1 /*
2 * Copyright (c) 2010, 2025, 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<typename T>
36 class Queue {
37 class QueueNode : public CHeapObj<mtCompiler> {
38 T* _value;
39 QueueNode* _next;
40 public:
41 QueueNode(T* value, QueueNode* next) : _value(value), _next(next) { }
42 T* value() const { return _value; }
43 void set_next(QueueNode* next) { _next = next; }
44 QueueNode* next() const { return _next; }
45 };
46
47 QueueNode* _head;
48 QueueNode* _tail;
49 int _processing;
50
51 void push_unlocked(T* value) {
52 QueueNode* n = new QueueNode(value, nullptr);
53 if (_tail != nullptr) {
54 _tail->set_next(n);
55 }
56 _tail = n;
57 if (_head == nullptr) {
58 _head = _tail;
59 }
60 }
61 T* pop_unlocked() {
62 QueueNode* n = _head;
63 if (_head != nullptr) {
64 _head = _head->next();
65 }
66 if (_head == nullptr) {
67 _tail = _head;
68 }
69 T* value = nullptr;
70 if (n != nullptr) {
71 value = n->value();
72 delete n;
73 }
74 return value;
75 }
76 public:
77 Queue() : _head(nullptr), _tail(nullptr) { }
78 void push(T* value, Monitor* lock, JavaThread* current) {
79 MonitorLocker locker(current, lock);
80 push_unlocked(value);
81 locker.notify_all();
82 }
83
84 bool is_empty_unlocked() const { return _head == nullptr; }
85 bool is_processing_unlocked() const { return _processing > 0; }
86
87 T* pop(Monitor* lock, JavaThread* current) {
88 MonitorLocker locker(current, lock);
89 while (is_empty_unlocked() && !CompileBroker::is_compilation_disabled_forever()) {
90 locker.wait();
91 }
92 T* value = pop_unlocked();
93 return value;
94 }
95
96 T* try_pop(Monitor* lock, JavaThread* current) {
97 MonitorLocker locker(current, lock);
98 T* value = pop_unlocked();
99 return value;
100 }
101 void print_on(outputStream* st);
102 };
103 } // namespace CompilationPolicyUtils
104
105 class CompileTask;
106 class CompileQueue;
107 /*
108 * The system supports 5 execution levels:
109 * * level 0 - interpreter (Profiling is tracked by a MethodData object, or MDO in short)
110 * * level 1 - C1 with full optimization (no profiling)
111 * * level 2 - C1 with invocation and backedge counters
112 * * level 3 - C1 with full profiling (level 2 + All other MDO profiling information)
113 * * level 4 - C2 with full profile guided optimization
114 *
115 * The MethodData object is created by both the interpreter or either compiler to store any
116 * profiling information collected on a method (ciMethod::ensure_method_data() for C1 and C2
117 * and CompilationPolicy::create_mdo() for the interpreter). Both the interpreter and code
118 * compiled by C1 at level 3 will constantly update profiling information in the MDO during
119 * execution. The information in the MDO is then used by C1 and C2 during compilation, via
120 * the compiler interface (ciMethodXXX).
121 * See ciMethod.cpp and ciMethodData.cpp for information transfer from an MDO to the compilers
122 * through the compiler interface.
123 *
124 * Levels 0, 2 and 3 periodically notify the runtime about the current value of the counters
125 * (invocation counters and backedge counters). The frequency of these notifications is
126 * different at each level. These notifications are used by the policy to decide what transition
127 * to make.
128 *
129 * Execution starts at level 0 (interpreter), then the policy can decide either to compile the
130 * method at level 3 or level 2. The decision is based on the following factors:
131 * 1. The length of the C2 queue determines the next level. The observation is that level 2
132 * is generally faster than level 3 by about 30%, therefore we would want to minimize the time
133 * a method spends at level 3. We should only spend the time at level 3 that is necessary to get
134 * adequate profiling. So, if the C2 queue is long enough it is more beneficial to go first to
135 * level 2, because if we transitioned to level 3 we would be stuck there until our C2 compile
136 * request makes its way through the long queue. When the load on C2 recedes we are going to
137 * recompile at level 3 and start gathering profiling information.
138 * 2. The length of C1 queue is used to dynamically adjust the thresholds, so as to introduce
139 * additional filtering if the compiler is overloaded. The rationale is that by the time a
140 * method gets compiled it can become unused, so it doesn't make sense to put too much onto the
141 * queue.
142 *
143 * After profiling is completed at level 3 the transition is made to level 4. Again, the length
144 * of the C2 queue is used as a feedback to adjust the thresholds.
145 *
146 * After the first C1 compile some basic information is determined about the code like the number
147 * of the blocks and the number of the loops. Based on that it can be decided that a method
148 * is trivial and compiling it with C1 will yield the same code. In this case the method is
149 * compiled at level 1 instead of 4.
150 *
151 * We also support profiling at level 0. If C1 is slow enough to produce the level 3 version of
152 * the code and the C2 queue is sufficiently small we can decide to start profiling in the
153 * interpreter (and continue profiling in the compiled code once the level 3 version arrives).
154 * If the profiling at level 0 is fully completed before level 3 version is produced, a level 2
155 * version is compiled instead in order to run faster waiting for a level 4 version.
156 *
157 * Compile queues are implemented as priority queues - for each method in the queue we compute
158 * the event rate (the number of invocation and backedge counter increments per unit of time).
159 * When getting an element off the queue we pick the one with the largest rate. Maintaining the
160 * rate also allows us to remove stale methods (the ones that got on the queue but stopped
161 * being used shortly after that).
162 */
163
164 /* Command line options:
165 * - Tier?InvokeNotifyFreqLog and Tier?BackedgeNotifyFreqLog control the frequency of method
166 * invocation and backedge notifications. Basically every n-th invocation or backedge a mutator thread
167 * makes a call into the runtime.
168 *
169 * - Tier?InvocationThreshold, Tier?CompileThreshold, Tier?BackEdgeThreshold, Tier?MinInvocationThreshold control
170 * compilation thresholds.
171 * Level 2 thresholds are not used and are provided for option-compatibility and potential future use.
172 * Other thresholds work as follows:
173 *
174 * Transition from interpreter (level 0) to C1 with full profiling (level 3) happens when
175 * the following predicate is true (X is the level):
176 *
177 * i > TierXInvocationThreshold * s || (i > TierXMinInvocationThreshold * s && i + b > TierXCompileThreshold * s),
178 *
179 * where $i$ is the number of method invocations, $b$ number of backedges and $s$ is the scaling
180 * coefficient that will be discussed further.
181 * The intuition is to equalize the time that is spend profiling each method.
182 * The same predicate is used to control the transition from level 3 to level 4 (C2). It should be
183 * noted though that the thresholds are relative. Moreover i and b for the 0->3 transition come
184 * from Method* and for 3->4 transition they come from MDO (since profiled invocations are
185 * counted separately). Finally, if a method does not contain anything worth profiling, a transition
186 * from level 3 to level 4 occurs without considering thresholds (e.g., with fewer invocations than
187 * what is specified by Tier4InvocationThreshold).
188 *
189 * OSR transitions are controlled simply with b > TierXBackEdgeThreshold * s predicates.
190 *
191 * - Tier?LoadFeedback options are used to automatically scale the predicates described above depending
192 * on the compiler load. The scaling coefficients are computed as follows:
193 *
194 * s = queue_size_X / (TierXLoadFeedback * compiler_count_X) + 1,
195 *
196 * where queue_size_X is the current size of the compiler queue of level X, and compiler_count_X
197 * is the number of level X compiler threads.
198 *
199 * Basically these parameters describe how many methods should be in the compile queue
200 * per compiler thread before the scaling coefficient increases by one.
201 *
202 * This feedback provides the mechanism to automatically control the flow of compilation requests
203 * depending on the machine speed, mutator load and other external factors.
204 *
205 * - Tier3DelayOn and Tier3DelayOff parameters control another important feedback loop.
206 * Consider the following observation: a method compiled with full profiling (level 3)
207 * is about 30% slower than a method at level 2 (just invocation and backedge counters, no MDO).
208 * Normally, the following transitions will occur: 0->3->4. The problem arises when the C2 queue
209 * gets congested and the 3->4 transition is delayed. While the method is the C2 queue it continues
210 * executing at level 3 for much longer time than is required by the predicate and at suboptimal speed.
211 * The idea is to dynamically change the behavior of the system in such a way that if a substantial
212 * load on C2 is detected we would first do the 0->2 transition allowing a method to run faster.
213 * And then when the load decreases to allow 2->3 transitions.
214 *
215 * Tier3Delay* parameters control this switching mechanism.
216 * Tier3DelayOn is the number of methods in the C2 queue per compiler thread after which the policy
217 * no longer does 0->3 transitions but does 0->2 transitions instead.
218 * Tier3DelayOff switches the original behavior back when the number of methods in the C2 queue
219 * per compiler thread falls below the specified amount.
220 * The hysteresis is necessary to avoid jitter.
221 *
222 * - TieredCompileTaskTimeout is the amount of time an idle method can spend in the compile queue.
223 * Basically, since we use the event rate d(i + b)/dt as a value of priority when selecting a method to
224 * compile from the compile queue, we also can detect stale methods for which the rate has been
225 * 0 for some time in the same iteration. Stale methods can appear in the queue when an application
226 * abruptly changes its behavior.
227 *
228 * - TieredStopAtLevel, is used mostly for testing. It allows to bypass the policy logic and stick
229 * to a given level. For example it's useful to set TieredStopAtLevel = 1 in order to compile everything
230 * with pure c1.
231 *
232 * - Tier0ProfilingStartPercentage allows the interpreter to start profiling when the inequalities in the
233 * 0->3 predicate are already exceeded by the given percentage but the level 3 version of the
234 * method is still not ready. We can even go directly from level 0 to 4 if c1 doesn't produce a compiled
235 * version in time. This reduces the overall transition to level 4 and decreases the startup time.
236 * Note that this behavior is also guarded by the Tier3Delay mechanism: when the c2 queue is too long
237 * these is not reason to start profiling prematurely.
238 *
239 * - TieredRateUpdateMinTime and TieredRateUpdateMaxTime are parameters of the rate computation.
240 * Basically, the rate is not computed more frequently than TieredRateUpdateMinTime and is considered
241 * to be zero if no events occurred in TieredRateUpdateMaxTime.
242 */
243
244 class CompilationPolicy : AllStatic {
245 friend class CallPredicate;
246 friend class LoopPredicate;
247 friend class RecompilationPolicy;
248
249 typedef CompilationPolicyUtils::Queue<InstanceKlass> TrainingReplayQueue;
250
251 static int64_t _start_time;
252 static int _c1_count, _c2_count, _ac_count;
253 static double _increase_threshold_at_ratio;
254 static TrainingReplayQueue _training_replay_queue;
255
256 // Set carry flags in the counters (in Method* and MDO).
257 inline static void handle_counter_overflow(const methodHandle& method);
258 #ifdef ASSERT
259 // Verify that a level is consistent with the compilation mode
260 static bool verify_level(CompLevel level);
261 #endif
262 // Clamp the request level according to various constraints.
263 inline static CompLevel limit_level(CompLevel level);
264 // Common transition function. Given a predicate determines if a method should transition to another level.
265 template<typename Predicate>
266 static CompLevel common(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD, bool disable_feedback = false);
267
268 template<typename Predicate>
269 static CompLevel transition_from_none(const methodHandle& method, CompLevel cur_level, bool delay_profiling, bool disable_feedback);
270 template<typename Predicate>
271 static CompLevel transition_from_limited_profile(const methodHandle& method, CompLevel cur_level, bool delay_profiling, bool disable_feedback);
272 template<typename Predicate>
273 static CompLevel transition_from_full_profile(const methodHandle& method, CompLevel cur_level);
274 template<typename Predicate>
275 static CompLevel standard_transition(const methodHandle& method, CompLevel cur_level, bool delayprof, bool disable_feedback);
276
277 static CompLevel trained_transition_from_none(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD);
278 static CompLevel trained_transition_from_limited_profile(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD);
279 static CompLevel trained_transition_from_full_profile(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD);
280 static CompLevel trained_transition(const methodHandle& method, CompLevel cur_level, MethodTrainingData* mtd, JavaThread* THREAD);
281
282 // Transition functions.
283 // call_event determines if a method should be compiled at a different
284 // level with a regular invocation entry.
285 static CompLevel call_event(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD);
286 // loop_event checks if a method should be OSR compiled at a different
287 // level.
288 static CompLevel loop_event(const methodHandle& method, CompLevel cur_level, JavaThread* THREAD);
289 static void print_counters_on(outputStream* st, const char* prefix, Method* m);
290 static void print_training_data_on(outputStream* st, const char* prefix, Method* method);
291 // Has a method been long around?
292 // We don't remove old methods from the compile queue even if they have
293 // very low activity (see select_task()).
294 inline static bool is_old(const methodHandle& method);
295 // Was a given method inactive for a given number of milliseconds.
296 // If it is, we would remove it from the queue (see select_task()).
297 inline static bool is_stale(int64_t t, int64_t timeout, const methodHandle& method);
298 // Compute the weight of the method for the compilation scheduling
299 inline static double weight(Method* method);
300 // Apply heuristics and return true if x should be compiled before y
301 inline static bool compare_methods(Method* x, Method* y);
302 inline static bool compare_tasks(CompileTask* x, CompileTask* y);
303 // Compute event rate for a given method. The rate is the number of event (invocations + backedges)
304 // per millisecond.
305 inline static void update_rate(int64_t t, const methodHandle& method);
306 // Compute threshold scaling coefficient
307 inline static double threshold_scale(CompLevel level, int feedback_k);
308 // If a method is old enough and is still in the interpreter we would want to
309 // start profiling without waiting for the compiled method to arrive. This function
310 // determines whether we should do that.
311 inline static bool should_create_mdo(const methodHandle& method, CompLevel cur_level);
312 // Create MDO if necessary.
313 static void create_mdo(const methodHandle& mh, JavaThread* THREAD);
314 // Is method profiled enough?
315 static bool is_method_profiled(const methodHandle& method);
316
317 static void set_c1_count(int x) { _c1_count = x; }
318 static void set_c2_count(int x) { _c2_count = x; }
319 static void set_ac_count(int x) { _ac_count = x; }
320
321 enum EventType { CALL, LOOP, COMPILE, FORCE_COMPILE, FORCE_RECOMPILE, REMOVE_FROM_QUEUE, UPDATE_IN_QUEUE, REPROFILE, MAKE_NOT_ENTRANT };
322 static void print_event_on(outputStream *st, EventType type, Method* m, Method* im, int bci, CompLevel level);
323 static void print_event(EventType type, Method* m, Method* im, int bci, CompLevel level);
324 // Check if the method can be compiled, change level if necessary
325 static void compile(const methodHandle& mh, int bci, CompLevel level, TRAPS);
326 // Simple methods are as good being compiled with C1 as C2.
327 // This function tells if it's such a function.
328 inline static bool is_trivial(const methodHandle& method);
329 // Force method to be compiled at CompLevel_simple?
330 inline static bool force_comp_at_level_simple(const methodHandle& method);
331
332 // Get a compilation level for a given method.
333 static CompLevel comp_level(Method* method);
334 static void method_invocation_event(const methodHandle& method, const methodHandle& inlinee,
335 CompLevel level, nmethod* nm, TRAPS);
336 static void method_back_branch_event(const methodHandle& method, const methodHandle& inlinee,
337 int bci, CompLevel level, nmethod* nm, TRAPS);
338
339 static void set_increase_threshold_at_ratio() { _increase_threshold_at_ratio = 100 / (100 - (double)IncreaseFirstTierCompileThresholdAt); }
340 static void set_start_time(int64_t t) { _start_time = t; }
341 static int64_t start_time() { return _start_time; }
342
343 // m must be compiled before executing it
344 static bool must_be_compiled(const methodHandle& m, int comp_level = CompLevel_any);
345 static void maybe_compile_early(const methodHandle& m, TRAPS);
346 static void replay_training_at_init_impl(InstanceKlass* klass, JavaThread* current);
347 public:
348 static int min_invocations() { return Tier4MinInvocationThreshold; }
349 static int c1_count() { return _c1_count; }
350 static int c2_count() { return _c2_count; }
351 static int ac_count() { return _ac_count; }
352 static int compiler_count(CompLevel comp_level);
353 // If m must_be_compiled then request a compilation from the CompileBroker.
354 // This supports the -Xcomp option.
355 static void compile_if_required(const methodHandle& m, TRAPS);
356
357 static void replay_training_at_init(InstanceKlass* klass, JavaThread* current);
358 static void replay_training_at_init_loop(JavaThread* current);
359
360 // m is allowed to be compiled
361 static bool can_be_compiled(const methodHandle& m, int comp_level = CompLevel_any);
362 // m is allowed to be osr compiled
363 static bool can_be_osr_compiled(const methodHandle& m, int comp_level = CompLevel_any);
364 static bool is_compilation_enabled();
365
366 static CompileTask* select_task_helper(CompileQueue* compile_queue);
367 // Return initial compile level to use with Xcomp (depends on compilation mode).
368 static void reprofile(ScopeDesc* trap_scope, bool is_osr);
369 static nmethod* event(const methodHandle& method, const methodHandle& inlinee,
370 int branch_bci, int bci, CompLevel comp_level, nmethod* nm, TRAPS);
371 // Select task is called by CompileBroker. We should return a task or nullptr.
372 static CompileTask* select_task(CompileQueue* compile_queue, JavaThread* THREAD);
373 // Tell the runtime if we think a given method is adequately profiled.
374 static bool is_mature(MethodData* mdo);
375 // Initialize: set compiler thread count
376 static void initialize();
377 static bool should_not_inline(ciEnv* env, ciMethod* callee);
378
379 // Return desired initial compilation level for Xcomp
380 static CompLevel initial_compile_level(const methodHandle& method);
381 // Return highest level possible
382 static CompLevel highest_compile_level();
383 static void dump();
384
385 static void sample_load_average();
386 static bool have_recompilation_work();
387 static bool recompilation_step(int step, TRAPS);
388 };
389
390 #endif // SHARE_COMPILER_COMPILATIONPOLICY_HPP