1 /*
2 * Copyright (c) 1997, 2021, 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 #include "precompiled.hpp"
26 #include "jvm.h"
27 #include "classfile/javaClasses.inline.hpp"
28 #include "classfile/symbolTable.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "classfile/vmClasses.hpp"
31 #include "code/codeCache.hpp"
32 #include "code/debugInfoRec.hpp"
33 #include "code/nmethod.hpp"
34 #include "code/pcDesc.hpp"
35 #include "code/scopeDesc.hpp"
36 #include "compiler/compilationPolicy.hpp"
37 #include "gc/shared/collectedHeap.hpp"
38 #include "interpreter/bytecode.hpp"
39 #include "interpreter/interpreter.hpp"
40 #include "interpreter/oopMapCache.hpp"
41 #include "memory/allocation.inline.hpp"
42 #include "memory/oopFactory.hpp"
43 #include "memory/resourceArea.hpp"
44 #include "memory/universe.hpp"
45 #include "oops/constantPool.hpp"
46 #include "oops/method.hpp"
47 #include "oops/objArrayKlass.hpp"
48 #include "oops/objArrayOop.inline.hpp"
49 #include "oops/oop.inline.hpp"
50 #include "oops/fieldStreams.inline.hpp"
51 #include "oops/typeArrayOop.inline.hpp"
52 #include "oops/verifyOopClosure.hpp"
53 #include "prims/jvmtiDeferredUpdates.hpp"
54 #include "prims/jvmtiExport.hpp"
55 #include "prims/jvmtiThreadState.hpp"
56 #include "prims/vectorSupport.hpp"
57 #include "prims/methodHandles.hpp"
58 #include "runtime/atomic.hpp"
59 #include "runtime/biasedLocking.hpp"
60 #include "runtime/deoptimization.hpp"
61 #include "runtime/escapeBarrier.hpp"
62 #include "runtime/fieldDescriptor.hpp"
63 #include "runtime/fieldDescriptor.inline.hpp"
64 #include "runtime/frame.inline.hpp"
65 #include "runtime/handles.inline.hpp"
66 #include "runtime/interfaceSupport.inline.hpp"
67 #include "runtime/jniHandles.inline.hpp"
68 #include "runtime/keepStackGCProcessed.hpp"
69 #include "runtime/objectMonitor.inline.hpp"
70 #include "runtime/osThread.hpp"
71 #include "runtime/safepointVerifiers.hpp"
72 #include "runtime/sharedRuntime.hpp"
73 #include "runtime/signature.hpp"
74 #include "runtime/stackFrameStream.inline.hpp"
75 #include "runtime/stackWatermarkSet.hpp"
76 #include "runtime/stubRoutines.hpp"
77 #include "runtime/synchronizer.hpp"
78 #include "runtime/thread.hpp"
79 #include "runtime/threadSMR.hpp"
80 #include "runtime/threadWXSetters.inline.hpp"
81 #include "runtime/vframe.hpp"
82 #include "runtime/vframeArray.hpp"
83 #include "runtime/vframe_hp.hpp"
84 #include "runtime/vmOperations.hpp"
85 #include "utilities/events.hpp"
86 #include "utilities/macros.hpp"
87 #include "utilities/preserveException.hpp"
88 #include "utilities/xmlstream.hpp"
89 #if INCLUDE_JFR
90 #include "jfr/jfrEvents.hpp"
91 #include "jfr/metadata/jfrSerializer.hpp"
92 #endif
93
94 bool DeoptimizationMarker::_is_active = false;
95
96 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
97 int caller_adjustment,
98 int caller_actual_parameters,
99 int number_of_frames,
100 intptr_t* frame_sizes,
101 address* frame_pcs,
102 BasicType return_type,
103 int exec_mode) {
104 _size_of_deoptimized_frame = size_of_deoptimized_frame;
105 _caller_adjustment = caller_adjustment;
106 _caller_actual_parameters = caller_actual_parameters;
107 _number_of_frames = number_of_frames;
108 _frame_sizes = frame_sizes;
109 _frame_pcs = frame_pcs;
110 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
111 _return_type = return_type;
112 _initial_info = 0;
113 // PD (x86 only)
114 _counter_temp = 0;
115 _unpack_kind = exec_mode;
116 _sender_sp_temp = 0;
117
118 _total_frame_sizes = size_of_frames();
119 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
120 }
121
122
123 Deoptimization::UnrollBlock::~UnrollBlock() {
124 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
125 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
126 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
127 }
128
129
130 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
131 assert(register_number < RegisterMap::reg_count, "checking register number");
132 return &_register_block[register_number * 2];
133 }
134
135
136
137 int Deoptimization::UnrollBlock::size_of_frames() const {
138 // Acount first for the adjustment of the initial frame
139 int result = _caller_adjustment;
140 for (int index = 0; index < number_of_frames(); index++) {
141 result += frame_sizes()[index];
142 }
143 return result;
144 }
145
146
147 void Deoptimization::UnrollBlock::print() {
148 ResourceMark rm;
149 stringStream st;
150 st.print_cr("UnrollBlock");
151 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
152 st.print( " frame_sizes: ");
153 for (int index = 0; index < number_of_frames(); index++) {
154 st.print(INTX_FORMAT " ", frame_sizes()[index]);
155 }
156 st.cr();
157 tty->print_raw(st.as_string());
158 }
159
160
161 // In order to make fetch_unroll_info work properly with escape
162 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY.
163 // The actual reallocation of previously eliminated objects occurs in realloc_objects,
164 // which is called from the method fetch_unroll_info_helper below.
165 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode))
166 // fetch_unroll_info() is called at the beginning of the deoptimization
167 // handler. Note this fact before we start generating temporary frames
168 // that can confuse an asynchronous stack walker. This counter is
169 // decremented at the end of unpack_frames().
170 if (TraceDeoptimization) {
171 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(current));
172 }
173 current->inc_in_deopt_handler();
174
175 if (exec_mode == Unpack_exception) {
176 // When we get here, a callee has thrown an exception into a deoptimized
177 // frame. That throw might have deferred stack watermark checking until
178 // after unwinding. So we deal with such deferred requests here.
179 StackWatermarkSet::after_unwind(current);
180 }
181
182 return fetch_unroll_info_helper(current, exec_mode);
183 JRT_END
184
185 #if COMPILER2_OR_JVMCI
186 #ifndef PRODUCT
187 // print information about reallocated objects
188 static void print_objects(JavaThread* deoptee_thread,
189 GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
190 ResourceMark rm;
191 stringStream st; // change to logStream with logging
192 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
193 fieldDescriptor fd;
194
195 for (int i = 0; i < objects->length(); i++) {
196 ObjectValue* sv = (ObjectValue*) objects->at(i);
197 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
198 Handle obj = sv->value();
199
200 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
201 k->print_value_on(&st);
202 assert(obj.not_null() || realloc_failures, "reallocation was missed");
203 if (obj.is_null()) {
204 st.print(" allocation failed");
205 } else {
206 st.print(" allocated (%d bytes)", obj->size() * HeapWordSize);
207 }
208 st.cr();
209
210 if (Verbose && !obj.is_null()) {
211 k->oop_print_on(obj(), &st);
212 }
213 }
214 tty->print_raw(st.as_string());
215 }
216 #endif
217
218 static bool rematerialize_objects(JavaThread* thread, int exec_mode, CompiledMethod* compiled_method,
219 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
220 bool& deoptimized_objects) {
221 bool realloc_failures = false;
222 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
223
224 JavaThread* deoptee_thread = chunk->at(0)->thread();
225 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread),
226 "a frame can only be deoptimized by the owner thread");
227
228 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
229
230 // The flag return_oop() indicates call sites which return oop
231 // in compiled code. Such sites include java method calls,
232 // runtime calls (for example, used to allocate new objects/arrays
233 // on slow code path) and any other calls generated in compiled code.
234 // It is not guaranteed that we can get such information here only
235 // by analyzing bytecode in deoptimized frames. This is why this flag
236 // is set during method compilation (see Compile::Process_OopMap_Node()).
237 // If the previous frame was popped or if we are dispatching an exception,
238 // we don't have an oop result.
239 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
240 Handle return_value;
241 if (save_oop_result) {
242 // Reallocation may trigger GC. If deoptimization happened on return from
243 // call which returns oop we need to save it since it is not in oopmap.
244 oop result = deoptee.saved_oop_result(&map);
245 assert(oopDesc::is_oop_or_null(result), "must be oop");
246 return_value = Handle(thread, result);
247 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
248 if (TraceDeoptimization) {
249 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
250 }
251 }
252 if (objects != NULL) {
253 if (exec_mode == Deoptimization::Unpack_none) {
254 assert(thread->thread_state() == _thread_in_vm, "assumption");
255 JavaThread* THREAD = thread; // For exception macros.
256 // Clear pending OOM if reallocation fails and return true indicating allocation failure
257 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
258 deoptimized_objects = true;
259 } else {
260 JavaThread* current = thread; // For JRT_BLOCK
261 JRT_BLOCK
262 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
263 JRT_END
264 }
265 bool skip_internal = (compiled_method != NULL) && !compiled_method->is_compiled_by_jvmci();
266 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
267 #ifndef PRODUCT
268 if (TraceDeoptimization) {
269 print_objects(deoptee_thread, objects, realloc_failures);
270 }
271 #endif
272 }
273 if (save_oop_result) {
274 // Restore result.
275 deoptee.set_saved_oop_result(&map, return_value());
276 }
277 return realloc_failures;
278 }
279
280 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
281 frame& deoptee, int exec_mode, bool& deoptimized_objects) {
282 JavaThread* deoptee_thread = chunk->at(0)->thread();
283 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
284 assert(thread == Thread::current(), "should be");
285 HandleMark hm(thread);
286 #ifndef PRODUCT
287 bool first = true;
288 #endif
289 for (int i = 0; i < chunk->length(); i++) {
290 compiledVFrame* cvf = chunk->at(i);
291 assert (cvf->scope() != NULL,"expect only compiled java frames");
292 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
293 if (monitors->is_nonempty()) {
294 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee,
295 exec_mode, realloc_failures);
296 deoptimized_objects = deoptimized_objects || relocked;
297 #ifndef PRODUCT
298 if (PrintDeoptimizationDetails) {
299 ResourceMark rm;
300 stringStream st;
301 for (int j = 0; j < monitors->length(); j++) {
302 MonitorInfo* mi = monitors->at(j);
303 if (mi->eliminated()) {
304 if (first) {
305 first = false;
306 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
307 }
308 if (exec_mode == Deoptimization::Unpack_none) {
309 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
310 if (monitor != NULL && monitor->object() == mi->owner()) {
311 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
312 continue;
313 }
314 }
315 if (mi->owner_is_scalar_replaced()) {
316 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
317 st.print_cr(" failed reallocation for klass %s", k->external_name());
318 } else {
319 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
320 }
321 }
322 }
323 tty->print_raw(st.as_string());
324 }
325 #endif // !PRODUCT
326 }
327 }
328 }
329
330 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
331 // The given vframes cover one physical frame.
332 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
333 bool& realloc_failures) {
334 frame deoptee = chunk->at(0)->fr();
335 JavaThread* deoptee_thread = chunk->at(0)->thread();
336 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
337 RegisterMap map(chunk->at(0)->register_map());
338 bool deoptimized_objects = false;
339
340 bool const jvmci_enabled = JVMCI_ONLY(UseJVMCICompiler) NOT_JVMCI(false);
341
342 // Reallocate the non-escaping objects and restore their fields.
343 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations)
344 || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
345 realloc_failures = rematerialize_objects(thread, Unpack_none, cm, deoptee, map, chunk, deoptimized_objects);
346 }
347
348 // Revoke biases of objects with eliminated locks in the given frame.
349 Deoptimization::revoke_for_object_deoptimization(deoptee_thread, deoptee, &map, thread);
350
351 // MonitorInfo structures used in eliminate_locks are not GC safe.
352 NoSafepointVerifier no_safepoint;
353
354 // Now relock objects if synchronization on them was eliminated.
355 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
356 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
357 }
358 return deoptimized_objects;
359 }
360 #endif // COMPILER2_OR_JVMCI
361
362 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
363 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) {
364 // When we get here we are about to unwind the deoptee frame. In order to
365 // catch not yet safe to use frames, the following stack watermark barrier
366 // poll will make such frames safe to use.
367 StackWatermarkSet::before_unwind(current);
368
369 // Note: there is a safepoint safety issue here. No matter whether we enter
370 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
371 // the vframeArray is created.
372 //
373
374 // Allocate our special deoptimization ResourceMark
375 DeoptResourceMark* dmark = new DeoptResourceMark(current);
376 assert(current->deopt_mark() == NULL, "Pending deopt!");
377 current->set_deopt_mark(dmark);
378
379 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect
380 RegisterMap map(current, true);
381 RegisterMap dummy_map(current, false);
382 // Now get the deoptee with a valid map
383 frame deoptee = stub_frame.sender(&map);
384 // Set the deoptee nmethod
385 assert(current->deopt_compiled_method() == NULL, "Pending deopt!");
386 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
387 current->set_deopt_compiled_method(cm);
388
389 if (VerifyStack) {
390 current->validate_frame_layout();
391 }
392
393 // Create a growable array of VFrames where each VFrame represents an inlined
394 // Java frame. This storage is allocated with the usual system arena.
395 assert(deoptee.is_compiled_frame(), "Wrong frame type");
396 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
397 vframe* vf = vframe::new_vframe(&deoptee, &map, current);
398 while (!vf->is_top()) {
399 assert(vf->is_compiled_frame(), "Wrong frame type");
400 chunk->push(compiledVFrame::cast(vf));
401 vf = vf->sender();
402 }
403 assert(vf->is_compiled_frame(), "Wrong frame type");
404 chunk->push(compiledVFrame::cast(vf));
405
406 bool realloc_failures = false;
407
408 #if COMPILER2_OR_JVMCI
409 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
410
411 // Reallocate the non-escaping objects and restore their fields. Then
412 // relock objects if synchronization on them was eliminated.
413 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations)
414 || EliminateAutoBox || EnableVectorAggressiveReboxing )) {
415 bool unused;
416 realloc_failures = rematerialize_objects(current, exec_mode, cm, deoptee, map, chunk, unused);
417 }
418 #endif // COMPILER2_OR_JVMCI
419
420 // Revoke biases, done with in java state.
421 // No safepoints allowed after this
422 revoke_from_deopt_handler(current, deoptee, &map);
423
424 // Ensure that no safepoint is taken after pointers have been stored
425 // in fields of rematerialized objects. If a safepoint occurs from here on
426 // out the java state residing in the vframeArray will be missed.
427 // Locks may be rebaised in a safepoint.
428 NoSafepointVerifier no_safepoint;
429
430 #if COMPILER2_OR_JVMCI
431 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
432 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
433 bool unused;
434 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
435 }
436 #endif // COMPILER2_OR_JVMCI
437
438 ScopeDesc* trap_scope = chunk->at(0)->scope();
439 Handle exceptionObject;
440 if (trap_scope->rethrow_exception()) {
441 if (PrintDeoptimizationDetails) {
442 tty->print_cr("Exception to be rethrown in the interpreter for method %s::%s at bci %d", trap_scope->method()->method_holder()->name()->as_C_string(), trap_scope->method()->name()->as_C_string(), trap_scope->bci());
443 }
444 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
445 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw");
446 ScopeValue* topOfStack = expressions->top();
447 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
448 guarantee(exceptionObject() != NULL, "exception oop can not be null");
449 }
450
451 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
452 #if COMPILER2_OR_JVMCI
453 if (realloc_failures) {
454 pop_frames_failed_reallocs(current, array);
455 }
456 #endif
457
458 assert(current->vframe_array_head() == NULL, "Pending deopt!");
459 current->set_vframe_array_head(array);
460
461 // Now that the vframeArray has been created if we have any deferred local writes
462 // added by jvmti then we can free up that structure as the data is now in the
463 // vframeArray
464
465 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
466
467 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
468 CodeBlob* cb = stub_frame.cb();
469 // Verify we have the right vframeArray
470 assert(cb->frame_size() >= 0, "Unexpected frame size");
471 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
472
473 // If the deopt call site is a MethodHandle invoke call site we have
474 // to adjust the unpack_sp.
475 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
476 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
477 unpack_sp = deoptee.unextended_sp();
478
479 #ifdef ASSERT
480 assert(cb->is_deoptimization_stub() ||
481 cb->is_uncommon_trap_stub() ||
482 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
483 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
484 "unexpected code blob: %s", cb->name());
485 #endif
486
487 // This is a guarantee instead of an assert because if vframe doesn't match
488 // we will unpack the wrong deoptimized frame and wind up in strange places
489 // where it will be very difficult to figure out what went wrong. Better
490 // to die an early death here than some very obscure death later when the
491 // trail is cold.
492 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
493 // in that it will fail to detect a problem when there is one. This needs
494 // more work in tiger timeframe.
495 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
496
497 int number_of_frames = array->frames();
498
499 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
500 // virtual activation, which is the reverse of the elements in the vframes array.
501 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
502 // +1 because we always have an interpreter return address for the final slot.
503 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
504 int popframe_extra_args = 0;
505 // Create an interpreter return address for the stub to use as its return
506 // address so the skeletal frames are perfectly walkable
507 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
508
509 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
510 // activation be put back on the expression stack of the caller for reexecution
511 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) {
512 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words());
513 }
514
515 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
516 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
517 // than simply use array->sender.pc(). This requires us to walk the current set of frames
518 //
519 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
520 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
521
522 // It's possible that the number of parameters at the call site is
523 // different than number of arguments in the callee when method
524 // handles are used. If the caller is interpreted get the real
525 // value so that the proper amount of space can be added to it's
526 // frame.
527 bool caller_was_method_handle = false;
528 if (deopt_sender.is_interpreted_frame()) {
529 methodHandle method(current, deopt_sender.interpreter_frame_method());
530 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
531 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
532 // Method handle invokes may involve fairly arbitrary chains of
533 // calls so it's impossible to know how much actual space the
534 // caller has for locals.
535 caller_was_method_handle = true;
536 }
537 }
538
539 //
540 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
541 // frame_sizes/frame_pcs[1] next oldest frame (int)
542 // frame_sizes/frame_pcs[n] youngest frame (int)
543 //
544 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
545 // owns the space for the return address to it's caller). Confusing ain't it.
546 //
547 // The vframe array can address vframes with indices running from
548 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
549 // When we create the skeletal frames we need the oldest frame to be in the zero slot
550 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
551 // so things look a little strange in this loop.
552 //
553 int callee_parameters = 0;
554 int callee_locals = 0;
555 for (int index = 0; index < array->frames(); index++ ) {
556 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
557 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
558 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
559 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
560 callee_locals,
561 index == 0,
562 popframe_extra_args);
563 // This pc doesn't have to be perfect just good enough to identify the frame
564 // as interpreted so the skeleton frame will be walkable
565 // The correct pc will be set when the skeleton frame is completely filled out
566 // The final pc we store in the loop is wrong and will be overwritten below
567 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
568
569 callee_parameters = array->element(index)->method()->size_of_parameters();
570 callee_locals = array->element(index)->method()->max_locals();
571 popframe_extra_args = 0;
572 }
573
574 // Compute whether the root vframe returns a float or double value.
575 BasicType return_type;
576 {
577 methodHandle method(current, array->element(0)->method());
578 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
579 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
580 }
581
582 // Compute information for handling adapters and adjusting the frame size of the caller.
583 int caller_adjustment = 0;
584
585 // Compute the amount the oldest interpreter frame will have to adjust
586 // its caller's stack by. If the caller is a compiled frame then
587 // we pretend that the callee has no parameters so that the
588 // extension counts for the full amount of locals and not just
589 // locals-parms. This is because without a c2i adapter the parm
590 // area as created by the compiled frame will not be usable by
591 // the interpreter. (Depending on the calling convention there
592 // may not even be enough space).
593
594 // QQQ I'd rather see this pushed down into last_frame_adjust
595 // and have it take the sender (aka caller).
596
597 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
598 caller_adjustment = last_frame_adjust(0, callee_locals);
599 } else if (callee_locals > callee_parameters) {
600 // The caller frame may need extending to accommodate
601 // non-parameter locals of the first unpacked interpreted frame.
602 // Compute that adjustment.
603 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
604 }
605
606 // If the sender is deoptimized the we must retrieve the address of the handler
607 // since the frame will "magically" show the original pc before the deopt
608 // and we'd undo the deopt.
609
610 frame_pcs[0] = deopt_sender.raw_pc();
611
612 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
613
614 #if INCLUDE_JVMCI
615 if (exceptionObject() != NULL) {
616 current->set_exception_oop(exceptionObject());
617 exec_mode = Unpack_exception;
618 }
619 #endif
620
621 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
622 assert(current->has_pending_exception(), "should have thrown OOME");
623 current->set_exception_oop(current->pending_exception());
624 current->clear_pending_exception();
625 exec_mode = Unpack_exception;
626 }
627
628 #if INCLUDE_JVMCI
629 if (current->frames_to_pop_failed_realloc() > 0) {
630 current->set_pending_monitorenter(false);
631 }
632 #endif
633
634 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
635 caller_adjustment * BytesPerWord,
636 caller_was_method_handle ? 0 : callee_parameters,
637 number_of_frames,
638 frame_sizes,
639 frame_pcs,
640 return_type,
641 exec_mode);
642 // On some platforms, we need a way to pass some platform dependent
643 // information to the unpacking code so the skeletal frames come out
644 // correct (initial fp value, unextended sp, ...)
645 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
646
647 if (array->frames() > 1) {
648 if (VerifyStack && TraceDeoptimization) {
649 tty->print_cr("Deoptimizing method containing inlining");
650 }
651 }
652
653 array->set_unroll_block(info);
654 return info;
655 }
656
657 // Called to cleanup deoptimization data structures in normal case
658 // after unpacking to stack and when stack overflow error occurs
659 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
660 vframeArray *array) {
661
662 // Get array if coming from exception
663 if (array == NULL) {
664 array = thread->vframe_array_head();
665 }
666 thread->set_vframe_array_head(NULL);
667
668 // Free the previous UnrollBlock
669 vframeArray* old_array = thread->vframe_array_last();
670 thread->set_vframe_array_last(array);
671
672 if (old_array != NULL) {
673 UnrollBlock* old_info = old_array->unroll_block();
674 old_array->set_unroll_block(NULL);
675 delete old_info;
676 delete old_array;
677 }
678
679 // Deallocate any resource creating in this routine and any ResourceObjs allocated
680 // inside the vframeArray (StackValueCollections)
681
682 delete thread->deopt_mark();
683 thread->set_deopt_mark(NULL);
684 thread->set_deopt_compiled_method(NULL);
685
686
687 if (JvmtiExport::can_pop_frame()) {
688 // Regardless of whether we entered this routine with the pending
689 // popframe condition bit set, we should always clear it now
690 thread->clear_popframe_condition();
691 }
692
693 // unpack_frames() is called at the end of the deoptimization handler
694 // and (in C2) at the end of the uncommon trap handler. Note this fact
695 // so that an asynchronous stack walker can work again. This counter is
696 // incremented at the beginning of fetch_unroll_info() and (in C2) at
697 // the beginning of uncommon_trap().
698 thread->dec_in_deopt_handler();
699 }
700
701 // Moved from cpu directories because none of the cpus has callee save values.
702 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
703 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
704
705 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
706 // the days we had adapter frames. When we deoptimize a situation where a
707 // compiled caller calls a compiled caller will have registers it expects
708 // to survive the call to the callee. If we deoptimize the callee the only
709 // way we can restore these registers is to have the oldest interpreter
710 // frame that we create restore these values. That is what this routine
711 // will accomplish.
712
713 // At the moment we have modified c2 to not have any callee save registers
714 // so this problem does not exist and this routine is just a place holder.
715
716 assert(f->is_interpreted_frame(), "must be interpreted");
717 }
718
719 #ifndef PRODUCT
720 static bool falls_through(Bytecodes::Code bc) {
721 switch (bc) {
722 // List may be incomplete. Here we really only care about bytecodes where compiled code
723 // can deoptimize.
724 case Bytecodes::_goto:
725 case Bytecodes::_goto_w:
726 case Bytecodes::_athrow:
727 return false;
728 default:
729 return true;
730 }
731 }
732 #endif
733
734 // Return BasicType of value being returned
735 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
736
737 // We are already active in the special DeoptResourceMark any ResourceObj's we
738 // allocate will be freed at the end of the routine.
739
740 // JRT_LEAF methods don't normally allocate handles and there is a
741 // NoHandleMark to enforce that. It is actually safe to use Handles
742 // in a JRT_LEAF method, and sometimes desirable, but to do so we
743 // must use ResetNoHandleMark to bypass the NoHandleMark, and
744 // then use a HandleMark to ensure any Handles we do create are
745 // cleaned up in this scope.
746 ResetNoHandleMark rnhm;
747 HandleMark hm(thread);
748
749 frame stub_frame = thread->last_frame();
750
751 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
752 // must point to the vframeArray for the unpack frame.
753 vframeArray* array = thread->vframe_array_head();
754
755 #ifndef PRODUCT
756 if (TraceDeoptimization) {
757 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d",
758 p2i(thread), p2i(array), exec_mode);
759 }
760 #endif
761 Events::log_deopt_message(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
762 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode);
763
764 UnrollBlock* info = array->unroll_block();
765
766 // We set the last_Java frame. But the stack isn't really parsable here. So we
767 // clear it to make sure JFR understands not to try and walk stacks from events
768 // in here.
769 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
770 thread->frame_anchor()->set_last_Java_sp(NULL);
771
772 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
773 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
774
775 thread->frame_anchor()->set_last_Java_sp(sp);
776
777 BasicType bt = info->return_type();
778
779 // If we have an exception pending, claim that the return type is an oop
780 // so the deopt_blob does not overwrite the exception_oop.
781
782 if (exec_mode == Unpack_exception)
783 bt = T_OBJECT;
784
785 // Cleanup thread deopt data
786 cleanup_deopt_info(thread, array);
787
788 #ifndef PRODUCT
789 if (VerifyStack) {
790 ResourceMark res_mark;
791 // Clear pending exception to not break verification code (restored afterwards)
792 PreserveExceptionMark pm(thread);
793
794 thread->validate_frame_layout();
795
796 // Verify that the just-unpacked frames match the interpreter's
797 // notions of expression stack and locals
798 vframeArray* cur_array = thread->vframe_array_last();
799 RegisterMap rm(thread, false);
800 rm.set_include_argument_oops(false);
801 bool is_top_frame = true;
802 int callee_size_of_parameters = 0;
803 int callee_max_locals = 0;
804 for (int i = 0; i < cur_array->frames(); i++) {
805 vframeArrayElement* el = cur_array->element(i);
806 frame* iframe = el->iframe();
807 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
808
809 // Get the oop map for this bci
810 InterpreterOopMap mask;
811 int cur_invoke_parameter_size = 0;
812 bool try_next_mask = false;
813 int next_mask_expression_stack_size = -1;
814 int top_frame_expression_stack_adjustment = 0;
815 methodHandle mh(thread, iframe->interpreter_frame_method());
816 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
817 BytecodeStream str(mh, iframe->interpreter_frame_bci());
818 int max_bci = mh->code_size();
819 // Get to the next bytecode if possible
820 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
821 // Check to see if we can grab the number of outgoing arguments
822 // at an uncommon trap for an invoke (where the compiler
823 // generates debug info before the invoke has executed)
824 Bytecodes::Code cur_code = str.next();
825 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere;
826 if (Bytecodes::is_invoke(cur_code)) {
827 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
828 cur_invoke_parameter_size = invoke.size_of_parameters();
829 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
830 callee_size_of_parameters++;
831 }
832 }
833 if (str.bci() < max_bci) {
834 next_code = str.next();
835 if (next_code >= 0) {
836 // The interpreter oop map generator reports results before
837 // the current bytecode has executed except in the case of
838 // calls. It seems to be hard to tell whether the compiler
839 // has emitted debug information matching the "state before"
840 // a given bytecode or the state after, so we try both
841 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) {
842 // Get expression stack size for the next bytecode
843 InterpreterOopMap next_mask;
844 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
845 next_mask_expression_stack_size = next_mask.expression_stack_size();
846 if (Bytecodes::is_invoke(next_code)) {
847 Bytecode_invoke invoke(mh, str.bci());
848 next_mask_expression_stack_size += invoke.size_of_parameters();
849 }
850 // Need to subtract off the size of the result type of
851 // the bytecode because this is not described in the
852 // debug info but returned to the interpreter in the TOS
853 // caching register
854 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
855 if (bytecode_result_type != T_ILLEGAL) {
856 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
857 }
858 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
859 try_next_mask = true;
860 }
861 }
862 }
863
864 // Verify stack depth and oops in frame
865 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
866 if (!(
867 /* SPARC */
868 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
869 /* x86 */
870 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
871 (try_next_mask &&
872 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
873 top_frame_expression_stack_adjustment))) ||
874 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
875 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
876 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
877 )) {
878 {
879 // Print out some information that will help us debug the problem
880 tty->print_cr("Wrong number of expression stack elements during deoptimization");
881 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
882 tty->print_cr(" Current code %s", Bytecodes::name(cur_code));
883 if (try_next_mask) {
884 tty->print_cr(" Next code %s", Bytecodes::name(next_code));
885 }
886 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
887 iframe->interpreter_frame_expression_stack_size());
888 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
889 tty->print_cr(" try_next_mask = %d", try_next_mask);
890 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
891 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
892 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
893 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
894 tty->print_cr(" exec_mode = %d", exec_mode);
895 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
896 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
897 tty->print_cr(" Interpreted frames:");
898 for (int k = 0; k < cur_array->frames(); k++) {
899 vframeArrayElement* el = cur_array->element(k);
900 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
901 }
902 cur_array->print_on_2(tty);
903 }
904 guarantee(false, "wrong number of expression stack elements during deopt");
905 }
906 VerifyOopClosure verify;
907 iframe->oops_interpreted_do(&verify, &rm, false);
908 callee_size_of_parameters = mh->size_of_parameters();
909 callee_max_locals = mh->max_locals();
910 is_top_frame = false;
911 }
912 }
913 #endif /* !PRODUCT */
914
915 return bt;
916 JRT_END
917
918 class DeoptimizeMarkedClosure : public HandshakeClosure {
919 public:
920 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
921 void do_thread(Thread* thread) {
922 JavaThread* jt = thread->as_Java_thread();
923 jt->deoptimize_marked_methods();
924 }
925 };
926
927 void Deoptimization::deoptimize_all_marked(nmethod* nmethod_only) {
928 ResourceMark rm;
929 DeoptimizationMarker dm;
930
931 // Make the dependent methods not entrant
932 if (nmethod_only != NULL) {
933 nmethod_only->mark_for_deoptimization();
934 nmethod_only->make_not_entrant();
935 } else {
936 MutexLocker mu(SafepointSynchronize::is_at_safepoint() ? NULL : CodeCache_lock, Mutex::_no_safepoint_check_flag);
937 CodeCache::make_marked_nmethods_not_entrant();
938 }
939
940 DeoptimizeMarkedClosure deopt;
941 if (SafepointSynchronize::is_at_safepoint()) {
942 Threads::java_threads_do(&deopt);
943 } else {
944 Handshake::execute(&deopt);
945 }
946 }
947
948 Deoptimization::DeoptAction Deoptimization::_unloaded_action
949 = Deoptimization::Action_reinterpret;
950
951 #if COMPILER2_OR_JVMCI
952 template<typename CacheType>
953 class BoxCacheBase : public CHeapObj<mtCompiler> {
954 protected:
955 static InstanceKlass* find_cache_klass(Symbol* klass_name) {
956 ResourceMark rm;
957 char* klass_name_str = klass_name->as_C_string();
958 InstanceKlass* ik = SystemDictionary::find_instance_klass(klass_name, Handle(), Handle());
959 guarantee(ik != NULL, "%s must be loaded", klass_name_str);
960 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
961 CacheType::compute_offsets(ik);
962 return ik;
963 }
964 };
965
966 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
967 PrimitiveType _low;
968 PrimitiveType _high;
969 jobject _cache;
970 protected:
971 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
972 BoxCache(Thread* thread) {
973 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(CacheType::symbol());
974 objArrayOop cache = CacheType::cache(ik);
975 assert(cache->length() > 0, "Empty cache");
976 _low = BoxType::value(cache->obj_at(0));
977 _high = _low + cache->length() - 1;
978 _cache = JNIHandles::make_global(Handle(thread, cache));
979 }
980 ~BoxCache() {
981 JNIHandles::destroy_global(_cache);
982 }
983 public:
984 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
985 if (_singleton == NULL) {
986 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
987 if (!Atomic::replace_if_null(&_singleton, s)) {
988 delete s;
989 }
990 }
991 return _singleton;
992 }
993 oop lookup(PrimitiveType value) {
994 if (_low <= value && value <= _high) {
995 int offset = value - _low;
996 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
997 }
998 return NULL;
999 }
1000 oop lookup_raw(intptr_t raw_value) {
1001 // Have to cast to avoid little/big-endian problems.
1002 if (sizeof(PrimitiveType) > sizeof(jint)) {
1003 jlong value = (jlong)raw_value;
1004 return lookup(value);
1005 }
1006 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1007 return lookup(value);
1008 }
1009 };
1010
1011 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1012 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1013 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1014 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1015 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1016
1017 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL;
1018 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL;
1019 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL;
1020 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL;
1021 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL;
1022
1023 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1024 jobject _true_cache;
1025 jobject _false_cache;
1026 protected:
1027 static BooleanBoxCache *_singleton;
1028 BooleanBoxCache(Thread *thread) {
1029 InstanceKlass* ik = find_cache_klass(java_lang_Boolean::symbol());
1030 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1031 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1032 }
1033 ~BooleanBoxCache() {
1034 JNIHandles::destroy_global(_true_cache);
1035 JNIHandles::destroy_global(_false_cache);
1036 }
1037 public:
1038 static BooleanBoxCache* singleton(Thread* thread) {
1039 if (_singleton == NULL) {
1040 BooleanBoxCache* s = new BooleanBoxCache(thread);
1041 if (!Atomic::replace_if_null(&_singleton, s)) {
1042 delete s;
1043 }
1044 }
1045 return _singleton;
1046 }
1047 oop lookup_raw(intptr_t raw_value) {
1048 // Have to cast to avoid little/big-endian problems.
1049 jboolean value = (jboolean)*((jint*)&raw_value);
1050 return lookup(value);
1051 }
1052 oop lookup(jboolean value) {
1053 if (value != 0) {
1054 return JNIHandles::resolve_non_null(_true_cache);
1055 }
1056 return JNIHandles::resolve_non_null(_false_cache);
1057 }
1058 };
1059
1060 BooleanBoxCache* BooleanBoxCache::_singleton = NULL;
1061
1062 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS) {
1063 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1064 BasicType box_type = vmClasses::box_klass_type(k);
1065 if (box_type != T_OBJECT) {
1066 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1067 switch(box_type) {
1068 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1069 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1070 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1071 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1072 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1073 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1074 default:;
1075 }
1076 }
1077 return NULL;
1078 }
1079
1080 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1081 Handle pending_exception(THREAD, thread->pending_exception());
1082 const char* exception_file = thread->exception_file();
1083 int exception_line = thread->exception_line();
1084 thread->clear_pending_exception();
1085
1086 bool failures = false;
1087
1088 for (int i = 0; i < objects->length(); i++) {
1089 assert(objects->at(i)->is_object(), "invalid debug information");
1090 ObjectValue* sv = (ObjectValue*) objects->at(i);
1091
1092 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1093 oop obj = NULL;
1094
1095 if (k->is_instance_klass()) {
1096 if (sv->is_auto_box()) {
1097 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1098 obj = get_cached_box(abv, fr, reg_map, THREAD);
1099 if (obj != NULL) {
1100 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1101 abv->set_cached(true);
1102 }
1103 }
1104
1105 InstanceKlass* ik = InstanceKlass::cast(k);
1106 if (obj == NULL) {
1107 #ifdef COMPILER2
1108 if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1109 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1110 } else {
1111 obj = ik->allocate_instance(THREAD);
1112 }
1113 #else
1114 obj = ik->allocate_instance(THREAD);
1115 #endif // COMPILER2
1116 }
1117 } else if (k->is_typeArray_klass()) {
1118 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1119 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1120 int len = sv->field_size() / type2size[ak->element_type()];
1121 obj = ak->allocate(len, THREAD);
1122 } else if (k->is_objArray_klass()) {
1123 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1124 obj = ak->allocate(sv->field_size(), THREAD);
1125 }
1126
1127 if (obj == NULL) {
1128 failures = true;
1129 }
1130
1131 assert(sv->value().is_null(), "redundant reallocation");
1132 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
1133 CLEAR_PENDING_EXCEPTION;
1134 sv->set_value(obj);
1135 }
1136
1137 if (failures) {
1138 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1139 } else if (pending_exception.not_null()) {
1140 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1141 }
1142
1143 return failures;
1144 }
1145
1146 #if INCLUDE_JVMCI
1147 /**
1148 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1149 * we need to somehow be able to recover the actual kind to be able to write the correct
1150 * amount of bytes.
1151 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1152 * the entries at index n + 1 to n + i are 'markers'.
1153 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1154 * expected form of the array would be:
1155 *
1156 * {b0, b1, b2, b3, INT, marker, b6, b7}
1157 *
1158 * Thus, in order to get back the size of the entry, we simply need to count the number
1159 * of marked entries
1160 *
1161 * @param virtualArray the virtualized byte array
1162 * @param i index of the virtual entry we are recovering
1163 * @return The number of bytes the entry spans
1164 */
1165 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1166 int index = i;
1167 while (++index < virtualArray->field_size() &&
1168 virtualArray->field_at(index)->is_marker()) {}
1169 return index - i;
1170 }
1171
1172 /**
1173 * If there was a guarantee for byte array to always start aligned to a long, we could
1174 * do a simple check on the parity of the index. Unfortunately, that is not always the
1175 * case. Thus, we check alignment of the actual address we are writing to.
1176 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1177 * write a long to index 3.
1178 */
1179 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1180 jbyte* res = obj->byte_at_addr(index);
1181 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1182 return res;
1183 }
1184
1185 static void byte_array_put(typeArrayOop obj, intptr_t val, int index, int byte_count) {
1186 switch (byte_count) {
1187 case 1:
1188 obj->byte_at_put(index, (jbyte) *((jint *) &val));
1189 break;
1190 case 2:
1191 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) *((jint *) &val);
1192 break;
1193 case 4:
1194 *((jint *) check_alignment_get_addr(obj, index, 4)) = (jint) *((jint *) &val);
1195 break;
1196 case 8:
1197 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) *((jlong *) &val);
1198 break;
1199 default:
1200 ShouldNotReachHere();
1201 }
1202 }
1203 #endif // INCLUDE_JVMCI
1204
1205
1206 // restore elements of an eliminated type array
1207 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1208 int index = 0;
1209 intptr_t val;
1210
1211 for (int i = 0; i < sv->field_size(); i++) {
1212 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1213 switch(type) {
1214 case T_LONG: case T_DOUBLE: {
1215 assert(value->type() == T_INT, "Agreement.");
1216 StackValue* low =
1217 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1218 #ifdef _LP64
1219 jlong res = (jlong)low->get_int();
1220 #else
1221 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1222 #endif
1223 obj->long_at_put(index, res);
1224 break;
1225 }
1226
1227 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1228 case T_INT: case T_FLOAT: { // 4 bytes.
1229 assert(value->type() == T_INT, "Agreement.");
1230 bool big_value = false;
1231 if (i + 1 < sv->field_size() && type == T_INT) {
1232 if (sv->field_at(i)->is_location()) {
1233 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1234 if (type == Location::dbl || type == Location::lng) {
1235 big_value = true;
1236 }
1237 } else if (sv->field_at(i)->is_constant_int()) {
1238 ScopeValue* next_scope_field = sv->field_at(i + 1);
1239 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1240 big_value = true;
1241 }
1242 }
1243 }
1244
1245 if (big_value) {
1246 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1247 #ifdef _LP64
1248 jlong res = (jlong)low->get_int();
1249 #else
1250 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1251 #endif
1252 obj->int_at_put(index, (jint)*((jint*)&res));
1253 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
1254 } else {
1255 val = value->get_int();
1256 obj->int_at_put(index, (jint)*((jint*)&val));
1257 }
1258 break;
1259 }
1260
1261 case T_SHORT:
1262 assert(value->type() == T_INT, "Agreement.");
1263 val = value->get_int();
1264 obj->short_at_put(index, (jshort)*((jint*)&val));
1265 break;
1266
1267 case T_CHAR:
1268 assert(value->type() == T_INT, "Agreement.");
1269 val = value->get_int();
1270 obj->char_at_put(index, (jchar)*((jint*)&val));
1271 break;
1272
1273 case T_BYTE: {
1274 assert(value->type() == T_INT, "Agreement.");
1275 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1276 val = value->get_int();
1277 #if INCLUDE_JVMCI
1278 int byte_count = count_number_of_bytes_for_entry(sv, i);
1279 byte_array_put(obj, val, index, byte_count);
1280 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1281 i += byte_count - 1; // Balance the loop counter.
1282 index += byte_count;
1283 // index has been updated so continue at top of loop
1284 continue;
1285 #else
1286 obj->byte_at_put(index, (jbyte)*((jint*)&val));
1287 break;
1288 #endif // INCLUDE_JVMCI
1289 }
1290
1291 case T_BOOLEAN: {
1292 assert(value->type() == T_INT, "Agreement.");
1293 val = value->get_int();
1294 obj->bool_at_put(index, (jboolean)*((jint*)&val));
1295 break;
1296 }
1297
1298 default:
1299 ShouldNotReachHere();
1300 }
1301 index++;
1302 }
1303 }
1304
1305 // restore fields of an eliminated object array
1306 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1307 for (int i = 0; i < sv->field_size(); i++) {
1308 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1309 assert(value->type() == T_OBJECT, "object element expected");
1310 obj->obj_at_put(i, value->get_obj()());
1311 }
1312 }
1313
1314 class ReassignedField {
1315 public:
1316 int _offset;
1317 BasicType _type;
1318 public:
1319 ReassignedField() {
1320 _offset = 0;
1321 _type = T_ILLEGAL;
1322 }
1323 };
1324
1325 int compare(ReassignedField* left, ReassignedField* right) {
1326 return left->_offset - right->_offset;
1327 }
1328
1329 // Restore fields of an eliminated instance object using the same field order
1330 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1331 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
1332 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1333 InstanceKlass* ik = klass;
1334 while (ik != NULL) {
1335 for (AllFieldStream fs(ik); !fs.done(); fs.next()) {
1336 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
1337 ReassignedField field;
1338 field._offset = fs.offset();
1339 field._type = Signature::basic_type(fs.signature());
1340 fields->append(field);
1341 }
1342 }
1343 ik = ik->superklass();
1344 }
1345 fields->sort(compare);
1346 for (int i = 0; i < fields->length(); i++) {
1347 intptr_t val;
1348 ScopeValue* scope_field = sv->field_at(svIndex);
1349 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1350 int offset = fields->at(i)._offset;
1351 BasicType type = fields->at(i)._type;
1352 switch (type) {
1353 case T_OBJECT: case T_ARRAY:
1354 assert(value->type() == T_OBJECT, "Agreement.");
1355 obj->obj_field_put(offset, value->get_obj()());
1356 break;
1357
1358 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1359 case T_INT: case T_FLOAT: { // 4 bytes.
1360 assert(value->type() == T_INT, "Agreement.");
1361 bool big_value = false;
1362 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1363 if (scope_field->is_location()) {
1364 Location::Type type = ((LocationValue*) scope_field)->location().type();
1365 if (type == Location::dbl || type == Location::lng) {
1366 big_value = true;
1367 }
1368 }
1369 if (scope_field->is_constant_int()) {
1370 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1371 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1372 big_value = true;
1373 }
1374 }
1375 }
1376
1377 if (big_value) {
1378 i++;
1379 assert(i < fields->length(), "second T_INT field needed");
1380 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1381 } else {
1382 val = value->get_int();
1383 obj->int_field_put(offset, (jint)*((jint*)&val));
1384 break;
1385 }
1386 }
1387 /* no break */
1388
1389 case T_LONG: case T_DOUBLE: {
1390 assert(value->type() == T_INT, "Agreement.");
1391 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1392 #ifdef _LP64
1393 jlong res = (jlong)low->get_int();
1394 #else
1395 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1396 #endif
1397 obj->long_field_put(offset, res);
1398 break;
1399 }
1400
1401 case T_SHORT:
1402 assert(value->type() == T_INT, "Agreement.");
1403 val = value->get_int();
1404 obj->short_field_put(offset, (jshort)*((jint*)&val));
1405 break;
1406
1407 case T_CHAR:
1408 assert(value->type() == T_INT, "Agreement.");
1409 val = value->get_int();
1410 obj->char_field_put(offset, (jchar)*((jint*)&val));
1411 break;
1412
1413 case T_BYTE:
1414 assert(value->type() == T_INT, "Agreement.");
1415 val = value->get_int();
1416 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1417 break;
1418
1419 case T_BOOLEAN:
1420 assert(value->type() == T_INT, "Agreement.");
1421 val = value->get_int();
1422 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1423 break;
1424
1425 default:
1426 ShouldNotReachHere();
1427 }
1428 svIndex++;
1429 }
1430 return svIndex;
1431 }
1432
1433 // restore fields of all eliminated objects and arrays
1434 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1435 for (int i = 0; i < objects->length(); i++) {
1436 ObjectValue* sv = (ObjectValue*) objects->at(i);
1437 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1438 Handle obj = sv->value();
1439 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1440 if (PrintDeoptimizationDetails) {
1441 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1442 }
1443 if (obj.is_null()) {
1444 continue;
1445 }
1446
1447 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1448 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1449 continue;
1450 }
1451 #ifdef COMPILER2
1452 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1453 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1454 ScopeValue* payload = sv->field_at(0);
1455 if (payload->is_location() &&
1456 payload->as_LocationValue()->location().type() == Location::vector) {
1457 if (PrintDeoptimizationDetails) {
1458 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1459 if (Verbose) {
1460 Handle obj = sv->value();
1461 k->oop_print_on(obj(), tty);
1462 }
1463 }
1464 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1465 }
1466 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1467 // which could be restored after vector object allocation.
1468 }
1469 #endif
1470 if (k->is_instance_klass()) {
1471 InstanceKlass* ik = InstanceKlass::cast(k);
1472 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1473 } else if (k->is_typeArray_klass()) {
1474 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1475 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1476 } else if (k->is_objArray_klass()) {
1477 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1478 }
1479 }
1480 }
1481
1482
1483 // relock objects for which synchronization was eliminated
1484 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1485 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1486 bool relocked_objects = false;
1487 for (int i = 0; i < monitors->length(); i++) {
1488 MonitorInfo* mon_info = monitors->at(i);
1489 if (mon_info->eliminated()) {
1490 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1491 relocked_objects = true;
1492 if (!mon_info->owner_is_scalar_replaced()) {
1493 Handle obj(thread, mon_info->owner());
1494 markWord mark = obj->mark();
1495 if (UseBiasedLocking && mark.has_bias_pattern()) {
1496 // New allocated objects may have the mark set to anonymously biased.
1497 // Also the deoptimized method may called methods with synchronization
1498 // where the thread-local object is bias locked to the current thread.
1499 assert(mark.is_biased_anonymously() ||
1500 mark.biased_locker() == deoptee_thread, "should be locked to current thread");
1501 // Reset mark word to unbiased prototype.
1502 markWord unbiased_prototype = markWord::prototype().set_age(mark.age());
1503 obj->set_mark(unbiased_prototype);
1504 } else if (exec_mode == Unpack_none) {
1505 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1506 // With exec_mode == Unpack_none obj may be thread local and locked in
1507 // a callee frame. In this case the bias was revoked before in revoke_for_object_deoptimization().
1508 // Make the lock in the callee a recursive lock and restore the displaced header.
1509 markWord dmw = mark.displaced_mark_helper();
1510 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) NULL));
1511 obj->set_mark(dmw);
1512 }
1513 if (mark.has_monitor()) {
1514 // defer relocking if the deoptee thread is currently waiting for obj
1515 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1516 if (waiting_monitor != NULL && waiting_monitor->object() == obj()) {
1517 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1518 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1519 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1520 continue;
1521 }
1522 }
1523 }
1524 if (LockingMode == LM_LIGHTWEIGHT && exec_mode == Unpack_none) {
1525 // We have lost information about the correct state of the lock stack.
1526 // Inflate the locks instead. Enter then inflate to avoid races with
1527 // deflation.
1528 ObjectSynchronizer::enter_for(obj, nullptr, deoptee_thread);
1529 assert(mon_info->owner()->is_locked(), "object must be locked now");
1530 ObjectMonitor* mon = ObjectSynchronizer::inflate_for(deoptee_thread, obj(), ObjectSynchronizer::inflate_cause_vm_internal);
1531 assert(mon->owner() == deoptee_thread, "must be");
1532 } else {
1533 BasicLock* lock = mon_info->lock();
1534 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1535 assert(mon_info->owner()->is_locked(), "object must be locked now");
1536 }
1537 }
1538 }
1539 }
1540 return relocked_objects;
1541 }
1542 #endif // COMPILER2_OR_JVMCI
1543
1544 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1545 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1546
1547 #ifndef PRODUCT
1548 if (PrintDeoptimizationDetails) {
1549 ResourceMark rm;
1550 stringStream st;
1551 st.print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1552 fr.print_on(&st);
1553 st.print_cr(" Virtual frames (innermost first):");
1554 for (int index = 0; index < chunk->length(); index++) {
1555 compiledVFrame* vf = chunk->at(index);
1556 st.print(" %2d - ", index);
1557 vf->print_value_on(&st);
1558 int bci = chunk->at(index)->raw_bci();
1559 const char* code_name;
1560 if (bci == SynchronizationEntryBCI) {
1561 code_name = "sync entry";
1562 } else {
1563 Bytecodes::Code code = vf->method()->code_at(bci);
1564 code_name = Bytecodes::name(code);
1565 }
1566 st.print(" - %s", code_name);
1567 st.print_cr(" @ bci %d ", bci);
1568 if (Verbose) {
1569 vf->print_on(&st);
1570 st.cr();
1571 }
1572 }
1573 tty->print_raw(st.as_string());
1574 }
1575 #endif
1576
1577 // Register map for next frame (used for stack crawl). We capture
1578 // the state of the deopt'ing frame's caller. Thus if we need to
1579 // stuff a C2I adapter we can properly fill in the callee-save
1580 // register locations.
1581 frame caller = fr.sender(reg_map);
1582 int frame_size = caller.sp() - fr.sp();
1583
1584 frame sender = caller;
1585
1586 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1587 // the vframeArray containing the unpacking information is allocated in the C heap.
1588 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1589 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1590
1591 // Compare the vframeArray to the collected vframes
1592 assert(array->structural_compare(thread, chunk), "just checking");
1593
1594 #ifndef PRODUCT
1595 if (PrintDeoptimizationDetails) {
1596 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1597 }
1598 #endif // PRODUCT
1599
1600 return array;
1601 }
1602
1603 #if COMPILER2_OR_JVMCI
1604 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1605 // Reallocation of some scalar replaced objects failed. Record
1606 // that we need to pop all the interpreter frames for the
1607 // deoptimized compiled frame.
1608 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1609 thread->set_frames_to_pop_failed_realloc(array->frames());
1610 // Unlock all monitors here otherwise the interpreter will see a
1611 // mix of locked and unlocked monitors (because of failed
1612 // reallocations of synchronized objects) and be confused.
1613 for (int i = 0; i < array->frames(); i++) {
1614 MonitorChunk* monitors = array->element(i)->monitors();
1615 if (monitors != NULL) {
1616 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1617 BasicObjectLock* src = monitors->at(j);
1618 if (src->obj() != NULL) {
1619 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1620 }
1621 }
1622 array->element(i)->free_monitors(thread);
1623 #ifdef ASSERT
1624 array->element(i)->set_removed_monitors();
1625 #endif
1626 }
1627 }
1628 }
1629 #endif
1630
1631 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke,
1632 bool only_eliminated) {
1633 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1634 Thread* thread = Thread::current();
1635 for (int i = 0; i < monitors->length(); i++) {
1636 MonitorInfo* mon_info = monitors->at(i);
1637 if (mon_info->eliminated() == only_eliminated &&
1638 !mon_info->owner_is_scalar_replaced() &&
1639 mon_info->owner() != NULL) {
1640 objects_to_revoke->append(Handle(thread, mon_info->owner()));
1641 }
1642 }
1643 }
1644
1645 static void get_monitors_from_stack(GrowableArray<Handle>* objects_to_revoke, JavaThread* thread,
1646 frame fr, RegisterMap* map, bool only_eliminated) {
1647 // Unfortunately we don't have a RegisterMap available in most of
1648 // the places we want to call this routine so we need to walk the
1649 // stack again to update the register map.
1650 if (map == NULL || !map->update_map()) {
1651 StackFrameStream sfs(thread, true /* update */, true /* process_frames */);
1652 bool found = false;
1653 while (!found && !sfs.is_done()) {
1654 frame* cur = sfs.current();
1655 sfs.next();
1656 found = cur->id() == fr.id();
1657 }
1658 assert(found, "frame to be deoptimized not found on target thread's stack");
1659 map = sfs.register_map();
1660 }
1661
1662 vframe* vf = vframe::new_vframe(&fr, map, thread);
1663 compiledVFrame* cvf = compiledVFrame::cast(vf);
1664 // Revoke monitors' biases in all scopes
1665 while (!cvf->is_top()) {
1666 collect_monitors(cvf, objects_to_revoke, only_eliminated);
1667 cvf = compiledVFrame::cast(cvf->sender());
1668 }
1669 collect_monitors(cvf, objects_to_revoke, only_eliminated);
1670 }
1671
1672 void Deoptimization::revoke_from_deopt_handler(JavaThread* thread, frame fr, RegisterMap* map) {
1673 if (!UseBiasedLocking) {
1674 return;
1675 }
1676 assert(thread == Thread::current(), "should be");
1677 ResourceMark rm(thread);
1678 HandleMark hm(thread);
1679 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1680 get_monitors_from_stack(objects_to_revoke, thread, fr, map, false);
1681
1682 int len = objects_to_revoke->length();
1683 for (int i = 0; i < len; i++) {
1684 oop obj = (objects_to_revoke->at(i))();
1685 BiasedLocking::revoke_own_lock(thread, objects_to_revoke->at(i));
1686 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
1687 }
1688 }
1689
1690 // Revoke the bias of objects with eliminated locking to prepare subsequent relocking.
1691 void Deoptimization::revoke_for_object_deoptimization(JavaThread* deoptee_thread, frame fr,
1692 RegisterMap* map, JavaThread* thread) {
1693 if (!UseBiasedLocking) {
1694 return;
1695 }
1696 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1697 assert(KeepStackGCProcessedMark::stack_is_kept_gc_processed(deoptee_thread), "must be");
1698 // Collect monitors but only those with eliminated locking.
1699 get_monitors_from_stack(objects_to_revoke, deoptee_thread, fr, map, true);
1700
1701 int len = objects_to_revoke->length();
1702 for (int i = 0; i < len; i++) {
1703 oop obj = (objects_to_revoke->at(i))();
1704 markWord mark = obj->mark();
1705 if (!mark.has_bias_pattern() ||
1706 mark.is_biased_anonymously() || // eliminated locking does not bias an object if it wasn't before
1707 !obj->klass()->prototype_header().has_bias_pattern() || // bulk revoke ignores eliminated monitors
1708 (obj->klass()->prototype_header().bias_epoch() != mark.bias_epoch())) { // bulk rebias ignores eliminated monitors
1709 // We reach here regularly if there's just eliminated locking on obj.
1710 // We must not call BiasedLocking::revoke_own_lock() in this case, as we
1711 // would hit assertions because it is a prerequisite that there has to be
1712 // non-eliminated locking on obj by deoptee_thread.
1713 // Luckily we don't have to revoke here because obj has to be a
1714 // non-escaping obj and can be relocked without revoking the bias. See
1715 // Deoptimization::relock_objects().
1716 continue;
1717 }
1718 BiasedLocking::revoke(thread, objects_to_revoke->at(i));
1719 assert(!objects_to_revoke->at(i)->mark().has_bias_pattern(), "biases should be revoked by now");
1720 }
1721 }
1722
1723 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1724 assert(fr.can_be_deoptimized(), "checking frame type");
1725
1726 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1727
1728 if (LogCompilation && xtty != NULL) {
1729 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1730 assert(cm != NULL, "only compiled methods can deopt");
1731
1732 ttyLocker ttyl;
1733 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1734 cm->log_identity(xtty);
1735 xtty->end_head();
1736 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1737 xtty->begin_elem("jvms bci='%d'", sd->bci());
1738 xtty->method(sd->method());
1739 xtty->end_elem();
1740 if (sd->is_top()) break;
1741 }
1742 xtty->tail("deoptimized");
1743 }
1744
1745 // Patch the compiled method so that when execution returns to it we will
1746 // deopt the execution state and return to the interpreter.
1747 fr.deoptimize(thread);
1748 }
1749
1750 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1751 // Deoptimize only if the frame comes from compile code.
1752 // Do not deoptimize the frame which is already patched
1753 // during the execution of the loops below.
1754 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1755 return;
1756 }
1757 ResourceMark rm;
1758 DeoptimizationMarker dm;
1759 deoptimize_single_frame(thread, fr, reason);
1760 }
1761
1762 #if INCLUDE_JVMCI
1763 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) {
1764 // there is no exception handler for this pc => deoptimize
1765 cm->make_not_entrant();
1766
1767 // Use Deoptimization::deoptimize for all of its side-effects:
1768 // gathering traps statistics, logging...
1769 // it also patches the return pc but we do not care about that
1770 // since we return a continuation to the deopt_blob below.
1771 JavaThread* thread = JavaThread::current();
1772 RegisterMap reg_map(thread, false);
1773 frame runtime_frame = thread->last_frame();
1774 frame caller_frame = runtime_frame.sender(®_map);
1775 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method");
1776 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1777 compiledVFrame* cvf = compiledVFrame::cast(vf);
1778 ScopeDesc* imm_scope = cvf->scope();
1779 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1780 if (imm_mdo != NULL) {
1781 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), NULL);
1782 if (pdata != NULL && pdata->is_BitData()) {
1783 BitData* bit_data = (BitData*) pdata;
1784 bit_data->set_exception_seen();
1785 }
1786 }
1787
1788 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1789
1790 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true);
1791 if (trap_mdo != NULL) {
1792 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1793 }
1794
1795 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1796 }
1797 #endif
1798
1799 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1800 assert(thread == Thread::current() ||
1801 thread->is_handshake_safe_for(Thread::current()) ||
1802 SafepointSynchronize::is_at_safepoint(),
1803 "can only deoptimize other thread at a safepoint/handshake");
1804 // Compute frame and register map based on thread and sp.
1805 RegisterMap reg_map(thread, false);
1806 frame fr = thread->last_frame();
1807 while (fr.id() != id) {
1808 fr = fr.sender(®_map);
1809 }
1810 deoptimize(thread, fr, reason);
1811 }
1812
1813
1814 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1815 Thread* current = Thread::current();
1816 if (thread == current || thread->is_handshake_safe_for(current)) {
1817 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1818 } else {
1819 VM_DeoptimizeFrame deopt(thread, id, reason);
1820 VMThread::execute(&deopt);
1821 }
1822 }
1823
1824 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1825 deoptimize_frame(thread, id, Reason_constraint);
1826 }
1827
1828 // JVMTI PopFrame support
1829 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1830 {
1831 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1832 }
1833 JRT_END
1834
1835 MethodData*
1836 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1837 bool create_if_missing) {
1838 JavaThread* THREAD = thread; // For exception macros.
1839 MethodData* mdo = m()->method_data();
1840 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1841 // Build an MDO. Ignore errors like OutOfMemory;
1842 // that simply means we won't have an MDO to update.
1843 Method::build_interpreter_method_data(m, THREAD);
1844 if (HAS_PENDING_EXCEPTION) {
1845 // Only metaspace OOM is expected. No Java code executed.
1846 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
1847 CLEAR_PENDING_EXCEPTION;
1848 }
1849 mdo = m()->method_data();
1850 }
1851 return mdo;
1852 }
1853
1854 #if COMPILER2_OR_JVMCI
1855 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1856 // In case of an unresolved klass entry, load the class.
1857 // This path is exercised from case _ldc in Parse::do_one_bytecode,
1858 // and probably nowhere else.
1859 // Even that case would benefit from simply re-interpreting the
1860 // bytecode, without paying special attention to the class index.
1861 // So this whole "class index" feature should probably be removed.
1862
1863 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1864 Klass* tk = constant_pool->klass_at(index, THREAD);
1865 if (HAS_PENDING_EXCEPTION) {
1866 // Exception happened during classloading. We ignore the exception here, since it
1867 // is going to be rethrown since the current activation is going to be deoptimized and
1868 // the interpreter will re-execute the bytecode.
1869 // Do not clear probable Async Exceptions.
1870 CLEAR_PENDING_NONASYNC_EXCEPTION;
1871 // Class loading called java code which may have caused a stack
1872 // overflow. If the exception was thrown right before the return
1873 // to the runtime the stack is no longer guarded. Reguard the
1874 // stack otherwise if we return to the uncommon trap blob and the
1875 // stack bang causes a stack overflow we crash.
1876 JavaThread* jt = THREAD;
1877 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
1878 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1879 }
1880 return;
1881 }
1882
1883 assert(!constant_pool->tag_at(index).is_symbol(),
1884 "no symbolic names here, please");
1885 }
1886
1887 #if INCLUDE_JFR
1888
1889 class DeoptReasonSerializer : public JfrSerializer {
1890 public:
1891 void serialize(JfrCheckpointWriter& writer) {
1892 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1893 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1894 writer.write_key((u8)i);
1895 writer.write(Deoptimization::trap_reason_name(i));
1896 }
1897 }
1898 };
1899
1900 class DeoptActionSerializer : public JfrSerializer {
1901 public:
1902 void serialize(JfrCheckpointWriter& writer) {
1903 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1904 writer.write_count(nof_actions);
1905 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1906 writer.write_key(i);
1907 writer.write(Deoptimization::trap_action_name((int)i));
1908 }
1909 }
1910 };
1911
1912 static void register_serializers() {
1913 static int critical_section = 0;
1914 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1915 return;
1916 }
1917 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1918 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1919 }
1920
1921 static void post_deoptimization_event(CompiledMethod* nm,
1922 const Method* method,
1923 int trap_bci,
1924 int instruction,
1925 Deoptimization::DeoptReason reason,
1926 Deoptimization::DeoptAction action) {
1927 assert(nm != NULL, "invariant");
1928 assert(method != NULL, "invariant");
1929 if (EventDeoptimization::is_enabled()) {
1930 static bool serializers_registered = false;
1931 if (!serializers_registered) {
1932 register_serializers();
1933 serializers_registered = true;
1934 }
1935 EventDeoptimization event;
1936 event.set_compileId(nm->compile_id());
1937 event.set_compiler(nm->compiler_type());
1938 event.set_method(method);
1939 event.set_lineNumber(method->line_number_from_bci(trap_bci));
1940 event.set_bci(trap_bci);
1941 event.set_instruction(instruction);
1942 event.set_reason(reason);
1943 event.set_action(action);
1944 event.commit();
1945 }
1946 }
1947
1948 #endif // INCLUDE_JFR
1949
1950 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
1951 HandleMark hm(current);
1952
1953 // uncommon_trap() is called at the beginning of the uncommon trap
1954 // handler. Note this fact before we start generating temporary frames
1955 // that can confuse an asynchronous stack walker. This counter is
1956 // decremented at the end of unpack_frames().
1957 current->inc_in_deopt_handler();
1958
1959 // We need to update the map if we have biased locking.
1960 #if INCLUDE_JVMCI
1961 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1962 RegisterMap reg_map(current, true);
1963 #else
1964 RegisterMap reg_map(current, UseBiasedLocking);
1965 #endif
1966 frame stub_frame = current->last_frame();
1967 frame fr = stub_frame.sender(®_map);
1968 // Make sure the calling nmethod is not getting deoptimized and removed
1969 // before we are done with it.
1970 nmethodLocker nl(fr.pc());
1971
1972 // Log a message
1973 Events::log_deopt_message(current, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1974 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1975
1976 {
1977 ResourceMark rm;
1978
1979 DeoptReason reason = trap_request_reason(trap_request);
1980 DeoptAction action = trap_request_action(trap_request);
1981 #if INCLUDE_JVMCI
1982 int debug_id = trap_request_debug_id(trap_request);
1983 #endif
1984 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1985
1986 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
1987 compiledVFrame* cvf = compiledVFrame::cast(vf);
1988
1989 CompiledMethod* nm = cvf->code();
1990
1991 ScopeDesc* trap_scope = cvf->scope();
1992
1993 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
1994
1995 if (TraceDeoptimization || is_receiver_constraint_failure) {
1996 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"), trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
1997 #if INCLUDE_JVMCI
1998 , debug_id
1999 #endif
2000 );
2001 }
2002
2003 methodHandle trap_method(current, trap_scope->method());
2004 int trap_bci = trap_scope->bci();
2005 #if INCLUDE_JVMCI
2006 jlong speculation = current->pending_failed_speculation();
2007 if (nm->is_compiled_by_jvmci()) {
2008 nm->as_nmethod()->update_speculation(current);
2009 } else {
2010 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2011 }
2012
2013 if (trap_bci == SynchronizationEntryBCI) {
2014 trap_bci = 0;
2015 current->set_pending_monitorenter(true);
2016 }
2017
2018 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2019 current->set_pending_transfer_to_interpreter(true);
2020 }
2021 #endif
2022
2023 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2024 // Record this event in the histogram.
2025 gather_statistics(reason, action, trap_bc);
2026
2027 // Ensure that we can record deopt. history:
2028 // Need MDO to record RTM code generation state.
2029 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
2030
2031 methodHandle profiled_method;
2032 #if INCLUDE_JVMCI
2033 if (nm->is_compiled_by_jvmci()) {
2034 profiled_method = methodHandle(current, nm->method());
2035 } else {
2036 profiled_method = trap_method;
2037 }
2038 #else
2039 profiled_method = trap_method;
2040 #endif
2041
2042 MethodData* trap_mdo =
2043 get_method_data(current, profiled_method, create_if_missing);
2044
2045 JFR_ONLY(post_deoptimization_event(nm, trap_method(), trap_bci, trap_bc, reason, action);)
2046
2047 // Log a message
2048 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2049 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
2050 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2051
2052 // Print a bunch of diagnostics, if requested.
2053 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2054 ResourceMark rm;
2055 ttyLocker ttyl;
2056 char buf[100];
2057 if (xtty != NULL) {
2058 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
2059 os::current_thread_id(),
2060 format_trap_request(buf, sizeof(buf), trap_request));
2061 #if INCLUDE_JVMCI
2062 if (speculation != 0) {
2063 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2064 }
2065 #endif
2066 nm->log_identity(xtty);
2067 }
2068 Symbol* class_name = NULL;
2069 bool unresolved = false;
2070 if (unloaded_class_index >= 0) {
2071 constantPoolHandle constants (current, trap_method->constants());
2072 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2073 class_name = constants->klass_name_at(unloaded_class_index);
2074 unresolved = true;
2075 if (xtty != NULL)
2076 xtty->print(" unresolved='1'");
2077 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2078 class_name = constants->symbol_at(unloaded_class_index);
2079 }
2080 if (xtty != NULL)
2081 xtty->name(class_name);
2082 }
2083 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
2084 // Dump the relevant MDO state.
2085 // This is the deopt count for the current reason, any previous
2086 // reasons or recompiles seen at this point.
2087 int dcnt = trap_mdo->trap_count(reason);
2088 if (dcnt != 0)
2089 xtty->print(" count='%d'", dcnt);
2090 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2091 int dos = (pdata == NULL)? 0: pdata->trap_state();
2092 if (dos != 0) {
2093 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2094 if (trap_state_is_recompiled(dos)) {
2095 int recnt2 = trap_mdo->overflow_recompile_count();
2096 if (recnt2 != 0)
2097 xtty->print(" recompiles2='%d'", recnt2);
2098 }
2099 }
2100 }
2101 if (xtty != NULL) {
2102 xtty->stamp();
2103 xtty->end_head();
2104 }
2105 if (TraceDeoptimization) { // make noise on the tty
2106 tty->print("Uncommon trap occurred in");
2107 nm->method()->print_short_name(tty);
2108 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2109 #if INCLUDE_JVMCI
2110 if (nm->is_nmethod()) {
2111 const char* installed_code_name = nm->as_nmethod()->jvmci_name();
2112 if (installed_code_name != NULL) {
2113 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name);
2114 }
2115 }
2116 #endif
2117 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2118 p2i(fr.pc()),
2119 os::current_thread_id(),
2120 trap_reason_name(reason),
2121 trap_action_name(action),
2122 unloaded_class_index
2123 #if INCLUDE_JVMCI
2124 , debug_id
2125 #endif
2126 );
2127 if (class_name != NULL) {
2128 tty->print(unresolved ? " unresolved class: " : " symbol: ");
2129 class_name->print_symbol_on(tty);
2130 }
2131 tty->cr();
2132 }
2133 if (xtty != NULL) {
2134 // Log the precise location of the trap.
2135 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2136 xtty->begin_elem("jvms bci='%d'", sd->bci());
2137 xtty->method(sd->method());
2138 xtty->end_elem();
2139 if (sd->is_top()) break;
2140 }
2141 xtty->tail("uncommon_trap");
2142 }
2143 }
2144 // (End diagnostic printout.)
2145
2146 if (is_receiver_constraint_failure) {
2147 fatal("missing receiver type check");
2148 }
2149
2150 // Load class if necessary
2151 if (unloaded_class_index >= 0) {
2152 constantPoolHandle constants(current, trap_method->constants());
2153 load_class_by_index(constants, unloaded_class_index, THREAD);
2154 }
2155
2156 // Flush the nmethod if necessary and desirable.
2157 //
2158 // We need to avoid situations where we are re-flushing the nmethod
2159 // because of a hot deoptimization site. Repeated flushes at the same
2160 // point need to be detected by the compiler and avoided. If the compiler
2161 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2162 // module must take measures to avoid an infinite cycle of recompilation
2163 // and deoptimization. There are several such measures:
2164 //
2165 // 1. If a recompilation is ordered a second time at some site X
2166 // and for the same reason R, the action is adjusted to 'reinterpret',
2167 // to give the interpreter time to exercise the method more thoroughly.
2168 // If this happens, the method's overflow_recompile_count is incremented.
2169 //
2170 // 2. If the compiler fails to reduce the deoptimization rate, then
2171 // the method's overflow_recompile_count will begin to exceed the set
2172 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2173 // is adjusted to 'make_not_compilable', and the method is abandoned
2174 // to the interpreter. This is a performance hit for hot methods,
2175 // but is better than a disastrous infinite cycle of recompilations.
2176 // (Actually, only the method containing the site X is abandoned.)
2177 //
2178 // 3. In parallel with the previous measures, if the total number of
2179 // recompilations of a method exceeds the much larger set limit
2180 // PerMethodRecompilationCutoff, the method is abandoned.
2181 // This should only happen if the method is very large and has
2182 // many "lukewarm" deoptimizations. The code which enforces this
2183 // limit is elsewhere (class nmethod, class Method).
2184 //
2185 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2186 // to recompile at each bytecode independently of the per-BCI cutoff.
2187 //
2188 // The decision to update code is up to the compiler, and is encoded
2189 // in the Action_xxx code. If the compiler requests Action_none
2190 // no trap state is changed, no compiled code is changed, and the
2191 // computation suffers along in the interpreter.
2192 //
2193 // The other action codes specify various tactics for decompilation
2194 // and recompilation. Action_maybe_recompile is the loosest, and
2195 // allows the compiled code to stay around until enough traps are seen,
2196 // and until the compiler gets around to recompiling the trapping method.
2197 //
2198 // The other actions cause immediate removal of the present code.
2199
2200 // Traps caused by injected profile shouldn't pollute trap counts.
2201 bool injected_profile_trap = trap_method->has_injected_profile() &&
2202 (reason == Reason_intrinsic || reason == Reason_unreached);
2203
2204 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2205 bool make_not_entrant = false;
2206 bool make_not_compilable = false;
2207 bool reprofile = false;
2208 switch (action) {
2209 case Action_none:
2210 // Keep the old code.
2211 update_trap_state = false;
2212 break;
2213 case Action_maybe_recompile:
2214 // Do not need to invalidate the present code, but we can
2215 // initiate another
2216 // Start compiler without (necessarily) invalidating the nmethod.
2217 // The system will tolerate the old code, but new code should be
2218 // generated when possible.
2219 break;
2220 case Action_reinterpret:
2221 // Go back into the interpreter for a while, and then consider
2222 // recompiling form scratch.
2223 make_not_entrant = true;
2224 // Reset invocation counter for outer most method.
2225 // This will allow the interpreter to exercise the bytecodes
2226 // for a while before recompiling.
2227 // By contrast, Action_make_not_entrant is immediate.
2228 //
2229 // Note that the compiler will track null_check, null_assert,
2230 // range_check, and class_check events and log them as if they
2231 // had been traps taken from compiled code. This will update
2232 // the MDO trap history so that the next compilation will
2233 // properly detect hot trap sites.
2234 reprofile = true;
2235 break;
2236 case Action_make_not_entrant:
2237 // Request immediate recompilation, and get rid of the old code.
2238 // Make them not entrant, so next time they are called they get
2239 // recompiled. Unloaded classes are loaded now so recompile before next
2240 // time they are called. Same for uninitialized. The interpreter will
2241 // link the missing class, if any.
2242 make_not_entrant = true;
2243 break;
2244 case Action_make_not_compilable:
2245 // Give up on compiling this method at all.
2246 make_not_entrant = true;
2247 make_not_compilable = true;
2248 break;
2249 default:
2250 ShouldNotReachHere();
2251 }
2252
2253 // Setting +ProfileTraps fixes the following, on all platforms:
2254 // 4852688: ProfileInterpreter is off by default for ia64. The result is
2255 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2256 // recompile relies on a MethodData* to record heroic opt failures.
2257
2258 // Whether the interpreter is producing MDO data or not, we also need
2259 // to use the MDO to detect hot deoptimization points and control
2260 // aggressive optimization.
2261 bool inc_recompile_count = false;
2262 ProfileData* pdata = NULL;
2263 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != NULL) {
2264 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2265 uint this_trap_count = 0;
2266 bool maybe_prior_trap = false;
2267 bool maybe_prior_recompile = false;
2268 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2269 #if INCLUDE_JVMCI
2270 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2271 #endif
2272 nm->method(),
2273 //outputs:
2274 this_trap_count,
2275 maybe_prior_trap,
2276 maybe_prior_recompile);
2277 // Because the interpreter also counts null, div0, range, and class
2278 // checks, these traps from compiled code are double-counted.
2279 // This is harmless; it just means that the PerXTrapLimit values
2280 // are in effect a little smaller than they look.
2281
2282 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2283 if (per_bc_reason != Reason_none) {
2284 // Now take action based on the partially known per-BCI history.
2285 if (maybe_prior_trap
2286 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2287 // If there are too many traps at this BCI, force a recompile.
2288 // This will allow the compiler to see the limit overflow, and
2289 // take corrective action, if possible. The compiler generally
2290 // does not use the exact PerBytecodeTrapLimit value, but instead
2291 // changes its tactics if it sees any traps at all. This provides
2292 // a little hysteresis, delaying a recompile until a trap happens
2293 // several times.
2294 //
2295 // Actually, since there is only one bit of counter per BCI,
2296 // the possible per-BCI counts are {0,1,(per-method count)}.
2297 // This produces accurate results if in fact there is only
2298 // one hot trap site, but begins to get fuzzy if there are
2299 // many sites. For example, if there are ten sites each
2300 // trapping two or more times, they each get the blame for
2301 // all of their traps.
2302 make_not_entrant = true;
2303 }
2304
2305 // Detect repeated recompilation at the same BCI, and enforce a limit.
2306 if (make_not_entrant && maybe_prior_recompile) {
2307 // More than one recompile at this point.
2308 inc_recompile_count = maybe_prior_trap;
2309 }
2310 } else {
2311 // For reasons which are not recorded per-bytecode, we simply
2312 // force recompiles unconditionally.
2313 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2314 make_not_entrant = true;
2315 }
2316
2317 // Go back to the compiler if there are too many traps in this method.
2318 if (this_trap_count >= per_method_trap_limit(reason)) {
2319 // If there are too many traps in this method, force a recompile.
2320 // This will allow the compiler to see the limit overflow, and
2321 // take corrective action, if possible.
2322 // (This condition is an unlikely backstop only, because the
2323 // PerBytecodeTrapLimit is more likely to take effect first,
2324 // if it is applicable.)
2325 make_not_entrant = true;
2326 }
2327
2328 // Here's more hysteresis: If there has been a recompile at
2329 // this trap point already, run the method in the interpreter
2330 // for a while to exercise it more thoroughly.
2331 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2332 reprofile = true;
2333 }
2334 }
2335
2336 // Take requested actions on the method:
2337
2338 // Recompile
2339 if (make_not_entrant) {
2340 if (!nm->make_not_entrant()) {
2341 return; // the call did not change nmethod's state
2342 }
2343
2344 if (pdata != NULL) {
2345 // Record the recompilation event, if any.
2346 int tstate0 = pdata->trap_state();
2347 int tstate1 = trap_state_set_recompiled(tstate0, true);
2348 if (tstate1 != tstate0)
2349 pdata->set_trap_state(tstate1);
2350 }
2351
2352 #if INCLUDE_RTM_OPT
2353 // Restart collecting RTM locking abort statistic if the method
2354 // is recompiled for a reason other than RTM state change.
2355 // Assume that in new recompiled code the statistic could be different,
2356 // for example, due to different inlining.
2357 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
2358 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
2359 trap_mdo->atomic_set_rtm_state(ProfileRTM);
2360 }
2361 #endif
2362 // For code aging we count traps separately here, using make_not_entrant()
2363 // as a guard against simultaneous deopts in multiple threads.
2364 if (reason == Reason_tenured && trap_mdo != NULL) {
2365 trap_mdo->inc_tenure_traps();
2366 }
2367 }
2368
2369 if (inc_recompile_count) {
2370 trap_mdo->inc_overflow_recompile_count();
2371 if ((uint)trap_mdo->overflow_recompile_count() >
2372 (uint)PerBytecodeRecompilationCutoff) {
2373 // Give up on the method containing the bad BCI.
2374 if (trap_method() == nm->method()) {
2375 make_not_compilable = true;
2376 } else {
2377 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2378 // But give grace to the enclosing nm->method().
2379 }
2380 }
2381 }
2382
2383 // Reprofile
2384 if (reprofile) {
2385 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2386 }
2387
2388 // Give up compiling
2389 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2390 assert(make_not_entrant, "consistent");
2391 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2392 }
2393
2394 } // Free marked resources
2395
2396 }
2397 JRT_END
2398
2399 ProfileData*
2400 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2401 int trap_bci,
2402 Deoptimization::DeoptReason reason,
2403 bool update_total_trap_count,
2404 #if INCLUDE_JVMCI
2405 bool is_osr,
2406 #endif
2407 Method* compiled_method,
2408 //outputs:
2409 uint& ret_this_trap_count,
2410 bool& ret_maybe_prior_trap,
2411 bool& ret_maybe_prior_recompile) {
2412 bool maybe_prior_trap = false;
2413 bool maybe_prior_recompile = false;
2414 uint this_trap_count = 0;
2415 if (update_total_trap_count) {
2416 uint idx = reason;
2417 #if INCLUDE_JVMCI
2418 if (is_osr) {
2419 // Upper half of history array used for traps in OSR compilations
2420 idx += Reason_TRAP_HISTORY_LENGTH;
2421 }
2422 #endif
2423 uint prior_trap_count = trap_mdo->trap_count(idx);
2424 this_trap_count = trap_mdo->inc_trap_count(idx);
2425
2426 // If the runtime cannot find a place to store trap history,
2427 // it is estimated based on the general condition of the method.
2428 // If the method has ever been recompiled, or has ever incurred
2429 // a trap with the present reason , then this BCI is assumed
2430 // (pessimistically) to be the culprit.
2431 maybe_prior_trap = (prior_trap_count != 0);
2432 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2433 }
2434 ProfileData* pdata = NULL;
2435
2436
2437 // For reasons which are recorded per bytecode, we check per-BCI data.
2438 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2439 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2440 if (per_bc_reason != Reason_none) {
2441 // Find the profile data for this BCI. If there isn't one,
2442 // try to allocate one from the MDO's set of spares.
2443 // This will let us detect a repeated trap at this point.
2444 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
2445
2446 if (pdata != NULL) {
2447 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2448 if (LogCompilation && xtty != NULL) {
2449 ttyLocker ttyl;
2450 // no more room for speculative traps in this MDO
2451 xtty->elem("speculative_traps_oom");
2452 }
2453 }
2454 // Query the trap state of this profile datum.
2455 int tstate0 = pdata->trap_state();
2456 if (!trap_state_has_reason(tstate0, per_bc_reason))
2457 maybe_prior_trap = false;
2458 if (!trap_state_is_recompiled(tstate0))
2459 maybe_prior_recompile = false;
2460
2461 // Update the trap state of this profile datum.
2462 int tstate1 = tstate0;
2463 // Record the reason.
2464 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2465 // Store the updated state on the MDO, for next time.
2466 if (tstate1 != tstate0)
2467 pdata->set_trap_state(tstate1);
2468 } else {
2469 if (LogCompilation && xtty != NULL) {
2470 ttyLocker ttyl;
2471 // Missing MDP? Leave a small complaint in the log.
2472 xtty->elem("missing_mdp bci='%d'", trap_bci);
2473 }
2474 }
2475 }
2476
2477 // Return results:
2478 ret_this_trap_count = this_trap_count;
2479 ret_maybe_prior_trap = maybe_prior_trap;
2480 ret_maybe_prior_recompile = maybe_prior_recompile;
2481 return pdata;
2482 }
2483
2484 void
2485 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2486 ResourceMark rm;
2487 // Ignored outputs:
2488 uint ignore_this_trap_count;
2489 bool ignore_maybe_prior_trap;
2490 bool ignore_maybe_prior_recompile;
2491 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2492 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2493 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2494 query_update_method_data(trap_mdo, trap_bci,
2495 (DeoptReason)reason,
2496 update_total_counts,
2497 #if INCLUDE_JVMCI
2498 false,
2499 #endif
2500 NULL,
2501 ignore_this_trap_count,
2502 ignore_maybe_prior_trap,
2503 ignore_maybe_prior_recompile);
2504 }
2505
2506 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) {
2507 // Enable WXWrite: current function is called from methods compiled by C2 directly
2508 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2509
2510 if (TraceDeoptimization) {
2511 tty->print("Uncommon trap ");
2512 }
2513 // Still in Java no safepoints
2514 {
2515 // This enters VM and may safepoint
2516 uncommon_trap_inner(current, trap_request);
2517 }
2518 HandleMark hm(current);
2519 return fetch_unroll_info_helper(current, exec_mode);
2520 }
2521
2522 // Local derived constants.
2523 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2524 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2525 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2526
2527 //---------------------------trap_state_reason---------------------------------
2528 Deoptimization::DeoptReason
2529 Deoptimization::trap_state_reason(int trap_state) {
2530 // This assert provides the link between the width of DataLayout::trap_bits
2531 // and the encoding of "recorded" reasons. It ensures there are enough
2532 // bits to store all needed reasons in the per-BCI MDO profile.
2533 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2534 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2535 trap_state -= recompile_bit;
2536 if (trap_state == DS_REASON_MASK) {
2537 return Reason_many;
2538 } else {
2539 assert((int)Reason_none == 0, "state=0 => Reason_none");
2540 return (DeoptReason)trap_state;
2541 }
2542 }
2543 //-------------------------trap_state_has_reason-------------------------------
2544 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2545 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2546 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2547 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2548 trap_state -= recompile_bit;
2549 if (trap_state == DS_REASON_MASK) {
2550 return -1; // true, unspecifically (bottom of state lattice)
2551 } else if (trap_state == reason) {
2552 return 1; // true, definitely
2553 } else if (trap_state == 0) {
2554 return 0; // false, definitely (top of state lattice)
2555 } else {
2556 return 0; // false, definitely
2557 }
2558 }
2559 //-------------------------trap_state_add_reason-------------------------------
2560 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2561 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2562 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2563 trap_state -= recompile_bit;
2564 if (trap_state == DS_REASON_MASK) {
2565 return trap_state + recompile_bit; // already at state lattice bottom
2566 } else if (trap_state == reason) {
2567 return trap_state + recompile_bit; // the condition is already true
2568 } else if (trap_state == 0) {
2569 return reason + recompile_bit; // no condition has yet been true
2570 } else {
2571 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2572 }
2573 }
2574 //-----------------------trap_state_is_recompiled------------------------------
2575 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2576 return (trap_state & DS_RECOMPILE_BIT) != 0;
2577 }
2578 //-----------------------trap_state_set_recompiled-----------------------------
2579 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2580 if (z) return trap_state | DS_RECOMPILE_BIT;
2581 else return trap_state & ~DS_RECOMPILE_BIT;
2582 }
2583 //---------------------------format_trap_state---------------------------------
2584 // This is used for debugging and diagnostics, including LogFile output.
2585 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2586 int trap_state) {
2587 assert(buflen > 0, "sanity");
2588 DeoptReason reason = trap_state_reason(trap_state);
2589 bool recomp_flag = trap_state_is_recompiled(trap_state);
2590 // Re-encode the state from its decoded components.
2591 int decoded_state = 0;
2592 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2593 decoded_state = trap_state_add_reason(decoded_state, reason);
2594 if (recomp_flag)
2595 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2596 // If the state re-encodes properly, format it symbolically.
2597 // Because this routine is used for debugging and diagnostics,
2598 // be robust even if the state is a strange value.
2599 size_t len;
2600 if (decoded_state != trap_state) {
2601 // Random buggy state that doesn't decode??
2602 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2603 } else {
2604 len = jio_snprintf(buf, buflen, "%s%s",
2605 trap_reason_name(reason),
2606 recomp_flag ? " recompiled" : "");
2607 }
2608 return buf;
2609 }
2610
2611
2612 //--------------------------------statics--------------------------------------
2613 const char* Deoptimization::_trap_reason_name[] = {
2614 // Note: Keep this in sync. with enum DeoptReason.
2615 "none",
2616 "null_check",
2617 "null_assert" JVMCI_ONLY("_or_unreached0"),
2618 "range_check",
2619 "class_check",
2620 "array_check",
2621 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2622 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2623 "profile_predicate",
2624 "unloaded",
2625 "uninitialized",
2626 "initialized",
2627 "unreached",
2628 "unhandled",
2629 "constraint",
2630 "div0_check",
2631 "age",
2632 "predicate",
2633 "loop_limit_check",
2634 "speculate_class_check",
2635 "speculate_null_check",
2636 "speculate_null_assert",
2637 "rtm_state_change",
2638 "unstable_if",
2639 "unstable_fused_if",
2640 "receiver_constraint",
2641 #if INCLUDE_JVMCI
2642 "aliasing",
2643 "transfer_to_interpreter",
2644 "not_compiled_exception_handler",
2645 "unresolved",
2646 "jsr_mismatch",
2647 #endif
2648 "tenured"
2649 };
2650 const char* Deoptimization::_trap_action_name[] = {
2651 // Note: Keep this in sync. with enum DeoptAction.
2652 "none",
2653 "maybe_recompile",
2654 "reinterpret",
2655 "make_not_entrant",
2656 "make_not_compilable"
2657 };
2658
2659 const char* Deoptimization::trap_reason_name(int reason) {
2660 // Check that every reason has a name
2661 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2662
2663 if (reason == Reason_many) return "many";
2664 if ((uint)reason < Reason_LIMIT)
2665 return _trap_reason_name[reason];
2666 static char buf[20];
2667 sprintf(buf, "reason%d", reason);
2668 return buf;
2669 }
2670 const char* Deoptimization::trap_action_name(int action) {
2671 // Check that every action has a name
2672 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2673
2674 if ((uint)action < Action_LIMIT)
2675 return _trap_action_name[action];
2676 static char buf[20];
2677 sprintf(buf, "action%d", action);
2678 return buf;
2679 }
2680
2681 // This is used for debugging and diagnostics, including LogFile output.
2682 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2683 int trap_request) {
2684 jint unloaded_class_index = trap_request_index(trap_request);
2685 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2686 const char* action = trap_action_name(trap_request_action(trap_request));
2687 #if INCLUDE_JVMCI
2688 int debug_id = trap_request_debug_id(trap_request);
2689 #endif
2690 size_t len;
2691 if (unloaded_class_index < 0) {
2692 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2693 reason, action
2694 #if INCLUDE_JVMCI
2695 ,debug_id
2696 #endif
2697 );
2698 } else {
2699 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2700 reason, action, unloaded_class_index
2701 #if INCLUDE_JVMCI
2702 ,debug_id
2703 #endif
2704 );
2705 }
2706 return buf;
2707 }
2708
2709 juint Deoptimization::_deoptimization_hist
2710 [Deoptimization::Reason_LIMIT]
2711 [1 + Deoptimization::Action_LIMIT]
2712 [Deoptimization::BC_CASE_LIMIT]
2713 = {0};
2714
2715 enum {
2716 LSB_BITS = 8,
2717 LSB_MASK = right_n_bits(LSB_BITS)
2718 };
2719
2720 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2721 Bytecodes::Code bc) {
2722 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2723 assert(action >= 0 && action < Action_LIMIT, "oob");
2724 _deoptimization_hist[Reason_none][0][0] += 1; // total
2725 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2726 juint* cases = _deoptimization_hist[reason][1+action];
2727 juint* bc_counter_addr = NULL;
2728 juint bc_counter = 0;
2729 // Look for an unused counter, or an exact match to this BC.
2730 if (bc != Bytecodes::_illegal) {
2731 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2732 juint* counter_addr = &cases[bc_case];
2733 juint counter = *counter_addr;
2734 if ((counter == 0 && bc_counter_addr == NULL)
2735 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2736 // this counter is either free or is already devoted to this BC
2737 bc_counter_addr = counter_addr;
2738 bc_counter = counter | bc;
2739 }
2740 }
2741 }
2742 if (bc_counter_addr == NULL) {
2743 // Overflow, or no given bytecode.
2744 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2745 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2746 }
2747 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2748 }
2749
2750 jint Deoptimization::total_deoptimization_count() {
2751 return _deoptimization_hist[Reason_none][0][0];
2752 }
2753
2754 void Deoptimization::print_statistics() {
2755 juint total = total_deoptimization_count();
2756 juint account = total;
2757 if (total != 0) {
2758 ttyLocker ttyl;
2759 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2760 tty->print_cr("Deoptimization traps recorded:");
2761 #define PRINT_STAT_LINE(name, r) \
2762 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2763 PRINT_STAT_LINE("total", total);
2764 // For each non-zero entry in the histogram, print the reason,
2765 // the action, and (if specifically known) the type of bytecode.
2766 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2767 for (int action = 0; action < Action_LIMIT; action++) {
2768 juint* cases = _deoptimization_hist[reason][1+action];
2769 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2770 juint counter = cases[bc_case];
2771 if (counter != 0) {
2772 char name[1*K];
2773 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2774 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2775 bc = Bytecodes::_illegal;
2776 sprintf(name, "%s/%s/%s",
2777 trap_reason_name(reason),
2778 trap_action_name(action),
2779 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2780 juint r = counter >> LSB_BITS;
2781 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2782 account -= r;
2783 }
2784 }
2785 }
2786 }
2787 if (account != 0) {
2788 PRINT_STAT_LINE("unaccounted", account);
2789 }
2790 #undef PRINT_STAT_LINE
2791 if (xtty != NULL) xtty->tail("statistics");
2792 }
2793 }
2794
2795 #else // COMPILER2_OR_JVMCI
2796
2797
2798 // Stubs for C1 only system.
2799 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2800 return false;
2801 }
2802
2803 const char* Deoptimization::trap_reason_name(int reason) {
2804 return "unknown";
2805 }
2806
2807 void Deoptimization::print_statistics() {
2808 // no output
2809 }
2810
2811 void
2812 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2813 // no udpate
2814 }
2815
2816 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2817 return 0;
2818 }
2819
2820 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2821 Bytecodes::Code bc) {
2822 // no update
2823 }
2824
2825 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2826 int trap_state) {
2827 jio_snprintf(buf, buflen, "#%d", trap_state);
2828 return buf;
2829 }
2830
2831 #endif // COMPILER2_OR_JVMCI