//===- HexagonBitSimplify.cpp ---------------------------------------------===// // // 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 // //===----------------------------------------------------------------------===// #include "BitTracker.h" #include "HexagonBitTracker.h" #include "HexagonInstrInfo.h" #include "HexagonRegisterInfo.h" #include "HexagonSubtarget.h" #include "llvm/ADT/BitVector.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/GraphTraits.h" #include "llvm/ADT/STLExtras.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/StringRef.h" #include "llvm/CodeGen/MachineBasicBlock.h" #include "llvm/CodeGen/MachineDominators.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineInstr.h" #include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/CodeGen/MachineOperand.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/TargetRegisterInfo.h" #include "llvm/IR/DebugLoc.h" #include "llvm/InitializePasses.h" #include "llvm/MC/MCInstrDesc.h" #include "llvm/Pass.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/MathExtras.h" #include "llvm/Support/raw_ostream.h" #include #include #include #include #include #include #include #include #define DEBUG_TYPE "hexbit" using namespace llvm; static cl::opt PreserveTiedOps("hexbit-keep-tied", cl::Hidden, cl::init(true), cl::desc("Preserve subregisters in tied operands")); static cl::opt GenExtract("hexbit-extract", cl::Hidden, cl::init(true), cl::desc("Generate extract instructions")); static cl::opt GenBitSplit("hexbit-bitsplit", cl::Hidden, cl::init(true), cl::desc("Generate bitsplit instructions")); static cl::opt MaxExtract("hexbit-max-extract", cl::Hidden, cl::init(std::numeric_limits::max())); static unsigned CountExtract = 0; static cl::opt MaxBitSplit("hexbit-max-bitsplit", cl::Hidden, cl::init(std::numeric_limits::max())); static unsigned CountBitSplit = 0; static cl::opt RegisterSetLimit("hexbit-registerset-limit", cl::Hidden, cl::init(1000)); namespace llvm { void initializeHexagonBitSimplifyPass(PassRegistry& Registry); FunctionPass *createHexagonBitSimplify(); } // end namespace llvm namespace { // Set of virtual registers, based on BitVector. struct RegisterSet { RegisterSet() = default; explicit RegisterSet(unsigned s, bool t = false) : Bits(s, t) {} RegisterSet(const RegisterSet &RS) = default; void clear() { Bits.clear(); LRU.clear(); } unsigned count() const { return Bits.count(); } unsigned find_first() const { int First = Bits.find_first(); if (First < 0) return 0; return x2v(First); } unsigned find_next(unsigned Prev) const { int Next = Bits.find_next(v2x(Prev)); if (Next < 0) return 0; return x2v(Next); } RegisterSet &insert(unsigned R) { unsigned Idx = v2x(R); ensure(Idx); bool Exists = Bits.test(Idx); Bits.set(Idx); if (!Exists) { LRU.push_back(Idx); if (LRU.size() > RegisterSetLimit) { unsigned T = LRU.front(); Bits.reset(T); LRU.pop_front(); } } return *this; } RegisterSet &remove(unsigned R) { unsigned Idx = v2x(R); if (Idx < Bits.size()) { bool Exists = Bits.test(Idx); Bits.reset(Idx); if (Exists) { auto F = llvm::find(LRU, Idx); assert(F != LRU.end()); LRU.erase(F); } } return *this; } RegisterSet &insert(const RegisterSet &Rs) { for (unsigned R = Rs.find_first(); R; R = Rs.find_next(R)) insert(R); return *this; } RegisterSet &remove(const RegisterSet &Rs) { for (unsigned R = Rs.find_first(); R; R = Rs.find_next(R)) remove(R); return *this; } bool operator[](unsigned R) const { unsigned Idx = v2x(R); return Idx < Bits.size() ? Bits[Idx] : false; } bool has(unsigned R) const { unsigned Idx = v2x(R); if (Idx >= Bits.size()) return false; return Bits.test(Idx); } bool empty() const { return !Bits.any(); } bool includes(const RegisterSet &Rs) const { // A.test(B) <=> A-B != {} return !Rs.Bits.test(Bits); } bool intersects(const RegisterSet &Rs) const { return Bits.anyCommon(Rs.Bits); } private: BitVector Bits; std::deque LRU; void ensure(unsigned Idx) { if (Bits.size() <= Idx) Bits.resize(std::max(Idx+1, 32U)); } static inline unsigned v2x(unsigned v) { return Register::virtReg2Index(v); } static inline unsigned x2v(unsigned x) { return Register::index2VirtReg(x); } }; struct PrintRegSet { PrintRegSet(const RegisterSet &S, const TargetRegisterInfo *RI) : RS(S), TRI(RI) {} friend raw_ostream &operator<< (raw_ostream &OS, const PrintRegSet &P); private: const RegisterSet &RS; const TargetRegisterInfo *TRI; }; raw_ostream &operator<< (raw_ostream &OS, const PrintRegSet &P) LLVM_ATTRIBUTE_UNUSED; raw_ostream &operator<< (raw_ostream &OS, const PrintRegSet &P) { OS << '{'; for (unsigned R = P.RS.find_first(); R; R = P.RS.find_next(R)) OS << ' ' << printReg(R, P.TRI); OS << " }"; return OS; } class Transformation; class HexagonBitSimplify : public MachineFunctionPass { public: static char ID; HexagonBitSimplify() : MachineFunctionPass(ID) {} StringRef getPassName() const override { return "Hexagon bit simplification"; } void getAnalysisUsage(AnalysisUsage &AU) const override { AU.addRequired(); AU.addPreserved(); MachineFunctionPass::getAnalysisUsage(AU); } bool runOnMachineFunction(MachineFunction &MF) override; static void getInstrDefs(const MachineInstr &MI, RegisterSet &Defs); static void getInstrUses(const MachineInstr &MI, RegisterSet &Uses); static bool isEqual(const BitTracker::RegisterCell &RC1, uint16_t B1, const BitTracker::RegisterCell &RC2, uint16_t B2, uint16_t W); static bool isZero(const BitTracker::RegisterCell &RC, uint16_t B, uint16_t W); static bool getConst(const BitTracker::RegisterCell &RC, uint16_t B, uint16_t W, uint64_t &U); static bool replaceReg(Register OldR, Register NewR, MachineRegisterInfo &MRI); static bool getSubregMask(const BitTracker::RegisterRef &RR, unsigned &Begin, unsigned &Width, MachineRegisterInfo &MRI); static bool replaceRegWithSub(Register OldR, Register NewR, unsigned NewSR, MachineRegisterInfo &MRI); static bool replaceSubWithSub(Register OldR, unsigned OldSR, Register NewR, unsigned NewSR, MachineRegisterInfo &MRI); static bool parseRegSequence(const MachineInstr &I, BitTracker::RegisterRef &SL, BitTracker::RegisterRef &SH, const MachineRegisterInfo &MRI); static bool getUsedBitsInStore(unsigned Opc, BitVector &Bits, uint16_t Begin); static bool getUsedBits(unsigned Opc, unsigned OpN, BitVector &Bits, uint16_t Begin, const HexagonInstrInfo &HII); static const TargetRegisterClass *getFinalVRegClass( const BitTracker::RegisterRef &RR, MachineRegisterInfo &MRI); static bool isTransparentCopy(const BitTracker::RegisterRef &RD, const BitTracker::RegisterRef &RS, MachineRegisterInfo &MRI); private: MachineDominatorTree *MDT = nullptr; bool visitBlock(MachineBasicBlock &B, Transformation &T, RegisterSet &AVs); static bool hasTiedUse(unsigned Reg, MachineRegisterInfo &MRI, unsigned NewSub = Hexagon::NoSubRegister); }; using HBS = HexagonBitSimplify; // The purpose of this class is to provide a common facility to traverse // the function top-down or bottom-up via the dominator tree, and keep // track of the available registers. class Transformation { public: bool TopDown; Transformation(bool TD) : TopDown(TD) {} virtual ~Transformation() = default; virtual bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) = 0; }; } // end anonymous namespace char HexagonBitSimplify::ID = 0; INITIALIZE_PASS_BEGIN(HexagonBitSimplify, "hexagon-bit-simplify", "Hexagon bit simplification", false, false) INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) INITIALIZE_PASS_END(HexagonBitSimplify, "hexagon-bit-simplify", "Hexagon bit simplification", false, false) bool HexagonBitSimplify::visitBlock(MachineBasicBlock &B, Transformation &T, RegisterSet &AVs) { bool Changed = false; if (T.TopDown) Changed = T.processBlock(B, AVs); RegisterSet Defs; for (auto &I : B) getInstrDefs(I, Defs); RegisterSet NewAVs = AVs; NewAVs.insert(Defs); for (auto *DTN : children(MDT->getNode(&B))) Changed |= visitBlock(*(DTN->getBlock()), T, NewAVs); if (!T.TopDown) Changed |= T.processBlock(B, AVs); return Changed; } // // Utility functions: // void HexagonBitSimplify::getInstrDefs(const MachineInstr &MI, RegisterSet &Defs) { for (auto &Op : MI.operands()) { if (!Op.isReg() || !Op.isDef()) continue; Register R = Op.getReg(); if (!R.isVirtual()) continue; Defs.insert(R); } } void HexagonBitSimplify::getInstrUses(const MachineInstr &MI, RegisterSet &Uses) { for (auto &Op : MI.operands()) { if (!Op.isReg() || !Op.isUse()) continue; Register R = Op.getReg(); if (!R.isVirtual()) continue; Uses.insert(R); } } // Check if all the bits in range [B, E) in both cells are equal. bool HexagonBitSimplify::isEqual(const BitTracker::RegisterCell &RC1, uint16_t B1, const BitTracker::RegisterCell &RC2, uint16_t B2, uint16_t W) { for (uint16_t i = 0; i < W; ++i) { // If RC1[i] is "bottom", it cannot be proven equal to RC2[i]. if (RC1[B1+i].Type == BitTracker::BitValue::Ref && RC1[B1+i].RefI.Reg == 0) return false; // Same for RC2[i]. if (RC2[B2+i].Type == BitTracker::BitValue::Ref && RC2[B2+i].RefI.Reg == 0) return false; if (RC1[B1+i] != RC2[B2+i]) return false; } return true; } bool HexagonBitSimplify::isZero(const BitTracker::RegisterCell &RC, uint16_t B, uint16_t W) { assert(B < RC.width() && B+W <= RC.width()); for (uint16_t i = B; i < B+W; ++i) if (!RC[i].is(0)) return false; return true; } bool HexagonBitSimplify::getConst(const BitTracker::RegisterCell &RC, uint16_t B, uint16_t W, uint64_t &U) { assert(B < RC.width() && B+W <= RC.width()); int64_t T = 0; for (uint16_t i = B+W; i > B; --i) { const BitTracker::BitValue &BV = RC[i-1]; T <<= 1; if (BV.is(1)) T |= 1; else if (!BV.is(0)) return false; } U = T; return true; } bool HexagonBitSimplify::replaceReg(Register OldR, Register NewR, MachineRegisterInfo &MRI) { if (!OldR.isVirtual() || !NewR.isVirtual()) return false; auto Begin = MRI.use_begin(OldR), End = MRI.use_end(); decltype(End) NextI; for (auto I = Begin; I != End; I = NextI) { NextI = std::next(I); I->setReg(NewR); } return Begin != End; } bool HexagonBitSimplify::replaceRegWithSub(Register OldR, Register NewR, unsigned NewSR, MachineRegisterInfo &MRI) { if (!OldR.isVirtual() || !NewR.isVirtual()) return false; if (hasTiedUse(OldR, MRI, NewSR)) return false; auto Begin = MRI.use_begin(OldR), End = MRI.use_end(); decltype(End) NextI; for (auto I = Begin; I != End; I = NextI) { NextI = std::next(I); I->setReg(NewR); I->setSubReg(NewSR); } return Begin != End; } bool HexagonBitSimplify::replaceSubWithSub(Register OldR, unsigned OldSR, Register NewR, unsigned NewSR, MachineRegisterInfo &MRI) { if (!OldR.isVirtual() || !NewR.isVirtual()) return false; if (OldSR != NewSR && hasTiedUse(OldR, MRI, NewSR)) return false; auto Begin = MRI.use_begin(OldR), End = MRI.use_end(); decltype(End) NextI; for (auto I = Begin; I != End; I = NextI) { NextI = std::next(I); if (I->getSubReg() != OldSR) continue; I->setReg(NewR); I->setSubReg(NewSR); } return Begin != End; } // For a register ref (pair Reg:Sub), set Begin to the position of the LSB // of Sub in Reg, and set Width to the size of Sub in bits. Return true, // if this succeeded, otherwise return false. bool HexagonBitSimplify::getSubregMask(const BitTracker::RegisterRef &RR, unsigned &Begin, unsigned &Width, MachineRegisterInfo &MRI) { const TargetRegisterClass *RC = MRI.getRegClass(RR.Reg); if (RR.Sub == 0) { Begin = 0; Width = MRI.getTargetRegisterInfo()->getRegSizeInBits(*RC); return true; } Begin = 0; switch (RC->getID()) { case Hexagon::DoubleRegsRegClassID: case Hexagon::HvxWRRegClassID: Width = MRI.getTargetRegisterInfo()->getRegSizeInBits(*RC) / 2; if (RR.Sub == Hexagon::isub_hi || RR.Sub == Hexagon::vsub_hi) Begin = Width; break; default: return false; } return true; } // For a REG_SEQUENCE, set SL to the low subregister and SH to the high // subregister. bool HexagonBitSimplify::parseRegSequence(const MachineInstr &I, BitTracker::RegisterRef &SL, BitTracker::RegisterRef &SH, const MachineRegisterInfo &MRI) { assert(I.getOpcode() == TargetOpcode::REG_SEQUENCE); unsigned Sub1 = I.getOperand(2).getImm(), Sub2 = I.getOperand(4).getImm(); auto &DstRC = *MRI.getRegClass(I.getOperand(0).getReg()); auto &HRI = static_cast( *MRI.getTargetRegisterInfo()); unsigned SubLo = HRI.getHexagonSubRegIndex(DstRC, Hexagon::ps_sub_lo); unsigned SubHi = HRI.getHexagonSubRegIndex(DstRC, Hexagon::ps_sub_hi); assert((Sub1 == SubLo && Sub2 == SubHi) || (Sub1 == SubHi && Sub2 == SubLo)); if (Sub1 == SubLo && Sub2 == SubHi) { SL = I.getOperand(1); SH = I.getOperand(3); return true; } if (Sub1 == SubHi && Sub2 == SubLo) { SH = I.getOperand(1); SL = I.getOperand(3); return true; } return false; } // All stores (except 64-bit stores) take a 32-bit register as the source // of the value to be stored. If the instruction stores into a location // that is shorter than 32 bits, some bits of the source register are not // used. For each store instruction, calculate the set of used bits in // the source register, and set appropriate bits in Bits. Return true if // the bits are calculated, false otherwise. bool HexagonBitSimplify::getUsedBitsInStore(unsigned Opc, BitVector &Bits, uint16_t Begin) { using namespace Hexagon; switch (Opc) { // Store byte case S2_storerb_io: // memb(Rs32+#s11:0)=Rt32 case S2_storerbnew_io: // memb(Rs32+#s11:0)=Nt8.new case S2_pstorerbt_io: // if (Pv4) memb(Rs32+#u6:0)=Rt32 case S2_pstorerbf_io: // if (!Pv4) memb(Rs32+#u6:0)=Rt32 case S4_pstorerbtnew_io: // if (Pv4.new) memb(Rs32+#u6:0)=Rt32 case S4_pstorerbfnew_io: // if (!Pv4.new) memb(Rs32+#u6:0)=Rt32 case S2_pstorerbnewt_io: // if (Pv4) memb(Rs32+#u6:0)=Nt8.new case S2_pstorerbnewf_io: // if (!Pv4) memb(Rs32+#u6:0)=Nt8.new case S4_pstorerbnewtnew_io: // if (Pv4.new) memb(Rs32+#u6:0)=Nt8.new case S4_pstorerbnewfnew_io: // if (!Pv4.new) memb(Rs32+#u6:0)=Nt8.new case S2_storerb_pi: // memb(Rx32++#s4:0)=Rt32 case S2_storerbnew_pi: // memb(Rx32++#s4:0)=Nt8.new case S2_pstorerbt_pi: // if (Pv4) memb(Rx32++#s4:0)=Rt32 case S2_pstorerbf_pi: // if (!Pv4) memb(Rx32++#s4:0)=Rt32 case S2_pstorerbtnew_pi: // if (Pv4.new) memb(Rx32++#s4:0)=Rt32 case S2_pstorerbfnew_pi: // if (!Pv4.new) memb(Rx32++#s4:0)=Rt32 case S2_pstorerbnewt_pi: // if (Pv4) memb(Rx32++#s4:0)=Nt8.new case S2_pstorerbnewf_pi: // if (!Pv4) memb(Rx32++#s4:0)=Nt8.new case S2_pstorerbnewtnew_pi: // if (Pv4.new) memb(Rx32++#s4:0)=Nt8.new case S2_pstorerbnewfnew_pi: // if (!Pv4.new) memb(Rx32++#s4:0)=Nt8.new case S4_storerb_ap: // memb(Re32=#U6)=Rt32 case S4_storerbnew_ap: // memb(Re32=#U6)=Nt8.new case S2_storerb_pr: // memb(Rx32++Mu2)=Rt32 case S2_storerbnew_pr: // memb(Rx32++Mu2)=Nt8.new case S4_storerb_ur: // memb(Ru32<<#u2+#U6)=Rt32 case S4_storerbnew_ur: // memb(Ru32<<#u2+#U6)=Nt8.new case S2_storerb_pbr: // memb(Rx32++Mu2:brev)=Rt32 case S2_storerbnew_pbr: // memb(Rx32++Mu2:brev)=Nt8.new case S2_storerb_pci: // memb(Rx32++#s4:0:circ(Mu2))=Rt32 case S2_storerbnew_pci: // memb(Rx32++#s4:0:circ(Mu2))=Nt8.new case S2_storerb_pcr: // memb(Rx32++I:circ(Mu2))=Rt32 case S2_storerbnew_pcr: // memb(Rx32++I:circ(Mu2))=Nt8.new case S4_storerb_rr: // memb(Rs32+Ru32<<#u2)=Rt32 case S4_storerbnew_rr: // memb(Rs32+Ru32<<#u2)=Nt8.new case S4_pstorerbt_rr: // if (Pv4) memb(Rs32+Ru32<<#u2)=Rt32 case S4_pstorerbf_rr: // if (!Pv4) memb(Rs32+Ru32<<#u2)=Rt32 case S4_pstorerbtnew_rr: // if (Pv4.new) memb(Rs32+Ru32<<#u2)=Rt32 case S4_pstorerbfnew_rr: // if (!Pv4.new) memb(Rs32+Ru32<<#u2)=Rt32 case S4_pstorerbnewt_rr: // if (Pv4) memb(Rs32+Ru32<<#u2)=Nt8.new case S4_pstorerbnewf_rr: // if (!Pv4) memb(Rs32+Ru32<<#u2)=Nt8.new case S4_pstorerbnewtnew_rr: // if (Pv4.new) memb(Rs32+Ru32<<#u2)=Nt8.new case S4_pstorerbnewfnew_rr: // if (!Pv4.new) memb(Rs32+Ru32<<#u2)=Nt8.new case S2_storerbgp: // memb(gp+#u16:0)=Rt32 case S2_storerbnewgp: // memb(gp+#u16:0)=Nt8.new case S4_pstorerbt_abs: // if (Pv4) memb(#u6)=Rt32 case S4_pstorerbf_abs: // if (!Pv4) memb(#u6)=Rt32 case S4_pstorerbtnew_abs: // if (Pv4.new) memb(#u6)=Rt32 case S4_pstorerbfnew_abs: // if (!Pv4.new) memb(#u6)=Rt32 case S4_pstorerbnewt_abs: // if (Pv4) memb(#u6)=Nt8.new case S4_pstorerbnewf_abs: // if (!Pv4) memb(#u6)=Nt8.new case S4_pstorerbnewtnew_abs: // if (Pv4.new) memb(#u6)=Nt8.new case S4_pstorerbnewfnew_abs: // if (!Pv4.new) memb(#u6)=Nt8.new Bits.set(Begin, Begin+8); return true; // Store low half case S2_storerh_io: // memh(Rs32+#s11:1)=Rt32 case S2_storerhnew_io: // memh(Rs32+#s11:1)=Nt8.new case S2_pstorerht_io: // if (Pv4) memh(Rs32+#u6:1)=Rt32 case S2_pstorerhf_io: // if (!Pv4) memh(Rs32+#u6:1)=Rt32 case S4_pstorerhtnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Rt32 case S4_pstorerhfnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Rt32 case S2_pstorerhnewt_io: // if (Pv4) memh(Rs32+#u6:1)=Nt8.new case S2_pstorerhnewf_io: // if (!Pv4) memh(Rs32+#u6:1)=Nt8.new case S4_pstorerhnewtnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Nt8.new case S4_pstorerhnewfnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Nt8.new case S2_storerh_pi: // memh(Rx32++#s4:1)=Rt32 case S2_storerhnew_pi: // memh(Rx32++#s4:1)=Nt8.new case S2_pstorerht_pi: // if (Pv4) memh(Rx32++#s4:1)=Rt32 case S2_pstorerhf_pi: // if (!Pv4) memh(Rx32++#s4:1)=Rt32 case S2_pstorerhtnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Rt32 case S2_pstorerhfnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Rt32 case S2_pstorerhnewt_pi: // if (Pv4) memh(Rx32++#s4:1)=Nt8.new case S2_pstorerhnewf_pi: // if (!Pv4) memh(Rx32++#s4:1)=Nt8.new case S2_pstorerhnewtnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Nt8.new case S2_pstorerhnewfnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Nt8.new case S4_storerh_ap: // memh(Re32=#U6)=Rt32 case S4_storerhnew_ap: // memh(Re32=#U6)=Nt8.new case S2_storerh_pr: // memh(Rx32++Mu2)=Rt32 case S2_storerhnew_pr: // memh(Rx32++Mu2)=Nt8.new case S4_storerh_ur: // memh(Ru32<<#u2+#U6)=Rt32 case S4_storerhnew_ur: // memh(Ru32<<#u2+#U6)=Nt8.new case S2_storerh_pbr: // memh(Rx32++Mu2:brev)=Rt32 case S2_storerhnew_pbr: // memh(Rx32++Mu2:brev)=Nt8.new case S2_storerh_pci: // memh(Rx32++#s4:1:circ(Mu2))=Rt32 case S2_storerhnew_pci: // memh(Rx32++#s4:1:circ(Mu2))=Nt8.new case S2_storerh_pcr: // memh(Rx32++I:circ(Mu2))=Rt32 case S2_storerhnew_pcr: // memh(Rx32++I:circ(Mu2))=Nt8.new case S4_storerh_rr: // memh(Rs32+Ru32<<#u2)=Rt32 case S4_pstorerht_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Rt32 case S4_pstorerhf_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Rt32 case S4_pstorerhtnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Rt32 case S4_pstorerhfnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Rt32 case S4_storerhnew_rr: // memh(Rs32+Ru32<<#u2)=Nt8.new case S4_pstorerhnewt_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Nt8.new case S4_pstorerhnewf_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Nt8.new case S4_pstorerhnewtnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Nt8.new case S4_pstorerhnewfnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Nt8.new case S2_storerhgp: // memh(gp+#u16:1)=Rt32 case S2_storerhnewgp: // memh(gp+#u16:1)=Nt8.new case S4_pstorerht_abs: // if (Pv4) memh(#u6)=Rt32 case S4_pstorerhf_abs: // if (!Pv4) memh(#u6)=Rt32 case S4_pstorerhtnew_abs: // if (Pv4.new) memh(#u6)=Rt32 case S4_pstorerhfnew_abs: // if (!Pv4.new) memh(#u6)=Rt32 case S4_pstorerhnewt_abs: // if (Pv4) memh(#u6)=Nt8.new case S4_pstorerhnewf_abs: // if (!Pv4) memh(#u6)=Nt8.new case S4_pstorerhnewtnew_abs: // if (Pv4.new) memh(#u6)=Nt8.new case S4_pstorerhnewfnew_abs: // if (!Pv4.new) memh(#u6)=Nt8.new Bits.set(Begin, Begin+16); return true; // Store high half case S2_storerf_io: // memh(Rs32+#s11:1)=Rt.H32 case S2_pstorerft_io: // if (Pv4) memh(Rs32+#u6:1)=Rt.H32 case S2_pstorerff_io: // if (!Pv4) memh(Rs32+#u6:1)=Rt.H32 case S4_pstorerftnew_io: // if (Pv4.new) memh(Rs32+#u6:1)=Rt.H32 case S4_pstorerffnew_io: // if (!Pv4.new) memh(Rs32+#u6:1)=Rt.H32 case S2_storerf_pi: // memh(Rx32++#s4:1)=Rt.H32 case S2_pstorerft_pi: // if (Pv4) memh(Rx32++#s4:1)=Rt.H32 case S2_pstorerff_pi: // if (!Pv4) memh(Rx32++#s4:1)=Rt.H32 case S2_pstorerftnew_pi: // if (Pv4.new) memh(Rx32++#s4:1)=Rt.H32 case S2_pstorerffnew_pi: // if (!Pv4.new) memh(Rx32++#s4:1)=Rt.H32 case S4_storerf_ap: // memh(Re32=#U6)=Rt.H32 case S2_storerf_pr: // memh(Rx32++Mu2)=Rt.H32 case S4_storerf_ur: // memh(Ru32<<#u2+#U6)=Rt.H32 case S2_storerf_pbr: // memh(Rx32++Mu2:brev)=Rt.H32 case S2_storerf_pci: // memh(Rx32++#s4:1:circ(Mu2))=Rt.H32 case S2_storerf_pcr: // memh(Rx32++I:circ(Mu2))=Rt.H32 case S4_storerf_rr: // memh(Rs32+Ru32<<#u2)=Rt.H32 case S4_pstorerft_rr: // if (Pv4) memh(Rs32+Ru32<<#u2)=Rt.H32 case S4_pstorerff_rr: // if (!Pv4) memh(Rs32+Ru32<<#u2)=Rt.H32 case S4_pstorerftnew_rr: // if (Pv4.new) memh(Rs32+Ru32<<#u2)=Rt.H32 case S4_pstorerffnew_rr: // if (!Pv4.new) memh(Rs32+Ru32<<#u2)=Rt.H32 case S2_storerfgp: // memh(gp+#u16:1)=Rt.H32 case S4_pstorerft_abs: // if (Pv4) memh(#u6)=Rt.H32 case S4_pstorerff_abs: // if (!Pv4) memh(#u6)=Rt.H32 case S4_pstorerftnew_abs: // if (Pv4.new) memh(#u6)=Rt.H32 case S4_pstorerffnew_abs: // if (!Pv4.new) memh(#u6)=Rt.H32 Bits.set(Begin+16, Begin+32); return true; } return false; } // For an instruction with opcode Opc, calculate the set of bits that it // uses in a register in operand OpN. This only calculates the set of used // bits for cases where it does not depend on any operands (as is the case // in shifts, for example). For concrete instructions from a program, the // operand may be a subregister of a larger register, while Bits would // correspond to the larger register in its entirety. Because of that, // the parameter Begin can be used to indicate which bit of Bits should be // considered the LSB of the operand. bool HexagonBitSimplify::getUsedBits(unsigned Opc, unsigned OpN, BitVector &Bits, uint16_t Begin, const HexagonInstrInfo &HII) { using namespace Hexagon; const MCInstrDesc &D = HII.get(Opc); if (D.mayStore()) { if (OpN == D.getNumOperands()-1) return getUsedBitsInStore(Opc, Bits, Begin); return false; } switch (Opc) { // One register source. Used bits: R1[0-7]. case A2_sxtb: case A2_zxtb: case A4_cmpbeqi: case A4_cmpbgti: case A4_cmpbgtui: if (OpN == 1) { Bits.set(Begin, Begin+8); return true; } break; // One register source. Used bits: R1[0-15]. case A2_aslh: case A2_sxth: case A2_zxth: case A4_cmpheqi: case A4_cmphgti: case A4_cmphgtui: if (OpN == 1) { Bits.set(Begin, Begin+16); return true; } break; // One register source. Used bits: R1[16-31]. case A2_asrh: if (OpN == 1) { Bits.set(Begin+16, Begin+32); return true; } break; // Two register sources. Used bits: R1[0-7], R2[0-7]. case A4_cmpbeq: case A4_cmpbgt: case A4_cmpbgtu: if (OpN == 1) { Bits.set(Begin, Begin+8); return true; } break; // Two register sources. Used bits: R1[0-15], R2[0-15]. case A4_cmpheq: case A4_cmphgt: case A4_cmphgtu: case A2_addh_h16_ll: case A2_addh_h16_sat_ll: case A2_addh_l16_ll: case A2_addh_l16_sat_ll: case A2_combine_ll: case A2_subh_h16_ll: case A2_subh_h16_sat_ll: case A2_subh_l16_ll: case A2_subh_l16_sat_ll: case M2_mpy_acc_ll_s0: case M2_mpy_acc_ll_s1: case M2_mpy_acc_sat_ll_s0: case M2_mpy_acc_sat_ll_s1: case M2_mpy_ll_s0: case M2_mpy_ll_s1: case M2_mpy_nac_ll_s0: case M2_mpy_nac_ll_s1: case M2_mpy_nac_sat_ll_s0: case M2_mpy_nac_sat_ll_s1: case M2_mpy_rnd_ll_s0: case M2_mpy_rnd_ll_s1: case M2_mpy_sat_ll_s0: case M2_mpy_sat_ll_s1: case M2_mpy_sat_rnd_ll_s0: case M2_mpy_sat_rnd_ll_s1: case M2_mpyd_acc_ll_s0: case M2_mpyd_acc_ll_s1: case M2_mpyd_ll_s0: case M2_mpyd_ll_s1: case M2_mpyd_nac_ll_s0: case M2_mpyd_nac_ll_s1: case M2_mpyd_rnd_ll_s0: case M2_mpyd_rnd_ll_s1: case M2_mpyu_acc_ll_s0: case M2_mpyu_acc_ll_s1: case M2_mpyu_ll_s0: case M2_mpyu_ll_s1: case M2_mpyu_nac_ll_s0: case M2_mpyu_nac_ll_s1: case M2_mpyud_acc_ll_s0: case M2_mpyud_acc_ll_s1: case M2_mpyud_ll_s0: case M2_mpyud_ll_s1: case M2_mpyud_nac_ll_s0: case M2_mpyud_nac_ll_s1: if (OpN == 1 || OpN == 2) { Bits.set(Begin, Begin+16); return true; } break; // Two register sources. Used bits: R1[0-15], R2[16-31]. case A2_addh_h16_lh: case A2_addh_h16_sat_lh: case A2_combine_lh: case A2_subh_h16_lh: case A2_subh_h16_sat_lh: case M2_mpy_acc_lh_s0: case M2_mpy_acc_lh_s1: case M2_mpy_acc_sat_lh_s0: case M2_mpy_acc_sat_lh_s1: case M2_mpy_lh_s0: case M2_mpy_lh_s1: case M2_mpy_nac_lh_s0: case M2_mpy_nac_lh_s1: case M2_mpy_nac_sat_lh_s0: case M2_mpy_nac_sat_lh_s1: case M2_mpy_rnd_lh_s0: case M2_mpy_rnd_lh_s1: case M2_mpy_sat_lh_s0: case M2_mpy_sat_lh_s1: case M2_mpy_sat_rnd_lh_s0: case M2_mpy_sat_rnd_lh_s1: case M2_mpyd_acc_lh_s0: case M2_mpyd_acc_lh_s1: case M2_mpyd_lh_s0: case M2_mpyd_lh_s1: case M2_mpyd_nac_lh_s0: case M2_mpyd_nac_lh_s1: case M2_mpyd_rnd_lh_s0: case M2_mpyd_rnd_lh_s1: case M2_mpyu_acc_lh_s0: case M2_mpyu_acc_lh_s1: case M2_mpyu_lh_s0: case M2_mpyu_lh_s1: case M2_mpyu_nac_lh_s0: case M2_mpyu_nac_lh_s1: case M2_mpyud_acc_lh_s0: case M2_mpyud_acc_lh_s1: case M2_mpyud_lh_s0: case M2_mpyud_lh_s1: case M2_mpyud_nac_lh_s0: case M2_mpyud_nac_lh_s1: // These four are actually LH. case A2_addh_l16_hl: case A2_addh_l16_sat_hl: case A2_subh_l16_hl: case A2_subh_l16_sat_hl: if (OpN == 1) { Bits.set(Begin, Begin+16); return true; } if (OpN == 2) { Bits.set(Begin+16, Begin+32); return true; } break; // Two register sources, used bits: R1[16-31], R2[0-15]. case A2_addh_h16_hl: case A2_addh_h16_sat_hl: case A2_combine_hl: case A2_subh_h16_hl: case A2_subh_h16_sat_hl: case M2_mpy_acc_hl_s0: case M2_mpy_acc_hl_s1: case M2_mpy_acc_sat_hl_s0: case M2_mpy_acc_sat_hl_s1: case M2_mpy_hl_s0: case M2_mpy_hl_s1: case M2_mpy_nac_hl_s0: case M2_mpy_nac_hl_s1: case M2_mpy_nac_sat_hl_s0: case M2_mpy_nac_sat_hl_s1: case M2_mpy_rnd_hl_s0: case M2_mpy_rnd_hl_s1: case M2_mpy_sat_hl_s0: case M2_mpy_sat_hl_s1: case M2_mpy_sat_rnd_hl_s0: case M2_mpy_sat_rnd_hl_s1: case M2_mpyd_acc_hl_s0: case M2_mpyd_acc_hl_s1: case M2_mpyd_hl_s0: case M2_mpyd_hl_s1: case M2_mpyd_nac_hl_s0: case M2_mpyd_nac_hl_s1: case M2_mpyd_rnd_hl_s0: case M2_mpyd_rnd_hl_s1: case M2_mpyu_acc_hl_s0: case M2_mpyu_acc_hl_s1: case M2_mpyu_hl_s0: case M2_mpyu_hl_s1: case M2_mpyu_nac_hl_s0: case M2_mpyu_nac_hl_s1: case M2_mpyud_acc_hl_s0: case M2_mpyud_acc_hl_s1: case M2_mpyud_hl_s0: case M2_mpyud_hl_s1: case M2_mpyud_nac_hl_s0: case M2_mpyud_nac_hl_s1: if (OpN == 1) { Bits.set(Begin+16, Begin+32); return true; } if (OpN == 2) { Bits.set(Begin, Begin+16); return true; } break; // Two register sources, used bits: R1[16-31], R2[16-31]. case A2_addh_h16_hh: case A2_addh_h16_sat_hh: case A2_combine_hh: case A2_subh_h16_hh: case A2_subh_h16_sat_hh: case M2_mpy_acc_hh_s0: case M2_mpy_acc_hh_s1: case M2_mpy_acc_sat_hh_s0: case M2_mpy_acc_sat_hh_s1: case M2_mpy_hh_s0: case M2_mpy_hh_s1: case M2_mpy_nac_hh_s0: case M2_mpy_nac_hh_s1: case M2_mpy_nac_sat_hh_s0: case M2_mpy_nac_sat_hh_s1: case M2_mpy_rnd_hh_s0: case M2_mpy_rnd_hh_s1: case M2_mpy_sat_hh_s0: case M2_mpy_sat_hh_s1: case M2_mpy_sat_rnd_hh_s0: case M2_mpy_sat_rnd_hh_s1: case M2_mpyd_acc_hh_s0: case M2_mpyd_acc_hh_s1: case M2_mpyd_hh_s0: case M2_mpyd_hh_s1: case M2_mpyd_nac_hh_s0: case M2_mpyd_nac_hh_s1: case M2_mpyd_rnd_hh_s0: case M2_mpyd_rnd_hh_s1: case M2_mpyu_acc_hh_s0: case M2_mpyu_acc_hh_s1: case M2_mpyu_hh_s0: case M2_mpyu_hh_s1: case M2_mpyu_nac_hh_s0: case M2_mpyu_nac_hh_s1: case M2_mpyud_acc_hh_s0: case M2_mpyud_acc_hh_s1: case M2_mpyud_hh_s0: case M2_mpyud_hh_s1: case M2_mpyud_nac_hh_s0: case M2_mpyud_nac_hh_s1: if (OpN == 1 || OpN == 2) { Bits.set(Begin+16, Begin+32); return true; } break; } return false; } // Calculate the register class that matches Reg:Sub. For example, if // %1 is a double register, then %1:isub_hi would match the "int" // register class. const TargetRegisterClass *HexagonBitSimplify::getFinalVRegClass( const BitTracker::RegisterRef &RR, MachineRegisterInfo &MRI) { if (!RR.Reg.isVirtual()) return nullptr; auto *RC = MRI.getRegClass(RR.Reg); if (RR.Sub == 0) return RC; auto &HRI = static_cast( *MRI.getTargetRegisterInfo()); auto VerifySR = [&HRI] (const TargetRegisterClass *RC, unsigned Sub) -> void { (void)HRI; assert(Sub == HRI.getHexagonSubRegIndex(*RC, Hexagon::ps_sub_lo) || Sub == HRI.getHexagonSubRegIndex(*RC, Hexagon::ps_sub_hi)); }; switch (RC->getID()) { case Hexagon::DoubleRegsRegClassID: VerifySR(RC, RR.Sub); return &Hexagon::IntRegsRegClass; case Hexagon::HvxWRRegClassID: VerifySR(RC, RR.Sub); return &Hexagon::HvxVRRegClass; } return nullptr; } // Check if RD could be replaced with RS at any possible use of RD. // For example a predicate register cannot be replaced with a integer // register, but a 64-bit register with a subregister can be replaced // with a 32-bit register. bool HexagonBitSimplify::isTransparentCopy(const BitTracker::RegisterRef &RD, const BitTracker::RegisterRef &RS, MachineRegisterInfo &MRI) { if (!RD.Reg.isVirtual() || !RS.Reg.isVirtual()) return false; // Return false if one (or both) classes are nullptr. auto *DRC = getFinalVRegClass(RD, MRI); if (!DRC) return false; return DRC == getFinalVRegClass(RS, MRI); } bool HexagonBitSimplify::hasTiedUse(unsigned Reg, MachineRegisterInfo &MRI, unsigned NewSub) { if (!PreserveTiedOps) return false; return llvm::any_of(MRI.use_operands(Reg), [NewSub] (const MachineOperand &Op) -> bool { return Op.getSubReg() != NewSub && Op.isTied(); }); } namespace { class DeadCodeElimination { public: DeadCodeElimination(MachineFunction &mf, MachineDominatorTree &mdt) : MF(mf), HII(*MF.getSubtarget().getInstrInfo()), MDT(mdt), MRI(mf.getRegInfo()) {} bool run() { return runOnNode(MDT.getRootNode()); } private: bool isDead(unsigned R) const; bool runOnNode(MachineDomTreeNode *N); MachineFunction &MF; const HexagonInstrInfo &HII; MachineDominatorTree &MDT; MachineRegisterInfo &MRI; }; } // end anonymous namespace bool DeadCodeElimination::isDead(unsigned R) const { for (const MachineOperand &MO : MRI.use_operands(R)) { const MachineInstr *UseI = MO.getParent(); if (UseI->isDebugValue()) continue; if (UseI->isPHI()) { assert(!UseI->getOperand(0).getSubReg()); Register DR = UseI->getOperand(0).getReg(); if (DR == R) continue; } return false; } return true; } bool DeadCodeElimination::runOnNode(MachineDomTreeNode *N) { bool Changed = false; for (auto *DTN : children(N)) Changed |= runOnNode(DTN); MachineBasicBlock *B = N->getBlock(); std::vector Instrs; for (MachineInstr &MI : llvm::reverse(*B)) Instrs.push_back(&MI); for (auto *MI : Instrs) { unsigned Opc = MI->getOpcode(); // Do not touch lifetime markers. This is why the target-independent DCE // cannot be used. if (Opc == TargetOpcode::LIFETIME_START || Opc == TargetOpcode::LIFETIME_END) continue; bool Store = false; if (MI->isInlineAsm()) continue; // Delete PHIs if possible. if (!MI->isPHI() && !MI->isSafeToMove(nullptr, Store)) continue; bool AllDead = true; SmallVector Regs; for (auto &Op : MI->operands()) { if (!Op.isReg() || !Op.isDef()) continue; Register R = Op.getReg(); if (!R.isVirtual() || !isDead(R)) { AllDead = false; break; } Regs.push_back(R); } if (!AllDead) continue; B->erase(MI); for (unsigned i = 0, n = Regs.size(); i != n; ++i) MRI.markUsesInDebugValueAsUndef(Regs[i]); Changed = true; } return Changed; } namespace { // Eliminate redundant instructions // // This transformation will identify instructions where the output register // is the same as one of its input registers. This only works on instructions // that define a single register (unlike post-increment loads, for example). // The equality check is actually more detailed: the code calculates which // bits of the output are used, and only compares these bits with the input // registers. // If the output matches an input, the instruction is replaced with COPY. // The copies will be removed by another transformation. class RedundantInstrElimination : public Transformation { public: RedundantInstrElimination(BitTracker &bt, const HexagonInstrInfo &hii, const HexagonRegisterInfo &hri, MachineRegisterInfo &mri) : Transformation(true), HII(hii), HRI(hri), MRI(mri), BT(bt) {} bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; private: bool isLossyShiftLeft(const MachineInstr &MI, unsigned OpN, unsigned &LostB, unsigned &LostE); bool isLossyShiftRight(const MachineInstr &MI, unsigned OpN, unsigned &LostB, unsigned &LostE); bool computeUsedBits(unsigned Reg, BitVector &Bits); bool computeUsedBits(const MachineInstr &MI, unsigned OpN, BitVector &Bits, uint16_t Begin); bool usedBitsEqual(BitTracker::RegisterRef RD, BitTracker::RegisterRef RS); const HexagonInstrInfo &HII; const HexagonRegisterInfo &HRI; MachineRegisterInfo &MRI; BitTracker &BT; }; } // end anonymous namespace // Check if the instruction is a lossy shift left, where the input being // shifted is the operand OpN of MI. If true, [LostB, LostE) is the range // of bit indices that are lost. bool RedundantInstrElimination::isLossyShiftLeft(const MachineInstr &MI, unsigned OpN, unsigned &LostB, unsigned &LostE) { using namespace Hexagon; unsigned Opc = MI.getOpcode(); unsigned ImN, RegN, Width; switch (Opc) { case S2_asl_i_p: ImN = 2; RegN = 1; Width = 64; break; case S2_asl_i_p_acc: case S2_asl_i_p_and: case S2_asl_i_p_nac: case S2_asl_i_p_or: case S2_asl_i_p_xacc: ImN = 3; RegN = 2; Width = 64; break; case S2_asl_i_r: ImN = 2; RegN = 1; Width = 32; break; case S2_addasl_rrri: case S4_andi_asl_ri: case S4_ori_asl_ri: case S4_addi_asl_ri: case S4_subi_asl_ri: case S2_asl_i_r_acc: case S2_asl_i_r_and: case S2_asl_i_r_nac: case S2_asl_i_r_or: case S2_asl_i_r_sat: case S2_asl_i_r_xacc: ImN = 3; RegN = 2; Width = 32; break; default: return false; } if (RegN != OpN) return false; assert(MI.getOperand(ImN).isImm()); unsigned S = MI.getOperand(ImN).getImm(); if (S == 0) return false; LostB = Width-S; LostE = Width; return true; } // Check if the instruction is a lossy shift right, where the input being // shifted is the operand OpN of MI. If true, [LostB, LostE) is the range // of bit indices that are lost. bool RedundantInstrElimination::isLossyShiftRight(const MachineInstr &MI, unsigned OpN, unsigned &LostB, unsigned &LostE) { using namespace Hexagon; unsigned Opc = MI.getOpcode(); unsigned ImN, RegN; switch (Opc) { case S2_asr_i_p: case S2_lsr_i_p: ImN = 2; RegN = 1; break; case S2_asr_i_p_acc: case S2_asr_i_p_and: case S2_asr_i_p_nac: case S2_asr_i_p_or: case S2_lsr_i_p_acc: case S2_lsr_i_p_and: case S2_lsr_i_p_nac: case S2_lsr_i_p_or: case S2_lsr_i_p_xacc: ImN = 3; RegN = 2; break; case S2_asr_i_r: case S2_lsr_i_r: ImN = 2; RegN = 1; break; case S4_andi_lsr_ri: case S4_ori_lsr_ri: case S4_addi_lsr_ri: case S4_subi_lsr_ri: case S2_asr_i_r_acc: case S2_asr_i_r_and: case S2_asr_i_r_nac: case S2_asr_i_r_or: case S2_lsr_i_r_acc: case S2_lsr_i_r_and: case S2_lsr_i_r_nac: case S2_lsr_i_r_or: case S2_lsr_i_r_xacc: ImN = 3; RegN = 2; break; default: return false; } if (RegN != OpN) return false; assert(MI.getOperand(ImN).isImm()); unsigned S = MI.getOperand(ImN).getImm(); LostB = 0; LostE = S; return true; } // Calculate the bit vector that corresponds to the used bits of register Reg. // The vector Bits has the same size, as the size of Reg in bits. If the cal- // culation fails (i.e. the used bits are unknown), it returns false. Other- // wise, it returns true and sets the corresponding bits in Bits. bool RedundantInstrElimination::computeUsedBits(unsigned Reg, BitVector &Bits) { BitVector Used(Bits.size()); RegisterSet Visited; std::vector Pending; Pending.push_back(Reg); for (unsigned i = 0; i < Pending.size(); ++i) { unsigned R = Pending[i]; if (Visited.has(R)) continue; Visited.insert(R); for (auto I = MRI.use_begin(R), E = MRI.use_end(); I != E; ++I) { BitTracker::RegisterRef UR = *I; unsigned B, W; if (!HBS::getSubregMask(UR, B, W, MRI)) return false; MachineInstr &UseI = *I->getParent(); if (UseI.isPHI() || UseI.isCopy()) { Register DefR = UseI.getOperand(0).getReg(); if (!DefR.isVirtual()) return false; Pending.push_back(DefR); } else { if (!computeUsedBits(UseI, I.getOperandNo(), Used, B)) return false; } } } Bits |= Used; return true; } // Calculate the bits used by instruction MI in a register in operand OpN. // Return true/false if the calculation succeeds/fails. If is succeeds, set // used bits in Bits. This function does not reset any bits in Bits, so // subsequent calls over different instructions will result in the union // of the used bits in all these instructions. // The register in question may be used with a sub-register, whereas Bits // holds the bits for the entire register. To keep track of that, the // argument Begin indicates where in Bits is the lowest-significant bit // of the register used in operand OpN. For example, in instruction: // %1 = S2_lsr_i_r %2:isub_hi, 10 // the operand 1 is a 32-bit register, which happens to be a subregister // of the 64-bit register %2, and that subregister starts at position 32. // In this case Begin=32, since Bits[32] would be the lowest-significant bit // of %2:isub_hi. bool RedundantInstrElimination::computeUsedBits(const MachineInstr &MI, unsigned OpN, BitVector &Bits, uint16_t Begin) { unsigned Opc = MI.getOpcode(); BitVector T(Bits.size()); bool GotBits = HBS::getUsedBits(Opc, OpN, T, Begin, HII); // Even if we don't have bits yet, we could still provide some information // if the instruction is a lossy shift: the lost bits will be marked as // not used. unsigned LB, LE; if (isLossyShiftLeft(MI, OpN, LB, LE) || isLossyShiftRight(MI, OpN, LB, LE)) { assert(MI.getOperand(OpN).isReg()); BitTracker::RegisterRef RR = MI.getOperand(OpN); const TargetRegisterClass *RC = HBS::getFinalVRegClass(RR, MRI); uint16_t Width = HRI.getRegSizeInBits(*RC); if (!GotBits) T.set(Begin, Begin+Width); assert(LB <= LE && LB < Width && LE <= Width); T.reset(Begin+LB, Begin+LE); GotBits = true; } if (GotBits) Bits |= T; return GotBits; } // Calculates the used bits in RD ("defined register"), and checks if these // bits in RS ("used register") and RD are identical. bool RedundantInstrElimination::usedBitsEqual(BitTracker::RegisterRef RD, BitTracker::RegisterRef RS) { const BitTracker::RegisterCell &DC = BT.lookup(RD.Reg); const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); unsigned DB, DW; if (!HBS::getSubregMask(RD, DB, DW, MRI)) return false; unsigned SB, SW; if (!HBS::getSubregMask(RS, SB, SW, MRI)) return false; if (SW != DW) return false; BitVector Used(DC.width()); if (!computeUsedBits(RD.Reg, Used)) return false; for (unsigned i = 0; i != DW; ++i) if (Used[i+DB] && DC[DB+i] != SC[SB+i]) return false; return true; } bool RedundantInstrElimination::processBlock(MachineBasicBlock &B, const RegisterSet&) { if (!BT.reached(&B)) return false; bool Changed = false; for (auto I = B.begin(), E = B.end(); I != E; ++I) { MachineInstr *MI = &*I; if (MI->getOpcode() == TargetOpcode::COPY) continue; if (MI->isPHI() || MI->hasUnmodeledSideEffects() || MI->isInlineAsm()) continue; unsigned NumD = MI->getDesc().getNumDefs(); if (NumD != 1) continue; BitTracker::RegisterRef RD = MI->getOperand(0); if (!BT.has(RD.Reg)) continue; const BitTracker::RegisterCell &DC = BT.lookup(RD.Reg); auto At = MachineBasicBlock::iterator(MI); // Find a source operand that is equal to the result. for (auto &Op : MI->uses()) { if (!Op.isReg()) continue; BitTracker::RegisterRef RS = Op; if (!BT.has(RS.Reg)) continue; if (!HBS::isTransparentCopy(RD, RS, MRI)) continue; unsigned BN, BW; if (!HBS::getSubregMask(RS, BN, BW, MRI)) continue; const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); if (!usedBitsEqual(RD, RS) && !HBS::isEqual(DC, 0, SC, BN, BW)) continue; // If found, replace the instruction with a COPY. const DebugLoc &DL = MI->getDebugLoc(); const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); Register NewR = MRI.createVirtualRegister(FRC); MachineInstr *CopyI = BuildMI(B, At, DL, HII.get(TargetOpcode::COPY), NewR) .addReg(RS.Reg, 0, RS.Sub); HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); // This pass can create copies between registers that don't have the // exact same values. Updating the tracker has to involve updating // all dependent cells. Example: // %1 = inst %2 ; %1 != %2, but used bits are equal // // %3 = copy %2 ; <- inserted // ... = %3 ; <- replaced from %2 // Indirectly, we can create a "copy" between %1 and %2 even // though their exact values do not match. BT.visit(*CopyI); Changed = true; break; } } return Changed; } namespace { // Recognize instructions that produce constant values known at compile-time. // Replace them with register definitions that load these constants directly. class ConstGeneration : public Transformation { public: ConstGeneration(BitTracker &bt, const HexagonInstrInfo &hii, MachineRegisterInfo &mri) : Transformation(true), HII(hii), MRI(mri), BT(bt) {} bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; static bool isTfrConst(const MachineInstr &MI); private: Register genTfrConst(const TargetRegisterClass *RC, int64_t C, MachineBasicBlock &B, MachineBasicBlock::iterator At, DebugLoc &DL); const HexagonInstrInfo &HII; MachineRegisterInfo &MRI; BitTracker &BT; }; } // end anonymous namespace bool ConstGeneration::isTfrConst(const MachineInstr &MI) { unsigned Opc = MI.getOpcode(); switch (Opc) { case Hexagon::A2_combineii: case Hexagon::A4_combineii: case Hexagon::A2_tfrsi: case Hexagon::A2_tfrpi: case Hexagon::PS_true: case Hexagon::PS_false: case Hexagon::CONST32: case Hexagon::CONST64: return true; } return false; } // Generate a transfer-immediate instruction that is appropriate for the // register class and the actual value being transferred. Register ConstGeneration::genTfrConst(const TargetRegisterClass *RC, int64_t C, MachineBasicBlock &B, MachineBasicBlock::iterator At, DebugLoc &DL) { Register Reg = MRI.createVirtualRegister(RC); if (RC == &Hexagon::IntRegsRegClass) { BuildMI(B, At, DL, HII.get(Hexagon::A2_tfrsi), Reg) .addImm(int32_t(C)); return Reg; } if (RC == &Hexagon::DoubleRegsRegClass) { if (isInt<8>(C)) { BuildMI(B, At, DL, HII.get(Hexagon::A2_tfrpi), Reg) .addImm(C); return Reg; } unsigned Lo = Lo_32(C), Hi = Hi_32(C); if (isInt<8>(Lo) || isInt<8>(Hi)) { unsigned Opc = isInt<8>(Lo) ? Hexagon::A2_combineii : Hexagon::A4_combineii; BuildMI(B, At, DL, HII.get(Opc), Reg) .addImm(int32_t(Hi)) .addImm(int32_t(Lo)); return Reg; } MachineFunction *MF = B.getParent(); auto &HST = MF->getSubtarget(); // Disable CONST64 for tiny core since it takes a LD resource. if (!HST.isTinyCore() || MF->getFunction().hasOptSize()) { BuildMI(B, At, DL, HII.get(Hexagon::CONST64), Reg) .addImm(C); return Reg; } } if (RC == &Hexagon::PredRegsRegClass) { unsigned Opc; if (C == 0) Opc = Hexagon::PS_false; else if ((C & 0xFF) == 0xFF) Opc = Hexagon::PS_true; else return 0; BuildMI(B, At, DL, HII.get(Opc), Reg); return Reg; } return 0; } bool ConstGeneration::processBlock(MachineBasicBlock &B, const RegisterSet&) { if (!BT.reached(&B)) return false; bool Changed = false; RegisterSet Defs; for (auto I = B.begin(), E = B.end(); I != E; ++I) { if (isTfrConst(*I)) continue; Defs.clear(); HBS::getInstrDefs(*I, Defs); if (Defs.count() != 1) continue; Register DR = Defs.find_first(); if (!DR.isVirtual()) continue; uint64_t U; const BitTracker::RegisterCell &DRC = BT.lookup(DR); if (HBS::getConst(DRC, 0, DRC.width(), U)) { int64_t C = U; DebugLoc DL = I->getDebugLoc(); auto At = I->isPHI() ? B.getFirstNonPHI() : I; Register ImmReg = genTfrConst(MRI.getRegClass(DR), C, B, At, DL); if (ImmReg) { HBS::replaceReg(DR, ImmReg, MRI); BT.put(ImmReg, DRC); Changed = true; } } } return Changed; } namespace { // Identify pairs of available registers which hold identical values. // In such cases, only one of them needs to be calculated, the other one // will be defined as a copy of the first. class CopyGeneration : public Transformation { public: CopyGeneration(BitTracker &bt, const HexagonInstrInfo &hii, const HexagonRegisterInfo &hri, MachineRegisterInfo &mri) : Transformation(true), HII(hii), HRI(hri), MRI(mri), BT(bt) {} bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; private: bool findMatch(const BitTracker::RegisterRef &Inp, BitTracker::RegisterRef &Out, const RegisterSet &AVs); const HexagonInstrInfo &HII; const HexagonRegisterInfo &HRI; MachineRegisterInfo &MRI; BitTracker &BT; RegisterSet Forbidden; }; // Eliminate register copies RD = RS, by replacing the uses of RD with // with uses of RS. class CopyPropagation : public Transformation { public: CopyPropagation(const HexagonRegisterInfo &hri, MachineRegisterInfo &mri) : Transformation(false), HRI(hri), MRI(mri) {} bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; static bool isCopyReg(unsigned Opc, bool NoConv); private: bool propagateRegCopy(MachineInstr &MI); const HexagonRegisterInfo &HRI; MachineRegisterInfo &MRI; }; } // end anonymous namespace /// Check if there is a register in AVs that is identical to Inp. If so, /// set Out to the found register. The output may be a pair Reg:Sub. bool CopyGeneration::findMatch(const BitTracker::RegisterRef &Inp, BitTracker::RegisterRef &Out, const RegisterSet &AVs) { if (!BT.has(Inp.Reg)) return false; const BitTracker::RegisterCell &InpRC = BT.lookup(Inp.Reg); auto *FRC = HBS::getFinalVRegClass(Inp, MRI); unsigned B, W; if (!HBS::getSubregMask(Inp, B, W, MRI)) return false; for (Register R = AVs.find_first(); R; R = AVs.find_next(R)) { if (!BT.has(R) || Forbidden[R]) continue; const BitTracker::RegisterCell &RC = BT.lookup(R); unsigned RW = RC.width(); if (W == RW) { if (FRC != MRI.getRegClass(R)) continue; if (!HBS::isTransparentCopy(R, Inp, MRI)) continue; if (!HBS::isEqual(InpRC, B, RC, 0, W)) continue; Out.Reg = R; Out.Sub = 0; return true; } // Check if there is a super-register, whose part (with a subregister) // is equal to the input. // Only do double registers for now. if (W*2 != RW) continue; if (MRI.getRegClass(R) != &Hexagon::DoubleRegsRegClass) continue; if (HBS::isEqual(InpRC, B, RC, 0, W)) Out.Sub = Hexagon::isub_lo; else if (HBS::isEqual(InpRC, B, RC, W, W)) Out.Sub = Hexagon::isub_hi; else continue; Out.Reg = R; if (HBS::isTransparentCopy(Out, Inp, MRI)) return true; } return false; } bool CopyGeneration::processBlock(MachineBasicBlock &B, const RegisterSet &AVs) { if (!BT.reached(&B)) return false; RegisterSet AVB(AVs); bool Changed = false; RegisterSet Defs; for (auto I = B.begin(), E = B.end(); I != E; ++I, AVB.insert(Defs)) { Defs.clear(); HBS::getInstrDefs(*I, Defs); unsigned Opc = I->getOpcode(); if (CopyPropagation::isCopyReg(Opc, false) || ConstGeneration::isTfrConst(*I)) continue; DebugLoc DL = I->getDebugLoc(); auto At = I->isPHI() ? B.getFirstNonPHI() : I; for (Register R = Defs.find_first(); R; R = Defs.find_next(R)) { BitTracker::RegisterRef MR; auto *FRC = HBS::getFinalVRegClass(R, MRI); if (findMatch(R, MR, AVB)) { Register NewR = MRI.createVirtualRegister(FRC); BuildMI(B, At, DL, HII.get(TargetOpcode::COPY), NewR) .addReg(MR.Reg, 0, MR.Sub); BT.put(BitTracker::RegisterRef(NewR), BT.get(MR)); HBS::replaceReg(R, NewR, MRI); Forbidden.insert(R); continue; } if (FRC == &Hexagon::DoubleRegsRegClass || FRC == &Hexagon::HvxWRRegClass) { // Try to generate REG_SEQUENCE. unsigned SubLo = HRI.getHexagonSubRegIndex(*FRC, Hexagon::ps_sub_lo); unsigned SubHi = HRI.getHexagonSubRegIndex(*FRC, Hexagon::ps_sub_hi); BitTracker::RegisterRef TL = { R, SubLo }; BitTracker::RegisterRef TH = { R, SubHi }; BitTracker::RegisterRef ML, MH; if (findMatch(TL, ML, AVB) && findMatch(TH, MH, AVB)) { auto *FRC = HBS::getFinalVRegClass(R, MRI); Register NewR = MRI.createVirtualRegister(FRC); BuildMI(B, At, DL, HII.get(TargetOpcode::REG_SEQUENCE), NewR) .addReg(ML.Reg, 0, ML.Sub) .addImm(SubLo) .addReg(MH.Reg, 0, MH.Sub) .addImm(SubHi); BT.put(BitTracker::RegisterRef(NewR), BT.get(R)); HBS::replaceReg(R, NewR, MRI); Forbidden.insert(R); } } } } return Changed; } bool CopyPropagation::isCopyReg(unsigned Opc, bool NoConv) { switch (Opc) { case TargetOpcode::COPY: case TargetOpcode::REG_SEQUENCE: case Hexagon::A4_combineir: case Hexagon::A4_combineri: return true; case Hexagon::A2_tfr: case Hexagon::A2_tfrp: case Hexagon::A2_combinew: case Hexagon::V6_vcombine: return NoConv; default: break; } return false; } bool CopyPropagation::propagateRegCopy(MachineInstr &MI) { bool Changed = false; unsigned Opc = MI.getOpcode(); BitTracker::RegisterRef RD = MI.getOperand(0); assert(MI.getOperand(0).getSubReg() == 0); switch (Opc) { case TargetOpcode::COPY: case Hexagon::A2_tfr: case Hexagon::A2_tfrp: { BitTracker::RegisterRef RS = MI.getOperand(1); if (!HBS::isTransparentCopy(RD, RS, MRI)) break; if (RS.Sub != 0) Changed = HBS::replaceRegWithSub(RD.Reg, RS.Reg, RS.Sub, MRI); else Changed = HBS::replaceReg(RD.Reg, RS.Reg, MRI); break; } case TargetOpcode::REG_SEQUENCE: { BitTracker::RegisterRef SL, SH; if (HBS::parseRegSequence(MI, SL, SH, MRI)) { const TargetRegisterClass &RC = *MRI.getRegClass(RD.Reg); unsigned SubLo = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_lo); unsigned SubHi = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_hi); Changed = HBS::replaceSubWithSub(RD.Reg, SubLo, SL.Reg, SL.Sub, MRI); Changed |= HBS::replaceSubWithSub(RD.Reg, SubHi, SH.Reg, SH.Sub, MRI); } break; } case Hexagon::A2_combinew: case Hexagon::V6_vcombine: { const TargetRegisterClass &RC = *MRI.getRegClass(RD.Reg); unsigned SubLo = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_lo); unsigned SubHi = HRI.getHexagonSubRegIndex(RC, Hexagon::ps_sub_hi); BitTracker::RegisterRef RH = MI.getOperand(1), RL = MI.getOperand(2); Changed = HBS::replaceSubWithSub(RD.Reg, SubLo, RL.Reg, RL.Sub, MRI); Changed |= HBS::replaceSubWithSub(RD.Reg, SubHi, RH.Reg, RH.Sub, MRI); break; } case Hexagon::A4_combineir: case Hexagon::A4_combineri: { unsigned SrcX = (Opc == Hexagon::A4_combineir) ? 2 : 1; unsigned Sub = (Opc == Hexagon::A4_combineir) ? Hexagon::isub_lo : Hexagon::isub_hi; BitTracker::RegisterRef RS = MI.getOperand(SrcX); Changed = HBS::replaceSubWithSub(RD.Reg, Sub, RS.Reg, RS.Sub, MRI); break; } } return Changed; } bool CopyPropagation::processBlock(MachineBasicBlock &B, const RegisterSet&) { std::vector Instrs; for (MachineInstr &MI : llvm::reverse(B)) Instrs.push_back(&MI); bool Changed = false; for (auto *I : Instrs) { unsigned Opc = I->getOpcode(); if (!CopyPropagation::isCopyReg(Opc, true)) continue; Changed |= propagateRegCopy(*I); } return Changed; } namespace { // Recognize patterns that can be simplified and replace them with the // simpler forms. // This is by no means complete class BitSimplification : public Transformation { public: BitSimplification(BitTracker &bt, const MachineDominatorTree &mdt, const HexagonInstrInfo &hii, const HexagonRegisterInfo &hri, MachineRegisterInfo &mri, MachineFunction &mf) : Transformation(true), MDT(mdt), HII(hii), HRI(hri), MRI(mri), MF(mf), BT(bt) {} bool processBlock(MachineBasicBlock &B, const RegisterSet &AVs) override; private: struct RegHalf : public BitTracker::RegisterRef { bool Low; // Low/High halfword. }; bool matchHalf(unsigned SelfR, const BitTracker::RegisterCell &RC, unsigned B, RegHalf &RH); bool validateReg(BitTracker::RegisterRef R, unsigned Opc, unsigned OpNum); bool matchPackhl(unsigned SelfR, const BitTracker::RegisterCell &RC, BitTracker::RegisterRef &Rs, BitTracker::RegisterRef &Rt); unsigned getCombineOpcode(bool HLow, bool LLow); bool genStoreUpperHalf(MachineInstr *MI); bool genStoreImmediate(MachineInstr *MI); bool genPackhl(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC); bool genExtractHalf(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC); bool genCombineHalf(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC); bool genExtractLow(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC); bool genBitSplit(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC, const RegisterSet &AVs); bool simplifyTstbit(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC); bool simplifyExtractLow(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC, const RegisterSet &AVs); bool simplifyRCmp0(MachineInstr *MI, BitTracker::RegisterRef RD); // Cache of created instructions to avoid creating duplicates. // XXX Currently only used by genBitSplit. std::vector NewMIs; const MachineDominatorTree &MDT; const HexagonInstrInfo &HII; const HexagonRegisterInfo &HRI; MachineRegisterInfo &MRI; MachineFunction &MF; BitTracker &BT; }; } // end anonymous namespace // Check if the bits [B..B+16) in register cell RC form a valid halfword, // i.e. [0..16), [16..32), etc. of some register. If so, return true and // set the information about the found register in RH. bool BitSimplification::matchHalf(unsigned SelfR, const BitTracker::RegisterCell &RC, unsigned B, RegHalf &RH) { // XXX This could be searching in the set of available registers, in case // the match is not exact. // Match 16-bit chunks, where the RC[B..B+15] references exactly one // register and all the bits B..B+15 match between RC and the register. // This is meant to match "v1[0-15]", where v1 = { [0]:0 [1-15]:v1... }, // and RC = { [0]:0 [1-15]:v1[1-15]... }. bool Low = false; unsigned I = B; while (I < B+16 && RC[I].num()) I++; if (I == B+16) return false; Register Reg = RC[I].RefI.Reg; unsigned P = RC[I].RefI.Pos; // The RefI.Pos will be advanced by I-B. if (P < I-B) return false; unsigned Pos = P - (I-B); if (Reg == 0 || Reg == SelfR) // Don't match "self". return false; if (!Reg.isVirtual()) return false; if (!BT.has(Reg)) return false; const BitTracker::RegisterCell &SC = BT.lookup(Reg); if (Pos+16 > SC.width()) return false; for (unsigned i = 0; i < 16; ++i) { const BitTracker::BitValue &RV = RC[i+B]; if (RV.Type == BitTracker::BitValue::Ref) { if (RV.RefI.Reg != Reg) return false; if (RV.RefI.Pos != i+Pos) return false; continue; } if (RC[i+B] != SC[i+Pos]) return false; } unsigned Sub = 0; switch (Pos) { case 0: Sub = Hexagon::isub_lo; Low = true; break; case 16: Sub = Hexagon::isub_lo; Low = false; break; case 32: Sub = Hexagon::isub_hi; Low = true; break; case 48: Sub = Hexagon::isub_hi; Low = false; break; default: return false; } RH.Reg = Reg; RH.Sub = Sub; RH.Low = Low; // If the subregister is not valid with the register, set it to 0. if (!HBS::getFinalVRegClass(RH, MRI)) RH.Sub = 0; return true; } bool BitSimplification::validateReg(BitTracker::RegisterRef R, unsigned Opc, unsigned OpNum) { auto *OpRC = HII.getRegClass(HII.get(Opc), OpNum, &HRI, MF); auto *RRC = HBS::getFinalVRegClass(R, MRI); return OpRC->hasSubClassEq(RRC); } // Check if RC matches the pattern of a S2_packhl. If so, return true and // set the inputs Rs and Rt. bool BitSimplification::matchPackhl(unsigned SelfR, const BitTracker::RegisterCell &RC, BitTracker::RegisterRef &Rs, BitTracker::RegisterRef &Rt) { RegHalf L1, H1, L2, H2; if (!matchHalf(SelfR, RC, 0, L2) || !matchHalf(SelfR, RC, 16, L1)) return false; if (!matchHalf(SelfR, RC, 32, H2) || !matchHalf(SelfR, RC, 48, H1)) return false; // Rs = H1.L1, Rt = H2.L2 if (H1.Reg != L1.Reg || H1.Sub != L1.Sub || H1.Low || !L1.Low) return false; if (H2.Reg != L2.Reg || H2.Sub != L2.Sub || H2.Low || !L2.Low) return false; Rs = H1; Rt = H2; return true; } unsigned BitSimplification::getCombineOpcode(bool HLow, bool LLow) { return HLow ? LLow ? Hexagon::A2_combine_ll : Hexagon::A2_combine_lh : LLow ? Hexagon::A2_combine_hl : Hexagon::A2_combine_hh; } // If MI stores the upper halfword of a register (potentially obtained via // shifts or extracts), replace it with a storerf instruction. This could // cause the "extraction" code to become dead. bool BitSimplification::genStoreUpperHalf(MachineInstr *MI) { unsigned Opc = MI->getOpcode(); if (Opc != Hexagon::S2_storerh_io) return false; MachineOperand &ValOp = MI->getOperand(2); BitTracker::RegisterRef RS = ValOp; if (!BT.has(RS.Reg)) return false; const BitTracker::RegisterCell &RC = BT.lookup(RS.Reg); RegHalf H; if (!matchHalf(0, RC, 0, H)) return false; if (H.Low) return false; MI->setDesc(HII.get(Hexagon::S2_storerf_io)); ValOp.setReg(H.Reg); ValOp.setSubReg(H.Sub); return true; } // If MI stores a value known at compile-time, and the value is within a range // that avoids using constant-extenders, replace it with a store-immediate. bool BitSimplification::genStoreImmediate(MachineInstr *MI) { unsigned Opc = MI->getOpcode(); unsigned Align = 0; switch (Opc) { case Hexagon::S2_storeri_io: Align++; [[fallthrough]]; case Hexagon::S2_storerh_io: Align++; [[fallthrough]]; case Hexagon::S2_storerb_io: break; default: return false; } // Avoid stores to frame-indices (due to an unknown offset). if (!MI->getOperand(0).isReg()) return false; MachineOperand &OffOp = MI->getOperand(1); if (!OffOp.isImm()) return false; int64_t Off = OffOp.getImm(); // Offset is u6:a. Sadly, there is no isShiftedUInt(n,x). if (!isUIntN(6+Align, Off) || (Off & ((1<getOperand(2); if (!BT.has(RS.Reg)) return false; const BitTracker::RegisterCell &RC = BT.lookup(RS.Reg); uint64_t U; if (!HBS::getConst(RC, 0, RC.width(), U)) return false; // Only consider 8-bit values to avoid constant-extenders. int V; switch (Opc) { case Hexagon::S2_storerb_io: V = int8_t(U); break; case Hexagon::S2_storerh_io: V = int16_t(U); break; case Hexagon::S2_storeri_io: V = int32_t(U); break; default: // Opc is already checked above to be one of the three store instructions. // This silences a -Wuninitialized false positive on GCC 5.4. llvm_unreachable("Unexpected store opcode"); } if (!isInt<8>(V)) return false; MI->removeOperand(2); switch (Opc) { case Hexagon::S2_storerb_io: MI->setDesc(HII.get(Hexagon::S4_storeirb_io)); break; case Hexagon::S2_storerh_io: MI->setDesc(HII.get(Hexagon::S4_storeirh_io)); break; case Hexagon::S2_storeri_io: MI->setDesc(HII.get(Hexagon::S4_storeiri_io)); break; } MI->addOperand(MachineOperand::CreateImm(V)); return true; } // If MI is equivalent o S2_packhl, generate the S2_packhl. MI could be the // last instruction in a sequence that results in something equivalent to // the pack-halfwords. The intent is to cause the entire sequence to become // dead. bool BitSimplification::genPackhl(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { unsigned Opc = MI->getOpcode(); if (Opc == Hexagon::S2_packhl) return false; BitTracker::RegisterRef Rs, Rt; if (!matchPackhl(RD.Reg, RC, Rs, Rt)) return false; if (!validateReg(Rs, Hexagon::S2_packhl, 1) || !validateReg(Rt, Hexagon::S2_packhl, 2)) return false; MachineBasicBlock &B = *MI->getParent(); Register NewR = MRI.createVirtualRegister(&Hexagon::DoubleRegsRegClass); DebugLoc DL = MI->getDebugLoc(); auto At = MI->isPHI() ? B.getFirstNonPHI() : MachineBasicBlock::iterator(MI); BuildMI(B, At, DL, HII.get(Hexagon::S2_packhl), NewR) .addReg(Rs.Reg, 0, Rs.Sub) .addReg(Rt.Reg, 0, Rt.Sub); HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); BT.put(BitTracker::RegisterRef(NewR), RC); return true; } // If MI produces halfword of the input in the low half of the output, // replace it with zero-extend or extractu. bool BitSimplification::genExtractHalf(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { RegHalf L; // Check for halfword in low 16 bits, zeros elsewhere. if (!matchHalf(RD.Reg, RC, 0, L) || !HBS::isZero(RC, 16, 16)) return false; unsigned Opc = MI->getOpcode(); MachineBasicBlock &B = *MI->getParent(); DebugLoc DL = MI->getDebugLoc(); // Prefer zxth, since zxth can go in any slot, while extractu only in // slots 2 and 3. unsigned NewR = 0; auto At = MI->isPHI() ? B.getFirstNonPHI() : MachineBasicBlock::iterator(MI); if (L.Low && Opc != Hexagon::A2_zxth) { if (validateReg(L, Hexagon::A2_zxth, 1)) { NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); BuildMI(B, At, DL, HII.get(Hexagon::A2_zxth), NewR) .addReg(L.Reg, 0, L.Sub); } } else if (!L.Low && Opc != Hexagon::S2_lsr_i_r) { if (validateReg(L, Hexagon::S2_lsr_i_r, 1)) { NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); BuildMI(B, MI, DL, HII.get(Hexagon::S2_lsr_i_r), NewR) .addReg(L.Reg, 0, L.Sub) .addImm(16); } } if (NewR == 0) return false; HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); BT.put(BitTracker::RegisterRef(NewR), RC); return true; } // If MI is equivalent to a combine(.L/.H, .L/.H) replace with with the // combine. bool BitSimplification::genCombineHalf(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { RegHalf L, H; // Check for combine h/l if (!matchHalf(RD.Reg, RC, 0, L) || !matchHalf(RD.Reg, RC, 16, H)) return false; // Do nothing if this is just a reg copy. if (L.Reg == H.Reg && L.Sub == H.Sub && !H.Low && L.Low) return false; unsigned Opc = MI->getOpcode(); unsigned COpc = getCombineOpcode(H.Low, L.Low); if (COpc == Opc) return false; if (!validateReg(H, COpc, 1) || !validateReg(L, COpc, 2)) return false; MachineBasicBlock &B = *MI->getParent(); DebugLoc DL = MI->getDebugLoc(); Register NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); auto At = MI->isPHI() ? B.getFirstNonPHI() : MachineBasicBlock::iterator(MI); BuildMI(B, At, DL, HII.get(COpc), NewR) .addReg(H.Reg, 0, H.Sub) .addReg(L.Reg, 0, L.Sub); HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); BT.put(BitTracker::RegisterRef(NewR), RC); return true; } // If MI resets high bits of a register and keeps the lower ones, replace it // with zero-extend byte/half, and-immediate, or extractu, as appropriate. bool BitSimplification::genExtractLow(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { unsigned Opc = MI->getOpcode(); switch (Opc) { case Hexagon::A2_zxtb: case Hexagon::A2_zxth: case Hexagon::S2_extractu: return false; } if (Opc == Hexagon::A2_andir && MI->getOperand(2).isImm()) { int32_t Imm = MI->getOperand(2).getImm(); if (isInt<10>(Imm)) return false; } if (MI->hasUnmodeledSideEffects() || MI->isInlineAsm()) return false; unsigned W = RC.width(); while (W > 0 && RC[W-1].is(0)) W--; if (W == 0 || W == RC.width()) return false; unsigned NewOpc = (W == 8) ? Hexagon::A2_zxtb : (W == 16) ? Hexagon::A2_zxth : (W < 10) ? Hexagon::A2_andir : Hexagon::S2_extractu; MachineBasicBlock &B = *MI->getParent(); DebugLoc DL = MI->getDebugLoc(); for (auto &Op : MI->uses()) { if (!Op.isReg()) continue; BitTracker::RegisterRef RS = Op; if (!BT.has(RS.Reg)) continue; const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); unsigned BN, BW; if (!HBS::getSubregMask(RS, BN, BW, MRI)) continue; if (BW < W || !HBS::isEqual(RC, 0, SC, BN, W)) continue; if (!validateReg(RS, NewOpc, 1)) continue; Register NewR = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); auto At = MI->isPHI() ? B.getFirstNonPHI() : MachineBasicBlock::iterator(MI); auto MIB = BuildMI(B, At, DL, HII.get(NewOpc), NewR) .addReg(RS.Reg, 0, RS.Sub); if (NewOpc == Hexagon::A2_andir) MIB.addImm((1 << W) - 1); else if (NewOpc == Hexagon::S2_extractu) MIB.addImm(W).addImm(0); HBS::replaceSubWithSub(RD.Reg, RD.Sub, NewR, 0, MRI); BT.put(BitTracker::RegisterRef(NewR), RC); return true; } return false; } bool BitSimplification::genBitSplit(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC, const RegisterSet &AVs) { if (!GenBitSplit) return false; if (MaxBitSplit.getNumOccurrences()) { if (CountBitSplit >= MaxBitSplit) return false; } unsigned Opc = MI->getOpcode(); switch (Opc) { case Hexagon::A4_bitsplit: case Hexagon::A4_bitspliti: return false; } unsigned W = RC.width(); if (W != 32) return false; auto ctlz = [] (const BitTracker::RegisterCell &C) -> unsigned { unsigned Z = C.width(); while (Z > 0 && C[Z-1].is(0)) --Z; return C.width() - Z; }; // Count the number of leading zeros in the target RC. unsigned Z = ctlz(RC); if (Z == 0 || Z == W) return false; // A simplistic analysis: assume the source register (the one being split) // is fully unknown, and that all its bits are self-references. const BitTracker::BitValue &B0 = RC[0]; if (B0.Type != BitTracker::BitValue::Ref) return false; unsigned SrcR = B0.RefI.Reg; unsigned SrcSR = 0; unsigned Pos = B0.RefI.Pos; // All the non-zero bits should be consecutive bits from the same register. for (unsigned i = 1; i < W-Z; ++i) { const BitTracker::BitValue &V = RC[i]; if (V.Type != BitTracker::BitValue::Ref) return false; if (V.RefI.Reg != SrcR || V.RefI.Pos != Pos+i) return false; } // Now, find the other bitfield among AVs. for (unsigned S = AVs.find_first(); S; S = AVs.find_next(S)) { // The number of leading zeros here should be the number of trailing // non-zeros in RC. unsigned SRC = MRI.getRegClass(S)->getID(); if (SRC != Hexagon::IntRegsRegClassID && SRC != Hexagon::DoubleRegsRegClassID) continue; if (!BT.has(S)) continue; const BitTracker::RegisterCell &SC = BT.lookup(S); if (SC.width() != W || ctlz(SC) != W-Z) continue; // The Z lower bits should now match SrcR. const BitTracker::BitValue &S0 = SC[0]; if (S0.Type != BitTracker::BitValue::Ref || S0.RefI.Reg != SrcR) continue; unsigned P = S0.RefI.Pos; if (Pos <= P && (Pos + W-Z) != P) continue; if (P < Pos && (P + Z) != Pos) continue; // The starting bitfield position must be at a subregister boundary. if (std::min(P, Pos) != 0 && std::min(P, Pos) != 32) continue; unsigned I; for (I = 1; I < Z; ++I) { const BitTracker::BitValue &V = SC[I]; if (V.Type != BitTracker::BitValue::Ref) break; if (V.RefI.Reg != SrcR || V.RefI.Pos != P+I) break; } if (I != Z) continue; // Generate bitsplit where S is defined. if (MaxBitSplit.getNumOccurrences()) CountBitSplit++; MachineInstr *DefS = MRI.getVRegDef(S); assert(DefS != nullptr); DebugLoc DL = DefS->getDebugLoc(); MachineBasicBlock &B = *DefS->getParent(); auto At = DefS->isPHI() ? B.getFirstNonPHI() : MachineBasicBlock::iterator(DefS); if (MRI.getRegClass(SrcR)->getID() == Hexagon::DoubleRegsRegClassID) SrcSR = (std::min(Pos, P) == 32) ? Hexagon::isub_hi : Hexagon::isub_lo; if (!validateReg({SrcR,SrcSR}, Hexagon::A4_bitspliti, 1)) continue; unsigned ImmOp = Pos <= P ? W-Z : Z; // Find an existing bitsplit instruction if one already exists. unsigned NewR = 0; for (MachineInstr *In : NewMIs) { if (In->getOpcode() != Hexagon::A4_bitspliti) continue; MachineOperand &Op1 = In->getOperand(1); if (Op1.getReg() != SrcR || Op1.getSubReg() != SrcSR) continue; if (In->getOperand(2).getImm() != ImmOp) continue; // Check if the target register is available here. MachineOperand &Op0 = In->getOperand(0); MachineInstr *DefI = MRI.getVRegDef(Op0.getReg()); assert(DefI != nullptr); if (!MDT.dominates(DefI, &*At)) continue; // Found one that can be reused. assert(Op0.getSubReg() == 0); NewR = Op0.getReg(); break; } if (!NewR) { NewR = MRI.createVirtualRegister(&Hexagon::DoubleRegsRegClass); auto NewBS = BuildMI(B, At, DL, HII.get(Hexagon::A4_bitspliti), NewR) .addReg(SrcR, 0, SrcSR) .addImm(ImmOp); NewMIs.push_back(NewBS); } if (Pos <= P) { HBS::replaceRegWithSub(RD.Reg, NewR, Hexagon::isub_lo, MRI); HBS::replaceRegWithSub(S, NewR, Hexagon::isub_hi, MRI); } else { HBS::replaceRegWithSub(S, NewR, Hexagon::isub_lo, MRI); HBS::replaceRegWithSub(RD.Reg, NewR, Hexagon::isub_hi, MRI); } return true; } return false; } // Check for tstbit simplification opportunity, where the bit being checked // can be tracked back to another register. For example: // %2 = S2_lsr_i_r %1, 5 // %3 = S2_tstbit_i %2, 0 // => // %3 = S2_tstbit_i %1, 5 bool BitSimplification::simplifyTstbit(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC) { unsigned Opc = MI->getOpcode(); if (Opc != Hexagon::S2_tstbit_i) return false; unsigned BN = MI->getOperand(2).getImm(); BitTracker::RegisterRef RS = MI->getOperand(1); unsigned F, W; DebugLoc DL = MI->getDebugLoc(); if (!BT.has(RS.Reg) || !HBS::getSubregMask(RS, F, W, MRI)) return false; MachineBasicBlock &B = *MI->getParent(); auto At = MI->isPHI() ? B.getFirstNonPHI() : MachineBasicBlock::iterator(MI); const BitTracker::RegisterCell &SC = BT.lookup(RS.Reg); const BitTracker::BitValue &V = SC[F+BN]; if (V.Type == BitTracker::BitValue::Ref && V.RefI.Reg != RS.Reg) { const TargetRegisterClass *TC = MRI.getRegClass(V.RefI.Reg); // Need to map V.RefI.Reg to a 32-bit register, i.e. if it is // a double register, need to use a subregister and adjust bit // number. unsigned P = std::numeric_limits::max(); BitTracker::RegisterRef RR(V.RefI.Reg, 0); if (TC == &Hexagon::DoubleRegsRegClass) { P = V.RefI.Pos; RR.Sub = Hexagon::isub_lo; if (P >= 32) { P -= 32; RR.Sub = Hexagon::isub_hi; } } else if (TC == &Hexagon::IntRegsRegClass) { P = V.RefI.Pos; } if (P != std::numeric_limits::max()) { Register NewR = MRI.createVirtualRegister(&Hexagon::PredRegsRegClass); BuildMI(B, At, DL, HII.get(Hexagon::S2_tstbit_i), NewR) .addReg(RR.Reg, 0, RR.Sub) .addImm(P); HBS::replaceReg(RD.Reg, NewR, MRI); BT.put(NewR, RC); return true; } } else if (V.is(0) || V.is(1)) { Register NewR = MRI.createVirtualRegister(&Hexagon::PredRegsRegClass); unsigned NewOpc = V.is(0) ? Hexagon::PS_false : Hexagon::PS_true; BuildMI(B, At, DL, HII.get(NewOpc), NewR); HBS::replaceReg(RD.Reg, NewR, MRI); return true; } return false; } // Detect whether RD is a bitfield extract (sign- or zero-extended) of // some register from the AVs set. Create a new corresponding instruction // at the location of MI. The intent is to recognize situations where // a sequence of instructions performs an operation that is equivalent to // an extract operation, such as a shift left followed by a shift right. bool BitSimplification::simplifyExtractLow(MachineInstr *MI, BitTracker::RegisterRef RD, const BitTracker::RegisterCell &RC, const RegisterSet &AVs) { if (!GenExtract) return false; if (MaxExtract.getNumOccurrences()) { if (CountExtract >= MaxExtract) return false; CountExtract++; } unsigned W = RC.width(); unsigned RW = W; unsigned Len; bool Signed; // The code is mostly class-independent, except for the part that generates // the extract instruction, and establishes the source register (in case it // needs to use a subregister). const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); if (FRC != &Hexagon::IntRegsRegClass && FRC != &Hexagon::DoubleRegsRegClass) return false; assert(RD.Sub == 0); // Observation: // If the cell has a form of 00..0xx..x with k zeros and n remaining // bits, this could be an extractu of the n bits, but it could also be // an extractu of a longer field which happens to have 0s in the top // bit positions. // The same logic applies to sign-extended fields. // // Do not check for the extended extracts, since it would expand the // search space quite a bit. The search may be expensive as it is. const BitTracker::BitValue &TopV = RC[W-1]; // Eliminate candidates that have self-referential bits, since they // cannot be extracts from other registers. Also, skip registers that // have compile-time constant values. bool IsConst = true; for (unsigned I = 0; I != W; ++I) { const BitTracker::BitValue &V = RC[I]; if (V.Type == BitTracker::BitValue::Ref && V.RefI.Reg == RD.Reg) return false; IsConst = IsConst && (V.is(0) || V.is(1)); } if (IsConst) return false; if (TopV.is(0) || TopV.is(1)) { bool S = TopV.is(1); for (--W; W > 0 && RC[W-1].is(S); --W) ; Len = W; Signed = S; // The sign bit must be a part of the field being extended. if (Signed) ++Len; } else { // This could still be a sign-extended extract. assert(TopV.Type == BitTracker::BitValue::Ref); if (TopV.RefI.Reg == RD.Reg || TopV.RefI.Pos == W-1) return false; for (--W; W > 0 && RC[W-1] == TopV; --W) ; // The top bits of RC are copies of TopV. One occurrence of TopV will // be a part of the field. Len = W + 1; Signed = true; } // This would be just a copy. It should be handled elsewhere. if (Len == RW) return false; LLVM_DEBUG({ dbgs() << __func__ << " on reg: " << printReg(RD.Reg, &HRI, RD.Sub) << ", MI: " << *MI; dbgs() << "Cell: " << RC << '\n'; dbgs() << "Expected bitfield size: " << Len << " bits, " << (Signed ? "sign" : "zero") << "-extended\n"; }); bool Changed = false; for (unsigned R = AVs.find_first(); R != 0; R = AVs.find_next(R)) { if (!BT.has(R)) continue; const BitTracker::RegisterCell &SC = BT.lookup(R); unsigned SW = SC.width(); // The source can be longer than the destination, as long as its size is // a multiple of the size of the destination. Also, we would need to be // able to refer to the subregister in the source that would be of the // same size as the destination, but only check the sizes here. if (SW < RW || (SW % RW) != 0) continue; // The field can start at any offset in SC as long as it contains Len // bits and does not cross subregister boundary (if the source register // is longer than the destination). unsigned Off = 0; while (Off <= SW-Len) { unsigned OE = (Off+Len)/RW; if (OE != Off/RW) { // The assumption here is that if the source (R) is longer than the // destination, then the destination is a sequence of words of // size RW, and each such word in R can be accessed via a subregister. // // If the beginning and the end of the field cross the subregister // boundary, advance to the next subregister. Off = OE*RW; continue; } if (HBS::isEqual(RC, 0, SC, Off, Len)) break; ++Off; } if (Off > SW-Len) continue; // Found match. unsigned ExtOpc = 0; if (Off == 0) { if (Len == 8) ExtOpc = Signed ? Hexagon::A2_sxtb : Hexagon::A2_zxtb; else if (Len == 16) ExtOpc = Signed ? Hexagon::A2_sxth : Hexagon::A2_zxth; else if (Len < 10 && !Signed) ExtOpc = Hexagon::A2_andir; } if (ExtOpc == 0) { ExtOpc = Signed ? (RW == 32 ? Hexagon::S4_extract : Hexagon::S4_extractp) : (RW == 32 ? Hexagon::S2_extractu : Hexagon::S2_extractup); } unsigned SR = 0; // This only recognizes isub_lo and isub_hi. if (RW != SW && RW*2 != SW) continue; if (RW != SW) SR = (Off/RW == 0) ? Hexagon::isub_lo : Hexagon::isub_hi; Off = Off % RW; if (!validateReg({R,SR}, ExtOpc, 1)) continue; // Don't generate the same instruction as the one being optimized. if (MI->getOpcode() == ExtOpc) { // All possible ExtOpc's have the source in operand(1). const MachineOperand &SrcOp = MI->getOperand(1); if (SrcOp.getReg() == R) continue; } DebugLoc DL = MI->getDebugLoc(); MachineBasicBlock &B = *MI->getParent(); Register NewR = MRI.createVirtualRegister(FRC); auto At = MI->isPHI() ? B.getFirstNonPHI() : MachineBasicBlock::iterator(MI); auto MIB = BuildMI(B, At, DL, HII.get(ExtOpc), NewR) .addReg(R, 0, SR); switch (ExtOpc) { case Hexagon::A2_sxtb: case Hexagon::A2_zxtb: case Hexagon::A2_sxth: case Hexagon::A2_zxth: break; case Hexagon::A2_andir: MIB.addImm((1u << Len) - 1); break; case Hexagon::S4_extract: case Hexagon::S2_extractu: case Hexagon::S4_extractp: case Hexagon::S2_extractup: MIB.addImm(Len) .addImm(Off); break; default: llvm_unreachable("Unexpected opcode"); } HBS::replaceReg(RD.Reg, NewR, MRI); BT.put(BitTracker::RegisterRef(NewR), RC); Changed = true; break; } return Changed; } bool BitSimplification::simplifyRCmp0(MachineInstr *MI, BitTracker::RegisterRef RD) { unsigned Opc = MI->getOpcode(); if (Opc != Hexagon::A4_rcmpeqi && Opc != Hexagon::A4_rcmpneqi) return false; MachineOperand &CmpOp = MI->getOperand(2); if (!CmpOp.isImm() || CmpOp.getImm() != 0) return false; const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); if (FRC != &Hexagon::IntRegsRegClass && FRC != &Hexagon::DoubleRegsRegClass) return false; assert(RD.Sub == 0); MachineBasicBlock &B = *MI->getParent(); const DebugLoc &DL = MI->getDebugLoc(); auto At = MI->isPHI() ? B.getFirstNonPHI() : MachineBasicBlock::iterator(MI); bool KnownZ = true; bool KnownNZ = false; BitTracker::RegisterRef SR = MI->getOperand(1); if (!BT.has(SR.Reg)) return false; const BitTracker::RegisterCell &SC = BT.lookup(SR.Reg); unsigned F, W; if (!HBS::getSubregMask(SR, F, W, MRI)) return false; for (uint16_t I = F; I != F+W; ++I) { const BitTracker::BitValue &V = SC[I]; if (!V.is(0)) KnownZ = false; if (V.is(1)) KnownNZ = true; } auto ReplaceWithConst = [&](int C) { Register NewR = MRI.createVirtualRegister(FRC); BuildMI(B, At, DL, HII.get(Hexagon::A2_tfrsi), NewR) .addImm(C); HBS::replaceReg(RD.Reg, NewR, MRI); BitTracker::RegisterCell NewRC(W); for (uint16_t I = 0; I != W; ++I) { NewRC[I] = BitTracker::BitValue(C & 1); C = unsigned(C) >> 1; } BT.put(BitTracker::RegisterRef(NewR), NewRC); return true; }; auto IsNonZero = [] (const MachineOperand &Op) { if (Op.isGlobal() || Op.isBlockAddress()) return true; if (Op.isImm()) return Op.getImm() != 0; if (Op.isCImm()) return !Op.getCImm()->isZero(); if (Op.isFPImm()) return !Op.getFPImm()->isZero(); return false; }; auto IsZero = [] (const MachineOperand &Op) { if (Op.isGlobal() || Op.isBlockAddress()) return false; if (Op.isImm()) return Op.getImm() == 0; if (Op.isCImm()) return Op.getCImm()->isZero(); if (Op.isFPImm()) return Op.getFPImm()->isZero(); return false; }; // If the source register is known to be 0 or non-0, the comparison can // be folded to a load of a constant. if (KnownZ || KnownNZ) { assert(KnownZ != KnownNZ && "Register cannot be both 0 and non-0"); return ReplaceWithConst(KnownZ == (Opc == Hexagon::A4_rcmpeqi)); } // Special case: if the compare comes from a C2_muxii, then we know the // two possible constants that can be the source value. MachineInstr *InpDef = MRI.getVRegDef(SR.Reg); if (!InpDef) return false; if (SR.Sub == 0 && InpDef->getOpcode() == Hexagon::C2_muxii) { MachineOperand &Src1 = InpDef->getOperand(2); MachineOperand &Src2 = InpDef->getOperand(3); // Check if both are non-zero. bool KnownNZ1 = IsNonZero(Src1), KnownNZ2 = IsNonZero(Src2); if (KnownNZ1 && KnownNZ2) return ReplaceWithConst(Opc == Hexagon::A4_rcmpneqi); // Check if both are zero. bool KnownZ1 = IsZero(Src1), KnownZ2 = IsZero(Src2); if (KnownZ1 && KnownZ2) return ReplaceWithConst(Opc == Hexagon::A4_rcmpeqi); // If for both operands we know that they are either 0 or non-0, // replace the comparison with a C2_muxii, using the same predicate // register, but with operands substituted with 0/1 accordingly. if ((KnownZ1 || KnownNZ1) && (KnownZ2 || KnownNZ2)) { Register NewR = MRI.createVirtualRegister(FRC); BuildMI(B, At, DL, HII.get(Hexagon::C2_muxii), NewR) .addReg(InpDef->getOperand(1).getReg()) .addImm(KnownZ1 == (Opc == Hexagon::A4_rcmpeqi)) .addImm(KnownZ2 == (Opc == Hexagon::A4_rcmpeqi)); HBS::replaceReg(RD.Reg, NewR, MRI); // Create a new cell with only the least significant bit unknown. BitTracker::RegisterCell NewRC(W); NewRC[0] = BitTracker::BitValue::self(); NewRC.fill(1, W, BitTracker::BitValue::Zero); BT.put(BitTracker::RegisterRef(NewR), NewRC); return true; } } return false; } bool BitSimplification::processBlock(MachineBasicBlock &B, const RegisterSet &AVs) { if (!BT.reached(&B)) return false; bool Changed = false; RegisterSet AVB = AVs; RegisterSet Defs; for (auto I = B.begin(), E = B.end(); I != E; ++I, AVB.insert(Defs)) { MachineInstr *MI = &*I; Defs.clear(); HBS::getInstrDefs(*MI, Defs); unsigned Opc = MI->getOpcode(); if (Opc == TargetOpcode::COPY || Opc == TargetOpcode::REG_SEQUENCE) continue; if (MI->mayStore()) { bool T = genStoreUpperHalf(MI); T = T || genStoreImmediate(MI); Changed |= T; continue; } if (Defs.count() != 1) continue; const MachineOperand &Op0 = MI->getOperand(0); if (!Op0.isReg() || !Op0.isDef()) continue; BitTracker::RegisterRef RD = Op0; if (!BT.has(RD.Reg)) continue; const TargetRegisterClass *FRC = HBS::getFinalVRegClass(RD, MRI); const BitTracker::RegisterCell &RC = BT.lookup(RD.Reg); if (FRC->getID() == Hexagon::DoubleRegsRegClassID) { bool T = genPackhl(MI, RD, RC); T = T || simplifyExtractLow(MI, RD, RC, AVB); Changed |= T; continue; } if (FRC->getID() == Hexagon::IntRegsRegClassID) { bool T = genBitSplit(MI, RD, RC, AVB); T = T || simplifyExtractLow(MI, RD, RC, AVB); T = T || genExtractHalf(MI, RD, RC); T = T || genCombineHalf(MI, RD, RC); T = T || genExtractLow(MI, RD, RC); T = T || simplifyRCmp0(MI, RD); Changed |= T; continue; } if (FRC->getID() == Hexagon::PredRegsRegClassID) { bool T = simplifyTstbit(MI, RD, RC); Changed |= T; continue; } } return Changed; } bool HexagonBitSimplify::runOnMachineFunction(MachineFunction &MF) { if (skipFunction(MF.getFunction())) return false; auto &HST = MF.getSubtarget(); auto &HRI = *HST.getRegisterInfo(); auto &HII = *HST.getInstrInfo(); MDT = &getAnalysis(); MachineRegisterInfo &MRI = MF.getRegInfo(); bool Changed; Changed = DeadCodeElimination(MF, *MDT).run(); const HexagonEvaluator HE(HRI, MRI, HII, MF); BitTracker BT(HE, MF); LLVM_DEBUG(BT.trace(true)); BT.run(); MachineBasicBlock &Entry = MF.front(); RegisterSet AIG; // Available registers for IG. ConstGeneration ImmG(BT, HII, MRI); Changed |= visitBlock(Entry, ImmG, AIG); RegisterSet ARE; // Available registers for RIE. RedundantInstrElimination RIE(BT, HII, HRI, MRI); bool Ried = visitBlock(Entry, RIE, ARE); if (Ried) { Changed = true; BT.run(); } RegisterSet ACG; // Available registers for CG. CopyGeneration CopyG(BT, HII, HRI, MRI); Changed |= visitBlock(Entry, CopyG, ACG); RegisterSet ACP; // Available registers for CP. CopyPropagation CopyP(HRI, MRI); Changed |= visitBlock(Entry, CopyP, ACP); Changed = DeadCodeElimination(MF, *MDT).run() || Changed; BT.run(); RegisterSet ABS; // Available registers for BS. BitSimplification BitS(BT, *MDT, HII, HRI, MRI, MF); Changed |= visitBlock(Entry, BitS, ABS); Changed = DeadCodeElimination(MF, *MDT).run() || Changed; if (Changed) { for (auto &B : MF) for (auto &I : B) I.clearKillInfo(); DeadCodeElimination(MF, *MDT).run(); } return Changed; } // Recognize loops where the code at the end of the loop matches the code // before the entry of the loop, and the matching code is such that is can // be simplified. This pass relies on the bit simplification above and only // prepares code in a way that can be handled by the bit simplifcation. // // This is the motivating testcase (and explanation): // // { // loop0(.LBB0_2, r1) // %for.body.preheader // r5:4 = memd(r0++#8) // } // { // r3 = lsr(r4, #16) // r7:6 = combine(r5, r5) // } // { // r3 = insert(r5, #16, #16) // r7:6 = vlsrw(r7:6, #16) // } // .LBB0_2: // { // memh(r2+#4) = r5 // memh(r2+#6) = r6 # R6 is really R5.H // } // { // r2 = add(r2, #8) // memh(r2+#0) = r4 // memh(r2+#2) = r3 # R3 is really R4.H // } // { // r5:4 = memd(r0++#8) // } // { # "Shuffling" code that sets up R3 and R6 // r3 = lsr(r4, #16) # so that their halves can be stored in the // r7:6 = combine(r5, r5) # next iteration. This could be folded into // } # the stores if the code was at the beginning // { # of the loop iteration. Since the same code // r3 = insert(r5, #16, #16) # precedes the loop, it can actually be moved // r7:6 = vlsrw(r7:6, #16) # there. // }:endloop0 // // // The outcome: // // { // loop0(.LBB0_2, r1) // r5:4 = memd(r0++#8) // } // .LBB0_2: // { // memh(r2+#4) = r5 // memh(r2+#6) = r5.h // } // { // r2 = add(r2, #8) // memh(r2+#0) = r4 // memh(r2+#2) = r4.h // } // { // r5:4 = memd(r0++#8) // }:endloop0 namespace llvm { FunctionPass *createHexagonLoopRescheduling(); void initializeHexagonLoopReschedulingPass(PassRegistry&); } // end namespace llvm namespace { class HexagonLoopRescheduling : public MachineFunctionPass { public: static char ID; HexagonLoopRescheduling() : MachineFunctionPass(ID) { initializeHexagonLoopReschedulingPass(*PassRegistry::getPassRegistry()); } bool runOnMachineFunction(MachineFunction &MF) override; private: const HexagonInstrInfo *HII = nullptr; const HexagonRegisterInfo *HRI = nullptr; MachineRegisterInfo *MRI = nullptr; BitTracker *BTP = nullptr; struct LoopCand { LoopCand(MachineBasicBlock *lb, MachineBasicBlock *pb, MachineBasicBlock *eb) : LB(lb), PB(pb), EB(eb) {} MachineBasicBlock *LB, *PB, *EB; }; using InstrList = std::vector; struct InstrGroup { BitTracker::RegisterRef Inp, Out; InstrList Ins; }; struct PhiInfo { PhiInfo(MachineInstr &P, MachineBasicBlock &B); unsigned DefR; BitTracker::RegisterRef LR, PR; // Loop Register, Preheader Register MachineBasicBlock *LB, *PB; // Loop Block, Preheader Block }; static unsigned getDefReg(const MachineInstr *MI); bool isConst(unsigned Reg) const; bool isBitShuffle(const MachineInstr *MI, unsigned DefR) const; bool isStoreInput(const MachineInstr *MI, unsigned DefR) const; bool isShuffleOf(unsigned OutR, unsigned InpR) const; bool isSameShuffle(unsigned OutR1, unsigned InpR1, unsigned OutR2, unsigned &InpR2) const; void moveGroup(InstrGroup &G, MachineBasicBlock &LB, MachineBasicBlock &PB, MachineBasicBlock::iterator At, unsigned OldPhiR, unsigned NewPredR); bool processLoop(LoopCand &C); }; } // end anonymous namespace char HexagonLoopRescheduling::ID = 0; INITIALIZE_PASS(HexagonLoopRescheduling, "hexagon-loop-resched", "Hexagon Loop Rescheduling", false, false) HexagonLoopRescheduling::PhiInfo::PhiInfo(MachineInstr &P, MachineBasicBlock &B) { DefR = HexagonLoopRescheduling::getDefReg(&P); LB = &B; PB = nullptr; for (unsigned i = 1, n = P.getNumOperands(); i < n; i += 2) { const MachineOperand &OpB = P.getOperand(i+1); if (OpB.getMBB() == &B) { LR = P.getOperand(i); continue; } PB = OpB.getMBB(); PR = P.getOperand(i); } } unsigned HexagonLoopRescheduling::getDefReg(const MachineInstr *MI) { RegisterSet Defs; HBS::getInstrDefs(*MI, Defs); if (Defs.count() != 1) return 0; return Defs.find_first(); } bool HexagonLoopRescheduling::isConst(unsigned Reg) const { if (!BTP->has(Reg)) return false; const BitTracker::RegisterCell &RC = BTP->lookup(Reg); for (unsigned i = 0, w = RC.width(); i < w; ++i) { const BitTracker::BitValue &V = RC[i]; if (!V.is(0) && !V.is(1)) return false; } return true; } bool HexagonLoopRescheduling::isBitShuffle(const MachineInstr *MI, unsigned DefR) const { unsigned Opc = MI->getOpcode(); switch (Opc) { case TargetOpcode::COPY: case Hexagon::S2_lsr_i_r: case Hexagon::S2_asr_i_r: case Hexagon::S2_asl_i_r: case Hexagon::S2_lsr_i_p: case Hexagon::S2_asr_i_p: case Hexagon::S2_asl_i_p: case Hexagon::S2_insert: case Hexagon::A2_or: case Hexagon::A2_orp: case Hexagon::A2_and: case Hexagon::A2_andp: case Hexagon::A2_combinew: case Hexagon::A4_combineri: case Hexagon::A4_combineir: case Hexagon::A2_combineii: case Hexagon::A4_combineii: case Hexagon::A2_combine_ll: case Hexagon::A2_combine_lh: case Hexagon::A2_combine_hl: case Hexagon::A2_combine_hh: return true; } return false; } bool HexagonLoopRescheduling::isStoreInput(const MachineInstr *MI, unsigned InpR) const { for (unsigned i = 0, n = MI->getNumOperands(); i < n; ++i) { const MachineOperand &Op = MI->getOperand(i); if (!Op.isReg()) continue; if (Op.getReg() == InpR) return i == n-1; } return false; } bool HexagonLoopRescheduling::isShuffleOf(unsigned OutR, unsigned InpR) const { if (!BTP->has(OutR) || !BTP->has(InpR)) return false; const BitTracker::RegisterCell &OutC = BTP->lookup(OutR); for (unsigned i = 0, w = OutC.width(); i < w; ++i) { const BitTracker::BitValue &V = OutC[i]; if (V.Type != BitTracker::BitValue::Ref) continue; if (V.RefI.Reg != InpR) return false; } return true; } bool HexagonLoopRescheduling::isSameShuffle(unsigned OutR1, unsigned InpR1, unsigned OutR2, unsigned &InpR2) const { if (!BTP->has(OutR1) || !BTP->has(InpR1) || !BTP->has(OutR2)) return false; const BitTracker::RegisterCell &OutC1 = BTP->lookup(OutR1); const BitTracker::RegisterCell &OutC2 = BTP->lookup(OutR2); unsigned W = OutC1.width(); unsigned MatchR = 0; if (W != OutC2.width()) return false; for (unsigned i = 0; i < W; ++i) { const BitTracker::BitValue &V1 = OutC1[i], &V2 = OutC2[i]; if (V1.Type != V2.Type || V1.Type == BitTracker::BitValue::One) return false; if (V1.Type != BitTracker::BitValue::Ref) continue; if (V1.RefI.Pos != V2.RefI.Pos) return false; if (V1.RefI.Reg != InpR1) return false; if (V2.RefI.Reg == 0 || V2.RefI.Reg == OutR2) return false; if (!MatchR) MatchR = V2.RefI.Reg; else if (V2.RefI.Reg != MatchR) return false; } InpR2 = MatchR; return true; } void HexagonLoopRescheduling::moveGroup(InstrGroup &G, MachineBasicBlock &LB, MachineBasicBlock &PB, MachineBasicBlock::iterator At, unsigned OldPhiR, unsigned NewPredR) { DenseMap RegMap; const TargetRegisterClass *PhiRC = MRI->getRegClass(NewPredR); Register PhiR = MRI->createVirtualRegister(PhiRC); BuildMI(LB, At, At->getDebugLoc(), HII->get(TargetOpcode::PHI), PhiR) .addReg(NewPredR) .addMBB(&PB) .addReg(G.Inp.Reg) .addMBB(&LB); RegMap.insert(std::make_pair(G.Inp.Reg, PhiR)); for (const MachineInstr *SI : llvm::reverse(G.Ins)) { unsigned DR = getDefReg(SI); const TargetRegisterClass *RC = MRI->getRegClass(DR); Register NewDR = MRI->createVirtualRegister(RC); DebugLoc DL = SI->getDebugLoc(); auto MIB = BuildMI(LB, At, DL, HII->get(SI->getOpcode()), NewDR); for (unsigned j = 0, m = SI->getNumOperands(); j < m; ++j) { const MachineOperand &Op = SI->getOperand(j); if (!Op.isReg()) { MIB.add(Op); continue; } if (!Op.isUse()) continue; unsigned UseR = RegMap[Op.getReg()]; MIB.addReg(UseR, 0, Op.getSubReg()); } RegMap.insert(std::make_pair(DR, NewDR)); } HBS::replaceReg(OldPhiR, RegMap[G.Out.Reg], *MRI); } bool HexagonLoopRescheduling::processLoop(LoopCand &C) { LLVM_DEBUG(dbgs() << "Processing loop in " << printMBBReference(*C.LB) << "\n"); std::vector Phis; for (auto &I : *C.LB) { if (!I.isPHI()) break; unsigned PR = getDefReg(&I); if (isConst(PR)) continue; bool BadUse = false, GoodUse = false; for (const MachineOperand &MO : MRI->use_operands(PR)) { const MachineInstr *UseI = MO.getParent(); if (UseI->getParent() != C.LB) { BadUse = true; break; } if (isBitShuffle(UseI, PR) || isStoreInput(UseI, PR)) GoodUse = true; } if (BadUse || !GoodUse) continue; Phis.push_back(PhiInfo(I, *C.LB)); } LLVM_DEBUG({ dbgs() << "Phis: {"; for (auto &I : Phis) { dbgs() << ' ' << printReg(I.DefR, HRI) << "=phi(" << printReg(I.PR.Reg, HRI, I.PR.Sub) << ":b" << I.PB->getNumber() << ',' << printReg(I.LR.Reg, HRI, I.LR.Sub) << ":b" << I.LB->getNumber() << ')'; } dbgs() << " }\n"; }); if (Phis.empty()) return false; bool Changed = false; InstrList ShufIns; // Go backwards in the block: for each bit shuffling instruction, check // if that instruction could potentially be moved to the front of the loop: // the output of the loop cannot be used in a non-shuffling instruction // in this loop. for (MachineInstr &MI : llvm::reverse(*C.LB)) { if (MI.isTerminator()) continue; if (MI.isPHI()) break; RegisterSet Defs; HBS::getInstrDefs(MI, Defs); if (Defs.count() != 1) continue; Register DefR = Defs.find_first(); if (!DefR.isVirtual()) continue; if (!isBitShuffle(&MI, DefR)) continue; bool BadUse = false; for (auto UI = MRI->use_begin(DefR), UE = MRI->use_end(); UI != UE; ++UI) { MachineInstr *UseI = UI->getParent(); if (UseI->getParent() == C.LB) { if (UseI->isPHI()) { // If the use is in a phi node in this loop, then it should be // the value corresponding to the back edge. unsigned Idx = UI.getOperandNo(); if (UseI->getOperand(Idx+1).getMBB() != C.LB) BadUse = true; } else { if (!llvm::is_contained(ShufIns, UseI)) BadUse = true; } } else { // There is a use outside of the loop, but there is no epilog block // suitable for a copy-out. if (C.EB == nullptr) BadUse = true; } if (BadUse) break; } if (BadUse) continue; ShufIns.push_back(&MI); } // Partition the list of shuffling instructions into instruction groups, // where each group has to be moved as a whole (i.e. a group is a chain of // dependent instructions). A group produces a single live output register, // which is meant to be the input of the loop phi node (although this is // not checked here yet). It also uses a single register as its input, // which is some value produced in the loop body. After moving the group // to the beginning of the loop, that input register would need to be // the loop-carried register (through a phi node) instead of the (currently // loop-carried) output register. using InstrGroupList = std::vector; InstrGroupList Groups; for (unsigned i = 0, n = ShufIns.size(); i < n; ++i) { MachineInstr *SI = ShufIns[i]; if (SI == nullptr) continue; InstrGroup G; G.Ins.push_back(SI); G.Out.Reg = getDefReg(SI); RegisterSet Inputs; HBS::getInstrUses(*SI, Inputs); for (unsigned j = i+1; j < n; ++j) { MachineInstr *MI = ShufIns[j]; if (MI == nullptr) continue; RegisterSet Defs; HBS::getInstrDefs(*MI, Defs); // If this instruction does not define any pending inputs, skip it. if (!Defs.intersects(Inputs)) continue; // Otherwise, add it to the current group and remove the inputs that // are defined by MI. G.Ins.push_back(MI); Inputs.remove(Defs); // Then add all registers used by MI. HBS::getInstrUses(*MI, Inputs); ShufIns[j] = nullptr; } // Only add a group if it requires at most one register. if (Inputs.count() > 1) continue; auto LoopInpEq = [G] (const PhiInfo &P) -> bool { return G.Out.Reg == P.LR.Reg; }; if (llvm::none_of(Phis, LoopInpEq)) continue; G.Inp.Reg = Inputs.find_first(); Groups.push_back(G); } LLVM_DEBUG({ for (unsigned i = 0, n = Groups.size(); i < n; ++i) { InstrGroup &G = Groups[i]; dbgs() << "Group[" << i << "] inp: " << printReg(G.Inp.Reg, HRI, G.Inp.Sub) << " out: " << printReg(G.Out.Reg, HRI, G.Out.Sub) << "\n"; for (const MachineInstr *MI : G.Ins) dbgs() << " " << MI; } }); for (InstrGroup &G : Groups) { if (!isShuffleOf(G.Out.Reg, G.Inp.Reg)) continue; auto LoopInpEq = [G] (const PhiInfo &P) -> bool { return G.Out.Reg == P.LR.Reg; }; auto F = llvm::find_if(Phis, LoopInpEq); if (F == Phis.end()) continue; unsigned PrehR = 0; if (!isSameShuffle(G.Out.Reg, G.Inp.Reg, F->PR.Reg, PrehR)) { const MachineInstr *DefPrehR = MRI->getVRegDef(F->PR.Reg); unsigned Opc = DefPrehR->getOpcode(); if (Opc != Hexagon::A2_tfrsi && Opc != Hexagon::A2_tfrpi) continue; if (!DefPrehR->getOperand(1).isImm()) continue; if (DefPrehR->getOperand(1).getImm() != 0) continue; const TargetRegisterClass *RC = MRI->getRegClass(G.Inp.Reg); if (RC != MRI->getRegClass(F->PR.Reg)) { PrehR = MRI->createVirtualRegister(RC); unsigned TfrI = (RC == &Hexagon::IntRegsRegClass) ? Hexagon::A2_tfrsi : Hexagon::A2_tfrpi; auto T = C.PB->getFirstTerminator(); DebugLoc DL = (T != C.PB->end()) ? T->getDebugLoc() : DebugLoc(); BuildMI(*C.PB, T, DL, HII->get(TfrI), PrehR) .addImm(0); } else { PrehR = F->PR.Reg; } } // isSameShuffle could match with PrehR being of a wider class than // G.Inp.Reg, for example if G shuffles the low 32 bits of its input, // it would match for the input being a 32-bit register, and PrehR // being a 64-bit register (where the low 32 bits match). This could // be handled, but for now skip these cases. if (MRI->getRegClass(PrehR) != MRI->getRegClass(G.Inp.Reg)) continue; moveGroup(G, *F->LB, *F->PB, F->LB->getFirstNonPHI(), F->DefR, PrehR); Changed = true; } return Changed; } bool HexagonLoopRescheduling::runOnMachineFunction(MachineFunction &MF) { if (skipFunction(MF.getFunction())) return false; auto &HST = MF.getSubtarget(); HII = HST.getInstrInfo(); HRI = HST.getRegisterInfo(); MRI = &MF.getRegInfo(); const HexagonEvaluator HE(*HRI, *MRI, *HII, MF); BitTracker BT(HE, MF); LLVM_DEBUG(BT.trace(true)); BT.run(); BTP = &BT; std::vector Cand; for (auto &B : MF) { if (B.pred_size() != 2 || B.succ_size() != 2) continue; MachineBasicBlock *PB = nullptr; bool IsLoop = false; for (MachineBasicBlock *Pred : B.predecessors()) { if (Pred != &B) PB = Pred; else IsLoop = true; } if (!IsLoop) continue; MachineBasicBlock *EB = nullptr; for (MachineBasicBlock *Succ : B.successors()) { if (Succ == &B) continue; // Set EP to the epilog block, if it has only 1 predecessor (i.e. the // edge from B to EP is non-critical. if (Succ->pred_size() == 1) EB = Succ; break; } Cand.push_back(LoopCand(&B, PB, EB)); } bool Changed = false; for (auto &C : Cand) Changed |= processLoop(C); return Changed; } //===----------------------------------------------------------------------===// // Public Constructor Functions //===----------------------------------------------------------------------===// FunctionPass *llvm::createHexagonLoopRescheduling() { return new HexagonLoopRescheduling(); } FunctionPass *llvm::createHexagonBitSimplify() { return new HexagonBitSimplify(); }