1 /*
2 * Copyright (c) 2010, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #ifndef SHARE_COMPILER_COMPILATIONPOLICY_HPP
26 #define SHARE_COMPILER_COMPILATIONPOLICY_HPP
27
28 #include "code/nmethod.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "oops/methodData.hpp"
31 #include "oops/trainingData.hpp"
32 #include "utilities/globalDefinitions.hpp"
33
34 namespace CompilationPolicyUtils {
35 template<int SAMPLE_COUNT = 256>
36 class WeightedMovingAverage {
37 int _current;
38 int _samples[SAMPLE_COUNT];
39 int64_t _timestamps[SAMPLE_COUNT];
40
41 void sample(int s, int64_t t) {
42 assert(s >= 0, "Negative sample values are not supported");
43 _samples[_current] = s;
44 _timestamps[_current] = t;
45 if (++_current >= SAMPLE_COUNT) {
46 _current = 0;
47 }
48 }
49
50 // Since sampling happens at irregular invervals the solution is to
51 // discount the older samples proportionally to the time between
52 // the now and the time of the sample.
53 double value(int64_t t) const {
54 double decay_speed = 1;
55 double weighted_sum = 0;
56 int count = 0;
57 for (int i = 0; i < SAMPLE_COUNT; i++) {
58 if (_samples[i] >= 0) {
59 count++;
60 double delta_t = (t - _timestamps[i]) / 1000.0; // in seconds
61 if (delta_t < 1) delta_t = 1;
62 weighted_sum += (double) _samples[i] / (delta_t * decay_speed);
63 }
64 }
65 if (count > 0) {
66 return weighted_sum / count;
67 } else {
68 return 0;
69 }
70 }
71 static int64_t time() {
72 return nanos_to_millis(os::javaTimeNanos());
73 }
74 public:
75 WeightedMovingAverage() : _current(0) {
76 for (int i = 0; i < SAMPLE_COUNT; i++) {
77 _samples[i] = -1;
78 }
79 }
80 void sample(int s) { sample(s, time()); }
81 double value() const { return value(time()); }
82 };
83
84 template<typename T>
85 class Queue {
86 class QueueNode : public CHeapObj<mtCompiler> {
87 T* _value;
88 QueueNode* _next;
89 public:
90 QueueNode(T* value, QueueNode* next) : _value(value), _next(next) { }
91 T* value() const { return _value; }
92 void set_next(QueueNode* next) { _next = next; }
93 QueueNode* next() const { return _next; }
94 };
95
96 QueueNode* _head;
97 QueueNode* _tail;
98
99 void push_unlocked(T* value) {
100 QueueNode* n = new QueueNode(value, nullptr);
101 if (_tail != nullptr) {
102 _tail->set_next(n);
103 }
104 _tail = n;
105 if (_head == nullptr) {
106 _head = _tail;
107 }
108 }
109 T* pop_unlocked() {
110 QueueNode* n = _head;
111 if (_head != nullptr) {
112 _head = _head->next();
113 }
114 if (_head == nullptr) {
115 _tail = _head;
116 }
117 T* value = nullptr;
118 if (n != nullptr) {
119 value = n->value();
120 delete n;
121 }
122 return value;
123 }
124 public:
125 Queue() : _head(nullptr), _tail(nullptr) { }
126 void push(T* value, Monitor* lock, TRAPS) {
127 MonitorLocker locker(THREAD, lock);
128 push_unlocked(value);
129 locker.notify_all();
130 }
131
132 bool is_empty_unlocked() const { return _head == nullptr; }
133
134 T* pop(Monitor* lock, TRAPS) {
135 MonitorLocker locker(THREAD, lock);
136 while(is_empty_unlocked() && !CompileBroker::is_compilation_disabled_forever()) {
137 locker.notify_all(); // notify that queue is empty
138 locker.wait();
139 }
140 T* value = pop_unlocked();
141 return value;
142 }
143
144 T* try_pop(Monitor* lock, TRAPS) {
145 MonitorLocker locker(THREAD, lock);
146 T* value = nullptr;
147 if (!is_empty_unlocked()) {
148 value = pop_unlocked();
149 }
150 return value;
151 }
152
153 void print_on(outputStream* st);
154 };
155 } // namespace CompilationPolicyUtils
156
157 class CompileTask;
158 class CompileQueue;
159 /*
160 * The system supports 5 execution levels:
161 * * level 0 - interpreter (Profiling is tracked by a MethodData object, or MDO in short)
162 * * level 1 - C1 with full optimization (no profiling)
163 * * level 2 - C1 with invocation and backedge counters
164 * * level 3 - C1 with full profiling (level 2 + All other MDO profiling information)
165 * * level 4 - C2 with full profile guided optimization
166 *
167 * The MethodData object is created by both the interpreter or either compiler to store any
168 * profiling information collected on a method (ciMethod::ensure_method_data() for C1 and C2
169 * and CompilationPolicy::create_mdo() for the interpreter). Both the interpreter and code
170 * compiled by C1 at level 3 will constantly update profiling information in the MDO during
171 * execution. The information in the MDO is then used by C1 and C2 during compilation, via
172 * the compiler interface (ciMethodXXX).
173 * See ciMethod.cpp and ciMethodData.cpp for information transfer from an MDO to the compilers
174 * through the compiler interface.
175 *
176 * Levels 0, 2 and 3 periodically notify the runtime about the current value of the counters
177 * (invocation counters and backedge counters). The frequency of these notifications is
178 * different at each level. These notifications are used by the policy to decide what transition
179 * to make.
180 *
181 * Execution starts at level 0 (interpreter), then the policy can decide either to compile the
182 * method at level 3 or level 2. The decision is based on the following factors:
183 * 1. The length of the C2 queue determines the next level. The observation is that level 2
184 * is generally faster than level 3 by about 30%, therefore we would want to minimize the time
185 * a method spends at level 3. We should only spend the time at level 3 that is necessary to get
186 * adequate profiling. So, if the C2 queue is long enough it is more beneficial to go first to
187 * level 2, because if we transitioned to level 3 we would be stuck there until our C2 compile
188 * request makes its way through the long queue. When the load on C2 recedes we are going to
189 * recompile at level 3 and start gathering profiling information.
190 * 2. The length of C1 queue is used to dynamically adjust the thresholds, so as to introduce
191 * additional filtering if the compiler is overloaded. The rationale is that by the time a
192 * method gets compiled it can become unused, so it doesn't make sense to put too much onto the
193 * queue.
194 *
195 * After profiling is completed at level 3 the transition is made to level 4. Again, the length
196 * of the C2 queue is used as a feedback to adjust the thresholds.
197 *
198 * After the first C1 compile some basic information is determined about the code like the number
199 * of the blocks and the number of the loops. Based on that it can be decided that a method
200 * is trivial and compiling it with C1 will yield the same code. In this case the method is
201 * compiled at level 1 instead of 4.
202 *
203 * We also support profiling at level 0. If C1 is slow enough to produce the level 3 version of
204 * the code and the C2 queue is sufficiently small we can decide to start profiling in the
205 * interpreter (and continue profiling in the compiled code once the level 3 version arrives).
206 * If the profiling at level 0 is fully completed before level 3 version is produced, a level 2
207 * version is compiled instead in order to run faster waiting for a level 4 version.
208 *
209 * Compile queues are implemented as priority queues - for each method in the queue we compute
210 * the event rate (the number of invocation and backedge counter increments per unit of time).
211 * When getting an element off the queue we pick the one with the largest rate. Maintaining the
212 * rate also allows us to remove stale methods (the ones that got on the queue but stopped
213 * being used shortly after that).
214 */
215
216 /* Command line options:
217 * - Tier?InvokeNotifyFreqLog and Tier?BackedgeNotifyFreqLog control the frequency of method
218 * invocation and backedge notifications. Basically every n-th invocation or backedge a mutator thread
219 * makes a call into the runtime.
220 *
221 * - Tier?InvocationThreshold, Tier?CompileThreshold, Tier?BackEdgeThreshold, Tier?MinInvocationThreshold control
222 * compilation thresholds.
223 * Level 2 thresholds are not used and are provided for option-compatibility and potential future use.
224 * Other thresholds work as follows:
225 *
226 * Transition from interpreter (level 0) to C1 with full profiling (level 3) happens when
227 * the following predicate is true (X is the level):
228 *
229 * i > TierXInvocationThreshold * s || (i > TierXMinInvocationThreshold * s && i + b > TierXCompileThreshold * s),
230 *
231 * where $i$ is the number of method invocations, $b$ number of backedges and $s$ is the scaling
232 * coefficient that will be discussed further.
233 * The intuition is to equalize the time that is spend profiling each method.
234 * The same predicate is used to control the transition from level 3 to level 4 (C2). It should be
235 * noted though that the thresholds are relative. Moreover i and b for the 0->3 transition come
236 * from Method* and for 3->4 transition they come from MDO (since profiled invocations are
237 * counted separately). Finally, if a method does not contain anything worth profiling, a transition
238 * from level 3 to level 4 occurs without considering thresholds (e.g., with fewer invocations than
239 * what is specified by Tier4InvocationThreshold).
240 *
241 * OSR transitions are controlled simply with b > TierXBackEdgeThreshold * s predicates.
242 *
243 * - Tier?LoadFeedback options are used to automatically scale the predicates described above depending
244 * on the compiler load. The scaling coefficients are computed as follows:
245 *
246 * s = queue_size_X / (TierXLoadFeedback * compiler_count_X) + 1,
247 *
248 * where queue_size_X is the current size of the compiler queue of level X, and compiler_count_X
249 * is the number of level X compiler threads.
250 *
251 * Basically these parameters describe how many methods should be in the compile queue
252 * per compiler thread before the scaling coefficient increases by one.
253 *
254 * This feedback provides the mechanism to automatically control the flow of compilation requests
255 * depending on the machine speed, mutator load and other external factors.
256 *
257 * - Tier3DelayOn and Tier3DelayOff parameters control another important feedback loop.
258 * Consider the following observation: a method compiled with full profiling (level 3)
259 * is about 30% slower than a method at level 2 (just invocation and backedge counters, no MDO).
260 * Normally, the following transitions will occur: 0->3->4. The problem arises when the C2 queue
261 * gets congested and the 3->4 transition is delayed. While the method is the C2 queue it continues
262 * executing at level 3 for much longer time than is required by the predicate and at suboptimal speed.
263 * The idea is to dynamically change the behavior of the system in such a way that if a substantial
264 * load on C2 is detected we would first do the 0->2 transition allowing a method to run faster.
265 * And then when the load decreases to allow 2->3 transitions.
266 *
267 * Tier3Delay* parameters control this switching mechanism.
268 * Tier3DelayOn is the number of methods in the C2 queue per compiler thread after which the policy
269 * no longer does 0->3 transitions but does 0->2 transitions instead.
270 * Tier3DelayOff switches the original behavior back when the number of methods in the C2 queue
271 * per compiler thread falls below the specified amount.
272 * The hysteresis is necessary to avoid jitter.
273 *
274 * - TieredCompileTaskTimeout is the amount of time an idle method can spend in the compile queue.
275 * Basically, since we use the event rate d(i + b)/dt as a value of priority when selecting a method to
276 * compile from the compile queue, we also can detect stale methods for which the rate has been
277 * 0 for some time in the same iteration. Stale methods can appear in the queue when an application
278 * abruptly changes its behavior.
279 *
280 * - TieredStopAtLevel, is used mostly for testing. It allows to bypass the policy logic and stick
281 * to a given level. For example it's useful to set TieredStopAtLevel = 1 in order to compile everything
282 * with pure c1.
283 *
284 * - Tier0ProfilingStartPercentage allows the interpreter to start profiling when the inequalities in the
285 * 0->3 predicate are already exceeded by the given percentage but the level 3 version of the
286 * method is still not ready. We can even go directly from level 0 to 4 if c1 doesn't produce a compiled
287 * version in time. This reduces the overall transition to level 4 and decreases the startup time.
288 * Note that this behavior is also guarded by the Tier3Delay mechanism: when the c2 queue is too long
289 * these is not reason to start profiling prematurely.
290 *
291 * - TieredRateUpdateMinTime and TieredRateUpdateMaxTime are parameters of the rate computation.
292 * Basically, the rate is not computed more frequently than TieredRateUpdateMinTime and is considered
293 * to be zero if no events occurred in TieredRateUpdateMaxTime.
294 */
295
296 class CompilationPolicy : AllStatic {
297 friend class CallPredicate;
298 friend class LoopPredicate;
299
300 typedef CompilationPolicyUtils::WeightedMovingAverage<> LoadAverage;
301 typedef CompilationPolicyUtils::Queue<InstanceKlass> TrainingReplayQueue;
302
303 static int64_t _start_time;
304 static int _c1_count, _c2_count, _c3_count, _sc_count;
305 static double _increase_threshold_at_ratio;
306 static LoadAverage _load_average;
307 static volatile bool _recompilation_done;
308 static TrainingReplayQueue _training_replay_queue;
309
310 // Set carry flags in the counters (in Method* and MDO).
311 inline static void handle_counter_overflow(const methodHandle& method);
312 // Verify that a level is consistent with the compilation mode
313 static bool verify_level(CompLevel level);
314 // Clamp the request level according to various constraints.
315 inline static CompLevel limit_level(CompLevel level);
316 // Common transition function. Given a predicate determines if a method should transition to another level.
317 template<typename Predicate>
318 static CompLevel common(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD, bool disable_feedback = false);
319
320 template<typename Predicate>
321 static CompLevel transition_from_none(const methodHandle& method, CompLevel cur_level, bool delay_profiling, bool disable_feedback);
322 template<typename Predicate>
323 static CompLevel transition_from_limited_profile(const methodHandle& method, CompLevel cur_level, bool delay_profiling, bool disable_feedback);
324 template<typename Predicate>
325 static CompLevel transition_from_full_profile(const methodHandle& method, CompLevel cur_level);
326 template<typename Predicate>
327 static CompLevel standard_transition(const methodHandle& method, CompLevel cur_level, bool delayprof, bool disable_feedback);
328
329 static CompLevel trained_transition_from_none(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD);
330 static CompLevel trained_transition_from_limited_profile(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD);
331 static CompLevel trained_transition_from_full_profile(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD);
332 static CompLevel trained_transition(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD);
333
334 // Transition functions.
335 // call_event determines if a method should be compiled at a different
336 // level with a regular invocation entry.
337 static CompLevel call_event(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD);
338 // loop_event checks if a method should be OSR compiled at a different
339 // level.
340 static CompLevel loop_event(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD);
341 static void print_counters(const char* prefix, Method* m);
342 static void print_training_data(const char* prefix, Method* method);
343 // Has a method been long around?
344 // We don't remove old methods from the compile queue even if they have
345 // very low activity (see select_task()).
346 inline static bool is_old(const methodHandle& method);
347 // Was a given method inactive for a given number of milliseconds.
348 // If it is, we would remove it from the queue (see select_task()).
349 inline static bool is_stale(int64_t t, int64_t timeout, const methodHandle& method);
350 // Compute the weight of the method for the compilation scheduling
351 inline static double weight(Method* method);
352 // Apply heuristics and return true if x should be compiled before y
353 inline static bool compare_methods(Method* x, Method* y);
354 inline static bool compare_tasks(CompileTask* x, CompileTask* y);
355 // Compute event rate for a given method. The rate is the number of event (invocations + backedges)
356 // per millisecond.
357 inline static void update_rate(int64_t t, const methodHandle& method);
358 // Compute threshold scaling coefficient
359 inline static double threshold_scale(CompLevel level, int feedback_k);
360 // If a method is old enough and is still in the interpreter we would want to
361 // start profiling without waiting for the compiled method to arrive. This function
362 // determines whether we should do that.
363 inline static bool should_create_mdo(const methodHandle& method, CompLevel cur_level);
364 // Create MDO if necessary.
365 static void create_mdo(const methodHandle& mh, JavaThread* THREAD);
366 // Is method profiled enough?
367 static bool is_method_profiled(const methodHandle& method);
368
369 static void set_c1_count(int x) { _c1_count = x; }
370 static void set_c2_count(int x) { _c2_count = x; }
371 static void set_c3_count(int x) { _c3_count = x; }
372 static void set_sc_count(int x) { _sc_count = x; }
373
374 enum EventType { CALL, LOOP, COMPILE, FORCE_COMPILE, FORCE_RECOMPILE, REMOVE_FROM_QUEUE, UPDATE_IN_QUEUE, REPROFILE, MAKE_NOT_ENTRANT };
375 static void print_event(EventType type, Method* m, Method* im, int bci, CompLevel level);
376 // Check if the method can be compiled, change level if necessary
377 static void compile(const methodHandle& mh, int bci, CompLevel level, TRAPS);
378 // Simple methods are as good being compiled with C1 as C2.
379 // This function tells if it's such a function.
380 inline static bool is_trivial(const methodHandle& method);
381 // Force method to be compiled at CompLevel_simple?
382 inline static bool force_comp_at_level_simple(const methodHandle& method);
383
384 // Get a compilation level for a given method.
385 static CompLevel comp_level(Method* method);
386 static void method_invocation_event(const methodHandle& method, const methodHandle& inlinee,
387 CompLevel level, nmethod* nm, TRAPS);
388 static void method_back_branch_event(const methodHandle& method, const methodHandle& inlinee,
389 int bci, CompLevel level, nmethod* nm, TRAPS);
390
391 static void set_increase_threshold_at_ratio() { _increase_threshold_at_ratio = 100 / (100 - (double)IncreaseFirstTierCompileThresholdAt); }
392 static void set_start_time(int64_t t) { _start_time = t; }
393 static int64_t start_time() { return _start_time; }
394
395 // m must be compiled before executing it
396 static bool must_be_compiled(const methodHandle& m, int comp_level = CompLevel_any);
397 static void maybe_compile_early(const methodHandle& m, TRAPS);
398 static void maybe_compile_early_after_init(const methodHandle& m, TRAPS);
399 static void replay_training_at_init_impl(InstanceKlass* klass, TRAPS);
400 public:
401 static int min_invocations() { return Tier4MinInvocationThreshold; }
402 static int c1_count() { return _c1_count; }
403 static int c2_count() { return _c2_count; }
404 static int c3_count() { return _c3_count; }
405 static int sc_count() { return _sc_count; }
406 static int compiler_count(CompLevel comp_level);
407 // If m must_be_compiled then request a compilation from the CompileBroker.
408 // This supports the -Xcomp option.
409 static void compile_if_required(const methodHandle& m, TRAPS);
410
411 static void replay_training_at_init(bool is_on_shutdown, TRAPS);
412 static void replay_training_at_init(InstanceKlass* klass, TRAPS);
413 static void replay_training_at_init_loop(TRAPS);
414
415 // m is allowed to be compiled
416 static bool can_be_compiled(const methodHandle& m, int comp_level = CompLevel_any);
417 // m is allowed to be osr compiled
418 static bool can_be_osr_compiled(const methodHandle& m, int comp_level = CompLevel_any);
419 static bool is_compilation_enabled();
420
421 static CompileTask* select_task_helper(CompileQueue* compile_queue);
422 // Return initial compile level to use with Xcomp (depends on compilation mode).
423 static void reprofile(ScopeDesc* trap_scope, bool is_osr);
424 static nmethod* event(const methodHandle& method, const methodHandle& inlinee,
425 int branch_bci, int bci, CompLevel comp_level, nmethod* nm, TRAPS);
426 // Select task is called by CompileBroker. We should return a task or nullptr.
427 static CompileTask* select_task(CompileQueue* compile_queue, JavaThread* THREAD);
428 // Tell the runtime if we think a given method is adequately profiled.
429 static bool is_mature(MethodData* mdo);
430 // Initialize: set compiler thread count
431 static void initialize();
432 static bool should_not_inline(ciEnv* env, ciMethod* callee);
433
434 // Return desired initial compilation level for Xcomp
435 static CompLevel initial_compile_level(const methodHandle& method);
436 // Return highest level possible
437 static CompLevel highest_compile_level();
438 static void dump();
439
440 static void sample_load_average();
441 static bool have_recompilation_work();
442 static bool recompilation_step(int step, TRAPS);
443 };
444
445 #endif // SHARE_COMPILER_COMPILATIONPOLICY_HPP