1 /*
2 * Copyright (c) 1997, 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_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(nmethod* nm, bool inc_recompile_counts = true);
61 // Record this as a dependent method.
62 void dependent(nmethod* nm);
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 Reason_auto_vectorization_check, // compiler generated (speculative) auto vectorization checks failed
102
103 // recorded per method
104 Reason_unloaded, // unloaded class or constant pool entry
105 Reason_uninitialized, // bad class state (uninitialized)
106 Reason_initialized, // class has been fully initialized
107 Reason_unreached, // code is not reached, compiler
108 Reason_unhandled, // arbitrary compiler limitation
109 Reason_constraint, // arbitrary runtime constraint violated
110 Reason_div0_check, // a null_check due to division by zero
111 Reason_age, // nmethod too old; tier threshold reached
112 Reason_predicate, // compiler generated predicate failed
113 Reason_loop_limit_check, // compiler generated loop limits check failed
114 Reason_speculate_class_check, // saw unexpected object class from type speculation
115 Reason_speculate_null_check, // saw unexpected null from type speculation
116 Reason_speculate_null_assert, // saw unexpected null from type speculation
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(nmethod* nm);
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 bool realloc_inline_type_result(InlineKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS);
206 static void reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type);
207 static void reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj);
208 static void reassign_flat_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, flatArrayOop obj, FlatArrayKlass* vak, bool skip_internal, TRAPS);
209 static void reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS);
210 static bool relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
211 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures);
212 static void pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array);
213 #endif // COMPILER2_OR_JVMCI
214
215 public:
216 static vframeArray* create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures);
217
218 // Interface used for unpacking deoptimized frames
219
220 // UnrollBlock is returned by fetch_unroll_info() to the deoptimization handler (blob).
221 // This is only a CheapObj to ease debugging after a deopt failure
222 class UnrollBlock : public CHeapObj<mtCompiler> {
223 friend class VMStructs;
224 friend class JVMCIVMStructs;
225 private:
226 int _size_of_deoptimized_frame; // Size, in bytes, of current deoptimized frame
227 int _caller_adjustment; // Adjustment, in bytes, to caller's SP by initial interpreted frame
228 int _number_of_frames; // Number frames to unroll
229 int _total_frame_sizes; // Total of number*sizes frames
230 intptr_t* _frame_sizes; // Array of frame sizes, in bytes, for unrolling the stack
231 address* _frame_pcs; // Array of frame pc's, in bytes, for unrolling the stack
232 intptr_t* _register_block; // Block for storing callee-saved registers.
233 BasicType _return_type; // Tells if we have to restore double or long return value
234 intptr_t _initial_info; // Platform dependent data for the sender frame (was FP on x86)
235 int _caller_actual_parameters; // The number of actual arguments at the
236 // interpreted caller of the deoptimized frame
237 int _unpack_kind; // exec_mode that can be changed during fetch_unroll_info
238
239 // The following fields are used as temps during the unpacking phase
240 // (which is tight on registers, especially on x86). They really ought
241 // to be PD variables but that involves moving this class into its own
242 // file to use the pd include mechanism. Maybe in a later cleanup ...
243 intptr_t _counter_temp; // SHOULD BE PD VARIABLE (x86 frame count temp)
244 intptr_t _sender_sp_temp; // SHOULD BE PD VARIABLE (x86 sender_sp)
245 public:
246 // Constructor
247 UnrollBlock(int size_of_deoptimized_frame,
248 int caller_adjustment,
249 int caller_actual_parameters,
250 int number_of_frames,
251 intptr_t* frame_sizes,
252 address* frames_pcs,
253 BasicType return_type,
254 int unpack_kind);
255 ~UnrollBlock();
256
257 // Accessors
258 intptr_t* frame_sizes() const { return _frame_sizes; }
259 int number_of_frames() const { return _number_of_frames; }
260
261 // Returns the total size of frames
262 int size_of_frames() const;
263
264 void set_initial_info(intptr_t info) { _initial_info = info; }
265
266 int caller_actual_parameters() const { return _caller_actual_parameters; }
267
268 // Accessors used by the code generator for the unpack stub.
269 static ByteSize size_of_deoptimized_frame_offset() { return byte_offset_of(UnrollBlock, _size_of_deoptimized_frame); }
270 static ByteSize caller_adjustment_offset() { return byte_offset_of(UnrollBlock, _caller_adjustment); }
271 static ByteSize number_of_frames_offset() { return byte_offset_of(UnrollBlock, _number_of_frames); }
272 static ByteSize frame_sizes_offset() { return byte_offset_of(UnrollBlock, _frame_sizes); }
273 static ByteSize total_frame_sizes_offset() { return byte_offset_of(UnrollBlock, _total_frame_sizes); }
274 static ByteSize frame_pcs_offset() { return byte_offset_of(UnrollBlock, _frame_pcs); }
275 static ByteSize counter_temp_offset() { return byte_offset_of(UnrollBlock, _counter_temp); }
276 static ByteSize initial_info_offset() { return byte_offset_of(UnrollBlock, _initial_info); }
277 static ByteSize unpack_kind_offset() { return byte_offset_of(UnrollBlock, _unpack_kind); }
278 static ByteSize sender_sp_temp_offset() { return byte_offset_of(UnrollBlock, _sender_sp_temp); }
279
280 BasicType return_type() const { return _return_type; }
281 void print();
282 };
283
284 //** Returns an UnrollBlock continuing information
285 // how to make room for the resulting interpreter frames.
286 // Called by assembly stub after execution has returned to
287 // deoptimized frame.
288 // @argument thread. Thread where stub_frame resides.
289 // @see OptoRuntime::deoptimization_fetch_unroll_info_C
290 static UnrollBlock* fetch_unroll_info(JavaThread* current, int exec_mode);
291
292 //** Unpacks vframeArray onto execution stack
293 // Called by assembly stub after execution has returned to
294 // deoptimized frame and after the stack unrolling.
295 // @argument thread. Thread where stub_frame resides.
296 // @argument exec_mode. Determines how execution should be continued in top frame.
297 // 0 means continue after current byte code
298 // 1 means exception has happened, handle exception
299 // 2 means reexecute current bytecode (for uncommon traps).
300 // @see OptoRuntime::deoptimization_unpack_frames_C
301 // Return BasicType of call return type, if any
302 static BasicType unpack_frames(JavaThread* thread, int exec_mode);
303
304 // Cleans up deoptimization bits on thread after unpacking or in the
305 // case of an exception.
306 static void cleanup_deopt_info(JavaThread *thread,
307 vframeArray * array);
308
309 // Restores callee saved values from deoptimized frame into oldest interpreter frame
310 // so caller of the deoptimized frame will get back the values it expects.
311 static void unwind_callee_save_values(frame* f, vframeArray* vframe_array);
312
313 //** Performs an uncommon trap for compiled code.
314 // The top most compiler frame is converted into interpreter frames
315 static UnrollBlock* uncommon_trap(JavaThread* current, jint unloaded_class_index, jint exec_mode);
316 // Helper routine that enters the VM and may block
317 static void uncommon_trap_inner(JavaThread* current, jint unloaded_class_index);
318
319 //** Deoptimizes the frame identified by id.
320 // Only called from VMDeoptimizeFrame
321 // @argument thread. Thread where stub_frame resides.
322 // @argument id. id of frame that should be deoptimized.
323 static void deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason);
324
325 // if thread is not the current thread then execute
326 // VM_DeoptimizeFrame otherwise deoptimize directly.
327 static void deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason);
328 static void deoptimize_frame(JavaThread* thread, intptr_t* id);
329
330 // Statistics
331 static void gather_statistics(DeoptReason reason, DeoptAction action,
332 Bytecodes::Code bc = Bytecodes::_illegal);
333 static void print_statistics();
334
335 // How much room to adjust the last frame's SP by, to make space for
336 // the callee's interpreter frame (which expects locals to be next to
337 // incoming arguments)
338 static int last_frame_adjust(int callee_parameters, int callee_locals);
339
340 // trap_request codes
341 static DeoptReason trap_request_reason(int trap_request) {
342 if (trap_request < 0)
343 return (DeoptReason)
344 ((~(trap_request) >> _reason_shift) & right_n_bits(_reason_bits));
345 else
346 // standard reason for unloaded CP entry
347 return Reason_unloaded;
348 }
349 static DeoptAction trap_request_action(int trap_request) {
350 if (trap_request < 0)
351 return (DeoptAction)
352 ((~(trap_request) >> _action_shift) & right_n_bits(_action_bits));
353 else
354 // standard action for unloaded CP entry
355 return _unloaded_action;
356 }
357 static int trap_request_debug_id(int trap_request) {
358 if (trap_request < 0) {
359 return ((~(trap_request) >> _debug_id_shift) & right_n_bits(_debug_id_bits));
360 } else {
361 // standard action for unloaded CP entry
362 return 0;
363 }
364 }
365 static int trap_request_index(int trap_request) {
366 if (trap_request < 0)
367 return -1;
368 else
369 return trap_request;
370 }
371 static int make_trap_request(DeoptReason reason, DeoptAction action,
372 int index = -1) {
373 assert((1 << _reason_bits) >= Reason_LIMIT, "enough bits");
374 assert((1 << _action_bits) >= Action_LIMIT, "enough bits");
375 int trap_request;
376 if (index != -1)
377 trap_request = index;
378 else
379 trap_request = (~(((reason) << _reason_shift)
380 + ((action) << _action_shift)));
381 assert(reason == trap_request_reason(trap_request), "valid reason");
382 assert(action == trap_request_action(trap_request), "valid action");
383 assert(index == trap_request_index(trap_request), "valid index");
384 return trap_request;
385 }
386
387 // The trap_state stored in a MDO is decoded here.
388 // It records two items of information.
389 // reason: If a deoptimization happened here, what its reason was,
390 // or if there were multiple deopts with differing reasons.
391 // recompiled: If a deoptimization here triggered a recompilation.
392 // Note that not all reasons are recorded per-bci.
393 static DeoptReason trap_state_reason(int trap_state);
394 static int trap_state_has_reason(int trap_state, int reason);
395 static int trap_state_add_reason(int trap_state, int reason);
396 static bool trap_state_is_recompiled(int trap_state);
397 static int trap_state_set_recompiled(int trap_state, bool z);
398 static const char* format_trap_state(char* buf, size_t buflen,
399 int trap_state);
400
401 static bool reason_is_recorded_per_bytecode(DeoptReason reason) {
402 return reason > Reason_none && reason <= Reason_RECORDED_LIMIT;
403 }
404
405 static DeoptReason reason_recorded_per_bytecode_if_any(DeoptReason reason) {
406 if (reason_is_recorded_per_bytecode(reason))
407 return reason;
408 else if (reason == Reason_div0_check) // null check due to divide-by-zero?
409 return Reason_null_check; // recorded per BCI as a null check
410 else if (reason == Reason_speculate_class_check)
411 return Reason_class_check;
412 else if (reason == Reason_speculate_null_check)
413 return Reason_null_check;
414 else if (reason == Reason_speculate_null_assert)
415 return Reason_null_assert;
416 else if (reason == Reason_unstable_if)
417 return Reason_intrinsic;
418 else if (reason == Reason_unstable_fused_if)
419 return Reason_range_check;
420 else
421 return Reason_none;
422 }
423
424 static bool reason_is_speculate(int reason) {
425 if (reason == Reason_speculate_class_check ||
426 reason == Reason_speculate_null_check ||
427 reason == Reason_speculate_null_assert) {
428 return true;
429 }
430 return false;
431 }
432
433 static DeoptReason reason_null_check(bool speculative) {
434 return speculative ? Deoptimization::Reason_speculate_null_check : Deoptimization::Reason_null_check;
435 }
436
437 static DeoptReason reason_class_check(bool speculative) {
438 return speculative ? Deoptimization::Reason_speculate_class_check : Deoptimization::Reason_class_check;
439 }
440
441 static DeoptReason reason_null_assert(bool speculative) {
442 return speculative ? Deoptimization::Reason_speculate_null_assert : Deoptimization::Reason_null_assert;
443 }
444
445 static uint per_method_trap_limit(int reason) {
446 return reason_is_speculate(reason) ? (uint)PerMethodSpecTrapLimit : (uint)PerMethodTrapLimit;
447 }
448
449 static const char* trap_reason_name(int reason);
450 static const char* trap_action_name(int action);
451 // Format like reason='foo' action='bar' index='123'.
452 // This is suitable both for XML and for tty output.
453 static const char* format_trap_request(char* buf, size_t buflen,
454 int trap_request);
455
456 static jint total_deoptimization_count();
457 static jint deoptimization_count(const char* reason_str, const char* action_str);
458
459 // JVMTI PopFrame support
460
461 // Preserves incoming arguments to the popped frame when it is
462 // returning to a deoptimized caller
463 static void popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address);
464
465 static MethodData* get_method_data(JavaThread* thread, const methodHandle& m, bool create_if_missing);
466 private:
467 // Update the mdo's count and per-BCI reason bits, returning previous state:
468 static ProfileData* query_update_method_data(MethodData* trap_mdo,
469 int trap_bci,
470 DeoptReason reason,
471 bool update_total_trap_count,
472 #if INCLUDE_JVMCI
473 bool is_osr,
474 #endif
475 Method* compiled_method,
476 //outputs:
477 uint& ret_this_trap_count,
478 bool& ret_maybe_prior_trap,
479 bool& ret_maybe_prior_recompile);
480 // class loading support for uncommon trap
481 static void load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS);
482
483 static UnrollBlock* fetch_unroll_info_helper(JavaThread* current, int exec_mode);
484
485 static DeoptAction _unloaded_action; // == Action_reinterpret;
486 static const char* _trap_reason_name[];
487 static const char* _trap_action_name[];
488
489 static juint _deoptimization_hist[Reason_LIMIT][1+Action_LIMIT][BC_CASE_LIMIT];
490 // Note: Histogram array size is 1-2 Kb.
491
492 public:
493 static void update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason);
494 };
495
496 #endif // SHARE_RUNTIME_DEOPTIMIZATION_HPP