1 //===-------------- PPCMIPeephole.cpp - MI Peephole Cleanups -------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===---------------------------------------------------------------------===// 9 // 10 // This pass performs peephole optimizations to clean up ugly code 11 // sequences at the MachineInstruction layer. It runs at the end of 12 // the SSA phases, following VSX swap removal. A pass of dead code 13 // elimination follows this one for quick clean-up of any dead 14 // instructions introduced here. Although we could do this as callbacks 15 // from the generic peephole pass, this would have a couple of bad 16 // effects: it might remove optimization opportunities for VSX swap 17 // removal, and it would miss cleanups made possible following VSX 18 // swap removal. 19 // 20 //===---------------------------------------------------------------------===// 21 22 #include "PPC.h" 23 #include "PPCInstrBuilder.h" 24 #include "PPCInstrInfo.h" 25 #include "PPCTargetMachine.h" 26 #include "llvm/ADT/Statistic.h" 27 #include "llvm/CodeGen/MachineDominators.h" 28 #include "llvm/CodeGen/MachineFunctionPass.h" 29 #include "llvm/CodeGen/MachineInstrBuilder.h" 30 #include "llvm/CodeGen/MachineRegisterInfo.h" 31 #include "llvm/Support/Debug.h" 32 #include "MCTargetDesc/PPCPredicates.h" 33 34 using namespace llvm; 35 36 #define DEBUG_TYPE "ppc-mi-peepholes" 37 38 STATISTIC(NumOptADDLIs, "Number of optimized ADD instruction fed by LI"); 39 40 namespace llvm { 41 void initializePPCMIPeepholePass(PassRegistry&); 42 } 43 44 namespace { 45 46 struct PPCMIPeephole : public MachineFunctionPass { 47 48 static char ID; 49 const PPCInstrInfo *TII; 50 MachineFunction *MF; 51 MachineRegisterInfo *MRI; 52 53 PPCMIPeephole() : MachineFunctionPass(ID) { 54 initializePPCMIPeepholePass(*PassRegistry::getPassRegistry()); 55 } 56 57 private: 58 MachineDominatorTree *MDT; 59 60 // Initialize class variables. 61 void initialize(MachineFunction &MFParm); 62 63 // Perform peepholes. 64 bool simplifyCode(void); 65 66 // Perform peepholes. 67 bool eliminateRedundantCompare(void); 68 69 // Find the "true" register represented by SrcReg (following chains 70 // of copies and subreg_to_reg operations). 71 unsigned lookThruCopyLike(unsigned SrcReg); 72 73 public: 74 75 void getAnalysisUsage(AnalysisUsage &AU) const override { 76 AU.addRequired<MachineDominatorTree>(); 77 AU.addPreserved<MachineDominatorTree>(); 78 MachineFunctionPass::getAnalysisUsage(AU); 79 } 80 81 // Main entry point for this pass. 82 bool runOnMachineFunction(MachineFunction &MF) override { 83 if (skipFunction(*MF.getFunction())) 84 return false; 85 initialize(MF); 86 return simplifyCode(); 87 } 88 }; 89 90 // Initialize class variables. 91 void PPCMIPeephole::initialize(MachineFunction &MFParm) { 92 MF = &MFParm; 93 MRI = &MF->getRegInfo(); 94 MDT = &getAnalysis<MachineDominatorTree>(); 95 TII = MF->getSubtarget<PPCSubtarget>().getInstrInfo(); 96 DEBUG(dbgs() << "*** PowerPC MI peephole pass ***\n\n"); 97 DEBUG(MF->dump()); 98 } 99 100 static MachineInstr *getVRegDefOrNull(MachineOperand *Op, 101 MachineRegisterInfo *MRI) { 102 assert(Op && "Invalid Operand!"); 103 if (!Op->isReg()) 104 return nullptr; 105 106 unsigned Reg = Op->getReg(); 107 if (!TargetRegisterInfo::isVirtualRegister(Reg)) 108 return nullptr; 109 110 return MRI->getVRegDef(Reg); 111 } 112 113 // Perform peephole optimizations. 114 bool PPCMIPeephole::simplifyCode(void) { 115 bool Simplified = false; 116 MachineInstr* ToErase = nullptr; 117 118 for (MachineBasicBlock &MBB : *MF) { 119 for (MachineInstr &MI : MBB) { 120 121 // If the previous instruction was marked for elimination, 122 // remove it now. 123 if (ToErase) { 124 ToErase->eraseFromParent(); 125 ToErase = nullptr; 126 } 127 128 // Ignore debug instructions. 129 if (MI.isDebugValue()) 130 continue; 131 132 // Per-opcode peepholes. 133 switch (MI.getOpcode()) { 134 135 default: 136 break; 137 138 case PPC::XXPERMDI: { 139 // Perform simplifications of 2x64 vector swaps and splats. 140 // A swap is identified by an immediate value of 2, and a splat 141 // is identified by an immediate value of 0 or 3. 142 int Immed = MI.getOperand(3).getImm(); 143 144 if (Immed != 1) { 145 146 // For each of these simplifications, we need the two source 147 // regs to match. Unfortunately, MachineCSE ignores COPY and 148 // SUBREG_TO_REG, so for example we can see 149 // XXPERMDI t, SUBREG_TO_REG(s), SUBREG_TO_REG(s), immed. 150 // We have to look through chains of COPY and SUBREG_TO_REG 151 // to find the real source values for comparison. 152 unsigned TrueReg1 = lookThruCopyLike(MI.getOperand(1).getReg()); 153 unsigned TrueReg2 = lookThruCopyLike(MI.getOperand(2).getReg()); 154 155 if (TrueReg1 == TrueReg2 156 && TargetRegisterInfo::isVirtualRegister(TrueReg1)) { 157 MachineInstr *DefMI = MRI->getVRegDef(TrueReg1); 158 unsigned DefOpc = DefMI ? DefMI->getOpcode() : 0; 159 160 // If this is a splat fed by a splatting load, the splat is 161 // redundant. Replace with a copy. This doesn't happen directly due 162 // to code in PPCDAGToDAGISel.cpp, but it can happen when converting 163 // a load of a double to a vector of 64-bit integers. 164 auto isConversionOfLoadAndSplat = [=]() -> bool { 165 if (DefOpc != PPC::XVCVDPSXDS && DefOpc != PPC::XVCVDPUXDS) 166 return false; 167 unsigned DefReg = lookThruCopyLike(DefMI->getOperand(1).getReg()); 168 if (TargetRegisterInfo::isVirtualRegister(DefReg)) { 169 MachineInstr *LoadMI = MRI->getVRegDef(DefReg); 170 if (LoadMI && LoadMI->getOpcode() == PPC::LXVDSX) 171 return true; 172 } 173 return false; 174 }; 175 if (DefMI && (Immed == 0 || Immed == 3)) { 176 if (DefOpc == PPC::LXVDSX || isConversionOfLoadAndSplat()) { 177 DEBUG(dbgs() 178 << "Optimizing load-and-splat/splat " 179 "to load-and-splat/copy: "); 180 DEBUG(MI.dump()); 181 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 182 MI.getOperand(0).getReg()) 183 .add(MI.getOperand(1)); 184 ToErase = &MI; 185 Simplified = true; 186 } 187 } 188 189 // If this is a splat or a swap fed by another splat, we 190 // can replace it with a copy. 191 if (DefOpc == PPC::XXPERMDI) { 192 unsigned FeedImmed = DefMI->getOperand(3).getImm(); 193 unsigned FeedReg1 194 = lookThruCopyLike(DefMI->getOperand(1).getReg()); 195 unsigned FeedReg2 196 = lookThruCopyLike(DefMI->getOperand(2).getReg()); 197 198 if ((FeedImmed == 0 || FeedImmed == 3) && FeedReg1 == FeedReg2) { 199 DEBUG(dbgs() 200 << "Optimizing splat/swap or splat/splat " 201 "to splat/copy: "); 202 DEBUG(MI.dump()); 203 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 204 MI.getOperand(0).getReg()) 205 .add(MI.getOperand(1)); 206 ToErase = &MI; 207 Simplified = true; 208 } 209 210 // If this is a splat fed by a swap, we can simplify modify 211 // the splat to splat the other value from the swap's input 212 // parameter. 213 else if ((Immed == 0 || Immed == 3) 214 && FeedImmed == 2 && FeedReg1 == FeedReg2) { 215 DEBUG(dbgs() << "Optimizing swap/splat => splat: "); 216 DEBUG(MI.dump()); 217 MI.getOperand(1).setReg(DefMI->getOperand(1).getReg()); 218 MI.getOperand(2).setReg(DefMI->getOperand(2).getReg()); 219 MI.getOperand(3).setImm(3 - Immed); 220 Simplified = true; 221 } 222 223 // If this is a swap fed by a swap, we can replace it 224 // with a copy from the first swap's input. 225 else if (Immed == 2 && FeedImmed == 2 && FeedReg1 == FeedReg2) { 226 DEBUG(dbgs() << "Optimizing swap/swap => copy: "); 227 DEBUG(MI.dump()); 228 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 229 MI.getOperand(0).getReg()) 230 .add(DefMI->getOperand(1)); 231 ToErase = &MI; 232 Simplified = true; 233 } 234 } else if ((Immed == 0 || Immed == 3) && DefOpc == PPC::XXPERMDIs && 235 (DefMI->getOperand(2).getImm() == 0 || 236 DefMI->getOperand(2).getImm() == 3)) { 237 // Splat fed by another splat - switch the output of the first 238 // and remove the second. 239 DefMI->getOperand(0).setReg(MI.getOperand(0).getReg()); 240 ToErase = &MI; 241 Simplified = true; 242 DEBUG(dbgs() << "Removing redundant splat: "); 243 DEBUG(MI.dump()); 244 } 245 } 246 } 247 break; 248 } 249 case PPC::VSPLTB: 250 case PPC::VSPLTH: 251 case PPC::XXSPLTW: { 252 unsigned MyOpcode = MI.getOpcode(); 253 unsigned OpNo = MyOpcode == PPC::XXSPLTW ? 1 : 2; 254 unsigned TrueReg = lookThruCopyLike(MI.getOperand(OpNo).getReg()); 255 if (!TargetRegisterInfo::isVirtualRegister(TrueReg)) 256 break; 257 MachineInstr *DefMI = MRI->getVRegDef(TrueReg); 258 if (!DefMI) 259 break; 260 unsigned DefOpcode = DefMI->getOpcode(); 261 auto isConvertOfSplat = [=]() -> bool { 262 if (DefOpcode != PPC::XVCVSPSXWS && DefOpcode != PPC::XVCVSPUXWS) 263 return false; 264 unsigned ConvReg = DefMI->getOperand(1).getReg(); 265 if (!TargetRegisterInfo::isVirtualRegister(ConvReg)) 266 return false; 267 MachineInstr *Splt = MRI->getVRegDef(ConvReg); 268 return Splt && (Splt->getOpcode() == PPC::LXVWSX || 269 Splt->getOpcode() == PPC::XXSPLTW); 270 }; 271 bool AlreadySplat = (MyOpcode == DefOpcode) || 272 (MyOpcode == PPC::VSPLTB && DefOpcode == PPC::VSPLTBs) || 273 (MyOpcode == PPC::VSPLTH && DefOpcode == PPC::VSPLTHs) || 274 (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::XXSPLTWs) || 275 (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::LXVWSX) || 276 (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::MTVSRWS)|| 277 (MyOpcode == PPC::XXSPLTW && isConvertOfSplat()); 278 // If the instruction[s] that feed this splat have already splat 279 // the value, this splat is redundant. 280 if (AlreadySplat) { 281 DEBUG(dbgs() << "Changing redundant splat to a copy: "); 282 DEBUG(MI.dump()); 283 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 284 MI.getOperand(0).getReg()) 285 .add(MI.getOperand(OpNo)); 286 ToErase = &MI; 287 Simplified = true; 288 } 289 // Splat fed by a shift. Usually when we align value to splat into 290 // vector element zero. 291 if (DefOpcode == PPC::XXSLDWI) { 292 unsigned ShiftRes = DefMI->getOperand(0).getReg(); 293 unsigned ShiftOp1 = DefMI->getOperand(1).getReg(); 294 unsigned ShiftOp2 = DefMI->getOperand(2).getReg(); 295 unsigned ShiftImm = DefMI->getOperand(3).getImm(); 296 unsigned SplatImm = MI.getOperand(2).getImm(); 297 if (ShiftOp1 == ShiftOp2) { 298 unsigned NewElem = (SplatImm + ShiftImm) & 0x3; 299 if (MRI->hasOneNonDBGUse(ShiftRes)) { 300 DEBUG(dbgs() << "Removing redundant shift: "); 301 DEBUG(DefMI->dump()); 302 ToErase = DefMI; 303 } 304 Simplified = true; 305 DEBUG(dbgs() << "Changing splat immediate from " << SplatImm << 306 " to " << NewElem << " in instruction: "); 307 DEBUG(MI.dump()); 308 MI.getOperand(1).setReg(ShiftOp1); 309 MI.getOperand(2).setImm(NewElem); 310 } 311 } 312 break; 313 } 314 case PPC::XVCVDPSP: { 315 // If this is a DP->SP conversion fed by an FRSP, the FRSP is redundant. 316 unsigned TrueReg = lookThruCopyLike(MI.getOperand(1).getReg()); 317 if (!TargetRegisterInfo::isVirtualRegister(TrueReg)) 318 break; 319 MachineInstr *DefMI = MRI->getVRegDef(TrueReg); 320 321 // This can occur when building a vector of single precision or integer 322 // values. 323 if (DefMI && DefMI->getOpcode() == PPC::XXPERMDI) { 324 unsigned DefsReg1 = lookThruCopyLike(DefMI->getOperand(1).getReg()); 325 unsigned DefsReg2 = lookThruCopyLike(DefMI->getOperand(2).getReg()); 326 if (!TargetRegisterInfo::isVirtualRegister(DefsReg1) || 327 !TargetRegisterInfo::isVirtualRegister(DefsReg2)) 328 break; 329 MachineInstr *P1 = MRI->getVRegDef(DefsReg1); 330 MachineInstr *P2 = MRI->getVRegDef(DefsReg2); 331 332 if (!P1 || !P2) 333 break; 334 335 // Remove the passed FRSP instruction if it only feeds this MI and 336 // set any uses of that FRSP (in this MI) to the source of the FRSP. 337 auto removeFRSPIfPossible = [&](MachineInstr *RoundInstr) { 338 if (RoundInstr->getOpcode() == PPC::FRSP && 339 MRI->hasOneNonDBGUse(RoundInstr->getOperand(0).getReg())) { 340 Simplified = true; 341 unsigned ConvReg1 = RoundInstr->getOperand(1).getReg(); 342 unsigned FRSPDefines = RoundInstr->getOperand(0).getReg(); 343 MachineInstr &Use = *(MRI->use_instr_begin(FRSPDefines)); 344 for (int i = 0, e = Use.getNumOperands(); i < e; ++i) 345 if (Use.getOperand(i).isReg() && 346 Use.getOperand(i).getReg() == FRSPDefines) 347 Use.getOperand(i).setReg(ConvReg1); 348 DEBUG(dbgs() << "Removing redundant FRSP:\n"); 349 DEBUG(RoundInstr->dump()); 350 DEBUG(dbgs() << "As it feeds instruction:\n"); 351 DEBUG(MI.dump()); 352 DEBUG(dbgs() << "Through instruction:\n"); 353 DEBUG(DefMI->dump()); 354 RoundInstr->eraseFromParent(); 355 } 356 }; 357 358 // If the input to XVCVDPSP is a vector that was built (even 359 // partially) out of FRSP's, the FRSP(s) can safely be removed 360 // since this instruction performs the same operation. 361 if (P1 != P2) { 362 removeFRSPIfPossible(P1); 363 removeFRSPIfPossible(P2); 364 break; 365 } 366 removeFRSPIfPossible(P1); 367 } 368 break; 369 } 370 371 // TODO: Any instruction that has an immediate form fed only by a PHI 372 // whose operands are all load immediate can be folded away. We currently 373 // do this for ADD instructions, but should expand it to arithmetic and 374 // binary instructions with immediate forms in the future. 375 case PPC::ADD4: 376 case PPC::ADD8: { 377 auto isSingleUsePHI = [&](MachineOperand *PhiOp) { 378 assert(PhiOp && "Invalid Operand!"); 379 MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI); 380 381 return DefPhiMI && (DefPhiMI->getOpcode() == PPC::PHI) && 382 MRI->hasOneNonDBGUse(DefPhiMI->getOperand(0).getReg()); 383 }; 384 385 auto dominatesAllSingleUseLIs = [&](MachineOperand *DominatorOp, 386 MachineOperand *PhiOp) { 387 assert(PhiOp && "Invalid Operand!"); 388 assert(DominatorOp && "Invalid Operand!"); 389 MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI); 390 MachineInstr *DefDomMI = getVRegDefOrNull(DominatorOp, MRI); 391 392 // Note: the vregs only show up at odd indices position of PHI Node, 393 // the even indices position save the BB info. 394 for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) { 395 MachineInstr *LiMI = 396 getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI); 397 if (!LiMI || !MRI->hasOneNonDBGUse(LiMI->getOperand(0).getReg()) || 398 !MDT->dominates(DefDomMI, LiMI) || 399 (LiMI->getOpcode() != PPC::LI && LiMI->getOpcode() != PPC::LI8)) 400 return false; 401 } 402 403 return true; 404 }; 405 406 MachineOperand Op1 = MI.getOperand(1); 407 MachineOperand Op2 = MI.getOperand(2); 408 if (isSingleUsePHI(&Op2) && dominatesAllSingleUseLIs(&Op1, &Op2)) 409 std::swap(Op1, Op2); 410 else if (!isSingleUsePHI(&Op1) || !dominatesAllSingleUseLIs(&Op2, &Op1)) 411 break; // We don't have an ADD fed by LI's that can be transformed 412 413 // Now we know that Op1 is the PHI node and Op2 is the dominator 414 unsigned DominatorReg = Op2.getReg(); 415 416 const TargetRegisterClass *TRC = MI.getOpcode() == PPC::ADD8 417 ? &PPC::G8RC_and_G8RC_NOX0RegClass 418 : &PPC::GPRC_and_GPRC_NOR0RegClass; 419 MRI->setRegClass(DominatorReg, TRC); 420 421 // replace LIs with ADDIs 422 MachineInstr *DefPhiMI = getVRegDefOrNull(&Op1, MRI); 423 for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) { 424 MachineInstr *LiMI = getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI); 425 DEBUG(dbgs() << "Optimizing LI to ADDI: "); 426 DEBUG(LiMI->dump()); 427 428 // There could be repeated registers in the PHI, e.g: %vreg1<def> = 429 // PHI %vreg6, <BB#2>, %vreg8, <BB#3>, %vreg8, <BB#6>; So if we've 430 // already replaced the def instruction, skip. 431 if (LiMI->getOpcode() == PPC::ADDI || LiMI->getOpcode() == PPC::ADDI8) 432 continue; 433 434 assert((LiMI->getOpcode() == PPC::LI || 435 LiMI->getOpcode() == PPC::LI8) && 436 "Invalid Opcode!"); 437 auto LiImm = LiMI->getOperand(1).getImm(); // save the imm of LI 438 LiMI->RemoveOperand(1); // remove the imm of LI 439 LiMI->setDesc(TII->get(LiMI->getOpcode() == PPC::LI ? PPC::ADDI 440 : PPC::ADDI8)); 441 MachineInstrBuilder(*LiMI->getParent()->getParent(), *LiMI) 442 .addReg(DominatorReg) 443 .addImm(LiImm); // restore the imm of LI 444 DEBUG(LiMI->dump()); 445 } 446 447 // Replace ADD with COPY 448 DEBUG(dbgs() << "Optimizing ADD to COPY: "); 449 DEBUG(MI.dump()); 450 BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), 451 MI.getOperand(0).getReg()) 452 .add(Op1); 453 ToErase = &MI; 454 Simplified = true; 455 NumOptADDLIs++; 456 break; 457 } 458 } 459 } 460 461 // If the last instruction was marked for elimination, 462 // remove it now. 463 if (ToErase) { 464 ToErase->eraseFromParent(); 465 ToErase = nullptr; 466 } 467 } 468 469 // We try to eliminate redundant compare instruction. 470 Simplified |= eliminateRedundantCompare(); 471 472 return Simplified; 473 } 474 475 // helper functions for eliminateRedundantCompare 476 static bool isEqOrNe(MachineInstr *BI) { 477 PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); 478 unsigned PredCond = PPC::getPredicateCondition(Pred); 479 return (PredCond == PPC::PRED_EQ || PredCond == PPC::PRED_NE); 480 } 481 482 static bool isSupportedCmpOp(unsigned opCode) { 483 return (opCode == PPC::CMPLD || opCode == PPC::CMPD || 484 opCode == PPC::CMPLW || opCode == PPC::CMPW || 485 opCode == PPC::CMPLDI || opCode == PPC::CMPDI || 486 opCode == PPC::CMPLWI || opCode == PPC::CMPWI); 487 } 488 489 static bool is64bitCmpOp(unsigned opCode) { 490 return (opCode == PPC::CMPLD || opCode == PPC::CMPD || 491 opCode == PPC::CMPLDI || opCode == PPC::CMPDI); 492 } 493 494 static bool isSignedCmpOp(unsigned opCode) { 495 return (opCode == PPC::CMPD || opCode == PPC::CMPW || 496 opCode == PPC::CMPDI || opCode == PPC::CMPWI); 497 } 498 499 static unsigned getSignedCmpOpCode(unsigned opCode) { 500 if (opCode == PPC::CMPLD) return PPC::CMPD; 501 if (opCode == PPC::CMPLW) return PPC::CMPW; 502 if (opCode == PPC::CMPLDI) return PPC::CMPDI; 503 if (opCode == PPC::CMPLWI) return PPC::CMPWI; 504 return opCode; 505 } 506 507 // We can decrement immediate x in (GE x) by changing it to (GT x-1) or 508 // (LT x) to (LE x-1) 509 static unsigned getPredicateToDecImm(MachineInstr *BI, MachineInstr *CMPI) { 510 uint64_t Imm = CMPI->getOperand(2).getImm(); 511 bool SignedCmp = isSignedCmpOp(CMPI->getOpcode()); 512 if ((!SignedCmp && Imm == 0) || (SignedCmp && Imm == 0x8000)) 513 return 0; 514 515 PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); 516 unsigned PredCond = PPC::getPredicateCondition(Pred); 517 unsigned PredHint = PPC::getPredicateHint(Pred); 518 if (PredCond == PPC::PRED_GE) 519 return PPC::getPredicate(PPC::PRED_GT, PredHint); 520 if (PredCond == PPC::PRED_LT) 521 return PPC::getPredicate(PPC::PRED_LE, PredHint); 522 523 return 0; 524 } 525 526 // We can increment immediate x in (GT x) by changing it to (GE x+1) or 527 // (LE x) to (LT x+1) 528 static unsigned getPredicateToIncImm(MachineInstr *BI, MachineInstr *CMPI) { 529 uint64_t Imm = CMPI->getOperand(2).getImm(); 530 bool SignedCmp = isSignedCmpOp(CMPI->getOpcode()); 531 if ((!SignedCmp && Imm == 0xFFFF) || (SignedCmp && Imm == 0x7FFF)) 532 return 0; 533 534 PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); 535 unsigned PredCond = PPC::getPredicateCondition(Pred); 536 unsigned PredHint = PPC::getPredicateHint(Pred); 537 if (PredCond == PPC::PRED_GT) 538 return PPC::getPredicate(PPC::PRED_GE, PredHint); 539 if (PredCond == PPC::PRED_LE) 540 return PPC::getPredicate(PPC::PRED_LT, PredHint); 541 542 return 0; 543 } 544 545 // This takes a Phi node and returns a register value for the spefied BB. 546 static unsigned getIncomingRegForBlock(MachineInstr *Phi, 547 MachineBasicBlock *MBB) { 548 for (unsigned I = 2, E = Phi->getNumOperands() + 1; I != E; I += 2) { 549 MachineOperand &MO = Phi->getOperand(I); 550 if (MO.getMBB() == MBB) 551 return Phi->getOperand(I-1).getReg(); 552 } 553 llvm_unreachable("invalid src basic block for this Phi node\n"); 554 return 0; 555 } 556 557 // This function tracks the source of the register through register copy. 558 // If BB1 and BB2 are non-NULL, we also track PHI instruction in BB2 559 // assuming that the control comes from BB1 into BB2. 560 static unsigned getSrcVReg(unsigned Reg, MachineBasicBlock *BB1, 561 MachineBasicBlock *BB2, MachineRegisterInfo *MRI) { 562 unsigned SrcReg = Reg; 563 while (1) { 564 unsigned NextReg = SrcReg; 565 MachineInstr *Inst = MRI->getVRegDef(SrcReg); 566 if (BB1 && Inst->getOpcode() == PPC::PHI && Inst->getParent() == BB2) { 567 NextReg = getIncomingRegForBlock(Inst, BB1); 568 // We track through PHI only once to avoid infinite loop. 569 BB1 = nullptr; 570 } 571 else if (Inst->isFullCopy()) 572 NextReg = Inst->getOperand(1).getReg(); 573 if (NextReg == SrcReg || !TargetRegisterInfo::isVirtualRegister(NextReg)) 574 break; 575 SrcReg = NextReg; 576 } 577 return SrcReg; 578 } 579 580 static bool eligibleForCompareElimination(MachineBasicBlock &MBB, 581 MachineBasicBlock *&PredMBB, 582 MachineBasicBlock *&MBBtoMoveCmp, 583 MachineRegisterInfo *MRI) { 584 585 auto isEligibleBB = [&](MachineBasicBlock &BB) { 586 auto BII = BB.getFirstInstrTerminator(); 587 // We optimize BBs ending with a conditional branch. 588 // We check only for BCC here, not BCCLR, because BCCLR 589 // will be formed only later in the pipeline. 590 if (BB.succ_size() == 2 && 591 BII != BB.instr_end() && 592 (*BII).getOpcode() == PPC::BCC && 593 (*BII).getOperand(1).isReg()) { 594 // We optimize only if the condition code is used only by one BCC. 595 unsigned CndReg = (*BII).getOperand(1).getReg(); 596 if (!TargetRegisterInfo::isVirtualRegister(CndReg) || 597 !MRI->hasOneNonDBGUse(CndReg)) 598 return false; 599 600 // We skip this BB if a physical register is used in comparison. 601 MachineInstr *CMPI = MRI->getVRegDef(CndReg); 602 for (MachineOperand &MO : CMPI->operands()) 603 if (MO.isReg() && !TargetRegisterInfo::isVirtualRegister(MO.getReg())) 604 return false; 605 606 return true; 607 } 608 return false; 609 }; 610 611 // If this BB has more than one successor, we can create a new BB and 612 // move the compare instruction in the new BB. 613 // So far, we do not move compare instruction to a BB having multiple 614 // successors to avoid potentially increasing code size. 615 auto isEligibleForMoveCmp = [](MachineBasicBlock &BB) { 616 return BB.succ_size() == 1; 617 }; 618 619 if (!isEligibleBB(MBB)) 620 return false; 621 622 unsigned NumPredBBs = MBB.pred_size(); 623 if (NumPredBBs == 1) { 624 MachineBasicBlock *TmpMBB = *MBB.pred_begin(); 625 if (isEligibleBB(*TmpMBB)) { 626 PredMBB = TmpMBB; 627 MBBtoMoveCmp = nullptr; 628 return true; 629 } 630 } 631 else if (NumPredBBs == 2) { 632 // We check for partially redundant case. 633 // So far, we support cases with only two predecessors 634 // to avoid increasing the number of instructions. 635 MachineBasicBlock::pred_iterator PI = MBB.pred_begin(); 636 MachineBasicBlock *Pred1MBB = *PI; 637 MachineBasicBlock *Pred2MBB = *(PI+1); 638 639 if (isEligibleBB(*Pred1MBB) && isEligibleForMoveCmp(*Pred2MBB)) { 640 // We assume Pred1MBB is the BB containing the compare to be merged and 641 // Pred2MBB is the BB to which we will append a compare instruction. 642 // Hence we can proceed as is. 643 } 644 else if (isEligibleBB(*Pred2MBB) && isEligibleForMoveCmp(*Pred1MBB)) { 645 // We need to swap Pred1MBB and Pred2MBB to canonicalize. 646 std::swap(Pred1MBB, Pred2MBB); 647 } 648 else return false; 649 650 // Here, Pred2MBB is the BB to which we need to append a compare inst. 651 // We cannot move the compare instruction if operands are not available 652 // in Pred2MBB (i.e. defined in MBB by an instruction other than PHI). 653 MachineInstr *BI = &*MBB.getFirstInstrTerminator(); 654 MachineInstr *CMPI = MRI->getVRegDef(BI->getOperand(1).getReg()); 655 for (int I = 1; I <= 2; I++) 656 if (CMPI->getOperand(I).isReg()) { 657 MachineInstr *Inst = MRI->getVRegDef(CMPI->getOperand(I).getReg()); 658 if (Inst->getParent() == &MBB && Inst->getOpcode() != PPC::PHI) 659 return false; 660 } 661 662 PredMBB = Pred1MBB; 663 MBBtoMoveCmp = Pred2MBB; 664 return true; 665 } 666 667 return false; 668 } 669 670 // If multiple conditional branches are executed based on the (essentially) 671 // same comparison, we merge compare instructions into one and make multiple 672 // conditional branches on this comparison. 673 // For example, 674 // if (a == 0) { ... } 675 // else if (a < 0) { ... } 676 // can be executed by one compare and two conditional branches instead of 677 // two pairs of a compare and a conditional branch. 678 // 679 // This method merges two compare instructions in two MBBs and modifies the 680 // compare and conditional branch instructions if needed. 681 // For the above example, the input for this pass looks like: 682 // cmplwi r3, 0 683 // beq 0, .LBB0_3 684 // cmpwi r3, -1 685 // bgt 0, .LBB0_4 686 // So, before merging two compares, we need to modify these instructions as 687 // cmpwi r3, 0 ; cmplwi and cmpwi yield same result for beq 688 // beq 0, .LBB0_3 689 // cmpwi r3, 0 ; greather than -1 means greater or equal to 0 690 // bge 0, .LBB0_4 691 692 bool PPCMIPeephole::eliminateRedundantCompare(void) { 693 bool Simplified = false; 694 695 for (MachineBasicBlock &MBB2 : *MF) { 696 MachineBasicBlock *MBB1 = nullptr, *MBBtoMoveCmp = nullptr; 697 698 // For fully redundant case, we select two basic blocks MBB1 and MBB2 699 // as an optimization target if 700 // - both MBBs end with a conditional branch, 701 // - MBB1 is the only predecessor of MBB2, and 702 // - compare does not take a physical register as a operand in both MBBs. 703 // In this case, eligibleForCompareElimination sets MBBtoMoveCmp nullptr. 704 // 705 // As partially redundant case, we additionally handle if MBB2 has one 706 // additional predecessor, which has only one successor (MBB2). 707 // In this case, we move the compre instruction originally in MBB2 into 708 // MBBtoMoveCmp. This partially redundant case is typically appear by 709 // compiling a while loop; here, MBBtoMoveCmp is the loop preheader. 710 // 711 // Overview of CFG of related basic blocks 712 // Fully redundant case Partially redundant case 713 // -------- ---------------- -------- 714 // | MBB1 | (w/ 2 succ) | MBBtoMoveCmp | | MBB1 | (w/ 2 succ) 715 // -------- ---------------- -------- 716 // | \ (w/ 1 succ) \ | \ 717 // | \ \ | \ 718 // | \ | 719 // -------- -------- 720 // | MBB2 | (w/ 1 pred | MBB2 | (w/ 2 pred 721 // -------- and 2 succ) -------- and 2 succ) 722 // | \ | \ 723 // | \ | \ 724 // 725 if (!eligibleForCompareElimination(MBB2, MBB1, MBBtoMoveCmp, MRI)) 726 continue; 727 728 MachineInstr *BI1 = &*MBB1->getFirstInstrTerminator(); 729 MachineInstr *CMPI1 = MRI->getVRegDef(BI1->getOperand(1).getReg()); 730 731 MachineInstr *BI2 = &*MBB2.getFirstInstrTerminator(); 732 MachineInstr *CMPI2 = MRI->getVRegDef(BI2->getOperand(1).getReg()); 733 bool IsPartiallyRedundant = (MBBtoMoveCmp != nullptr); 734 735 // We cannot optimize an unsupported compare opcode or 736 // a mix of 32-bit and 64-bit comaprisons 737 if (!isSupportedCmpOp(CMPI1->getOpcode()) || 738 !isSupportedCmpOp(CMPI2->getOpcode()) || 739 is64bitCmpOp(CMPI1->getOpcode()) != is64bitCmpOp(CMPI2->getOpcode())) 740 continue; 741 742 unsigned NewOpCode = 0; 743 unsigned NewPredicate1 = 0, NewPredicate2 = 0; 744 int16_t Imm1 = 0, NewImm1 = 0, Imm2 = 0, NewImm2 = 0; 745 bool SwapOperands = false; 746 747 if (CMPI1->getOpcode() != CMPI2->getOpcode()) { 748 // Typically, unsigned comparison is used for equality check, but 749 // we replace it with a signed comparison if the comparison 750 // to be merged is a signed comparison. 751 // In other cases of opcode mismatch, we cannot optimize this. 752 if (isEqOrNe(BI2) && 753 CMPI1->getOpcode() == getSignedCmpOpCode(CMPI2->getOpcode())) 754 NewOpCode = CMPI1->getOpcode(); 755 else if (isEqOrNe(BI1) && 756 getSignedCmpOpCode(CMPI1->getOpcode()) == CMPI2->getOpcode()) 757 NewOpCode = CMPI2->getOpcode(); 758 else continue; 759 } 760 761 if (CMPI1->getOperand(2).isReg() && CMPI2->getOperand(2).isReg()) { 762 // In case of comparisons between two registers, these two registers 763 // must be same to merge two comparisons. 764 unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(), 765 nullptr, nullptr, MRI); 766 unsigned Cmp1Operand2 = getSrcVReg(CMPI1->getOperand(2).getReg(), 767 nullptr, nullptr, MRI); 768 unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(), 769 MBB1, &MBB2, MRI); 770 unsigned Cmp2Operand2 = getSrcVReg(CMPI2->getOperand(2).getReg(), 771 MBB1, &MBB2, MRI); 772 773 if (Cmp1Operand1 == Cmp2Operand1 && Cmp1Operand2 == Cmp2Operand2) { 774 // Same pair of registers in the same order; ready to merge as is. 775 } 776 else if (Cmp1Operand1 == Cmp2Operand2 && Cmp1Operand2 == Cmp2Operand1) { 777 // Same pair of registers in different order. 778 // We reverse the predicate to merge compare instructions. 779 PPC::Predicate Pred = (PPC::Predicate)BI2->getOperand(0).getImm(); 780 NewPredicate2 = (unsigned)PPC::getSwappedPredicate(Pred); 781 // In case of partial redundancy, we need to swap operands 782 // in another compare instruction. 783 SwapOperands = true; 784 } 785 else continue; 786 } 787 else if (CMPI1->getOperand(2).isImm() && CMPI2->getOperand(2).isImm()){ 788 // In case of comparisons between a register and an immediate, 789 // the operand register must be same for two compare instructions. 790 unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(), 791 nullptr, nullptr, MRI); 792 unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(), 793 MBB1, &MBB2, MRI); 794 if (Cmp1Operand1 != Cmp2Operand1) 795 continue; 796 797 NewImm1 = Imm1 = (int16_t)CMPI1->getOperand(2).getImm(); 798 NewImm2 = Imm2 = (int16_t)CMPI2->getOperand(2).getImm(); 799 800 // If immediate are not same, we try to adjust by changing predicate; 801 // e.g. GT imm means GE (imm+1). 802 if (Imm1 != Imm2 && (!isEqOrNe(BI2) || !isEqOrNe(BI1))) { 803 int Diff = Imm1 - Imm2; 804 if (Diff < -2 || Diff > 2) 805 continue; 806 807 unsigned PredToInc1 = getPredicateToIncImm(BI1, CMPI1); 808 unsigned PredToDec1 = getPredicateToDecImm(BI1, CMPI1); 809 unsigned PredToInc2 = getPredicateToIncImm(BI2, CMPI2); 810 unsigned PredToDec2 = getPredicateToDecImm(BI2, CMPI2); 811 if (Diff == 2) { 812 if (PredToInc2 && PredToDec1) { 813 NewPredicate2 = PredToInc2; 814 NewPredicate1 = PredToDec1; 815 NewImm2++; 816 NewImm1--; 817 } 818 } 819 else if (Diff == 1) { 820 if (PredToInc2) { 821 NewImm2++; 822 NewPredicate2 = PredToInc2; 823 } 824 else if (PredToDec1) { 825 NewImm1--; 826 NewPredicate1 = PredToDec1; 827 } 828 } 829 else if (Diff == -1) { 830 if (PredToDec2) { 831 NewImm2--; 832 NewPredicate2 = PredToDec2; 833 } 834 else if (PredToInc1) { 835 NewImm1++; 836 NewPredicate1 = PredToInc1; 837 } 838 } 839 else if (Diff == -2) { 840 if (PredToDec2 && PredToInc1) { 841 NewPredicate2 = PredToDec2; 842 NewPredicate1 = PredToInc1; 843 NewImm2--; 844 NewImm1++; 845 } 846 } 847 } 848 849 // We cannnot merge two compares if the immediates are not same. 850 if (NewImm2 != NewImm1) 851 continue; 852 } 853 854 DEBUG(dbgs() << "Optimize two pairs of compare and branch:\n"); 855 DEBUG(CMPI1->dump()); 856 DEBUG(BI1->dump()); 857 DEBUG(CMPI2->dump()); 858 DEBUG(BI2->dump()); 859 860 // We adjust opcode, predicates and immediate as we determined above. 861 if (NewOpCode != 0 && NewOpCode != CMPI1->getOpcode()) { 862 CMPI1->setDesc(TII->get(NewOpCode)); 863 } 864 if (NewPredicate1) { 865 BI1->getOperand(0).setImm(NewPredicate1); 866 } 867 if (NewPredicate2) { 868 BI2->getOperand(0).setImm(NewPredicate2); 869 } 870 if (NewImm1 != Imm1) { 871 CMPI1->getOperand(2).setImm(NewImm1); 872 } 873 874 if (IsPartiallyRedundant) { 875 // We touch up the compare instruction in MBB2 and move it to 876 // a previous BB to handle partially redundant case. 877 if (SwapOperands) { 878 unsigned Op1 = CMPI2->getOperand(1).getReg(); 879 unsigned Op2 = CMPI2->getOperand(2).getReg(); 880 CMPI2->getOperand(1).setReg(Op2); 881 CMPI2->getOperand(2).setReg(Op1); 882 } 883 if (NewImm2 != Imm2) 884 CMPI2->getOperand(2).setImm(NewImm2); 885 886 for (int I = 1; I <= 2; I++) { 887 if (CMPI2->getOperand(I).isReg()) { 888 MachineInstr *Inst = MRI->getVRegDef(CMPI2->getOperand(I).getReg()); 889 if (Inst->getParent() != &MBB2) 890 continue; 891 892 assert(Inst->getOpcode() == PPC::PHI && 893 "We cannot support if an operand comes from this BB."); 894 unsigned SrcReg = getIncomingRegForBlock(Inst, MBBtoMoveCmp); 895 CMPI2->getOperand(I).setReg(SrcReg); 896 } 897 } 898 auto I = MachineBasicBlock::iterator(MBBtoMoveCmp->getFirstTerminator()); 899 MBBtoMoveCmp->splice(I, &MBB2, MachineBasicBlock::iterator(CMPI2)); 900 901 DebugLoc DL = CMPI2->getDebugLoc(); 902 unsigned NewVReg = MRI->createVirtualRegister(&PPC::CRRCRegClass); 903 BuildMI(MBB2, MBB2.begin(), DL, 904 TII->get(PPC::PHI), NewVReg) 905 .addReg(BI1->getOperand(1).getReg()).addMBB(MBB1) 906 .addReg(BI2->getOperand(1).getReg()).addMBB(MBBtoMoveCmp); 907 BI2->getOperand(1).setReg(NewVReg); 908 } 909 else { 910 // We finally eliminate compare instruction in MBB2. 911 BI2->getOperand(1).setReg(BI1->getOperand(1).getReg()); 912 CMPI2->eraseFromParent(); 913 } 914 BI2->getOperand(1).setIsKill(true); 915 BI1->getOperand(1).setIsKill(false); 916 917 DEBUG(dbgs() << "into a compare and two branches:\n"); 918 DEBUG(CMPI1->dump()); 919 DEBUG(BI1->dump()); 920 DEBUG(BI2->dump()); 921 if (IsPartiallyRedundant) { 922 DEBUG(dbgs() << "The following compare is moved into BB#" << 923 MBBtoMoveCmp->getNumber() << " to handle partial redundancy.\n"); 924 DEBUG(CMPI2->dump()); 925 } 926 927 Simplified = true; 928 } 929 930 return Simplified; 931 } 932 933 // This is used to find the "true" source register for an 934 // XXPERMDI instruction, since MachineCSE does not handle the 935 // "copy-like" operations (Copy and SubregToReg). Returns 936 // the original SrcReg unless it is the target of a copy-like 937 // operation, in which case we chain backwards through all 938 // such operations to the ultimate source register. If a 939 // physical register is encountered, we stop the search. 940 unsigned PPCMIPeephole::lookThruCopyLike(unsigned SrcReg) { 941 942 while (true) { 943 944 MachineInstr *MI = MRI->getVRegDef(SrcReg); 945 if (!MI->isCopyLike()) 946 return SrcReg; 947 948 unsigned CopySrcReg; 949 if (MI->isCopy()) 950 CopySrcReg = MI->getOperand(1).getReg(); 951 else { 952 assert(MI->isSubregToReg() && "bad opcode for lookThruCopyLike"); 953 CopySrcReg = MI->getOperand(2).getReg(); 954 } 955 956 if (!TargetRegisterInfo::isVirtualRegister(CopySrcReg)) 957 return CopySrcReg; 958 959 SrcReg = CopySrcReg; 960 } 961 } 962 963 } // end default namespace 964 965 INITIALIZE_PASS_BEGIN(PPCMIPeephole, DEBUG_TYPE, 966 "PowerPC MI Peephole Optimization", false, false) 967 INITIALIZE_PASS_END(PPCMIPeephole, DEBUG_TYPE, 968 "PowerPC MI Peephole Optimization", false, false) 969 970 char PPCMIPeephole::ID = 0; 971 FunctionPass* 972 llvm::createPPCMIPeepholePass() { return new PPCMIPeephole(); } 973 974