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