xref: /freebsd/contrib/llvm-project/llvm/lib/Support/Parallel.cpp (revision 29fc4075e69fd27de0cded313ac6000165d99f8b)
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 #if LLVM_ENABLE_THREADS
23 
24 namespace llvm {
25 namespace parallel {
26 namespace detail {
27 
28 namespace {
29 
30 /// An abstract class that takes closures and runs them asynchronously.
31 class Executor {
32 public:
33   virtual ~Executor() = default;
34   virtual void add(std::function<void()> func) = 0;
35 
36   static Executor *getDefaultExecutor();
37 };
38 
39 /// An implementation of an Executor that runs closures on a thread pool
40 ///   in filo order.
41 class ThreadPoolExecutor : public Executor {
42 public:
43   explicit ThreadPoolExecutor(ThreadPoolStrategy S = hardware_concurrency()) {
44     unsigned ThreadCount = S.compute_thread_count();
45     // Spawn all but one of the threads in another thread as spawning threads
46     // can take a while.
47     Threads.reserve(ThreadCount);
48     Threads.resize(1);
49     std::lock_guard<std::mutex> Lock(Mutex);
50     Threads[0] = std::thread([this, ThreadCount, S] {
51       for (unsigned I = 1; I < ThreadCount; ++I) {
52         Threads.emplace_back([=] { work(S, I); });
53         if (Stop)
54           break;
55       }
56       ThreadsCreated.set_value();
57       work(S, 0);
58     });
59   }
60 
61   void stop() {
62     {
63       std::lock_guard<std::mutex> Lock(Mutex);
64       if (Stop)
65         return;
66       Stop = true;
67     }
68     Cond.notify_all();
69     ThreadsCreated.get_future().wait();
70   }
71 
72   ~ThreadPoolExecutor() override {
73     stop();
74     std::thread::id CurrentThreadId = std::this_thread::get_id();
75     for (std::thread &T : Threads)
76       if (T.get_id() == CurrentThreadId)
77         T.detach();
78       else
79         T.join();
80   }
81 
82   struct Creator {
83     static void *call() { return new ThreadPoolExecutor(strategy); }
84   };
85   struct Deleter {
86     static void call(void *Ptr) { ((ThreadPoolExecutor *)Ptr)->stop(); }
87   };
88 
89   void add(std::function<void()> F) override {
90     {
91       std::lock_guard<std::mutex> Lock(Mutex);
92       WorkStack.push(F);
93     }
94     Cond.notify_one();
95   }
96 
97 private:
98   void work(ThreadPoolStrategy S, unsigned ThreadID) {
99     S.apply_thread_strategy(ThreadID);
100     while (true) {
101       std::unique_lock<std::mutex> Lock(Mutex);
102       Cond.wait(Lock, [&] { return Stop || !WorkStack.empty(); });
103       if (Stop)
104         break;
105       auto Task = WorkStack.top();
106       WorkStack.pop();
107       Lock.unlock();
108       Task();
109     }
110   }
111 
112   std::atomic<bool> Stop{false};
113   std::stack<std::function<void()>> WorkStack;
114   std::mutex Mutex;
115   std::condition_variable Cond;
116   std::promise<void> ThreadsCreated;
117   std::vector<std::thread> Threads;
118 };
119 
120 Executor *Executor::getDefaultExecutor() {
121   // The ManagedStatic enables the ThreadPoolExecutor to be stopped via
122   // llvm_shutdown() which allows a "clean" fast exit, e.g. via _exit(). This
123   // stops the thread pool and waits for any worker thread creation to complete
124   // but does not wait for the threads to finish. The wait for worker thread
125   // creation to complete is important as it prevents intermittent crashes on
126   // Windows due to a race condition between thread creation and process exit.
127   //
128   // The ThreadPoolExecutor will only be destroyed when the static unique_ptr to
129   // it is destroyed, i.e. in a normal full exit. The ThreadPoolExecutor
130   // destructor ensures it has been stopped and waits for worker threads to
131   // finish. The wait is important as it prevents intermittent crashes on
132   // Windows when the process is doing a full exit.
133   //
134   // The Windows crashes appear to only occur with the MSVC static runtimes and
135   // are more frequent with the debug static runtime.
136   //
137   // This also prevents intermittent deadlocks on exit with the MinGW runtime.
138 
139   static ManagedStatic<ThreadPoolExecutor, ThreadPoolExecutor::Creator,
140                        ThreadPoolExecutor::Deleter>
141       ManagedExec;
142   static std::unique_ptr<ThreadPoolExecutor> Exec(&(*ManagedExec));
143   return Exec.get();
144 }
145 } // namespace
146 
147 static std::atomic<int> TaskGroupInstances;
148 
149 // Latch::sync() called by the dtor may cause one thread to block. If is a dead
150 // lock if all threads in the default executor are blocked. To prevent the dead
151 // lock, only allow the first TaskGroup to run tasks parallelly. In the scenario
152 // of nested parallel_for_each(), only the outermost one runs parallelly.
153 TaskGroup::TaskGroup() : Parallel(TaskGroupInstances++ == 0) {}
154 TaskGroup::~TaskGroup() {
155   // We must ensure that all the workloads have finished before decrementing the
156   // instances count.
157   L.sync();
158   --TaskGroupInstances;
159 }
160 
161 void TaskGroup::spawn(std::function<void()> F) {
162   if (Parallel) {
163     L.inc();
164     Executor::getDefaultExecutor()->add([&, F] {
165       F();
166       L.dec();
167     });
168   } else {
169     F();
170   }
171 }
172 
173 } // namespace detail
174 } // namespace parallel
175 } // namespace llvm
176 #endif // LLVM_ENABLE_THREADS
177 
178 void llvm::parallelForEachN(size_t Begin, size_t End,
179                             llvm::function_ref<void(size_t)> Fn) {
180   // If we have zero or one items, then do not incur the overhead of spinning up
181   // a task group.  They are surprisingly expensive, and because they do not
182   // support nested parallelism, a single entry task group can block parallel
183   // execution underneath them.
184 #if LLVM_ENABLE_THREADS
185   auto NumItems = End - Begin;
186   if (NumItems > 1 && parallel::strategy.ThreadsRequested != 1) {
187     // Limit the number of tasks to MaxTasksPerGroup to limit job scheduling
188     // overhead on large inputs.
189     auto TaskSize = NumItems / parallel::detail::MaxTasksPerGroup;
190     if (TaskSize == 0)
191       TaskSize = 1;
192 
193     parallel::detail::TaskGroup TG;
194     for (; Begin + TaskSize < End; Begin += TaskSize) {
195       TG.spawn([=, &Fn] {
196         for (size_t I = Begin, E = Begin + TaskSize; I != E; ++I)
197           Fn(I);
198       });
199     }
200     for (; Begin != End; ++Begin)
201       Fn(Begin);
202     return;
203   }
204 #endif
205 
206   for (; Begin != End; ++Begin)
207     Fn(Begin);
208 }
209