1
2
3 /*
4 * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved.
5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 *
7 * This code is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 only, as
9 * published by the Free Software Foundation.
10 *
11 * This code is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 * version 2 for more details (a copy is included in the LICENSE file that
15 * accompanied this code).
16 *
17 * You should have received a copy of the GNU General Public License version
18 * 2 along with this work; if not, write to the Free Software Foundation,
19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 *
21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22 * or visit www.oracle.com if you need additional information or have any
23 * questions.
24 *
25 */
26
27 #include "precompiled.hpp"
28 #include "jvm.h"
29 #include "classfile/javaClasses.inline.hpp"
30 #include "classfile/symbolTable.hpp"
31 #include "classfile/systemDictionary.hpp"
32 #include "code/codeCache.hpp"
33 #include "code/debugInfoRec.hpp"
34 #include "code/nmethod.hpp"
35 #include "code/pcDesc.hpp"
36 #include "code/scopeDesc.hpp"
37 #include "compiler/compilationPolicy.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/jvmtiThreadState.hpp"
54 #include "runtime/atomic.hpp"
55 #include "runtime/biasedLocking.hpp"
56 #include "runtime/deoptimization.hpp"
57 #include "runtime/fieldDescriptor.hpp"
58 #include "runtime/fieldDescriptor.inline.hpp"
59 #include "runtime/frame.inline.hpp"
60 #include "runtime/handles.inline.hpp"
61 #include "runtime/interfaceSupport.inline.hpp"
62 #include "runtime/jniHandles.inline.hpp"
63 #include "runtime/objectMonitor.inline.hpp"
64 #include "runtime/safepointVerifiers.hpp"
65 #include "runtime/sharedRuntime.hpp"
66 #include "runtime/signature.hpp"
67 #include "runtime/stubRoutines.hpp"
68 #include "runtime/thread.hpp"
69 #include "runtime/threadSMR.hpp"
70 #include "runtime/vframe.hpp"
71 #include "runtime/vframeArray.hpp"
72 #include "runtime/vframe_hp.hpp"
73 #include "utilities/events.hpp"
74 #include "utilities/macros.hpp"
75 #include "utilities/preserveException.hpp"
76 #include "utilities/xmlstream.hpp"
77 #if INCLUDE_JFR
78 #include "jfr/jfrEvents.hpp"
79 #include "jfr/metadata/jfrSerializer.hpp"
80 #endif
81
82 bool DeoptimizationMarker::_is_active = false;
83
84 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
85 int caller_adjustment,
86 int caller_actual_parameters,
87 int number_of_frames,
88 intptr_t* frame_sizes,
89 address* frame_pcs,
90 BasicType return_type,
91 int exec_mode) {
92 _size_of_deoptimized_frame = size_of_deoptimized_frame;
93 _caller_adjustment = caller_adjustment;
94 _caller_actual_parameters = caller_actual_parameters;
95 _number_of_frames = number_of_frames;
96 _frame_sizes = frame_sizes;
97 _frame_pcs = frame_pcs;
98 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
99 _return_type = return_type;
100 _initial_info = 0;
101 // PD (x86 only)
102 _counter_temp = 0;
103 _unpack_kind = exec_mode;
104 _sender_sp_temp = 0;
105
106 _total_frame_sizes = size_of_frames();
107 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
108 }
109
110
111 Deoptimization::UnrollBlock::~UnrollBlock() {
112 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
113 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
114 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
115 }
116
117
118 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
119 assert(register_number < RegisterMap::reg_count, "checking register number");
120 return &_register_block[register_number * 2];
121 }
122
123
124
125 int Deoptimization::UnrollBlock::size_of_frames() const {
126 // Acount first for the adjustment of the initial frame
127 int result = _caller_adjustment;
128 for (int index = 0; index < number_of_frames(); index++) {
129 result += frame_sizes()[index];
130 }
131 return result;
132 }
133
134
135 void Deoptimization::UnrollBlock::print() {
136 ttyLocker ttyl;
137 tty->print_cr("UnrollBlock");
138 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
139 tty->print( " frame_sizes: ");
140 for (int index = 0; index < number_of_frames(); index++) {
141 tty->print(INTX_FORMAT " ", frame_sizes()[index]);
142 }
143 tty->cr();
144 }
145
146
147 // In order to make fetch_unroll_info work properly with escape
148 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
149 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
150 // of previously eliminated objects occurs in realloc_objects, which is
151 // called from the method fetch_unroll_info_helper below.
152 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode))
153 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
154 // but makes the entry a little slower. There is however a little dance we have to
155 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
156
157 // fetch_unroll_info() is called at the beginning of the deoptimization
158 // handler. Note this fact before we start generating temporary frames
159 // that can confuse an asynchronous stack walker. This counter is
160 // decremented at the end of unpack_frames().
161 if (TraceDeoptimization) {
162 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread));
163 }
164 thread->inc_in_deopt_handler();
165
166 return fetch_unroll_info_helper(thread, exec_mode);
167 JRT_END
168
169 #if COMPILER2_OR_JVMCI
170 static bool eliminate_allocations(JavaThread* thread, int exec_mode, CompiledMethod* compiled_method,
171 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
172 bool& deoptimized_objects) {
173 bool realloc_failures = false;
174 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
175
176 JavaThread* deoptee_thread = chunk->at(0)->thread();
177 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread), "a frame can only be deoptimized by the owner thread");
178
179 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
180
181 // The flag return_oop() indicates call sites which return oop
182 // in compiled code. Such sites include java method calls,
183 // runtime calls (for example, used to allocate new objects/arrays
184 // on slow code path) and any other calls generated in compiled code.
185 // It is not guaranteed that we can get such information here only
186 // by analyzing bytecode in deoptimized frames. This is why this flag
187 // is set during method compilation (see Compile::Process_OopMap_Node()).
188 // If the previous frame was popped or if we are dispatching an exception,
189 // we don't have an oop result.
190 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
191 Handle return_value;
192 if (save_oop_result) {
193 // Reallocation may trigger GC. If deoptimization happened on return from
194 // call which returns oop we need to save it since it is not in oopmap.
195 oop result = deoptee.saved_oop_result(&map);
196 assert(oopDesc::is_oop_or_null(result), "must be oop");
197 return_value = Handle(thread, result);
198 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
199 if (TraceDeoptimization) {
200 ttyLocker ttyl;
201 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
202 }
203 }
204 if (objects != NULL) {
205 if (exec_mode == Deoptimization::Unpack_none) {
206 assert(thread->thread_state() == _thread_in_vm, "assumption");
207 Thread* THREAD = thread;
208 // Clear pending OOM if reallocation fails and return true indicating allocation failure
209 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
210 deoptimized_objects = true;
211 } else {
212 JRT_BLOCK
213 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
214 JRT_END
215 }
216 bool skip_internal = (compiled_method != NULL) && !compiled_method->is_compiled_by_jvmci();
217 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
218 #ifndef PRODUCT
219 if (TraceDeoptimization) {
220 ttyLocker ttyl;
221 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
222 Deoptimization::print_objects(objects, realloc_failures);
223 }
224 #endif
225 }
226 if (save_oop_result) {
227 // Restore result.
228 deoptee.set_saved_oop_result(&map, return_value());
229 }
230 return realloc_failures;
231 }
232
233 static void eliminate_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
234 frame& deoptee, int exec_mode, bool& deoptimized_objects) {
235 JavaThread* deoptee_thread = chunk->at(0)->thread();
236 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
237
238 #ifndef PRODUCT
239 bool first = true;
240 #endif
241 for (int i = 0; i < chunk->length(); i++) {
242 compiledVFrame* cvf = chunk->at(i);
243 assert (cvf->scope() != NULL,"expect only compiled java frames");
244 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
245 if (monitors->is_nonempty()) {
246 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee, exec_mode, realloc_failures);
247 deoptimized_objects = deoptimized_objects || relocked;
248 #ifndef PRODUCT
249 if (PrintDeoptimizationDetails) {
250 ttyLocker ttyl;
251 for (int j = 0; j < monitors->length(); j++) {
252 MonitorInfo* mi = monitors->at(j);
253 if (mi->eliminated()) {
254 if (first) {
255 first = false;
256 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
257 }
258 if (exec_mode == Deoptimization::Unpack_none) {
259 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
260 if (monitor != NULL && (oop)monitor->object() == mi->owner()) {
261 tty->print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
262 continue;
263 }
264 }
265 if (mi->owner_is_scalar_replaced()) {
266 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
267 tty->print_cr(" failed reallocation for klass %s", k->external_name());
268 } else {
269 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
270 }
271 }
272 }
273 }
274 #endif // !PRODUCT
275 }
276 }
277 }
278
279 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
280 // The given vframes cover one physical frame.
281 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool& realloc_failures) {
282 frame deoptee = chunk->at(0)->fr();
283 JavaThread* deoptee_thread = chunk->at(0)->thread();
284 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
285 RegisterMap map(chunk->at(0)->register_map());
286 bool deoptimized_objects = false;
287
288 // Reallocate the non-escaping objects and restore their fields. Then
289 // relock objects if synchronization on them was eliminated.
290 if (DoEscapeAnalysis && EliminateAllocations) {
291 realloc_failures = eliminate_allocations(thread, Unpack_none, cm, deoptee, map, chunk, deoptimized_objects);
292 }
293
294 // Revoke biases of objects with eliminated locks in the given frame.
295 Deoptimization::revoke_for_object_deoptimization(deoptee_thread, deoptee, &map, thread);
296
297 // MonitorInfo structures used in eliminate_locks are not GC safe.
298 NoSafepointVerifier no_safepoint;
299
300 if ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) {
301 eliminate_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
302 }
303 return deoptimized_objects;
304 }
305 #endif // COMPILER2_OR_JVMCI
306
307 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
308 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) {
309
310 // Note: there is a safepoint safety issue here. No matter whether we enter
311 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
312 // the vframeArray is created.
313 //
314
315 // Allocate our special deoptimization ResourceMark
316 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
317 assert(thread->deopt_mark() == NULL, "Pending deopt!");
318 thread->set_deopt_mark(dmark);
319
320 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
321 RegisterMap map(thread, true);
322 RegisterMap dummy_map(thread, false);
323 // Now get the deoptee with a valid map
324 frame deoptee = stub_frame.sender(&map);
325 // Set the deoptee nmethod
326 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!");
327 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
328 thread->set_deopt_compiled_method(cm);
329
330 if (VerifyStack) {
331 thread->validate_frame_layout();
332 }
333
334 // Create a growable array of VFrames where each VFrame represents an inlined
335 // Java frame. This storage is allocated with the usual system arena.
336 assert(deoptee.is_compiled_frame(), "Wrong frame type");
337 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
338 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
339 while (!vf->is_top()) {
340 assert(vf->is_compiled_frame(), "Wrong frame type");
341 chunk->push(compiledVFrame::cast(vf));
342 vf = vf->sender();
343 }
344 assert(vf->is_compiled_frame(), "Wrong frame type");
345 chunk->push(compiledVFrame::cast(vf));
346
347 bool realloc_failures = false;
348
349 #if COMPILER2_OR_JVMCI
350 #if INCLUDE_JVMCI
351 bool jvmci_enabled = true;
352 #else
353 bool jvmci_enabled = false;
354 #endif
355
356 // Reallocate the non-escaping objects and restore their fields. Then
357 // relock objects if synchronization on them was eliminated.
358 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations) )) {
359 bool unused;
360 realloc_failures = eliminate_allocations(thread, exec_mode, cm, deoptee, map, chunk, unused);
361 }
362 #endif // COMPILER2_OR_JVMCI
363
364 // Revoke biases, done with in java state.
365 // No safepoints allowed after this
366 revoke_from_deopt_handler(thread, deoptee, &map);
367
368 // Ensure that no safepoint is taken after pointers have been stored
369 // in fields of rematerialized objects. If a safepoint occurs from here on
370 // out the java state residing in the vframeArray will be missed.
371 // Locks may be rebaised in a safepoint.
372 NoSafepointVerifier no_safepoint;
373
374 #if COMPILER2_OR_JVMCI
375 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
376 && !EscapeBarrier::objs_are_deoptimized(thread, deoptee.id())) {
377 bool unused;
378 eliminate_locks(thread, chunk, realloc_failures, deoptee, exec_mode, unused);
379 }
380 #endif // COMPILER2_OR_JVMCI
381
382 ScopeDesc* trap_scope = chunk->at(0)->scope();
383 Handle exceptionObject;
384 if (trap_scope->rethrow_exception()) {
385 if (PrintDeoptimizationDetails) {
386 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());
387 }
388 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
389 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw");
390 ScopeValue* topOfStack = expressions->top();
391 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
392 guarantee(exceptionObject() != NULL, "exception oop can not be null");
393 }
394
395 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
396 #if COMPILER2_OR_JVMCI
397 if (realloc_failures) {
398 pop_frames_failed_reallocs(thread, array);
399 }
400 #endif
401
402 assert(thread->vframe_array_head() == NULL, "Pending deopt!");
403 thread->set_vframe_array_head(array);
404
405 // Now that the vframeArray has been created if we have any deferred local writes
406 // added by jvmti then we can free up that structure as the data is now in the
407 // vframeArray
408
409 if (JvmtiDeferredUpdates::deferred_locals(thread) != NULL) {
410 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = JvmtiDeferredUpdates::deferred_locals(thread);
411 int i = 0;
412 do {
413 // Because of inlining we could have multiple vframes for a single frame
414 // and several of the vframes could have deferred writes. Find them all.
415 if (list->at(i)->id() == array->original().id()) {
416 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
417 list->remove_at(i);
418 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
419 delete dlv;
420 } else {
421 i++;
422 }
423 } while ( i < list->length() );
424 if (list->length() == 0) {
425 JvmtiDeferredUpdates* updates = thread->deferred_updates();
426 thread->set_deferred_updates(NULL);
427 // free deferred updates.
428 delete updates;
429 }
430
431 }
432
433 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
434 CodeBlob* cb = stub_frame.cb();
435 // Verify we have the right vframeArray
436 assert(cb->frame_size() >= 0, "Unexpected frame size");
437 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
438
439 // If the deopt call site is a MethodHandle invoke call site we have
440 // to adjust the unpack_sp.
441 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
442 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
443 unpack_sp = deoptee.unextended_sp();
444
445 #ifdef ASSERT
446 assert(cb->is_deoptimization_stub() ||
447 cb->is_uncommon_trap_stub() ||
448 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
449 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
450 "unexpected code blob: %s", cb->name());
451 #endif
452
453 // This is a guarantee instead of an assert because if vframe doesn't match
454 // we will unpack the wrong deoptimized frame and wind up in strange places
455 // where it will be very difficult to figure out what went wrong. Better
456 // to die an early death here than some very obscure death later when the
457 // trail is cold.
458 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
459 // in that it will fail to detect a problem when there is one. This needs
460 // more work in tiger timeframe.
461 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
462
463 int number_of_frames = array->frames();
464
465 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
466 // virtual activation, which is the reverse of the elements in the vframes array.
467 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
468 // +1 because we always have an interpreter return address for the final slot.
469 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
470 int popframe_extra_args = 0;
471 // Create an interpreter return address for the stub to use as its return
472 // address so the skeletal frames are perfectly walkable
473 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
474
475 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
476 // activation be put back on the expression stack of the caller for reexecution
477 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
478 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
479 }
480
481 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
482 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
483 // than simply use array->sender.pc(). This requires us to walk the current set of frames
484 //
485 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
486 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
487
488 // It's possible that the number of parameters at the call site is
489 // different than number of arguments in the callee when method
490 // handles are used. If the caller is interpreted get the real
491 // value so that the proper amount of space can be added to it's
492 // frame.
493 bool caller_was_method_handle = false;
494 if (deopt_sender.is_interpreted_frame()) {
495 methodHandle method(thread, deopt_sender.interpreter_frame_method());
496 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
497 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
498 // Method handle invokes may involve fairly arbitrary chains of
499 // calls so it's impossible to know how much actual space the
500 // caller has for locals.
501 caller_was_method_handle = true;
502 }
503 }
504
505 //
506 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
507 // frame_sizes/frame_pcs[1] next oldest frame (int)
508 // frame_sizes/frame_pcs[n] youngest frame (int)
509 //
510 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
511 // owns the space for the return address to it's caller). Confusing ain't it.
512 //
513 // The vframe array can address vframes with indices running from
514 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
515 // When we create the skeletal frames we need the oldest frame to be in the zero slot
516 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
517 // so things look a little strange in this loop.
518 //
519 int callee_parameters = 0;
520 int callee_locals = 0;
521 for (int index = 0; index < array->frames(); index++ ) {
522 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
523 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
524 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
525 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
526 callee_locals,
527 index == 0,
528 popframe_extra_args);
529 // This pc doesn't have to be perfect just good enough to identify the frame
530 // as interpreted so the skeleton frame will be walkable
531 // The correct pc will be set when the skeleton frame is completely filled out
532 // The final pc we store in the loop is wrong and will be overwritten below
533 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
534
535 callee_parameters = array->element(index)->method()->size_of_parameters();
536 callee_locals = array->element(index)->method()->max_locals();
537 popframe_extra_args = 0;
538 }
539
540 // Compute whether the root vframe returns a float or double value.
541 BasicType return_type;
542 {
543 methodHandle method(thread, array->element(0)->method());
544 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
545 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
546 }
547
548 // Compute information for handling adapters and adjusting the frame size of the caller.
549 int caller_adjustment = 0;
550
551 // Compute the amount the oldest interpreter frame will have to adjust
552 // its caller's stack by. If the caller is a compiled frame then
553 // we pretend that the callee has no parameters so that the
554 // extension counts for the full amount of locals and not just
555 // locals-parms. This is because without a c2i adapter the parm
556 // area as created by the compiled frame will not be usable by
557 // the interpreter. (Depending on the calling convention there
558 // may not even be enough space).
559
560 // QQQ I'd rather see this pushed down into last_frame_adjust
561 // and have it take the sender (aka caller).
562
563 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
564 caller_adjustment = last_frame_adjust(0, callee_locals);
565 } else if (callee_locals > callee_parameters) {
566 // The caller frame may need extending to accommodate
567 // non-parameter locals of the first unpacked interpreted frame.
568 // Compute that adjustment.
569 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
570 }
571
572 // If the sender is deoptimized the we must retrieve the address of the handler
573 // since the frame will "magically" show the original pc before the deopt
574 // and we'd undo the deopt.
575
576 frame_pcs[0] = deopt_sender.raw_pc();
577
578 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
579
580 #if INCLUDE_JVMCI
581 if (exceptionObject() != NULL) {
582 thread->set_exception_oop(exceptionObject());
583 exec_mode = Unpack_exception;
584 }
585 #endif
586
587 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
588 assert(thread->has_pending_exception(), "should have thrown OOME");
589 thread->set_exception_oop(thread->pending_exception());
590 thread->clear_pending_exception();
591 exec_mode = Unpack_exception;
592 }
593
594 #if INCLUDE_JVMCI
595 if (thread->frames_to_pop_failed_realloc() > 0) {
596 thread->set_pending_monitorenter(false);
597 }
598 #endif
599
600 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
601 caller_adjustment * BytesPerWord,
602 caller_was_method_handle ? 0 : callee_parameters,
603 number_of_frames,
604 frame_sizes,
605 frame_pcs,
606 return_type,
607 exec_mode);
608 // On some platforms, we need a way to pass some platform dependent
609 // information to the unpacking code so the skeletal frames come out
610 // correct (initial fp value, unextended sp, ...)
611 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
612
613 if (array->frames() > 1) {
614 if (VerifyStack && TraceDeoptimization) {
615 ttyLocker ttyl;
616 tty->print_cr("Deoptimizing method containing inlining");
617 }
618 }
619
620 array->set_unroll_block(info);
621 return info;
622 }
623
624 // Called to cleanup deoptimization data structures in normal case
625 // after unpacking to stack and when stack overflow error occurs
626 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
627 vframeArray *array) {
628
629 // Get array if coming from exception
630 if (array == NULL) {
631 array = thread->vframe_array_head();
632 }
633 thread->set_vframe_array_head(NULL);
634
635 // Free the previous UnrollBlock
636 vframeArray* old_array = thread->vframe_array_last();
637 thread->set_vframe_array_last(array);
638
639 if (old_array != NULL) {
640 UnrollBlock* old_info = old_array->unroll_block();
641 old_array->set_unroll_block(NULL);
642 delete old_info;
643 delete old_array;
644 }
645
646 // Deallocate any resource creating in this routine and any ResourceObjs allocated
647 // inside the vframeArray (StackValueCollections)
648
649 delete thread->deopt_mark();
650 thread->set_deopt_mark(NULL);
651 thread->set_deopt_compiled_method(NULL);
652
653
654 if (JvmtiExport::can_pop_frame()) {
655 // Regardless of whether we entered this routine with the pending
656 // popframe condition bit set, we should always clear it now
657 thread->clear_popframe_condition();
658 }
659
660 // unpack_frames() is called at the end of the deoptimization handler
661 // and (in C2) at the end of the uncommon trap handler. Note this fact
662 // so that an asynchronous stack walker can work again. This counter is
663 // incremented at the beginning of fetch_unroll_info() and (in C2) at
664 // the beginning of uncommon_trap().
665 thread->dec_in_deopt_handler();
666 }
667
668 // Moved from cpu directories because none of the cpus has callee save values.
669 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
670 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
671
672 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
673 // the days we had adapter frames. When we deoptimize a situation where a
674 // compiled caller calls a compiled caller will have registers it expects
675 // to survive the call to the callee. If we deoptimize the callee the only
676 // way we can restore these registers is to have the oldest interpreter
677 // frame that we create restore these values. That is what this routine
678 // will accomplish.
679
680 // At the moment we have modified c2 to not have any callee save registers
681 // so this problem does not exist and this routine is just a place holder.
682
683 assert(f->is_interpreted_frame(), "must be interpreted");
684 }
685
686 // Return BasicType of value being returned
687 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
688
689 // We are already active in the special DeoptResourceMark any ResourceObj's we
690 // allocate will be freed at the end of the routine.
691
692 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
693 // but makes the entry a little slower. There is however a little dance we have to
694 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
695 ResetNoHandleMark rnhm; // No-op in release/product versions
696 HandleMark hm;
697
698 frame stub_frame = thread->last_frame();
699
700 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
701 // must point to the vframeArray for the unpack frame.
702 vframeArray* array = thread->vframe_array_head();
703
704 #ifndef PRODUCT
705 if (TraceDeoptimization) {
706 ttyLocker ttyl;
707 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d",
708 p2i(thread), p2i(array), exec_mode);
709 }
710 #endif
711 Events::log_deopt_message(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
712 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode);
713
714 UnrollBlock* info = array->unroll_block();
715
716 // We set the last_Java frame. But the stack isn't really parsable here. So we
717 // clear it to make sure JFR understands not to try and walk stacks from events
718 // in here.
719 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
720 thread->frame_anchor()->set_last_Java_sp(NULL);
721
722 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
723 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
724
725 thread->frame_anchor()->set_last_Java_sp(sp);
726
727 BasicType bt = info->return_type();
728
729 // If we have an exception pending, claim that the return type is an oop
730 // so the deopt_blob does not overwrite the exception_oop.
731
732 if (exec_mode == Unpack_exception)
733 bt = T_OBJECT;
734
735 // Cleanup thread deopt data
736 cleanup_deopt_info(thread, array);
737
738 #ifndef PRODUCT
739 if (VerifyStack) {
740 ResourceMark res_mark;
741 // Clear pending exception to not break verification code (restored afterwards)
742 PRESERVE_EXCEPTION_MARK;
743
744 thread->validate_frame_layout();
745
746 // Verify that the just-unpacked frames match the interpreter's
747 // notions of expression stack and locals
748 vframeArray* cur_array = thread->vframe_array_last();
749 RegisterMap rm(thread, false);
750 rm.set_include_argument_oops(false);
751 bool is_top_frame = true;
752 int callee_size_of_parameters = 0;
753 int callee_max_locals = 0;
754 for (int i = 0; i < cur_array->frames(); i++) {
755 vframeArrayElement* el = cur_array->element(i);
756 frame* iframe = el->iframe();
757 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
758
759 // Get the oop map for this bci
760 InterpreterOopMap mask;
761 int cur_invoke_parameter_size = 0;
762 bool try_next_mask = false;
763 int next_mask_expression_stack_size = -1;
764 int top_frame_expression_stack_adjustment = 0;
765 methodHandle mh(thread, iframe->interpreter_frame_method());
766 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
767 BytecodeStream str(mh, iframe->interpreter_frame_bci());
768 int max_bci = mh->code_size();
769 // Get to the next bytecode if possible
770 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
771 // Check to see if we can grab the number of outgoing arguments
772 // at an uncommon trap for an invoke (where the compiler
773 // generates debug info before the invoke has executed)
774 Bytecodes::Code cur_code = str.next();
775 if (Bytecodes::is_invoke(cur_code)) {
776 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
777 cur_invoke_parameter_size = invoke.size_of_parameters();
778 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
779 callee_size_of_parameters++;
780 }
781 }
782 if (str.bci() < max_bci) {
783 Bytecodes::Code next_code = str.next();
784 if (next_code >= 0) {
785 // The interpreter oop map generator reports results before
786 // the current bytecode has executed except in the case of
787 // calls. It seems to be hard to tell whether the compiler
788 // has emitted debug information matching the "state before"
789 // a given bytecode or the state after, so we try both
790 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) {
791 // Get expression stack size for the next bytecode
792 InterpreterOopMap next_mask;
793 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
794 next_mask_expression_stack_size = next_mask.expression_stack_size();
795 if (Bytecodes::is_invoke(next_code)) {
796 Bytecode_invoke invoke(mh, str.bci());
797 next_mask_expression_stack_size += invoke.size_of_parameters();
798 }
799 // Need to subtract off the size of the result type of
800 // the bytecode because this is not described in the
801 // debug info but returned to the interpreter in the TOS
802 // caching register
803 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
804 if (bytecode_result_type != T_ILLEGAL) {
805 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
806 }
807 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
808 try_next_mask = true;
809 }
810 }
811 }
812
813 // Verify stack depth and oops in frame
814 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
815 if (!(
816 /* SPARC */
817 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
818 /* x86 */
819 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
820 (try_next_mask &&
821 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
822 top_frame_expression_stack_adjustment))) ||
823 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
824 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
825 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
826 )) {
827 {
828 ttyLocker ttyl;
829
830 // Print out some information that will help us debug the problem
831 tty->print_cr("Wrong number of expression stack elements during deoptimization");
832 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
833 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
834 iframe->interpreter_frame_expression_stack_size());
835 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
836 tty->print_cr(" try_next_mask = %d", try_next_mask);
837 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
838 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
839 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
840 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
841 tty->print_cr(" exec_mode = %d", exec_mode);
842 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
843 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
844 tty->print_cr(" Interpreted frames:");
845 for (int k = 0; k < cur_array->frames(); k++) {
846 vframeArrayElement* el = cur_array->element(k);
847 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
848 }
849 cur_array->print_on_2(tty);
850 } // release tty lock before calling guarantee
851 guarantee(false, "wrong number of expression stack elements during deopt");
852 }
853 VerifyOopClosure verify;
854 iframe->oops_interpreted_do(&verify, &rm, false);
855 callee_size_of_parameters = mh->size_of_parameters();
856 callee_max_locals = mh->max_locals();
857 is_top_frame = false;
858 }
859 }
860 #endif /* !PRODUCT */
861
862 return bt;
863 JRT_END
864
865 class DeoptimizeMarkedClosure : public HandshakeClosure {
866 public:
867 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
868 void do_thread(Thread* thread) {
869 JavaThread* jt = (JavaThread*)thread;
870 jt->deoptimize_marked_methods();
871 }
872 };
873
874 void Deoptimization::deoptimize_all_marked(nmethod* nmethod_only) {
875 ResourceMark rm;
876 DeoptimizationMarker dm;
877
878 // Make the dependent methods not entrant
879 if (nmethod_only != NULL) {
880 nmethod_only->mark_for_deoptimization();
881 nmethod_only->make_not_entrant();
882 } else {
883 MutexLocker mu(SafepointSynchronize::is_at_safepoint() ? NULL : CodeCache_lock, Mutex::_no_safepoint_check_flag);
884 CodeCache::make_marked_nmethods_not_entrant();
885 }
886
887 DeoptimizeMarkedClosure deopt;
888 if (SafepointSynchronize::is_at_safepoint()) {
889 Threads::java_threads_do(&deopt);
890 } else {
891 Handshake::execute(&deopt);
892 }
893 }
894
895 Deoptimization::DeoptAction Deoptimization::_unloaded_action
896 = Deoptimization::Action_reinterpret;
897
898
899
900 #if INCLUDE_JVMCI || INCLUDE_AOT
901 template<typename CacheType>
902 class BoxCacheBase : public CHeapObj<mtCompiler> {
903 protected:
904 static InstanceKlass* find_cache_klass(Symbol* klass_name, TRAPS) {
905 ResourceMark rm;
906 char* klass_name_str = klass_name->as_C_string();
907 Klass* k = SystemDictionary::find(klass_name, Handle(), Handle(), THREAD);
908 guarantee(k != NULL, "%s must be loaded", klass_name_str);
909 InstanceKlass* ik = InstanceKlass::cast(k);
910 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
911 CacheType::compute_offsets(ik);
912 return ik;
913 }
914 };
915
916 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
917 PrimitiveType _low;
918 PrimitiveType _high;
919 jobject _cache;
920 protected:
921 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
922 BoxCache(Thread* thread) {
923 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(CacheType::symbol(), thread);
924 objArrayOop cache = CacheType::cache(ik);
925 assert(cache->length() > 0, "Empty cache");
926 _low = BoxType::value(cache->obj_at(0));
927 _high = _low + cache->length() - 1;
928 _cache = JNIHandles::make_global(Handle(thread, cache));
929 }
930 ~BoxCache() {
931 JNIHandles::destroy_global(_cache);
932 }
933 public:
934 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
935 if (_singleton == NULL) {
936 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
937 if (!Atomic::replace_if_null(&_singleton, s)) {
938 delete s;
939 }
940 }
941 return _singleton;
942 }
943 oop lookup(PrimitiveType value) {
944 if (_low <= value && value <= _high) {
945 int offset = value - _low;
946 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
947 }
948 return NULL;
949 }
950 oop lookup_raw(intptr_t raw_value) {
951 // Have to cast to avoid little/big-endian problems.
952 if (sizeof(PrimitiveType) > sizeof(jint)) {
953 jlong value = (jlong)raw_value;
954 return lookup(value);
955 }
956 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
957 return lookup(value);
958 }
959 };
960
961 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
962 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
963 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
964 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
965 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
966
967 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL;
968 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL;
969 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL;
970 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL;
971 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL;
972
973 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
974 jobject _true_cache;
975 jobject _false_cache;
976 protected:
977 static BooleanBoxCache *_singleton;
978 BooleanBoxCache(Thread *thread) {
979 InstanceKlass* ik = find_cache_klass(java_lang_Boolean::symbol(), thread);
980 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
981 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
982 }
983 ~BooleanBoxCache() {
984 JNIHandles::destroy_global(_true_cache);
985 JNIHandles::destroy_global(_false_cache);
986 }
987 public:
988 static BooleanBoxCache* singleton(Thread* thread) {
989 if (_singleton == NULL) {
990 BooleanBoxCache* s = new BooleanBoxCache(thread);
991 if (!Atomic::replace_if_null(&_singleton, s)) {
992 delete s;
993 }
994 }
995 return _singleton;
996 }
997 oop lookup_raw(intptr_t raw_value) {
998 // Have to cast to avoid little/big-endian problems.
999 jboolean value = (jboolean)*((jint*)&raw_value);
1000 return lookup(value);
1001 }
1002 oop lookup(jboolean value) {
1003 if (value != 0) {
1004 return JNIHandles::resolve_non_null(_true_cache);
1005 }
1006 return JNIHandles::resolve_non_null(_false_cache);
1007 }
1008 };
1009
1010 BooleanBoxCache* BooleanBoxCache::_singleton = NULL;
1011
1012 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS) {
1013 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1014 BasicType box_type = SystemDictionary::box_klass_type(k);
1015 if (box_type != T_OBJECT) {
1016 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1017 switch(box_type) {
1018 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1019 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1020 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1021 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1022 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1023 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1024 default:;
1025 }
1026 }
1027 return NULL;
1028 }
1029 #endif // INCLUDE_JVMCI || INCLUDE_AOT
1030
1031 #if COMPILER2_OR_JVMCI
1032 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1033 Handle pending_exception(THREAD, thread->pending_exception());
1034 const char* exception_file = thread->exception_file();
1035 int exception_line = thread->exception_line();
1036 thread->clear_pending_exception();
1037
1038 bool failures = false;
1039
1040 for (int i = 0; i < objects->length(); i++) {
1041 assert(objects->at(i)->is_object(), "invalid debug information");
1042 ObjectValue* sv = (ObjectValue*) objects->at(i);
1043
1044 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1045 oop obj = NULL;
1046
1047 if (k->is_instance_klass()) {
1048 #if INCLUDE_JVMCI || INCLUDE_AOT
1049 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null();
1050 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1051 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1052 obj = get_cached_box(abv, fr, reg_map, THREAD);
1053 if (obj != NULL) {
1054 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1055 abv->set_cached(true);
1056 }
1057 }
1058 #endif // INCLUDE_JVMCI || INCLUDE_AOT
1059 InstanceKlass* ik = InstanceKlass::cast(k);
1060 if (obj == NULL) {
1061 obj = ik->allocate_instance(THREAD);
1062 }
1063 } else if (k->is_typeArray_klass()) {
1064 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1065 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1066 int len = sv->field_size() / type2size[ak->element_type()];
1067 obj = ak->allocate(len, THREAD);
1068 } else if (k->is_objArray_klass()) {
1069 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1070 obj = ak->allocate(sv->field_size(), THREAD);
1071 }
1072
1073 if (obj == NULL) {
1074 failures = true;
1075 }
1076
1077 assert(sv->value().is_null(), "redundant reallocation");
1078 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
1079 CLEAR_PENDING_EXCEPTION;
1080 sv->set_value(obj);
1081 }
1082
1083 if (failures) {
1084 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1085 } else if (pending_exception.not_null()) {
1086 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1087 }
1088
1089 return failures;
1090 }
1091
1092 #if INCLUDE_JVMCI
1093 /**
1094 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1095 * we need to somehow be able to recover the actual kind to be able to write the correct
1096 * amount of bytes.
1097 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1098 * the entries at index n + 1 to n + i are 'markers'.
1099 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1100 * expected form of the array would be:
1101 *
1102 * {b0, b1, b2, b3, INT, marker, b6, b7}
1103 *
1104 * Thus, in order to get back the size of the entry, we simply need to count the number
1105 * of marked entries
1106 *
1107 * @param virtualArray the virtualized byte array
1108 * @param i index of the virtual entry we are recovering
1109 * @return The number of bytes the entry spans
1110 */
1111 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1112 int index = i;
1113 while (++index < virtualArray->field_size() &&
1114 virtualArray->field_at(index)->is_marker()) {}
1115 return index - i;
1116 }
1117
1118 /**
1119 * If there was a guarantee for byte array to always start aligned to a long, we could
1120 * do a simple check on the parity of the index. Unfortunately, that is not always the
1121 * case. Thus, we check alignment of the actual address we are writing to.
1122 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1123 * write a long to index 3.
1124 */
1125 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1126 jbyte* res = obj->byte_at_addr(index);
1127 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1128 return res;
1129 }
1130
1131 static void byte_array_put(typeArrayOop obj, intptr_t val, int index, int byte_count) {
1132 switch (byte_count) {
1133 case 1:
1134 obj->byte_at_put(index, (jbyte) *((jint *) &val));
1135 break;
1136 case 2:
1137 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) *((jint *) &val);
1138 break;
1139 case 4:
1140 *((jint *) check_alignment_get_addr(obj, index, 4)) = (jint) *((jint *) &val);
1141 break;
1142 case 8:
1143 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) *((jlong *) &val);
1144 break;
1145 default:
1146 ShouldNotReachHere();
1147 }
1148 }
1149 #endif // INCLUDE_JVMCI
1150
1151
1152 // restore elements of an eliminated type array
1153 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1154 int index = 0;
1155 intptr_t val;
1156
1157 for (int i = 0; i < sv->field_size(); i++) {
1158 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1159 switch(type) {
1160 case T_LONG: case T_DOUBLE: {
1161 assert(value->type() == T_INT, "Agreement.");
1162 StackValue* low =
1163 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1164 #ifdef _LP64
1165 jlong res = (jlong)low->get_int();
1166 #else
1167 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1168 #endif
1169 obj->long_at_put(index, res);
1170 break;
1171 }
1172
1173 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1174 case T_INT: case T_FLOAT: { // 4 bytes.
1175 assert(value->type() == T_INT, "Agreement.");
1176 bool big_value = false;
1177 if (i + 1 < sv->field_size() && type == T_INT) {
1178 if (sv->field_at(i)->is_location()) {
1179 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1180 if (type == Location::dbl || type == Location::lng) {
1181 big_value = true;
1182 }
1183 } else if (sv->field_at(i)->is_constant_int()) {
1184 ScopeValue* next_scope_field = sv->field_at(i + 1);
1185 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1186 big_value = true;
1187 }
1188 }
1189 }
1190
1191 if (big_value) {
1192 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1193 #ifdef _LP64
1194 jlong res = (jlong)low->get_int();
1195 #else
1196 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1197 #endif
1198 obj->int_at_put(index, (jint)*((jint*)&res));
1199 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
1200 } else {
1201 val = value->get_int();
1202 obj->int_at_put(index, (jint)*((jint*)&val));
1203 }
1204 break;
1205 }
1206
1207 case T_SHORT:
1208 assert(value->type() == T_INT, "Agreement.");
1209 val = value->get_int();
1210 obj->short_at_put(index, (jshort)*((jint*)&val));
1211 break;
1212
1213 case T_CHAR:
1214 assert(value->type() == T_INT, "Agreement.");
1215 val = value->get_int();
1216 obj->char_at_put(index, (jchar)*((jint*)&val));
1217 break;
1218
1219 case T_BYTE: {
1220 assert(value->type() == T_INT, "Agreement.");
1221 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1222 val = value->get_int();
1223 #if INCLUDE_JVMCI
1224 int byte_count = count_number_of_bytes_for_entry(sv, i);
1225 byte_array_put(obj, val, index, byte_count);
1226 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1227 i += byte_count - 1; // Balance the loop counter.
1228 index += byte_count;
1229 // index has been updated so continue at top of loop
1230 continue;
1231 #else
1232 obj->byte_at_put(index, (jbyte)*((jint*)&val));
1233 break;
1234 #endif // INCLUDE_JVMCI
1235 }
1236
1237 case T_BOOLEAN: {
1238 assert(value->type() == T_INT, "Agreement.");
1239 val = value->get_int();
1240 obj->bool_at_put(index, (jboolean)*((jint*)&val));
1241 break;
1242 }
1243
1244 default:
1245 ShouldNotReachHere();
1246 }
1247 index++;
1248 }
1249 }
1250
1251 // restore fields of an eliminated object array
1252 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1253 for (int i = 0; i < sv->field_size(); i++) {
1254 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1255 assert(value->type() == T_OBJECT, "object element expected");
1256 obj->obj_at_put(i, value->get_obj()());
1257 }
1258 }
1259
1260 class ReassignedField {
1261 public:
1262 int _offset;
1263 BasicType _type;
1264 public:
1265 ReassignedField() {
1266 _offset = 0;
1267 _type = T_ILLEGAL;
1268 }
1269 };
1270
1271 int compare(ReassignedField* left, ReassignedField* right) {
1272 return left->_offset - right->_offset;
1273 }
1274
1275 // Restore fields of an eliminated instance object using the same field order
1276 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1277 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
1278 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1279 InstanceKlass* ik = klass;
1280 while (ik != NULL) {
1281 for (AllFieldStream fs(ik); !fs.done(); fs.next()) {
1282 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
1283 ReassignedField field;
1284 field._offset = fs.offset();
1285 field._type = Signature::basic_type(fs.signature());
1286 fields->append(field);
1287 }
1288 }
1289 ik = ik->superklass();
1290 }
1291 fields->sort(compare);
1292 for (int i = 0; i < fields->length(); i++) {
1293 intptr_t val;
1294 ScopeValue* scope_field = sv->field_at(svIndex);
1295 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1296 int offset = fields->at(i)._offset;
1297 BasicType type = fields->at(i)._type;
1298 switch (type) {
1299 case T_OBJECT: case T_ARRAY:
1300 assert(value->type() == T_OBJECT, "Agreement.");
1301 obj->obj_field_put(offset, value->get_obj()());
1302 break;
1303
1304 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1305 case T_INT: case T_FLOAT: { // 4 bytes.
1306 assert(value->type() == T_INT, "Agreement.");
1307 bool big_value = false;
1308 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1309 if (scope_field->is_location()) {
1310 Location::Type type = ((LocationValue*) scope_field)->location().type();
1311 if (type == Location::dbl || type == Location::lng) {
1312 big_value = true;
1313 }
1314 }
1315 if (scope_field->is_constant_int()) {
1316 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1317 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1318 big_value = true;
1319 }
1320 }
1321 }
1322
1323 if (big_value) {
1324 i++;
1325 assert(i < fields->length(), "second T_INT field needed");
1326 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1327 } else {
1328 val = value->get_int();
1329 obj->int_field_put(offset, (jint)*((jint*)&val));
1330 break;
1331 }
1332 }
1333 /* no break */
1334
1335 case T_LONG: case T_DOUBLE: {
1336 assert(value->type() == T_INT, "Agreement.");
1337 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1338 #ifdef _LP64
1339 jlong res = (jlong)low->get_int();
1340 #else
1341 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1342 #endif
1343 obj->long_field_put(offset, res);
1344 break;
1345 }
1346
1347 case T_SHORT:
1348 assert(value->type() == T_INT, "Agreement.");
1349 val = value->get_int();
1350 obj->short_field_put(offset, (jshort)*((jint*)&val));
1351 break;
1352
1353 case T_CHAR:
1354 assert(value->type() == T_INT, "Agreement.");
1355 val = value->get_int();
1356 obj->char_field_put(offset, (jchar)*((jint*)&val));
1357 break;
1358
1359 case T_BYTE:
1360 assert(value->type() == T_INT, "Agreement.");
1361 val = value->get_int();
1362 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1363 break;
1364
1365 case T_BOOLEAN:
1366 assert(value->type() == T_INT, "Agreement.");
1367 val = value->get_int();
1368 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1369 break;
1370
1371 default:
1372 ShouldNotReachHere();
1373 }
1374 svIndex++;
1375 }
1376 return svIndex;
1377 }
1378
1379 // restore fields of all eliminated objects and arrays
1380 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1381 for (int i = 0; i < objects->length(); i++) {
1382 ObjectValue* sv = (ObjectValue*) objects->at(i);
1383 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1384 Handle obj = sv->value();
1385 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1386 if (PrintDeoptimizationDetails) {
1387 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1388 }
1389 if (obj.is_null()) {
1390 continue;
1391 }
1392 #if INCLUDE_JVMCI || INCLUDE_AOT
1393 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1394 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1395 continue;
1396 }
1397 #endif // INCLUDE_JVMCI || INCLUDE_AOT
1398 if (k->is_instance_klass()) {
1399 InstanceKlass* ik = InstanceKlass::cast(k);
1400 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1401 } else if (k->is_typeArray_klass()) {
1402 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1403 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1404 } else if (k->is_objArray_klass()) {
1405 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1406 }
1407 }
1408 }
1409
1410
1411 // relock objects for which synchronization was eliminated
1412 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1413 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1414 bool relocked_objects = false;
1415 for (int i = 0; i < monitors->length(); i++) {
1416 MonitorInfo* mon_info = monitors->at(i);
1417 if (mon_info->eliminated()) {
1418 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1419 relocked_objects = true;
1420 if (!mon_info->owner_is_scalar_replaced()) {
1421 Handle obj(thread, mon_info->owner());
1422 markWord mark = obj->mark();
1423 if (UseBiasedLocking && mark.has_bias_pattern()) {
1424 // New allocated objects may have the mark set to anonymously biased.
1425 // Also the deoptimized method may called methods with synchronization
1426 // where the thread-local object is bias locked to the current thread.
1427 assert(mark.is_biased_anonymously() ||
1428 mark.biased_locker() == deoptee_thread, "should be locked to current thread");
1429 // Reset mark word to unbiased prototype.
1430 markWord unbiased_prototype = markWord::prototype().set_age(mark.age());
1431 obj->set_mark(unbiased_prototype);
1432 } else if (exec_mode == Unpack_none) {
1433 if (mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1434 // With exec_mode == Unpack_none obj may be thread local and locked in
1435 // a callee frame. In this case the bias was revoked before in revoke_for_object_deoptimization().
1436 // Make the lock in the callee a recursive lock and restore the displaced header.
1437 markWord dmw = mark.displaced_mark_helper();
1438 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) NULL));
1439 obj->set_mark(dmw);
1440 }
1441 if (mark.has_monitor()) {
1442 // defer relocking if the deoptee thread is currently waiting for obj
1443 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1444 if (waiting_monitor != NULL && (oop)waiting_monitor->object() == obj()) {
1445 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1446 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1447 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1448 continue;
1449 }
1450 }
1451 }
1452 BasicLock* lock = mon_info->lock();
1453 ObjectSynchronizer::enter(obj, lock, deoptee_thread);
1454 assert(mon_info->owner()->is_locked(), "object must be locked now");
1455 }
1456 }
1457 }
1458 return relocked_objects;
1459 }
1460
1461
1462 #ifndef PRODUCT
1463 // print information about reallocated objects
1464 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1465 fieldDescriptor fd;
1466
1467 for (int i = 0; i < objects->length(); i++) {
1468 ObjectValue* sv = (ObjectValue*) objects->at(i);
1469 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1470 Handle obj = sv->value();
1471
1472 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1473 k->print_value();
1474 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1475 if (obj.is_null()) {
1476 tty->print(" allocation failed");
1477 } else {
1478 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1479 }
1480 tty->cr();
1481
1482 if (Verbose && !obj.is_null()) {
1483 k->oop_print_on(obj(), tty);
1484 }
1485 }
1486 }
1487 #endif
1488 #endif // COMPILER2_OR_JVMCI
1489
1490 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1491 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1492
1493 #ifndef PRODUCT
1494 if (PrintDeoptimizationDetails) {
1495 ttyLocker ttyl;
1496 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1497 fr.print_on(tty);
1498 tty->print_cr(" Virtual frames (innermost first):");
1499 for (int index = 0; index < chunk->length(); index++) {
1500 compiledVFrame* vf = chunk->at(index);
1501 tty->print(" %2d - ", index);
1502 vf->print_value();
1503 int bci = chunk->at(index)->raw_bci();
1504 const char* code_name;
1505 if (bci == SynchronizationEntryBCI) {
1506 code_name = "sync entry";
1507 } else {
1508 Bytecodes::Code code = vf->method()->code_at(bci);
1509 code_name = Bytecodes::name(code);
1510 }
1511 tty->print(" - %s", code_name);
1512 tty->print_cr(" @ bci %d ", bci);
1513 if (Verbose) {
1514 vf->print();
1515 tty->cr();
1516 }
1517 }
1518 }
1519 #endif
1520
1521 // Register map for next frame (used for stack crawl). We capture
1522 // the state of the deopt'ing frame's caller. Thus if we need to
1523 // stuff a C2I adapter we can properly fill in the callee-save
1524 // register locations.
1525 frame caller = fr.sender(reg_map);
1526 int frame_size = caller.sp() - fr.sp();
1527
1528 frame sender = caller;
1529
1530 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1531 // the vframeArray containing the unpacking information is allocated in the C heap.
1532 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1533 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1534
1535 // Compare the vframeArray to the collected vframes
1536 assert(array->structural_compare(thread, chunk), "just checking");
1537
1538 #ifndef PRODUCT
1539 if (PrintDeoptimizationDetails) {
1540 ttyLocker ttyl;
1541 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1542 }
1543 #endif // PRODUCT
1544
1545 return array;
1546 }
1547
1548 #if COMPILER2_OR_JVMCI
1549 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1550 // Reallocation of some scalar replaced objects failed. Record
1551 // that we need to pop all the interpreter frames for the
1552 // deoptimized compiled frame.
1553 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1554 thread->set_frames_to_pop_failed_realloc(array->frames());
1555 // Unlock all monitors here otherwise the interpreter will see a
1556 // mix of locked and unlocked monitors (because of failed
1557 // reallocations of synchronized objects) and be confused.
1558 for (int i = 0; i < array->frames(); i++) {
1559 MonitorChunk* monitors = array->element(i)->monitors();
1560 if (monitors != NULL) {
1561 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1562 BasicObjectLock* src = monitors->at(j);
1563 if (src->obj() != NULL) {
1564 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1565 }
1566 }
1567 array->element(i)->free_monitors(thread);
1568 #ifdef ASSERT
1569 array->element(i)->set_removed_monitors();
1570 #endif
1571 }
1572 }
1573 }
1574 #endif
1575
1576 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke, bool only_eliminated) {
1577 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1578 Thread* thread = Thread::current();
1579 for (int i = 0; i < monitors->length(); i++) {
1580 MonitorInfo* mon_info = monitors->at(i);
1581 if ((mon_info->eliminated() == only_eliminated) && !mon_info->owner_is_scalar_replaced() && mon_info->owner() != NULL) {
1582 objects_to_revoke->append(Handle(thread, mon_info->owner()));
1583 }
1584 }
1585 }
1586
1587 static void get_monitors_from_stack(GrowableArray<Handle>* objects_to_revoke, JavaThread* thread, frame fr, RegisterMap* map,
1588 bool only_eliminated) {
1589 // Unfortunately we don't have a RegisterMap available in most of
1590 // the places we want to call this routine so we need to walk the
1591 // stack again to update the register map.
1592 if (map == NULL || !map->update_map()) {
1593 StackFrameStream sfs(thread, true);
1594 bool found = false;
1595 while (!found && !sfs.is_done()) {
1596 frame* cur = sfs.current();
1597 sfs.next();
1598 found = cur->id() == fr.id();
1599 }
1600 assert(found, "frame to be deoptimized not found on target thread's stack");
1601 map = sfs.register_map();
1602 }
1603
1604 vframe* vf = vframe::new_vframe(&fr, map, thread);
1605 compiledVFrame* cvf = compiledVFrame::cast(vf);
1606 // Revoke monitors' biases in all scopes
1607 while (!cvf->is_top()) {
1608 collect_monitors(cvf, objects_to_revoke, only_eliminated);
1609 cvf = compiledVFrame::cast(cvf->sender());
1610 }
1611 collect_monitors(cvf, objects_to_revoke, only_eliminated);
1612 }
1613
1614 void Deoptimization::revoke_from_deopt_handler(JavaThread* thread, frame fr, RegisterMap* map) {
1615 if (!UseBiasedLocking) {
1616 return;
1617 }
1618 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1619 get_monitors_from_stack(objects_to_revoke, thread, fr, map, false);
1620
1621 int len = objects_to_revoke->length();
1622 for (int i = 0; i < len; i++) {
1623 oop obj = (objects_to_revoke->at(i))();
1624 BiasedLocking::revoke_own_lock(objects_to_revoke->at(i), thread);
1625 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
1626 }
1627 }
1628
1629 // Revoke the bias of objects with eliminated locking to prepare subsequent relocking.
1630 void Deoptimization::revoke_for_object_deoptimization(JavaThread* deoptee_thread, frame fr, RegisterMap* map, JavaThread* thread) {
1631 if (!UseBiasedLocking) {
1632 return;
1633 }
1634 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1635 // Collect monitors, but only those with eliminated locking.
1636 get_monitors_from_stack(objects_to_revoke, deoptee_thread, fr, map, true);
1637
1638 int len = objects_to_revoke->length();
1639 for (int i = 0; i < len; i++) {
1640 oop obj = (objects_to_revoke->at(i))();
1641 markWord mark = obj->mark();
1642 if (!mark.has_bias_pattern()
1643 || mark.is_biased_anonymously() // eliminated locking does not bias an object if it wasn't before
1644 || !obj->klass()->prototype_header().has_bias_pattern() // bulk revoke ignores eliminated monitors
1645 || (obj->klass()->prototype_header().bias_epoch() != mark.bias_epoch())) { // bulk rebias ignores eliminated monitors
1646 // We reach here regularly if there's just eliminated locking on obj.
1647 // We must not call BiasedLocking::revoke_own_lock() in this case, as we would hit assertions, because it is a
1648 // prerequisite that there has to be non-eliminated locking on obj by deoptee_thread.
1649 // Luckily we don't have to revoke here, because obj has to be a non-escaping obj and can be relocked without
1650 // revoking the bias. See Deoptimization::relock_objects().
1651 continue;
1652 }
1653 BiasedLocking::revoke(objects_to_revoke->at(i), thread);
1654 assert(!objects_to_revoke->at(i)->mark().has_bias_pattern(), "biases should be revoked by now");
1655 }
1656 }
1657
1658 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1659 assert(fr.can_be_deoptimized(), "checking frame type");
1660
1661 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1662
1663 if (LogCompilation && xtty != NULL) {
1664 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1665 assert(cm != NULL, "only compiled methods can deopt");
1666
1667 ttyLocker ttyl;
1668 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1669 cm->log_identity(xtty);
1670 xtty->end_head();
1671 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1672 xtty->begin_elem("jvms bci='%d'", sd->bci());
1673 xtty->method(sd->method());
1674 xtty->end_elem();
1675 if (sd->is_top()) break;
1676 }
1677 xtty->tail("deoptimized");
1678 }
1679
1680 // Patch the compiled method so that when execution returns to it we will
1681 // deopt the execution state and return to the interpreter.
1682 fr.deoptimize(thread);
1683 }
1684
1685 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1686 // Deoptimize only if the frame comes from compile code.
1687 // Do not deoptimize the frame which is already patched
1688 // during the execution of the loops below.
1689 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1690 return;
1691 }
1692 ResourceMark rm;
1693 DeoptimizationMarker dm;
1694 deoptimize_single_frame(thread, fr, reason);
1695 }
1696
1697 #if INCLUDE_JVMCI
1698 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) {
1699 // there is no exception handler for this pc => deoptimize
1700 cm->make_not_entrant();
1701
1702 // Use Deoptimization::deoptimize for all of its side-effects:
1703 // gathering traps statistics, logging...
1704 // it also patches the return pc but we do not care about that
1705 // since we return a continuation to the deopt_blob below.
1706 JavaThread* thread = JavaThread::current();
1707 RegisterMap reg_map(thread, false);
1708 frame runtime_frame = thread->last_frame();
1709 frame caller_frame = runtime_frame.sender(®_map);
1710 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method");
1711 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1712
1713 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true);
1714 if (trap_mdo != NULL) {
1715 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1716 }
1717
1718 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1719 }
1720 #endif
1721
1722 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1723 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1724 "can only deoptimize other thread at a safepoint");
1725 // Compute frame and register map based on thread and sp.
1726 RegisterMap reg_map(thread, false);
1727 frame fr = thread->last_frame();
1728 while (fr.id() != id) {
1729 fr = fr.sender(®_map);
1730 }
1731 deoptimize(thread, fr, reason);
1732 }
1733
1734
1735 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1736 if (thread == Thread::current()) {
1737 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1738 } else {
1739 VM_DeoptimizeFrame deopt(thread, id, reason);
1740 VMThread::execute(&deopt);
1741 }
1742 }
1743
1744 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1745 deoptimize_frame(thread, id, Reason_constraint);
1746 }
1747
1748 // JVMTI PopFrame support
1749 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1750 {
1751 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1752 }
1753 JRT_END
1754
1755 MethodData*
1756 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1757 bool create_if_missing) {
1758 Thread* THREAD = thread;
1759 MethodData* mdo = m()->method_data();
1760 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1761 // Build an MDO. Ignore errors like OutOfMemory;
1762 // that simply means we won't have an MDO to update.
1763 Method::build_interpreter_method_data(m, THREAD);
1764 if (HAS_PENDING_EXCEPTION) {
1765 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1766 CLEAR_PENDING_EXCEPTION;
1767 }
1768 mdo = m()->method_data();
1769 }
1770 return mdo;
1771 }
1772
1773 #if COMPILER2_OR_JVMCI
1774 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1775 // In case of an unresolved klass entry, load the class.
1776 // This path is exercised from case _ldc in Parse::do_one_bytecode,
1777 // and probably nowhere else.
1778 // Even that case would benefit from simply re-interpreting the
1779 // bytecode, without paying special attention to the class index.
1780 // So this whole "class index" feature should probably be removed.
1781
1782 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1783 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1784 return;
1785 }
1786
1787 assert(!constant_pool->tag_at(index).is_symbol(),
1788 "no symbolic names here, please");
1789 }
1790
1791
1792 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1793 EXCEPTION_MARK;
1794 load_class_by_index(constant_pool, index, THREAD);
1795 if (HAS_PENDING_EXCEPTION) {
1796 // Exception happened during classloading. We ignore the exception here, since it
1797 // is going to be rethrown since the current activation is going to be deoptimized and
1798 // the interpreter will re-execute the bytecode.
1799 CLEAR_PENDING_EXCEPTION;
1800 // Class loading called java code which may have caused a stack
1801 // overflow. If the exception was thrown right before the return
1802 // to the runtime the stack is no longer guarded. Reguard the
1803 // stack otherwise if we return to the uncommon trap blob and the
1804 // stack bang causes a stack overflow we crash.
1805 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1806 JavaThread* thread = (JavaThread*)THREAD;
1807 bool guard_pages_enabled = thread->stack_guards_enabled();
1808 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1809 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1810 }
1811 }
1812
1813 #if INCLUDE_JFR
1814
1815 class DeoptReasonSerializer : public JfrSerializer {
1816 public:
1817 void serialize(JfrCheckpointWriter& writer) {
1818 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1819 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1820 writer.write_key((u8)i);
1821 writer.write(Deoptimization::trap_reason_name(i));
1822 }
1823 }
1824 };
1825
1826 class DeoptActionSerializer : public JfrSerializer {
1827 public:
1828 void serialize(JfrCheckpointWriter& writer) {
1829 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1830 writer.write_count(nof_actions);
1831 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1832 writer.write_key(i);
1833 writer.write(Deoptimization::trap_action_name((int)i));
1834 }
1835 }
1836 };
1837
1838 static void register_serializers() {
1839 static int critical_section = 0;
1840 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1841 return;
1842 }
1843 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1844 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1845 }
1846
1847 static void post_deoptimization_event(CompiledMethod* nm,
1848 const Method* method,
1849 int trap_bci,
1850 int instruction,
1851 Deoptimization::DeoptReason reason,
1852 Deoptimization::DeoptAction action) {
1853 assert(nm != NULL, "invariant");
1854 assert(method != NULL, "invariant");
1855 if (EventDeoptimization::is_enabled()) {
1856 static bool serializers_registered = false;
1857 if (!serializers_registered) {
1858 register_serializers();
1859 serializers_registered = true;
1860 }
1861 EventDeoptimization event;
1862 event.set_compileId(nm->compile_id());
1863 event.set_compiler(nm->compiler_type());
1864 event.set_method(method);
1865 event.set_lineNumber(method->line_number_from_bci(trap_bci));
1866 event.set_bci(trap_bci);
1867 event.set_instruction(instruction);
1868 event.set_reason(reason);
1869 event.set_action(action);
1870 event.commit();
1871 }
1872 }
1873
1874 #endif // INCLUDE_JFR
1875
1876 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1877 HandleMark hm;
1878
1879 // uncommon_trap() is called at the beginning of the uncommon trap
1880 // handler. Note this fact before we start generating temporary frames
1881 // that can confuse an asynchronous stack walker. This counter is
1882 // decremented at the end of unpack_frames().
1883 thread->inc_in_deopt_handler();
1884
1885 // We need to update the map if we have biased locking.
1886 #if INCLUDE_JVMCI
1887 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1888 RegisterMap reg_map(thread, true);
1889 #else
1890 RegisterMap reg_map(thread, UseBiasedLocking);
1891 #endif
1892 frame stub_frame = thread->last_frame();
1893 frame fr = stub_frame.sender(®_map);
1894 // Make sure the calling nmethod is not getting deoptimized and removed
1895 // before we are done with it.
1896 nmethodLocker nl(fr.pc());
1897
1898 // Log a message
1899 Events::log_deopt_message(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1900 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1901
1902 {
1903 ResourceMark rm;
1904
1905 DeoptReason reason = trap_request_reason(trap_request);
1906 DeoptAction action = trap_request_action(trap_request);
1907 #if INCLUDE_JVMCI
1908 int debug_id = trap_request_debug_id(trap_request);
1909 #endif
1910 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1911
1912 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1913 compiledVFrame* cvf = compiledVFrame::cast(vf);
1914
1915 CompiledMethod* nm = cvf->code();
1916
1917 ScopeDesc* trap_scope = cvf->scope();
1918
1919 if (TraceDeoptimization) {
1920 ttyLocker ttyl;
1921 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()
1922 #if INCLUDE_JVMCI
1923 , debug_id
1924 #endif
1925 );
1926 }
1927
1928 methodHandle trap_method(THREAD, trap_scope->method());
1929 int trap_bci = trap_scope->bci();
1930 #if INCLUDE_JVMCI
1931 jlong speculation = thread->pending_failed_speculation();
1932 if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods
1933 nm->as_nmethod()->update_speculation(thread);
1934 } else {
1935 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
1936 }
1937
1938 if (trap_bci == SynchronizationEntryBCI) {
1939 trap_bci = 0;
1940 thread->set_pending_monitorenter(true);
1941 }
1942
1943 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1944 thread->set_pending_transfer_to_interpreter(true);
1945 }
1946 #endif
1947
1948 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1949 // Record this event in the histogram.
1950 gather_statistics(reason, action, trap_bc);
1951
1952 // Ensure that we can record deopt. history:
1953 // Need MDO to record RTM code generation state.
1954 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1955
1956 methodHandle profiled_method;
1957 #if INCLUDE_JVMCI
1958 if (nm->is_compiled_by_jvmci()) {
1959 profiled_method = methodHandle(THREAD, nm->method());
1960 } else {
1961 profiled_method = trap_method;
1962 }
1963 #else
1964 profiled_method = trap_method;
1965 #endif
1966
1967 MethodData* trap_mdo =
1968 get_method_data(thread, profiled_method, create_if_missing);
1969
1970 JFR_ONLY(post_deoptimization_event(nm, trap_method(), trap_bci, trap_bc, reason, action);)
1971
1972 // Log a message
1973 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
1974 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1975 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
1976
1977 // Print a bunch of diagnostics, if requested.
1978 if (TraceDeoptimization || LogCompilation) {
1979 ResourceMark rm;
1980 ttyLocker ttyl;
1981 char buf[100];
1982 if (xtty != NULL) {
1983 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1984 os::current_thread_id(),
1985 format_trap_request(buf, sizeof(buf), trap_request));
1986 #if INCLUDE_JVMCI
1987 if (speculation != 0) {
1988 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
1989 }
1990 #endif
1991 nm->log_identity(xtty);
1992 }
1993 Symbol* class_name = NULL;
1994 bool unresolved = false;
1995 if (unloaded_class_index >= 0) {
1996 constantPoolHandle constants (THREAD, trap_method->constants());
1997 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1998 class_name = constants->klass_name_at(unloaded_class_index);
1999 unresolved = true;
2000 if (xtty != NULL)
2001 xtty->print(" unresolved='1'");
2002 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2003 class_name = constants->symbol_at(unloaded_class_index);
2004 }
2005 if (xtty != NULL)
2006 xtty->name(class_name);
2007 }
2008 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
2009 // Dump the relevant MDO state.
2010 // This is the deopt count for the current reason, any previous
2011 // reasons or recompiles seen at this point.
2012 int dcnt = trap_mdo->trap_count(reason);
2013 if (dcnt != 0)
2014 xtty->print(" count='%d'", dcnt);
2015 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2016 int dos = (pdata == NULL)? 0: pdata->trap_state();
2017 if (dos != 0) {
2018 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2019 if (trap_state_is_recompiled(dos)) {
2020 int recnt2 = trap_mdo->overflow_recompile_count();
2021 if (recnt2 != 0)
2022 xtty->print(" recompiles2='%d'", recnt2);
2023 }
2024 }
2025 }
2026 if (xtty != NULL) {
2027 xtty->stamp();
2028 xtty->end_head();
2029 }
2030 if (TraceDeoptimization) { // make noise on the tty
2031 tty->print("Uncommon trap occurred in");
2032 nm->method()->print_short_name(tty);
2033 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2034 #if INCLUDE_JVMCI
2035 if (nm->is_nmethod()) {
2036 const char* installed_code_name = nm->as_nmethod()->jvmci_name();
2037 if (installed_code_name != NULL) {
2038 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name);
2039 }
2040 }
2041 #endif
2042 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2043 p2i(fr.pc()),
2044 os::current_thread_id(),
2045 trap_reason_name(reason),
2046 trap_action_name(action),
2047 unloaded_class_index
2048 #if INCLUDE_JVMCI
2049 , debug_id
2050 #endif
2051 );
2052 if (class_name != NULL) {
2053 tty->print(unresolved ? " unresolved class: " : " symbol: ");
2054 class_name->print_symbol_on(tty);
2055 }
2056 tty->cr();
2057 }
2058 if (xtty != NULL) {
2059 // Log the precise location of the trap.
2060 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2061 xtty->begin_elem("jvms bci='%d'", sd->bci());
2062 xtty->method(sd->method());
2063 xtty->end_elem();
2064 if (sd->is_top()) break;
2065 }
2066 xtty->tail("uncommon_trap");
2067 }
2068 }
2069 // (End diagnostic printout.)
2070
2071 // Load class if necessary
2072 if (unloaded_class_index >= 0) {
2073 constantPoolHandle constants(THREAD, trap_method->constants());
2074 load_class_by_index(constants, unloaded_class_index);
2075 }
2076
2077 // Flush the nmethod if necessary and desirable.
2078 //
2079 // We need to avoid situations where we are re-flushing the nmethod
2080 // because of a hot deoptimization site. Repeated flushes at the same
2081 // point need to be detected by the compiler and avoided. If the compiler
2082 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2083 // module must take measures to avoid an infinite cycle of recompilation
2084 // and deoptimization. There are several such measures:
2085 //
2086 // 1. If a recompilation is ordered a second time at some site X
2087 // and for the same reason R, the action is adjusted to 'reinterpret',
2088 // to give the interpreter time to exercise the method more thoroughly.
2089 // If this happens, the method's overflow_recompile_count is incremented.
2090 //
2091 // 2. If the compiler fails to reduce the deoptimization rate, then
2092 // the method's overflow_recompile_count will begin to exceed the set
2093 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2094 // is adjusted to 'make_not_compilable', and the method is abandoned
2095 // to the interpreter. This is a performance hit for hot methods,
2096 // but is better than a disastrous infinite cycle of recompilations.
2097 // (Actually, only the method containing the site X is abandoned.)
2098 //
2099 // 3. In parallel with the previous measures, if the total number of
2100 // recompilations of a method exceeds the much larger set limit
2101 // PerMethodRecompilationCutoff, the method is abandoned.
2102 // This should only happen if the method is very large and has
2103 // many "lukewarm" deoptimizations. The code which enforces this
2104 // limit is elsewhere (class nmethod, class Method).
2105 //
2106 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2107 // to recompile at each bytecode independently of the per-BCI cutoff.
2108 //
2109 // The decision to update code is up to the compiler, and is encoded
2110 // in the Action_xxx code. If the compiler requests Action_none
2111 // no trap state is changed, no compiled code is changed, and the
2112 // computation suffers along in the interpreter.
2113 //
2114 // The other action codes specify various tactics for decompilation
2115 // and recompilation. Action_maybe_recompile is the loosest, and
2116 // allows the compiled code to stay around until enough traps are seen,
2117 // and until the compiler gets around to recompiling the trapping method.
2118 //
2119 // The other actions cause immediate removal of the present code.
2120
2121 // Traps caused by injected profile shouldn't pollute trap counts.
2122 bool injected_profile_trap = trap_method->has_injected_profile() &&
2123 (reason == Reason_intrinsic || reason == Reason_unreached);
2124
2125 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2126 bool make_not_entrant = false;
2127 bool make_not_compilable = false;
2128 bool reprofile = false;
2129 switch (action) {
2130 case Action_none:
2131 // Keep the old code.
2132 update_trap_state = false;
2133 break;
2134 case Action_maybe_recompile:
2135 // Do not need to invalidate the present code, but we can
2136 // initiate another
2137 // Start compiler without (necessarily) invalidating the nmethod.
2138 // The system will tolerate the old code, but new code should be
2139 // generated when possible.
2140 break;
2141 case Action_reinterpret:
2142 // Go back into the interpreter for a while, and then consider
2143 // recompiling form scratch.
2144 make_not_entrant = true;
2145 // Reset invocation counter for outer most method.
2146 // This will allow the interpreter to exercise the bytecodes
2147 // for a while before recompiling.
2148 // By contrast, Action_make_not_entrant is immediate.
2149 //
2150 // Note that the compiler will track null_check, null_assert,
2151 // range_check, and class_check events and log them as if they
2152 // had been traps taken from compiled code. This will update
2153 // the MDO trap history so that the next compilation will
2154 // properly detect hot trap sites.
2155 reprofile = true;
2156 break;
2157 case Action_make_not_entrant:
2158 // Request immediate recompilation, and get rid of the old code.
2159 // Make them not entrant, so next time they are called they get
2160 // recompiled. Unloaded classes are loaded now so recompile before next
2161 // time they are called. Same for uninitialized. The interpreter will
2162 // link the missing class, if any.
2163 make_not_entrant = true;
2164 break;
2165 case Action_make_not_compilable:
2166 // Give up on compiling this method at all.
2167 make_not_entrant = true;
2168 make_not_compilable = true;
2169 break;
2170 default:
2171 ShouldNotReachHere();
2172 }
2173
2174 // Setting +ProfileTraps fixes the following, on all platforms:
2175 // 4852688: ProfileInterpreter is off by default for ia64. The result is
2176 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2177 // recompile relies on a MethodData* to record heroic opt failures.
2178
2179 // Whether the interpreter is producing MDO data or not, we also need
2180 // to use the MDO to detect hot deoptimization points and control
2181 // aggressive optimization.
2182 bool inc_recompile_count = false;
2183 ProfileData* pdata = NULL;
2184 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) {
2185 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
2186 uint this_trap_count = 0;
2187 bool maybe_prior_trap = false;
2188 bool maybe_prior_recompile = false;
2189 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2190 #if INCLUDE_JVMCI
2191 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2192 #endif
2193 nm->method(),
2194 //outputs:
2195 this_trap_count,
2196 maybe_prior_trap,
2197 maybe_prior_recompile);
2198 // Because the interpreter also counts null, div0, range, and class
2199 // checks, these traps from compiled code are double-counted.
2200 // This is harmless; it just means that the PerXTrapLimit values
2201 // are in effect a little smaller than they look.
2202
2203 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2204 if (per_bc_reason != Reason_none) {
2205 // Now take action based on the partially known per-BCI history.
2206 if (maybe_prior_trap
2207 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2208 // If there are too many traps at this BCI, force a recompile.
2209 // This will allow the compiler to see the limit overflow, and
2210 // take corrective action, if possible. The compiler generally
2211 // does not use the exact PerBytecodeTrapLimit value, but instead
2212 // changes its tactics if it sees any traps at all. This provides
2213 // a little hysteresis, delaying a recompile until a trap happens
2214 // several times.
2215 //
2216 // Actually, since there is only one bit of counter per BCI,
2217 // the possible per-BCI counts are {0,1,(per-method count)}.
2218 // This produces accurate results if in fact there is only
2219 // one hot trap site, but begins to get fuzzy if there are
2220 // many sites. For example, if there are ten sites each
2221 // trapping two or more times, they each get the blame for
2222 // all of their traps.
2223 make_not_entrant = true;
2224 }
2225
2226 // Detect repeated recompilation at the same BCI, and enforce a limit.
2227 if (make_not_entrant && maybe_prior_recompile) {
2228 // More than one recompile at this point.
2229 inc_recompile_count = maybe_prior_trap;
2230 }
2231 } else {
2232 // For reasons which are not recorded per-bytecode, we simply
2233 // force recompiles unconditionally.
2234 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2235 make_not_entrant = true;
2236 }
2237
2238 // Go back to the compiler if there are too many traps in this method.
2239 if (this_trap_count >= per_method_trap_limit(reason)) {
2240 // If there are too many traps in this method, force a recompile.
2241 // This will allow the compiler to see the limit overflow, and
2242 // take corrective action, if possible.
2243 // (This condition is an unlikely backstop only, because the
2244 // PerBytecodeTrapLimit is more likely to take effect first,
2245 // if it is applicable.)
2246 make_not_entrant = true;
2247 }
2248
2249 // Here's more hysteresis: If there has been a recompile at
2250 // this trap point already, run the method in the interpreter
2251 // for a while to exercise it more thoroughly.
2252 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2253 reprofile = true;
2254 }
2255 }
2256
2257 // Take requested actions on the method:
2258
2259 // Recompile
2260 if (make_not_entrant) {
2261 if (!nm->make_not_entrant()) {
2262 return; // the call did not change nmethod's state
2263 }
2264
2265 if (pdata != NULL) {
2266 // Record the recompilation event, if any.
2267 int tstate0 = pdata->trap_state();
2268 int tstate1 = trap_state_set_recompiled(tstate0, true);
2269 if (tstate1 != tstate0)
2270 pdata->set_trap_state(tstate1);
2271 }
2272
2273 #if INCLUDE_RTM_OPT
2274 // Restart collecting RTM locking abort statistic if the method
2275 // is recompiled for a reason other than RTM state change.
2276 // Assume that in new recompiled code the statistic could be different,
2277 // for example, due to different inlining.
2278 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
2279 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
2280 trap_mdo->atomic_set_rtm_state(ProfileRTM);
2281 }
2282 #endif
2283 // For code aging we count traps separately here, using make_not_entrant()
2284 // as a guard against simultaneous deopts in multiple threads.
2285 if (reason == Reason_tenured && trap_mdo != NULL) {
2286 trap_mdo->inc_tenure_traps();
2287 }
2288 }
2289
2290 if (inc_recompile_count) {
2291 trap_mdo->inc_overflow_recompile_count();
2292 if ((uint)trap_mdo->overflow_recompile_count() >
2293 (uint)PerBytecodeRecompilationCutoff) {
2294 // Give up on the method containing the bad BCI.
2295 if (trap_method() == nm->method()) {
2296 make_not_compilable = true;
2297 } else {
2298 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2299 // But give grace to the enclosing nm->method().
2300 }
2301 }
2302 }
2303
2304 // Reprofile
2305 if (reprofile) {
2306 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
2307 }
2308
2309 // Give up compiling
2310 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2311 assert(make_not_entrant, "consistent");
2312 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2313 }
2314
2315 } // Free marked resources
2316
2317 }
2318 JRT_END
2319
2320 ProfileData*
2321 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2322 int trap_bci,
2323 Deoptimization::DeoptReason reason,
2324 bool update_total_trap_count,
2325 #if INCLUDE_JVMCI
2326 bool is_osr,
2327 #endif
2328 Method* compiled_method,
2329 //outputs:
2330 uint& ret_this_trap_count,
2331 bool& ret_maybe_prior_trap,
2332 bool& ret_maybe_prior_recompile) {
2333 bool maybe_prior_trap = false;
2334 bool maybe_prior_recompile = false;
2335 uint this_trap_count = 0;
2336 if (update_total_trap_count) {
2337 uint idx = reason;
2338 #if INCLUDE_JVMCI
2339 if (is_osr) {
2340 idx += Reason_LIMIT;
2341 }
2342 #endif
2343 uint prior_trap_count = trap_mdo->trap_count(idx);
2344 this_trap_count = trap_mdo->inc_trap_count(idx);
2345
2346 // If the runtime cannot find a place to store trap history,
2347 // it is estimated based on the general condition of the method.
2348 // If the method has ever been recompiled, or has ever incurred
2349 // a trap with the present reason , then this BCI is assumed
2350 // (pessimistically) to be the culprit.
2351 maybe_prior_trap = (prior_trap_count != 0);
2352 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2353 }
2354 ProfileData* pdata = NULL;
2355
2356
2357 // For reasons which are recorded per bytecode, we check per-BCI data.
2358 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2359 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2360 if (per_bc_reason != Reason_none) {
2361 // Find the profile data for this BCI. If there isn't one,
2362 // try to allocate one from the MDO's set of spares.
2363 // This will let us detect a repeated trap at this point.
2364 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
2365
2366 if (pdata != NULL) {
2367 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2368 if (LogCompilation && xtty != NULL) {
2369 ttyLocker ttyl;
2370 // no more room for speculative traps in this MDO
2371 xtty->elem("speculative_traps_oom");
2372 }
2373 }
2374 // Query the trap state of this profile datum.
2375 int tstate0 = pdata->trap_state();
2376 if (!trap_state_has_reason(tstate0, per_bc_reason))
2377 maybe_prior_trap = false;
2378 if (!trap_state_is_recompiled(tstate0))
2379 maybe_prior_recompile = false;
2380
2381 // Update the trap state of this profile datum.
2382 int tstate1 = tstate0;
2383 // Record the reason.
2384 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2385 // Store the updated state on the MDO, for next time.
2386 if (tstate1 != tstate0)
2387 pdata->set_trap_state(tstate1);
2388 } else {
2389 if (LogCompilation && xtty != NULL) {
2390 ttyLocker ttyl;
2391 // Missing MDP? Leave a small complaint in the log.
2392 xtty->elem("missing_mdp bci='%d'", trap_bci);
2393 }
2394 }
2395 }
2396
2397 // Return results:
2398 ret_this_trap_count = this_trap_count;
2399 ret_maybe_prior_trap = maybe_prior_trap;
2400 ret_maybe_prior_recompile = maybe_prior_recompile;
2401 return pdata;
2402 }
2403
2404 void
2405 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2406 ResourceMark rm;
2407 // Ignored outputs:
2408 uint ignore_this_trap_count;
2409 bool ignore_maybe_prior_trap;
2410 bool ignore_maybe_prior_recompile;
2411 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2412 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2413 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2414 query_update_method_data(trap_mdo, trap_bci,
2415 (DeoptReason)reason,
2416 update_total_counts,
2417 #if INCLUDE_JVMCI
2418 false,
2419 #endif
2420 NULL,
2421 ignore_this_trap_count,
2422 ignore_maybe_prior_trap,
2423 ignore_maybe_prior_recompile);
2424 }
2425
2426 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2427 if (TraceDeoptimization) {
2428 tty->print("Uncommon trap ");
2429 }
2430 // Still in Java no safepoints
2431 {
2432 // This enters VM and may safepoint
2433 uncommon_trap_inner(thread, trap_request);
2434 }
2435 return fetch_unroll_info_helper(thread, exec_mode);
2436 }
2437
2438 // Local derived constants.
2439 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2440 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2441 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2442
2443 //---------------------------trap_state_reason---------------------------------
2444 Deoptimization::DeoptReason
2445 Deoptimization::trap_state_reason(int trap_state) {
2446 // This assert provides the link between the width of DataLayout::trap_bits
2447 // and the encoding of "recorded" reasons. It ensures there are enough
2448 // bits to store all needed reasons in the per-BCI MDO profile.
2449 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2450 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2451 trap_state -= recompile_bit;
2452 if (trap_state == DS_REASON_MASK) {
2453 return Reason_many;
2454 } else {
2455 assert((int)Reason_none == 0, "state=0 => Reason_none");
2456 return (DeoptReason)trap_state;
2457 }
2458 }
2459 //-------------------------trap_state_has_reason-------------------------------
2460 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2461 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2462 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2463 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2464 trap_state -= recompile_bit;
2465 if (trap_state == DS_REASON_MASK) {
2466 return -1; // true, unspecifically (bottom of state lattice)
2467 } else if (trap_state == reason) {
2468 return 1; // true, definitely
2469 } else if (trap_state == 0) {
2470 return 0; // false, definitely (top of state lattice)
2471 } else {
2472 return 0; // false, definitely
2473 }
2474 }
2475 //-------------------------trap_state_add_reason-------------------------------
2476 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2477 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2478 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2479 trap_state -= recompile_bit;
2480 if (trap_state == DS_REASON_MASK) {
2481 return trap_state + recompile_bit; // already at state lattice bottom
2482 } else if (trap_state == reason) {
2483 return trap_state + recompile_bit; // the condition is already true
2484 } else if (trap_state == 0) {
2485 return reason + recompile_bit; // no condition has yet been true
2486 } else {
2487 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2488 }
2489 }
2490 //-----------------------trap_state_is_recompiled------------------------------
2491 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2492 return (trap_state & DS_RECOMPILE_BIT) != 0;
2493 }
2494 //-----------------------trap_state_set_recompiled-----------------------------
2495 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2496 if (z) return trap_state | DS_RECOMPILE_BIT;
2497 else return trap_state & ~DS_RECOMPILE_BIT;
2498 }
2499 //---------------------------format_trap_state---------------------------------
2500 // This is used for debugging and diagnostics, including LogFile output.
2501 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2502 int trap_state) {
2503 assert(buflen > 0, "sanity");
2504 DeoptReason reason = trap_state_reason(trap_state);
2505 bool recomp_flag = trap_state_is_recompiled(trap_state);
2506 // Re-encode the state from its decoded components.
2507 int decoded_state = 0;
2508 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2509 decoded_state = trap_state_add_reason(decoded_state, reason);
2510 if (recomp_flag)
2511 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2512 // If the state re-encodes properly, format it symbolically.
2513 // Because this routine is used for debugging and diagnostics,
2514 // be robust even if the state is a strange value.
2515 size_t len;
2516 if (decoded_state != trap_state) {
2517 // Random buggy state that doesn't decode??
2518 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2519 } else {
2520 len = jio_snprintf(buf, buflen, "%s%s",
2521 trap_reason_name(reason),
2522 recomp_flag ? " recompiled" : "");
2523 }
2524 return buf;
2525 }
2526
2527
2528 //--------------------------------statics--------------------------------------
2529 const char* Deoptimization::_trap_reason_name[] = {
2530 // Note: Keep this in sync. with enum DeoptReason.
2531 "none",
2532 "null_check",
2533 "null_assert" JVMCI_ONLY("_or_unreached0"),
2534 "range_check",
2535 "class_check",
2536 "array_check",
2537 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2538 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2539 "profile_predicate",
2540 "unloaded",
2541 "uninitialized",
2542 "initialized",
2543 "unreached",
2544 "unhandled",
2545 "constraint",
2546 "div0_check",
2547 "age",
2548 "predicate",
2549 "loop_limit_check",
2550 "speculate_class_check",
2551 "speculate_null_check",
2552 "speculate_null_assert",
2553 "rtm_state_change",
2554 "unstable_if",
2555 "unstable_fused_if",
2556 #if INCLUDE_JVMCI
2557 "aliasing",
2558 "transfer_to_interpreter",
2559 "not_compiled_exception_handler",
2560 "unresolved",
2561 "jsr_mismatch",
2562 #endif
2563 "tenured"
2564 };
2565 const char* Deoptimization::_trap_action_name[] = {
2566 // Note: Keep this in sync. with enum DeoptAction.
2567 "none",
2568 "maybe_recompile",
2569 "reinterpret",
2570 "make_not_entrant",
2571 "make_not_compilable"
2572 };
2573
2574 const char* Deoptimization::trap_reason_name(int reason) {
2575 // Check that every reason has a name
2576 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2577
2578 if (reason == Reason_many) return "many";
2579 if ((uint)reason < Reason_LIMIT)
2580 return _trap_reason_name[reason];
2581 static char buf[20];
2582 sprintf(buf, "reason%d", reason);
2583 return buf;
2584 }
2585 const char* Deoptimization::trap_action_name(int action) {
2586 // Check that every action has a name
2587 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2588
2589 if ((uint)action < Action_LIMIT)
2590 return _trap_action_name[action];
2591 static char buf[20];
2592 sprintf(buf, "action%d", action);
2593 return buf;
2594 }
2595
2596 // This is used for debugging and diagnostics, including LogFile output.
2597 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2598 int trap_request) {
2599 jint unloaded_class_index = trap_request_index(trap_request);
2600 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2601 const char* action = trap_action_name(trap_request_action(trap_request));
2602 #if INCLUDE_JVMCI
2603 int debug_id = trap_request_debug_id(trap_request);
2604 #endif
2605 size_t len;
2606 if (unloaded_class_index < 0) {
2607 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2608 reason, action
2609 #if INCLUDE_JVMCI
2610 ,debug_id
2611 #endif
2612 );
2613 } else {
2614 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2615 reason, action, unloaded_class_index
2616 #if INCLUDE_JVMCI
2617 ,debug_id
2618 #endif
2619 );
2620 }
2621 return buf;
2622 }
2623
2624 juint Deoptimization::_deoptimization_hist
2625 [Deoptimization::Reason_LIMIT]
2626 [1 + Deoptimization::Action_LIMIT]
2627 [Deoptimization::BC_CASE_LIMIT]
2628 = {0};
2629
2630 enum {
2631 LSB_BITS = 8,
2632 LSB_MASK = right_n_bits(LSB_BITS)
2633 };
2634
2635 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2636 Bytecodes::Code bc) {
2637 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2638 assert(action >= 0 && action < Action_LIMIT, "oob");
2639 _deoptimization_hist[Reason_none][0][0] += 1; // total
2640 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2641 juint* cases = _deoptimization_hist[reason][1+action];
2642 juint* bc_counter_addr = NULL;
2643 juint bc_counter = 0;
2644 // Look for an unused counter, or an exact match to this BC.
2645 if (bc != Bytecodes::_illegal) {
2646 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2647 juint* counter_addr = &cases[bc_case];
2648 juint counter = *counter_addr;
2649 if ((counter == 0 && bc_counter_addr == NULL)
2650 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2651 // this counter is either free or is already devoted to this BC
2652 bc_counter_addr = counter_addr;
2653 bc_counter = counter | bc;
2654 }
2655 }
2656 }
2657 if (bc_counter_addr == NULL) {
2658 // Overflow, or no given bytecode.
2659 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2660 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2661 }
2662 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2663 }
2664
2665 jint Deoptimization::total_deoptimization_count() {
2666 return _deoptimization_hist[Reason_none][0][0];
2667 }
2668
2669 void Deoptimization::print_statistics() {
2670 juint total = total_deoptimization_count();
2671 juint account = total;
2672 if (total != 0) {
2673 ttyLocker ttyl;
2674 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2675 tty->print_cr("Deoptimization traps recorded:");
2676 #define PRINT_STAT_LINE(name, r) \
2677 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2678 PRINT_STAT_LINE("total", total);
2679 // For each non-zero entry in the histogram, print the reason,
2680 // the action, and (if specifically known) the type of bytecode.
2681 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2682 for (int action = 0; action < Action_LIMIT; action++) {
2683 juint* cases = _deoptimization_hist[reason][1+action];
2684 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2685 juint counter = cases[bc_case];
2686 if (counter != 0) {
2687 char name[1*K];
2688 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2689 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2690 bc = Bytecodes::_illegal;
2691 sprintf(name, "%s/%s/%s",
2692 trap_reason_name(reason),
2693 trap_action_name(action),
2694 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2695 juint r = counter >> LSB_BITS;
2696 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2697 account -= r;
2698 }
2699 }
2700 }
2701 }
2702 if (account != 0) {
2703 PRINT_STAT_LINE("unaccounted", account);
2704 }
2705 #undef PRINT_STAT_LINE
2706 if (xtty != NULL) xtty->tail("statistics");
2707 }
2708 }
2709
2710 #ifdef ASSERT
2711 // Revert optimizations based on escape analysis.
2712 void Deoptimization::deoptimize_objects_alot_loop() {
2713 JavaThread* ct = JavaThread::current();
2714 HandleMark hm(ct);
2715 if (DeoptimizeObjectsALotThreadCount == 1) {
2716 // Revert everything at once
2717 while (!ct->is_terminated()) {
2718 { // Begin new scope for escape barrier
2719 HandleMarkCleaner hmc(ct);
2720 ResourceMark rm(ct);
2721 EscapeBarrier eb(ct, true);
2722 eb.deoptimize_objects_all_threads();
2723 }
2724 // Now sleep after the escape barriers destructor resumed the java threads.
2725 ct->sleep(DeoptimizeObjectsALotInterval);
2726 }
2727 } else {
2728 // Revert everything for one deoptee_thread which gets selected round robin
2729 JavaThread* deoptee_thread = NULL;
2730 while (!ct->is_terminated()) {
2731 { // Begin new scope for escape barrier
2732 HandleMarkCleaner hmc(ct);
2733 ResourceMark rm(ct);
2734 ThreadsListHandle tlh;
2735 int idx = (1 + tlh.list()->find_index_of_JavaThread(deoptee_thread)) % tlh.list()->length();
2736 deoptee_thread = tlh.list()->thread_at(idx);
2737 EscapeBarrier eb(ct, deoptee_thread, true);
2738 eb.deoptimize_objects(100);
2739 }
2740 // Now sleep after the escape barriers destructor resumed deoptee_thread.
2741 ct->sleep(DeoptimizeObjectsALotInterval);
2742 }
2743 }
2744 }
2745 #endif // !ASSERT
2746
2747 // Returns true iff objects were reallocated and relocked because of access through JVMTI
2748 bool EscapeBarrier::objs_are_deoptimized(JavaThread* thread, intptr_t* fr_id) {
2749 // first/oldest update holds the flag
2750 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = JvmtiDeferredUpdates::deferred_locals(thread);
2751 bool result = false;
2752 if (list != NULL) {
2753 for (int i = 0; i < list->length(); i++) {
2754 if (list->at(i)->matches(fr_id)) {
2755 result = list->at(i)->objects_are_deoptimized();
2756 break;
2757 }
2758 }
2759 }
2760 return result;
2761 }
2762
2763 // Deoptimize frames with non escaping objects. Deoptimize objects with optimizations based on
2764 // escape analysis. Do it for all frames within the given depth and continue from there until the
2765 // entry frame is reached, because thread local objects passed as arguments might escape from callee
2766 // frames within the given depth.
2767 bool EscapeBarrier::deoptimize_objects(int depth) {
2768 if (barrier_active() && deoptee_thread()->has_last_Java_frame()) {
2769 ResourceMark rm(calling_thread());
2770 HandleMark hm;
2771 RegisterMap reg_map(deoptee_thread());
2772 vframe* vf = deoptee_thread()->last_java_vframe(®_map);
2773 int cur_depth = 0;
2774 while (vf != NULL && ((cur_depth <= depth) || !vf->is_entry_frame())) {
2775 if (vf->is_compiled_frame()) {
2776 compiledVFrame* cvf = compiledVFrame::cast(vf);
2777 // Deoptimize frame and local objects if any exist.
2778 // If cvf is deeper than depth, then we deoptimize iff local objects are passed as args.
2779 bool should_deopt = cur_depth <= depth ? cvf->not_global_escape_in_scope() : cvf->arg_escape();
2780 if (should_deopt && !deoptimize_objects(cvf->fr().id())) {
2781 // reallocation of scalar replaced objects failed, because heap is exhausted
2782 return false;
2783 }
2784 }
2785
2786 // move to next physical frame
2787 while(!vf->is_top()) {
2788 cur_depth++;
2789 vf = vf->sender();
2790 }
2791 cur_depth++;
2792 vf = vf->sender();
2793 }
2794 }
2795 return true;
2796 }
2797
2798 bool EscapeBarrier::deoptimize_objects_all_threads() {
2799 if (!barrier_active()) return true;
2800 ResourceMark rm(calling_thread());
2801 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2802 if (jt->has_last_Java_frame()) {
2803 RegisterMap reg_map(jt);
2804 vframe* vf = jt->last_java_vframe(®_map);
2805 assert(jt->frame_anchor()->walkable(),
2806 "The stack of JavaThread " PTR_FORMAT " is not walkable. Thread state is %d",
2807 p2i(jt), jt->thread_state());
2808 while (vf != NULL) {
2809 if (vf->is_compiled_frame()) {
2810 compiledVFrame* cvf = compiledVFrame::cast(vf);
2811 if ((cvf->not_global_escape_in_scope() || cvf->arg_escape()) &&
2812 !deoptimize_objects_internal(jt, cvf->fr().id())) {
2813 return false; // reallocation failure
2814 }
2815 }
2816 // move to next physical frame
2817 while(!vf->is_top()) {
2818 vf = vf->sender();
2819 }
2820 vf = vf->sender();
2821 }
2822 }
2823 }
2824 return true; // success
2825 }
2826
2827 bool EscapeBarrier::_deoptimizing_objects_for_all_threads = false;
2828 bool EscapeBarrier::_self_deoptimization_in_progress = false;
2829
2830 class EscapeBarrierSuspendHandshake : public HandshakeClosure {
2831 JavaThread* _excluded_thread;
2832 public:
2833 EscapeBarrierSuspendHandshake(JavaThread* excluded_thread, const char* name) : HandshakeClosure(name), _excluded_thread(excluded_thread) { }
2834 void do_thread(Thread* th) {
2835 if (th->is_Java_thread() && !th->is_hidden_from_external_view() && (th != _excluded_thread)) {
2836 th->set_ea_obj_deopt_flag();
2837 }
2838 }
2839 };
2840
2841 void EscapeBarrier::sync_and_suspend_one() {
2842 assert(_calling_thread != NULL, "calling thread must not be NULL");
2843 assert(_deoptee_thread != NULL, "deoptee thread must not be NULL");
2844 assert(barrier_active(), "should not call");
2845
2846 // Sync with other threads that might be doing deoptimizations
2847 {
2848 // Need to switch to _thread_blocked for the wait() call
2849 ThreadBlockInVM tbivm(_calling_thread);
2850 MonitorLocker ml(_calling_thread, EscapeBarrier_lock, Mutex::_no_safepoint_check_flag);
2851 while (_self_deoptimization_in_progress || _deoptee_thread->is_ea_obj_deopt_suspend()) {
2852 ml.wait();
2853 }
2854
2855 if (self_deopt()) {
2856 _self_deoptimization_in_progress = true;
2857 return;
2858 }
2859
2860 // set suspend flag for target thread
2861 _deoptee_thread->set_ea_obj_deopt_flag();
2862 }
2863
2864 // suspend target thread
2865 uint32_t debug_bits = 0;
2866 if (!_deoptee_thread->is_thread_fully_suspended(false, &debug_bits)) {
2867 EscapeBarrierSuspendHandshake sh(NULL, "EscapeBarrierSuspendOne");
2868 Handshake::execute(&sh, _deoptee_thread);
2869 }
2870 assert(!_deoptee_thread->has_last_Java_frame() || _deoptee_thread->frame_anchor()->walkable(),
2871 "stack should be walkable now");
2872 }
2873
2874 void EscapeBarrier::sync_and_suspend_all() {
2875 assert(barrier_active(), "should not call");
2876 assert(_calling_thread != NULL, "calling thread must not be NULL");
2877 assert(all_threads(), "sanity");
2878
2879 // Sync with other threads that might be doing deoptimizations
2880 {
2881 // Need to switch to _thread_blocked for the wait() call
2882 ThreadBlockInVM tbivm(_calling_thread);
2883 MonitorLocker ml(_calling_thread, EscapeBarrier_lock, Mutex::_no_safepoint_check_flag);
2884
2885 bool deopt_in_progress;
2886 do {
2887 deopt_in_progress = _self_deoptimization_in_progress;
2888 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2889 deopt_in_progress = (deopt_in_progress || jt->is_ea_obj_deopt_suspend());
2890 if (deopt_in_progress) {
2891 break;
2892 }
2893 }
2894 if (deopt_in_progress) {
2895 ml.wait(); // then check again
2896 }
2897 } while(deopt_in_progress);
2898
2899 _self_deoptimization_in_progress = true;
2900 _deoptimizing_objects_for_all_threads = true;
2901 }
2902
2903 EscapeBarrierSuspendHandshake sh(_calling_thread, "EscapeBarrierSuspendAll");
2904 Handshake::execute(&sh);
2905 #ifdef ASSERT
2906 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2907 if (jt->is_hidden_from_external_view()) continue;
2908 assert(!jt->has_last_Java_frame() || jt->frame_anchor()->walkable(),
2909 "The stack of JavaThread " PTR_FORMAT " is not walkable. Thread state is %d",
2910 p2i(jt), jt->thread_state());
2911 }
2912 #endif // ASSERT
2913 }
2914
2915 void EscapeBarrier::resume_one() {
2916 assert(barrier_active(), "should not call");
2917 assert(!all_threads(), "use resume_all()");
2918 MonitorLocker ml(_calling_thread, EscapeBarrier_lock, Mutex::_no_safepoint_check_flag);
2919 if (self_deopt()) {
2920 assert(_self_deoptimization_in_progress, "incorrect synchronization");
2921 _self_deoptimization_in_progress = false;
2922 } else {
2923 _deoptee_thread->clear_ea_obj_deopt_flag();
2924 }
2925 ml.notify_all();
2926 }
2927
2928 void EscapeBarrier::resume_all() {
2929 assert(barrier_active(), "should not call");
2930 assert(all_threads(), "use resume_one()");
2931 MonitorLocker ml(_calling_thread, EscapeBarrier_lock, Mutex::_no_safepoint_check_flag);
2932 assert(_self_deoptimization_in_progress, "incorrect synchronization");
2933 _deoptimizing_objects_for_all_threads = false;
2934 _self_deoptimization_in_progress = false;
2935 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2936 jt->clear_ea_obj_deopt_flag();
2937 }
2938 ml.notify_all();
2939 }
2940
2941 void EscapeBarrier::thread_added(JavaThread* jt) {
2942 if (!jt->is_hidden_from_external_view()) {
2943 MutexLocker ml(EscapeBarrier_lock, Mutex::_no_safepoint_check_flag);
2944 if (_deoptimizing_objects_for_all_threads) {
2945 jt->set_ea_obj_deopt_flag();
2946 } else {
2947 jt->clear_ea_obj_deopt_flag();
2948 }
2949 }
2950 }
2951
2952 // Remember that objects were reallocated and relocked for the compiled frame with the given id
2953 static void set_objs_are_deoptimized(JavaThread* thread, intptr_t* fr_id) {
2954 // set in first/oldest update
2955 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = JvmtiDeferredUpdates::deferred_locals(thread);
2956 DEBUG_ONLY(bool found = false);
2957 if (list != NULL) {
2958 for (int i = 0; i < list->length(); i++) {
2959 if (list->at(i)->matches(fr_id)) {
2960 DEBUG_ONLY(found = true);
2961 list->at(i)->set_objs_are_deoptimized();
2962 break;
2963 }
2964 }
2965 }
2966 assert(found, "variable set should exist at least for one vframe");
2967 }
2968
2969 // Deoptimize the given frame and deoptimize objects with optimizations based on escape analysis,
2970 // i.e. reallocate scalar replaced objects on the heap and relock objects if locking has been
2971 // eliminated.
2972 // Deoptimized objects are kept as JVMTI deferred updates until the compiled frame is replaced with interpreter frames.
2973 // Returns false iff at least one reallocation failed.
2974 bool EscapeBarrier::deoptimize_objects_internal(JavaThread* deoptee, intptr_t* fr_id) {
2975 if (!barrier_active()) return true;
2976
2977 JavaThread* ct = calling_thread();
2978 bool realloc_failures = false;
2979
2980 if (!objs_are_deoptimized(deoptee, fr_id)) {
2981 // Make sure the frame identified by fr_id is deoptimized and fetch its last vframe
2982 compiledVFrame* last_cvf;
2983 bool fr_is_deoptimized;
2984 do {
2985 StackFrameStream fst(deoptee);
2986 while (fst.current()->id() != fr_id && !fst.is_done()) {
2987 fst.next();
2988 }
2989 assert(fst.current()->id() == fr_id, "frame not found");
2990 assert(fst.current()->is_compiled_frame(), "only compiled frames can contain stack allocated objects");
2991 fr_is_deoptimized = fst.current()->is_deoptimized_frame();
2992 if (!fr_is_deoptimized) {
2993 // Execution must not continue in the compiled method, so we deoptimize the frame.
2994 Deoptimization::deoptimize_frame(deoptee, fr_id);
2995 } else {
2996 last_cvf = compiledVFrame::cast(vframe::new_vframe(fst.current(), fst.register_map(), deoptee));
2997 }
2998 } while(!fr_is_deoptimized);
2999
3000 // collect inlined frames
3001 compiledVFrame* cvf = last_cvf;
3002 GrowableArray<compiledVFrame*>* vfs = new GrowableArray<compiledVFrame*>(10);
3003 while (!cvf->is_top()) {
3004 vfs->push(cvf);
3005 cvf = compiledVFrame::cast(cvf->sender());
3006 }
3007 vfs->push(cvf);
3008
3009 // reallocate and relock optimized objects
3010 bool deoptimized_objects = Deoptimization::deoptimize_objects_internal(ct, vfs, realloc_failures);
3011 if (!realloc_failures && deoptimized_objects) {
3012 // now do the updates
3013 for (int frame_index = 0; frame_index < vfs->length(); frame_index++) {
3014 cvf = vfs->at(frame_index);
3015
3016 // locals
3017 GrowableArray<ScopeValue*>* scopeLocals = cvf->scope()->locals();
3018 StackValueCollection* locals = cvf->locals();
3019 if (locals != NULL) {
3020 for (int i2 = 0; i2 < locals->size(); i2++) {
3021 StackValue* var = locals->at(i2);
3022 if (var->type() == T_OBJECT && scopeLocals->at(i2)->is_object()) {
3023 jvalue val;
3024 val.l = cast_from_oop<jobject>(locals->at(i2)->get_obj()());
3025 cvf->update_local(T_OBJECT, i2, val);
3026 }
3027 }
3028 }
3029
3030 // expressions
3031 GrowableArray<ScopeValue*>* scopeExpressions = cvf->scope()->expressions();
3032 StackValueCollection* expressions = cvf->expressions();
3033 if (expressions != NULL) {
3034 for (int i2 = 0; i2 < expressions->size(); i2++) {
3035 StackValue* var = expressions->at(i2);
3036 if (var->type() == T_OBJECT && scopeExpressions->at(i2)->is_object()) {
3037 jvalue val;
3038 val.l = cast_from_oop<jobject>(expressions->at(i2)->get_obj()());
3039 cvf->update_stack(T_OBJECT, i2, val);
3040 }
3041 }
3042 }
3043
3044 // monitors
3045 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
3046 if (monitors != NULL) {
3047 for (int i2 = 0; i2 < monitors->length(); i2++) {
3048 if (monitors->at(i2)->eliminated()) {
3049 assert(!monitors->at(i2)->owner_is_scalar_replaced(), "reallocation failure, should not update");
3050 cvf->update_monitor(i2, monitors->at(i2));
3051 }
3052 }
3053 }
3054 }
3055 set_objs_are_deoptimized(deoptee, fr_id);
3056 }
3057 }
3058 return !realloc_failures;
3059 }
3060
3061 #else // COMPILER2_OR_JVMCI
3062
3063
3064 // Stubs for C1 only system.
3065 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
3066 return false;
3067 }
3068
3069 const char* Deoptimization::trap_reason_name(int reason) {
3070 return "unknown";
3071 }
3072
3073 void Deoptimization::print_statistics() {
3074 // no output
3075 }
3076
3077 void
3078 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
3079 // no udpate
3080 }
3081
3082 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
3083 return 0;
3084 }
3085
3086 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
3087 Bytecodes::Code bc) {
3088 // no update
3089 }
3090
3091 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
3092 int trap_state) {
3093 jio_snprintf(buf, buflen, "#%d", trap_state);
3094 return buf;
3095 }
3096
3097 #endif // COMPILER2_OR_JVMCI