1//===- SplitModule.cpp - Split a module into partitions -------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file defines the function llvm::SplitModule, which splits a module
10// into multiple linkable partitions. It can be used to implement parallel code
11// generation for link-time optimization.
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/Transforms/Utils/SplitModule.h"
16#include "llvm/ADT/DenseMap.h"
17#include "llvm/ADT/EquivalenceClasses.h"
18#include "llvm/ADT/SmallPtrSet.h"
19#include "llvm/ADT/SmallVector.h"
20#include "llvm/ADT/StringRef.h"
21#include "llvm/IR/Comdat.h"
22#include "llvm/IR/Constant.h"
23#include "llvm/IR/Constants.h"
24#include "llvm/IR/Function.h"
25#include "llvm/IR/GlobalAlias.h"
26#include "llvm/IR/GlobalObject.h"
27#include "llvm/IR/GlobalValue.h"
28#include "llvm/IR/GlobalVariable.h"
29#include "llvm/IR/Instruction.h"
30#include "llvm/IR/Module.h"
31#include "llvm/IR/User.h"
32#include "llvm/IR/Value.h"
33#include "llvm/Support/Casting.h"
34#include "llvm/Support/Debug.h"
35#include "llvm/Support/ErrorHandling.h"
36#include "llvm/Support/MD5.h"
37#include "llvm/Support/raw_ostream.h"
38#include "llvm/Transforms/Utils/Cloning.h"
39#include "llvm/Transforms/Utils/ValueMapper.h"
40#include <algorithm>
41#include <cassert>
42#include <iterator>
43#include <memory>
44#include <queue>
45#include <utility>
46#include <vector>
47
48using namespace llvm;
49
50#define DEBUG_TYPE "split-module"
51
52namespace {
53
54using ClusterMapType = EquivalenceClasses<const GlobalValue *>;
55using ComdatMembersType = DenseMap<const Comdat *, const GlobalValue *>;
56using ClusterIDMapType = DenseMap<const GlobalValue *, unsigned>;
57
58} // end anonymous namespace
59
60static void addNonConstUser(ClusterMapType &GVtoClusterMap,
61 const GlobalValue *GV, const User *U) {
62 assert((!isa<Constant>(U) || isa<GlobalValue>(U)) && "Bad user");
63
64 if (const Instruction *I = dyn_cast<Instruction>(Val: U)) {
65 const GlobalValue *F = I->getParent()->getParent();
66 GVtoClusterMap.unionSets(V1: GV, V2: F);
67 } else if (const GlobalValue *GVU = dyn_cast<GlobalValue>(Val: U)) {
68 GVtoClusterMap.unionSets(V1: GV, V2: GVU);
69 } else {
70 llvm_unreachable("Underimplemented use case");
71 }
72}
73
74// Adds all GlobalValue users of V to the same cluster as GV.
75static void addAllGlobalValueUsers(ClusterMapType &GVtoClusterMap,
76 const GlobalValue *GV, const Value *V) {
77 for (const auto *U : V->users()) {
78 SmallVector<const User *, 4> Worklist;
79 Worklist.push_back(Elt: U);
80 while (!Worklist.empty()) {
81 const User *UU = Worklist.pop_back_val();
82 // For each constant that is not a GV (a pure const) recurse.
83 if (isa<Constant>(Val: UU) && !isa<GlobalValue>(Val: UU)) {
84 Worklist.append(in_start: UU->user_begin(), in_end: UU->user_end());
85 continue;
86 }
87 addNonConstUser(GVtoClusterMap, GV, U: UU);
88 }
89 }
90}
91
92static const GlobalObject *getGVPartitioningRoot(const GlobalValue *GV) {
93 const GlobalObject *GO = GV->getAliaseeObject();
94 if (const auto *GI = dyn_cast_or_null<GlobalIFunc>(Val: GO))
95 GO = GI->getResolverFunction();
96 return GO;
97}
98
99// Find partitions for module in the way that no locals need to be
100// globalized.
101// Try to balance pack those partitions into N files since this roughly equals
102// thread balancing for the backend codegen step.
103static void findPartitions(Module &M, ClusterIDMapType &ClusterIDMap,
104 unsigned N) {
105 // At this point module should have the proper mix of globals and locals.
106 // As we attempt to partition this module, we must not change any
107 // locals to globals.
108 LLVM_DEBUG(dbgs() << "Partition module with (" << M.size() << ")functions\n");
109 ClusterMapType GVtoClusterMap;
110 ComdatMembersType ComdatMembers;
111
112 auto recordGVSet = [&GVtoClusterMap, &ComdatMembers](GlobalValue &GV) {
113 if (GV.isDeclaration())
114 return;
115
116 if (!GV.hasName())
117 GV.setName("__llvmsplit_unnamed");
118
119 // Comdat groups must not be partitioned. For comdat groups that contain
120 // locals, record all their members here so we can keep them together.
121 // Comdat groups that only contain external globals are already handled by
122 // the MD5-based partitioning.
123 if (const Comdat *C = GV.getComdat()) {
124 auto &Member = ComdatMembers[C];
125 if (Member)
126 GVtoClusterMap.unionSets(V1: Member, V2: &GV);
127 else
128 Member = &GV;
129 }
130
131 // Aliases should not be separated from their aliasees and ifuncs should
132 // not be separated from their resolvers regardless of linkage.
133 if (const GlobalObject *Root = getGVPartitioningRoot(GV: &GV))
134 if (&GV != Root)
135 GVtoClusterMap.unionSets(V1: &GV, V2: Root);
136
137 if (const Function *F = dyn_cast<Function>(Val: &GV)) {
138 for (const BasicBlock &BB : *F) {
139 BlockAddress *BA = BlockAddress::lookup(BB: &BB);
140 if (!BA || !BA->isConstantUsed())
141 continue;
142 addAllGlobalValueUsers(GVtoClusterMap, GV: F, V: BA);
143 }
144 }
145
146 if (GV.hasLocalLinkage())
147 addAllGlobalValueUsers(GVtoClusterMap, GV: &GV, V: &GV);
148 };
149
150 llvm::for_each(Range: M.functions(), F: recordGVSet);
151 llvm::for_each(Range: M.globals(), F: recordGVSet);
152 llvm::for_each(Range: M.aliases(), F: recordGVSet);
153
154 // Assigned all GVs to merged clusters while balancing number of objects in
155 // each.
156 auto CompareClusters = [](const std::pair<unsigned, unsigned> &a,
157 const std::pair<unsigned, unsigned> &b) {
158 if (a.second || b.second)
159 return a.second > b.second;
160 else
161 return a.first > b.first;
162 };
163
164 std::priority_queue<std::pair<unsigned, unsigned>,
165 std::vector<std::pair<unsigned, unsigned>>,
166 decltype(CompareClusters)>
167 BalancinQueue(CompareClusters);
168 // Pre-populate priority queue with N slot blanks.
169 for (unsigned i = 0; i < N; ++i)
170 BalancinQueue.push(x: std::make_pair(x&: i, y: 0));
171
172 using SortType = std::pair<unsigned, ClusterMapType::iterator>;
173
174 SmallVector<SortType, 64> Sets;
175 SmallPtrSet<const GlobalValue *, 32> Visited;
176
177 // To guarantee determinism, we have to sort SCC according to size.
178 // When size is the same, use leader's name.
179 for (ClusterMapType::iterator I = GVtoClusterMap.begin(),
180 E = GVtoClusterMap.end(); I != E; ++I)
181 if (I->isLeader())
182 Sets.push_back(
183 Elt: std::make_pair(x: std::distance(first: GVtoClusterMap.member_begin(I),
184 last: GVtoClusterMap.member_end()), y&: I));
185
186 llvm::sort(C&: Sets, Comp: [](const SortType &a, const SortType &b) {
187 if (a.first == b.first)
188 return a.second->getData()->getName() > b.second->getData()->getName();
189 else
190 return a.first > b.first;
191 });
192
193 for (auto &I : Sets) {
194 unsigned CurrentClusterID = BalancinQueue.top().first;
195 unsigned CurrentClusterSize = BalancinQueue.top().second;
196 BalancinQueue.pop();
197
198 LLVM_DEBUG(dbgs() << "Root[" << CurrentClusterID << "] cluster_size("
199 << I.first << ") ----> " << I.second->getData()->getName()
200 << "\n");
201
202 for (ClusterMapType::member_iterator MI =
203 GVtoClusterMap.findLeader(I: I.second);
204 MI != GVtoClusterMap.member_end(); ++MI) {
205 if (!Visited.insert(Ptr: *MI).second)
206 continue;
207 LLVM_DEBUG(dbgs() << "----> " << (*MI)->getName()
208 << ((*MI)->hasLocalLinkage() ? " l " : " e ") << "\n");
209 Visited.insert(Ptr: *MI);
210 ClusterIDMap[*MI] = CurrentClusterID;
211 CurrentClusterSize++;
212 }
213 // Add this set size to the number of entries in this cluster.
214 BalancinQueue.push(x: std::make_pair(x&: CurrentClusterID, y&: CurrentClusterSize));
215 }
216}
217
218static void externalize(GlobalValue *GV) {
219 if (GV->hasLocalLinkage()) {
220 GV->setLinkage(GlobalValue::ExternalLinkage);
221 GV->setVisibility(GlobalValue::HiddenVisibility);
222 }
223
224 // Unnamed entities must be named consistently between modules. setName will
225 // give a distinct name to each such entity.
226 if (!GV->hasName())
227 GV->setName("__llvmsplit_unnamed");
228}
229
230// Returns whether GV should be in partition (0-based) I of N.
231static bool isInPartition(const GlobalValue *GV, unsigned I, unsigned N) {
232 if (const GlobalObject *Root = getGVPartitioningRoot(GV))
233 GV = Root;
234
235 StringRef Name;
236 if (const Comdat *C = GV->getComdat())
237 Name = C->getName();
238 else
239 Name = GV->getName();
240
241 // Partition by MD5 hash. We only need a few bits for evenness as the number
242 // of partitions will generally be in the 1-2 figure range; the low 16 bits
243 // are enough.
244 MD5 H;
245 MD5::MD5Result R;
246 H.update(Str: Name);
247 H.final(Result&: R);
248 return (R[0] | (R[1] << 8)) % N == I;
249}
250
251void llvm::SplitModule(
252 Module &M, unsigned N,
253 function_ref<void(std::unique_ptr<Module> MPart)> ModuleCallback,
254 bool PreserveLocals) {
255 if (!PreserveLocals) {
256 for (Function &F : M)
257 externalize(GV: &F);
258 for (GlobalVariable &GV : M.globals())
259 externalize(GV: &GV);
260 for (GlobalAlias &GA : M.aliases())
261 externalize(GV: &GA);
262 for (GlobalIFunc &GIF : M.ifuncs())
263 externalize(GV: &GIF);
264 }
265
266 // This performs splitting without a need for externalization, which might not
267 // always be possible.
268 ClusterIDMapType ClusterIDMap;
269 findPartitions(M, ClusterIDMap, N);
270
271 // FIXME: We should be able to reuse M as the last partition instead of
272 // cloning it. Note that the callers at the moment expect the module to
273 // be preserved, so will need some adjustments as well.
274 for (unsigned I = 0; I < N; ++I) {
275 ValueToValueMapTy VMap;
276 std::unique_ptr<Module> MPart(
277 CloneModule(M, VMap, ShouldCloneDefinition: [&](const GlobalValue *GV) {
278 if (ClusterIDMap.count(Val: GV))
279 return (ClusterIDMap[GV] == I);
280 else
281 return isInPartition(GV, I, N);
282 }));
283 if (I != 0)
284 MPart->setModuleInlineAsm("");
285 ModuleCallback(std::move(MPart));
286 }
287}
288

source code of llvm/lib/Transforms/Utils/SplitModule.cpp