Krzysztof Parzyszek | 6f4000e | 2016-01-12 17:01:16 +0000 | [diff] [blame^] | 1 | //===--- RDFDeadCode.cpp --------------------------------------------------===// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | // |
| 10 | // RDF-based generic dead code elimination. |
| 11 | |
| 12 | #include "RDFGraph.h" |
| 13 | #include "RDFLiveness.h" |
| 14 | #include "RDFDeadCode.h" |
| 15 | |
| 16 | #include "llvm/ADT/SetVector.h" |
| 17 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 18 | #include "llvm/CodeGen/MachineFunction.h" |
| 19 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 20 | |
| 21 | using namespace llvm; |
| 22 | using namespace rdf; |
| 23 | |
| 24 | // Check if the given instruction has observable side-effects, i.e. if |
| 25 | // it should be considered "live". It is safe for this function to be |
| 26 | // overly conservative (i.e. return "true" for all instructions), but it |
| 27 | // is not safe to return "false" for an instruction that should not be |
| 28 | // considered removable. |
| 29 | bool DeadCodeElimination::isLiveInstr(const MachineInstr *MI) const { |
| 30 | if (MI->mayStore() || MI->isBranch() || MI->isCall() || MI->isReturn()) |
| 31 | return true; |
| 32 | if (MI->hasOrderedMemoryRef() || MI->hasUnmodeledSideEffects()) |
| 33 | return true; |
| 34 | if (MI->isPHI()) |
| 35 | return false; |
| 36 | for (auto &Op : MI->operands()) |
| 37 | if (Op.isReg() && MRI.isReserved(Op.getReg())) |
| 38 | return true; |
| 39 | return false; |
| 40 | } |
| 41 | |
| 42 | void DeadCodeElimination::scanInstr(NodeAddr<InstrNode*> IA, |
| 43 | SetVector<NodeId> &WorkQ) { |
| 44 | if (!DFG.IsCode<NodeAttrs::Stmt>(IA)) |
| 45 | return; |
| 46 | if (!isLiveInstr(NodeAddr<StmtNode*>(IA).Addr->getCode())) |
| 47 | return; |
| 48 | for (NodeAddr<RefNode*> RA : IA.Addr->members(DFG)) { |
| 49 | if (!LiveNodes.count(RA.Id)) |
| 50 | WorkQ.insert(RA.Id); |
| 51 | } |
| 52 | } |
| 53 | |
| 54 | void DeadCodeElimination::processDef(NodeAddr<DefNode*> DA, |
| 55 | SetVector<NodeId> &WorkQ) { |
| 56 | NodeAddr<InstrNode*> IA = DA.Addr->getOwner(DFG); |
| 57 | for (NodeAddr<UseNode*> UA : IA.Addr->members_if(DFG.IsUse, DFG)) { |
| 58 | if (!LiveNodes.count(UA.Id)) |
| 59 | WorkQ.insert(UA.Id); |
| 60 | } |
| 61 | for (NodeAddr<DefNode*> TA : DFG.getRelatedRefs(IA, DA)) |
| 62 | LiveNodes.insert(TA.Id); |
| 63 | } |
| 64 | |
| 65 | void DeadCodeElimination::processUse(NodeAddr<UseNode*> UA, |
| 66 | SetVector<NodeId> &WorkQ) { |
| 67 | for (NodeAddr<DefNode*> DA : LV.getAllReachingDefs(UA)) { |
| 68 | if (!LiveNodes.count(DA.Id)) |
| 69 | WorkQ.insert(DA.Id); |
| 70 | } |
| 71 | } |
| 72 | |
| 73 | // Traverse the DFG and collect the set dead RefNodes and the set of |
| 74 | // dead instructions. Return "true" if any of these sets is non-empty, |
| 75 | // "false" otherwise. |
| 76 | bool DeadCodeElimination::collect() { |
| 77 | // This function works by first finding all live nodes. The dead nodes |
| 78 | // are then the complement of the set of live nodes. |
| 79 | // |
| 80 | // Assume that all nodes are dead. Identify instructions which must be |
| 81 | // considered live, i.e. instructions with observable side-effects, such |
| 82 | // as calls and stores. All arguments of such instructions are considered |
| 83 | // live. For each live def, all operands used in the corresponding |
| 84 | // instruction are considered live. For each live use, all its reaching |
| 85 | // defs are considered live. |
| 86 | LiveNodes.clear(); |
| 87 | SetVector<NodeId> WorkQ; |
| 88 | for (NodeAddr<BlockNode*> BA : DFG.getFunc().Addr->members(DFG)) |
| 89 | for (NodeAddr<InstrNode*> IA : BA.Addr->members(DFG)) |
| 90 | scanInstr(IA, WorkQ); |
| 91 | |
| 92 | while (!WorkQ.empty()) { |
| 93 | NodeId N = *WorkQ.begin(); |
| 94 | WorkQ.remove(N); |
| 95 | LiveNodes.insert(N); |
| 96 | auto RA = DFG.addr<RefNode*>(N); |
| 97 | if (DFG.IsDef(RA)) |
| 98 | processDef(RA, WorkQ); |
| 99 | else |
| 100 | processUse(RA, WorkQ); |
| 101 | } |
| 102 | |
| 103 | if (trace()) { |
| 104 | dbgs() << "Live nodes:\n"; |
| 105 | for (NodeId N : LiveNodes) { |
| 106 | auto RA = DFG.addr<RefNode*>(N); |
| 107 | dbgs() << PrintNode<RefNode*>(RA, DFG) << "\n"; |
| 108 | } |
| 109 | } |
| 110 | |
| 111 | auto IsDead = [this] (NodeAddr<InstrNode*> IA) -> bool { |
| 112 | for (NodeAddr<DefNode*> DA : IA.Addr->members_if(DFG.IsDef, DFG)) |
| 113 | if (LiveNodes.count(DA.Id)) |
| 114 | return false; |
| 115 | return true; |
| 116 | }; |
| 117 | |
| 118 | for (NodeAddr<BlockNode*> BA : DFG.getFunc().Addr->members(DFG)) { |
| 119 | for (NodeAddr<InstrNode*> IA : BA.Addr->members(DFG)) { |
| 120 | for (NodeAddr<RefNode*> RA : IA.Addr->members(DFG)) |
| 121 | if (!LiveNodes.count(RA.Id)) |
| 122 | DeadNodes.insert(RA.Id); |
| 123 | if (DFG.IsCode<NodeAttrs::Stmt>(IA)) |
| 124 | if (isLiveInstr(NodeAddr<StmtNode*>(IA).Addr->getCode())) |
| 125 | continue; |
| 126 | if (IsDead(IA)) { |
| 127 | DeadInstrs.insert(IA.Id); |
| 128 | if (trace()) |
| 129 | dbgs() << "Dead instr: " << PrintNode<InstrNode*>(IA, DFG) << "\n"; |
| 130 | } |
| 131 | } |
| 132 | } |
| 133 | |
| 134 | return !DeadNodes.empty(); |
| 135 | } |
| 136 | |
| 137 | // Erase the nodes given in the Nodes set from DFG. In addition to removing |
| 138 | // them from the DFG, if a node corresponds to a statement, the corresponding |
| 139 | // machine instruction is erased from the function. |
| 140 | bool DeadCodeElimination::erase(const SetVector<NodeId> &Nodes) { |
| 141 | if (Nodes.empty()) |
| 142 | return false; |
| 143 | |
| 144 | // Prepare the actual set of ref nodes to remove: ref nodes from Nodes |
| 145 | // are included directly, for each InstrNode in Nodes, include the set |
| 146 | // of all RefNodes from it. |
| 147 | NodeList DRNs, DINs; |
| 148 | for (auto I : Nodes) { |
| 149 | auto BA = DFG.addr<NodeBase*>(I); |
| 150 | uint16_t Type = BA.Addr->getType(); |
| 151 | if (Type == NodeAttrs::Ref) { |
| 152 | DRNs.push_back(DFG.addr<RefNode*>(I)); |
| 153 | continue; |
| 154 | } |
| 155 | |
| 156 | // If it's a code node, add all ref nodes from it. |
| 157 | uint16_t Kind = BA.Addr->getKind(); |
| 158 | if (Kind == NodeAttrs::Stmt || Kind == NodeAttrs::Phi) { |
| 159 | for (auto N : NodeAddr<CodeNode*>(BA).Addr->members(DFG)) |
| 160 | DRNs.push_back(N); |
| 161 | DINs.push_back(DFG.addr<InstrNode*>(I)); |
| 162 | } else { |
| 163 | llvm_unreachable("Unexpected code node"); |
| 164 | return false; |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | // Sort the list so that use nodes are removed first. This makes the |
| 169 | // "unlink" functions a bit faster. |
| 170 | auto UsesFirst = [] (NodeAddr<RefNode*> A, NodeAddr<RefNode*> B) -> bool { |
| 171 | uint16_t KindA = A.Addr->getKind(), KindB = B.Addr->getKind(); |
| 172 | if (KindA == NodeAttrs::Use && KindB == NodeAttrs::Def) |
| 173 | return true; |
| 174 | if (KindA == NodeAttrs::Def && KindB == NodeAttrs::Use) |
| 175 | return false; |
| 176 | return A.Id < B.Id; |
| 177 | }; |
| 178 | std::sort(DRNs.begin(), DRNs.end(), UsesFirst); |
| 179 | |
| 180 | if (trace()) |
| 181 | dbgs() << "Removing dead ref nodes:\n"; |
| 182 | for (NodeAddr<RefNode*> RA : DRNs) { |
| 183 | if (trace()) |
| 184 | dbgs() << " " << PrintNode<RefNode*>(RA, DFG) << '\n'; |
| 185 | if (DFG.IsUse(RA)) |
| 186 | DFG.unlinkUse(RA); |
| 187 | else if (DFG.IsDef(RA)) |
| 188 | DFG.unlinkDef(RA); |
| 189 | } |
| 190 | |
| 191 | // Now, remove all dead instruction nodes. |
| 192 | for (NodeAddr<InstrNode*> IA : DINs) { |
| 193 | NodeAddr<BlockNode*> BA = IA.Addr->getOwner(DFG); |
| 194 | BA.Addr->removeMember(IA, DFG); |
| 195 | if (!DFG.IsCode<NodeAttrs::Stmt>(IA)) |
| 196 | continue; |
| 197 | |
| 198 | MachineInstr *MI = NodeAddr<StmtNode*>(IA).Addr->getCode(); |
| 199 | if (trace()) |
| 200 | dbgs() << "erasing: " << *MI; |
| 201 | MI->eraseFromParent(); |
| 202 | } |
| 203 | return true; |
| 204 | } |