//===-------------- PPCMIPeephole.cpp - MI Peephole Cleanups -------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===---------------------------------------------------------------------===// // // This pass performs peephole optimizations to clean up ugly code // sequences at the MachineInstruction layer. It runs at the end of // the SSA phases, following VSX swap removal. A pass of dead code // elimination follows this one for quick clean-up of any dead // instructions introduced here. Although we could do this as callbacks // from the generic peephole pass, this would have a couple of bad // effects: it might remove optimization opportunities for VSX swap // removal, and it would miss cleanups made possible following VSX // swap removal. // //===---------------------------------------------------------------------===// #include "MCTargetDesc/PPCMCTargetDesc.h" #include "MCTargetDesc/PPCPredicates.h" #include "PPC.h" #include "PPCInstrBuilder.h" #include "PPCInstrInfo.h" #include "PPCMachineFunctionInfo.h" #include "PPCTargetMachine.h" #include "llvm/ADT/Statistic.h" #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" #include "llvm/CodeGen/MachineDominators.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/CodeGen/MachinePostDominators.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/InitializePasses.h" #include "llvm/Support/Debug.h" using namespace llvm; #define DEBUG_TYPE "ppc-mi-peepholes" STATISTIC(RemoveTOCSave, "Number of TOC saves removed"); STATISTIC(MultiTOCSaves, "Number of functions with multiple TOC saves that must be kept"); STATISTIC(NumTOCSavesInPrologue, "Number of TOC saves placed in the prologue"); STATISTIC(NumEliminatedSExt, "Number of eliminated sign-extensions"); STATISTIC(NumEliminatedZExt, "Number of eliminated zero-extensions"); STATISTIC(NumOptADDLIs, "Number of optimized ADD instruction fed by LI"); STATISTIC(NumConvertedToImmediateForm, "Number of instructions converted to their immediate form"); STATISTIC(NumFunctionsEnteredInMIPeephole, "Number of functions entered in PPC MI Peepholes"); STATISTIC(NumFixedPointIterations, "Number of fixed-point iterations converting reg-reg instructions " "to reg-imm ones"); STATISTIC(NumRotatesCollapsed, "Number of pairs of rotate left, clear left/right collapsed"); STATISTIC(NumEXTSWAndSLDICombined, "Number of pairs of EXTSW and SLDI combined as EXTSWSLI"); STATISTIC(NumLoadImmZeroFoldedAndRemoved, "Number of LI(8) reg, 0 that are folded to r0 and removed"); static cl::opt FixedPointRegToImm("ppc-reg-to-imm-fixed-point", cl::Hidden, cl::init(true), cl::desc("Iterate to a fixed point when attempting to " "convert reg-reg instructions to reg-imm")); static cl::opt ConvertRegReg("ppc-convert-rr-to-ri", cl::Hidden, cl::init(true), cl::desc("Convert eligible reg+reg instructions to reg+imm")); static cl::opt EnableSExtElimination("ppc-eliminate-signext", cl::desc("enable elimination of sign-extensions"), cl::init(false), cl::Hidden); static cl::opt EnableZExtElimination("ppc-eliminate-zeroext", cl::desc("enable elimination of zero-extensions"), cl::init(false), cl::Hidden); namespace { struct PPCMIPeephole : public MachineFunctionPass { static char ID; const PPCInstrInfo *TII; MachineFunction *MF; MachineRegisterInfo *MRI; PPCMIPeephole() : MachineFunctionPass(ID) { initializePPCMIPeepholePass(*PassRegistry::getPassRegistry()); } private: MachineDominatorTree *MDT; MachinePostDominatorTree *MPDT; MachineBlockFrequencyInfo *MBFI; uint64_t EntryFreq; // Initialize class variables. void initialize(MachineFunction &MFParm); // Perform peepholes. bool simplifyCode(void); // Perform peepholes. bool eliminateRedundantCompare(void); bool eliminateRedundantTOCSaves(std::map &TOCSaves); bool combineSEXTAndSHL(MachineInstr &MI, MachineInstr *&ToErase); bool emitRLDICWhenLoweringJumpTables(MachineInstr &MI); void UpdateTOCSaves(std::map &TOCSaves, MachineInstr *MI); public: void getAnalysisUsage(AnalysisUsage &AU) const override { AU.addRequired(); AU.addRequired(); AU.addRequired(); AU.addPreserved(); AU.addPreserved(); AU.addPreserved(); MachineFunctionPass::getAnalysisUsage(AU); } // Main entry point for this pass. bool runOnMachineFunction(MachineFunction &MF) override { initialize(MF); // At this point, TOC pointer should not be used in a function that uses // PC-Relative addressing. assert((MF.getRegInfo().use_empty(PPC::X2) || !MF.getSubtarget().isUsingPCRelativeCalls()) && "TOC pointer used in a function using PC-Relative addressing!"); if (skipFunction(MF.getFunction())) return false; return simplifyCode(); } }; // Initialize class variables. void PPCMIPeephole::initialize(MachineFunction &MFParm) { MF = &MFParm; MRI = &MF->getRegInfo(); MDT = &getAnalysis(); MPDT = &getAnalysis(); MBFI = &getAnalysis(); EntryFreq = MBFI->getEntryFreq(); TII = MF->getSubtarget().getInstrInfo(); LLVM_DEBUG(dbgs() << "*** PowerPC MI peephole pass ***\n\n"); LLVM_DEBUG(MF->dump()); } static MachineInstr *getVRegDefOrNull(MachineOperand *Op, MachineRegisterInfo *MRI) { assert(Op && "Invalid Operand!"); if (!Op->isReg()) return nullptr; Register Reg = Op->getReg(); if (!Register::isVirtualRegister(Reg)) return nullptr; return MRI->getVRegDef(Reg); } // This function returns number of known zero bits in output of MI // starting from the most significant bit. static unsigned getKnownLeadingZeroCount(MachineInstr *MI, const PPCInstrInfo *TII) { unsigned Opcode = MI->getOpcode(); if (Opcode == PPC::RLDICL || Opcode == PPC::RLDICL_rec || Opcode == PPC::RLDCL || Opcode == PPC::RLDCL_rec) return MI->getOperand(3).getImm(); if ((Opcode == PPC::RLDIC || Opcode == PPC::RLDIC_rec) && MI->getOperand(3).getImm() <= 63 - MI->getOperand(2).getImm()) return MI->getOperand(3).getImm(); if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINM_rec || Opcode == PPC::RLWNM || Opcode == PPC::RLWNM_rec || Opcode == PPC::RLWINM8 || Opcode == PPC::RLWNM8) && MI->getOperand(3).getImm() <= MI->getOperand(4).getImm()) return 32 + MI->getOperand(3).getImm(); if (Opcode == PPC::ANDI_rec) { uint16_t Imm = MI->getOperand(2).getImm(); return 48 + countLeadingZeros(Imm); } if (Opcode == PPC::CNTLZW || Opcode == PPC::CNTLZW_rec || Opcode == PPC::CNTTZW || Opcode == PPC::CNTTZW_rec || Opcode == PPC::CNTLZW8 || Opcode == PPC::CNTTZW8) // The result ranges from 0 to 32. return 58; if (Opcode == PPC::CNTLZD || Opcode == PPC::CNTLZD_rec || Opcode == PPC::CNTTZD || Opcode == PPC::CNTTZD_rec) // The result ranges from 0 to 64. return 57; if (Opcode == PPC::LHZ || Opcode == PPC::LHZX || Opcode == PPC::LHZ8 || Opcode == PPC::LHZX8 || Opcode == PPC::LHZU || Opcode == PPC::LHZUX || Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8) return 48; if (Opcode == PPC::LBZ || Opcode == PPC::LBZX || Opcode == PPC::LBZ8 || Opcode == PPC::LBZX8 || Opcode == PPC::LBZU || Opcode == PPC::LBZUX || Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8) return 56; if (TII->isZeroExtended(*MI)) return 32; return 0; } // This function maintains a map for the pairs // Each time a new TOC save is encountered, it checks if any of the existing // ones are dominated by the new one. If so, it marks the existing one as // redundant by setting it's entry in the map as false. It then adds the new // instruction to the map with either true or false depending on if any // existing instructions dominated the new one. void PPCMIPeephole::UpdateTOCSaves( std::map &TOCSaves, MachineInstr *MI) { assert(TII->isTOCSaveMI(*MI) && "Expecting a TOC save instruction here"); assert(MF->getSubtarget().isELFv2ABI() && "TOC-save removal only supported on ELFv2"); PPCFunctionInfo *FI = MF->getInfo(); MachineBasicBlock *Entry = &MF->front(); uint64_t CurrBlockFreq = MBFI->getBlockFreq(MI->getParent()).getFrequency(); // If the block in which the TOC save resides is in a block that // post-dominates Entry, or a block that is hotter than entry (keep in mind // that early MachineLICM has already run so the TOC save won't be hoisted) // we can just do the save in the prologue. if (CurrBlockFreq > EntryFreq || MPDT->dominates(MI->getParent(), Entry)) FI->setMustSaveTOC(true); // If we are saving the TOC in the prologue, all the TOC saves can be removed // from the code. if (FI->mustSaveTOC()) { for (auto &TOCSave : TOCSaves) TOCSave.second = false; // Add new instruction to map. TOCSaves[MI] = false; return; } bool Keep = true; for (auto It = TOCSaves.begin(); It != TOCSaves.end(); It++ ) { MachineInstr *CurrInst = It->first; // If new instruction dominates an existing one, mark existing one as // redundant. if (It->second && MDT->dominates(MI, CurrInst)) It->second = false; // Check if the new instruction is redundant. if (MDT->dominates(CurrInst, MI)) { Keep = false; break; } } // Add new instruction to map. TOCSaves[MI] = Keep; } // This function returns a list of all PHI nodes in the tree starting from // the RootPHI node. We perform a BFS traversal to get an ordered list of nodes. // The list initially only contains the root PHI. When we visit a PHI node, we // add it to the list. We continue to look for other PHI node operands while // there are nodes to visit in the list. The function returns false if the // optimization cannot be applied on this tree. static bool collectUnprimedAccPHIs(MachineRegisterInfo *MRI, MachineInstr *RootPHI, SmallVectorImpl &PHIs) { PHIs.push_back(RootPHI); unsigned VisitedIndex = 0; while (VisitedIndex < PHIs.size()) { MachineInstr *VisitedPHI = PHIs[VisitedIndex]; for (unsigned PHIOp = 1, NumOps = VisitedPHI->getNumOperands(); PHIOp != NumOps; PHIOp += 2) { Register RegOp = VisitedPHI->getOperand(PHIOp).getReg(); if (!Register::isVirtualRegister(RegOp)) return false; MachineInstr *Instr = MRI->getVRegDef(RegOp); // While collecting the PHI nodes, we check if they can be converted (i.e. // all the operands are either copies, implicit defs or PHI nodes). unsigned Opcode = Instr->getOpcode(); if (Opcode == PPC::COPY) { Register Reg = Instr->getOperand(1).getReg(); if (!Register::isVirtualRegister(Reg) || MRI->getRegClass(Reg) != &PPC::ACCRCRegClass) return false; } else if (Opcode != PPC::IMPLICIT_DEF && Opcode != PPC::PHI) return false; // If we detect a cycle in the PHI nodes, we exit. It would be // possible to change cycles as well, but that would add a lot // of complexity for a case that is unlikely to occur with MMA // code. if (Opcode != PPC::PHI) continue; if (llvm::is_contained(PHIs, Instr)) return false; PHIs.push_back(Instr); } VisitedIndex++; } return true; } // This function changes the unprimed accumulator PHI nodes in the PHIs list to // primed accumulator PHI nodes. The list is traversed in reverse order to // change all the PHI operands of a PHI node before changing the node itself. // We keep a map to associate each changed PHI node to its non-changed form. static void convertUnprimedAccPHIs(const PPCInstrInfo *TII, MachineRegisterInfo *MRI, SmallVectorImpl &PHIs, Register Dst) { DenseMap ChangedPHIMap; for (auto It = PHIs.rbegin(), End = PHIs.rend(); It != End; ++It) { MachineInstr *PHI = *It; SmallVector, 4> PHIOps; // We check if the current PHI node can be changed by looking at its // operands. If all the operands are either copies from primed // accumulators, implicit definitions or other unprimed accumulator // PHI nodes, we change it. for (unsigned PHIOp = 1, NumOps = PHI->getNumOperands(); PHIOp != NumOps; PHIOp += 2) { Register RegOp = PHI->getOperand(PHIOp).getReg(); MachineInstr *PHIInput = MRI->getVRegDef(RegOp); unsigned Opcode = PHIInput->getOpcode(); assert((Opcode == PPC::COPY || Opcode == PPC::IMPLICIT_DEF || Opcode == PPC::PHI) && "Unexpected instruction"); if (Opcode == PPC::COPY) { assert(MRI->getRegClass(PHIInput->getOperand(1).getReg()) == &PPC::ACCRCRegClass && "Unexpected register class"); PHIOps.push_back({PHIInput->getOperand(1), PHI->getOperand(PHIOp + 1)}); } else if (Opcode == PPC::IMPLICIT_DEF) { Register AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass); BuildMI(*PHIInput->getParent(), PHIInput, PHIInput->getDebugLoc(), TII->get(PPC::IMPLICIT_DEF), AccReg); PHIOps.push_back({MachineOperand::CreateReg(AccReg, false), PHI->getOperand(PHIOp + 1)}); } else if (Opcode == PPC::PHI) { // We found a PHI operand. At this point we know this operand // has already been changed so we get its associated changed form // from the map. assert(ChangedPHIMap.count(PHIInput) == 1 && "This PHI node should have already been changed."); MachineInstr *PrimedAccPHI = ChangedPHIMap.lookup(PHIInput); PHIOps.push_back({MachineOperand::CreateReg( PrimedAccPHI->getOperand(0).getReg(), false), PHI->getOperand(PHIOp + 1)}); } } Register AccReg = Dst; // If the PHI node we are changing is the root node, the register it defines // will be the destination register of the original copy (of the PHI def). // For all other PHI's in the list, we need to create another primed // accumulator virtual register as the PHI will no longer define the // unprimed accumulator. if (PHI != PHIs[0]) AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass); MachineInstrBuilder NewPHI = BuildMI( *PHI->getParent(), PHI, PHI->getDebugLoc(), TII->get(PPC::PHI), AccReg); for (auto RegMBB : PHIOps) NewPHI.add(RegMBB.first).add(RegMBB.second); ChangedPHIMap[PHI] = NewPHI.getInstr(); } } // Perform peephole optimizations. bool PPCMIPeephole::simplifyCode(void) { bool Simplified = false; MachineInstr* ToErase = nullptr; std::map TOCSaves; const TargetRegisterInfo *TRI = &TII->getRegisterInfo(); NumFunctionsEnteredInMIPeephole++; if (ConvertRegReg) { // Fixed-point conversion of reg/reg instructions fed by load-immediate // into reg/imm instructions. FIXME: This is expensive, control it with // an option. bool SomethingChanged = false; do { NumFixedPointIterations++; SomethingChanged = false; for (MachineBasicBlock &MBB : *MF) { for (MachineInstr &MI : MBB) { if (MI.isDebugInstr()) continue; if (TII->convertToImmediateForm(MI)) { // We don't erase anything in case the def has other uses. Let DCE // remove it if it can be removed. LLVM_DEBUG(dbgs() << "Converted instruction to imm form: "); LLVM_DEBUG(MI.dump()); NumConvertedToImmediateForm++; SomethingChanged = true; Simplified = true; continue; } } } } while (SomethingChanged && FixedPointRegToImm); } for (MachineBasicBlock &MBB : *MF) { for (MachineInstr &MI : MBB) { // If the previous instruction was marked for elimination, // remove it now. if (ToErase) { ToErase->eraseFromParent(); ToErase = nullptr; } // Ignore debug instructions. if (MI.isDebugInstr()) continue; // Per-opcode peepholes. switch (MI.getOpcode()) { default: break; case PPC::COPY: { Register Src = MI.getOperand(1).getReg(); Register Dst = MI.getOperand(0).getReg(); if (!Register::isVirtualRegister(Src) || !Register::isVirtualRegister(Dst)) break; if (MRI->getRegClass(Src) != &PPC::UACCRCRegClass || MRI->getRegClass(Dst) != &PPC::ACCRCRegClass) break; // We are copying an unprimed accumulator to a primed accumulator. // If the input to the copy is a PHI that is fed only by (i) copies in // the other direction (ii) implicitly defined unprimed accumulators or // (iii) other PHI nodes satisfying (i) and (ii), we can change // the PHI to a PHI on primed accumulators (as long as we also change // its operands). To detect and change such copies, we first get a list // of all the PHI nodes starting from the root PHI node in BFS order. // We then visit all these PHI nodes to check if they can be changed to // primed accumulator PHI nodes and if so, we change them. MachineInstr *RootPHI = MRI->getVRegDef(Src); if (RootPHI->getOpcode() != PPC::PHI) break; SmallVector PHIs; if (!collectUnprimedAccPHIs(MRI, RootPHI, PHIs)) break; convertUnprimedAccPHIs(TII, MRI, PHIs, Dst); ToErase = &MI; break; } case PPC::LI: case PPC::LI8: { // If we are materializing a zero, look for any use operands for which // zero means immediate zero. All such operands can be replaced with // PPC::ZERO. if (!MI.getOperand(1).isImm() || MI.getOperand(1).getImm() != 0) break; unsigned MIDestReg = MI.getOperand(0).getReg(); for (MachineInstr& UseMI : MRI->use_instructions(MIDestReg)) Simplified |= TII->onlyFoldImmediate(UseMI, MI, MIDestReg); if (MRI->use_nodbg_empty(MIDestReg)) { ++NumLoadImmZeroFoldedAndRemoved; ToErase = &MI; } break; } case PPC::STD: { MachineFrameInfo &MFI = MF->getFrameInfo(); if (MFI.hasVarSizedObjects() || !MF->getSubtarget().isELFv2ABI()) break; // When encountering a TOC save instruction, call UpdateTOCSaves // to add it to the TOCSaves map and mark any existing TOC saves // it dominates as redundant. if (TII->isTOCSaveMI(MI)) UpdateTOCSaves(TOCSaves, &MI); break; } case PPC::XXPERMDI: { // Perform simplifications of 2x64 vector swaps and splats. // A swap is identified by an immediate value of 2, and a splat // is identified by an immediate value of 0 or 3. int Immed = MI.getOperand(3).getImm(); if (Immed == 1) break; // For each of these simplifications, we need the two source // regs to match. Unfortunately, MachineCSE ignores COPY and // SUBREG_TO_REG, so for example we can see // XXPERMDI t, SUBREG_TO_REG(s), SUBREG_TO_REG(s), immed. // We have to look through chains of COPY and SUBREG_TO_REG // to find the real source values for comparison. unsigned TrueReg1 = TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI); unsigned TrueReg2 = TRI->lookThruCopyLike(MI.getOperand(2).getReg(), MRI); if (!(TrueReg1 == TrueReg2 && Register::isVirtualRegister(TrueReg1))) break; MachineInstr *DefMI = MRI->getVRegDef(TrueReg1); if (!DefMI) break; unsigned DefOpc = DefMI->getOpcode(); // If this is a splat fed by a splatting load, the splat is // redundant. Replace with a copy. This doesn't happen directly due // to code in PPCDAGToDAGISel.cpp, but it can happen when converting // a load of a double to a vector of 64-bit integers. auto isConversionOfLoadAndSplat = [=]() -> bool { if (DefOpc != PPC::XVCVDPSXDS && DefOpc != PPC::XVCVDPUXDS) return false; unsigned FeedReg1 = TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI); if (Register::isVirtualRegister(FeedReg1)) { MachineInstr *LoadMI = MRI->getVRegDef(FeedReg1); if (LoadMI && LoadMI->getOpcode() == PPC::LXVDSX) return true; } return false; }; if ((Immed == 0 || Immed == 3) && (DefOpc == PPC::LXVDSX || isConversionOfLoadAndSplat())) { LLVM_DEBUG(dbgs() << "Optimizing load-and-splat/splat " "to load-and-splat/copy: "); LLVM_DEBUG(MI.dump()); BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), MI.getOperand(0).getReg()) .add(MI.getOperand(1)); ToErase = &MI; Simplified = true; } // If this is a splat or a swap fed by another splat, we // can replace it with a copy. if (DefOpc == PPC::XXPERMDI) { unsigned DefReg1 = DefMI->getOperand(1).getReg(); unsigned DefReg2 = DefMI->getOperand(2).getReg(); unsigned DefImmed = DefMI->getOperand(3).getImm(); // If the two inputs are not the same register, check to see if // they originate from the same virtual register after only // copy-like instructions. if (DefReg1 != DefReg2) { unsigned FeedReg1 = TRI->lookThruCopyLike(DefReg1, MRI); unsigned FeedReg2 = TRI->lookThruCopyLike(DefReg2, MRI); if (!(FeedReg1 == FeedReg2 && Register::isVirtualRegister(FeedReg1))) break; } if (DefImmed == 0 || DefImmed == 3) { LLVM_DEBUG(dbgs() << "Optimizing splat/swap or splat/splat " "to splat/copy: "); LLVM_DEBUG(MI.dump()); BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), MI.getOperand(0).getReg()) .add(MI.getOperand(1)); ToErase = &MI; Simplified = true; } // If this is a splat fed by a swap, we can simplify modify // the splat to splat the other value from the swap's input // parameter. else if ((Immed == 0 || Immed == 3) && DefImmed == 2) { LLVM_DEBUG(dbgs() << "Optimizing swap/splat => splat: "); LLVM_DEBUG(MI.dump()); MI.getOperand(1).setReg(DefReg1); MI.getOperand(2).setReg(DefReg2); MI.getOperand(3).setImm(3 - Immed); Simplified = true; } // If this is a swap fed by a swap, we can replace it // with a copy from the first swap's input. else if (Immed == 2 && DefImmed == 2) { LLVM_DEBUG(dbgs() << "Optimizing swap/swap => copy: "); LLVM_DEBUG(MI.dump()); BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), MI.getOperand(0).getReg()) .add(DefMI->getOperand(1)); ToErase = &MI; Simplified = true; } } else if ((Immed == 0 || Immed == 3) && DefOpc == PPC::XXPERMDIs && (DefMI->getOperand(2).getImm() == 0 || DefMI->getOperand(2).getImm() == 3)) { // Splat fed by another splat - switch the output of the first // and remove the second. DefMI->getOperand(0).setReg(MI.getOperand(0).getReg()); ToErase = &MI; Simplified = true; LLVM_DEBUG(dbgs() << "Removing redundant splat: "); LLVM_DEBUG(MI.dump()); } break; } case PPC::VSPLTB: case PPC::VSPLTH: case PPC::XXSPLTW: { unsigned MyOpcode = MI.getOpcode(); unsigned OpNo = MyOpcode == PPC::XXSPLTW ? 1 : 2; unsigned TrueReg = TRI->lookThruCopyLike(MI.getOperand(OpNo).getReg(), MRI); if (!Register::isVirtualRegister(TrueReg)) break; MachineInstr *DefMI = MRI->getVRegDef(TrueReg); if (!DefMI) break; unsigned DefOpcode = DefMI->getOpcode(); auto isConvertOfSplat = [=]() -> bool { if (DefOpcode != PPC::XVCVSPSXWS && DefOpcode != PPC::XVCVSPUXWS) return false; Register ConvReg = DefMI->getOperand(1).getReg(); if (!Register::isVirtualRegister(ConvReg)) return false; MachineInstr *Splt = MRI->getVRegDef(ConvReg); return Splt && (Splt->getOpcode() == PPC::LXVWSX || Splt->getOpcode() == PPC::XXSPLTW); }; bool AlreadySplat = (MyOpcode == DefOpcode) || (MyOpcode == PPC::VSPLTB && DefOpcode == PPC::VSPLTBs) || (MyOpcode == PPC::VSPLTH && DefOpcode == PPC::VSPLTHs) || (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::XXSPLTWs) || (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::LXVWSX) || (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::MTVSRWS)|| (MyOpcode == PPC::XXSPLTW && isConvertOfSplat()); // If the instruction[s] that feed this splat have already splat // the value, this splat is redundant. if (AlreadySplat) { LLVM_DEBUG(dbgs() << "Changing redundant splat to a copy: "); LLVM_DEBUG(MI.dump()); BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), MI.getOperand(0).getReg()) .add(MI.getOperand(OpNo)); ToErase = &MI; Simplified = true; } // Splat fed by a shift. Usually when we align value to splat into // vector element zero. if (DefOpcode == PPC::XXSLDWI) { Register ShiftRes = DefMI->getOperand(0).getReg(); Register ShiftOp1 = DefMI->getOperand(1).getReg(); Register ShiftOp2 = DefMI->getOperand(2).getReg(); unsigned ShiftImm = DefMI->getOperand(3).getImm(); unsigned SplatImm = MI.getOperand(2).getImm(); if (ShiftOp1 == ShiftOp2) { unsigned NewElem = (SplatImm + ShiftImm) & 0x3; if (MRI->hasOneNonDBGUse(ShiftRes)) { LLVM_DEBUG(dbgs() << "Removing redundant shift: "); LLVM_DEBUG(DefMI->dump()); ToErase = DefMI; } Simplified = true; LLVM_DEBUG(dbgs() << "Changing splat immediate from " << SplatImm << " to " << NewElem << " in instruction: "); LLVM_DEBUG(MI.dump()); MI.getOperand(1).setReg(ShiftOp1); MI.getOperand(2).setImm(NewElem); } } break; } case PPC::XVCVDPSP: { // If this is a DP->SP conversion fed by an FRSP, the FRSP is redundant. unsigned TrueReg = TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI); if (!Register::isVirtualRegister(TrueReg)) break; MachineInstr *DefMI = MRI->getVRegDef(TrueReg); // This can occur when building a vector of single precision or integer // values. if (DefMI && DefMI->getOpcode() == PPC::XXPERMDI) { unsigned DefsReg1 = TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI); unsigned DefsReg2 = TRI->lookThruCopyLike(DefMI->getOperand(2).getReg(), MRI); if (!Register::isVirtualRegister(DefsReg1) || !Register::isVirtualRegister(DefsReg2)) break; MachineInstr *P1 = MRI->getVRegDef(DefsReg1); MachineInstr *P2 = MRI->getVRegDef(DefsReg2); if (!P1 || !P2) break; // Remove the passed FRSP/XSRSP instruction if it only feeds this MI // and set any uses of that FRSP/XSRSP (in this MI) to the source of // the FRSP/XSRSP. auto removeFRSPIfPossible = [&](MachineInstr *RoundInstr) { unsigned Opc = RoundInstr->getOpcode(); if ((Opc == PPC::FRSP || Opc == PPC::XSRSP) && MRI->hasOneNonDBGUse(RoundInstr->getOperand(0).getReg())) { Simplified = true; Register ConvReg1 = RoundInstr->getOperand(1).getReg(); Register FRSPDefines = RoundInstr->getOperand(0).getReg(); MachineInstr &Use = *(MRI->use_instr_nodbg_begin(FRSPDefines)); for (int i = 0, e = Use.getNumOperands(); i < e; ++i) if (Use.getOperand(i).isReg() && Use.getOperand(i).getReg() == FRSPDefines) Use.getOperand(i).setReg(ConvReg1); LLVM_DEBUG(dbgs() << "Removing redundant FRSP/XSRSP:\n"); LLVM_DEBUG(RoundInstr->dump()); LLVM_DEBUG(dbgs() << "As it feeds instruction:\n"); LLVM_DEBUG(MI.dump()); LLVM_DEBUG(dbgs() << "Through instruction:\n"); LLVM_DEBUG(DefMI->dump()); RoundInstr->eraseFromParent(); } }; // If the input to XVCVDPSP is a vector that was built (even // partially) out of FRSP's, the FRSP(s) can safely be removed // since this instruction performs the same operation. if (P1 != P2) { removeFRSPIfPossible(P1); removeFRSPIfPossible(P2); break; } removeFRSPIfPossible(P1); } break; } case PPC::EXTSH: case PPC::EXTSH8: case PPC::EXTSH8_32_64: { if (!EnableSExtElimination) break; Register NarrowReg = MI.getOperand(1).getReg(); if (!Register::isVirtualRegister(NarrowReg)) break; MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg); // If we've used a zero-extending load that we will sign-extend, // just do a sign-extending load. if (SrcMI->getOpcode() == PPC::LHZ || SrcMI->getOpcode() == PPC::LHZX) { if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg())) break; auto is64Bit = [] (unsigned Opcode) { return Opcode == PPC::EXTSH8; }; auto isXForm = [] (unsigned Opcode) { return Opcode == PPC::LHZX; }; auto getSextLoadOp = [] (bool is64Bit, bool isXForm) { if (is64Bit) if (isXForm) return PPC::LHAX8; else return PPC::LHA8; else if (isXForm) return PPC::LHAX; else return PPC::LHA; }; unsigned Opc = getSextLoadOp(is64Bit(MI.getOpcode()), isXForm(SrcMI->getOpcode())); LLVM_DEBUG(dbgs() << "Zero-extending load\n"); LLVM_DEBUG(SrcMI->dump()); LLVM_DEBUG(dbgs() << "and sign-extension\n"); LLVM_DEBUG(MI.dump()); LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n"); SrcMI->setDesc(TII->get(Opc)); SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg()); ToErase = &MI; Simplified = true; NumEliminatedSExt++; } break; } case PPC::EXTSW: case PPC::EXTSW_32: case PPC::EXTSW_32_64: { if (!EnableSExtElimination) break; Register NarrowReg = MI.getOperand(1).getReg(); if (!Register::isVirtualRegister(NarrowReg)) break; MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg); // If we've used a zero-extending load that we will sign-extend, // just do a sign-extending load. if (SrcMI->getOpcode() == PPC::LWZ || SrcMI->getOpcode() == PPC::LWZX) { if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg())) break; auto is64Bit = [] (unsigned Opcode) { return Opcode == PPC::EXTSW || Opcode == PPC::EXTSW_32_64; }; auto isXForm = [] (unsigned Opcode) { return Opcode == PPC::LWZX; }; auto getSextLoadOp = [] (bool is64Bit, bool isXForm) { if (is64Bit) if (isXForm) return PPC::LWAX; else return PPC::LWA; else if (isXForm) return PPC::LWAX_32; else return PPC::LWA_32; }; unsigned Opc = getSextLoadOp(is64Bit(MI.getOpcode()), isXForm(SrcMI->getOpcode())); LLVM_DEBUG(dbgs() << "Zero-extending load\n"); LLVM_DEBUG(SrcMI->dump()); LLVM_DEBUG(dbgs() << "and sign-extension\n"); LLVM_DEBUG(MI.dump()); LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n"); SrcMI->setDesc(TII->get(Opc)); SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg()); ToErase = &MI; Simplified = true; NumEliminatedSExt++; } else if (MI.getOpcode() == PPC::EXTSW_32_64 && TII->isSignExtended(*SrcMI)) { // We can eliminate EXTSW if the input is known to be already // sign-extended. LLVM_DEBUG(dbgs() << "Removing redundant sign-extension\n"); Register TmpReg = MF->getRegInfo().createVirtualRegister(&PPC::G8RCRegClass); BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::IMPLICIT_DEF), TmpReg); BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::INSERT_SUBREG), MI.getOperand(0).getReg()) .addReg(TmpReg) .addReg(NarrowReg) .addImm(PPC::sub_32); ToErase = &MI; Simplified = true; NumEliminatedSExt++; } break; } case PPC::RLDICL: { // We can eliminate RLDICL (e.g. for zero-extension) // if all bits to clear are already zero in the input. // This code assume following code sequence for zero-extension. // %6 = COPY %5:sub_32; (optional) // %8 = IMPLICIT_DEF; // %7 = INSERT_SUBREG %8, %6, sub_32; if (!EnableZExtElimination) break; if (MI.getOperand(2).getImm() != 0) break; Register SrcReg = MI.getOperand(1).getReg(); if (!Register::isVirtualRegister(SrcReg)) break; MachineInstr *SrcMI = MRI->getVRegDef(SrcReg); if (!(SrcMI && SrcMI->getOpcode() == PPC::INSERT_SUBREG && SrcMI->getOperand(0).isReg() && SrcMI->getOperand(1).isReg())) break; MachineInstr *ImpDefMI, *SubRegMI; ImpDefMI = MRI->getVRegDef(SrcMI->getOperand(1).getReg()); SubRegMI = MRI->getVRegDef(SrcMI->getOperand(2).getReg()); if (ImpDefMI->getOpcode() != PPC::IMPLICIT_DEF) break; SrcMI = SubRegMI; if (SubRegMI->getOpcode() == PPC::COPY) { Register CopyReg = SubRegMI->getOperand(1).getReg(); if (Register::isVirtualRegister(CopyReg)) SrcMI = MRI->getVRegDef(CopyReg); } unsigned KnownZeroCount = getKnownLeadingZeroCount(SrcMI, TII); if (MI.getOperand(3).getImm() <= KnownZeroCount) { LLVM_DEBUG(dbgs() << "Removing redundant zero-extension\n"); BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), MI.getOperand(0).getReg()) .addReg(SrcReg); ToErase = &MI; Simplified = true; NumEliminatedZExt++; } break; } // TODO: Any instruction that has an immediate form fed only by a PHI // whose operands are all load immediate can be folded away. We currently // do this for ADD instructions, but should expand it to arithmetic and // binary instructions with immediate forms in the future. case PPC::ADD4: case PPC::ADD8: { auto isSingleUsePHI = [&](MachineOperand *PhiOp) { assert(PhiOp && "Invalid Operand!"); MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI); return DefPhiMI && (DefPhiMI->getOpcode() == PPC::PHI) && MRI->hasOneNonDBGUse(DefPhiMI->getOperand(0).getReg()); }; auto dominatesAllSingleUseLIs = [&](MachineOperand *DominatorOp, MachineOperand *PhiOp) { assert(PhiOp && "Invalid Operand!"); assert(DominatorOp && "Invalid Operand!"); MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI); MachineInstr *DefDomMI = getVRegDefOrNull(DominatorOp, MRI); // Note: the vregs only show up at odd indices position of PHI Node, // the even indices position save the BB info. for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) { MachineInstr *LiMI = getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI); if (!LiMI || (LiMI->getOpcode() != PPC::LI && LiMI->getOpcode() != PPC::LI8) || !MRI->hasOneNonDBGUse(LiMI->getOperand(0).getReg()) || !MDT->dominates(DefDomMI, LiMI)) return false; } return true; }; MachineOperand Op1 = MI.getOperand(1); MachineOperand Op2 = MI.getOperand(2); if (isSingleUsePHI(&Op2) && dominatesAllSingleUseLIs(&Op1, &Op2)) std::swap(Op1, Op2); else if (!isSingleUsePHI(&Op1) || !dominatesAllSingleUseLIs(&Op2, &Op1)) break; // We don't have an ADD fed by LI's that can be transformed // Now we know that Op1 is the PHI node and Op2 is the dominator Register DominatorReg = Op2.getReg(); const TargetRegisterClass *TRC = MI.getOpcode() == PPC::ADD8 ? &PPC::G8RC_and_G8RC_NOX0RegClass : &PPC::GPRC_and_GPRC_NOR0RegClass; MRI->setRegClass(DominatorReg, TRC); // replace LIs with ADDIs MachineInstr *DefPhiMI = getVRegDefOrNull(&Op1, MRI); for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) { MachineInstr *LiMI = getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI); LLVM_DEBUG(dbgs() << "Optimizing LI to ADDI: "); LLVM_DEBUG(LiMI->dump()); // There could be repeated registers in the PHI, e.g: %1 = // PHI %6, <%bb.2>, %8, <%bb.3>, %8, <%bb.6>; So if we've // already replaced the def instruction, skip. if (LiMI->getOpcode() == PPC::ADDI || LiMI->getOpcode() == PPC::ADDI8) continue; assert((LiMI->getOpcode() == PPC::LI || LiMI->getOpcode() == PPC::LI8) && "Invalid Opcode!"); auto LiImm = LiMI->getOperand(1).getImm(); // save the imm of LI LiMI->RemoveOperand(1); // remove the imm of LI LiMI->setDesc(TII->get(LiMI->getOpcode() == PPC::LI ? PPC::ADDI : PPC::ADDI8)); MachineInstrBuilder(*LiMI->getParent()->getParent(), *LiMI) .addReg(DominatorReg) .addImm(LiImm); // restore the imm of LI LLVM_DEBUG(LiMI->dump()); } // Replace ADD with COPY LLVM_DEBUG(dbgs() << "Optimizing ADD to COPY: "); LLVM_DEBUG(MI.dump()); BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY), MI.getOperand(0).getReg()) .add(Op1); ToErase = &MI; Simplified = true; NumOptADDLIs++; break; } case PPC::RLDICR: { Simplified |= emitRLDICWhenLoweringJumpTables(MI) || combineSEXTAndSHL(MI, ToErase); break; } case PPC::RLWINM: case PPC::RLWINM_rec: case PPC::RLWINM8: case PPC::RLWINM8_rec: { Simplified = TII->combineRLWINM(MI, &ToErase); if (Simplified) ++NumRotatesCollapsed; break; } } } // If the last instruction was marked for elimination, // remove it now. if (ToErase) { ToErase->eraseFromParent(); ToErase = nullptr; } } // Eliminate all the TOC save instructions which are redundant. Simplified |= eliminateRedundantTOCSaves(TOCSaves); PPCFunctionInfo *FI = MF->getInfo(); if (FI->mustSaveTOC()) NumTOCSavesInPrologue++; // We try to eliminate redundant compare instruction. Simplified |= eliminateRedundantCompare(); return Simplified; } // helper functions for eliminateRedundantCompare static bool isEqOrNe(MachineInstr *BI) { PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); unsigned PredCond = PPC::getPredicateCondition(Pred); return (PredCond == PPC::PRED_EQ || PredCond == PPC::PRED_NE); } static bool isSupportedCmpOp(unsigned opCode) { return (opCode == PPC::CMPLD || opCode == PPC::CMPD || opCode == PPC::CMPLW || opCode == PPC::CMPW || opCode == PPC::CMPLDI || opCode == PPC::CMPDI || opCode == PPC::CMPLWI || opCode == PPC::CMPWI); } static bool is64bitCmpOp(unsigned opCode) { return (opCode == PPC::CMPLD || opCode == PPC::CMPD || opCode == PPC::CMPLDI || opCode == PPC::CMPDI); } static bool isSignedCmpOp(unsigned opCode) { return (opCode == PPC::CMPD || opCode == PPC::CMPW || opCode == PPC::CMPDI || opCode == PPC::CMPWI); } static unsigned getSignedCmpOpCode(unsigned opCode) { if (opCode == PPC::CMPLD) return PPC::CMPD; if (opCode == PPC::CMPLW) return PPC::CMPW; if (opCode == PPC::CMPLDI) return PPC::CMPDI; if (opCode == PPC::CMPLWI) return PPC::CMPWI; return opCode; } // We can decrement immediate x in (GE x) by changing it to (GT x-1) or // (LT x) to (LE x-1) static unsigned getPredicateToDecImm(MachineInstr *BI, MachineInstr *CMPI) { uint64_t Imm = CMPI->getOperand(2).getImm(); bool SignedCmp = isSignedCmpOp(CMPI->getOpcode()); if ((!SignedCmp && Imm == 0) || (SignedCmp && Imm == 0x8000)) return 0; PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); unsigned PredCond = PPC::getPredicateCondition(Pred); unsigned PredHint = PPC::getPredicateHint(Pred); if (PredCond == PPC::PRED_GE) return PPC::getPredicate(PPC::PRED_GT, PredHint); if (PredCond == PPC::PRED_LT) return PPC::getPredicate(PPC::PRED_LE, PredHint); return 0; } // We can increment immediate x in (GT x) by changing it to (GE x+1) or // (LE x) to (LT x+1) static unsigned getPredicateToIncImm(MachineInstr *BI, MachineInstr *CMPI) { uint64_t Imm = CMPI->getOperand(2).getImm(); bool SignedCmp = isSignedCmpOp(CMPI->getOpcode()); if ((!SignedCmp && Imm == 0xFFFF) || (SignedCmp && Imm == 0x7FFF)) return 0; PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm(); unsigned PredCond = PPC::getPredicateCondition(Pred); unsigned PredHint = PPC::getPredicateHint(Pred); if (PredCond == PPC::PRED_GT) return PPC::getPredicate(PPC::PRED_GE, PredHint); if (PredCond == PPC::PRED_LE) return PPC::getPredicate(PPC::PRED_LT, PredHint); return 0; } // This takes a Phi node and returns a register value for the specified BB. static unsigned getIncomingRegForBlock(MachineInstr *Phi, MachineBasicBlock *MBB) { for (unsigned I = 2, E = Phi->getNumOperands() + 1; I != E; I += 2) { MachineOperand &MO = Phi->getOperand(I); if (MO.getMBB() == MBB) return Phi->getOperand(I-1).getReg(); } llvm_unreachable("invalid src basic block for this Phi node\n"); return 0; } // This function tracks the source of the register through register copy. // If BB1 and BB2 are non-NULL, we also track PHI instruction in BB2 // assuming that the control comes from BB1 into BB2. static unsigned getSrcVReg(unsigned Reg, MachineBasicBlock *BB1, MachineBasicBlock *BB2, MachineRegisterInfo *MRI) { unsigned SrcReg = Reg; while (1) { unsigned NextReg = SrcReg; MachineInstr *Inst = MRI->getVRegDef(SrcReg); if (BB1 && Inst->getOpcode() == PPC::PHI && Inst->getParent() == BB2) { NextReg = getIncomingRegForBlock(Inst, BB1); // We track through PHI only once to avoid infinite loop. BB1 = nullptr; } else if (Inst->isFullCopy()) NextReg = Inst->getOperand(1).getReg(); if (NextReg == SrcReg || !Register::isVirtualRegister(NextReg)) break; SrcReg = NextReg; } return SrcReg; } static bool eligibleForCompareElimination(MachineBasicBlock &MBB, MachineBasicBlock *&PredMBB, MachineBasicBlock *&MBBtoMoveCmp, MachineRegisterInfo *MRI) { auto isEligibleBB = [&](MachineBasicBlock &BB) { auto BII = BB.getFirstInstrTerminator(); // We optimize BBs ending with a conditional branch. // We check only for BCC here, not BCCLR, because BCCLR // will be formed only later in the pipeline. if (BB.succ_size() == 2 && BII != BB.instr_end() && (*BII).getOpcode() == PPC::BCC && (*BII).getOperand(1).isReg()) { // We optimize only if the condition code is used only by one BCC. Register CndReg = (*BII).getOperand(1).getReg(); if (!Register::isVirtualRegister(CndReg) || !MRI->hasOneNonDBGUse(CndReg)) return false; MachineInstr *CMPI = MRI->getVRegDef(CndReg); // We assume compare and branch are in the same BB for ease of analysis. if (CMPI->getParent() != &BB) return false; // We skip this BB if a physical register is used in comparison. for (MachineOperand &MO : CMPI->operands()) if (MO.isReg() && !Register::isVirtualRegister(MO.getReg())) return false; return true; } return false; }; // If this BB has more than one successor, we can create a new BB and // move the compare instruction in the new BB. // So far, we do not move compare instruction to a BB having multiple // successors to avoid potentially increasing code size. auto isEligibleForMoveCmp = [](MachineBasicBlock &BB) { return BB.succ_size() == 1; }; if (!isEligibleBB(MBB)) return false; unsigned NumPredBBs = MBB.pred_size(); if (NumPredBBs == 1) { MachineBasicBlock *TmpMBB = *MBB.pred_begin(); if (isEligibleBB(*TmpMBB)) { PredMBB = TmpMBB; MBBtoMoveCmp = nullptr; return true; } } else if (NumPredBBs == 2) { // We check for partially redundant case. // So far, we support cases with only two predecessors // to avoid increasing the number of instructions. MachineBasicBlock::pred_iterator PI = MBB.pred_begin(); MachineBasicBlock *Pred1MBB = *PI; MachineBasicBlock *Pred2MBB = *(PI+1); if (isEligibleBB(*Pred1MBB) && isEligibleForMoveCmp(*Pred2MBB)) { // We assume Pred1MBB is the BB containing the compare to be merged and // Pred2MBB is the BB to which we will append a compare instruction. // Hence we can proceed as is. } else if (isEligibleBB(*Pred2MBB) && isEligibleForMoveCmp(*Pred1MBB)) { // We need to swap Pred1MBB and Pred2MBB to canonicalize. std::swap(Pred1MBB, Pred2MBB); } else return false; // Here, Pred2MBB is the BB to which we need to append a compare inst. // We cannot move the compare instruction if operands are not available // in Pred2MBB (i.e. defined in MBB by an instruction other than PHI). MachineInstr *BI = &*MBB.getFirstInstrTerminator(); MachineInstr *CMPI = MRI->getVRegDef(BI->getOperand(1).getReg()); for (int I = 1; I <= 2; I++) if (CMPI->getOperand(I).isReg()) { MachineInstr *Inst = MRI->getVRegDef(CMPI->getOperand(I).getReg()); if (Inst->getParent() == &MBB && Inst->getOpcode() != PPC::PHI) return false; } PredMBB = Pred1MBB; MBBtoMoveCmp = Pred2MBB; return true; } return false; } // This function will iterate over the input map containing a pair of TOC save // instruction and a flag. The flag will be set to false if the TOC save is // proven redundant. This function will erase from the basic block all the TOC // saves marked as redundant. bool PPCMIPeephole::eliminateRedundantTOCSaves( std::map &TOCSaves) { bool Simplified = false; int NumKept = 0; for (auto TOCSave : TOCSaves) { if (!TOCSave.second) { TOCSave.first->eraseFromParent(); RemoveTOCSave++; Simplified = true; } else { NumKept++; } } if (NumKept > 1) MultiTOCSaves++; return Simplified; } // If multiple conditional branches are executed based on the (essentially) // same comparison, we merge compare instructions into one and make multiple // conditional branches on this comparison. // For example, // if (a == 0) { ... } // else if (a < 0) { ... } // can be executed by one compare and two conditional branches instead of // two pairs of a compare and a conditional branch. // // This method merges two compare instructions in two MBBs and modifies the // compare and conditional branch instructions if needed. // For the above example, the input for this pass looks like: // cmplwi r3, 0 // beq 0, .LBB0_3 // cmpwi r3, -1 // bgt 0, .LBB0_4 // So, before merging two compares, we need to modify these instructions as // cmpwi r3, 0 ; cmplwi and cmpwi yield same result for beq // beq 0, .LBB0_3 // cmpwi r3, 0 ; greather than -1 means greater or equal to 0 // bge 0, .LBB0_4 bool PPCMIPeephole::eliminateRedundantCompare(void) { bool Simplified = false; for (MachineBasicBlock &MBB2 : *MF) { MachineBasicBlock *MBB1 = nullptr, *MBBtoMoveCmp = nullptr; // For fully redundant case, we select two basic blocks MBB1 and MBB2 // as an optimization target if // - both MBBs end with a conditional branch, // - MBB1 is the only predecessor of MBB2, and // - compare does not take a physical register as a operand in both MBBs. // In this case, eligibleForCompareElimination sets MBBtoMoveCmp nullptr. // // As partially redundant case, we additionally handle if MBB2 has one // additional predecessor, which has only one successor (MBB2). // In this case, we move the compare instruction originally in MBB2 into // MBBtoMoveCmp. This partially redundant case is typically appear by // compiling a while loop; here, MBBtoMoveCmp is the loop preheader. // // Overview of CFG of related basic blocks // Fully redundant case Partially redundant case // -------- ---------------- -------- // | MBB1 | (w/ 2 succ) | MBBtoMoveCmp | | MBB1 | (w/ 2 succ) // -------- ---------------- -------- // | \ (w/ 1 succ) \ | \ // | \ \ | \ // | \ | // -------- -------- // | MBB2 | (w/ 1 pred | MBB2 | (w/ 2 pred // -------- and 2 succ) -------- and 2 succ) // | \ | \ // | \ | \ // if (!eligibleForCompareElimination(MBB2, MBB1, MBBtoMoveCmp, MRI)) continue; MachineInstr *BI1 = &*MBB1->getFirstInstrTerminator(); MachineInstr *CMPI1 = MRI->getVRegDef(BI1->getOperand(1).getReg()); MachineInstr *BI2 = &*MBB2.getFirstInstrTerminator(); MachineInstr *CMPI2 = MRI->getVRegDef(BI2->getOperand(1).getReg()); bool IsPartiallyRedundant = (MBBtoMoveCmp != nullptr); // We cannot optimize an unsupported compare opcode or // a mix of 32-bit and 64-bit comaprisons if (!isSupportedCmpOp(CMPI1->getOpcode()) || !isSupportedCmpOp(CMPI2->getOpcode()) || is64bitCmpOp(CMPI1->getOpcode()) != is64bitCmpOp(CMPI2->getOpcode())) continue; unsigned NewOpCode = 0; unsigned NewPredicate1 = 0, NewPredicate2 = 0; int16_t Imm1 = 0, NewImm1 = 0, Imm2 = 0, NewImm2 = 0; bool SwapOperands = false; if (CMPI1->getOpcode() != CMPI2->getOpcode()) { // Typically, unsigned comparison is used for equality check, but // we replace it with a signed comparison if the comparison // to be merged is a signed comparison. // In other cases of opcode mismatch, we cannot optimize this. // We cannot change opcode when comparing against an immediate // if the most significant bit of the immediate is one // due to the difference in sign extension. auto CmpAgainstImmWithSignBit = [](MachineInstr *I) { if (!I->getOperand(2).isImm()) return false; int16_t Imm = (int16_t)I->getOperand(2).getImm(); return Imm < 0; }; if (isEqOrNe(BI2) && !CmpAgainstImmWithSignBit(CMPI2) && CMPI1->getOpcode() == getSignedCmpOpCode(CMPI2->getOpcode())) NewOpCode = CMPI1->getOpcode(); else if (isEqOrNe(BI1) && !CmpAgainstImmWithSignBit(CMPI1) && getSignedCmpOpCode(CMPI1->getOpcode()) == CMPI2->getOpcode()) NewOpCode = CMPI2->getOpcode(); else continue; } if (CMPI1->getOperand(2).isReg() && CMPI2->getOperand(2).isReg()) { // In case of comparisons between two registers, these two registers // must be same to merge two comparisons. unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(), nullptr, nullptr, MRI); unsigned Cmp1Operand2 = getSrcVReg(CMPI1->getOperand(2).getReg(), nullptr, nullptr, MRI); unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(), MBB1, &MBB2, MRI); unsigned Cmp2Operand2 = getSrcVReg(CMPI2->getOperand(2).getReg(), MBB1, &MBB2, MRI); if (Cmp1Operand1 == Cmp2Operand1 && Cmp1Operand2 == Cmp2Operand2) { // Same pair of registers in the same order; ready to merge as is. } else if (Cmp1Operand1 == Cmp2Operand2 && Cmp1Operand2 == Cmp2Operand1) { // Same pair of registers in different order. // We reverse the predicate to merge compare instructions. PPC::Predicate Pred = (PPC::Predicate)BI2->getOperand(0).getImm(); NewPredicate2 = (unsigned)PPC::getSwappedPredicate(Pred); // In case of partial redundancy, we need to swap operands // in another compare instruction. SwapOperands = true; } else continue; } else if (CMPI1->getOperand(2).isImm() && CMPI2->getOperand(2).isImm()) { // In case of comparisons between a register and an immediate, // the operand register must be same for two compare instructions. unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(), nullptr, nullptr, MRI); unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(), MBB1, &MBB2, MRI); if (Cmp1Operand1 != Cmp2Operand1) continue; NewImm1 = Imm1 = (int16_t)CMPI1->getOperand(2).getImm(); NewImm2 = Imm2 = (int16_t)CMPI2->getOperand(2).getImm(); // If immediate are not same, we try to adjust by changing predicate; // e.g. GT imm means GE (imm+1). if (Imm1 != Imm2 && (!isEqOrNe(BI2) || !isEqOrNe(BI1))) { int Diff = Imm1 - Imm2; if (Diff < -2 || Diff > 2) continue; unsigned PredToInc1 = getPredicateToIncImm(BI1, CMPI1); unsigned PredToDec1 = getPredicateToDecImm(BI1, CMPI1); unsigned PredToInc2 = getPredicateToIncImm(BI2, CMPI2); unsigned PredToDec2 = getPredicateToDecImm(BI2, CMPI2); if (Diff == 2) { if (PredToInc2 && PredToDec1) { NewPredicate2 = PredToInc2; NewPredicate1 = PredToDec1; NewImm2++; NewImm1--; } } else if (Diff == 1) { if (PredToInc2) { NewImm2++; NewPredicate2 = PredToInc2; } else if (PredToDec1) { NewImm1--; NewPredicate1 = PredToDec1; } } else if (Diff == -1) { if (PredToDec2) { NewImm2--; NewPredicate2 = PredToDec2; } else if (PredToInc1) { NewImm1++; NewPredicate1 = PredToInc1; } } else if (Diff == -2) { if (PredToDec2 && PredToInc1) { NewPredicate2 = PredToDec2; NewPredicate1 = PredToInc1; NewImm2--; NewImm1++; } } } // We cannot merge two compares if the immediates are not same. if (NewImm2 != NewImm1) continue; } LLVM_DEBUG(dbgs() << "Optimize two pairs of compare and branch:\n"); LLVM_DEBUG(CMPI1->dump()); LLVM_DEBUG(BI1->dump()); LLVM_DEBUG(CMPI2->dump()); LLVM_DEBUG(BI2->dump()); // We adjust opcode, predicates and immediate as we determined above. if (NewOpCode != 0 && NewOpCode != CMPI1->getOpcode()) { CMPI1->setDesc(TII->get(NewOpCode)); } if (NewPredicate1) { BI1->getOperand(0).setImm(NewPredicate1); } if (NewPredicate2) { BI2->getOperand(0).setImm(NewPredicate2); } if (NewImm1 != Imm1) { CMPI1->getOperand(2).setImm(NewImm1); } if (IsPartiallyRedundant) { // We touch up the compare instruction in MBB2 and move it to // a previous BB to handle partially redundant case. if (SwapOperands) { Register Op1 = CMPI2->getOperand(1).getReg(); Register Op2 = CMPI2->getOperand(2).getReg(); CMPI2->getOperand(1).setReg(Op2); CMPI2->getOperand(2).setReg(Op1); } if (NewImm2 != Imm2) CMPI2->getOperand(2).setImm(NewImm2); for (int I = 1; I <= 2; I++) { if (CMPI2->getOperand(I).isReg()) { MachineInstr *Inst = MRI->getVRegDef(CMPI2->getOperand(I).getReg()); if (Inst->getParent() != &MBB2) continue; assert(Inst->getOpcode() == PPC::PHI && "We cannot support if an operand comes from this BB."); unsigned SrcReg = getIncomingRegForBlock(Inst, MBBtoMoveCmp); CMPI2->getOperand(I).setReg(SrcReg); } } auto I = MachineBasicBlock::iterator(MBBtoMoveCmp->getFirstTerminator()); MBBtoMoveCmp->splice(I, &MBB2, MachineBasicBlock::iterator(CMPI2)); DebugLoc DL = CMPI2->getDebugLoc(); Register NewVReg = MRI->createVirtualRegister(&PPC::CRRCRegClass); BuildMI(MBB2, MBB2.begin(), DL, TII->get(PPC::PHI), NewVReg) .addReg(BI1->getOperand(1).getReg()).addMBB(MBB1) .addReg(BI2->getOperand(1).getReg()).addMBB(MBBtoMoveCmp); BI2->getOperand(1).setReg(NewVReg); } else { // We finally eliminate compare instruction in MBB2. BI2->getOperand(1).setReg(BI1->getOperand(1).getReg()); CMPI2->eraseFromParent(); } BI2->getOperand(1).setIsKill(true); BI1->getOperand(1).setIsKill(false); LLVM_DEBUG(dbgs() << "into a compare and two branches:\n"); LLVM_DEBUG(CMPI1->dump()); LLVM_DEBUG(BI1->dump()); LLVM_DEBUG(BI2->dump()); if (IsPartiallyRedundant) { LLVM_DEBUG(dbgs() << "The following compare is moved into " << printMBBReference(*MBBtoMoveCmp) << " to handle partial redundancy.\n"); LLVM_DEBUG(CMPI2->dump()); } Simplified = true; } return Simplified; } // We miss the opportunity to emit an RLDIC when lowering jump tables // since ISEL sees only a single basic block. When selecting, the clear // and shift left will be in different blocks. bool PPCMIPeephole::emitRLDICWhenLoweringJumpTables(MachineInstr &MI) { if (MI.getOpcode() != PPC::RLDICR) return false; Register SrcReg = MI.getOperand(1).getReg(); if (!Register::isVirtualRegister(SrcReg)) return false; MachineInstr *SrcMI = MRI->getVRegDef(SrcReg); if (SrcMI->getOpcode() != PPC::RLDICL) return false; MachineOperand MOpSHSrc = SrcMI->getOperand(2); MachineOperand MOpMBSrc = SrcMI->getOperand(3); MachineOperand MOpSHMI = MI.getOperand(2); MachineOperand MOpMEMI = MI.getOperand(3); if (!(MOpSHSrc.isImm() && MOpMBSrc.isImm() && MOpSHMI.isImm() && MOpMEMI.isImm())) return false; uint64_t SHSrc = MOpSHSrc.getImm(); uint64_t MBSrc = MOpMBSrc.getImm(); uint64_t SHMI = MOpSHMI.getImm(); uint64_t MEMI = MOpMEMI.getImm(); uint64_t NewSH = SHSrc + SHMI; uint64_t NewMB = MBSrc - SHMI; if (NewMB > 63 || NewSH > 63) return false; // The bits cleared with RLDICL are [0, MBSrc). // The bits cleared with RLDICR are (MEMI, 63]. // After the sequence, the bits cleared are: // [0, MBSrc-SHMI) and (MEMI, 63). // // The bits cleared with RLDIC are [0, NewMB) and (63-NewSH, 63]. if ((63 - NewSH) != MEMI) return false; LLVM_DEBUG(dbgs() << "Converting pair: "); LLVM_DEBUG(SrcMI->dump()); LLVM_DEBUG(MI.dump()); MI.setDesc(TII->get(PPC::RLDIC)); MI.getOperand(1).setReg(SrcMI->getOperand(1).getReg()); MI.getOperand(2).setImm(NewSH); MI.getOperand(3).setImm(NewMB); MI.getOperand(1).setIsKill(SrcMI->getOperand(1).isKill()); SrcMI->getOperand(1).setIsKill(false); LLVM_DEBUG(dbgs() << "To: "); LLVM_DEBUG(MI.dump()); NumRotatesCollapsed++; // If SrcReg has no non-debug use it's safe to delete its def SrcMI. if (MRI->use_nodbg_empty(SrcReg)) { assert(!SrcMI->hasImplicitDef() && "Not expecting an implicit def with this instr."); SrcMI->eraseFromParent(); } return true; } // For case in LLVM IR // entry: // %iconv = sext i32 %index to i64 // br i1 undef label %true, label %false // true: // %ptr = getelementptr inbounds i32, i32* null, i64 %iconv // ... // PPCISelLowering::combineSHL fails to combine, because sext and shl are in // different BBs when conducting instruction selection. We can do a peephole // optimization to combine these two instructions into extswsli after // instruction selection. bool PPCMIPeephole::combineSEXTAndSHL(MachineInstr &MI, MachineInstr *&ToErase) { if (MI.getOpcode() != PPC::RLDICR) return false; if (!MF->getSubtarget().isISA3_0()) return false; assert(MI.getNumOperands() == 4 && "RLDICR should have 4 operands"); MachineOperand MOpSHMI = MI.getOperand(2); MachineOperand MOpMEMI = MI.getOperand(3); if (!(MOpSHMI.isImm() && MOpMEMI.isImm())) return false; uint64_t SHMI = MOpSHMI.getImm(); uint64_t MEMI = MOpMEMI.getImm(); if (SHMI + MEMI != 63) return false; Register SrcReg = MI.getOperand(1).getReg(); if (!Register::isVirtualRegister(SrcReg)) return false; MachineInstr *SrcMI = MRI->getVRegDef(SrcReg); if (SrcMI->getOpcode() != PPC::EXTSW && SrcMI->getOpcode() != PPC::EXTSW_32_64) return false; // If the register defined by extsw has more than one use, combination is not // needed. if (!MRI->hasOneNonDBGUse(SrcReg)) return false; assert(SrcMI->getNumOperands() == 2 && "EXTSW should have 2 operands"); assert(SrcMI->getOperand(1).isReg() && "EXTSW's second operand should be a register"); if (!Register::isVirtualRegister(SrcMI->getOperand(1).getReg())) return false; LLVM_DEBUG(dbgs() << "Combining pair: "); LLVM_DEBUG(SrcMI->dump()); LLVM_DEBUG(MI.dump()); MachineInstr *NewInstr = BuildMI(*MI.getParent(), &MI, MI.getDebugLoc(), SrcMI->getOpcode() == PPC::EXTSW ? TII->get(PPC::EXTSWSLI) : TII->get(PPC::EXTSWSLI_32_64), MI.getOperand(0).getReg()) .add(SrcMI->getOperand(1)) .add(MOpSHMI); (void)NewInstr; LLVM_DEBUG(dbgs() << "TO: "); LLVM_DEBUG(NewInstr->dump()); ++NumEXTSWAndSLDICombined; ToErase = &MI; // SrcMI, which is extsw, is of no use now, erase it. SrcMI->eraseFromParent(); return true; } } // end default namespace INITIALIZE_PASS_BEGIN(PPCMIPeephole, DEBUG_TYPE, "PowerPC MI Peephole Optimization", false, false) INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo) INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree) INITIALIZE_PASS_END(PPCMIPeephole, DEBUG_TYPE, "PowerPC MI Peephole Optimization", false, false) char PPCMIPeephole::ID = 0; FunctionPass* llvm::createPPCMIPeepholePass() { return new PPCMIPeephole(); }