1 //===- SplitModule.cpp - Split a module into partitions -------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file defines the function llvm::SplitModule, which splits a module 11 // into multiple linkable partitions. It can be used to implement parallel code 12 // generation for link-time optimization. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #define DEBUG_TYPE "split-module" 17 18 #include "llvm/Transforms/Utils/SplitModule.h" 19 #include "llvm/ADT/EquivalenceClasses.h" 20 #include "llvm/ADT/Hashing.h" 21 #include "llvm/ADT/MapVector.h" 22 #include "llvm/ADT/SetVector.h" 23 #include "llvm/IR/Function.h" 24 #include "llvm/IR/GlobalAlias.h" 25 #include "llvm/IR/GlobalObject.h" 26 #include "llvm/IR/GlobalValue.h" 27 #include "llvm/IR/Module.h" 28 #include "llvm/Support/CommandLine.h" 29 #include "llvm/Support/Debug.h" 30 #include "llvm/Support/MD5.h" 31 #include "llvm/Support/raw_ostream.h" 32 #include "llvm/Transforms/Utils/Cloning.h" 33 #include <queue> 34 35 using namespace llvm; 36 37 namespace { 38 typedef EquivalenceClasses<const GlobalValue *> ClusterMapType; 39 typedef DenseMap<const Comdat *, const GlobalValue *> ComdatMembersType; 40 typedef DenseMap<const GlobalValue *, unsigned> ClusterIDMapType; 41 } 42 43 static void addNonConstUser(ClusterMapType &GVtoClusterMap, 44 const GlobalValue *GV, const User *U) { 45 assert((!isa<Constant>(U) || isa<GlobalValue>(U)) && "Bad user"); 46 47 if (const Instruction *I = dyn_cast<Instruction>(U)) { 48 const GlobalValue *F = I->getParent()->getParent(); 49 GVtoClusterMap.unionSets(GV, F); 50 } else if (isa<GlobalAlias>(U) || isa<Function>(U) || 51 isa<GlobalVariable>(U)) { 52 GVtoClusterMap.unionSets(GV, cast<GlobalValue>(U)); 53 } else { 54 llvm_unreachable("Underimplemented use case"); 55 } 56 } 57 58 // Find partitions for module in the way that no locals need to be 59 // globalized. 60 // Try to balance pack those partitions into N files since this roughly equals 61 // thread balancing for the backend codegen step. 62 static void findPartitions(Module *M, ClusterIDMapType &ClusterIDMap, 63 unsigned N) { 64 // At this point module should have the proper mix of globals and locals. 65 // As we attempt to partition this module, we must not change any 66 // locals to globals. 67 68 DEBUG(dbgs() << "Partition module with (" << M->size() << ")functions\n"); 69 ClusterMapType GVtoClusterMap; 70 ComdatMembersType ComdatMembers; 71 72 auto recordGVSet = [&GVtoClusterMap, &ComdatMembers](GlobalValue &GV) { 73 if (GV.isDeclaration()) 74 return; 75 76 if (!GV.hasName()) 77 GV.setName("__llvmsplit_unnamed"); 78 79 // Comdat groups must not be partitioned. For comdat groups that contain 80 // locals, record all their members here so we can keep them together. 81 // Comdat groups that only contain external globals are already handled by 82 // the MD5-based partitioning. 83 if (const Comdat *C = GV.getComdat()) { 84 auto &Member = ComdatMembers[C]; 85 if (Member) 86 GVtoClusterMap.unionSets(Member, &GV); 87 else 88 Member = &GV; 89 } 90 91 // For aliases we should not separate them from their aliasees regardless 92 // of linkage. 93 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(&GV)) { 94 if (const GlobalObject *Base = GA->getBaseObject()) 95 GVtoClusterMap.unionSets(&GV, Base); 96 } 97 98 // Further only iterate over local GVs. 99 if (!GV.hasLocalLinkage()) 100 return; 101 102 for (auto *U : GV.users()) { 103 SmallVector<const User *, 4> Worklist; 104 Worklist.push_back(U); 105 while (!Worklist.empty()) { 106 const User *UU = Worklist.pop_back_val(); 107 // For each constant that is not a GV (a pure const) recurse. 108 if (isa<Constant>(UU) && !isa<GlobalValue>(UU)) { 109 Worklist.append(UU->user_begin(), UU->user_end()); 110 continue; 111 } 112 addNonConstUser(GVtoClusterMap, &GV, UU); 113 } 114 } 115 }; 116 117 std::for_each(M->begin(), M->end(), recordGVSet); 118 std::for_each(M->global_begin(), M->global_end(), recordGVSet); 119 std::for_each(M->alias_begin(), M->alias_end(), recordGVSet); 120 121 // Assigned all GVs to merged clusters while balancing number of objects in 122 // each. 123 auto CompareClusters = [](const std::pair<unsigned, unsigned> &a, 124 const std::pair<unsigned, unsigned> &b) { 125 if (a.second || b.second) 126 return a.second > b.second; 127 else 128 return a.first > b.first; 129 }; 130 131 std::priority_queue<std::pair<unsigned, unsigned>, 132 std::vector<std::pair<unsigned, unsigned>>, 133 decltype(CompareClusters)> 134 BalancinQueue(CompareClusters); 135 // Pre-populate priority queue with N slot blanks. 136 for (unsigned i = 0; i < N; ++i) 137 BalancinQueue.push(std::make_pair(i, 0)); 138 139 typedef std::pair<unsigned, ClusterMapType::iterator> SortType; 140 SmallVector<SortType, 64> Sets; 141 SmallPtrSet<const GlobalValue *, 64> Visited; 142 143 // To guarantee determinism, we have to sort SCC according to size. 144 // When size is the same, use leader's name. 145 for (ClusterMapType::iterator I = GVtoClusterMap.begin(), 146 E = GVtoClusterMap.end(); I != E; ++I) 147 if (I->isLeader()) 148 Sets.push_back( 149 std::make_pair(std::distance(GVtoClusterMap.member_begin(I), 150 GVtoClusterMap.member_end()), I)); 151 152 std::sort(Sets.begin(), Sets.end(), [](const SortType &a, const SortType &b) { 153 if (a.first == b.first) 154 return a.second->getData()->getName() > b.second->getData()->getName(); 155 else 156 return a.first > b.first; 157 }); 158 159 for (auto &I : Sets) { 160 unsigned CurrentClusterID = BalancinQueue.top().first; 161 unsigned CurrentClusterSize = BalancinQueue.top().second; 162 BalancinQueue.pop(); 163 164 DEBUG(dbgs() << "Root[" << CurrentClusterID << "] cluster_size(" << I.first 165 << ") ----> " << I.second->getData()->getName() << "\n"); 166 167 for (ClusterMapType::member_iterator MI = 168 GVtoClusterMap.findLeader(I.second); 169 MI != GVtoClusterMap.member_end(); ++MI) { 170 if (!Visited.insert(*MI).second) 171 continue; 172 DEBUG(dbgs() << "----> " << (*MI)->getName() 173 << ((*MI)->hasLocalLinkage() ? " l " : " e ") << "\n"); 174 Visited.insert(*MI); 175 ClusterIDMap[*MI] = CurrentClusterID; 176 CurrentClusterSize++; 177 } 178 // Add this set size to the number of entries in this cluster. 179 BalancinQueue.push(std::make_pair(CurrentClusterID, CurrentClusterSize)); 180 } 181 } 182 183 static void externalize(GlobalValue *GV) { 184 if (GV->hasLocalLinkage()) { 185 GV->setLinkage(GlobalValue::ExternalLinkage); 186 GV->setVisibility(GlobalValue::HiddenVisibility); 187 } 188 189 // Unnamed entities must be named consistently between modules. setName will 190 // give a distinct name to each such entity. 191 if (!GV->hasName()) 192 GV->setName("__llvmsplit_unnamed"); 193 } 194 195 // Returns whether GV should be in partition (0-based) I of N. 196 static bool isInPartition(const GlobalValue *GV, unsigned I, unsigned N) { 197 if (auto GA = dyn_cast<GlobalAlias>(GV)) 198 if (const GlobalObject *Base = GA->getBaseObject()) 199 GV = Base; 200 201 StringRef Name; 202 if (const Comdat *C = GV->getComdat()) 203 Name = C->getName(); 204 else 205 Name = GV->getName(); 206 207 // Partition by MD5 hash. We only need a few bits for evenness as the number 208 // of partitions will generally be in the 1-2 figure range; the low 16 bits 209 // are enough. 210 MD5 H; 211 MD5::MD5Result R; 212 H.update(Name); 213 H.final(R); 214 return (R[0] | (R[1] << 8)) % N == I; 215 } 216 217 void llvm::SplitModule( 218 std::unique_ptr<Module> M, unsigned N, 219 std::function<void(std::unique_ptr<Module> MPart)> ModuleCallback, 220 bool PreserveLocals) { 221 if (!PreserveLocals) { 222 for (Function &F : *M) 223 externalize(&F); 224 for (GlobalVariable &GV : M->globals()) 225 externalize(&GV); 226 for (GlobalAlias &GA : M->aliases()) 227 externalize(&GA); 228 229 // FIXME: We should be able to reuse M as the last partition instead of 230 // cloning it. 231 for (unsigned I = 0; I != N; ++I) { 232 ValueToValueMapTy VMap; 233 std::unique_ptr<Module> MPart( 234 CloneModule(M.get(), VMap, [=](const GlobalValue *GV) { 235 return isInPartition(GV, I, N); 236 })); 237 if (I != 0) 238 MPart->setModuleInlineAsm(""); 239 ModuleCallback(std::move(MPart)); 240 } 241 } else { 242 // This performs splitting without a need for externalization, which might not 243 // always be possible. 244 ClusterIDMapType ClusterIDMap; 245 findPartitions(M.get(), ClusterIDMap, N); 246 247 for (unsigned I = 0; I < N; ++I) { 248 ValueToValueMapTy VMap; 249 std::unique_ptr<Module> MPart( 250 CloneModule(M.get(), VMap, [&](const GlobalValue *GV) { 251 if (ClusterIDMap.count(GV)) 252 return (ClusterIDMap[GV] == I); 253 else 254 return isInPartition(GV, I, N); 255 })); 256 if (I != 0) 257 MPart->setModuleInlineAsm(""); 258 ModuleCallback(std::move(MPart)); 259 } 260 } 261 } 262