xref: /freebsd/contrib/llvm-project/llvm/lib/Support/Parallel.cpp (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57)
1 //===- llvm/Support/Parallel.cpp - Parallel algorithms --------------------===//
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 #include "llvm/Support/Parallel.h"
10 #include "llvm/Config/llvm-config.h"
11 #include "llvm/Support/ManagedStatic.h"
12 #include "llvm/Support/Threading.h"
13 
14 #include <atomic>
15 #include <future>
16 #include <stack>
17 #include <thread>
18 #include <vector>
19 
20 llvm::ThreadPoolStrategy llvm::parallel::strategy;
21 
22 namespace llvm {
23 namespace parallel {
24 #if LLVM_ENABLE_THREADS
25 
26 #ifdef _WIN32
27 static thread_local unsigned threadIndex;
28 
29 unsigned getThreadIndex() { return threadIndex; }
30 #else
31 thread_local unsigned threadIndex;
32 #endif
33 
34 namespace detail {
35 
36 namespace {
37 
38 /// An abstract class that takes closures and runs them asynchronously.
39 class Executor {
40 public:
41   virtual ~Executor() = default;
42   virtual void add(std::function<void()> func) = 0;
43 
44   static Executor *getDefaultExecutor();
45 };
46 
47 /// An implementation of an Executor that runs closures on a thread pool
48 ///   in filo order.
49 class ThreadPoolExecutor : public Executor {
50 public:
51   explicit ThreadPoolExecutor(ThreadPoolStrategy S = hardware_concurrency()) {
52     unsigned ThreadCount = S.compute_thread_count();
53     // Spawn all but one of the threads in another thread as spawning threads
54     // can take a while.
55     Threads.reserve(ThreadCount);
56     Threads.resize(1);
57     std::lock_guard<std::mutex> Lock(Mutex);
58     Threads[0] = std::thread([this, ThreadCount, S] {
59       for (unsigned I = 1; I < ThreadCount; ++I) {
60         Threads.emplace_back([=] { work(S, I); });
61         if (Stop)
62           break;
63       }
64       ThreadsCreated.set_value();
65       work(S, 0);
66     });
67   }
68 
69   void stop() {
70     {
71       std::lock_guard<std::mutex> Lock(Mutex);
72       if (Stop)
73         return;
74       Stop = true;
75     }
76     Cond.notify_all();
77     ThreadsCreated.get_future().wait();
78   }
79 
80   ~ThreadPoolExecutor() override {
81     stop();
82     std::thread::id CurrentThreadId = std::this_thread::get_id();
83     for (std::thread &T : Threads)
84       if (T.get_id() == CurrentThreadId)
85         T.detach();
86       else
87         T.join();
88   }
89 
90   struct Creator {
91     static void *call() { return new ThreadPoolExecutor(strategy); }
92   };
93   struct Deleter {
94     static void call(void *Ptr) { ((ThreadPoolExecutor *)Ptr)->stop(); }
95   };
96 
97   void add(std::function<void()> F) override {
98     {
99       std::lock_guard<std::mutex> Lock(Mutex);
100       WorkStack.push(std::move(F));
101     }
102     Cond.notify_one();
103   }
104 
105 private:
106   void work(ThreadPoolStrategy S, unsigned ThreadID) {
107     threadIndex = ThreadID;
108     S.apply_thread_strategy(ThreadID);
109     while (true) {
110       std::unique_lock<std::mutex> Lock(Mutex);
111       Cond.wait(Lock, [&] { return Stop || !WorkStack.empty(); });
112       if (Stop)
113         break;
114       auto Task = std::move(WorkStack.top());
115       WorkStack.pop();
116       Lock.unlock();
117       Task();
118     }
119   }
120 
121   std::atomic<bool> Stop{false};
122   std::stack<std::function<void()>> WorkStack;
123   std::mutex Mutex;
124   std::condition_variable Cond;
125   std::promise<void> ThreadsCreated;
126   std::vector<std::thread> Threads;
127 };
128 
129 Executor *Executor::getDefaultExecutor() {
130   // The ManagedStatic enables the ThreadPoolExecutor to be stopped via
131   // llvm_shutdown() which allows a "clean" fast exit, e.g. via _exit(). This
132   // stops the thread pool and waits for any worker thread creation to complete
133   // but does not wait for the threads to finish. The wait for worker thread
134   // creation to complete is important as it prevents intermittent crashes on
135   // Windows due to a race condition between thread creation and process exit.
136   //
137   // The ThreadPoolExecutor will only be destroyed when the static unique_ptr to
138   // it is destroyed, i.e. in a normal full exit. The ThreadPoolExecutor
139   // destructor ensures it has been stopped and waits for worker threads to
140   // finish. The wait is important as it prevents intermittent crashes on
141   // Windows when the process is doing a full exit.
142   //
143   // The Windows crashes appear to only occur with the MSVC static runtimes and
144   // are more frequent with the debug static runtime.
145   //
146   // This also prevents intermittent deadlocks on exit with the MinGW runtime.
147 
148   static ManagedStatic<ThreadPoolExecutor, ThreadPoolExecutor::Creator,
149                        ThreadPoolExecutor::Deleter>
150       ManagedExec;
151   static std::unique_ptr<ThreadPoolExecutor> Exec(&(*ManagedExec));
152   return Exec.get();
153 }
154 } // namespace
155 } // namespace detail
156 #endif
157 
158 static std::atomic<int> TaskGroupInstances;
159 
160 // Latch::sync() called by the dtor may cause one thread to block. If is a dead
161 // lock if all threads in the default executor are blocked. To prevent the dead
162 // lock, only allow the first TaskGroup to run tasks parallelly. In the scenario
163 // of nested parallel_for_each(), only the outermost one runs parallelly.
164 TaskGroup::TaskGroup() : Parallel(TaskGroupInstances++ == 0) {}
165 TaskGroup::~TaskGroup() {
166   // We must ensure that all the workloads have finished before decrementing the
167   // instances count.
168   L.sync();
169   --TaskGroupInstances;
170 }
171 
172 void TaskGroup::spawn(std::function<void()> F) {
173 #if LLVM_ENABLE_THREADS
174   if (Parallel) {
175     L.inc();
176     detail::Executor::getDefaultExecutor()->add([&, F = std::move(F)] {
177       F();
178       L.dec();
179     });
180     return;
181   }
182 #endif
183   F();
184 }
185 
186 void TaskGroup::execute(std::function<void()> F) {
187   if (parallel::strategy.ThreadsRequested == 1)
188     F();
189   else
190     spawn(F);
191 }
192 } // namespace parallel
193 } // namespace llvm
194 
195 void llvm::parallelFor(size_t Begin, size_t End,
196                        llvm::function_ref<void(size_t)> Fn) {
197   // If we have zero or one items, then do not incur the overhead of spinning up
198   // a task group.  They are surprisingly expensive, and because they do not
199   // support nested parallelism, a single entry task group can block parallel
200   // execution underneath them.
201 #if LLVM_ENABLE_THREADS
202   auto NumItems = End - Begin;
203   if (NumItems > 1 && parallel::strategy.ThreadsRequested != 1) {
204     // Limit the number of tasks to MaxTasksPerGroup to limit job scheduling
205     // overhead on large inputs.
206     auto TaskSize = NumItems / parallel::detail::MaxTasksPerGroup;
207     if (TaskSize == 0)
208       TaskSize = 1;
209 
210     parallel::TaskGroup TG;
211     for (; Begin + TaskSize < End; Begin += TaskSize) {
212       TG.spawn([=, &Fn] {
213         for (size_t I = Begin, E = Begin + TaskSize; I != E; ++I)
214           Fn(I);
215       });
216     }
217     if (Begin != End) {
218       TG.spawn([=, &Fn] {
219         for (size_t I = Begin; I != End; ++I)
220           Fn(I);
221       });
222     }
223     return;
224   }
225 #endif
226 
227   for (; Begin != End; ++Begin)
228     Fn(Begin);
229 }
230