//===- HexagonConstExtenders.cpp ------------------------------------------===// // // The LLVM Compiler Infrastructure // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// #include "HexagonInstrInfo.h" #include "HexagonRegisterInfo.h" #include "HexagonSubtarget.h" #include "llvm/ADT/SmallVector.h" #include "llvm/CodeGen/MachineDominators.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Pass.h" #include #include #include #include #define DEBUG_TYPE "hexagon-cext-opt" using namespace llvm; static cl::opt CountThreshold("hexagon-cext-threshold", cl::init(3), cl::Hidden, cl::ZeroOrMore, cl::desc("Minimum number of extenders to trigger replacement")); static cl::opt ReplaceLimit("hexagon-cext-limit", cl::init(0), cl::Hidden, cl::ZeroOrMore, cl::desc("Maximum number of replacements")); namespace llvm { void initializeHexagonConstExtendersPass(PassRegistry&); FunctionPass *createHexagonConstExtenders(); } static int32_t adjustUp(int32_t V, uint8_t A, uint8_t O) { assert(isPowerOf2_32(A)); int32_t U = (V & -A) + O; return U >= V ? U : U+A; } static int32_t adjustDown(int32_t V, uint8_t A, uint8_t O) { assert(isPowerOf2_32(A)); int32_t U = (V & -A) + O; return U <= V ? U : U-A; } namespace { struct OffsetRange { // The range of values between Min and Max that are of form Align*N+Offset, // for some integer N. Min and Max are required to be of that form as well, // except in the case of an empty range. int32_t Min = INT_MIN, Max = INT_MAX; uint8_t Align = 1; uint8_t Offset = 0; OffsetRange() = default; OffsetRange(int32_t L, int32_t H, uint8_t A, uint8_t O = 0) : Min(L), Max(H), Align(A), Offset(O) {} OffsetRange &intersect(OffsetRange A) { if (Align < A.Align) std::swap(*this, A); // Align >= A.Align. if (Offset >= A.Offset && (Offset - A.Offset) % A.Align == 0) { Min = adjustUp(std::max(Min, A.Min), Align, Offset); Max = adjustDown(std::min(Max, A.Max), Align, Offset); } else { // Make an empty range. Min = 0; Max = -1; } // Canonicalize empty ranges. if (Min > Max) std::tie(Min, Max, Align) = std::make_tuple(0, -1, 1); return *this; } OffsetRange &shift(int32_t S) { Min += S; Max += S; Offset = (Offset+S) % Align; return *this; } OffsetRange &extendBy(int32_t D) { // If D < 0, extend Min, otherwise extend Max. assert(D % Align == 0); if (D < 0) Min = (INT_MIN-D < Min) ? Min+D : INT_MIN; else Max = (INT_MAX-D > Max) ? Max+D : INT_MAX; return *this; } bool empty() const { return Min > Max; } bool contains(int32_t V) const { return Min <= V && V <= Max && (V-Offset) % Align == 0; } bool operator==(const OffsetRange &R) const { return Min == R.Min && Max == R.Max && Align == R.Align; } bool operator!=(const OffsetRange &R) const { return !operator==(R); } bool operator<(const OffsetRange &R) const { if (Min != R.Min) return Min < R.Min; if (Max != R.Max) return Max < R.Max; return Align < R.Align; } static OffsetRange zero() { return {0, 0, 1}; } }; struct RangeTree { struct Node { Node(const OffsetRange &R) : MaxEnd(R.Max), Range(R) {} unsigned Height = 1; unsigned Count = 1; int32_t MaxEnd; const OffsetRange &Range; Node *Left = nullptr, *Right = nullptr; }; Node *Root = nullptr; void add(const OffsetRange &R) { Root = add(Root, R); } void erase(const Node *N) { Root = remove(Root, N); delete N; } void order(SmallVectorImpl &Seq) const { order(Root, Seq); } SmallVector nodesWith(int32_t P, bool CheckAlign = true) { SmallVector Nodes; nodesWith(Root, P, CheckAlign, Nodes); return Nodes; } void dump() const; ~RangeTree() { SmallVector Nodes; order(Nodes); for (Node *N : Nodes) delete N; } private: void dump(const Node *N) const; void order(Node *N, SmallVectorImpl &Seq) const; void nodesWith(Node *N, int32_t P, bool CheckA, SmallVectorImpl &Seq) const; Node *add(Node *N, const OffsetRange &R); Node *remove(Node *N, const Node *D); Node *rotateLeft(Node *Lower, Node *Higher); Node *rotateRight(Node *Lower, Node *Higher); unsigned height(Node *N) { return N != nullptr ? N->Height : 0; } Node *update(Node *N) { assert(N != nullptr); N->Height = 1 + std::max(height(N->Left), height(N->Right)); if (N->Left) N->MaxEnd = std::max(N->MaxEnd, N->Left->MaxEnd); if (N->Right) N->MaxEnd = std::max(N->MaxEnd, N->Right->MaxEnd); return N; } Node *rebalance(Node *N) { assert(N != nullptr); int32_t Balance = height(N->Right) - height(N->Left); if (Balance < -1) return rotateRight(N->Left, N); if (Balance > 1) return rotateLeft(N->Right, N); return N; } }; struct Loc { MachineBasicBlock *Block = nullptr; MachineBasicBlock::iterator At; Loc(MachineBasicBlock *B, MachineBasicBlock::iterator It) : Block(B), At(It) { if (B->end() == It) { Pos = -1; } else { assert(It->getParent() == B); Pos = std::distance(B->begin(), It); } } bool operator<(Loc A) const { if (Block != A.Block) return Block->getNumber() < A.Block->getNumber(); if (A.Pos == -1) return Pos != A.Pos; return Pos != -1 && Pos < A.Pos; } private: int Pos = 0; }; struct HexagonConstExtenders : public MachineFunctionPass { static char ID; HexagonConstExtenders() : MachineFunctionPass(ID) {} void getAnalysisUsage(AnalysisUsage &AU) const override { AU.addRequired(); AU.addPreserved(); MachineFunctionPass::getAnalysisUsage(AU); } StringRef getPassName() const override { return "Hexagon constant-extender optimization"; } bool runOnMachineFunction(MachineFunction &MF) override; private: struct Register { Register() = default; Register(unsigned R, unsigned S) : Reg(R), Sub(S) {} Register(const MachineOperand &Op) : Reg(Op.getReg()), Sub(Op.getSubReg()) {} Register &operator=(const MachineOperand &Op) { if (Op.isReg()) { Reg = Op.getReg(); Sub = Op.getSubReg(); } else if (Op.isFI()) { Reg = TargetRegisterInfo::index2StackSlot(Op.getIndex()); } return *this; } bool isVReg() const { return Reg != 0 && !TargetRegisterInfo::isStackSlot(Reg) && TargetRegisterInfo::isVirtualRegister(Reg); } bool isSlot() const { return Reg != 0 && TargetRegisterInfo::isStackSlot(Reg); } operator MachineOperand() const { if (isVReg()) return MachineOperand::CreateReg(Reg, /*Def*/false, /*Imp*/false, /*Kill*/false, /*Dead*/false, /*Undef*/false, /*EarlyClobber*/false, Sub); if (TargetRegisterInfo::isStackSlot(Reg)) { int FI = TargetRegisterInfo::stackSlot2Index(Reg); return MachineOperand::CreateFI(FI); } llvm_unreachable("Cannot create MachineOperand"); } bool operator==(Register R) const { return Reg == R.Reg && Sub == R.Sub; } bool operator!=(Register R) const { return !operator==(R); } bool operator<(Register R) const { // For std::map. return Reg < R.Reg || (Reg == R.Reg && Sub < R.Sub); } unsigned Reg = 0, Sub = 0; }; struct ExtExpr { // A subexpression in which the extender is used. In general, this // represents an expression where adding D to the extender will be // equivalent to adding D to the expression as a whole. In other // words, expr(add(##V,D) = add(expr(##V),D). // The original motivation for this are the io/ur addressing modes, // where the offset is extended. Consider the io example: // In memw(Rs+##V), the ##V could be replaced by a register Rt to // form the rr mode: memw(Rt+Rs<<0). In such case, however, the // register Rt must have exactly the value of ##V. If there was // another instruction memw(Rs+##V+4), it would need a different Rt. // Now, if Rt was initialized as "##V+Rs<<0", both of these // instructions could use the same Rt, just with different offsets. // Here it's clear that "initializer+4" should be the same as if // the offset 4 was added to the ##V in the initializer. // The only kinds of expressions that support the requirement of // commuting with addition are addition and subtraction from ##V. // Include shifting the Rs to account for the ur addressing mode: // ##Val + Rs << S // ##Val - Rs Register Rs; unsigned S = 0; bool Neg = false; ExtExpr() = default; ExtExpr(Register RS, bool NG, unsigned SH) : Rs(RS), S(SH), Neg(NG) {} // Expression is trivial if it does not modify the extender. bool trivial() const { return Rs.Reg == 0; } bool operator==(const ExtExpr &Ex) const { return Rs == Ex.Rs && S == Ex.S && Neg == Ex.Neg; } bool operator!=(const ExtExpr &Ex) const { return !operator==(Ex); } bool operator<(const ExtExpr &Ex) const { if (Rs != Ex.Rs) return Rs < Ex.Rs; if (S != Ex.S) return S < Ex.S; return !Neg && Ex.Neg; } }; struct ExtDesc { MachineInstr *UseMI = nullptr; unsigned OpNum = -1u; // The subexpression in which the extender is used (e.g. address // computation). ExtExpr Expr; // Optional register that is assigned the value of Expr. Register Rd; // Def means that the output of the instruction may differ from the // original by a constant c, and that the difference can be corrected // by adding/subtracting c in all users of the defined register. bool IsDef = false; MachineOperand &getOp() { return UseMI->getOperand(OpNum); } const MachineOperand &getOp() const { return UseMI->getOperand(OpNum); } }; struct ExtRoot { union { const ConstantFP *CFP; // MO_FPImmediate const char *SymbolName; // MO_ExternalSymbol const GlobalValue *GV; // MO_GlobalAddress const BlockAddress *BA; // MO_BlockAddress int64_t ImmVal; // MO_Immediate, MO_TargetIndex, // and MO_ConstantPoolIndex } V; unsigned Kind; // Same as in MachineOperand. unsigned char TF; // TargetFlags. ExtRoot(const MachineOperand &Op); bool operator==(const ExtRoot &ER) const { return Kind == ER.Kind && V.ImmVal == ER.V.ImmVal; } bool operator!=(const ExtRoot &ER) const { return !operator==(ER); } bool operator<(const ExtRoot &ER) const; }; struct ExtValue : public ExtRoot { int32_t Offset; ExtValue(const MachineOperand &Op); ExtValue(const ExtDesc &ED) : ExtValue(ED.getOp()) {} ExtValue(const ExtRoot &ER, int32_t Off) : ExtRoot(ER), Offset(Off) {} bool operator<(const ExtValue &EV) const; bool operator==(const ExtValue &EV) const { return ExtRoot(*this) == ExtRoot(EV) && Offset == EV.Offset; } bool operator!=(const ExtValue &EV) const { return !operator==(EV); } explicit operator MachineOperand() const; }; using IndexList = SetVector; using ExtenderInit = std::pair; using AssignmentMap = std::map; using LocDefList = std::vector>; const HexagonInstrInfo *HII = nullptr; const HexagonRegisterInfo *HRI = nullptr; MachineDominatorTree *MDT = nullptr; MachineRegisterInfo *MRI = nullptr; std::vector Extenders; std::vector NewRegs; bool isStoreImmediate(unsigned Opc) const; bool isRegOffOpcode(unsigned ExtOpc) const ; unsigned getRegOffOpcode(unsigned ExtOpc) const; unsigned getDirectRegReplacement(unsigned ExtOpc) const; OffsetRange getOffsetRange(Register R, const MachineInstr &MI) const; OffsetRange getOffsetRange(const ExtDesc &ED) const; OffsetRange getOffsetRange(Register Rd) const; void recordExtender(MachineInstr &MI, unsigned OpNum); void collectInstr(MachineInstr &MI); void collect(MachineFunction &MF); void assignInits(const ExtRoot &ER, unsigned Begin, unsigned End, AssignmentMap &IMap); void calculatePlacement(const ExtenderInit &ExtI, const IndexList &Refs, LocDefList &Defs); Register insertInitializer(Loc DefL, const ExtenderInit &ExtI); bool replaceInstrExact(const ExtDesc &ED, Register ExtR); bool replaceInstrExpr(const ExtDesc &ED, const ExtenderInit &ExtI, Register ExtR, int32_t &Diff); bool replaceInstr(unsigned Idx, Register ExtR, const ExtenderInit &ExtI); bool replaceExtenders(const AssignmentMap &IMap); unsigned getOperandIndex(const MachineInstr &MI, const MachineOperand &Op) const; const MachineOperand &getPredicateOp(const MachineInstr &MI) const; const MachineOperand &getLoadResultOp(const MachineInstr &MI) const; const MachineOperand &getStoredValueOp(const MachineInstr &MI) const; friend struct PrintRegister; friend struct PrintExpr; friend struct PrintInit; friend struct PrintIMap; friend raw_ostream &operator<< (raw_ostream &OS, const struct PrintRegister &P); friend raw_ostream &operator<< (raw_ostream &OS, const struct PrintExpr &P); friend raw_ostream &operator<< (raw_ostream &OS, const struct PrintInit &P); friend raw_ostream &operator<< (raw_ostream &OS, const ExtDesc &ED); friend raw_ostream &operator<< (raw_ostream &OS, const ExtRoot &ER); friend raw_ostream &operator<< (raw_ostream &OS, const ExtValue &EV); friend raw_ostream &operator<< (raw_ostream &OS, const OffsetRange &OR); friend raw_ostream &operator<< (raw_ostream &OS, const struct PrintIMap &P); }; using HCE = HexagonConstExtenders; LLVM_ATTRIBUTE_UNUSED raw_ostream &operator<< (raw_ostream &OS, const OffsetRange &OR) { if (OR.Min > OR.Max) OS << '!'; OS << '[' << OR.Min << ',' << OR.Max << "]a" << unsigned(OR.Align) << '+' << unsigned(OR.Offset); return OS; } struct PrintRegister { PrintRegister(HCE::Register R, const HexagonRegisterInfo &I) : Rs(R), HRI(I) {} HCE::Register Rs; const HexagonRegisterInfo &HRI; }; LLVM_ATTRIBUTE_UNUSED raw_ostream &operator<< (raw_ostream &OS, const PrintRegister &P) { if (P.Rs.Reg != 0) OS << printReg(P.Rs.Reg, &P.HRI, P.Rs.Sub); else OS << "noreg"; return OS; } struct PrintExpr { PrintExpr(const HCE::ExtExpr &E, const HexagonRegisterInfo &I) : Ex(E), HRI(I) {} const HCE::ExtExpr &Ex; const HexagonRegisterInfo &HRI; }; LLVM_ATTRIBUTE_UNUSED raw_ostream &operator<< (raw_ostream &OS, const PrintExpr &P) { OS << "## " << (P.Ex.Neg ? "- " : "+ "); if (P.Ex.Rs.Reg != 0) OS << printReg(P.Ex.Rs.Reg, &P.HRI, P.Ex.Rs.Sub); else OS << "__"; OS << " << " << P.Ex.S; return OS; } struct PrintInit { PrintInit(const HCE::ExtenderInit &EI, const HexagonRegisterInfo &I) : ExtI(EI), HRI(I) {} const HCE::ExtenderInit &ExtI; const HexagonRegisterInfo &HRI; }; LLVM_ATTRIBUTE_UNUSED raw_ostream &operator<< (raw_ostream &OS, const PrintInit &P) { OS << '[' << P.ExtI.first << ", " << PrintExpr(P.ExtI.second, P.HRI) << ']'; return OS; } LLVM_ATTRIBUTE_UNUSED raw_ostream &operator<< (raw_ostream &OS, const HCE::ExtDesc &ED) { assert(ED.OpNum != -1u); const MachineBasicBlock &MBB = *ED.getOp().getParent()->getParent(); const MachineFunction &MF = *MBB.getParent(); const auto &HRI = *MF.getSubtarget().getRegisterInfo(); OS << "bb#" << MBB.getNumber() << ": "; if (ED.Rd.Reg != 0) OS << printReg(ED.Rd.Reg, &HRI, ED.Rd.Sub); else OS << "__"; OS << " = " << PrintExpr(ED.Expr, HRI); if (ED.IsDef) OS << ", def"; return OS; } LLVM_ATTRIBUTE_UNUSED raw_ostream &operator<< (raw_ostream &OS, const HCE::ExtRoot &ER) { switch (ER.Kind) { case MachineOperand::MO_Immediate: OS << "imm:" << ER.V.ImmVal; break; case MachineOperand::MO_FPImmediate: OS << "fpi:" << *ER.V.CFP; break; case MachineOperand::MO_ExternalSymbol: OS << "sym:" << *ER.V.SymbolName; break; case MachineOperand::MO_GlobalAddress: OS << "gad:" << ER.V.GV->getName(); break; case MachineOperand::MO_BlockAddress: OS << "blk:" << *ER.V.BA; break; case MachineOperand::MO_TargetIndex: OS << "tgi:" << ER.V.ImmVal; break; case MachineOperand::MO_ConstantPoolIndex: OS << "cpi:" << ER.V.ImmVal; break; case MachineOperand::MO_JumpTableIndex: OS << "jti:" << ER.V.ImmVal; break; default: OS << "???:" << ER.V.ImmVal; break; } return OS; } LLVM_ATTRIBUTE_UNUSED raw_ostream &operator<< (raw_ostream &OS, const HCE::ExtValue &EV) { OS << HCE::ExtRoot(EV) << " off:" << EV.Offset; return OS; } struct PrintIMap { PrintIMap(const HCE::AssignmentMap &M, const HexagonRegisterInfo &I) : IMap(M), HRI(I) {} const HCE::AssignmentMap &IMap; const HexagonRegisterInfo &HRI; }; LLVM_ATTRIBUTE_UNUSED raw_ostream &operator<< (raw_ostream &OS, const PrintIMap &P) { OS << "{\n"; for (const std::pair &Q : P.IMap) { OS << " " << PrintInit(Q.first, P.HRI) << " -> {"; for (unsigned I : Q.second) OS << ' ' << I; OS << " }\n"; } OS << "}\n"; return OS; } } INITIALIZE_PASS_BEGIN(HexagonConstExtenders, "hexagon-cext-opt", "Hexagon constant-extender optimization", false, false) INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) INITIALIZE_PASS_END(HexagonConstExtenders, "hexagon-cext-opt", "Hexagon constant-extender optimization", false, false) static unsigned ReplaceCounter = 0; char HCE::ID = 0; #ifndef NDEBUG LLVM_DUMP_METHOD void RangeTree::dump() const { dbgs() << "Root: " << Root << '\n'; if (Root) dump(Root); } LLVM_DUMP_METHOD void RangeTree::dump(const Node *N) const { dbgs() << "Node: " << N << '\n'; dbgs() << " Height: " << N->Height << '\n'; dbgs() << " Count: " << N->Count << '\n'; dbgs() << " MaxEnd: " << N->MaxEnd << '\n'; dbgs() << " Range: " << N->Range << '\n'; dbgs() << " Left: " << N->Left << '\n'; dbgs() << " Right: " << N->Right << "\n\n"; if (N->Left) dump(N->Left); if (N->Right) dump(N->Right); } #endif void RangeTree::order(Node *N, SmallVectorImpl &Seq) const { if (N == nullptr) return; order(N->Left, Seq); Seq.push_back(N); order(N->Right, Seq); } void RangeTree::nodesWith(Node *N, int32_t P, bool CheckA, SmallVectorImpl &Seq) const { if (N == nullptr || N->MaxEnd < P) return; nodesWith(N->Left, P, CheckA, Seq); if (N->Range.Min <= P) { if ((CheckA && N->Range.contains(P)) || (!CheckA && P <= N->Range.Max)) Seq.push_back(N); nodesWith(N->Right, P, CheckA, Seq); } } RangeTree::Node *RangeTree::add(Node *N, const OffsetRange &R) { if (N == nullptr) return new Node(R); if (N->Range == R) { N->Count++; return N; } if (R < N->Range) N->Left = add(N->Left, R); else N->Right = add(N->Right, R); return rebalance(update(N)); } RangeTree::Node *RangeTree::remove(Node *N, const Node *D) { assert(N != nullptr); if (N != D) { assert(N->Range != D->Range && "N and D should not be equal"); if (D->Range < N->Range) N->Left = remove(N->Left, D); else N->Right = remove(N->Right, D); return rebalance(update(N)); } // We got to the node we need to remove. If any of its children are // missing, simply replace it with the other child. if (N->Left == nullptr || N->Right == nullptr) return (N->Left == nullptr) ? N->Right : N->Left; // Find the rightmost child of N->Left, remove it and plug it in place // of N. Node *M = N->Left; while (M->Right) M = M->Right; M->Left = remove(N->Left, M); M->Right = N->Right; return rebalance(update(M)); } RangeTree::Node *RangeTree::rotateLeft(Node *Lower, Node *Higher) { assert(Higher->Right == Lower); // The Lower node is on the right from Higher. Make sure that Lower's // balance is greater to the right. Otherwise the rotation will create // an unbalanced tree again. if (height(Lower->Left) > height(Lower->Right)) Lower = rotateRight(Lower->Left, Lower); assert(height(Lower->Left) <= height(Lower->Right)); Higher->Right = Lower->Left; update(Higher); Lower->Left = Higher; update(Lower); return Lower; } RangeTree::Node *RangeTree::rotateRight(Node *Lower, Node *Higher) { assert(Higher->Left == Lower); // The Lower node is on the left from Higher. Make sure that Lower's // balance is greater to the left. Otherwise the rotation will create // an unbalanced tree again. if (height(Lower->Left) < height(Lower->Right)) Lower = rotateLeft(Lower->Right, Lower); assert(height(Lower->Left) >= height(Lower->Right)); Higher->Left = Lower->Right; update(Higher); Lower->Right = Higher; update(Lower); return Lower; } HCE::ExtRoot::ExtRoot(const MachineOperand &Op) { // Always store ImmVal, since it's the field used for comparisons. V.ImmVal = 0; if (Op.isImm()) ; // Keep 0. Do not use Op.getImm() for value here (treat 0 as the root). else if (Op.isFPImm()) V.CFP = Op.getFPImm(); else if (Op.isSymbol()) V.SymbolName = Op.getSymbolName(); else if (Op.isGlobal()) V.GV = Op.getGlobal(); else if (Op.isBlockAddress()) V.BA = Op.getBlockAddress(); else if (Op.isCPI() || Op.isTargetIndex() || Op.isJTI()) V.ImmVal = Op.getIndex(); else llvm_unreachable("Unexpected operand type"); Kind = Op.getType(); TF = Op.getTargetFlags(); } bool HCE::ExtRoot::operator< (const HCE::ExtRoot &ER) const { if (Kind != ER.Kind) return Kind < ER.Kind; switch (Kind) { case MachineOperand::MO_Immediate: case MachineOperand::MO_TargetIndex: case MachineOperand::MO_ConstantPoolIndex: case MachineOperand::MO_JumpTableIndex: return V.ImmVal < ER.V.ImmVal; case MachineOperand::MO_FPImmediate: { const APFloat &ThisF = V.CFP->getValueAPF(); const APFloat &OtherF = ER.V.CFP->getValueAPF(); return ThisF.bitcastToAPInt().ult(OtherF.bitcastToAPInt()); } case MachineOperand::MO_ExternalSymbol: return StringRef(V.SymbolName) < StringRef(ER.V.SymbolName); case MachineOperand::MO_GlobalAddress: // Do not use GUIDs, since they depend on the source path. Moving the // source file to a different directory could cause different GUID // values for a pair of given symbols. These symbols could then compare // "less" in one directory, but "greater" in another. assert(!V.GV->getName().empty() && !ER.V.GV->getName().empty()); return V.GV->getName() < ER.V.GV->getName(); case MachineOperand::MO_BlockAddress: { const BasicBlock *ThisB = V.BA->getBasicBlock(); const BasicBlock *OtherB = ER.V.BA->getBasicBlock(); assert(ThisB->getParent() == OtherB->getParent()); const Function &F = *ThisB->getParent(); return std::distance(F.begin(), ThisB->getIterator()) < std::distance(F.begin(), OtherB->getIterator()); } } return V.ImmVal < ER.V.ImmVal; } HCE::ExtValue::ExtValue(const MachineOperand &Op) : ExtRoot(Op) { if (Op.isImm()) Offset = Op.getImm(); else if (Op.isFPImm() || Op.isJTI()) Offset = 0; else if (Op.isSymbol() || Op.isGlobal() || Op.isBlockAddress() || Op.isCPI() || Op.isTargetIndex()) Offset = Op.getOffset(); else llvm_unreachable("Unexpected operand type"); } bool HCE::ExtValue::operator< (const HCE::ExtValue &EV) const { const ExtRoot &ER = *this; if (!(ER == ExtRoot(EV))) return ER < EV; return Offset < EV.Offset; } HCE::ExtValue::operator MachineOperand() const { switch (Kind) { case MachineOperand::MO_Immediate: return MachineOperand::CreateImm(V.ImmVal + Offset); case MachineOperand::MO_FPImmediate: assert(Offset == 0); return MachineOperand::CreateFPImm(V.CFP); case MachineOperand::MO_ExternalSymbol: assert(Offset == 0); return MachineOperand::CreateES(V.SymbolName, TF); case MachineOperand::MO_GlobalAddress: return MachineOperand::CreateGA(V.GV, Offset, TF); case MachineOperand::MO_BlockAddress: return MachineOperand::CreateBA(V.BA, Offset, TF); case MachineOperand::MO_TargetIndex: return MachineOperand::CreateTargetIndex(V.ImmVal, Offset, TF); case MachineOperand::MO_ConstantPoolIndex: return MachineOperand::CreateCPI(V.ImmVal, Offset, TF); case MachineOperand::MO_JumpTableIndex: assert(Offset == 0); return MachineOperand::CreateJTI(V.ImmVal, TF); default: llvm_unreachable("Unhandled kind"); } } bool HCE::isStoreImmediate(unsigned Opc) const { switch (Opc) { case Hexagon::S4_storeirbt_io: case Hexagon::S4_storeirbf_io: case Hexagon::S4_storeirht_io: case Hexagon::S4_storeirhf_io: case Hexagon::S4_storeirit_io: case Hexagon::S4_storeirif_io: case Hexagon::S4_storeirb_io: case Hexagon::S4_storeirh_io: case Hexagon::S4_storeiri_io: return true; default: break; } return false; } bool HCE::isRegOffOpcode(unsigned Opc) const { switch (Opc) { case Hexagon::L2_loadrub_io: case Hexagon::L2_loadrb_io: case Hexagon::L2_loadruh_io: case Hexagon::L2_loadrh_io: case Hexagon::L2_loadri_io: case Hexagon::L2_loadrd_io: case Hexagon::L2_loadbzw2_io: case Hexagon::L2_loadbzw4_io: case Hexagon::L2_loadbsw2_io: case Hexagon::L2_loadbsw4_io: case Hexagon::L2_loadalignh_io: case Hexagon::L2_loadalignb_io: case Hexagon::L2_ploadrubt_io: case Hexagon::L2_ploadrubf_io: case Hexagon::L2_ploadrbt_io: case Hexagon::L2_ploadrbf_io: case Hexagon::L2_ploadruht_io: case Hexagon::L2_ploadruhf_io: case Hexagon::L2_ploadrht_io: case Hexagon::L2_ploadrhf_io: case Hexagon::L2_ploadrit_io: case Hexagon::L2_ploadrif_io: case Hexagon::L2_ploadrdt_io: case Hexagon::L2_ploadrdf_io: case Hexagon::S2_storerb_io: case Hexagon::S2_storerh_io: case Hexagon::S2_storerf_io: case Hexagon::S2_storeri_io: case Hexagon::S2_storerd_io: case Hexagon::S2_pstorerbt_io: case Hexagon::S2_pstorerbf_io: case Hexagon::S2_pstorerht_io: case Hexagon::S2_pstorerhf_io: case Hexagon::S2_pstorerft_io: case Hexagon::S2_pstorerff_io: case Hexagon::S2_pstorerit_io: case Hexagon::S2_pstorerif_io: case Hexagon::S2_pstorerdt_io: case Hexagon::S2_pstorerdf_io: case Hexagon::A2_addi: return true; default: break; } return false; } unsigned HCE::getRegOffOpcode(unsigned ExtOpc) const { // If there exists an instruction that takes a register and offset, // that corresponds to the ExtOpc, return it, otherwise return 0. using namespace Hexagon; switch (ExtOpc) { case A2_tfrsi: return A2_addi; default: break; } const MCInstrDesc &D = HII->get(ExtOpc); if (D.mayLoad() || D.mayStore()) { uint64_t F = D.TSFlags; unsigned AM = (F >> HexagonII::AddrModePos) & HexagonII::AddrModeMask; switch (AM) { case HexagonII::Absolute: case HexagonII::AbsoluteSet: case HexagonII::BaseLongOffset: switch (ExtOpc) { case PS_loadrubabs: case L4_loadrub_ap: case L4_loadrub_ur: return L2_loadrub_io; case PS_loadrbabs: case L4_loadrb_ap: case L4_loadrb_ur: return L2_loadrb_io; case PS_loadruhabs: case L4_loadruh_ap: case L4_loadruh_ur: return L2_loadruh_io; case PS_loadrhabs: case L4_loadrh_ap: case L4_loadrh_ur: return L2_loadrh_io; case PS_loadriabs: case L4_loadri_ap: case L4_loadri_ur: return L2_loadri_io; case PS_loadrdabs: case L4_loadrd_ap: case L4_loadrd_ur: return L2_loadrd_io; case L4_loadbzw2_ap: case L4_loadbzw2_ur: return L2_loadbzw2_io; case L4_loadbzw4_ap: case L4_loadbzw4_ur: return L2_loadbzw4_io; case L4_loadbsw2_ap: case L4_loadbsw2_ur: return L2_loadbsw2_io; case L4_loadbsw4_ap: case L4_loadbsw4_ur: return L2_loadbsw4_io; case L4_loadalignh_ap: case L4_loadalignh_ur: return L2_loadalignh_io; case L4_loadalignb_ap: case L4_loadalignb_ur: return L2_loadalignb_io; case L4_ploadrubt_abs: return L2_ploadrubt_io; case L4_ploadrubf_abs: return L2_ploadrubf_io; case L4_ploadrbt_abs: return L2_ploadrbt_io; case L4_ploadrbf_abs: return L2_ploadrbf_io; case L4_ploadruht_abs: return L2_ploadruht_io; case L4_ploadruhf_abs: return L2_ploadruhf_io; case L4_ploadrht_abs: return L2_ploadrht_io; case L4_ploadrhf_abs: return L2_ploadrhf_io; case L4_ploadrit_abs: return L2_ploadrit_io; case L4_ploadrif_abs: return L2_ploadrif_io; case L4_ploadrdt_abs: return L2_ploadrdt_io; case L4_ploadrdf_abs: return L2_ploadrdf_io; case PS_storerbabs: case S4_storerb_ap: case S4_storerb_ur: return S2_storerb_io; case PS_storerhabs: case S4_storerh_ap: case S4_storerh_ur: return S2_storerh_io; case PS_storerfabs: case S4_storerf_ap: case S4_storerf_ur: return S2_storerf_io; case PS_storeriabs: case S4_storeri_ap: case S4_storeri_ur: return S2_storeri_io; case PS_storerdabs: case S4_storerd_ap: case S4_storerd_ur: return S2_storerd_io; case S4_pstorerbt_abs: return S2_pstorerbt_io; case S4_pstorerbf_abs: return S2_pstorerbf_io; case S4_pstorerht_abs: return S2_pstorerht_io; case S4_pstorerhf_abs: return S2_pstorerhf_io; case S4_pstorerft_abs: return S2_pstorerft_io; case S4_pstorerff_abs: return S2_pstorerff_io; case S4_pstorerit_abs: return S2_pstorerit_io; case S4_pstorerif_abs: return S2_pstorerif_io; case S4_pstorerdt_abs: return S2_pstorerdt_io; case S4_pstorerdf_abs: return S2_pstorerdf_io; default: break; } break; case HexagonII::BaseImmOffset: if (!isStoreImmediate(ExtOpc)) return ExtOpc; break; default: break; } } return 0; } unsigned HCE::getDirectRegReplacement(unsigned ExtOpc) const { switch (ExtOpc) { case Hexagon::A2_addi: return Hexagon::A2_add; case Hexagon::A2_andir: return Hexagon::A2_and; case Hexagon::A2_combineii: return Hexagon::A4_combineri; case Hexagon::A2_orir: return Hexagon::A2_or; case Hexagon::A2_paddif: return Hexagon::A2_paddf; case Hexagon::A2_paddit: return Hexagon::A2_paddt; case Hexagon::A2_subri: return Hexagon::A2_sub; case Hexagon::A2_tfrsi: return TargetOpcode::COPY; case Hexagon::A4_cmpbeqi: return Hexagon::A4_cmpbeq; case Hexagon::A4_cmpbgti: return Hexagon::A4_cmpbgt; case Hexagon::A4_cmpbgtui: return Hexagon::A4_cmpbgtu; case Hexagon::A4_cmpheqi: return Hexagon::A4_cmpheq; case Hexagon::A4_cmphgti: return Hexagon::A4_cmphgt; case Hexagon::A4_cmphgtui: return Hexagon::A4_cmphgtu; case Hexagon::A4_combineii: return Hexagon::A4_combineir; case Hexagon::A4_combineir: return TargetOpcode::REG_SEQUENCE; case Hexagon::A4_combineri: return TargetOpcode::REG_SEQUENCE; case Hexagon::A4_rcmpeqi: return Hexagon::A4_rcmpeq; case Hexagon::A4_rcmpneqi: return Hexagon::A4_rcmpneq; case Hexagon::C2_cmoveif: return Hexagon::A2_tfrpf; case Hexagon::C2_cmoveit: return Hexagon::A2_tfrpt; case Hexagon::C2_cmpeqi: return Hexagon::C2_cmpeq; case Hexagon::C2_cmpgti: return Hexagon::C2_cmpgt; case Hexagon::C2_cmpgtui: return Hexagon::C2_cmpgtu; case Hexagon::C2_muxii: return Hexagon::C2_muxir; case Hexagon::C2_muxir: return Hexagon::C2_mux; case Hexagon::C2_muxri: return Hexagon::C2_mux; case Hexagon::C4_cmpltei: return Hexagon::C4_cmplte; case Hexagon::C4_cmplteui: return Hexagon::C4_cmplteu; case Hexagon::C4_cmpneqi: return Hexagon::C4_cmpneq; case Hexagon::M2_accii: return Hexagon::M2_acci; // T -> T /* No M2_macsin */ case Hexagon::M2_macsip: return Hexagon::M2_maci; // T -> T case Hexagon::M2_mpysin: return Hexagon::M2_mpyi; case Hexagon::M2_mpysip: return Hexagon::M2_mpyi; case Hexagon::M2_mpysmi: return Hexagon::M2_mpyi; case Hexagon::M2_naccii: return Hexagon::M2_nacci; // T -> T case Hexagon::M4_mpyri_addi: return Hexagon::M4_mpyri_addr; case Hexagon::M4_mpyri_addr: return Hexagon::M4_mpyrr_addr; // _ -> T case Hexagon::M4_mpyrr_addi: return Hexagon::M4_mpyrr_addr; // _ -> T case Hexagon::S4_addaddi: return Hexagon::M2_acci; // _ -> T case Hexagon::S4_addi_asl_ri: return Hexagon::S2_asl_i_r_acc; // T -> T case Hexagon::S4_addi_lsr_ri: return Hexagon::S2_lsr_i_r_acc; // T -> T case Hexagon::S4_andi_asl_ri: return Hexagon::S2_asl_i_r_and; // T -> T case Hexagon::S4_andi_lsr_ri: return Hexagon::S2_lsr_i_r_and; // T -> T case Hexagon::S4_ori_asl_ri: return Hexagon::S2_asl_i_r_or; // T -> T case Hexagon::S4_ori_lsr_ri: return Hexagon::S2_lsr_i_r_or; // T -> T case Hexagon::S4_subaddi: return Hexagon::M2_subacc; // _ -> T case Hexagon::S4_subi_asl_ri: return Hexagon::S2_asl_i_r_nac; // T -> T case Hexagon::S4_subi_lsr_ri: return Hexagon::S2_lsr_i_r_nac; // T -> T // Store-immediates: case Hexagon::S4_storeirbf_io: return Hexagon::S2_pstorerbf_io; case Hexagon::S4_storeirb_io: return Hexagon::S2_storerb_io; case Hexagon::S4_storeirbt_io: return Hexagon::S2_pstorerbt_io; case Hexagon::S4_storeirhf_io: return Hexagon::S2_pstorerhf_io; case Hexagon::S4_storeirh_io: return Hexagon::S2_storerh_io; case Hexagon::S4_storeirht_io: return Hexagon::S2_pstorerht_io; case Hexagon::S4_storeirif_io: return Hexagon::S2_pstorerif_io; case Hexagon::S4_storeiri_io: return Hexagon::S2_storeri_io; case Hexagon::S4_storeirit_io: return Hexagon::S2_pstorerit_io; default: break; } return 0; } // Return the allowable deviation from the current value of Rb (i.e. the // range of values that can be added to the current value) which the // instruction MI can accommodate. // The instruction MI is a user of register Rb, which is defined via an // extender. It may be possible for MI to be tweaked to work for a register // defined with a slightly different value. For example // ... = L2_loadrub_io Rb, 1 // can be modifed to be // ... = L2_loadrub_io Rb', 0 // if Rb' = Rb+1. // The range for Rb would be [Min+1, Max+1], where [Min, Max] is a range // for L2_loadrub with offset 0. That means that Rb could be replaced with // Rc, where Rc-Rb belongs to [Min+1, Max+1]. OffsetRange HCE::getOffsetRange(Register Rb, const MachineInstr &MI) const { unsigned Opc = MI.getOpcode(); // Instructions that are constant-extended may be replaced with something // else that no longer offers the same range as the original. if (!isRegOffOpcode(Opc) || HII->isConstExtended(MI)) return OffsetRange::zero(); if (Opc == Hexagon::A2_addi) { const MachineOperand &Op1 = MI.getOperand(1), &Op2 = MI.getOperand(2); if (Rb != Register(Op1) || !Op2.isImm()) return OffsetRange::zero(); OffsetRange R = { -(1<<15)+1, (1<<15)-1, 1 }; return R.shift(Op2.getImm()); } // HII::getBaseAndOffsetPosition returns the increment position as "offset". if (HII->isPostIncrement(MI)) return OffsetRange::zero(); const MCInstrDesc &D = HII->get(Opc); assert(D.mayLoad() || D.mayStore()); unsigned BaseP, OffP; if (!HII->getBaseAndOffsetPosition(MI, BaseP, OffP) || Rb != Register(MI.getOperand(BaseP)) || !MI.getOperand(OffP).isImm()) return OffsetRange::zero(); uint64_t F = (D.TSFlags >> HexagonII::MemAccessSizePos) & HexagonII::MemAccesSizeMask; uint8_t A = HexagonII::getMemAccessSizeInBytes(HexagonII::MemAccessSize(F)); unsigned L = Log2_32(A); unsigned S = 10+L; // sint11_L int32_t Min = -alignDown((1<= 0 ? 0 : -Off; OffsetRange R = { Min, Max, A }; return R.shift(Off); } // Return the allowable deviation from the current value of the extender ED, // for which the instruction corresponding to ED can be modified without // using an extender. // The instruction uses the extender directly. It will be replaced with // another instruction, say MJ, where the extender will be replaced with a // register. MJ can allow some variability with respect to the value of // that register, as is the case with indexed memory instructions. OffsetRange HCE::getOffsetRange(const ExtDesc &ED) const { // The only way that there can be a non-zero range available is if // the instruction using ED will be converted to an indexed memory // instruction. unsigned IdxOpc = getRegOffOpcode(ED.UseMI->getOpcode()); switch (IdxOpc) { case 0: return OffsetRange::zero(); case Hexagon::A2_addi: // s16 return { -32767, 32767, 1 }; case Hexagon::A2_subri: // s10 return { -511, 511, 1 }; } if (!ED.UseMI->mayLoad() && !ED.UseMI->mayStore()) return OffsetRange::zero(); const MCInstrDesc &D = HII->get(IdxOpc); uint64_t F = (D.TSFlags >> HexagonII::MemAccessSizePos) & HexagonII::MemAccesSizeMask; uint8_t A = HexagonII::getMemAccessSizeInBytes(HexagonII::MemAccessSize(F)); unsigned L = Log2_32(A); unsigned S = 10+L; // sint11_L int32_t Min = -alignDown((1<use_operands(Rd.Reg)) { // Make sure that the register being used by this operand is identical // to the register that was defined: using a different subregister // precludes any non-trivial range. if (Rd != Register(Op)) return OffsetRange::zero(); Range.intersect(getOffsetRange(Rd, *Op.getParent())); } return Range; } void HCE::recordExtender(MachineInstr &MI, unsigned OpNum) { unsigned Opc = MI.getOpcode(); ExtDesc ED; ED.OpNum = OpNum; bool IsLoad = MI.mayLoad(); bool IsStore = MI.mayStore(); // Fixed stack slots have negative indexes, and they cannot be used // with TRI::stackSlot2Index and TRI::index2StackSlot. This is somewhat // unfortunate, but should not be a frequent thing. for (MachineOperand &Op : MI.operands()) if (Op.isFI() && Op.getIndex() < 0) return; if (IsLoad || IsStore) { unsigned AM = HII->getAddrMode(MI); switch (AM) { // (Re: ##Off + Rb<getName().empty()) return; Extenders.push_back(ED); } void HCE::collectInstr(MachineInstr &MI) { if (!HII->isConstExtended(MI)) return; // Skip some non-convertible instructions. unsigned Opc = MI.getOpcode(); switch (Opc) { case Hexagon::M2_macsin: // There is no Rx -= mpyi(Rs,Rt). case Hexagon::C4_addipc: case Hexagon::S4_or_andi: case Hexagon::S4_or_andix: case Hexagon::S4_or_ori: return; } recordExtender(MI, HII->getCExtOpNum(MI)); } void HCE::collect(MachineFunction &MF) { Extenders.clear(); for (MachineBasicBlock &MBB : MF) { // Skip unreachable blocks. if (MBB.getNumber() == -1) continue; for (MachineInstr &MI : MBB) collectInstr(MI); } } void HCE::assignInits(const ExtRoot &ER, unsigned Begin, unsigned End, AssignmentMap &IMap) { // Sanity check: make sure that all extenders in the range [Begin..End) // share the same root ER. for (unsigned I = Begin; I != End; ++I) assert(ER == ExtRoot(Extenders[I].getOp())); // Construct the list of ranges, such that for each P in Ranges[I], // a register Reg = ER+P can be used in place of Extender[I]. If the // instruction allows, uses in the form of Reg+Off are considered // (here, Off = required_value - P). std::vector Ranges(End-Begin); // For each extender that is a def, visit all uses of the defined register, // and produce an offset range that works for all uses. The def doesn't // have to be checked, because it can become dead if all uses can be updated // to use a different reg/offset. for (unsigned I = Begin; I != End; ++I) { const ExtDesc &ED = Extenders[I]; if (!ED.IsDef) continue; ExtValue EV(ED); LLVM_DEBUG(dbgs() << " =" << I << ". " << EV << " " << ED << '\n'); assert(ED.Rd.Reg != 0); Ranges[I-Begin] = getOffsetRange(ED.Rd).shift(EV.Offset); // A2_tfrsi is a special case: it will be replaced with A2_addi, which // has a 16-bit signed offset. This means that A2_tfrsi not only has a // range coming from its uses, but also from the fact that its replacement // has a range as well. if (ED.UseMI->getOpcode() == Hexagon::A2_tfrsi) { int32_t D = alignDown(32767, Ranges[I-Begin].Align); // XXX hardcoded Ranges[I-Begin].extendBy(-D).extendBy(D); } } // Visit all non-def extenders. For each one, determine the offset range // available for it. for (unsigned I = Begin; I != End; ++I) { const ExtDesc &ED = Extenders[I]; if (ED.IsDef) continue; ExtValue EV(ED); LLVM_DEBUG(dbgs() << " " << I << ". " << EV << " " << ED << '\n'); OffsetRange Dev = getOffsetRange(ED); Ranges[I-Begin].intersect(Dev.shift(EV.Offset)); } // Here for each I there is a corresponding Range[I]. Construct the // inverse map, that to each range will assign the set of indexes in // [Begin..End) that this range corresponds to. std::map RangeMap; for (unsigned I = Begin; I != End; ++I) RangeMap[Ranges[I-Begin]].insert(I); LLVM_DEBUG({ dbgs() << "Ranges\n"; for (unsigned I = Begin; I != End; ++I) dbgs() << " " << I << ". " << Ranges[I-Begin] << '\n'; dbgs() << "RangeMap\n"; for (auto &P : RangeMap) { dbgs() << " " << P.first << " ->"; for (unsigned I : P.second) dbgs() << ' ' << I; dbgs() << '\n'; } }); // Select the definition points, and generate the assignment between // these points and the uses. // For each candidate offset, keep a pair CandData consisting of // the total number of ranges containing that candidate, and the // vector of corresponding RangeTree nodes. using CandData = std::pair>; std::map CandMap; RangeTree Tree; for (const OffsetRange &R : Ranges) Tree.add(R); SmallVector Nodes; Tree.order(Nodes); auto MaxAlign = [](const SmallVectorImpl &Nodes, uint8_t Align, uint8_t Offset) { for (RangeTree::Node *N : Nodes) { if (N->Range.Align <= Align || N->Range.Offset < Offset) continue; if ((N->Range.Offset - Offset) % Align != 0) continue; Align = N->Range.Align; Offset = N->Range.Offset; } return std::make_pair(Align, Offset); }; // Construct the set of all potential definition points from the endpoints // of the ranges. If a given endpoint also belongs to a different range, // but with a higher alignment, also consider the more-highly-aligned // value of this endpoint. std::set CandSet; for (RangeTree::Node *N : Nodes) { const OffsetRange &R = N->Range; auto P0 = MaxAlign(Tree.nodesWith(R.Min, false), R.Align, R.Offset); CandSet.insert(R.Min); if (R.Align < P0.first) CandSet.insert(adjustUp(R.Min, P0.first, P0.second)); auto P1 = MaxAlign(Tree.nodesWith(R.Max, false), R.Align, R.Offset); CandSet.insert(R.Max); if (R.Align < P1.first) CandSet.insert(adjustDown(R.Max, P1.first, P1.second)); } // Build the assignment map: candidate C -> { list of extender indexes }. // This has to be done iteratively: // - pick the candidate that covers the maximum number of extenders, // - add the candidate to the map, // - remove the extenders from the pool. while (true) { using CMap = std::map; CMap Counts; for (auto It = CandSet.begin(), Et = CandSet.end(); It != Et; ) { auto &&V = Tree.nodesWith(*It); unsigned N = std::accumulate(V.begin(), V.end(), 0u, [](unsigned Acc, const RangeTree::Node *N) { return Acc + N->Count; }); if (N != 0) Counts.insert({*It, N}); It = (N != 0) ? std::next(It) : CandSet.erase(It); } if (Counts.empty()) break; // Find the best candidate with respect to the number of extenders covered. auto BestIt = std::max_element(Counts.begin(), Counts.end(), [](const CMap::value_type &A, const CMap::value_type &B) { return A.second < B.second || (A.second == B.second && A < B); }); int32_t Best = BestIt->first; ExtValue BestV(ER, Best); for (RangeTree::Node *N : Tree.nodesWith(Best)) { for (unsigned I : RangeMap[N->Range]) IMap[{BestV,Extenders[I].Expr}].insert(I); Tree.erase(N); } } LLVM_DEBUG(dbgs() << "IMap (before fixup) = " << PrintIMap(IMap, *HRI)); // There is some ambiguity in what initializer should be used, if the // descriptor's subexpression is non-trivial: it can be the entire // subexpression (which is what has been done so far), or it can be // the extender's value itself, if all corresponding extenders have the // exact value of the initializer (i.e. require offset of 0). // To reduce the number of initializers, merge such special cases. for (std::pair &P : IMap) { // Skip trivial initializers. if (P.first.second.trivial()) continue; // If the corresponding trivial initializer does not exist, skip this // entry. const ExtValue &EV = P.first.first; AssignmentMap::iterator F = IMap.find({EV, ExtExpr()}); if (F == IMap.end()) continue; // Finally, check if all extenders have the same value as the initializer. // Make sure that extenders that are a part of a stack address are not // merged with those that aren't. Stack addresses need an offset field // (to be used by frame index elimination), while non-stack expressions // can be replaced with forms (such as rr) that do not have such a field. // Example: // // Collected 3 extenders // =2. imm:0 off:32968 bb#2: %7 = ## + __ << 0, def // 0. imm:0 off:267 bb#0: __ = ## + SS#1 << 0 // 1. imm:0 off:267 bb#1: __ = ## + SS#1 << 0 // Ranges // 0. [-756,267]a1+0 // 1. [-756,267]a1+0 // 2. [201,65735]a1+0 // RangeMap // [-756,267]a1+0 -> 0 1 // [201,65735]a1+0 -> 2 // IMap (before fixup) = { // [imm:0 off:267, ## + __ << 0] -> { 2 } // [imm:0 off:267, ## + SS#1 << 0] -> { 0 1 } // } // IMap (after fixup) = { // [imm:0 off:267, ## + __ << 0] -> { 2 0 1 } // [imm:0 off:267, ## + SS#1 << 0] -> { } // } // Inserted def in bb#0 for initializer: [imm:0 off:267, ## + __ << 0] // %12:intregs = A2_tfrsi 267 // // The result was // %12:intregs = A2_tfrsi 267 // S4_pstorerbt_rr %3, %12, %stack.1, 0, killed %4 // Which became // r0 = #267 // if (p0.new) memb(r0+r29<<#4) = r2 bool IsStack = any_of(F->second, [this](unsigned I) { return Extenders[I].Expr.Rs.isSlot(); }); auto SameValue = [&EV,this,IsStack](unsigned I) { const ExtDesc &ED = Extenders[I]; return ED.Expr.Rs.isSlot() == IsStack && ExtValue(ED).Offset == EV.Offset; }; if (all_of(P.second, SameValue)) { F->second.insert(P.second.begin(), P.second.end()); P.second.clear(); } } LLVM_DEBUG(dbgs() << "IMap (after fixup) = " << PrintIMap(IMap, *HRI)); } void HCE::calculatePlacement(const ExtenderInit &ExtI, const IndexList &Refs, LocDefList &Defs) { if (Refs.empty()) return; // The placement calculation is somewhat simple right now: it finds a // single location for the def that dominates all refs. Since this may // place the def far from the uses, producing several locations for // defs that collectively dominate all refs could be better. // For now only do the single one. DenseSet Blocks; DenseSet RefMIs; const ExtDesc &ED0 = Extenders[Refs[0]]; MachineBasicBlock *DomB = ED0.UseMI->getParent(); RefMIs.insert(ED0.UseMI); Blocks.insert(DomB); for (unsigned i = 1, e = Refs.size(); i != e; ++i) { const ExtDesc &ED = Extenders[Refs[i]]; MachineBasicBlock *MBB = ED.UseMI->getParent(); RefMIs.insert(ED.UseMI); DomB = MDT->findNearestCommonDominator(DomB, MBB); Blocks.insert(MBB); } #ifndef NDEBUG // The block DomB should be dominated by the def of each register used // in the initializer. Register Rs = ExtI.second.Rs; // Only one reg allowed now. const MachineInstr *DefI = Rs.isVReg() ? MRI->getVRegDef(Rs.Reg) : nullptr; // This should be guaranteed given that the entire expression is used // at each instruction in Refs. Add an assertion just in case. assert(!DefI || MDT->dominates(DefI->getParent(), DomB)); #endif MachineBasicBlock::iterator It; if (Blocks.count(DomB)) { // Try to find the latest possible location for the def. MachineBasicBlock::iterator End = DomB->end(); for (It = DomB->begin(); It != End; ++It) if (RefMIs.count(&*It)) break; assert(It != End && "Should have found a ref in DomB"); } else { // DomB does not contain any refs. It = DomB->getFirstTerminator(); } Loc DefLoc(DomB, It); Defs.emplace_back(DefLoc, Refs); } HCE::Register HCE::insertInitializer(Loc DefL, const ExtenderInit &ExtI) { unsigned DefR = MRI->createVirtualRegister(&Hexagon::IntRegsRegClass); MachineBasicBlock &MBB = *DefL.Block; MachineBasicBlock::iterator At = DefL.At; DebugLoc dl = DefL.Block->findDebugLoc(DefL.At); const ExtValue &EV = ExtI.first; MachineOperand ExtOp(EV); const ExtExpr &Ex = ExtI.second; const MachineInstr *InitI = nullptr; if (Ex.Rs.isSlot()) { assert(Ex.S == 0 && "Cannot have a shift of a stack slot"); assert(!Ex.Neg && "Cannot subtract a stack slot"); // DefR = PS_fi Rb,##EV InitI = BuildMI(MBB, At, dl, HII->get(Hexagon::PS_fi), DefR) .add(MachineOperand(Ex.Rs)) .add(ExtOp); } else { assert((Ex.Rs.Reg == 0 || Ex.Rs.isVReg()) && "Expecting virtual register"); if (Ex.trivial()) { // DefR = ##EV InitI = BuildMI(MBB, At, dl, HII->get(Hexagon::A2_tfrsi), DefR) .add(ExtOp); } else if (Ex.S == 0) { if (Ex.Neg) { // DefR = sub(##EV,Rb) InitI = BuildMI(MBB, At, dl, HII->get(Hexagon::A2_subri), DefR) .add(ExtOp) .add(MachineOperand(Ex.Rs)); } else { // DefR = add(Rb,##EV) InitI = BuildMI(MBB, At, dl, HII->get(Hexagon::A2_addi), DefR) .add(MachineOperand(Ex.Rs)) .add(ExtOp); } } else { unsigned NewOpc = Ex.Neg ? Hexagon::S4_subi_asl_ri : Hexagon::S4_addi_asl_ri; // DefR = add(##EV,asl(Rb,S)) InitI = BuildMI(MBB, At, dl, HII->get(NewOpc), DefR) .add(ExtOp) .add(MachineOperand(Ex.Rs)) .addImm(Ex.S); } } assert(InitI); (void)InitI; LLVM_DEBUG(dbgs() << "Inserted def in bb#" << MBB.getNumber() << " for initializer: " << PrintInit(ExtI, *HRI) << "\n " << *InitI); return { DefR, 0 }; } // Replace the extender at index Idx with the register ExtR. bool HCE::replaceInstrExact(const ExtDesc &ED, Register ExtR) { MachineInstr &MI = *ED.UseMI; MachineBasicBlock &MBB = *MI.getParent(); MachineBasicBlock::iterator At = MI.getIterator(); DebugLoc dl = MI.getDebugLoc(); unsigned ExtOpc = MI.getOpcode(); // With a few exceptions, direct replacement amounts to creating an // instruction with a corresponding register opcode, with all operands // the same, except for the register used in place of the extender. unsigned RegOpc = getDirectRegReplacement(ExtOpc); if (RegOpc == TargetOpcode::REG_SEQUENCE) { if (ExtOpc == Hexagon::A4_combineri) BuildMI(MBB, At, dl, HII->get(RegOpc)) .add(MI.getOperand(0)) .add(MI.getOperand(1)) .addImm(Hexagon::isub_hi) .add(MachineOperand(ExtR)) .addImm(Hexagon::isub_lo); else if (ExtOpc == Hexagon::A4_combineir) BuildMI(MBB, At, dl, HII->get(RegOpc)) .add(MI.getOperand(0)) .add(MachineOperand(ExtR)) .addImm(Hexagon::isub_hi) .add(MI.getOperand(2)) .addImm(Hexagon::isub_lo); else llvm_unreachable("Unexpected opcode became REG_SEQUENCE"); MBB.erase(MI); return true; } if (ExtOpc == Hexagon::C2_cmpgei || ExtOpc == Hexagon::C2_cmpgeui) { unsigned NewOpc = ExtOpc == Hexagon::C2_cmpgei ? Hexagon::C2_cmplt : Hexagon::C2_cmpltu; BuildMI(MBB, At, dl, HII->get(NewOpc)) .add(MI.getOperand(0)) .add(MachineOperand(ExtR)) .add(MI.getOperand(1)); MBB.erase(MI); return true; } if (RegOpc != 0) { MachineInstrBuilder MIB = BuildMI(MBB, At, dl, HII->get(RegOpc)); unsigned RegN = ED.OpNum; // Copy all operands except the one that has the extender. for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { if (i != RegN) MIB.add(MI.getOperand(i)); else MIB.add(MachineOperand(ExtR)); } MIB.cloneMemRefs(MI); MBB.erase(MI); return true; } if ((MI.mayLoad() || MI.mayStore()) && !isStoreImmediate(ExtOpc)) { // For memory instructions, there is an asymmetry in the addressing // modes. Addressing modes allowing extenders can be replaced with // addressing modes that use registers, but the order of operands // (or even their number) may be different. // Replacements: // BaseImmOffset (io) -> BaseRegOffset (rr) // BaseLongOffset (ur) -> BaseRegOffset (rr) unsigned RegOpc, Shift; unsigned AM = HII->getAddrMode(MI); if (AM == HexagonII::BaseImmOffset) { RegOpc = HII->changeAddrMode_io_rr(ExtOpc); Shift = 0; } else if (AM == HexagonII::BaseLongOffset) { // Loads: Rd = L4_loadri_ur Rs, S, ## // Stores: S4_storeri_ur Rs, S, ##, Rt RegOpc = HII->changeAddrMode_ur_rr(ExtOpc); Shift = MI.getOperand(MI.mayLoad() ? 2 : 1).getImm(); } else { llvm_unreachable("Unexpected addressing mode"); } #ifndef NDEBUG if (RegOpc == -1u) { dbgs() << "\nExtOpc: " << HII->getName(ExtOpc) << " has no rr version\n"; llvm_unreachable("No corresponding rr instruction"); } #endif unsigned BaseP, OffP; HII->getBaseAndOffsetPosition(MI, BaseP, OffP); // Build an rr instruction: (RegOff + RegBase<<0) MachineInstrBuilder MIB = BuildMI(MBB, At, dl, HII->get(RegOpc)); // First, add the def for loads. if (MI.mayLoad()) MIB.add(getLoadResultOp(MI)); // Handle possible predication. if (HII->isPredicated(MI)) MIB.add(getPredicateOp(MI)); // Build the address. MIB.add(MachineOperand(ExtR)); // RegOff MIB.add(MI.getOperand(BaseP)); // RegBase MIB.addImm(Shift); // << Shift // Add the stored value for stores. if (MI.mayStore()) MIB.add(getStoredValueOp(MI)); MIB.cloneMemRefs(MI); MBB.erase(MI); return true; } #ifndef NDEBUG dbgs() << '\n' << MI; #endif llvm_unreachable("Unhandled exact replacement"); return false; } // Replace the extender ED with a form corresponding to the initializer ExtI. bool HCE::replaceInstrExpr(const ExtDesc &ED, const ExtenderInit &ExtI, Register ExtR, int32_t &Diff) { MachineInstr &MI = *ED.UseMI; MachineBasicBlock &MBB = *MI.getParent(); MachineBasicBlock::iterator At = MI.getIterator(); DebugLoc dl = MI.getDebugLoc(); unsigned ExtOpc = MI.getOpcode(); if (ExtOpc == Hexagon::A2_tfrsi) { // A2_tfrsi is a special case: it's replaced with A2_addi, which introduces // another range. One range is the one that's common to all tfrsi's uses, // this one is the range of immediates in A2_addi. When calculating ranges, // the addi's 16-bit argument was included, so now we need to make it such // that the produced value is in the range for the uses alone. // Most of the time, simply adding Diff will make the addi produce exact // result, but if Diff is outside of the 16-bit range, some adjustment // will be needed. unsigned IdxOpc = getRegOffOpcode(ExtOpc); assert(IdxOpc == Hexagon::A2_addi); // Clamp Diff to the 16 bit range. int32_t D = isInt<16>(Diff) ? Diff : (Diff > 0 ? 32767 : -32768); if (Diff > 32767) { // Split Diff into two values: one that is close to min/max int16, // and the other being the rest, and such that both have the same // "alignment" as Diff. uint32_t UD = Diff; OffsetRange R = getOffsetRange(MI.getOperand(0)); uint32_t A = std::min(R.Align, 1u << countTrailingZeros(UD)); D &= ~(A-1); } BuildMI(MBB, At, dl, HII->get(IdxOpc)) .add(MI.getOperand(0)) .add(MachineOperand(ExtR)) .addImm(D); Diff -= D; #ifndef NDEBUG // Make sure the output is within allowable range for uses. // "Diff" is a difference in the "opposite direction", i.e. Ext - DefV, // not DefV - Ext, as the getOffsetRange would calculate. OffsetRange Uses = getOffsetRange(MI.getOperand(0)); if (!Uses.contains(-Diff)) dbgs() << "Diff: " << -Diff << " out of range " << Uses << " for " << MI; assert(Uses.contains(-Diff)); #endif MBB.erase(MI); return true; } const ExtValue &EV = ExtI.first; (void)EV; const ExtExpr &Ex = ExtI.second; (void)Ex; if (ExtOpc == Hexagon::A2_addi || ExtOpc == Hexagon::A2_subri) { // If addi/subri are replaced with the exactly matching initializer, // they amount to COPY. // Check that the initializer is an exact match (for simplicity). #ifndef NDEBUG bool IsAddi = ExtOpc == Hexagon::A2_addi; const MachineOperand &RegOp = MI.getOperand(IsAddi ? 1 : 2); const MachineOperand &ImmOp = MI.getOperand(IsAddi ? 2 : 1); assert(Ex.Rs == RegOp && EV == ImmOp && Ex.Neg != IsAddi && "Initializer mismatch"); #endif BuildMI(MBB, At, dl, HII->get(TargetOpcode::COPY)) .add(MI.getOperand(0)) .add(MachineOperand(ExtR)); Diff = 0; MBB.erase(MI); return true; } if (ExtOpc == Hexagon::M2_accii || ExtOpc == Hexagon::M2_naccii || ExtOpc == Hexagon::S4_addaddi || ExtOpc == Hexagon::S4_subaddi) { // M2_accii: add(Rt,add(Rs,V)) (tied) // M2_naccii: sub(Rt,add(Rs,V)) // S4_addaddi: add(Rt,add(Rs,V)) // S4_subaddi: add(Rt,sub(V,Rs)) // Check that Rs and V match the initializer expression. The Rs+V is the // combination that is considered "subexpression" for V, although Rx+V // would also be valid. #ifndef NDEBUG bool IsSub = ExtOpc == Hexagon::S4_subaddi; Register Rs = MI.getOperand(IsSub ? 3 : 2); ExtValue V = MI.getOperand(IsSub ? 2 : 3); assert(EV == V && Rs == Ex.Rs && IsSub == Ex.Neg && "Initializer mismatch"); #endif unsigned NewOpc = ExtOpc == Hexagon::M2_naccii ? Hexagon::A2_sub : Hexagon::A2_add; BuildMI(MBB, At, dl, HII->get(NewOpc)) .add(MI.getOperand(0)) .add(MI.getOperand(1)) .add(MachineOperand(ExtR)); MBB.erase(MI); return true; } if (MI.mayLoad() || MI.mayStore()) { unsigned IdxOpc = getRegOffOpcode(ExtOpc); assert(IdxOpc && "Expecting indexed opcode"); MachineInstrBuilder MIB = BuildMI(MBB, At, dl, HII->get(IdxOpc)); // Construct the new indexed instruction. // First, add the def for loads. if (MI.mayLoad()) MIB.add(getLoadResultOp(MI)); // Handle possible predication. if (HII->isPredicated(MI)) MIB.add(getPredicateOp(MI)); // Build the address. MIB.add(MachineOperand(ExtR)); MIB.addImm(Diff); // Add the stored value for stores. if (MI.mayStore()) MIB.add(getStoredValueOp(MI)); MIB.cloneMemRefs(MI); MBB.erase(MI); return true; } #ifndef NDEBUG dbgs() << '\n' << PrintInit(ExtI, *HRI) << " " << MI; #endif llvm_unreachable("Unhandled expr replacement"); return false; } bool HCE::replaceInstr(unsigned Idx, Register ExtR, const ExtenderInit &ExtI) { if (ReplaceLimit.getNumOccurrences()) { if (ReplaceLimit <= ReplaceCounter) return false; ++ReplaceCounter; } const ExtDesc &ED = Extenders[Idx]; assert((!ED.IsDef || ED.Rd.Reg != 0) && "Missing Rd for def"); const ExtValue &DefV = ExtI.first; assert(ExtRoot(ExtValue(ED)) == ExtRoot(DefV) && "Extender root mismatch"); const ExtExpr &DefEx = ExtI.second; ExtValue EV(ED); int32_t Diff = EV.Offset - DefV.Offset; const MachineInstr &MI = *ED.UseMI; LLVM_DEBUG(dbgs() << __func__ << " Idx:" << Idx << " ExtR:" << PrintRegister(ExtR, *HRI) << " Diff:" << Diff << '\n'); // These two addressing modes must be converted into indexed forms // regardless of what the initializer looks like. bool IsAbs = false, IsAbsSet = false; if (MI.mayLoad() || MI.mayStore()) { unsigned AM = HII->getAddrMode(MI); IsAbs = AM == HexagonII::Absolute; IsAbsSet = AM == HexagonII::AbsoluteSet; } // If it's a def, remember all operands that need to be updated. // If ED is a def, and Diff is not 0, then all uses of the register Rd // defined by ED must be in the form (Rd, imm), i.e. the immediate offset // must follow the Rd in the operand list. std::vector> RegOps; if (ED.IsDef && Diff != 0) { for (MachineOperand &Op : MRI->use_operands(ED.Rd.Reg)) { MachineInstr &UI = *Op.getParent(); RegOps.push_back({&UI, getOperandIndex(UI, Op)}); } } // Replace the instruction. bool Replaced = false; if (Diff == 0 && DefEx.trivial() && !IsAbs && !IsAbsSet) Replaced = replaceInstrExact(ED, ExtR); else Replaced = replaceInstrExpr(ED, ExtI, ExtR, Diff); if (Diff != 0 && Replaced && ED.IsDef) { // Update offsets of the def's uses. for (std::pair P : RegOps) { unsigned J = P.second; assert(P.first->getNumOperands() > J+1 && P.first->getOperand(J+1).isImm()); MachineOperand &ImmOp = P.first->getOperand(J+1); ImmOp.setImm(ImmOp.getImm() + Diff); } // If it was an absolute-set instruction, the "set" part has been removed. // ExtR will now be the register with the extended value, and since all // users of Rd have been updated, all that needs to be done is to replace // Rd with ExtR. if (IsAbsSet) { assert(ED.Rd.Sub == 0 && ExtR.Sub == 0); MRI->replaceRegWith(ED.Rd.Reg, ExtR.Reg); } } return Replaced; } bool HCE::replaceExtenders(const AssignmentMap &IMap) { LocDefList Defs; bool Changed = false; for (const std::pair &P : IMap) { const IndexList &Idxs = P.second; if (Idxs.size() < CountThreshold) continue; Defs.clear(); calculatePlacement(P.first, Idxs, Defs); for (const std::pair &Q : Defs) { Register DefR = insertInitializer(Q.first, P.first); NewRegs.push_back(DefR.Reg); for (unsigned I : Q.second) Changed |= replaceInstr(I, DefR, P.first); } } return Changed; } unsigned HCE::getOperandIndex(const MachineInstr &MI, const MachineOperand &Op) const { for (unsigned i = 0, n = MI.getNumOperands(); i != n; ++i) if (&MI.getOperand(i) == &Op) return i; llvm_unreachable("Not an operand of MI"); } const MachineOperand &HCE::getPredicateOp(const MachineInstr &MI) const { assert(HII->isPredicated(MI)); for (const MachineOperand &Op : MI.operands()) { if (!Op.isReg() || !Op.isUse() || MRI->getRegClass(Op.getReg()) != &Hexagon::PredRegsRegClass) continue; assert(Op.getSubReg() == 0 && "Predicate register with a subregister"); return Op; } llvm_unreachable("Predicate operand not found"); } const MachineOperand &HCE::getLoadResultOp(const MachineInstr &MI) const { assert(MI.mayLoad()); return MI.getOperand(0); } const MachineOperand &HCE::getStoredValueOp(const MachineInstr &MI) const { assert(MI.mayStore()); return MI.getOperand(MI.getNumExplicitOperands()-1); } bool HCE::runOnMachineFunction(MachineFunction &MF) { if (skipFunction(MF.getFunction())) return false; if (MF.getFunction().hasPersonalityFn()) { LLVM_DEBUG(dbgs() << getPassName() << ": skipping " << MF.getName() << " due to exception handling\n"); return false; } LLVM_DEBUG(MF.print(dbgs() << "Before " << getPassName() << '\n', nullptr)); HII = MF.getSubtarget().getInstrInfo(); HRI = MF.getSubtarget().getRegisterInfo(); MDT = &getAnalysis(); MRI = &MF.getRegInfo(); AssignmentMap IMap; collect(MF); llvm::sort(Extenders, [this](const ExtDesc &A, const ExtDesc &B) { ExtValue VA(A), VB(B); if (VA != VB) return VA < VB; const MachineInstr *MA = A.UseMI; const MachineInstr *MB = B.UseMI; if (MA == MB) { // If it's the same instruction, compare operand numbers. return A.OpNum < B.OpNum; } const MachineBasicBlock *BA = MA->getParent(); const MachineBasicBlock *BB = MB->getParent(); assert(BA->getNumber() != -1 && BB->getNumber() != -1); if (BA != BB) return BA->getNumber() < BB->getNumber(); return MDT->dominates(MA, MB); }); bool Changed = false; LLVM_DEBUG(dbgs() << "Collected " << Extenders.size() << " extenders\n"); for (unsigned I = 0, E = Extenders.size(); I != E; ) { unsigned B = I; const ExtRoot &T = Extenders[B].getOp(); while (I != E && ExtRoot(Extenders[I].getOp()) == T) ++I; IMap.clear(); assignInits(T, B, I, IMap); Changed |= replaceExtenders(IMap); } LLVM_DEBUG({ if (Changed) MF.print(dbgs() << "After " << getPassName() << '\n', nullptr); else dbgs() << "No changes\n"; }); return Changed; } FunctionPass *llvm::createHexagonConstExtenders() { return new HexagonConstExtenders(); }