInstr *RootPHI = MRI->getVRegDef(Src); 472 if (RootPHI->getOpcode() != PPC::PHI) 473 break; 474 475 SmallVector<MachineInstr *, 4> PHIs; 476 if (!collectUnprimedAccPHIs(MRI, RootPHI, PHIs)) 477 break; 478 479 convertUnprimedAccPHIs(TII, MRI, PHIs, Dst); 480 481 ToErase = &MI; 482 break; 483 } 484 case PPC::LI: 485 case PPC::LI8: { 486 // If we are materializing a zero, look for any use operands for which 487 // zero means immediate zero. All such operands can be replaced with 488 // PPC::ZERO. 489 if (!MI.getOperand(1).isImm() || MI.getOperand(1).getImm() != 0) 490 break; 491 Register MIDestReg = MI.getOperand(0).getReg(); 492 for (MachineInstr& UseMI : MRI->use_instructions(MIDestReg)) 493 Simplified |= TII->onlyFoldImmediate(UseMI, MI, MIDestReg); 494 if (MRI->use_nodbg_empty(MIDestReg)) { 495 ++NumLoadImmZeroFoldedAndRemoved; 496 ToErase = &MI; 497 } 498 break; 499 } 500 case PPC::STW: 501 case PPC::STD: { 502 MachineFrameInfo &MFI = MF->getFrameInfo(); 503 if (MFI.hasVarSizedObjects() || 504 (!MF->getSubtarget<PPCSubtarget>().isELFv2ABI() && 505 !MF->getSubtarget<PPCSubtarget>().isAIXABI())) 506 break; 507 // When encountering a TOC save instruction, call UpdateTOCSaves 508 // to add it to the TOCSaves map and mark any existing TOC saves 509 // it dominates as redundant. 510 if (TII->isTOCSaveMI(MI)) 511 UpdateTOCSaves(TOCSaves, &MI); 512 break; 513 } 514 case PPC::XXPERMDI: { 515 // Perform simplifications of 2x64 vector swaps and splats. 516 // A swap is identified by an immediate value of 2, and a splat 517 // is identified by an immediate value of 0 or 3. 518 int Immed = MI.getOperand(3).getImm(); 519 520 if (Immed == 1) 521 break; 522 523 // For each of these simplifications, we need the two source 524 // regs to match. Unfortunately, MachineCSE ignores COPY and 525 // SUBREG_TO_REG, so for example we can see 526 // XXPERMDI t, SUBREG_TO_REG(s), SUBREG_TO_REG(s), immed. 527 // We have to look through chains of COPY and SUBREG_TO_REG 528 // to find the real source values for comparison. 529 Register TrueReg1 = 530 TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI); 531 Register TrueReg2 = 532 TRI->lookThruCopyLike(MI.getOperand(2).getReg(), MRI); 533 534 if (!(TrueReg1 == TrueReg2 && TrueReg1.isVirtual())) 535 break; 536 537 MachineInstr *DefMI = MRI->getVRegDef(TrueReg1); 538 539 if (!DefMI) 540 break; 541 542 unsigned DefOpc = DefMI->getOpcode(); 543 544 // If this is a splat fed by a splatting load, the splat is 545 // redundant. Replace with a copy. This doesn't happen directly due 546 // to code in PPCDAGToDAGISel.cpp, but it can happen when converting 547 // a load of a double to a vector of 64-bit integers. 548 auto isConversionOfLoadAndSplat = [=]() -> bool { 549 if (DefOpc != PPC::XVCVDPSXDS && DefOpc != PPC::XVCVDPUXDS) 550 return false; 551 Register FeedReg1 = 552 TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI); 553 if (FeedReg1.isVirtual()) { 554 MachineInstr *LoadMI = MRI->getVRegDef(FeedReg1); 555 if (LoadMI && LoadMI->getOpcode() == PPC::LXVDSX) 556 return true; 557 } 558 return false; 559 }; 560 if ((Immed == 0 || Immed == 3) && 561 (DefOpc == PPC::LXVDSX || isConversionOfLoadAndSplat())) { 562 LLVM_DEBUG(dbgs() << "Optimizing load-and-splat/splat " 563 "to load-and-splat/copy: "); 564 LLVM_DEBUG(MI.dump()); 565 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 566 MI.getOperand(0).getReg()) 567 .add(MI.getOperand(1)); 568 ToErase = &MI; 569 Simplified = true; 570 } 571 572 // If this is a splat or a swap fed by another splat, we 573 // can replace it with a copy. 574 if (DefOpc == PPC::XXPERMDI) { 575 Register DefReg1 = DefMI->getOperand(1).getReg(); 576 Register DefReg2 = DefMI->getOperand(2).getReg(); 577 unsigned DefImmed = DefMI->getOperand(3).getImm(); 578 579 // If the two inputs are not the same register, check to see if 580 // they originate from the same virtual register after only 581 // copy-like instructions. 582 if (DefReg1 != DefReg2) { 583 Register FeedReg1 = TRI->lookThruCopyLike(DefReg1, MRI); 584 Register FeedReg2 = TRI->lookThruCopyLike(DefReg2, MRI); 585 586 if (!(FeedReg1 == FeedReg2 && FeedReg1.isVirtual())) 587 break; 588 } 589 590 if (DefImmed == 0 || DefImmed == 3) { 591 LLVM_DEBUG(dbgs() << "Optimizing splat/swap or splat/splat " 592 "to splat/copy: "); 593 LLVM_DEBUG(MI.dump()); 594 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 595 MI.getOperand(0).getReg()) 596 .add(MI.getOperand(1)); 597 ToErase = &MI; 598 Simplified = true; 599 } 600 601 // If this is a splat fed by a swap, we can simplify modify 602 // the splat to splat the other value from the swap's input 603 // parameter. 604 else if ((Immed == 0 || Immed == 3) && DefImmed == 2) { 605 LLVM_DEBUG(dbgs() << "Optimizing swap/splat => splat: "); 606 LLVM_DEBUG(MI.dump()); 607 MI.getOperand(1).setReg(DefReg1); 608 MI.getOperand(2).setReg(DefReg2); 609 MI.getOperand(3).setImm(3 - Immed); 610 Simplified = true; 611 } 612 613 // If this is a swap fed by a swap, we can replace it 614 // with a copy from the first swap's input. 615 else if (Immed == 2 && DefImmed == 2) { 616 LLVM_DEBUG(dbgs() << "Optimizing swap/swap => copy: "); 617 LLVM_DEBUG(MI.dump()); 618 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 619 MI.getOperand(0).getReg()) 620 .add(DefMI->getOperand(1)); 621 ToErase = &MI; 622 Simplified = true; 623 } 624 } else if ((Immed == 0 || Immed == 3 || Immed == 2) && 625 DefOpc == PPC::XXPERMDIs && 626 (DefMI->getOperand(2).getImm() == 0 || 627 DefMI->getOperand(2).getImm() == 3)) { 628 ToErase = &MI; 629 Simplified = true; 630 // Swap of a splat, convert to copy. 631 if (Immed == 2) { 632 LLVM_DEBUG(dbgs() << "Optimizing swap(splat) => copy(splat): "); 633 LLVM_DEBUG(MI.dump()); 634 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 635 MI.getOperand(0).getReg()) 636 .add(MI.getOperand(1)); 637 break; 638 } 639 // Splat fed by another splat - switch the output of the first 640 // and remove the second. 641 DefMI->getOperand(0).setReg(MI.getOperand(0).getReg()); 642 LLVM_DEBUG(dbgs() << "Removing redundant splat: "); 643 LLVM_DEBUG(MI.dump()); 644 } else if (Immed == 2 && 645 (DefOpc == PPC::VSPLTB || DefOpc == PPC::VSPLTH || 646 DefOpc == PPC::VSPLTW || DefOpc == PPC::XXSPLTW)) { 647 // Swap of various vector splats, convert to copy. 648 ToErase = &MI; 649 Simplified = true; 650 LLVM_DEBUG(dbgs() << "Optimizing swap(vsplt[b|h|w]|xxspltw) => " 651 "copy(vsplt[b|h|w]|xxspltw): "); 652 LLVM_DEBUG(MI.dump()); 653 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 654 MI.getOperand(0).getReg()) 655 .add(MI.getOperand(1)); 656 } else if ((Immed == 0 || Immed == 3 || Immed == 2) && 657 TII->isLoadFromConstantPool(DefMI)) { 658 const Constant *C = TII->getConstantFromConstantPool(DefMI); 659 if (C && C->getType()->isVectorTy() && C->getSplatValue()) { 660 ToErase = &MI; 661 Simplified = true; 662 LLVM_DEBUG(dbgs() 663 << "Optimizing swap(splat pattern from constant-pool) " 664 "=> copy(splat pattern from constant-pool): "); 665 LLVM_DEBUG(MI.dump()); 666 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 667 MI.getOperand(0).getReg()) 668 .add(MI.getOperand(1)); 669 } 670 } 671 break; 672 } 673 case PPC::VSPLTB: 674 case PPC::VSPLTH: 675 case PPC::XXSPLTW: { 676 unsigned MyOpcode = MI.getOpcode(); 677 unsigned OpNo = MyOpcode == PPC::XXSPLTW ? 1 : 2; 678 Register TrueReg = 679 TRI->lookThruCopyLike(MI.getOperand(OpNo).getReg(), MRI); 680 if (!TrueReg.isVirtual()) 681 break; 682 MachineInstr *DefMI = MRI->getVRegDef(TrueReg); 683 if (!DefMI) 684 break; 685 unsigned DefOpcode = DefMI->getOpcode(); 686 auto isConvertOfSplat = [=]() -> bool { 687 if (DefOpcode != PPC::XVCVSPSXWS && DefOpcode != PPC::XVCVSPUXWS) 688 return false; 689 Register ConvReg = DefMI->getOperand(1).getReg(); 690 if (!ConvReg.isVirtual()) 691 return false; 692 MachineInstr *Splt = MRI->getVRegDef(ConvReg); 693 return Splt && (Splt->getOpcode() == PPC::LXVWSX || 694 Splt->getOpcode() == PPC::XXSPLTW); 695 }; 696 bool AlreadySplat = (MyOpcode == DefOpcode) || 697 (MyOpcode == PPC::VSPLTB && DefOpcode == PPC::VSPLTBs) || 698 (MyOpcode == PPC::VSPLTH && DefOpcode == PPC::VSPLTHs) || 699 (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::XXSPLTWs) || 700 (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::LXVWSX) || 701 (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::MTVSRWS)|| 702 (MyOpcode == PPC::XXSPLTW && isConvertOfSplat()); 703 // If the instruction[s] that feed this splat have already splat 704 // the value, this splat is redundant. 705 if (AlreadySplat) { 706 LLVM_DEBUG(dbgs() << "Changing redundant splat to a copy: "); 707 LLVM_DEBUG(MI.dump()); 708 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 709 MI.getOperand(0).getReg()) 710 .add(MI.getOperand(OpNo)); 711 ToErase = &MI; 712 Simplified = true; 713 } 714 // Splat fed by a shift. Usually when we align value to splat into 715 // vector element zero. 716 if (DefOpcode == PPC::XXSLDWI) { 717 Register ShiftRes = DefMI->getOperand(0).getReg(); 718 Register ShiftOp1 = DefMI->getOperand(1).getReg(); 719 Register ShiftOp2 = DefMI->getOperand(2).getReg(); 720 unsigned ShiftImm = DefMI->getOperand(3).getImm(); 721 unsigned SplatImm = 722 MI.getOperand(MyOpcode == PPC::XXSPLTW ? 2 : 1).getImm(); 723 if (ShiftOp1 == ShiftOp2) { 724 unsigned NewElem = (SplatImm + ShiftImm) & 0x3; 725 if (MRI->hasOneNonDBGUse(ShiftRes)) { 726 LLVM_DEBUG(dbgs() << "Removing redundant shift: "); 727 LLVM_DEBUG(DefMI->dump()); 728 ToErase = DefMI; 729 } 730 Simplified = true; 731 LLVM_DEBUG(dbgs() << "Changing splat immediate from " << SplatImm 732 << " to " << NewElem << " in instruction: "); 733 LLVM_DEBUG(MI.dump()); 734 MI.getOperand(1).setReg(ShiftOp1); 735 MI.getOperand(2).setImm(NewElem); 736 } 737 } 738 break; 739 } 740 case PPC::XVCVDPSP: { 741 // If this is a DP->SP conversion fed by an FRSP, the FRSP is redundant. 742 Register TrueReg = 743 TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI); 744 if (!TrueReg.isVirtual()) 745 break; 746 MachineInstr *DefMI = MRI->getVRegDef(TrueReg); 747 748 // This can occur when building a vector of single precision or integer 749 // values. 750 if (DefMI && DefMI->getOpcode() == PPC::XXPERMDI) { 751 Register DefsReg1 = 752 TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI); 753 Register DefsReg2 = 754 TRI->lookThruCopyLike(DefMI->getOperand(2).getReg(), MRI); 755 if (!DefsReg1.isVirtual() || !DefsReg2.isVirtual()) 756 break; 757 MachineInstr *P1 = MRI->getVRegDef(DefsReg1); 758 MachineInstr *P2 = MRI->getVRegDef(DefsReg2); 759 760 if (!P1 || !P2) 761 break; 762 763 // Remove the passed FRSP/XSRSP instruction if it only feeds this MI 764 // and set any uses of that FRSP/XSRSP (in this MI) to the source of 765 // the FRSP/XSRSP. 766 auto removeFRSPIfPossible = [&](MachineInstr *RoundInstr) { 767 unsigned Opc = RoundInstr->getOpcode(); 768 if ((Opc == PPC::FRSP || Opc == PPC::XSRSP) && 769 MRI->hasOneNonDBGUse(RoundInstr->getOperand(0).getReg())) { 770 Simplified = true; 771 Register ConvReg1 = RoundInstr->getOperand(1).getReg(); 772 Register FRSPDefines = RoundInstr->getOperand(0).getReg(); 773 MachineInstr &Use = *(MRI->use_instr_nodbg_begin(FRSPDefines)); 774 for (int i = 0, e = Use.getNumOperands(); i < e; ++i) 775 if (Use.getOperand(i).isReg() && 776 Use.getOperand(i).getReg() == FRSPDefines) 777 Use.getOperand(i).setReg(ConvReg1); 778 LLVM_DEBUG(dbgs() << "Removing redundant FRSP/XSRSP:\n"); 779 LLVM_DEBUG(RoundInstr->dump()); 780 LLVM_DEBUG(dbgs() << "As it feeds instruction:\n"); 781 LLVM_DEBUG(MI.dump()); 782 LLVM_DEBUG(dbgs() << "Through instruction:\n"); 783 LLVM_DEBUG(DefMI->dump()); 784 RoundInstr->eraseFromParent(); 785 } 786 }; 787 788 // If the input to XVCVDPSP is a vector that was built (even 789 // partially) out of FRSP's, the FRSP(s) can safely be removed 790 // since this instruction performs the same operation. 791 if (P1 != P2) { 792 removeFRSPIfPossible(P1); 793 removeFRSPIfPossible(P2); 794 break; 795 } 796 removeFRSPIfPossible(P1); 797 } 798 break; 799 } 800 case PPC::EXTSH: 801 case PPC::EXTSH8: 802 case PPC::EXTSH8_32_64: { 803 if (!EnableSExtElimination) break; 804 Register NarrowReg = MI.getOperand(1).getReg(); 805 if (!NarrowReg.isVirtual()) 806 break; 807 808 MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg); 809 unsigned SrcOpcode = SrcMI->getOpcode(); 810 // If we've used a zero-extending load that we will sign-extend, 811 // just do a sign-extending load. 812 if (SrcOpcode == PPC::LHZ || SrcOpcode == PPC::LHZX) { 813 if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg())) 814 break; 815 // Determine the new opcode. We need to make sure that if the original 816 // instruction has a 64 bit opcode we keep using a 64 bit opcode. 817 // Likewise if the source is X-Form the new opcode should also be 818 // X-Form. 819 unsigned Opc = PPC::LHA; 820 bool SourceIsXForm = SrcOpcode == PPC::LHZX; 821 bool MIIs64Bit = MI.getOpcode() == PPC::EXTSH8 || 822 MI.getOpcode() == PPC::EXTSH8_32_64; 823 824 if (SourceIsXForm && MIIs64Bit) 825 Opc = PPC::LHAX8; 826 else if (SourceIsXForm && !MIIs64Bit) 827 Opc = PPC::LHAX; 828 else if (MIIs64Bit) 829 Opc = PPC::LHA8; 830 831 LLVM_DEBUG(dbgs() << "Zero-extending load\n"); 832 LLVM_DEBUG(SrcMI->dump()); 833 LLVM_DEBUG(dbgs() << "and sign-extension\n"); 834 LLVM_DEBUG(MI.dump()); 835 LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n"); 836 SrcMI->setDesc(TII->get(Opc)); 837 SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg()); 838 ToErase = &MI; 839 Simplified = true; 840 NumEliminatedSExt++; 841 } 842 break; 843 } 844 case PPC::EXTSW: 845 case PPC::EXTSW_32: 846 case PPC::EXTSW_32_64: { 847 if (!EnableSExtElimination) break; 848 Register NarrowReg = MI.getOperand(1).getReg(); 849 if (!NarrowReg.isVirtual()) 850 break; 851 852 MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg); 853 unsigned SrcOpcode = SrcMI->getOpcode(); 854 // If we've used a zero-extending load that we will sign-extend, 855 // just do a sign-extending load. 856 if (SrcOpcode == PPC::LWZ || SrcOpcode == PPC::LWZX) { 857 if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg())) 858 break; 859 860 // The transformation from a zero-extending load to a sign-extending 861 // load is only legal when the displacement is a multiple of 4. 862 // If the displacement is not at least 4 byte aligned, don't perform 863 // the transformation. 864 bool IsWordAligned = false; 865 if (SrcMI->getOperand(1).isGlobal()) { 866 const GlobalObject *GO = 867 dyn_cast<GlobalObject>(SrcMI->getOperand(1).getGlobal()); 868 if (GO && GO->getAlign() && *GO->getAlign() >= 4 && 869 (SrcMI->getOperand(1).getOffset() % 4 == 0)) 870 IsWordAligned = true; 871 } else if (SrcMI->getOperand(1).isImm()) { 872 int64_t Value = SrcMI->getOperand(1).getImm(); 873 if (Value % 4 == 0) 874 IsWordAligned = true; 875 } 876 877 // Determine the new opcode. We need to make sure that if the original 878 // instruction has a 64 bit opcode we keep using a 64 bit opcode. 879 // Likewise if the source is X-Form the new opcode should also be 880 // X-Form. 881 unsigned Opc = PPC::LWA_32; 882 bool SourceIsXForm = SrcOpcode == PPC::LWZX; 883 bool MIIs64Bit = MI.getOpcode() == PPC::EXTSW || 884 MI.getOpcode() == PPC::EXTSW_32_64; 885 886 if (SourceIsXForm && MIIs64Bit) 887 Opc = PPC::LWAX; 888 else if (SourceIsXForm && !MIIs64Bit) 889 Opc = PPC::LWAX_32; 890 else if (MIIs64Bit) 891 Opc = PPC::LWA; 892 893 if (!IsWordAligned && (Opc == PPC::LWA || Opc == PPC::LWA_32)) 894 break; 895 896 LLVM_DEBUG(dbgs() << "Zero-extending load\n"); 897 LLVM_DEBUG(SrcMI->dump()); 898 LLVM_DEBUG(dbgs() << "and sign-extension\n"); 899 LLVM_DEBUG(MI.dump()); 900 LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n"); 901 SrcMI->setDesc(TII->get(Opc)); 902 SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg()); 903 ToErase = &MI; 904 Simplified = true; 905 NumEliminatedSExt++; 906 } else if (MI.getOpcode() == PPC::EXTSW_32_64 && 907 TII->isSignExtended(NarrowReg, MRI)) { 908 // We can eliminate EXTSW if the input is known to be already 909 // sign-extended. 910 LLVM_DEBUG(dbgs() << "Removing redundant sign-extension\n"); 911 Register TmpReg = 912 MF->getRegInfo().createVirtualRegister(&PPC::G8RCRegClass); 913 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::IMPLICIT_DEF), 914 TmpReg); 915 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::INSERT_SUBREG), 916 MI.getOperand(0).getReg()) 917 .addReg(TmpReg) 918 .addReg(NarrowReg) 919 .addImm(PPC::sub_32); 920 ToErase = &MI; 921 Simplified = true; 922 NumEliminatedSExt++; 923 } 924 break; 925 } 926 case PPC::RLDICL: { 927 // We can eliminate RLDICL (e.g. for zero-extension) 928 // if all bits to clear are already zero in the input. 929 // This code assume following code sequence for zero-extension. 930 // %6 = COPY %5:sub_32; (optional) 931 // %8 = IMPLICIT_DEF; 932 // %7<def,tied1> = INSERT_SUBREG %8<tied0>, %6, sub_32; 933 if (!EnableZExtElimination) break; 934 935 if (MI.getOperand(2).getImm() != 0) 936 break; 937 938 Register SrcReg = MI.getOperand(1).getReg(); 939 if (!SrcReg.isVirtual()) 940 break; 941 942 MachineInstr *SrcMI = MRI->getVRegDef(SrcReg); 943 if (!(SrcMI && SrcMI->getOpcode() == PPC::INSERT_SUBREG && 944 SrcMI->getOperand(0).isReg() && SrcMI->getOperand(1).isReg())) 945 break; 946 947 MachineInstr *ImpDefMI, *SubRegMI; 948 ImpDefMI = MRI->getVRegDef(SrcMI->getOperand(1).getReg()); 949 SubRegMI = MRI->getVRegDef(SrcMI->getOperand(2).getReg()); 950 if (ImpDefMI->getOpcode() != PPC::IMPLICIT_DEF) break; 951 952 SrcMI = SubRegMI; 953 if (SubRegMI->getOpcode() == PPC::COPY) { 954 Register CopyReg = SubRegMI->getOperand(1).getReg(); 955 if (CopyReg.isVirtual()) 956 SrcMI = MRI->getVRegDef(CopyReg); 957 } 958 if (!SrcMI->getOperand(0).isReg()) 959 break; 960 961 unsigned KnownZeroCount = 962 getKnownLeadingZeroCount(SrcMI->getOperand(0).getReg(), TII, MRI); 963 if (MI.getOperand(3).getImm() <= KnownZeroCount) { 964 LLVM_DEBUG(dbgs() << "Removing redundant zero-extension\n"); 965 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 966 MI.getOperand(0).getReg()) 967 .addReg(SrcReg); 968 ToErase = &MI; 969 Simplified = true; 970 NumEliminatedZExt++; 971 } 972 break; 973 } 974 975 // TODO: Any instruction that has an immediate form fed only by a PHI 976 // whose operands are all load immediate can be folded away. We currently 977 // do this for ADD instructions, but should expand it to arithmetic and 978 // binary instructions with immediate forms in the future. 979 case PPC::ADD4: 980 case PPC::ADD8: { 981 auto isSingleUsePHI = [&](MachineOperand *PhiOp) { 982 assert(PhiOp && "Invalid Operand!"); 983 MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI); 984 985 return DefPhiMI && (DefPhiMI->getOpcode() == PPC::PHI) && 986 MRI->hasOneNonDBGUse(DefPhiMI->getOperand(0).getReg()); 987 }; 988 989 auto dominatesAllSingleUseLIs = [&](MachineOperand *DominatorOp, 990 MachineOperand *PhiOp) { 991 assert(PhiOp && "Invalid Operand!"); 992 assert(DominatorOp && "Invalid Operand!"); 993 MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI); 994 MachineInstr *DefDomMI = getVRegDefOrNull(DominatorOp, MRI); 995 996 // Note: the vregs only show up at odd indices position of PHI Node, 997 // the even indices position save the BB info. 998 for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) { 999 MachineInstr *LiMI = 1000 getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI); 1001 if (!LiMI || 1002 (LiMI->getOpcode() != PPC::LI && LiMI->getOpcode() != PPC::LI8) 1003 || !MRI->hasOneNonDBGUse(LiMI->getOperand(0).getReg()) || 1004 !MDT->dominates(DefDomMI, LiMI)) 1005 return false; 1006 } 1007 1008 return true; 1009 }; 1010 1011 MachineOperand Op1 = MI.getOperand(1); 1012 MachineOperand Op2 = MI.getOperand(2); 1013 if (isSingleUsePHI(&Op2) && dominatesAllSingleUseLIs(&Op1, &Op2)) 1014 std::swap(Op1, Op2); 1015 else if (!isSingleUsePHI(&Op1) || !dominatesAllSingleUseLIs(&Op2, &Op1)) 1016 break; // We don't have an ADD fed by LI's that can be transformed 1017 1018 // Now we know that Op1 is the PHI node and Op2 is the dominator 1019 Register DominatorReg = Op2.getReg(); 1020 1021 const TargetRegisterClass *TRC = MI.getOpcode() == PPC::ADD8 1022 ? &PPC::G8RC_and_G8RC_NOX0RegClass 1023 : &PPC::GPRC_and_GPRC_NOR0RegClass; 1024 MRI->setRegClass(DominatorReg, TRC); 1025 1026 // replace LIs with ADDIs 1027 MachineInstr *DefPhiMI = getVRegDefOrNull(&Op1, MRI); 1028 for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) { 1029 MachineInstr *LiMI = getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI); 1030 LLVM_DEBUG(dbgs() << "Optimizing LI to ADDI: "); 1031 LLVM_DEBUG(LiMI->dump()); 1032 1033 // There could be repeated registers in the PHI, e.g: %1 = 1034 // PHI %6, <%bb.2>, %8, <%bb.3>, %8, <%bb.6>; So if we've 1035 // already replaced the def instruction, skip. 1036 if (LiMI->getOpcode() == PPC::ADDI || LiMI->getOpcode() == PPC::ADDI8) 1037 continue; 1038 1039 assert((LiMI->getOpcode() == PPC::LI || 1040 LiMI->getOpcode() == PPC::LI8) && 1041 "Invalid Opcode!"); 1042 auto LiImm = LiMI->getOperand(1).getImm(); // save the imm of LI 1043 LiMI->removeOperand(1); // remove the imm of LI 1044 LiMI->setDesc(TII->get(LiMI->getOpcode() == PPC::LI ? PPC::ADDI 1045 : PPC::ADDI8)); 1046 MachineInstrBuilder(*LiMI->getParent()->getParent(), *LiMI) 1047 .addReg(DominatorReg) 1048 .addImm(LiImm); // restore the imm of LI 1049 LLVM_DEBUG(LiMI->dump()); 1050 } 1051 1052 // Replace ADD with COPY 1053 LLVM_DEBUG(dbgs() << "Optimizing ADD to COPY: "); 1054 LLVM_DEBUG(MI.dump()); 1055 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 1056 MI.getOperand(0).getReg()) 1057 .add(Op1); 1058 ToErase = &MI; 1059 Simplified = true; 1060 NumOptADDLIs++; 1061 break; 1062 } 1063 case PPC::RLDICR: { 1064 Simplified |= emitRLDICWhenLoweringJumpTables(MI) || 1065 combineSEXTAndSHL(MI, ToErase); 1066 break; 1067 } 1068 case PPC::RLWINM: 1069 case PPC::RLWINM_rec: 1070 case PPC::RLWINM8: 1071 case PPC::RLWINM8_rec: { 1072 Simplified = TII->combineRLWINM(MI, &ToErase); 1073 if (Simplified) 1074 ++NumRotatesCollapsed; 1075 break; 1076 } 1077 // We will replace TD/TW/TDI/TWI with an unconditional trap if it will 1078 // always trap, we will delete the node if it will never trap. 1079 case PPC::TDI: 1080 case PPC::TWI: 1081 case PPC::TD: 1082 case PPC::TW: { 1083 if (!EnableTrapOptimization) break; 1084 MachineInstr *LiMI1 = getVRegDefOrNull(&MI.getOperand(1), MRI); 1085 MachineInstr *LiMI2 = getVRegDefOrNull(&MI.getOperand(2), MRI); 1086 bool IsOperand2Immediate = MI.getOperand(2).isImm(); 1087 // We can only do the optimization if we can get immediates 1088 // from both operands 1089 if (!(LiMI1 && (LiMI1->getOpcode() == PPC::LI || 1090 LiMI1->getOpcode() == PPC::LI8))) 1091 break; 1092 if (!IsOperand2Immediate && 1093 !(LiMI2 && (LiMI2->getOpcode() == PPC::LI || 1094 LiMI2->getOpcode() == PPC::LI8))) 1095 break; 1096 1097 auto ImmOperand0 = MI.getOperand(0).getImm(); 1098 auto ImmOperand1 = LiMI1->getOperand(1).getImm(); 1099 auto ImmOperand2 = IsOperand2Immediate ? MI.getOperand(2).getImm() 1100 : LiMI2->getOperand(1).getImm(); 1101 1102 // We will replace the MI with an unconditional trap if it will always 1103 // trap. 1104 if ((ImmOperand0 == 31) || 1105 ((ImmOperand0 & 0x10) && 1106 ((int64_t)ImmOperand1 < (int64_t)ImmOperand2)) || 1107 ((ImmOperand0 & 0x8) && 1108 ((int64_t)ImmOperand1 > (int64_t)ImmOperand2)) || 1109 ((ImmOperand0 & 0x2) && 1110 ((uint64_t)ImmOperand1 < (uint64_t)ImmOperand2)) || 1111 ((ImmOperand0 & 0x1) && 1112 ((uint64_t)ImmOperand1 > (uint64_t)ImmOperand2)) || 1113 ((ImmOperand0 & 0x4) && (ImmOperand1 == ImmOperand2))) { 1114 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::TRAP)); 1115 TrapOpt = true; 1116 } 1117 // We will delete the MI if it will never trap. 1118 ToErase = &MI; 1119 Simplified = true; 1120 break; 1121 } 1122 } 1123 } 1124 1125 // If the last instruction was marked for elimination, 1126 // remove it now. 1127 if (ToErase) { 1128 ToErase->eraseFromParent(); 1129 ToErase = nullptr; 1130 } 1131 // Reset TrapOpt to false at the end of the basic block. 1132 if (EnableTrapOptimization) 1133 TrapOpt = false; 1134 } 1135 1136 // Eliminate all the TOC save instructions which are redundant. 1137 Simplified |= eliminateRedundantTOCSaves(TOCSaves); 1138 PPCFunctionInfo *FI = MF->getInfo<PPCFunctionInfo>(); 1139 if (FI->mustSaveTOC()) 1140 NumTOCSavesInPrologue++; 1141 1142 // We try to eliminate redundant compare instruction. 1143 Simplified |= eliminateRedundantCompare(); 1144 1145 return Simplified; 1146 } 1147 1148 // helper functions for eliminateRedundantCompare 1149 static bool isEqOrNe(MachineInstr *BI) { 1150 PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); 1151 unsigned PredCond = PPC::getPredicateCondition(Pred); 1152 return (PredCond == PPC::PRED_EQ || PredCond == PPC::PRED_NE); 1153 } 1154 1155 static bool isSupportedCmpOp(unsigned opCode) { 1156 return (opCode == PPC::CMPLD || opCode == PPC::CMPD || 1157 opCode == PPC::CMPLW || opCode == PPC::CMPW || 1158 opCode == PPC::CMPLDI || opCode == PPC::CMPDI || 1159 opCode == PPC::CMPLWI || opCode == PPC::CMPWI); 1160 } 1161 1162 static bool is64bitCmpOp(unsigned opCode) { 1163 return (opCode == PPC::CMPLD || opCode == PPC::CMPD || 1164 opCode == PPC::CMPLDI || opCode == PPC::CMPDI); 1165 } 1166 1167 static bool isSignedCmpOp(unsigned opCode) { 1168 return (opCode == PPC::CMPD || opCode == PPC::CMPW || 1169 opCode == PPC::CMPDI || opCode == PPC::CMPWI); 1170 } 1171 1172 static unsigned getSignedCmpOpCode(unsigned opCode) { 1173 if (opCode == PPC::CMPLD) return PPC::CMPD; 1174 if (opCode == PPC::CMPLW) return PPC::CMPW; 1175 if (opCode == PPC::CMPLDI) return PPC::CMPDI; 1176 if (opCode == PPC::CMPLWI) return PPC::CMPWI; 1177 return opCode; 1178 } 1179 1180 // We can decrement immediate x in (GE x) by changing it to (GT x-1) or 1181 // (LT x) to (LE x-1) 1182 static unsigned getPredicateToDecImm(MachineInstr *BI, MachineInstr *CMPI) { 1183 uint64_t Imm = CMPI->getOperand(2).getImm(); 1184 bool SignedCmp = isSignedCmpOp(CMPI->getOpcode()); 1185 if ((!SignedCmp && Imm == 0) || (SignedCmp && Imm == 0x8000)) 1186 return 0; 1187 1188 PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); 1189 unsigned PredCond = PPC::getPredicateCondition(Pred); 1190 unsigned PredHint = PPC::getPredicateHint(Pred); 1191 if (PredCond == PPC::PRED_GE) 1192 return PPC::getPredicate(PPC::PRED_GT, PredHint); 1193 if (PredCond == PPC::PRED_LT) 1194 return PPC::getPredicate(PPC::PRED_LE, PredHint); 1195 1196 return 0; 1197 } 1198 1199 // We can increment immediate x in (GT x) by changing it to (GE x+1) or 1200 // (LE x) to (LT x+1) 1201 static unsigned getPredicateToIncImm(MachineInstr *BI, MachineInstr *CMPI) { 1202 uint64_t Imm = CMPI->getOperand(2).getImm(); 1203 bool SignedCmp = isSignedCmpOp(CMPI->getOpcode()); 1204 if ((!SignedCmp && Imm == 0xFFFF) || (SignedCmp && Imm == 0x7FFF)) 1205 return 0; 1206 1207 PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); 1208 unsigned PredCond = PPC::getPredicateCondition(Pred); 1209 unsigned PredHint = PPC::getPredicateHint(Pred); 1210 if (PredCond == PPC::PRED_GT) 1211 return PPC::getPredicate(PPC::PRED_GE, PredHint); 1212 if (PredCond == PPC::PRED_LE) 1213 return PPC::getPredicate(PPC::PRED_LT, PredHint); 1214 1215 return 0; 1216 } 1217 1218 // This takes a Phi node and returns a register value for the specified BB. 1219 static unsigned getIncomingRegForBlock(MachineInstr *Phi, 1220 MachineBasicBlock *MBB) { 1221 for (unsigned I = 2, E = Phi->getNumOperands() + 1; I != E; I += 2) { 1222 MachineOperand &MO = Phi->getOperand(I); 1223 if (MO.getMBB() == MBB) 1224 return Phi->getOperand(I-1).getReg(); 1225 } 1226 llvm_unreachable("invalid src basic block for this Phi node\n"); 1227 return 0; 1228 } 1229 1230 // This function tracks the source of the register through register copy. 1231 // If BB1 and BB2 are non-NULL, we also track PHI instruction in BB2 1232 // assuming that the control comes from BB1 into BB2. 1233 static unsigned getSrcVReg(unsigned Reg, MachineBasicBlock *BB1, 1234 MachineBasicBlock *BB2, MachineRegisterInfo *MRI) { 1235 unsigned SrcReg = Reg; 1236 while (true) { 1237 unsigned NextReg = SrcReg; 1238 MachineInstr *Inst = MRI->getVRegDef(SrcReg); 1239 if (BB1 && Inst->getOpcode() == PPC::PHI && Inst->getParent() == BB2) { 1240 NextReg = getIncomingRegForBlock(Inst, BB1); 1241 // We track through PHI only once to avoid infinite loop. 1242 BB1 = nullptr; 1243 } 1244 else if (Inst->isFullCopy()) 1245 NextReg = Inst->getOperand(1).getReg(); 1246 if (NextReg == SrcReg || !Register::isVirtualRegister(NextReg)) 1247 break; 1248 SrcReg = NextReg; 1249 } 1250 return SrcReg; 1251 } 1252 1253 static bool eligibleForCompareElimination(MachineBasicBlock &MBB, 1254 MachineBasicBlock *&PredMBB, 1255 MachineBasicBlock *&MBBtoMoveCmp, 1256 MachineRegisterInfo *MRI) { 1257 1258 auto isEligibleBB = [&](MachineBasicBlock &BB) { 1259 auto BII = BB.getFirstInstrTerminator(); 1260 // We optimize BBs ending with a conditional branch. 1261 // We check only for BCC here, not BCCLR, because BCCLR 1262 // will be formed only later in the pipeline. 1263 if (BB.succ_size() == 2 && 1264 BII != BB.instr_end() && 1265 (*BII).getOpcode() == PPC::BCC && 1266 (*BII).getOperand(1).isReg()) { 1267 // We optimize only if the condition code is used only by one BCC. 1268 Register CndReg = (*BII).getOperand(1).getReg(); 1269 if (!CndReg.isVirtual() || !MRI->hasOneNonDBGUse(CndReg)) 1270 return false; 1271 1272 MachineInstr *CMPI = MRI->getVRegDef(CndReg); 1273 // We assume compare and branch are in the same BB for ease of analysis. 1274 if (CMPI->getParent() != &BB) 1275 return false; 1276 1277 // We skip this BB if a physical register is used in comparison. 1278 for (MachineOperand &MO : CMPI->operands()) 1279 if (MO.isReg() && !MO.getReg().isVirtual()) 1280 return false; 1281 1282 return true; 1283 } 1284 return false; 1285 }; 1286 1287 // If this BB has more than one successor, we can create a new BB and 1288 // move the compare instruction in the new BB. 1289 // So far, we do not move compare instruction to a BB having multiple 1290 // successors to avoid potentially increasing code size. 1291 auto isEligibleForMoveCmp = [](MachineBasicBlock &BB) { 1292 return BB.succ_size() == 1; 1293 }; 1294 1295 if (!isEligibleBB(MBB)) 1296 return false; 1297 1298 unsigned NumPredBBs = MBB.pred_size(); 1299 if (NumPredBBs == 1) { 1300 MachineBasicBlock *TmpMBB = *MBB.pred_begin(); 1301 if (isEligibleBB(*TmpMBB)) { 1302 PredMBB = TmpMBB; 1303 MBBtoMoveCmp = nullptr; 1304 return true; 1305 } 1306 } 1307 else if (NumPredBBs == 2) { 1308 // We check for partially redundant case. 1309 // So far, we support cases with only two predecessors 1310 // to avoid increasing the number of instructions. 1311 MachineBasicBlock::pred_iterator PI = MBB.pred_begin(); 1312 MachineBasicBlock *Pred1MBB = *PI; 1313 MachineBasicBlock *Pred2MBB = *(PI+1); 1314 1315 if (isEligibleBB(*Pred1MBB) && isEligibleForMoveCmp(*Pred2MBB)) { 1316 // We assume Pred1MBB is the BB containing the compare to be merged and 1317 // Pred2MBB is the BB to which we will append a compare instruction. 1318 // Proceed as is if Pred1MBB is different from MBB. 1319 } 1320 else if (isEligibleBB(*Pred2MBB) && isEligibleForMoveCmp(*Pred1MBB)) { 1321 // We need to swap Pred1MBB and Pred2MBB to canonicalize. 1322 std::swap(Pred1MBB, Pred2MBB); 1323 } 1324 else return false; 1325 1326 if (Pred1MBB == &MBB) 1327 return false; 1328 1329 // Here, Pred2MBB is the BB to which we need to append a compare inst. 1330 // We cannot move the compare instruction if operands are not available 1331 // in Pred2MBB (i.e. defined in MBB by an instruction other than PHI). 1332 MachineInstr *BI = &*MBB.getFirstInstrTerminator(); 1333 MachineInstr *CMPI = MRI->getVRegDef(BI->getOperand(1).getReg()); 1334 for (int I = 1; I <= 2; I++) 1335 if (CMPI->getOperand(I).isReg()) { 1336 MachineInstr *Inst = MRI->getVRegDef(CMPI->getOperand(I).getReg()); 1337 if (Inst->getParent() == &MBB && Inst->getOpcode() != PPC::PHI) 1338 return false; 1339 } 1340 1341 PredMBB = Pred1MBB; 1342 MBBtoMoveCmp = Pred2MBB; 1343 return true; 1344 } 1345 1346 return false; 1347 } 1348 1349 // This function will iterate over the input map containing a pair of TOC save 1350 // instruction and a flag. The flag will be set to false if the TOC save is 1351 // proven redundant. This function will erase from the basic block all the TOC 1352 // saves marked as redundant. 1353 bool PPCMIPeephole::eliminateRedundantTOCSaves( 1354 std::map<MachineInstr *, bool> &TOCSaves) { 1355 bool Simplified = false; 1356 int NumKept = 0; 1357 for (auto TOCSave : TOCSaves) { 1358 if (!TOCSave.second) { 1359 TOCSave.first->eraseFromParent(); 1360 RemoveTOCSave++; 1361 Simplified = true; 1362 } else { 1363 NumKept++; 1364 } 1365 } 1366 1367 if (NumKept > 1) 1368 MultiTOCSaves++; 1369 1370 return Simplified; 1371 } 1372 1373 // If multiple conditional branches are executed based on the (essentially) 1374 // same comparison, we merge compare instructions into one and make multiple 1375 // conditional branches on this comparison. 1376 // For example, 1377 // if (a == 0) { ... } 1378 // else if (a < 0) { ... } 1379 // can be executed by one compare and two conditional branches instead of 1380 // two pairs of a compare and a conditional branch. 1381 // 1382 // This method merges two compare instructions in two MBBs and modifies the 1383 // compare and conditional branch instructions if needed. 1384 // For the above example, the input for this pass looks like: 1385 // cmplwi r3, 0 1386 // beq 0, .LBB0_3 1387 // cmpwi r3, -1 1388 // bgt 0, .LBB0_4 1389 // So, before merging two compares, we need to modify these instructions as 1390 // cmpwi r3, 0 ; cmplwi and cmpwi yield same result for beq 1391 // beq 0, .LBB0_3 1392 // cmpwi r3, 0 ; greather than -1 means greater or equal to 0 1393 // bge 0, .LBB0_4 1394 1395 bool PPCMIPeephole::eliminateRedundantCompare() { 1396 bool Simplified = false; 1397 1398 for (MachineBasicBlock &MBB2 : *MF) { 1399 MachineBasicBlock *MBB1 = nullptr, *MBBtoMoveCmp = nullptr; 1400 1401 // For fully redundant case, we select two basic blocks MBB1 and MBB2 1402 // as an optimization target if 1403 // - both MBBs end with a conditional branch, 1404 // - MBB1 is the only predecessor of MBB2, and 1405 // - compare does not take a physical register as a operand in both MBBs. 1406 // In this case, eligibleForCompareElimination sets MBBtoMoveCmp nullptr. 1407 // 1408 // As partially redundant case, we additionally handle if MBB2 has one 1409 // additional predecessor, which has only one successor (MBB2). 1410 // In this case, we move the compare instruction originally in MBB2 into 1411 // MBBtoMoveCmp. This partially redundant case is typically appear by 1412 // compiling a while loop; here, MBBtoMoveCmp is the loop preheader. 1413 // 1414 // Overview of CFG of related basic blocks 1415 // Fully redundant case Partially redundant case 1416 // -------- ---------------- -------- 1417 // | MBB1 | (w/ 2 succ) | MBBtoMoveCmp | | MBB1 | (w/ 2 succ) 1418 // -------- ---------------- -------- 1419 // | \ (w/ 1 succ) \ | \ 1420 // | \ \ | \ 1421 // | \ | 1422 // -------- -------- 1423 // | MBB2 | (w/ 1 pred | MBB2 | (w/ 2 pred 1424 // -------- and 2 succ) -------- and 2 succ) 1425 // | \ | \ 1426 // | \ | \ 1427 // 1428 if (!eligibleForCompareElimination(MBB2, MBB1, MBBtoMoveCmp, MRI)) 1429 continue; 1430 1431 MachineInstr *BI1 = &*MBB1->getFirstInstrTerminator(); 1432 MachineInstr *CMPI1 = MRI->getVRegDef(BI1->getOperand(1).getReg()); 1433 1434 MachineInstr *BI2 = &*MBB2.getFirstInstrTerminator(); 1435 MachineInstr *CMPI2 = MRI->getVRegDef(BI2->getOperand(1).getReg()); 1436 bool IsPartiallyRedundant = (MBBtoMoveCmp != nullptr); 1437 1438 // We cannot optimize an unsupported compare opcode or 1439 // a mix of 32-bit and 64-bit comparisons 1440 if (!isSupportedCmpOp(CMPI1->getOpcode()) || 1441 !isSupportedCmpOp(CMPI2->getOpcode()) || 1442 is64bitCmpOp(CMPI1->getOpcode()) != is64bitCmpOp(CMPI2->getOpcode())) 1443 continue; 1444 1445 unsigned NewOpCode = 0; 1446 unsigned NewPredicate1 = 0, NewPredicate2 = 0; 1447 int16_t Imm1 = 0, NewImm1 = 0, Imm2 = 0, NewImm2 = 0; 1448 bool SwapOperands = false; 1449 1450 if (CMPI1->getOpcode() != CMPI2->getOpcode()) { 1451 // Typically, unsigned comparison is used for equality check, but 1452 // we replace it with a signed comparison if the comparison 1453 // to be merged is a signed comparison. 1454 // In other cases of opcode mismatch, we cannot optimize this. 1455 1456 // We cannot change opcode when comparing against an immediate 1457 // if the most significant bit of the immediate is one 1458 // due to the difference in sign extension. 1459 auto CmpAgainstImmWithSignBit = [](MachineInstr *I) { 1460 if (!I->getOperand(2).isImm()) 1461 return false; 1462 int16_t Imm = (int16_t)I->getOperand(2).getImm(); 1463 return Imm < 0; 1464 }; 1465 1466 if (isEqOrNe(BI2) && !CmpAgainstImmWithSignBit(CMPI2) && 1467 CMPI1->getOpcode() == getSignedCmpOpCode(CMPI2->getOpcode())) 1468 NewOpCode = CMPI1->getOpcode(); 1469 else if (isEqOrNe(BI1) && !CmpAgainstImmWithSignBit(CMPI1) && 1470 getSignedCmpOpCode(CMPI1->getOpcode()) == CMPI2->getOpcode()) 1471 NewOpCode = CMPI2->getOpcode(); 1472 else continue; 1473 } 1474 1475 if (CMPI1->getOperand(2).isReg() && CMPI2->getOperand(2).isReg()) { 1476 // In case of comparisons between two registers, these two registers 1477 // must be same to merge two comparisons. 1478 unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(), 1479 nullptr, nullptr, MRI); 1480 unsigned Cmp1Operand2 = getSrcVReg(CMPI1->getOperand(2).getReg(), 1481 nullptr, nullptr, MRI); 1482 unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(), 1483 MBB1, &MBB2, MRI); 1484 unsigned Cmp2Operand2 = getSrcVReg(CMPI2->getOperand(2).getReg(), 1485 MBB1, &MBB2, MRI); 1486 1487 if (Cmp1Operand1 == Cmp2Operand1 && Cmp1Operand2 == Cmp2Operand2) { 1488 // Same pair of registers in the same order; ready to merge as is. 1489 } 1490 else if (Cmp1Operand1 == Cmp2Operand2 && Cmp1Operand2 == Cmp2Operand1) { 1491 // Same pair of registers in different order. 1492 // We reverse the predicate to merge compare instructions. 1493 PPC::Predicate Pred = (PPC::Predicate)BI2->getOperand(0).getImm(); 1494 NewPredicate2 = (unsigned)PPC::getSwappedPredicate(Pred); 1495 // In case of partial redundancy, we need to swap operands 1496 // in another compare instruction. 1497 SwapOperands = true; 1498 } 1499 else continue; 1500 } 1501 else if (CMPI1->getOperand(2).isImm() && CMPI2->getOperand(2).isImm()) { 1502 // In case of comparisons between a register and an immediate, 1503 // the operand register must be same for two compare instructions. 1504 unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(), 1505 nullptr, nullptr, MRI); 1506 unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(), 1507 MBB1, &MBB2, MRI); 1508 if (Cmp1Operand1 != Cmp2Operand1) 1509 continue; 1510 1511 NewImm1 = Imm1 = (int16_t)CMPI1->getOperand(2).getImm(); 1512 NewImm2 = Imm2 = (int16_t)CMPI2->getOperand(2).getImm(); 1513 1514 // If immediate are not same, we try to adjust by changing predicate; 1515 // e.g. GT imm means GE (imm+1). 1516 if (Imm1 != Imm2 && (!isEqOrNe(BI2) || !isEqOrNe(BI1))) { 1517 int Diff = Imm1 - Imm2; 1518 if (Diff < -2 || Diff > 2) 1519 continue; 1520 1521 unsigned PredToInc1 = getPredicateToIncImm(BI1, CMPI1); 1522 unsigned PredToDec1 = getPredicateToDecImm(BI1, CMPI1); 1523 unsigned PredToInc2 = getPredicateToIncImm(BI2, CMPI2); 1524 unsigned PredToDec2 = getPredicateToDecImm(BI2, CMPI2); 1525 if (Diff == 2) { 1526 if (PredToInc2 && PredToDec1) { 1527 NewPredicate2 = PredToInc2; 1528 NewPredicate1 = PredToDec1; 1529 NewImm2++; 1530 NewImm1--; 1531 } 1532 } 1533 else if (Diff == 1) { 1534 if (PredToInc2) { 1535 NewImm2++; 1536 NewPredicate2 = PredToInc2; 1537 } 1538 else if (PredToDec1) { 1539 NewImm1--; 1540 NewPredicate1 = PredToDec1; 1541 } 1542 } 1543 else if (Diff == -1) { 1544 if (PredToDec2) { 1545 NewImm2--; 1546 NewPredicate2 = PredToDec2; 1547 } 1548 else if (PredToInc1) { 1549 NewImm1++; 1550 NewPredicate1 = PredToInc1; 1551 } 1552 } 1553 else if (Diff == -2) { 1554 if (PredToDec2 && PredToInc1) { 1555 NewPredicate2 = PredToDec2; 1556 NewPredicate1 = PredToInc1; 1557 NewImm2--; 1558 NewImm1++; 1559 } 1560 } 1561 } 1562 1563 // We cannot merge two compares if the immediates are not same. 1564 if (NewImm2 != NewImm1) 1565 continue; 1566 } 1567 1568 LLVM_DEBUG(dbgs() << "Optimize two pairs of compare and branch:\n"); 1569 LLVM_DEBUG(CMPI1->dump()); 1570 LLVM_DEBUG(BI1->dump()); 1571 LLVM_DEBUG(CMPI2->dump()); 1572 LLVM_DEBUG(BI2->dump()); 1573 1574 // We adjust opcode, predicates and immediate as we determined above. 1575 if (NewOpCode != 0 && NewOpCode != CMPI1->getOpcode()) { 1576 CMPI1->setDesc(TII->get(NewOpCode)); 1577 } 1578 if (NewPredicate1) { 1579 BI1->getOperand(0).setImm(NewPredicate1); 1580 } 1581 if (NewPredicate2) { 1582 BI2->getOperand(0).setImm(NewPredicate2); 1583 } 1584 if (NewImm1 != Imm1) { 1585 CMPI1->getOperand(2).setImm(NewImm1); 1586 } 1587 1588 if (IsPartiallyRedundant) { 1589 // We touch up the compare instruction in MBB2 and move it to 1590 // a previous BB to handle partially redundant case. 1591 if (SwapOperands) { 1592 Register Op1 = CMPI2->getOperand(1).getReg(); 1593 Register Op2 = CMPI2->getOperand(2).getReg(); 1594 CMPI2->getOperand(1).setReg(Op2); 1595 CMPI2->getOperand(2).setReg(Op1); 1596 } 1597 if (NewImm2 != Imm2) 1598 CMPI2->getOperand(2).setImm(NewImm2); 1599 1600 for (int I = 1; I <= 2; I++) { 1601 if (CMPI2->getOperand(I).isReg()) { 1602 MachineInstr *Inst = MRI->getVRegDef(CMPI2->getOperand(I).getReg()); 1603 if (Inst->getParent() != &MBB2) 1604 continue; 1605 1606 assert(Inst->getOpcode() == PPC::PHI && 1607 "We cannot support if an operand comes from this BB."); 1608 unsigned SrcReg = getIncomingRegForBlock(Inst, MBBtoMoveCmp); 1609 CMPI2->getOperand(I).setReg(SrcReg); 1610 } 1611 } 1612 auto I = MachineBasicBlock::iterator(MBBtoMoveCmp->getFirstTerminator()); 1613 MBBtoMoveCmp->splice(I, &MBB2, MachineBasicBlock::iterator(CMPI2)); 1614 1615 DebugLoc DL = CMPI2->getDebugLoc(); 1616 Register NewVReg = MRI->createVirtualRegister(&PPC::CRRCRegClass); 1617 BuildMI(MBB2, MBB2.begin(), DL, 1618 TII->get(PPC::PHI), NewVReg) 1619 .addReg(BI1->getOperand(1).getReg()).addMBB(MBB1) 1620 .addReg(BI2->getOperand(1).getReg()).addMBB(MBBtoMoveCmp); 1621 BI2->getOperand(1).setReg(NewVReg); 1622 } 1623 else { 1624 // We finally eliminate compare instruction in MBB2. 1625 BI2->getOperand(1).setReg(BI1->getOperand(1).getReg()); 1626 CMPI2->eraseFromParent(); 1627 } 1628 BI2->getOperand(1).setIsKill(true); 1629 BI1->getOperand(1).setIsKill(false); 1630 1631 LLVM_DEBUG(dbgs() << "into a compare and two branches:\n"); 1632 LLVM_DEBUG(CMPI1->dump()); 1633 LLVM_DEBUG(BI1->dump()); 1634 LLVM_DEBUG(BI2->dump()); 1635 if (IsPartiallyRedundant) { 1636 LLVM_DEBUG(dbgs() << "The following compare is moved into " 1637 << printMBBReference(*MBBtoMoveCmp) 1638 << " to handle partial redundancy.\n"); 1639 LLVM_DEBUG(CMPI2->dump()); 1640 } 1641 1642 Simplified = true; 1643 } 1644 1645 return Simplified; 1646 } 1647 1648 // We miss the opportunity to emit an RLDIC when lowering jump tables 1649 // since ISEL sees only a single basic block. When selecting, the clear 1650 // and shift left will be in different blocks. 1651 bool PPCMIPeephole::emitRLDICWhenLoweringJumpTables(MachineInstr &MI) { 1652 if (MI.getOpcode() != PPC::RLDICR) 1653 return false; 1654 1655 Register SrcReg = MI.getOperand(1).getReg(); 1656 if (!SrcReg.isVirtual()) 1657 return false; 1658 1659 MachineInstr *SrcMI = MRI->getVRegDef(SrcReg); 1660 if (SrcMI->getOpcode() != PPC::RLDICL) 1661 return false; 1662 1663 MachineOperand MOpSHSrc = SrcMI->getOperand(2); 1664 MachineOperand MOpMBSrc = SrcMI->getOperand(3); 1665 MachineOperand MOpSHMI = MI.getOperand(2); 1666 MachineOperand MOpMEMI = MI.getOperand(3); 1667 if (!(MOpSHSrc.isImm() && MOpMBSrc.isImm() && MOpSHMI.isImm() && 1668 MOpMEMI.isImm())) 1669 return false; 1670 1671 uint64_t SHSrc = MOpSHSrc.getImm(); 1672 uint64_t MBSrc = MOpMBSrc.getImm(); 1673 uint64_t SHMI = MOpSHMI.getImm(); 1674 uint64_t MEMI = MOpMEMI.getImm(); 1675 uint64_t NewSH = SHSrc + SHMI; 1676 uint64_t NewMB = MBSrc - SHMI; 1677 if (NewMB > 63 || NewSH > 63) 1678 return false; 1679 1680 // The bits cleared with RLDICL are [0, MBSrc). 1681 // The bits cleared with RLDICR are (MEMI, 63]. 1682 // After the sequence, the bits cleared are: 1683 // [0, MBSrc-SHMI) and (MEMI, 63). 1684 // 1685 // The bits cleared with RLDIC are [0, NewMB) and (63-NewSH, 63]. 1686 if ((63 - NewSH) != MEMI) 1687 return false; 1688 1689 LLVM_DEBUG(dbgs() << "Converting pair: "); 1690 LLVM_DEBUG(SrcMI->dump()); 1691 LLVM_DEBUG(MI.dump()); 1692 1693 MI.setDesc(TII->get(PPC::RLDIC)); 1694 MI.getOperand(1).setReg(SrcMI->getOperand(1).getReg()); 1695 MI.getOperand(2).setImm(NewSH); 1696 MI.getOperand(3).setImm(NewMB); 1697 MI.getOperand(1).setIsKill(SrcMI->getOperand(1).isKill()); 1698 SrcMI->getOperand(1).setIsKill(false); 1699 1700 LLVM_DEBUG(dbgs() << "To: "); 1701 LLVM_DEBUG(MI.dump()); 1702 NumRotatesCollapsed++; 1703 // If SrcReg has no non-debug use it's safe to delete its def SrcMI. 1704 if (MRI->use_nodbg_empty(SrcReg)) { 1705 assert(!SrcMI->hasImplicitDef() && 1706 "Not expecting an implicit def with this instr."); 1707 SrcMI->eraseFromParent(); 1708 } 1709 return true; 1710 } 1711 1712 // For case in LLVM IR 1713 // entry: 1714 // %iconv = sext i32 %index to i64 1715 // br i1 undef label %true, label %false 1716 // true: 1717 // %ptr = getelementptr inbounds i32, i32* null, i64 %iconv 1718 // ... 1719 // PPCISelLowering::combineSHL fails to combine, because sext and shl are in 1720 // different BBs when conducting instruction selection. We can do a peephole 1721 // optimization to combine these two instructions into extswsli after 1722 // instruction selection. 1723 bool PPCMIPeephole::combineSEXTAndSHL(MachineInstr &MI, 1724 MachineInstr *&ToErase) { 1725 if (MI.getOpcode() != PPC::RLDICR) 1726 return false; 1727 1728 if (!MF->getSubtarget<PPCSubtarget>().isISA3_0()) 1729 return false; 1730 1731 assert(MI.getNumOperands() == 4 && "RLDICR should have 4 operands"); 1732 1733 MachineOperand MOpSHMI = MI.getOperand(2); 1734 MachineOperand MOpMEMI = MI.getOperand(3); 1735 if (!(MOpSHMI.isImm() && MOpMEMI.isImm())) 1736 return false; 1737 1738 uint64_t SHMI = MOpSHMI.getImm(); 1739 uint64_t MEMI = MOpMEMI.getImm(); 1740 if (SHMI + MEMI != 63) 1741 return false; 1742 1743 Register SrcReg = MI.getOperand(1).getReg(); 1744 if (!SrcReg.isVirtual()) 1745 return false; 1746 1747 MachineInstr *SrcMI = MRI->getVRegDef(SrcReg); 1748 if (SrcMI->getOpcode() != PPC::EXTSW && 1749 SrcMI->getOpcode() != PPC::EXTSW_32_64) 1750 return false; 1751 1752 // If the register defined by extsw has more than one use, combination is not 1753 // needed. 1754 if (!MRI->hasOneNonDBGUse(SrcReg)) 1755 return false; 1756 1757 assert(SrcMI->getNumOperands() == 2 && "EXTSW should have 2 operands"); 1758 assert(SrcMI->getOperand(1).isReg() && 1759 "EXTSW's second operand should be a register"); 1760 if (!SrcMI->getOperand(1).getReg().isVirtual()) 1761 return false; 1762 1763 LLVM_DEBUG(dbgs() << "Combining pair: "); 1764 LLVM_DEBUG(SrcMI->dump()); 1765 LLVM_DEBUG(MI.dump()); 1766 1767 MachineInstr *NewInstr = 1768 BuildMI(*MI.getParent(), &MI, MI.getDebugLoc(), 1769 SrcMI->getOpcode() == PPC::EXTSW ? TII->get(PPC::EXTSWSLI) 1770 : TII->get(PPC::EXTSWSLI_32_64), 1771 MI.getOperand(0).getReg()) 1772 .add(SrcMI->getOperand(1)) 1773 .add(MOpSHMI); 1774 (void)NewInstr; 1775 1776 LLVM_DEBUG(dbgs() << "TO: "); 1777 LLVM_DEBUG(NewInstr->dump()); 1778 ++NumEXTSWAndSLDICombined; 1779 ToErase = &MI; 1780 // SrcMI, which is extsw, is of no use now, erase it. 1781 SrcMI->eraseFromParent(); 1782 return true; 1783 } 1784 1785 } // end default namespace 1786 1787 INITIALIZE_PASS_BEGIN(PPCMIPeephole, DEBUG_TYPE, 1788 "PowerPC MI Peephole Optimization", false, false) 1789 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) 1790 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 1791 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree) 1792 INITIALIZE_PASS_END(PPCMIPeephole, DEBUG_TYPE, 1793 "PowerPC MI Peephole Optimization", false, false) 1794 1795 char PPCMIPeephole::ID = 0; 1796 FunctionPass* 1797 llvm::createPPCMIPeepholePass() { return new PPCMIPeephole(); } 1798