1 /*
2 * Copyright (c) 1997, 2023, 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_RUNTIME_DEOPTIMIZATION_HPP
26 #define SHARE_RUNTIME_DEOPTIMIZATION_HPP
27
28 #include "interpreter/bytecodes.hpp"
29 #include "memory/allocation.hpp"
30 #include "runtime/frame.hpp"
31
32 class ProfileData;
33 class vframeArray;
34 class MonitorInfo;
35 class MonitorValue;
36 class ObjectValue;
37 class AutoBoxObjectValue;
38 class ScopeValue;
39 class compiledVFrame;
40
41 template<class E> class GrowableArray;
42
43 class DeoptimizationScope {
44 private:
45 // What gen we have done the deopt handshake for.
46 static uint64_t _committed_deopt_gen;
47 // What gen to mark a method with, hence larger than _committed_deopt_gen.
48 static uint64_t _active_deopt_gen;
49 // Indicate an in-progress deopt handshake.
50 static bool _committing_in_progress;
51
52 // The required gen we need to execute/wait for
53 uint64_t _required_gen;
54 DEBUG_ONLY(bool _deopted;)
55
56 public:
57 DeoptimizationScope();
58 ~DeoptimizationScope();
59 // Mark a method, if already marked as dependent.
60 void mark(CompiledMethod* cm, bool inc_recompile_counts = true);
61 // Record this as a dependent method.
62 void dependent(CompiledMethod* cm);
63
64 // Execute the deoptimization.
65 // Make the nmethods not entrant, stackwalks and patch return pcs and sets post call nops.
66 void deoptimize_marked();
67 };
68
69 class Deoptimization : AllStatic {
70 friend class VMStructs;
71 friend class EscapeBarrier;
72
73 public:
74 // What condition caused the deoptimization?
75 // Note: Keep this enum in sync. with Deoptimization::_trap_reason_name.
76 enum DeoptReason {
77 Reason_many = -1, // indicates presence of several reasons
78 Reason_none = 0, // indicates absence of a relevant deopt.
79 // Next 8 reasons are recorded per bytecode in DataLayout::trap_bits.
80 // This is more complicated for JVMCI as JVMCI may deoptimize to *some* bytecode before the
81 // bytecode that actually caused the deopt (with inlining, JVMCI may even deoptimize to a
82 // bytecode in another method):
83 // - bytecode y in method b() causes deopt
84 // - JVMCI deoptimizes to bytecode x in method a()
85 // -> the deopt reason will be recorded for method a() at bytecode x
86 Reason_null_check, // saw unexpected null or zero divisor (@bci)
87 Reason_null_assert, // saw unexpected non-null or non-zero (@bci)
88 Reason_range_check, // saw unexpected array index (@bci)
89 Reason_class_check, // saw unexpected object class (@bci)
90 Reason_array_check, // saw unexpected array class (aastore @bci)
91 Reason_intrinsic, // saw unexpected operand to intrinsic (@bci)
92 Reason_bimorphic, // saw unexpected object class in bimorphic inlining (@bci)
93
94 #if INCLUDE_JVMCI
95 Reason_unreached0 = Reason_null_assert,
96 Reason_type_checked_inlining = Reason_intrinsic,
97 Reason_optimized_type_check = Reason_bimorphic,
98 #endif
99
100 Reason_profile_predicate, // compiler generated predicate moved from frequent branch in a loop failed
101
102 // recorded per method
103 Reason_unloaded, // unloaded class or constant pool entry
104 Reason_uninitialized, // bad class state (uninitialized)
105 Reason_initialized, // class has been fully initialized
106 Reason_unreached, // code is not reached, compiler
107 Reason_unhandled, // arbitrary compiler limitation
108 Reason_constraint, // arbitrary runtime constraint violated
109 Reason_div0_check, // a null_check due to division by zero
110 Reason_age, // nmethod too old; tier threshold reached
111 Reason_predicate, // compiler generated predicate failed
112 Reason_loop_limit_check, // compiler generated loop limits check failed
113 Reason_speculate_class_check, // saw unexpected object class from type speculation
114 Reason_speculate_null_check, // saw unexpected null from type speculation
115 Reason_speculate_null_assert, // saw unexpected null from type speculation
116 Reason_rtm_state_change, // rtm state change detected
117 Reason_unstable_if, // a branch predicted always false was taken
118 Reason_unstable_fused_if, // fused two ifs that had each one untaken branch. One is now taken.
119 Reason_receiver_constraint, // receiver subtype check failed
120 #if INCLUDE_JVMCI
121 Reason_aliasing, // optimistic assumption about aliasing failed
122 Reason_transfer_to_interpreter, // explicit transferToInterpreter()
123 Reason_not_compiled_exception_handler,
124 Reason_unresolved,
125 Reason_jsr_mismatch,
126 #endif
127
128 // Used to define MethodData::_trap_hist_limit where Reason_tenured isn't included
129 Reason_TRAP_HISTORY_LENGTH,
130
131 // Reason_tenured is counted separately, add normal counted Reasons above.
132 Reason_tenured = Reason_TRAP_HISTORY_LENGTH, // age of the code has reached the limit
133 Reason_LIMIT,
134
135 // Note: Reason_RECORDED_LIMIT should fit into 31 bits of
136 // DataLayout::trap_bits. This dependency is enforced indirectly
137 // via asserts, to avoid excessive direct header-to-header dependencies.
138 // See Deoptimization::trap_state_reason and class DataLayout.
139 Reason_RECORDED_LIMIT = Reason_profile_predicate, // some are not recorded per bc
140 };
141
142 // What action must be taken by the runtime?
143 // Note: Keep this enum in sync. with Deoptimization::_trap_action_name.
144 enum DeoptAction {
145 Action_none, // just interpret, do not invalidate nmethod
146 Action_maybe_recompile, // recompile the nmethod; need not invalidate
147 Action_reinterpret, // invalidate the nmethod, reset IC, maybe recompile
148 Action_make_not_entrant, // invalidate the nmethod, recompile (probably)
149 Action_make_not_compilable, // invalidate the nmethod and do not compile
150 Action_LIMIT
151 };
152
153 enum {
154 _action_bits = 3,
155 _reason_bits = 5,
156 _debug_id_bits = 23,
157 _action_shift = 0,
158 _reason_shift = _action_shift+_action_bits,
159 _debug_id_shift = _reason_shift+_reason_bits,
160 BC_CASE_LIMIT = PRODUCT_ONLY(1) NOT_PRODUCT(4) // for _deoptimization_hist
161 };
162
163 enum UnpackType {
164 Unpack_deopt = 0, // normal deoptimization, use pc computed in unpack_vframe_on_stack
165 Unpack_exception = 1, // exception is pending
166 Unpack_uncommon_trap = 2, // redo last byte code (C2 only)
167 Unpack_reexecute = 3, // reexecute bytecode (C1 only)
168 Unpack_none = 4, // not deoptimizing the frame, just reallocating/relocking for JVMTI
169 Unpack_LIMIT = 5
170 };
171
172 #if INCLUDE_JVMCI
173 // Can reconstruct virtualized unsafe large accesses to byte arrays.
174 static const int _support_large_access_byte_array_virtualization = 1;
175 #endif
176
177 // Make all nmethods that are marked_for_deoptimization not_entrant and deoptimize any live
178 // activations using those nmethods. Scan of the code cache is done to
179 // find all marked nmethods and they are made not_entrant.
180 static void deoptimize_all_marked();
181
182 public:
183 // Deoptimizes a frame lazily. Deopt happens on return to the frame.
184 static void deoptimize(JavaThread* thread, frame fr, DeoptReason reason = Reason_constraint);
185
186 #if INCLUDE_JVMCI
187 static address deoptimize_for_missing_exception_handler(CompiledMethod* cm);
188 static oop get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS);
189 #endif
190
191 private:
192 // Does the actual work for deoptimizing a single frame
193 static void deoptimize_single_frame(JavaThread* thread, frame fr, DeoptReason reason);
194
195 #if COMPILER2_OR_JVMCI
196 // Deoptimize objects, that is reallocate and relock them, just before they
197 // escape through JVMTI. The given vframes cover one physical frame.
198 static bool deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
199 bool& realloc_failures);
200
201 public:
202
203 // Support for restoring non-escaping objects
204 static bool realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS);
205 static void reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type);
206 static void reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj);
207 static void reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal);
208 static bool relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
209 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures);
210 static void pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array);
211 #endif // COMPILER2_OR_JVMCI
212
213 public:
214 static vframeArray* create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures);
215
216 // Interface used for unpacking deoptimized frames
217
218 // UnrollBlock is returned by fetch_unroll_info() to the deoptimization handler (blob).
219 // This is only a CheapObj to ease debugging after a deopt failure
220 class UnrollBlock : public CHeapObj<mtCompiler> {
221 friend class VMStructs;
222 friend class JVMCIVMStructs;
223 private:
224 int _size_of_deoptimized_frame; // Size, in bytes, of current deoptimized frame
225 int _caller_adjustment; // Adjustment, in bytes, to caller's SP by initial interpreted frame
226 int _number_of_frames; // Number frames to unroll
227 int _total_frame_sizes; // Total of number*sizes frames
228 intptr_t* _frame_sizes; // Array of frame sizes, in bytes, for unrolling the stack
229 address* _frame_pcs; // Array of frame pc's, in bytes, for unrolling the stack
230 intptr_t* _register_block; // Block for storing callee-saved registers.
231 BasicType _return_type; // Tells if we have to restore double or long return value
232 intptr_t _initial_info; // Platform dependent data for the sender frame (was FP on x86)
233 int _caller_actual_parameters; // The number of actual arguments at the
234 // interpreted caller of the deoptimized frame
235 int _unpack_kind; // exec_mode that can be changed during fetch_unroll_info
236
237 // The following fields are used as temps during the unpacking phase
238 // (which is tight on registers, especially on x86). They really ought
239 // to be PD variables but that involves moving this class into its own
240 // file to use the pd include mechanism. Maybe in a later cleanup ...
241 intptr_t _counter_temp; // SHOULD BE PD VARIABLE (x86 frame count temp)
242 intptr_t _sender_sp_temp; // SHOULD BE PD VARIABLE (x86 sender_sp)
243 public:
244 // Constructor
245 UnrollBlock(int size_of_deoptimized_frame,
246 int caller_adjustment,
247 int caller_actual_parameters,
248 int number_of_frames,
249 intptr_t* frame_sizes,
250 address* frames_pcs,
251 BasicType return_type,
252 int unpack_kind);
253 ~UnrollBlock();
254
255 // Accessors
256 intptr_t* frame_sizes() const { return _frame_sizes; }
257 int number_of_frames() const { return _number_of_frames; }
258
259 // Returns the total size of frames
260 int size_of_frames() const;
261
262 void set_initial_info(intptr_t info) { _initial_info = info; }
263
264 int caller_actual_parameters() const { return _caller_actual_parameters; }
265
266 // Accessors used by the code generator for the unpack stub.
267 static ByteSize size_of_deoptimized_frame_offset() { return byte_offset_of(UnrollBlock, _size_of_deoptimized_frame); }
268 static ByteSize caller_adjustment_offset() { return byte_offset_of(UnrollBlock, _caller_adjustment); }
269 static ByteSize number_of_frames_offset() { return byte_offset_of(UnrollBlock, _number_of_frames); }
270 static ByteSize frame_sizes_offset() { return byte_offset_of(UnrollBlock, _frame_sizes); }
271 static ByteSize total_frame_sizes_offset() { return byte_offset_of(UnrollBlock, _total_frame_sizes); }
272 static ByteSize frame_pcs_offset() { return byte_offset_of(UnrollBlock, _frame_pcs); }
273 static ByteSize counter_temp_offset() { return byte_offset_of(UnrollBlock, _counter_temp); }
274 static ByteSize initial_info_offset() { return byte_offset_of(UnrollBlock, _initial_info); }
275 static ByteSize unpack_kind_offset() { return byte_offset_of(UnrollBlock, _unpack_kind); }
276 static ByteSize sender_sp_temp_offset() { return byte_offset_of(UnrollBlock, _sender_sp_temp); }
277
278 BasicType return_type() const { return _return_type; }
279 void print();
280 };
281
282 //** Returns an UnrollBlock continuing information
283 // how to make room for the resulting interpreter frames.
284 // Called by assembly stub after execution has returned to
285 // deoptimized frame.
286 // @argument thread. Thread where stub_frame resides.
287 // @see OptoRuntime::deoptimization_fetch_unroll_info_C
288 static UnrollBlock* fetch_unroll_info(JavaThread* current, int exec_mode);
289
290 //** Unpacks vframeArray onto execution stack
291 // Called by assembly stub after execution has returned to
292 // deoptimized frame and after the stack unrolling.
293 // @argument thread. Thread where stub_frame resides.
294 // @argument exec_mode. Determines how execution should be continued in top frame.
295 // 0 means continue after current byte code
296 // 1 means exception has happened, handle exception
297 // 2 means reexecute current bytecode (for uncommon traps).
298 // @see OptoRuntime::deoptimization_unpack_frames_C
299 // Return BasicType of call return type, if any
300 static BasicType unpack_frames(JavaThread* thread, int exec_mode);
301
302 // Cleans up deoptimization bits on thread after unpacking or in the
303 // case of an exception.
304 static void cleanup_deopt_info(JavaThread *thread,
305 vframeArray * array);
306
307 // Restores callee saved values from deoptimized frame into oldest interpreter frame
308 // so caller of the deoptimized frame will get back the values it expects.
309 static void unwind_callee_save_values(frame* f, vframeArray* vframe_array);
310
311 //** Performs an uncommon trap for compiled code.
312 // The top most compiler frame is converted into interpreter frames
313 static UnrollBlock* uncommon_trap(JavaThread* current, jint unloaded_class_index, jint exec_mode);
314 // Helper routine that enters the VM and may block
315 static void uncommon_trap_inner(JavaThread* current, jint unloaded_class_index);
316
317 //** Deoptimizes the frame identified by id.
318 // Only called from VMDeoptimizeFrame
319 // @argument thread. Thread where stub_frame resides.
320 // @argument id. id of frame that should be deoptimized.
321 static void deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason);
322
323 // if thread is not the current thread then execute
324 // VM_DeoptimizeFrame otherwise deoptimize directly.
325 static void deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason);
326 static void deoptimize_frame(JavaThread* thread, intptr_t* id);
327
328 // Statistics
329 static void gather_statistics(DeoptReason reason, DeoptAction action,
330 Bytecodes::Code bc = Bytecodes::_illegal);
331 static void print_statistics();
332
333 // How much room to adjust the last frame's SP by, to make space for
334 // the callee's interpreter frame (which expects locals to be next to
335 // incoming arguments)
336 static int last_frame_adjust(int callee_parameters, int callee_locals);
337
338 // trap_request codes
339 static DeoptReason trap_request_reason(int trap_request) {
340 if (trap_request < 0)
341 return (DeoptReason)
342 ((~(trap_request) >> _reason_shift) & right_n_bits(_reason_bits));
343 else
344 // standard reason for unloaded CP entry
345 return Reason_unloaded;
346 }
347 static DeoptAction trap_request_action(int trap_request) {
348 if (trap_request < 0)
349 return (DeoptAction)
350 ((~(trap_request) >> _action_shift) & right_n_bits(_action_bits));
351 else
352 // standard action for unloaded CP entry
353 return _unloaded_action;
354 }
355 static int trap_request_debug_id(int trap_request) {
356 if (trap_request < 0) {
357 return ((~(trap_request) >> _debug_id_shift) & right_n_bits(_debug_id_bits));
358 } else {
359 // standard action for unloaded CP entry
360 return 0;
361 }
362 }
363 static int trap_request_index(int trap_request) {
364 if (trap_request < 0)
365 return -1;
366 else
367 return trap_request;
368 }
369 static int make_trap_request(DeoptReason reason, DeoptAction action,
370 int index = -1) {
371 assert((1 << _reason_bits) >= Reason_LIMIT, "enough bits");
372 assert((1 << _action_bits) >= Action_LIMIT, "enough bits");
373 int trap_request;
374 if (index != -1)
375 trap_request = index;
376 else
377 trap_request = (~(((reason) << _reason_shift)
378 + ((action) << _action_shift)));
379 assert(reason == trap_request_reason(trap_request), "valid reason");
380 assert(action == trap_request_action(trap_request), "valid action");
381 assert(index == trap_request_index(trap_request), "valid index");
382 return trap_request;
383 }
384
385 // The trap_state stored in a MDO is decoded here.
386 // It records two items of information.
387 // reason: If a deoptimization happened here, what its reason was,
388 // or if there were multiple deopts with differing reasons.
389 // recompiled: If a deoptimization here triggered a recompilation.
390 // Note that not all reasons are recorded per-bci.
391 static DeoptReason trap_state_reason(int trap_state);
392 static int trap_state_has_reason(int trap_state, int reason);
393 static int trap_state_add_reason(int trap_state, int reason);
394 static bool trap_state_is_recompiled(int trap_state);
395 static int trap_state_set_recompiled(int trap_state, bool z);
396 static const char* format_trap_state(char* buf, size_t buflen,
397 int trap_state);
398
399 static bool reason_is_recorded_per_bytecode(DeoptReason reason) {
400 return reason > Reason_none && reason <= Reason_RECORDED_LIMIT;
401 }
402
403 static DeoptReason reason_recorded_per_bytecode_if_any(DeoptReason reason) {
404 if (reason_is_recorded_per_bytecode(reason))
405 return reason;
406 else if (reason == Reason_div0_check) // null check due to divide-by-zero?
407 return Reason_null_check; // recorded per BCI as a null check
408 else if (reason == Reason_speculate_class_check)
409 return Reason_class_check;
410 else if (reason == Reason_speculate_null_check)
411 return Reason_null_check;
412 else if (reason == Reason_speculate_null_assert)
413 return Reason_null_assert;
414 else if (reason == Reason_unstable_if)
415 return Reason_intrinsic;
416 else if (reason == Reason_unstable_fused_if)
417 return Reason_range_check;
418 else
419 return Reason_none;
420 }
421
422 static bool reason_is_speculate(int reason) {
423 if (reason == Reason_speculate_class_check ||
424 reason == Reason_speculate_null_check ||
425 reason == Reason_speculate_null_assert) {
426 return true;
427 }
428 return false;
429 }
430
431 static DeoptReason reason_null_check(bool speculative) {
432 return speculative ? Deoptimization::Reason_speculate_null_check : Deoptimization::Reason_null_check;
433 }
434
435 static DeoptReason reason_class_check(bool speculative) {
436 return speculative ? Deoptimization::Reason_speculate_class_check : Deoptimization::Reason_class_check;
437 }
438
439 static DeoptReason reason_null_assert(bool speculative) {
440 return speculative ? Deoptimization::Reason_speculate_null_assert : Deoptimization::Reason_null_assert;
441 }
442
443 static uint per_method_trap_limit(int reason) {
444 return reason_is_speculate(reason) ? (uint)PerMethodSpecTrapLimit : (uint)PerMethodTrapLimit;
445 }
446
447 static const char* trap_reason_name(int reason);
448 static const char* trap_action_name(int action);
449 // Format like reason='foo' action='bar' index='123'.
450 // This is suitable both for XML and for tty output.
451 static const char* format_trap_request(char* buf, size_t buflen,
452 int trap_request);
453
454 static jint total_deoptimization_count();
455 static jint deoptimization_count(const char* reason_str, const char* action_str);
456
457 // JVMTI PopFrame support
458
459 // Preserves incoming arguments to the popped frame when it is
460 // returning to a deoptimized caller
461 static void popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address);
462
463 static MethodData* get_method_data(JavaThread* thread, const methodHandle& m, bool create_if_missing);
464 private:
465 // Update the mdo's count and per-BCI reason bits, returning previous state:
466 static ProfileData* query_update_method_data(MethodData* trap_mdo,
467 int trap_bci,
468 DeoptReason reason,
469 bool update_total_trap_count,
470 #if INCLUDE_JVMCI
471 bool is_osr,
472 #endif
473 Method* compiled_method,
474 //outputs:
475 uint& ret_this_trap_count,
476 bool& ret_maybe_prior_trap,
477 bool& ret_maybe_prior_recompile);
478 // class loading support for uncommon trap
479 static void load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS);
480
481 static UnrollBlock* fetch_unroll_info_helper(JavaThread* current, int exec_mode);
482
483 static DeoptAction _unloaded_action; // == Action_reinterpret;
484 static const char* _trap_reason_name[];
485 static const char* _trap_action_name[];
486
487 static juint _deoptimization_hist[Reason_LIMIT][1+Action_LIMIT][BC_CASE_LIMIT];
488 // Note: Histogram array size is 1-2 Kb.
489
490 public:
491 static void update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason);
492 };
493
494 #endif // SHARE_RUNTIME_DEOPTIMIZATION_HPP