LLVM  13.0.0git
Parallel.h
Go to the documentation of this file.
1 //===- llvm/Support/Parallel.h - 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 #ifndef LLVM_SUPPORT_PARALLEL_H
10 #define LLVM_SUPPORT_PARALLEL_H
11 
12 #include "llvm/ADT/STLExtras.h"
13 #include "llvm/Config/llvm-config.h"
14 #include "llvm/Support/Error.h"
16 #include "llvm/Support/Threading.h"
17 
18 #include <algorithm>
19 #include <condition_variable>
20 #include <functional>
21 #include <mutex>
22 
23 namespace llvm {
24 
25 namespace parallel {
26 
27 // Strategy for the default executor used by the parallel routines provided by
28 // this file. It defaults to using all hardware threads and should be
29 // initialized before the first use of parallel routines.
31 
32 namespace detail {
33 
34 #if LLVM_ENABLE_THREADS
35 
36 class Latch {
37  uint32_t Count;
38  mutable std::mutex Mutex;
39  mutable std::condition_variable Cond;
40 
41 public:
42  explicit Latch(uint32_t Count = 0) : Count(Count) {}
43  ~Latch() { sync(); }
44 
45  void inc() {
46  std::lock_guard<std::mutex> lock(Mutex);
47  ++Count;
48  }
49 
50  void dec() {
51  std::lock_guard<std::mutex> lock(Mutex);
52  if (--Count == 0)
53  Cond.notify_all();
54  }
55 
56  void sync() const {
57  std::unique_lock<std::mutex> lock(Mutex);
58  Cond.wait(lock, [&] { return Count == 0; });
59  }
60 };
61 
62 class TaskGroup {
63  Latch L;
64  bool Parallel;
65 
66 public:
67  TaskGroup();
68  ~TaskGroup();
69 
70  void spawn(std::function<void()> f);
71 
72  void sync() const { L.sync(); }
73 };
74 
76 
77 /// Inclusive median.
78 template <class RandomAccessIterator, class Comparator>
79 RandomAccessIterator medianOf3(RandomAccessIterator Start,
80  RandomAccessIterator End,
81  const Comparator &Comp) {
82  RandomAccessIterator Mid = Start + (std::distance(Start, End) / 2);
83  return Comp(*Start, *(End - 1))
84  ? (Comp(*Mid, *(End - 1)) ? (Comp(*Start, *Mid) ? Mid : Start)
85  : End - 1)
86  : (Comp(*Mid, *Start) ? (Comp(*(End - 1), *Mid) ? Mid : End - 1)
87  : Start);
88 }
89 
90 template <class RandomAccessIterator, class Comparator>
91 void parallel_quick_sort(RandomAccessIterator Start, RandomAccessIterator End,
92  const Comparator &Comp, TaskGroup &TG, size_t Depth) {
93  // Do a sequential sort for small inputs.
94  if (std::distance(Start, End) < detail::MinParallelSize || Depth == 0) {
95  llvm::sort(Start, End, Comp);
96  return;
97  }
98 
99  // Partition.
100  auto Pivot = medianOf3(Start, End, Comp);
101  // Move Pivot to End.
102  std::swap(*(End - 1), *Pivot);
103  Pivot = std::partition(Start, End - 1, [&Comp, End](decltype(*Start) V) {
104  return Comp(V, *(End - 1));
105  });
106  // Move Pivot to middle of partition.
107  std::swap(*Pivot, *(End - 1));
108 
109  // Recurse.
110  TG.spawn([=, &Comp, &TG] {
111  parallel_quick_sort(Start, Pivot, Comp, TG, Depth - 1);
112  });
113  parallel_quick_sort(Pivot + 1, End, Comp, TG, Depth - 1);
114 }
115 
116 template <class RandomAccessIterator, class Comparator>
117 void parallel_sort(RandomAccessIterator Start, RandomAccessIterator End,
118  const Comparator &Comp) {
119  TaskGroup TG;
120  parallel_quick_sort(Start, End, Comp, TG,
121  llvm::Log2_64(std::distance(Start, End)) + 1);
122 }
123 
124 // TaskGroup has a relatively high overhead, so we want to reduce
125 // the number of spawn() calls. We'll create up to 1024 tasks here.
126 // (Note that 1024 is an arbitrary number. This code probably needs
127 // improving to take the number of available cores into account.)
128 enum { MaxTasksPerGroup = 1024 };
129 
130 template <class IterTy, class FuncTy>
131 void parallel_for_each(IterTy Begin, IterTy End, FuncTy Fn) {
132  // Limit the number of tasks to MaxTasksPerGroup to limit job scheduling
133  // overhead on large inputs.
134  ptrdiff_t TaskSize = std::distance(Begin, End) / MaxTasksPerGroup;
135  if (TaskSize == 0)
136  TaskSize = 1;
137 
138  TaskGroup TG;
139  while (TaskSize < std::distance(Begin, End)) {
140  TG.spawn([=, &Fn] { std::for_each(Begin, Begin + TaskSize, Fn); });
141  Begin += TaskSize;
142  }
143  std::for_each(Begin, End, Fn);
144 }
145 
146 template <class IndexTy, class FuncTy>
147 void parallel_for_each_n(IndexTy Begin, IndexTy End, FuncTy Fn) {
148  // Limit the number of tasks to MaxTasksPerGroup to limit job scheduling
149  // overhead on large inputs.
150  ptrdiff_t TaskSize = (End - Begin) / MaxTasksPerGroup;
151  if (TaskSize == 0)
152  TaskSize = 1;
153 
154  TaskGroup TG;
155  IndexTy I = Begin;
156  for (; I + TaskSize < End; I += TaskSize) {
157  TG.spawn([=, &Fn] {
158  for (IndexTy J = I, E = I + TaskSize; J != E; ++J)
159  Fn(J);
160  });
161  }
162  for (IndexTy J = I; J < End; ++J)
163  Fn(J);
164 }
165 
166 template <class IterTy, class ResultTy, class ReduceFuncTy,
167  class TransformFuncTy>
168 ResultTy parallel_transform_reduce(IterTy Begin, IterTy End, ResultTy Init,
169  ReduceFuncTy Reduce,
170  TransformFuncTy Transform) {
171  // Limit the number of tasks to MaxTasksPerGroup to limit job scheduling
172  // overhead on large inputs.
173  size_t NumInputs = std::distance(Begin, End);
174  if (NumInputs == 0)
175  return std::move(Init);
176  size_t NumTasks = std::min(static_cast<size_t>(MaxTasksPerGroup), NumInputs);
177  std::vector<ResultTy> Results(NumTasks, Init);
178  {
179  // Each task processes either TaskSize or TaskSize+1 inputs. Any inputs
180  // remaining after dividing them equally amongst tasks are distributed as
181  // one extra input over the first tasks.
182  TaskGroup TG;
183  size_t TaskSize = NumInputs / NumTasks;
184  size_t RemainingInputs = NumInputs % NumTasks;
185  IterTy TBegin = Begin;
186  for (size_t TaskId = 0; TaskId < NumTasks; ++TaskId) {
187  IterTy TEnd = TBegin + TaskSize + (TaskId < RemainingInputs ? 1 : 0);
188  TG.spawn([=, &Transform, &Reduce, &Results] {
189  // Reduce the result of transformation eagerly within each task.
190  ResultTy R = Init;
191  for (IterTy It = TBegin; It != TEnd; ++It)
192  R = Reduce(R, Transform(*It));
193  Results[TaskId] = R;
194  });
195  TBegin = TEnd;
196  }
197  assert(TBegin == End);
198  }
199 
200  // Do a final reduction. There are at most 1024 tasks, so this only adds
201  // constant single-threaded overhead for large inputs. Hopefully most
202  // reductions are cheaper than the transformation.
203  ResultTy FinalResult = std::move(Results.front());
204  for (ResultTy &PartialResult :
205  makeMutableArrayRef(Results.data() + 1, Results.size() - 1))
206  FinalResult = Reduce(FinalResult, std::move(PartialResult));
207  return std::move(FinalResult);
208 }
209 
210 #endif
211 
212 } // namespace detail
213 } // namespace parallel
214 
215 template <class RandomAccessIterator,
216  class Comparator = std::less<
217  typename std::iterator_traits<RandomAccessIterator>::value_type>>
218 void parallelSort(RandomAccessIterator Start, RandomAccessIterator End,
219  const Comparator &Comp = Comparator()) {
220 #if LLVM_ENABLE_THREADS
221  if (parallel::strategy.ThreadsRequested != 1) {
222  parallel::detail::parallel_sort(Start, End, Comp);
223  return;
224  }
225 #endif
226  llvm::sort(Start, End, Comp);
227 }
228 
229 template <class IterTy, class FuncTy>
230 void parallelForEach(IterTy Begin, IterTy End, FuncTy Fn) {
231 #if LLVM_ENABLE_THREADS
232  if (parallel::strategy.ThreadsRequested != 1) {
233  parallel::detail::parallel_for_each(Begin, End, Fn);
234  return;
235  }
236 #endif
237  std::for_each(Begin, End, Fn);
238 }
239 
240 template <class FuncTy>
241 void parallelForEachN(size_t Begin, size_t End, FuncTy Fn) {
242 #if LLVM_ENABLE_THREADS
243  if (parallel::strategy.ThreadsRequested != 1) {
245  return;
246  }
247 #endif
248  for (size_t I = Begin; I != End; ++I)
249  Fn(I);
250 }
251 
252 template <class IterTy, class ResultTy, class ReduceFuncTy,
253  class TransformFuncTy>
254 ResultTy parallelTransformReduce(IterTy Begin, IterTy End, ResultTy Init,
255  ReduceFuncTy Reduce,
256  TransformFuncTy Transform) {
257 #if LLVM_ENABLE_THREADS
258  if (parallel::strategy.ThreadsRequested != 1) {
259  return parallel::detail::parallel_transform_reduce(Begin, End, Init, Reduce,
260  Transform);
261  }
262 #endif
263  for (IterTy I = Begin; I != End; ++I)
264  Init = Reduce(std::move(Init), Transform(*I));
265  return std::move(Init);
266 }
267 
268 // Range wrappers.
269 template <class RangeTy,
270  class Comparator = std::less<decltype(*std::begin(RangeTy()))>>
271 void parallelSort(RangeTy &&R, const Comparator &Comp = Comparator()) {
272  parallelSort(std::begin(R), std::end(R), Comp);
273 }
274 
275 template <class RangeTy, class FuncTy>
276 void parallelForEach(RangeTy &&R, FuncTy Fn) {
277  parallelForEach(std::begin(R), std::end(R), Fn);
278 }
279 
280 template <class RangeTy, class ResultTy, class ReduceFuncTy,
281  class TransformFuncTy>
282 ResultTy parallelTransformReduce(RangeTy &&R, ResultTy Init,
283  ReduceFuncTy Reduce,
284  TransformFuncTy Transform) {
285  return parallelTransformReduce(std::begin(R), std::end(R), Init, Reduce,
286  Transform);
287 }
288 
289 // Parallel for-each, but with error handling.
290 template <class RangeTy, class FuncTy>
291 Error parallelForEachError(RangeTy &&R, FuncTy Fn) {
292  // The transform_reduce algorithm requires that the initial value be copyable.
293  // Error objects are uncopyable. We only need to copy initial success values,
294  // so work around this mismatch via the C API. The C API represents success
295  // values with a null pointer. The joinErrors discards null values and joins
296  // multiple errors into an ErrorList.
299  [](LLVMErrorRef Lhs, LLVMErrorRef Rhs) {
300  return wrap(joinErrors(unwrap(Lhs), unwrap(Rhs)));
301  },
302  [&Fn](auto &&V) { return wrap(Fn(V)); }));
303 }
304 
305 } // namespace llvm
306 
307 #endif // LLVM_SUPPORT_PARALLEL_H
MathExtras.h
llvm
This class represents lattice values for constants.
Definition: AllocatorList.h:23
llvm::parallel::detail::TaskGroup::~TaskGroup
~TaskGroup()
Definition: Parallel.cpp:154
llvm::parallelForEachError
Error parallelForEachError(RangeTy &&R, FuncTy Fn)
Definition: Parallel.h:291
llvm::ThreadPoolStrategy
This tells how a thread pool will be used.
Definition: Threading.h:148
llvm::parallel::strategy
ThreadPoolStrategy strategy
Definition: Parallel.cpp:20
llvm::parallel::detail::parallel_for_each_n
void parallel_for_each_n(IndexTy Begin, IndexTy End, FuncTy Fn)
Definition: Parallel.h:147
llvm::Error::success
static ErrorSuccess success()
Create a success value.
Definition: Error.h:332
Error.h
llvm::Depth
@ Depth
Definition: SIMachineScheduler.h:34
llvm::parallelForEachN
void parallelForEachN(size_t Begin, size_t End, FuncTy Fn)
Definition: Parallel.h:241
llvm::sys::path::end
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:233
llvm::sys::path::begin
const_iterator begin(StringRef path, Style style=Style::native)
Get begin iterator over path.
Definition: Path.cpp:224
Results
Function Alias Analysis Results
Definition: AliasAnalysis.cpp:847
llvm::parallel::detail::TaskGroup::spawn
void spawn(std::function< void()> f)
Definition: Parallel.cpp:156
STLExtras.h
llvm::lltok::less
@ less
Definition: LLToken.h:32
llvm::parallel::detail::parallel_for_each
void parallel_for_each(IterTy Begin, IterTy End, FuncTy Fn)
Definition: Parallel.h:131
llvm::Log2_64
unsigned Log2_64(uint64_t Value)
Return the floor log base 2 of the specified value, -1 if the value is zero.
Definition: MathExtras.h:603
ptrdiff_t
f
Itanium Name Demangler i e convert the string _Z1fv into f()". You can also use the CRTP base ManglingParser to perform some simple analysis on the mangled name
E
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
llvm::parallel::detail::parallel_transform_reduce
ResultTy parallel_transform_reduce(IterTy Begin, IterTy End, ResultTy Init, ReduceFuncTy Reduce, TransformFuncTy Transform)
Definition: Parallel.h:168
llvm::unwrap
Attribute unwrap(LLVMAttributeRef Attr)
Definition: Attributes.h:229
llvm::parallel::detail::Latch
Definition: Parallel.h:36
llvm::sys::SmartMutex< false >
llvm::parallel::detail::TaskGroup::TaskGroup
TaskGroup()
Definition: Parallel.cpp:153
llvm::parallelForEach
void parallelForEach(IterTy Begin, IterTy End, FuncTy Fn)
Definition: Parallel.h:230
Threading.h
llvm::parallel::detail::Latch::~Latch
~Latch()
Definition: Parallel.h:43
llvm::for_each
UnaryFunction for_each(R &&Range, UnaryFunction F)
Provide wrappers to std::for_each which take ranges instead of having to pass begin/end explicitly.
Definition: STLExtras.h:1491
llvm::parallel::detail::MinParallelSize
const ptrdiff_t MinParallelSize
Definition: Parallel.h:75
move
compiles ldr LCPI1_0 ldr ldr mov lsr tst moveq r1 ldr LCPI1_1 and r0 bx lr It would be better to do something like to fold the shift into the conditional move
Definition: README.txt:546
llvm::parallel::detail::Latch::dec
void dec()
Definition: Parallel.h:50
I
#define I(x, y, z)
Definition: MD5.cpp:59
llvm::parallelTransformReduce
ResultTy parallelTransformReduce(IterTy Begin, IterTy End, ResultTy Init, ReduceFuncTy Reduce, TransformFuncTy Transform)
Definition: Parallel.h:254
llvm::joinErrors
Error joinErrors(Error E1, Error E2)
Concatenate errors.
Definition: Error.h:429
assert
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
std::swap
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:944
llvm::parallel::detail::TaskGroup::sync
void sync() const
Definition: Parallel.h:72
function
print Print MemDeps of function
Definition: MemDepPrinter.cpp:83
llvm::wrap
LLVMAttributeRef wrap(Attribute Attr)
Definition: Attributes.h:224
llvm::parallel::detail::Latch::Latch
Latch(uint32_t Count=0)
Definition: Parallel.h:42
llvm::min
Expected< ExpressionValue > min(const ExpressionValue &Lhs, const ExpressionValue &Rhs)
Definition: FileCheck.cpp:339
Cond
SmallVector< MachineOperand, 4 > Cond
Definition: BasicBlockSections.cpp:167
llvm::parallelSort
void parallelSort(RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp=Comparator())
Definition: Parallel.h:218
uint32_t
llvm::parallel::detail::MaxTasksPerGroup
@ MaxTasksPerGroup
Definition: Parallel.h:128
llvm::partition
auto partition(R &&Range, UnaryPredicate P)
Provide wrappers to std::partition which take ranges instead of having to pass begin/end explicitly.
Definition: STLExtras.h:1602
llvm::parallel::detail::TaskGroup
Definition: Parallel.h:62
llvm::Init
Definition: Record.h:271
llvm::parallel::detail::Latch::sync
void sync() const
Definition: Parallel.h:56
llvm::parallel::detail::parallel_sort
void parallel_sort(RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp)
Definition: Parallel.h:117
LLVMErrorRef
struct LLVMOpaqueError * LLVMErrorRef
Opaque reference to an error instance.
Definition: Error.h:26
llvm::Error
Lightweight error class with error context and mandatory checking.
Definition: Error.h:157
llvm::sort
void sort(IteratorTy Start, IteratorTy End)
Definition: STLExtras.h:1439
llvm::parallel::detail::parallel_quick_sort
void parallel_quick_sort(RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp, TaskGroup &TG, size_t Depth)
Definition: Parallel.h:91
llvm::makeMutableArrayRef
MutableArrayRef< T > makeMutableArrayRef(T &OneElt)
Construct a MutableArrayRef from a single element.
Definition: ArrayRef.h:532
llvm::parallel::detail::Latch::inc
void inc()
Definition: Parallel.h:45
llvm::parallel::detail::medianOf3
RandomAccessIterator medianOf3(RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp)
Inclusive median.
Definition: Parallel.h:79