1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
|
//===- bolt/Core/ParallelUtilities.cpp - Parallel utilities ---------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// Implementation of the class that manages parallel work on BinaryFunctions.
//
//===----------------------------------------------------------------------===//
#include "bolt/Core/ParallelUtilities.h"
#include "bolt/Core/BinaryContext.h"
#include "bolt/Core/BinaryFunction.h"
#include "llvm/Support/ThreadPool.h"
#include "llvm/Support/Timer.h"
#include <mutex>
#include <shared_mutex>
#define DEBUG_TYPE "par-utils"
namespace opts {
extern cl::OptionCategory BoltCategory;
cl::opt<unsigned>
ThreadCount("thread-count",
cl::desc("number of threads"),
cl::init(hardware_concurrency().compute_thread_count()),
cl::cat(BoltCategory));
cl::opt<bool>
NoThreads("no-threads",
cl::desc("disable multithreading"),
cl::init(false),
cl::cat(BoltCategory));
cl::opt<unsigned>
TaskCount("tasks-per-thread",
cl::desc("number of tasks to be created per thread"),
cl::init(20),
cl::cat(BoltCategory));
} // namespace opts
namespace llvm {
namespace bolt {
namespace ParallelUtilities {
namespace {
/// A single thread pool that is used to run parallel tasks
std::unique_ptr<ThreadPool> ThreadPoolPtr;
unsigned computeCostFor(const BinaryFunction &BF,
const PredicateTy &SkipPredicate,
const SchedulingPolicy &SchedPolicy) {
if (SchedPolicy == SchedulingPolicy::SP_TRIVIAL)
return 1;
if (SkipPredicate && SkipPredicate(BF))
return 0;
switch (SchedPolicy) {
case SchedulingPolicy::SP_CONSTANT:
return 1;
case SchedulingPolicy::SP_INST_LINEAR:
return BF.getSize();
case SchedulingPolicy::SP_INST_QUADRATIC:
return BF.getSize() * BF.getSize();
case SchedulingPolicy::SP_BB_LINEAR:
return BF.size();
case SchedulingPolicy::SP_BB_QUADRATIC:
return BF.size() * BF.size();
default:
llvm_unreachable("unsupported scheduling policy");
}
}
inline unsigned estimateTotalCost(const BinaryContext &BC,
const PredicateTy &SkipPredicate,
SchedulingPolicy &SchedPolicy) {
if (SchedPolicy == SchedulingPolicy::SP_TRIVIAL)
return BC.getBinaryFunctions().size();
unsigned TotalCost = 0;
for (auto &BFI : BC.getBinaryFunctions()) {
const BinaryFunction &BF = BFI.second;
TotalCost += computeCostFor(BF, SkipPredicate, SchedPolicy);
}
// Switch to trivial scheduling if total estimated work is zero
if (TotalCost == 0) {
outs() << "BOLT-WARNING: Running parallel work of 0 estimated cost, will "
"switch to trivial scheduling.\n";
SchedPolicy = SP_TRIVIAL;
TotalCost = BC.getBinaryFunctions().size();
}
return TotalCost;
}
} // namespace
ThreadPool &getThreadPool() {
if (ThreadPoolPtr.get())
return *ThreadPoolPtr;
ThreadPoolPtr = std::make_unique<ThreadPool>(
llvm::hardware_concurrency(opts::ThreadCount));
return *ThreadPoolPtr;
}
void runOnEachFunction(BinaryContext &BC, SchedulingPolicy SchedPolicy,
WorkFuncTy WorkFunction, PredicateTy SkipPredicate,
std::string LogName, bool ForceSequential,
unsigned TasksPerThread) {
if (BC.getBinaryFunctions().size() == 0)
return;
auto runBlock = [&](std::map<uint64_t, BinaryFunction>::iterator BlockBegin,
std::map<uint64_t, BinaryFunction>::iterator BlockEnd) {
Timer T(LogName, LogName);
LLVM_DEBUG(T.startTimer());
for (auto It = BlockBegin; It != BlockEnd; ++It) {
BinaryFunction &BF = It->second;
if (SkipPredicate && SkipPredicate(BF))
continue;
WorkFunction(BF);
}
LLVM_DEBUG(T.stopTimer());
};
if (opts::NoThreads || ForceSequential) {
runBlock(BC.getBinaryFunctions().begin(), BC.getBinaryFunctions().end());
return;
}
// Estimate the overall runtime cost using the scheduling policy
const unsigned TotalCost = estimateTotalCost(BC, SkipPredicate, SchedPolicy);
const unsigned BlocksCount = TasksPerThread * opts::ThreadCount;
const unsigned BlockCost =
TotalCost > BlocksCount ? TotalCost / BlocksCount : 1;
// Divide work into blocks of equal cost
ThreadPool &Pool = getThreadPool();
auto BlockBegin = BC.getBinaryFunctions().begin();
unsigned CurrentCost = 0;
for (auto It = BC.getBinaryFunctions().begin();
It != BC.getBinaryFunctions().end(); ++It) {
BinaryFunction &BF = It->second;
CurrentCost += computeCostFor(BF, SkipPredicate, SchedPolicy);
if (CurrentCost >= BlockCost) {
Pool.async(runBlock, BlockBegin, std::next(It));
BlockBegin = std::next(It);
CurrentCost = 0;
}
}
Pool.async(runBlock, BlockBegin, BC.getBinaryFunctions().end());
Pool.wait();
}
void runOnEachFunctionWithUniqueAllocId(
BinaryContext &BC, SchedulingPolicy SchedPolicy,
WorkFuncWithAllocTy WorkFunction, PredicateTy SkipPredicate,
std::string LogName, bool ForceSequential, unsigned TasksPerThread) {
if (BC.getBinaryFunctions().size() == 0)
return;
std::shared_timed_mutex MainLock;
auto runBlock = [&](std::map<uint64_t, BinaryFunction>::iterator BlockBegin,
std::map<uint64_t, BinaryFunction>::iterator BlockEnd,
MCPlusBuilder::AllocatorIdTy AllocId) {
Timer T(LogName, LogName);
LLVM_DEBUG(T.startTimer());
std::shared_lock<std::shared_timed_mutex> Lock(MainLock);
for (auto It = BlockBegin; It != BlockEnd; ++It) {
BinaryFunction &BF = It->second;
if (SkipPredicate && SkipPredicate(BF))
continue;
WorkFunction(BF, AllocId);
}
LLVM_DEBUG(T.stopTimer());
};
if (opts::NoThreads || ForceSequential) {
runBlock(BC.getBinaryFunctions().begin(), BC.getBinaryFunctions().end(), 0);
return;
}
// This lock is used to postpone task execution
std::unique_lock<std::shared_timed_mutex> Lock(MainLock);
// Estimate the overall runtime cost using the scheduling policy
const unsigned TotalCost = estimateTotalCost(BC, SkipPredicate, SchedPolicy);
const unsigned BlocksCount = TasksPerThread * opts::ThreadCount;
const unsigned BlockCost =
TotalCost > BlocksCount ? TotalCost / BlocksCount : 1;
// Divide work into blocks of equal cost
ThreadPool &Pool = getThreadPool();
auto BlockBegin = BC.getBinaryFunctions().begin();
unsigned CurrentCost = 0;
unsigned AllocId = 1;
for (auto It = BC.getBinaryFunctions().begin();
It != BC.getBinaryFunctions().end(); ++It) {
BinaryFunction &BF = It->second;
CurrentCost += computeCostFor(BF, SkipPredicate, SchedPolicy);
if (CurrentCost >= BlockCost) {
if (!BC.MIB->checkAllocatorExists(AllocId)) {
MCPlusBuilder::AllocatorIdTy Id =
BC.MIB->initializeNewAnnotationAllocator();
(void)Id;
assert(AllocId == Id && "unexpected allocator id created");
}
Pool.async(runBlock, BlockBegin, std::next(It), AllocId);
AllocId++;
BlockBegin = std::next(It);
CurrentCost = 0;
}
}
if (!BC.MIB->checkAllocatorExists(AllocId)) {
MCPlusBuilder::AllocatorIdTy Id =
BC.MIB->initializeNewAnnotationAllocator();
(void)Id;
assert(AllocId == Id && "unexpected allocator id created");
}
Pool.async(runBlock, BlockBegin, BC.getBinaryFunctions().end(), AllocId);
Lock.unlock();
Pool.wait();
}
} // namespace ParallelUtilities
} // namespace bolt
} // namespace llvm
|
