2 * Copyright 2016 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
19 #include <folly/Portability.h>
20 #include <folly/Preprocessor.h> // for FB_ANONYMOUS_VARIABLE
21 #include <folly/ScopeGuard.h>
22 #include <folly/Traits.h>
23 #include <folly/portability/GFlags.h>
24 #include <folly/portability/Time.h>
28 #include <boost/function_types/function_arity.hpp>
30 #include <glog/logging.h>
32 #include <type_traits>
34 DECLARE_bool(benchmark);
39 * Runs all benchmarks defined. Usually put in main().
44 * Runs all benchmarks defined if and only if the --benchmark flag has
45 * been passed to the program. Usually put in main().
47 inline bool runBenchmarksOnFlag() {
48 if (FLAGS_benchmark) {
51 return FLAGS_benchmark;
56 typedef std::pair<uint64_t, unsigned int> TimeIterPair;
59 * Adds a benchmark wrapped in a std::function. Only used
60 * internally. Pass by value is intentional.
62 void addBenchmarkImpl(const char* file,
64 std::function<TimeIterPair(unsigned int)>);
67 * Takes the difference between two timespec values. end is assumed to
70 inline uint64_t timespecDiff(timespec end, timespec start) {
71 if (end.tv_sec == start.tv_sec) {
72 assert(end.tv_nsec >= start.tv_nsec);
73 return end.tv_nsec - start.tv_nsec;
75 assert(end.tv_sec > start.tv_sec);
76 auto diff = uint64_t(end.tv_sec - start.tv_sec);
78 std::numeric_limits<uint64_t>::max() / 1000000000UL);
79 return diff * 1000000000UL
80 + end.tv_nsec - start.tv_nsec;
84 * Takes the difference between two sets of timespec values. The first
85 * two come from a high-resolution clock whereas the other two come
86 * from a low-resolution clock. The crux of the matter is that
87 * high-res values may be bogus as documented in
88 * http://linux.die.net/man/3/clock_gettime. The trouble is when the
89 * running process migrates from one CPU to another, which is more
90 * likely for long-running processes. Therefore we watch for high
91 * differences between the two timings.
93 * This function is subject to further improvements.
95 inline uint64_t timespecDiff(timespec end, timespec start,
96 timespec endCoarse, timespec startCoarse) {
97 auto fine = timespecDiff(end, start);
98 auto coarse = timespecDiff(endCoarse, startCoarse);
99 if (coarse - fine >= 1000000) {
100 // The fine time is in all likelihood bogus
106 } // namespace detail
109 * Supporting type for BENCHMARK_SUSPEND defined below.
111 struct BenchmarkSuspender {
112 BenchmarkSuspender() {
113 CHECK_EQ(0, clock_gettime(CLOCK_REALTIME, &start));
116 BenchmarkSuspender(const BenchmarkSuspender &) = delete;
117 BenchmarkSuspender(BenchmarkSuspender && rhs) noexcept {
122 BenchmarkSuspender& operator=(const BenchmarkSuspender &) = delete;
123 BenchmarkSuspender& operator=(BenchmarkSuspender && rhs) {
124 if (start.tv_nsec > 0 || start.tv_sec > 0) {
132 ~BenchmarkSuspender() {
133 if (start.tv_nsec > 0 || start.tv_sec > 0) {
139 assert(start.tv_nsec > 0 || start.tv_sec > 0);
145 assert(start.tv_nsec == 0 || start.tv_sec == 0);
146 CHECK_EQ(0, clock_gettime(CLOCK_REALTIME, &start));
150 auto dismissing(F f) -> typename std::result_of<F()>::type {
151 SCOPE_EXIT { rehire(); };
157 * This is for use inside of if-conditions, used in BENCHMARK macros.
158 * If-conditions bypass the explicit on operator bool.
160 explicit operator bool() const {
165 * Accumulates nanoseconds spent outside benchmark.
167 typedef uint64_t NanosecondsSpent;
168 static NanosecondsSpent nsSpent;
173 CHECK_EQ(0, clock_gettime(CLOCK_REALTIME, &end));
174 nsSpent += detail::timespecDiff(end, start);
182 * Adds a benchmark. Usually not called directly but instead through
183 * the macro BENCHMARK defined below. The lambda function involved
184 * must take exactly one parameter of type unsigned, and the benchmark
185 * uses it with counter semantics (iteration occurs inside the
188 template <typename Lambda>
189 typename std::enable_if<
190 boost::function_types::function_arity<decltype(&Lambda::operator())>::value
193 addBenchmark(const char* file, const char* name, Lambda&& lambda) {
194 auto execute = [=](unsigned int times) {
195 BenchmarkSuspender::nsSpent = 0;
199 // CORE MEASUREMENT STARTS
200 auto const r1 = clock_gettime(CLOCK_REALTIME, &start);
201 niter = lambda(times);
202 auto const r2 = clock_gettime(CLOCK_REALTIME, &end);
203 // CORE MEASUREMENT ENDS
208 return detail::TimeIterPair(
209 detail::timespecDiff(end, start) - BenchmarkSuspender::nsSpent,
213 detail::addBenchmarkImpl(file, name,
214 std::function<detail::TimeIterPair(unsigned int)>(execute));
218 * Adds a benchmark. Usually not called directly but instead through
219 * the macro BENCHMARK defined below. The lambda function involved
220 * must take zero parameters, and the benchmark calls it repeatedly
221 * (iteration occurs outside the function).
223 template <typename Lambda>
224 typename std::enable_if<
225 boost::function_types::function_arity<decltype(&Lambda::operator())>::value
228 addBenchmark(const char* file, const char* name, Lambda&& lambda) {
229 addBenchmark(file, name, [=](unsigned int times) {
230 unsigned int niter = 0;
231 while (times-- > 0) {
239 * Call doNotOptimizeAway(var) to ensure that var will be computed even
240 * post-optimization. Use it for variables that are computed during
241 * benchmarking but otherwise are useless. The compiler tends to do a
242 * good job at eliminating unused variables, and this function fools it
243 * into thinking var is in fact needed.
245 * Call makeUnpredictable(var) when you don't want the optimizer to use
246 * its knowledge of var to shape the following code. This is useful
247 * when constant propagation or power reduction is possible during your
248 * benchmark but not in real use cases.
253 #pragma optimize("", off)
255 inline void doNotOptimizeDependencySink(const void*) {}
257 #pragma optimize("", on)
260 void doNotOptimizeAway(const T& datum) {
261 doNotOptimizeDependencySink(&datum);
264 template <typename T>
265 void makeUnpredictable(T& datum) {
266 doNotOptimizeDependencySink(&datum);
272 template <typename T>
273 struct DoNotOptimizeAwayNeedsIndirect {
274 using Decayed = typename std::decay<T>::type;
276 // First two constraints ensure it can be an "r" operand.
277 // std::is_pointer check is because callers seem to expect that
278 // doNotOptimizeAway(&x) is equivalent to doNotOptimizeAway(x).
279 constexpr static bool value = !folly::IsTriviallyCopyable<Decayed>::value ||
280 sizeof(Decayed) > sizeof(long) || std::is_pointer<Decayed>::value;
282 } // detail namespace
284 template <typename T>
285 auto doNotOptimizeAway(const T& datum) -> typename std::enable_if<
286 !detail::DoNotOptimizeAwayNeedsIndirect<T>::value>::type {
287 // The "r" constraint forces the compiler to make datum available
288 // in a register to the asm block, which means that it must have
289 // computed/loaded it. We use this path for things that are <=
290 // sizeof(long) (they have to fit), trivial (otherwise the compiler
291 // doesn't want to put them in a register), and not a pointer (because
292 // doNotOptimizeAway(&foo) would otherwise be a foot gun that didn't
293 // necessarily compute foo).
295 // An earlier version of this method had a more permissive input operand
296 // constraint, but that caused unnecessary variation between clang and
298 asm volatile("" ::"r"(datum));
301 template <typename T>
302 auto doNotOptimizeAway(const T& datum) -> typename std::enable_if<
303 detail::DoNotOptimizeAwayNeedsIndirect<T>::value>::type {
304 // This version of doNotOptimizeAway tells the compiler that the asm
305 // block will read datum from memory, and that in addition it might read
306 // or write from any memory location. If the memory clobber could be
307 // separated into input and output that would be preferrable.
308 asm volatile("" ::"m"(datum) : "memory");
311 template <typename T>
312 auto makeUnpredictable(T& datum) -> typename std::enable_if<
313 !detail::DoNotOptimizeAwayNeedsIndirect<T>::value>::type {
314 asm volatile("" : "+r"(datum));
317 template <typename T>
318 auto makeUnpredictable(T& datum) -> typename std::enable_if<
319 detail::DoNotOptimizeAwayNeedsIndirect<T>::value>::type {
320 asm volatile("" ::"m"(datum) : "memory");
328 * Introduces a benchmark function. Used internally, see BENCHMARK and
331 #define BENCHMARK_IMPL(funName, stringName, rv, paramType, paramName) \
332 static void funName(paramType); \
333 static bool FB_ANONYMOUS_VARIABLE(follyBenchmarkUnused) = ( \
334 ::folly::addBenchmark(__FILE__, stringName, \
335 [](paramType paramName) -> unsigned { funName(paramName); \
338 static void funName(paramType paramName)
341 * Introduces a benchmark function with support for returning the actual
342 * number of iterations. Used internally, see BENCHMARK_MULTI and friends
345 #define BENCHMARK_MULTI_IMPL(funName, stringName, paramType, paramName) \
346 static unsigned funName(paramType); \
347 static bool FB_ANONYMOUS_VARIABLE(follyBenchmarkUnused) = ( \
348 ::folly::addBenchmark(__FILE__, stringName, \
349 [](paramType paramName) { return funName(paramName); }), \
351 static unsigned funName(paramType paramName)
354 * Introduces a benchmark function. Use with either one or two arguments.
355 * The first is the name of the benchmark. Use something descriptive, such
356 * as insertVectorBegin. The second argument may be missing, or could be a
357 * symbolic counter. The counter dictates how many internal iteration the
358 * benchmark does. Example:
360 * BENCHMARK(vectorPushBack) {
365 * BENCHMARK(insertVectorBegin, n) {
367 * FOR_EACH_RANGE (i, 0, n) {
368 * v.insert(v.begin(), 42);
372 #define BENCHMARK(name, ...) \
375 FB_STRINGIZE(name), \
376 FB_ARG_2_OR_1(1, ## __VA_ARGS__), \
377 FB_ONE_OR_NONE(unsigned, ## __VA_ARGS__), \
381 * Like BENCHMARK above, but allows the user to return the actual
382 * number of iterations executed in the function body. This can be
383 * useful if the benchmark function doesn't know upfront how many
384 * iterations it's going to run or if it runs through a certain
385 * number of test cases, e.g.:
387 * BENCHMARK_MULTI(benchmarkSomething) {
388 * std::vector<int> testCases { 0, 1, 1, 2, 3, 5 };
389 * for (int c : testCases) {
392 * return testCases.size();
395 #define BENCHMARK_MULTI(name, ...) \
396 BENCHMARK_MULTI_IMPL( \
398 FB_STRINGIZE(name), \
399 FB_ONE_OR_NONE(unsigned, ## __VA_ARGS__), \
403 * Defines a benchmark that passes a parameter to another one. This is
404 * common for benchmarks that need a "problem size" in addition to
405 * "number of iterations". Consider:
407 * void pushBack(uint n, size_t initialSize) {
409 * BENCHMARK_SUSPEND {
410 * v.resize(initialSize);
412 * FOR_EACH_RANGE (i, 0, n) {
416 * BENCHMARK_PARAM(pushBack, 0)
417 * BENCHMARK_PARAM(pushBack, 1000)
418 * BENCHMARK_PARAM(pushBack, 1000000)
420 * The benchmark above estimates the speed of push_back at different
421 * initial sizes of the vector. The framework will pass 0, 1000, and
422 * 1000000 for initialSize, and the iteration count for n.
424 #define BENCHMARK_PARAM(name, param) \
425 BENCHMARK_NAMED_PARAM(name, param, param)
428 * Same as BENCHMARK_PARAM, but allows one to return the actual number of
429 * iterations that have been run.
431 #define BENCHMARK_PARAM_MULTI(name, param) \
432 BENCHMARK_NAMED_PARAM_MULTI(name, param, param)
435 * Like BENCHMARK_PARAM(), but allows a custom name to be specified for each
436 * parameter, rather than using the parameter value.
438 * Useful when the parameter value is not a valid token for string pasting,
439 * of when you want to specify multiple parameter arguments.
443 * void addValue(uint n, int64_t bucketSize, int64_t min, int64_t max) {
444 * Histogram<int64_t> hist(bucketSize, min, max);
446 * FOR_EACH_RANGE (i, 0, n) {
447 * hist.addValue(num);
449 * if (num > max) { num = min; }
453 * BENCHMARK_NAMED_PARAM(addValue, 0_to_100, 1, 0, 100)
454 * BENCHMARK_NAMED_PARAM(addValue, 0_to_1000, 10, 0, 1000)
455 * BENCHMARK_NAMED_PARAM(addValue, 5k_to_20k, 250, 5000, 20000)
457 #define BENCHMARK_NAMED_PARAM(name, param_name, ...) \
459 FB_CONCATENATE(name, FB_CONCATENATE(_, param_name)), \
460 FB_STRINGIZE(name) "(" FB_STRINGIZE(param_name) ")", \
464 name(iters, ## __VA_ARGS__); \
468 * Same as BENCHMARK_NAMED_PARAM, but allows one to return the actual number
469 * of iterations that have been run.
471 #define BENCHMARK_NAMED_PARAM_MULTI(name, param_name, ...) \
472 BENCHMARK_MULTI_IMPL( \
473 FB_CONCATENATE(name, FB_CONCATENATE(_, param_name)), \
474 FB_STRINGIZE(name) "(" FB_STRINGIZE(param_name) ")", \
477 return name(iters, ## __VA_ARGS__); \
481 * Just like BENCHMARK, but prints the time relative to a
482 * baseline. The baseline is the most recent BENCHMARK() seen in
483 * the current scope. Example:
485 * // This is the baseline
486 * BENCHMARK(insertVectorBegin, n) {
488 * FOR_EACH_RANGE (i, 0, n) {
489 * v.insert(v.begin(), 42);
493 * BENCHMARK_RELATIVE(insertListBegin, n) {
495 * FOR_EACH_RANGE (i, 0, n) {
496 * s.insert(s.begin(), 42);
500 * Any number of relative benchmark can be associated with a
501 * baseline. Another BENCHMARK() occurrence effectively establishes a
504 #define BENCHMARK_RELATIVE(name, ...) \
507 "%" FB_STRINGIZE(name), \
508 FB_ARG_2_OR_1(1, ## __VA_ARGS__), \
509 FB_ONE_OR_NONE(unsigned, ## __VA_ARGS__), \
513 * Same as BENCHMARK_RELATIVE, but allows one to return the actual number
514 * of iterations that have been run.
516 #define BENCHMARK_RELATIVE_MULTI(name, ...) \
517 BENCHMARK_MULTI_IMPL( \
519 "%" FB_STRINGIZE(name), \
520 FB_ONE_OR_NONE(unsigned, ## __VA_ARGS__), \
524 * A combination of BENCHMARK_RELATIVE and BENCHMARK_PARAM.
526 #define BENCHMARK_RELATIVE_PARAM(name, param) \
527 BENCHMARK_RELATIVE_NAMED_PARAM(name, param, param)
530 * Same as BENCHMARK_RELATIVE_PARAM, but allows one to return the actual
531 * number of iterations that have been run.
533 #define BENCHMARK_RELATIVE_PARAM_MULTI(name, param) \
534 BENCHMARK_RELATIVE_NAMED_PARAM_MULTI(name, param, param)
537 * A combination of BENCHMARK_RELATIVE and BENCHMARK_NAMED_PARAM.
539 #define BENCHMARK_RELATIVE_NAMED_PARAM(name, param_name, ...) \
541 FB_CONCATENATE(name, FB_CONCATENATE(_, param_name)), \
542 "%" FB_STRINGIZE(name) "(" FB_STRINGIZE(param_name) ")", \
546 name(iters, ## __VA_ARGS__); \
550 * Same as BENCHMARK_RELATIVE_NAMED_PARAM, but allows one to return the
551 * actual number of iterations that have been run.
553 #define BENCHMARK_RELATIVE_NAMED_PARAM_MULTI(name, param_name, ...) \
554 BENCHMARK_MULTI_IMPL( \
555 FB_CONCATENATE(name, FB_CONCATENATE(_, param_name)), \
556 "%" FB_STRINGIZE(name) "(" FB_STRINGIZE(param_name) ")", \
559 return name(iters, ## __VA_ARGS__); \
563 * Draws a line of dashes.
565 #define BENCHMARK_DRAW_LINE() \
566 static bool FB_ANONYMOUS_VARIABLE(follyBenchmarkUnused) = ( \
567 ::folly::addBenchmark(__FILE__, "-", []() -> unsigned { return 0; }), \
571 * Allows execution of code that doesn't count torward the benchmark's
572 * time budget. Example:
574 * BENCHMARK_START_GROUP(insertVectorBegin, n) {
576 * BENCHMARK_SUSPEND {
579 * FOR_EACH_RANGE (i, 0, n) {
580 * v.insert(v.begin(), 42);
584 #define BENCHMARK_SUSPEND \
585 if (auto FB_ANONYMOUS_VARIABLE(BENCHMARK_SUSPEND) = \
586 ::folly::BenchmarkSuspender()) {} \