1 /*
2 * Copyright (c) 1997, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "code/codeCache.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "code/nmethod.hpp"
30 #include "code/pcDesc.hpp"
31 #include "code/scopeDesc.hpp"
32 #include "interpreter/bytecode.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "interpreter/oopMapCache.hpp"
35 #include "memory/allocation.inline.hpp"
36 #include "memory/oopFactory.hpp"
37 #include "memory/resourceArea.hpp"
38 #include "oops/method.hpp"
39 #include "oops/objArrayOop.inline.hpp"
40 #include "oops/oop.inline.hpp"
41 #include "oops/fieldStreams.hpp"
42 #include "oops/verifyOopClosure.hpp"
43 #include "prims/jvmtiThreadState.hpp"
44 #include "runtime/biasedLocking.hpp"
45 #include "runtime/compilationPolicy.hpp"
46 #include "runtime/deoptimization.hpp"
47 #include "runtime/interfaceSupport.hpp"
48 #include "runtime/sharedRuntime.hpp"
49 #include "runtime/signature.hpp"
50 #include "runtime/stubRoutines.hpp"
51 #include "runtime/thread.hpp"
52 #include "runtime/vframe.hpp"
53 #include "runtime/vframeArray.hpp"
54 #include "runtime/vframe_hp.hpp"
55 #include "utilities/events.hpp"
56 #include "utilities/xmlstream.hpp"
57
58 #if INCLUDE_JVMCI
59 #include "jvmci/jvmciRuntime.hpp"
60 #include "jvmci/jvmciJavaClasses.hpp"
61 #endif
62
63
64 bool DeoptimizationMarker::_is_active = false;
65
66 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
67 int caller_adjustment,
68 int caller_actual_parameters,
69 int number_of_frames,
70 intptr_t* frame_sizes,
71 address* frame_pcs,
72 BasicType return_type,
73 int exec_mode) {
74 _size_of_deoptimized_frame = size_of_deoptimized_frame;
75 _caller_adjustment = caller_adjustment;
76 _caller_actual_parameters = caller_actual_parameters;
77 _number_of_frames = number_of_frames;
78 _frame_sizes = frame_sizes;
79 _frame_pcs = frame_pcs;
80 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
81 _return_type = return_type;
82 _initial_info = 0;
83 // PD (x86 only)
84 _counter_temp = 0;
85 _unpack_kind = exec_mode;
86 _sender_sp_temp = 0;
87
88 _total_frame_sizes = size_of_frames();
89 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
90 }
91
92
93 Deoptimization::UnrollBlock::~UnrollBlock() {
94 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
95 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
96 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
97 }
98
99
100 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
101 assert(register_number < RegisterMap::reg_count, "checking register number");
102 return &_register_block[register_number * 2];
103 }
104
105
106
107 int Deoptimization::UnrollBlock::size_of_frames() const {
108 // Acount first for the adjustment of the initial frame
109 int result = _caller_adjustment;
110 for (int index = 0; index < number_of_frames(); index++) {
111 result += frame_sizes()[index];
112 }
113 return result;
114 }
115
116
117 void Deoptimization::UnrollBlock::print() {
118 ttyLocker ttyl;
119 tty->print_cr("UnrollBlock");
120 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
121 tty->print( " frame_sizes: ");
122 for (int index = 0; index < number_of_frames(); index++) {
123 tty->print(INTX_FORMAT " ", frame_sizes()[index]);
124 }
125 tty->cr();
126 }
127
128
129 // In order to make fetch_unroll_info work properly with escape
130 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
131 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
132 // of previously eliminated objects occurs in realloc_objects, which is
133 // called from the method fetch_unroll_info_helper below.
134 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode))
135 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
136 // but makes the entry a little slower. There is however a little dance we have to
137 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
138
139 // fetch_unroll_info() is called at the beginning of the deoptimization
140 // handler. Note this fact before we start generating temporary frames
141 // that can confuse an asynchronous stack walker. This counter is
142 // decremented at the end of unpack_frames().
143 if (TraceDeoptimization) {
144 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread));
145 }
146 thread->inc_in_deopt_handler();
147
148 return fetch_unroll_info_helper(thread, exec_mode);
149 JRT_END
150
151
152 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
153 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) {
154
155 // Note: there is a safepoint safety issue here. No matter whether we enter
156 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
157 // the vframeArray is created.
158 //
159
160 // Allocate our special deoptimization ResourceMark
161 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
162 assert(thread->deopt_mark() == NULL, "Pending deopt!");
163 thread->set_deopt_mark(dmark);
164
165 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
166 RegisterMap map(thread, true);
167 RegisterMap dummy_map(thread, false);
168 // Now get the deoptee with a valid map
169 frame deoptee = stub_frame.sender(&map);
170 // Set the deoptee nmethod
171 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!");
172 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
173 thread->set_deopt_compiled_method(cm);
174
175 if (VerifyStack) {
176 thread->validate_frame_layout();
177 }
178
179 // Create a growable array of VFrames where each VFrame represents an inlined
180 // Java frame. This storage is allocated with the usual system arena.
181 assert(deoptee.is_compiled_frame(), "Wrong frame type");
182 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
183 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
184 while (!vf->is_top()) {
185 assert(vf->is_compiled_frame(), "Wrong frame type");
186 chunk->push(compiledVFrame::cast(vf));
187 vf = vf->sender();
188 }
189 assert(vf->is_compiled_frame(), "Wrong frame type");
190 chunk->push(compiledVFrame::cast(vf));
191
192 ScopeDesc* trap_scope = chunk->at(0)->scope();
193 Handle exceptionObject;
194 if (trap_scope->rethrow_exception()) {
195 if (PrintDeoptimizationDetails) {
196 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());
197 }
198 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
199 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw");
200 ScopeValue* topOfStack = expressions->top();
201 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
202 assert(exceptionObject() != NULL, "exception oop can not be null");
203 }
204
205 bool realloc_failures = false;
206
207 #if defined(COMPILER2) || INCLUDE_JVMCI
208 // Reallocate the non-escaping objects and restore their fields. Then
209 // relock objects if synchronization on them was eliminated.
210 #ifndef INCLUDE_JVMCI
211 if (DoEscapeAnalysis || EliminateNestedLocks) {
212 if (EliminateAllocations) {
213 #endif // INCLUDE_JVMCI
214 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
215 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
216
217 // The flag return_oop() indicates call sites which return oop
218 // in compiled code. Such sites include java method calls,
219 // runtime calls (for example, used to allocate new objects/arrays
220 // on slow code path) and any other calls generated in compiled code.
221 // It is not guaranteed that we can get such information here only
222 // by analyzing bytecode in deoptimized frames. This is why this flag
223 // is set during method compilation (see Compile::Process_OopMap_Node()).
224 // If the previous frame was popped or if we are dispatching an exception,
225 // we don't have an oop result.
226 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt);
227 Handle return_value;
228 if (save_oop_result) {
229 // Reallocation may trigger GC. If deoptimization happened on return from
230 // call which returns oop we need to save it since it is not in oopmap.
231 oop result = deoptee.saved_oop_result(&map);
232 assert(result == NULL || result->is_oop(), "must be oop");
233 return_value = Handle(thread, result);
234 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
235 if (TraceDeoptimization) {
236 ttyLocker ttyl;
237 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
238 }
239 }
240 if (objects != NULL) {
241 JRT_BLOCK
242 realloc_failures = realloc_objects(thread, &deoptee, objects, THREAD);
243 JRT_END
244 bool skip_internal = (cm != NULL) && !cm->is_compiled_by_jvmci();
245 reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
246 #ifndef PRODUCT
247 if (TraceDeoptimization) {
248 ttyLocker ttyl;
249 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
250 print_objects(objects, realloc_failures);
251 }
252 #endif
253 }
254 if (save_oop_result) {
255 // Restore result.
256 deoptee.set_saved_oop_result(&map, return_value());
257 }
258 #ifndef INCLUDE_JVMCI
259 }
260 if (EliminateLocks) {
261 #endif // INCLUDE_JVMCI
262 #ifndef PRODUCT
263 bool first = true;
264 #endif
265 for (int i = 0; i < chunk->length(); i++) {
266 compiledVFrame* cvf = chunk->at(i);
267 assert (cvf->scope() != NULL,"expect only compiled java frames");
268 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
269 if (monitors->is_nonempty()) {
270 relock_objects(monitors, thread, realloc_failures);
271 #ifndef PRODUCT
272 if (PrintDeoptimizationDetails) {
273 ttyLocker ttyl;
274 for (int j = 0; j < monitors->length(); j++) {
275 MonitorInfo* mi = monitors->at(j);
276 if (mi->eliminated()) {
277 if (first) {
278 first = false;
279 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
280 }
281 if (mi->owner_is_scalar_replaced()) {
282 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
283 tty->print_cr(" failed reallocation for klass %s", k->external_name());
284 } else {
285 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
286 }
287 }
288 }
289 }
290 #endif // !PRODUCT
291 }
292 }
293 #ifndef INCLUDE_JVMCI
294 }
295 }
296 #endif // INCLUDE_JVMCI
297 #endif // COMPILER2 || INCLUDE_JVMCI
298
299 // Ensure that no safepoint is taken after pointers have been stored
300 // in fields of rematerialized objects. If a safepoint occurs from here on
301 // out the java state residing in the vframeArray will be missed.
302 NoSafepointVerifier no_safepoint;
303
304 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
305 #if defined(COMPILER2) || INCLUDE_JVMCI
306 if (realloc_failures) {
307 pop_frames_failed_reallocs(thread, array);
308 }
309 #endif
310
311 assert(thread->vframe_array_head() == NULL, "Pending deopt!");
312 thread->set_vframe_array_head(array);
313
314 // Now that the vframeArray has been created if we have any deferred local writes
315 // added by jvmti then we can free up that structure as the data is now in the
316 // vframeArray
317
318 if (thread->deferred_locals() != NULL) {
319 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
320 int i = 0;
321 do {
322 // Because of inlining we could have multiple vframes for a single frame
323 // and several of the vframes could have deferred writes. Find them all.
324 if (list->at(i)->id() == array->original().id()) {
325 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
326 list->remove_at(i);
327 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
328 delete dlv;
329 } else {
330 i++;
331 }
332 } while ( i < list->length() );
333 if (list->length() == 0) {
334 thread->set_deferred_locals(NULL);
335 // free the list and elements back to C heap.
336 delete list;
337 }
338
339 }
340
341 #ifndef SHARK
342 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
343 CodeBlob* cb = stub_frame.cb();
344 // Verify we have the right vframeArray
345 assert(cb->frame_size() >= 0, "Unexpected frame size");
346 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
347
348 // If the deopt call site is a MethodHandle invoke call site we have
349 // to adjust the unpack_sp.
350 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
351 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
352 unpack_sp = deoptee.unextended_sp();
353
354 #ifdef ASSERT
355 assert(cb->is_deoptimization_stub() ||
356 cb->is_uncommon_trap_stub() ||
357 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
358 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
359 "unexpected code blob: %s", cb->name());
360 #endif
361 #else
362 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
363 #endif // !SHARK
364
365 // This is a guarantee instead of an assert because if vframe doesn't match
366 // we will unpack the wrong deoptimized frame and wind up in strange places
367 // where it will be very difficult to figure out what went wrong. Better
368 // to die an early death here than some very obscure death later when the
369 // trail is cold.
370 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
371 // in that it will fail to detect a problem when there is one. This needs
372 // more work in tiger timeframe.
373 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
374
375 int number_of_frames = array->frames();
376
377 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
378 // virtual activation, which is the reverse of the elements in the vframes array.
379 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
380 // +1 because we always have an interpreter return address for the final slot.
381 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
382 int popframe_extra_args = 0;
383 // Create an interpreter return address for the stub to use as its return
384 // address so the skeletal frames are perfectly walkable
385 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
386
387 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
388 // activation be put back on the expression stack of the caller for reexecution
389 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
390 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
391 }
392
393 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
394 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
395 // than simply use array->sender.pc(). This requires us to walk the current set of frames
396 //
397 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
398 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
399
400 // It's possible that the number of parameters at the call site is
401 // different than number of arguments in the callee when method
402 // handles are used. If the caller is interpreted get the real
403 // value so that the proper amount of space can be added to it's
404 // frame.
405 bool caller_was_method_handle = false;
406 if (deopt_sender.is_interpreted_frame()) {
407 methodHandle method = deopt_sender.interpreter_frame_method();
408 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
409 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
410 // Method handle invokes may involve fairly arbitrary chains of
411 // calls so it's impossible to know how much actual space the
412 // caller has for locals.
413 caller_was_method_handle = true;
414 }
415 }
416
417 //
418 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
419 // frame_sizes/frame_pcs[1] next oldest frame (int)
420 // frame_sizes/frame_pcs[n] youngest frame (int)
421 //
422 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
423 // owns the space for the return address to it's caller). Confusing ain't it.
424 //
425 // The vframe array can address vframes with indices running from
426 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
427 // When we create the skeletal frames we need the oldest frame to be in the zero slot
428 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
429 // so things look a little strange in this loop.
430 //
431 int callee_parameters = 0;
432 int callee_locals = 0;
433 for (int index = 0; index < array->frames(); index++ ) {
434 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
435 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
436 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
437 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
438 callee_locals,
439 index == 0,
440 popframe_extra_args);
441 // This pc doesn't have to be perfect just good enough to identify the frame
442 // as interpreted so the skeleton frame will be walkable
443 // The correct pc will be set when the skeleton frame is completely filled out
444 // The final pc we store in the loop is wrong and will be overwritten below
445 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
446
447 callee_parameters = array->element(index)->method()->size_of_parameters();
448 callee_locals = array->element(index)->method()->max_locals();
449 popframe_extra_args = 0;
450 }
451
452 // Compute whether the root vframe returns a float or double value.
453 BasicType return_type;
454 {
455 HandleMark hm;
456 methodHandle method(thread, array->element(0)->method());
457 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
458 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
459 }
460
461 // Compute information for handling adapters and adjusting the frame size of the caller.
462 int caller_adjustment = 0;
463
464 // Compute the amount the oldest interpreter frame will have to adjust
465 // its caller's stack by. If the caller is a compiled frame then
466 // we pretend that the callee has no parameters so that the
467 // extension counts for the full amount of locals and not just
468 // locals-parms. This is because without a c2i adapter the parm
469 // area as created by the compiled frame will not be usable by
470 // the interpreter. (Depending on the calling convention there
471 // may not even be enough space).
472
473 // QQQ I'd rather see this pushed down into last_frame_adjust
474 // and have it take the sender (aka caller).
475
476 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
477 caller_adjustment = last_frame_adjust(0, callee_locals);
478 } else if (callee_locals > callee_parameters) {
479 // The caller frame may need extending to accommodate
480 // non-parameter locals of the first unpacked interpreted frame.
481 // Compute that adjustment.
482 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
483 }
484
485 // If the sender is deoptimized the we must retrieve the address of the handler
486 // since the frame will "magically" show the original pc before the deopt
487 // and we'd undo the deopt.
488
489 frame_pcs[0] = deopt_sender.raw_pc();
490
491 #ifndef SHARK
492 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
493 #endif // SHARK
494
495 #ifdef INCLUDE_JVMCI
496 if (exceptionObject() != NULL) {
497 thread->set_exception_oop(exceptionObject());
498 exec_mode = Unpack_exception;
499 }
500 #endif
501
502 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
503 assert(thread->has_pending_exception(), "should have thrown OOME");
504 thread->set_exception_oop(thread->pending_exception());
505 thread->clear_pending_exception();
506 exec_mode = Unpack_exception;
507 }
508
509 #if INCLUDE_JVMCI
510 if (thread->frames_to_pop_failed_realloc() > 0) {
511 thread->set_pending_monitorenter(false);
512 }
513 #endif
514
515 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
516 caller_adjustment * BytesPerWord,
517 caller_was_method_handle ? 0 : callee_parameters,
518 number_of_frames,
519 frame_sizes,
520 frame_pcs,
521 return_type,
522 exec_mode);
523 // On some platforms, we need a way to pass some platform dependent
524 // information to the unpacking code so the skeletal frames come out
525 // correct (initial fp value, unextended sp, ...)
526 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
527
528 if (array->frames() > 1) {
529 if (VerifyStack && TraceDeoptimization) {
530 ttyLocker ttyl;
531 tty->print_cr("Deoptimizing method containing inlining");
532 }
533 }
534
535 array->set_unroll_block(info);
536 return info;
537 }
538
539 // Called to cleanup deoptimization data structures in normal case
540 // after unpacking to stack and when stack overflow error occurs
541 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
542 vframeArray *array) {
543
544 // Get array if coming from exception
545 if (array == NULL) {
546 array = thread->vframe_array_head();
547 }
548 thread->set_vframe_array_head(NULL);
549
550 // Free the previous UnrollBlock
551 vframeArray* old_array = thread->vframe_array_last();
552 thread->set_vframe_array_last(array);
553
554 if (old_array != NULL) {
555 UnrollBlock* old_info = old_array->unroll_block();
556 old_array->set_unroll_block(NULL);
557 delete old_info;
558 delete old_array;
559 }
560
561 // Deallocate any resource creating in this routine and any ResourceObjs allocated
562 // inside the vframeArray (StackValueCollections)
563
564 delete thread->deopt_mark();
565 thread->set_deopt_mark(NULL);
566 thread->set_deopt_compiled_method(NULL);
567
568
569 if (JvmtiExport::can_pop_frame()) {
570 #ifndef CC_INTERP
571 // Regardless of whether we entered this routine with the pending
572 // popframe condition bit set, we should always clear it now
573 thread->clear_popframe_condition();
574 #else
575 // C++ interpreter will clear has_pending_popframe when it enters
576 // with method_resume. For deopt_resume2 we clear it now.
577 if (thread->popframe_forcing_deopt_reexecution())
578 thread->clear_popframe_condition();
579 #endif /* CC_INTERP */
580 }
581
582 // unpack_frames() is called at the end of the deoptimization handler
583 // and (in C2) at the end of the uncommon trap handler. Note this fact
584 // so that an asynchronous stack walker can work again. This counter is
585 // incremented at the beginning of fetch_unroll_info() and (in C2) at
586 // the beginning of uncommon_trap().
587 thread->dec_in_deopt_handler();
588 }
589
590 // Moved from cpu directories because none of the cpus has callee save values.
591 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
592 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
593
594 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
595 // the days we had adapter frames. When we deoptimize a situation where a
596 // compiled caller calls a compiled caller will have registers it expects
597 // to survive the call to the callee. If we deoptimize the callee the only
598 // way we can restore these registers is to have the oldest interpreter
599 // frame that we create restore these values. That is what this routine
600 // will accomplish.
601
602 // At the moment we have modified c2 to not have any callee save registers
603 // so this problem does not exist and this routine is just a place holder.
604
605 assert(f->is_interpreted_frame(), "must be interpreted");
606 }
607
608 // Return BasicType of value being returned
609 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
610
611 // We are already active int he special DeoptResourceMark any ResourceObj's we
612 // allocate will be freed at the end of the routine.
613
614 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
615 // but makes the entry a little slower. There is however a little dance we have to
616 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
617 ResetNoHandleMark rnhm; // No-op in release/product versions
618 HandleMark hm;
619
620 frame stub_frame = thread->last_frame();
621
622 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
623 // must point to the vframeArray for the unpack frame.
624 vframeArray* array = thread->vframe_array_head();
625
626 #ifndef PRODUCT
627 if (TraceDeoptimization) {
628 ttyLocker ttyl;
629 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d",
630 p2i(thread), p2i(array), exec_mode);
631 }
632 #endif
633 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
634 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode);
635
636 UnrollBlock* info = array->unroll_block();
637
638 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
639 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
640
641 BasicType bt = info->return_type();
642
643 // If we have an exception pending, claim that the return type is an oop
644 // so the deopt_blob does not overwrite the exception_oop.
645
646 if (exec_mode == Unpack_exception)
647 bt = T_OBJECT;
648
649 // Cleanup thread deopt data
650 cleanup_deopt_info(thread, array);
651
652 #ifndef PRODUCT
653 if (VerifyStack) {
654 ResourceMark res_mark;
655
656 thread->validate_frame_layout();
657
658 // Verify that the just-unpacked frames match the interpreter's
659 // notions of expression stack and locals
660 vframeArray* cur_array = thread->vframe_array_last();
661 RegisterMap rm(thread, false);
662 rm.set_include_argument_oops(false);
663 bool is_top_frame = true;
664 int callee_size_of_parameters = 0;
665 int callee_max_locals = 0;
666 for (int i = 0; i < cur_array->frames(); i++) {
667 vframeArrayElement* el = cur_array->element(i);
668 frame* iframe = el->iframe();
669 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
670
671 // Get the oop map for this bci
672 InterpreterOopMap mask;
673 int cur_invoke_parameter_size = 0;
674 bool try_next_mask = false;
675 int next_mask_expression_stack_size = -1;
676 int top_frame_expression_stack_adjustment = 0;
677 methodHandle mh(thread, iframe->interpreter_frame_method());
678 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
679 BytecodeStream str(mh);
680 str.set_start(iframe->interpreter_frame_bci());
681 int max_bci = mh->code_size();
682 // Get to the next bytecode if possible
683 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
684 // Check to see if we can grab the number of outgoing arguments
685 // at an uncommon trap for an invoke (where the compiler
686 // generates debug info before the invoke has executed)
687 Bytecodes::Code cur_code = str.next();
688 if (cur_code == Bytecodes::_invokevirtual ||
689 cur_code == Bytecodes::_invokespecial ||
690 cur_code == Bytecodes::_invokestatic ||
691 cur_code == Bytecodes::_invokeinterface ||
692 cur_code == Bytecodes::_invokedynamic) {
693 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
694 Symbol* signature = invoke.signature();
695 ArgumentSizeComputer asc(signature);
696 cur_invoke_parameter_size = asc.size();
697 if (invoke.has_receiver()) {
698 // Add in receiver
699 ++cur_invoke_parameter_size;
700 }
701 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
702 callee_size_of_parameters++;
703 }
704 }
705 if (str.bci() < max_bci) {
706 Bytecodes::Code bc = str.next();
707 if (bc >= 0) {
708 // The interpreter oop map generator reports results before
709 // the current bytecode has executed except in the case of
710 // calls. It seems to be hard to tell whether the compiler
711 // has emitted debug information matching the "state before"
712 // a given bytecode or the state after, so we try both
713 switch (cur_code) {
714 case Bytecodes::_invokevirtual:
715 case Bytecodes::_invokespecial:
716 case Bytecodes::_invokestatic:
717 case Bytecodes::_invokeinterface:
718 case Bytecodes::_invokedynamic:
719 case Bytecodes::_athrow:
720 break;
721 default: {
722 InterpreterOopMap next_mask;
723 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
724 next_mask_expression_stack_size = next_mask.expression_stack_size();
725 // Need to subtract off the size of the result type of
726 // the bytecode because this is not described in the
727 // debug info but returned to the interpreter in the TOS
728 // caching register
729 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
730 if (bytecode_result_type != T_ILLEGAL) {
731 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
732 }
733 assert(top_frame_expression_stack_adjustment >= 0, "");
734 try_next_mask = true;
735 break;
736 }
737 }
738 }
739 }
740
741 // Verify stack depth and oops in frame
742 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
743 if (!(
744 /* SPARC */
745 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
746 /* x86 */
747 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
748 (try_next_mask &&
749 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
750 top_frame_expression_stack_adjustment))) ||
751 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
752 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
753 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
754 )) {
755 ttyLocker ttyl;
756
757 // Print out some information that will help us debug the problem
758 tty->print_cr("Wrong number of expression stack elements during deoptimization");
759 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
760 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
761 iframe->interpreter_frame_expression_stack_size());
762 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
763 tty->print_cr(" try_next_mask = %d", try_next_mask);
764 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
765 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
766 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
767 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
768 tty->print_cr(" exec_mode = %d", exec_mode);
769 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
770 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
771 tty->print_cr(" Interpreted frames:");
772 for (int k = 0; k < cur_array->frames(); k++) {
773 vframeArrayElement* el = cur_array->element(k);
774 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
775 }
776 cur_array->print_on_2(tty);
777 guarantee(false, "wrong number of expression stack elements during deopt");
778 }
779 VerifyOopClosure verify;
780 iframe->oops_interpreted_do(&verify, &rm, false);
781 callee_size_of_parameters = mh->size_of_parameters();
782 callee_max_locals = mh->max_locals();
783 is_top_frame = false;
784 }
785 }
786 #endif /* !PRODUCT */
787
788
789 return bt;
790 JRT_END
791
792
793 int Deoptimization::deoptimize_dependents() {
794 Threads::deoptimized_wrt_marked_nmethods();
795 return 0;
796 }
797
798 Deoptimization::DeoptAction Deoptimization::_unloaded_action
799 = Deoptimization::Action_reinterpret;
800
801 #if defined(COMPILER2) || INCLUDE_JVMCI
802 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
803 Handle pending_exception(thread->pending_exception());
804 const char* exception_file = thread->exception_file();
805 int exception_line = thread->exception_line();
806 thread->clear_pending_exception();
807
808 bool failures = false;
809
810 for (int i = 0; i < objects->length(); i++) {
811 assert(objects->at(i)->is_object(), "invalid debug information");
812 ObjectValue* sv = (ObjectValue*) objects->at(i);
813
814 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
815 oop obj = NULL;
816
817 if (k->is_instance_klass()) {
818 InstanceKlass* ik = InstanceKlass::cast(k());
819 obj = ik->allocate_instance(THREAD);
820 } else if (k->is_typeArray_klass()) {
821 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
822 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
823 int len = sv->field_size() / type2size[ak->element_type()];
824 obj = ak->allocate(len, THREAD);
825 } else if (k->is_objArray_klass()) {
826 ObjArrayKlass* ak = ObjArrayKlass::cast(k());
827 obj = ak->allocate(sv->field_size(), THREAD);
828 }
829
830 if (obj == NULL) {
831 failures = true;
832 }
833
834 assert(sv->value().is_null(), "redundant reallocation");
835 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
836 CLEAR_PENDING_EXCEPTION;
837 sv->set_value(obj);
838 }
839
840 if (failures) {
841 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
842 } else if (pending_exception.not_null()) {
843 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
844 }
845
846 return failures;
847 }
848
849 // restore elements of an eliminated type array
850 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
851 int index = 0;
852 intptr_t val;
853
854 for (int i = 0; i < sv->field_size(); i++) {
855 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
856 switch(type) {
857 case T_LONG: case T_DOUBLE: {
858 assert(value->type() == T_INT, "Agreement.");
859 StackValue* low =
860 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
861 #ifdef _LP64
862 jlong res = (jlong)low->get_int();
863 #else
864 #ifdef SPARC
865 // For SPARC we have to swap high and low words.
866 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
867 #else
868 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
869 #endif //SPARC
870 #endif
871 obj->long_at_put(index, res);
872 break;
873 }
874
875 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
876 case T_INT: case T_FLOAT: { // 4 bytes.
877 assert(value->type() == T_INT, "Agreement.");
878 bool big_value = false;
879 if (i + 1 < sv->field_size() && type == T_INT) {
880 if (sv->field_at(i)->is_location()) {
881 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
882 if (type == Location::dbl || type == Location::lng) {
883 big_value = true;
884 }
885 } else if (sv->field_at(i)->is_constant_int()) {
886 ScopeValue* next_scope_field = sv->field_at(i + 1);
887 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
888 big_value = true;
889 }
890 }
891 }
892
893 if (big_value) {
894 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
895 #ifdef _LP64
896 jlong res = (jlong)low->get_int();
897 #else
898 #ifdef SPARC
899 // For SPARC we have to swap high and low words.
900 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
901 #else
902 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
903 #endif //SPARC
904 #endif
905 obj->int_at_put(index, (jint)*((jint*)&res));
906 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
907 } else {
908 val = value->get_int();
909 obj->int_at_put(index, (jint)*((jint*)&val));
910 }
911 break;
912 }
913
914 case T_SHORT:
915 assert(value->type() == T_INT, "Agreement.");
916 val = value->get_int();
917 obj->short_at_put(index, (jshort)*((jint*)&val));
918 break;
919
920 case T_CHAR:
921 assert(value->type() == T_INT, "Agreement.");
922 val = value->get_int();
923 obj->char_at_put(index, (jchar)*((jint*)&val));
924 break;
925
926 case T_BYTE:
927 assert(value->type() == T_INT, "Agreement.");
928 val = value->get_int();
929 obj->byte_at_put(index, (jbyte)*((jint*)&val));
930 break;
931
932 case T_BOOLEAN:
933 assert(value->type() == T_INT, "Agreement.");
934 val = value->get_int();
935 obj->bool_at_put(index, (jboolean)*((jint*)&val));
936 break;
937
938 default:
939 ShouldNotReachHere();
940 }
941 index++;
942 }
943 }
944
945
946 // restore fields of an eliminated object array
947 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
948 for (int i = 0; i < sv->field_size(); i++) {
949 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
950 assert(value->type() == T_OBJECT, "object element expected");
951 obj->obj_at_put(i, value->get_obj()());
952 }
953 }
954
955 class ReassignedField {
956 public:
957 int _offset;
958 BasicType _type;
959 public:
960 ReassignedField() {
961 _offset = 0;
962 _type = T_ILLEGAL;
963 }
964 };
965
966 int compare(ReassignedField* left, ReassignedField* right) {
967 return left->_offset - right->_offset;
968 }
969
970 // Restore fields of an eliminated instance object using the same field order
971 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
972 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
973 if (klass->superklass() != NULL) {
974 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal);
975 }
976
977 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
978 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
979 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
980 ReassignedField field;
981 field._offset = fs.offset();
982 field._type = FieldType::basic_type(fs.signature());
983 fields->append(field);
984 }
985 }
986 fields->sort(compare);
987 for (int i = 0; i < fields->length(); i++) {
988 intptr_t val;
989 ScopeValue* scope_field = sv->field_at(svIndex);
990 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
991 int offset = fields->at(i)._offset;
992 BasicType type = fields->at(i)._type;
993 switch (type) {
994 case T_OBJECT: case T_ARRAY:
995 assert(value->type() == T_OBJECT, "Agreement.");
996 obj->obj_field_put(offset, value->get_obj()());
997 break;
998
999 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1000 case T_INT: case T_FLOAT: { // 4 bytes.
1001 assert(value->type() == T_INT, "Agreement.");
1002 bool big_value = false;
1003 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1004 if (scope_field->is_location()) {
1005 Location::Type type = ((LocationValue*) scope_field)->location().type();
1006 if (type == Location::dbl || type == Location::lng) {
1007 big_value = true;
1008 }
1009 }
1010 if (scope_field->is_constant_int()) {
1011 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1012 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1013 big_value = true;
1014 }
1015 }
1016 }
1017
1018 if (big_value) {
1019 i++;
1020 assert(i < fields->length(), "second T_INT field needed");
1021 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1022 } else {
1023 val = value->get_int();
1024 obj->int_field_put(offset, (jint)*((jint*)&val));
1025 break;
1026 }
1027 }
1028 /* no break */
1029
1030 case T_LONG: case T_DOUBLE: {
1031 assert(value->type() == T_INT, "Agreement.");
1032 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1033 #ifdef _LP64
1034 jlong res = (jlong)low->get_int();
1035 #else
1036 #ifdef SPARC
1037 // For SPARC we have to swap high and low words.
1038 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1039 #else
1040 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1041 #endif //SPARC
1042 #endif
1043 obj->long_field_put(offset, res);
1044 break;
1045 }
1046
1047 case T_SHORT:
1048 assert(value->type() == T_INT, "Agreement.");
1049 val = value->get_int();
1050 obj->short_field_put(offset, (jshort)*((jint*)&val));
1051 break;
1052
1053 case T_CHAR:
1054 assert(value->type() == T_INT, "Agreement.");
1055 val = value->get_int();
1056 obj->char_field_put(offset, (jchar)*((jint*)&val));
1057 break;
1058
1059 case T_BYTE:
1060 assert(value->type() == T_INT, "Agreement.");
1061 val = value->get_int();
1062 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1063 break;
1064
1065 case T_BOOLEAN:
1066 assert(value->type() == T_INT, "Agreement.");
1067 val = value->get_int();
1068 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1069 break;
1070
1071 default:
1072 ShouldNotReachHere();
1073 }
1074 svIndex++;
1075 }
1076 return svIndex;
1077 }
1078
1079 // restore fields of all eliminated objects and arrays
1080 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1081 for (int i = 0; i < objects->length(); i++) {
1082 ObjectValue* sv = (ObjectValue*) objects->at(i);
1083 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
1084 Handle obj = sv->value();
1085 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1086 if (PrintDeoptimizationDetails) {
1087 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1088 }
1089 if (obj.is_null()) {
1090 continue;
1091 }
1092
1093 if (k->is_instance_klass()) {
1094 InstanceKlass* ik = InstanceKlass::cast(k());
1095 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1096 } else if (k->is_typeArray_klass()) {
1097 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
1098 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1099 } else if (k->is_objArray_klass()) {
1100 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1101 }
1102 }
1103 }
1104
1105
1106 // relock objects for which synchronization was eliminated
1107 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
1108 for (int i = 0; i < monitors->length(); i++) {
1109 MonitorInfo* mon_info = monitors->at(i);
1110 if (mon_info->eliminated()) {
1111 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1112 if (!mon_info->owner_is_scalar_replaced()) {
1113 Handle obj = Handle(mon_info->owner());
1114 markOop mark = obj->mark();
1115 if (UseBiasedLocking && mark->has_bias_pattern()) {
1116 // New allocated objects may have the mark set to anonymously biased.
1117 // Also the deoptimized method may called methods with synchronization
1118 // where the thread-local object is bias locked to the current thread.
1119 assert(mark->is_biased_anonymously() ||
1120 mark->biased_locker() == thread, "should be locked to current thread");
1121 // Reset mark word to unbiased prototype.
1122 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
1123 obj->set_mark(unbiased_prototype);
1124 }
1125 BasicLock* lock = mon_info->lock();
1126 ObjectSynchronizer::slow_enter(obj, lock, thread);
1127 assert(mon_info->owner()->is_locked(), "object must be locked now");
1128 }
1129 }
1130 }
1131 }
1132
1133
1134 #ifndef PRODUCT
1135 // print information about reallocated objects
1136 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1137 fieldDescriptor fd;
1138
1139 for (int i = 0; i < objects->length(); i++) {
1140 ObjectValue* sv = (ObjectValue*) objects->at(i);
1141 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
1142 Handle obj = sv->value();
1143
1144 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1145 k->print_value();
1146 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1147 if (obj.is_null()) {
1148 tty->print(" allocation failed");
1149 } else {
1150 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1151 }
1152 tty->cr();
1153
1154 if (Verbose && !obj.is_null()) {
1155 k->oop_print_on(obj(), tty);
1156 }
1157 }
1158 }
1159 #endif
1160 #endif // COMPILER2 || INCLUDE_JVMCI
1161
1162 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1163 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1164
1165 #ifndef PRODUCT
1166 if (PrintDeoptimizationDetails) {
1167 ttyLocker ttyl;
1168 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1169 fr.print_on(tty);
1170 tty->print_cr(" Virtual frames (innermost first):");
1171 for (int index = 0; index < chunk->length(); index++) {
1172 compiledVFrame* vf = chunk->at(index);
1173 tty->print(" %2d - ", index);
1174 vf->print_value();
1175 int bci = chunk->at(index)->raw_bci();
1176 const char* code_name;
1177 if (bci == SynchronizationEntryBCI) {
1178 code_name = "sync entry";
1179 } else {
1180 Bytecodes::Code code = vf->method()->code_at(bci);
1181 code_name = Bytecodes::name(code);
1182 }
1183 tty->print(" - %s", code_name);
1184 tty->print_cr(" @ bci %d ", bci);
1185 if (Verbose) {
1186 vf->print();
1187 tty->cr();
1188 }
1189 }
1190 }
1191 #endif
1192
1193 // Register map for next frame (used for stack crawl). We capture
1194 // the state of the deopt'ing frame's caller. Thus if we need to
1195 // stuff a C2I adapter we can properly fill in the callee-save
1196 // register locations.
1197 frame caller = fr.sender(reg_map);
1198 int frame_size = caller.sp() - fr.sp();
1199
1200 frame sender = caller;
1201
1202 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1203 // the vframeArray containing the unpacking information is allocated in the C heap.
1204 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1205 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1206
1207 // Compare the vframeArray to the collected vframes
1208 assert(array->structural_compare(thread, chunk), "just checking");
1209
1210 #ifndef PRODUCT
1211 if (PrintDeoptimizationDetails) {
1212 ttyLocker ttyl;
1213 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1214 }
1215 #endif // PRODUCT
1216
1217 return array;
1218 }
1219
1220 #if defined(COMPILER2) || INCLUDE_JVMCI
1221 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1222 // Reallocation of some scalar replaced objects failed. Record
1223 // that we need to pop all the interpreter frames for the
1224 // deoptimized compiled frame.
1225 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1226 thread->set_frames_to_pop_failed_realloc(array->frames());
1227 // Unlock all monitors here otherwise the interpreter will see a
1228 // mix of locked and unlocked monitors (because of failed
1229 // reallocations of synchronized objects) and be confused.
1230 for (int i = 0; i < array->frames(); i++) {
1231 MonitorChunk* monitors = array->element(i)->monitors();
1232 if (monitors != NULL) {
1233 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1234 BasicObjectLock* src = monitors->at(j);
1235 if (src->obj() != NULL) {
1236 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread);
1237 }
1238 }
1239 array->element(i)->free_monitors(thread);
1240 #ifdef ASSERT
1241 array->element(i)->set_removed_monitors();
1242 #endif
1243 }
1244 }
1245 }
1246 #endif
1247
1248 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1249 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1250 for (int i = 0; i < monitors->length(); i++) {
1251 MonitorInfo* mon_info = monitors->at(i);
1252 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1253 objects_to_revoke->append(Handle(mon_info->owner()));
1254 }
1255 }
1256 }
1257
1258
1259 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1260 if (!UseBiasedLocking) {
1261 return;
1262 }
1263
1264 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1265
1266 // Unfortunately we don't have a RegisterMap available in most of
1267 // the places we want to call this routine so we need to walk the
1268 // stack again to update the register map.
1269 if (map == NULL || !map->update_map()) {
1270 StackFrameStream sfs(thread, true);
1271 bool found = false;
1272 while (!found && !sfs.is_done()) {
1273 frame* cur = sfs.current();
1274 sfs.next();
1275 found = cur->id() == fr.id();
1276 }
1277 assert(found, "frame to be deoptimized not found on target thread's stack");
1278 map = sfs.register_map();
1279 }
1280
1281 vframe* vf = vframe::new_vframe(&fr, map, thread);
1282 compiledVFrame* cvf = compiledVFrame::cast(vf);
1283 // Revoke monitors' biases in all scopes
1284 while (!cvf->is_top()) {
1285 collect_monitors(cvf, objects_to_revoke);
1286 cvf = compiledVFrame::cast(cvf->sender());
1287 }
1288 collect_monitors(cvf, objects_to_revoke);
1289
1290 if (SafepointSynchronize::is_at_safepoint()) {
1291 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1292 } else {
1293 BiasedLocking::revoke(objects_to_revoke);
1294 }
1295 }
1296
1297
1298 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1299 if (!UseBiasedLocking) {
1300 return;
1301 }
1302
1303 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1304 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1305 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1306 if (jt->has_last_Java_frame()) {
1307 StackFrameStream sfs(jt, true);
1308 while (!sfs.is_done()) {
1309 frame* cur = sfs.current();
1310 if (cb->contains(cur->pc())) {
1311 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1312 compiledVFrame* cvf = compiledVFrame::cast(vf);
1313 // Revoke monitors' biases in all scopes
1314 while (!cvf->is_top()) {
1315 collect_monitors(cvf, objects_to_revoke);
1316 cvf = compiledVFrame::cast(cvf->sender());
1317 }
1318 collect_monitors(cvf, objects_to_revoke);
1319 }
1320 sfs.next();
1321 }
1322 }
1323 }
1324 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1325 }
1326
1327
1328 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1329 assert(fr.can_be_deoptimized(), "checking frame type");
1330
1331 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1332
1333 if (LogCompilation && xtty != NULL) {
1334 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1335 assert(cm != NULL, "only compiled methods can deopt");
1336
1337 ttyLocker ttyl;
1338 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1339 cm->log_identity(xtty);
1340 xtty->end_head();
1341 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1342 xtty->begin_elem("jvms bci='%d'", sd->bci());
1343 xtty->method(sd->method());
1344 xtty->end_elem();
1345 if (sd->is_top()) break;
1346 }
1347 xtty->tail("deoptimized");
1348 }
1349
1350 // Patch the compiled method so that when execution returns to it we will
1351 // deopt the execution state and return to the interpreter.
1352 fr.deoptimize(thread);
1353 }
1354
1355 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1356 deoptimize(thread, fr, map, Reason_constraint);
1357 }
1358
1359 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) {
1360 // Deoptimize only if the frame comes from compile code.
1361 // Do not deoptimize the frame which is already patched
1362 // during the execution of the loops below.
1363 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1364 return;
1365 }
1366 ResourceMark rm;
1367 DeoptimizationMarker dm;
1368 if (UseBiasedLocking) {
1369 revoke_biases_of_monitors(thread, fr, map);
1370 }
1371 deoptimize_single_frame(thread, fr, reason);
1372
1373 }
1374
1375
1376 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1377 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1378 "can only deoptimize other thread at a safepoint");
1379 // Compute frame and register map based on thread and sp.
1380 RegisterMap reg_map(thread, UseBiasedLocking);
1381 frame fr = thread->last_frame();
1382 while (fr.id() != id) {
1383 fr = fr.sender(®_map);
1384 }
1385 deoptimize(thread, fr, ®_map, reason);
1386 }
1387
1388
1389 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1390 if (thread == Thread::current()) {
1391 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1392 } else {
1393 VM_DeoptimizeFrame deopt(thread, id, reason);
1394 VMThread::execute(&deopt);
1395 }
1396 }
1397
1398 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1399 deoptimize_frame(thread, id, Reason_constraint);
1400 }
1401
1402 // JVMTI PopFrame support
1403 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1404 {
1405 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1406 }
1407 JRT_END
1408
1409 MethodData*
1410 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1411 bool create_if_missing) {
1412 Thread* THREAD = thread;
1413 MethodData* mdo = m()->method_data();
1414 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1415 // Build an MDO. Ignore errors like OutOfMemory;
1416 // that simply means we won't have an MDO to update.
1417 Method::build_interpreter_method_data(m, THREAD);
1418 if (HAS_PENDING_EXCEPTION) {
1419 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1420 CLEAR_PENDING_EXCEPTION;
1421 }
1422 mdo = m()->method_data();
1423 }
1424 return mdo;
1425 }
1426
1427 #if defined(COMPILER2) || defined(SHARK) || INCLUDE_JVMCI
1428 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1429 // in case of an unresolved klass entry, load the class.
1430 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1431 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1432 return;
1433 }
1434
1435 if (!constant_pool->tag_at(index).is_symbol()) return;
1436
1437 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1438 Symbol* symbol = constant_pool->symbol_at(index);
1439
1440 // class name?
1441 if (symbol->byte_at(0) != '(') {
1442 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1443 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1444 return;
1445 }
1446
1447 // then it must be a signature!
1448 ResourceMark rm(THREAD);
1449 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1450 if (ss.is_object()) {
1451 Symbol* class_name = ss.as_symbol(CHECK);
1452 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1453 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1454 }
1455 }
1456 }
1457
1458
1459 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1460 EXCEPTION_MARK;
1461 load_class_by_index(constant_pool, index, THREAD);
1462 if (HAS_PENDING_EXCEPTION) {
1463 // Exception happened during classloading. We ignore the exception here, since it
1464 // is going to be rethrown since the current activation is going to be deoptimized and
1465 // the interpreter will re-execute the bytecode.
1466 CLEAR_PENDING_EXCEPTION;
1467 // Class loading called java code which may have caused a stack
1468 // overflow. If the exception was thrown right before the return
1469 // to the runtime the stack is no longer guarded. Reguard the
1470 // stack otherwise if we return to the uncommon trap blob and the
1471 // stack bang causes a stack overflow we crash.
1472 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1473 JavaThread* thread = (JavaThread*)THREAD;
1474 bool guard_pages_enabled = thread->stack_guards_enabled();
1475 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1476 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1477 }
1478 }
1479
1480 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1481 HandleMark hm;
1482
1483 // uncommon_trap() is called at the beginning of the uncommon trap
1484 // handler. Note this fact before we start generating temporary frames
1485 // that can confuse an asynchronous stack walker. This counter is
1486 // decremented at the end of unpack_frames().
1487 thread->inc_in_deopt_handler();
1488
1489 // We need to update the map if we have biased locking.
1490 #if INCLUDE_JVMCI
1491 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1492 RegisterMap reg_map(thread, true);
1493 #else
1494 RegisterMap reg_map(thread, UseBiasedLocking);
1495 #endif
1496 frame stub_frame = thread->last_frame();
1497 frame fr = stub_frame.sender(®_map);
1498 // Make sure the calling nmethod is not getting deoptimized and removed
1499 // before we are done with it.
1500 nmethodLocker nl(fr.pc());
1501
1502 // Log a message
1503 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1504 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1505
1506 {
1507 ResourceMark rm;
1508
1509 // Revoke biases of any monitors in the frame to ensure we can migrate them
1510 revoke_biases_of_monitors(thread, fr, ®_map);
1511
1512 DeoptReason reason = trap_request_reason(trap_request);
1513 DeoptAction action = trap_request_action(trap_request);
1514 #if INCLUDE_JVMCI
1515 int debug_id = trap_request_debug_id(trap_request);
1516 #endif
1517 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1518
1519 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1520 compiledVFrame* cvf = compiledVFrame::cast(vf);
1521
1522 CompiledMethod* nm = cvf->code();
1523
1524 ScopeDesc* trap_scope = cvf->scope();
1525
1526 if (TraceDeoptimization) {
1527 ttyLocker ttyl;
1528 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()
1529 #if INCLUDE_JVMCI
1530 , debug_id
1531 #endif
1532 );
1533 }
1534
1535 methodHandle trap_method = trap_scope->method();
1536 int trap_bci = trap_scope->bci();
1537 #if INCLUDE_JVMCI
1538 oop speculation = thread->pending_failed_speculation();
1539 if (nm->is_compiled_by_jvmci()) {
1540 if (speculation != NULL) {
1541 oop speculation_log = nm->as_nmethod()->speculation_log();
1542 if (speculation_log != NULL) {
1543 if (TraceDeoptimization || TraceUncollectedSpeculations) {
1544 if (HotSpotSpeculationLog::lastFailed(speculation_log) != NULL) {
1545 tty->print_cr("A speculation that was not collected by the compiler is being overwritten");
1546 }
1547 }
1548 if (TraceDeoptimization) {
1549 tty->print_cr("Saving speculation to speculation log");
1550 }
1551 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation);
1552 } else {
1553 if (TraceDeoptimization) {
1554 tty->print_cr("Speculation present but no speculation log");
1555 }
1556 }
1557 thread->set_pending_failed_speculation(NULL);
1558 } else {
1559 if (TraceDeoptimization) {
1560 tty->print_cr("No speculation");
1561 }
1562 }
1563 } else {
1564 assert(speculation == NULL, "There should not be a speculation for method compiled by non-JVMCI compilers");
1565 }
1566
1567 if (trap_bci == SynchronizationEntryBCI) {
1568 trap_bci = 0;
1569 thread->set_pending_monitorenter(true);
1570 }
1571
1572 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1573 thread->set_pending_transfer_to_interpreter(true);
1574 }
1575 #endif
1576
1577 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1578 // Record this event in the histogram.
1579 gather_statistics(reason, action, trap_bc);
1580
1581 // Ensure that we can record deopt. history:
1582 // Need MDO to record RTM code generation state.
1583 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1584
1585 methodHandle profiled_method;
1586 #if INCLUDE_JVMCI
1587 if (nm->is_compiled_by_jvmci()) {
1588 profiled_method = nm->method();
1589 } else {
1590 profiled_method = trap_method;
1591 }
1592 #else
1593 profiled_method = trap_method;
1594 #endif
1595
1596 MethodData* trap_mdo =
1597 get_method_data(thread, profiled_method, create_if_missing);
1598
1599 // Log a message
1600 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d",
1601 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1602 trap_method->name_and_sig_as_C_string(), trap_bci);
1603
1604 // Print a bunch of diagnostics, if requested.
1605 if (TraceDeoptimization || LogCompilation) {
1606 ResourceMark rm;
1607 ttyLocker ttyl;
1608 char buf[100];
1609 if (xtty != NULL) {
1610 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1611 os::current_thread_id(),
1612 format_trap_request(buf, sizeof(buf), trap_request));
1613 nm->log_identity(xtty);
1614 }
1615 Symbol* class_name = NULL;
1616 bool unresolved = false;
1617 if (unloaded_class_index >= 0) {
1618 constantPoolHandle constants (THREAD, trap_method->constants());
1619 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1620 class_name = constants->klass_name_at(unloaded_class_index);
1621 unresolved = true;
1622 if (xtty != NULL)
1623 xtty->print(" unresolved='1'");
1624 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1625 class_name = constants->symbol_at(unloaded_class_index);
1626 }
1627 if (xtty != NULL)
1628 xtty->name(class_name);
1629 }
1630 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
1631 // Dump the relevant MDO state.
1632 // This is the deopt count for the current reason, any previous
1633 // reasons or recompiles seen at this point.
1634 int dcnt = trap_mdo->trap_count(reason);
1635 if (dcnt != 0)
1636 xtty->print(" count='%d'", dcnt);
1637 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1638 int dos = (pdata == NULL)? 0: pdata->trap_state();
1639 if (dos != 0) {
1640 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1641 if (trap_state_is_recompiled(dos)) {
1642 int recnt2 = trap_mdo->overflow_recompile_count();
1643 if (recnt2 != 0)
1644 xtty->print(" recompiles2='%d'", recnt2);
1645 }
1646 }
1647 }
1648 if (xtty != NULL) {
1649 xtty->stamp();
1650 xtty->end_head();
1651 }
1652 if (TraceDeoptimization) { // make noise on the tty
1653 tty->print("Uncommon trap occurred in");
1654 nm->method()->print_short_name(tty);
1655 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
1656 #if INCLUDE_JVMCI
1657 if (nm->is_nmethod()) {
1658 oop installedCode = nm->as_nmethod()->jvmci_installed_code();
1659 if (installedCode != NULL) {
1660 oop installedCodeName = NULL;
1661 if (installedCode->is_a(InstalledCode::klass())) {
1662 installedCodeName = InstalledCode::name(installedCode);
1663 }
1664 if (installedCodeName != NULL) {
1665 tty->print(" (JVMCI: installedCodeName=%s) ", java_lang_String::as_utf8_string(installedCodeName));
1666 } else {
1667 tty->print(" (JVMCI: installed code has no name) ");
1668 }
1669 } else if (nm->is_compiled_by_jvmci()) {
1670 tty->print(" (JVMCI: no installed code) ");
1671 }
1672 }
1673 #endif
1674 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
1675 p2i(fr.pc()),
1676 os::current_thread_id(),
1677 trap_reason_name(reason),
1678 trap_action_name(action),
1679 unloaded_class_index
1680 #if INCLUDE_JVMCI
1681 , debug_id
1682 #endif
1683 );
1684 if (class_name != NULL) {
1685 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1686 class_name->print_symbol_on(tty);
1687 }
1688 tty->cr();
1689 }
1690 if (xtty != NULL) {
1691 // Log the precise location of the trap.
1692 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1693 xtty->begin_elem("jvms bci='%d'", sd->bci());
1694 xtty->method(sd->method());
1695 xtty->end_elem();
1696 if (sd->is_top()) break;
1697 }
1698 xtty->tail("uncommon_trap");
1699 }
1700 }
1701 // (End diagnostic printout.)
1702
1703 // Load class if necessary
1704 if (unloaded_class_index >= 0) {
1705 constantPoolHandle constants(THREAD, trap_method->constants());
1706 load_class_by_index(constants, unloaded_class_index);
1707 }
1708
1709 // Flush the nmethod if necessary and desirable.
1710 //
1711 // We need to avoid situations where we are re-flushing the nmethod
1712 // because of a hot deoptimization site. Repeated flushes at the same
1713 // point need to be detected by the compiler and avoided. If the compiler
1714 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1715 // module must take measures to avoid an infinite cycle of recompilation
1716 // and deoptimization. There are several such measures:
1717 //
1718 // 1. If a recompilation is ordered a second time at some site X
1719 // and for the same reason R, the action is adjusted to 'reinterpret',
1720 // to give the interpreter time to exercise the method more thoroughly.
1721 // If this happens, the method's overflow_recompile_count is incremented.
1722 //
1723 // 2. If the compiler fails to reduce the deoptimization rate, then
1724 // the method's overflow_recompile_count will begin to exceed the set
1725 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1726 // is adjusted to 'make_not_compilable', and the method is abandoned
1727 // to the interpreter. This is a performance hit for hot methods,
1728 // but is better than a disastrous infinite cycle of recompilations.
1729 // (Actually, only the method containing the site X is abandoned.)
1730 //
1731 // 3. In parallel with the previous measures, if the total number of
1732 // recompilations of a method exceeds the much larger set limit
1733 // PerMethodRecompilationCutoff, the method is abandoned.
1734 // This should only happen if the method is very large and has
1735 // many "lukewarm" deoptimizations. The code which enforces this
1736 // limit is elsewhere (class nmethod, class Method).
1737 //
1738 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1739 // to recompile at each bytecode independently of the per-BCI cutoff.
1740 //
1741 // The decision to update code is up to the compiler, and is encoded
1742 // in the Action_xxx code. If the compiler requests Action_none
1743 // no trap state is changed, no compiled code is changed, and the
1744 // computation suffers along in the interpreter.
1745 //
1746 // The other action codes specify various tactics for decompilation
1747 // and recompilation. Action_maybe_recompile is the loosest, and
1748 // allows the compiled code to stay around until enough traps are seen,
1749 // and until the compiler gets around to recompiling the trapping method.
1750 //
1751 // The other actions cause immediate removal of the present code.
1752
1753 // Traps caused by injected profile shouldn't pollute trap counts.
1754 bool injected_profile_trap = trap_method->has_injected_profile() &&
1755 (reason == Reason_intrinsic || reason == Reason_unreached);
1756
1757 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
1758 bool make_not_entrant = false;
1759 bool make_not_compilable = false;
1760 bool reprofile = false;
1761 switch (action) {
1762 case Action_none:
1763 // Keep the old code.
1764 update_trap_state = false;
1765 break;
1766 case Action_maybe_recompile:
1767 // Do not need to invalidate the present code, but we can
1768 // initiate another
1769 // Start compiler without (necessarily) invalidating the nmethod.
1770 // The system will tolerate the old code, but new code should be
1771 // generated when possible.
1772 break;
1773 case Action_reinterpret:
1774 // Go back into the interpreter for a while, and then consider
1775 // recompiling form scratch.
1776 make_not_entrant = true;
1777 // Reset invocation counter for outer most method.
1778 // This will allow the interpreter to exercise the bytecodes
1779 // for a while before recompiling.
1780 // By contrast, Action_make_not_entrant is immediate.
1781 //
1782 // Note that the compiler will track null_check, null_assert,
1783 // range_check, and class_check events and log them as if they
1784 // had been traps taken from compiled code. This will update
1785 // the MDO trap history so that the next compilation will
1786 // properly detect hot trap sites.
1787 reprofile = true;
1788 break;
1789 case Action_make_not_entrant:
1790 // Request immediate recompilation, and get rid of the old code.
1791 // Make them not entrant, so next time they are called they get
1792 // recompiled. Unloaded classes are loaded now so recompile before next
1793 // time they are called. Same for uninitialized. The interpreter will
1794 // link the missing class, if any.
1795 make_not_entrant = true;
1796 break;
1797 case Action_make_not_compilable:
1798 // Give up on compiling this method at all.
1799 make_not_entrant = true;
1800 make_not_compilable = true;
1801 break;
1802 default:
1803 ShouldNotReachHere();
1804 }
1805
1806 // Setting +ProfileTraps fixes the following, on all platforms:
1807 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1808 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1809 // recompile relies on a MethodData* to record heroic opt failures.
1810
1811 // Whether the interpreter is producing MDO data or not, we also need
1812 // to use the MDO to detect hot deoptimization points and control
1813 // aggressive optimization.
1814 bool inc_recompile_count = false;
1815 ProfileData* pdata = NULL;
1816 if (ProfileTraps && update_trap_state && trap_mdo != NULL) {
1817 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
1818 uint this_trap_count = 0;
1819 bool maybe_prior_trap = false;
1820 bool maybe_prior_recompile = false;
1821 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
1822 #if INCLUDE_JVMCI
1823 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
1824 #endif
1825 nm->method(),
1826 //outputs:
1827 this_trap_count,
1828 maybe_prior_trap,
1829 maybe_prior_recompile);
1830 // Because the interpreter also counts null, div0, range, and class
1831 // checks, these traps from compiled code are double-counted.
1832 // This is harmless; it just means that the PerXTrapLimit values
1833 // are in effect a little smaller than they look.
1834
1835 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1836 if (per_bc_reason != Reason_none) {
1837 // Now take action based on the partially known per-BCI history.
1838 if (maybe_prior_trap
1839 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1840 // If there are too many traps at this BCI, force a recompile.
1841 // This will allow the compiler to see the limit overflow, and
1842 // take corrective action, if possible. The compiler generally
1843 // does not use the exact PerBytecodeTrapLimit value, but instead
1844 // changes its tactics if it sees any traps at all. This provides
1845 // a little hysteresis, delaying a recompile until a trap happens
1846 // several times.
1847 //
1848 // Actually, since there is only one bit of counter per BCI,
1849 // the possible per-BCI counts are {0,1,(per-method count)}.
1850 // This produces accurate results if in fact there is only
1851 // one hot trap site, but begins to get fuzzy if there are
1852 // many sites. For example, if there are ten sites each
1853 // trapping two or more times, they each get the blame for
1854 // all of their traps.
1855 make_not_entrant = true;
1856 }
1857
1858 // Detect repeated recompilation at the same BCI, and enforce a limit.
1859 if (make_not_entrant && maybe_prior_recompile) {
1860 // More than one recompile at this point.
1861 inc_recompile_count = maybe_prior_trap;
1862 }
1863 } else {
1864 // For reasons which are not recorded per-bytecode, we simply
1865 // force recompiles unconditionally.
1866 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1867 make_not_entrant = true;
1868 }
1869
1870 // Go back to the compiler if there are too many traps in this method.
1871 if (this_trap_count >= per_method_trap_limit(reason)) {
1872 // If there are too many traps in this method, force a recompile.
1873 // This will allow the compiler to see the limit overflow, and
1874 // take corrective action, if possible.
1875 // (This condition is an unlikely backstop only, because the
1876 // PerBytecodeTrapLimit is more likely to take effect first,
1877 // if it is applicable.)
1878 make_not_entrant = true;
1879 }
1880
1881 // Here's more hysteresis: If there has been a recompile at
1882 // this trap point already, run the method in the interpreter
1883 // for a while to exercise it more thoroughly.
1884 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1885 reprofile = true;
1886 }
1887 }
1888
1889 // Take requested actions on the method:
1890
1891 // Recompile
1892 if (make_not_entrant) {
1893 if (!nm->make_not_entrant()) {
1894 return; // the call did not change nmethod's state
1895 }
1896
1897 if (pdata != NULL) {
1898 // Record the recompilation event, if any.
1899 int tstate0 = pdata->trap_state();
1900 int tstate1 = trap_state_set_recompiled(tstate0, true);
1901 if (tstate1 != tstate0)
1902 pdata->set_trap_state(tstate1);
1903 }
1904
1905 #if INCLUDE_RTM_OPT
1906 // Restart collecting RTM locking abort statistic if the method
1907 // is recompiled for a reason other than RTM state change.
1908 // Assume that in new recompiled code the statistic could be different,
1909 // for example, due to different inlining.
1910 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
1911 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
1912 trap_mdo->atomic_set_rtm_state(ProfileRTM);
1913 }
1914 #endif
1915 // For code aging we count traps separately here, using make_not_entrant()
1916 // as a guard against simultaneous deopts in multiple threads.
1917 if (reason == Reason_tenured && trap_mdo != NULL) {
1918 trap_mdo->inc_tenure_traps();
1919 }
1920 }
1921
1922 if (inc_recompile_count) {
1923 trap_mdo->inc_overflow_recompile_count();
1924 if ((uint)trap_mdo->overflow_recompile_count() >
1925 (uint)PerBytecodeRecompilationCutoff) {
1926 // Give up on the method containing the bad BCI.
1927 if (trap_method() == nm->method()) {
1928 make_not_compilable = true;
1929 } else {
1930 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
1931 // But give grace to the enclosing nm->method().
1932 }
1933 }
1934 }
1935
1936 // Reprofile
1937 if (reprofile) {
1938 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1939 }
1940
1941 // Give up compiling
1942 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1943 assert(make_not_entrant, "consistent");
1944 nm->method()->set_not_compilable(CompLevel_full_optimization);
1945 }
1946
1947 } // Free marked resources
1948
1949 }
1950 JRT_END
1951
1952 ProfileData*
1953 Deoptimization::query_update_method_data(MethodData* trap_mdo,
1954 int trap_bci,
1955 Deoptimization::DeoptReason reason,
1956 bool update_total_trap_count,
1957 #if INCLUDE_JVMCI
1958 bool is_osr,
1959 #endif
1960 Method* compiled_method,
1961 //outputs:
1962 uint& ret_this_trap_count,
1963 bool& ret_maybe_prior_trap,
1964 bool& ret_maybe_prior_recompile) {
1965 bool maybe_prior_trap = false;
1966 bool maybe_prior_recompile = false;
1967 uint this_trap_count = 0;
1968 if (update_total_trap_count) {
1969 uint idx = reason;
1970 #if INCLUDE_JVMCI
1971 if (is_osr) {
1972 idx += Reason_LIMIT;
1973 }
1974 #endif
1975 uint prior_trap_count = trap_mdo->trap_count(idx);
1976 this_trap_count = trap_mdo->inc_trap_count(idx);
1977
1978 // If the runtime cannot find a place to store trap history,
1979 // it is estimated based on the general condition of the method.
1980 // If the method has ever been recompiled, or has ever incurred
1981 // a trap with the present reason , then this BCI is assumed
1982 // (pessimistically) to be the culprit.
1983 maybe_prior_trap = (prior_trap_count != 0);
1984 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1985 }
1986 ProfileData* pdata = NULL;
1987
1988
1989 // For reasons which are recorded per bytecode, we check per-BCI data.
1990 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1991 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
1992 if (per_bc_reason != Reason_none) {
1993 // Find the profile data for this BCI. If there isn't one,
1994 // try to allocate one from the MDO's set of spares.
1995 // This will let us detect a repeated trap at this point.
1996 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
1997
1998 if (pdata != NULL) {
1999 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2000 if (LogCompilation && xtty != NULL) {
2001 ttyLocker ttyl;
2002 // no more room for speculative traps in this MDO
2003 xtty->elem("speculative_traps_oom");
2004 }
2005 }
2006 // Query the trap state of this profile datum.
2007 int tstate0 = pdata->trap_state();
2008 if (!trap_state_has_reason(tstate0, per_bc_reason))
2009 maybe_prior_trap = false;
2010 if (!trap_state_is_recompiled(tstate0))
2011 maybe_prior_recompile = false;
2012
2013 // Update the trap state of this profile datum.
2014 int tstate1 = tstate0;
2015 // Record the reason.
2016 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2017 // Store the updated state on the MDO, for next time.
2018 if (tstate1 != tstate0)
2019 pdata->set_trap_state(tstate1);
2020 } else {
2021 if (LogCompilation && xtty != NULL) {
2022 ttyLocker ttyl;
2023 // Missing MDP? Leave a small complaint in the log.
2024 xtty->elem("missing_mdp bci='%d'", trap_bci);
2025 }
2026 }
2027 }
2028
2029 // Return results:
2030 ret_this_trap_count = this_trap_count;
2031 ret_maybe_prior_trap = maybe_prior_trap;
2032 ret_maybe_prior_recompile = maybe_prior_recompile;
2033 return pdata;
2034 }
2035
2036 void
2037 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2038 ResourceMark rm;
2039 // Ignored outputs:
2040 uint ignore_this_trap_count;
2041 bool ignore_maybe_prior_trap;
2042 bool ignore_maybe_prior_recompile;
2043 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2044 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2045 bool update_total_counts = JVMCI_ONLY(false) NOT_JVMCI(true);
2046 query_update_method_data(trap_mdo, trap_bci,
2047 (DeoptReason)reason,
2048 update_total_counts,
2049 #if INCLUDE_JVMCI
2050 false,
2051 #endif
2052 NULL,
2053 ignore_this_trap_count,
2054 ignore_maybe_prior_trap,
2055 ignore_maybe_prior_recompile);
2056 }
2057
2058 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2059 if (TraceDeoptimization) {
2060 tty->print("Uncommon trap ");
2061 }
2062 // Still in Java no safepoints
2063 {
2064 // This enters VM and may safepoint
2065 uncommon_trap_inner(thread, trap_request);
2066 }
2067 return fetch_unroll_info_helper(thread, exec_mode);
2068 }
2069
2070 // Local derived constants.
2071 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2072 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
2073 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2074
2075 //---------------------------trap_state_reason---------------------------------
2076 Deoptimization::DeoptReason
2077 Deoptimization::trap_state_reason(int trap_state) {
2078 // This assert provides the link between the width of DataLayout::trap_bits
2079 // and the encoding of "recorded" reasons. It ensures there are enough
2080 // bits to store all needed reasons in the per-BCI MDO profile.
2081 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2082 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2083 trap_state -= recompile_bit;
2084 if (trap_state == DS_REASON_MASK) {
2085 return Reason_many;
2086 } else {
2087 assert((int)Reason_none == 0, "state=0 => Reason_none");
2088 return (DeoptReason)trap_state;
2089 }
2090 }
2091 //-------------------------trap_state_has_reason-------------------------------
2092 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2093 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2094 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2095 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2096 trap_state -= recompile_bit;
2097 if (trap_state == DS_REASON_MASK) {
2098 return -1; // true, unspecifically (bottom of state lattice)
2099 } else if (trap_state == reason) {
2100 return 1; // true, definitely
2101 } else if (trap_state == 0) {
2102 return 0; // false, definitely (top of state lattice)
2103 } else {
2104 return 0; // false, definitely
2105 }
2106 }
2107 //-------------------------trap_state_add_reason-------------------------------
2108 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2109 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2110 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2111 trap_state -= recompile_bit;
2112 if (trap_state == DS_REASON_MASK) {
2113 return trap_state + recompile_bit; // already at state lattice bottom
2114 } else if (trap_state == reason) {
2115 return trap_state + recompile_bit; // the condition is already true
2116 } else if (trap_state == 0) {
2117 return reason + recompile_bit; // no condition has yet been true
2118 } else {
2119 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2120 }
2121 }
2122 //-----------------------trap_state_is_recompiled------------------------------
2123 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2124 return (trap_state & DS_RECOMPILE_BIT) != 0;
2125 }
2126 //-----------------------trap_state_set_recompiled-----------------------------
2127 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2128 if (z) return trap_state | DS_RECOMPILE_BIT;
2129 else return trap_state & ~DS_RECOMPILE_BIT;
2130 }
2131 //---------------------------format_trap_state---------------------------------
2132 // This is used for debugging and diagnostics, including LogFile output.
2133 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2134 int trap_state) {
2135 assert(buflen > 0, "sanity");
2136 DeoptReason reason = trap_state_reason(trap_state);
2137 bool recomp_flag = trap_state_is_recompiled(trap_state);
2138 // Re-encode the state from its decoded components.
2139 int decoded_state = 0;
2140 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2141 decoded_state = trap_state_add_reason(decoded_state, reason);
2142 if (recomp_flag)
2143 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2144 // If the state re-encodes properly, format it symbolically.
2145 // Because this routine is used for debugging and diagnostics,
2146 // be robust even if the state is a strange value.
2147 size_t len;
2148 if (decoded_state != trap_state) {
2149 // Random buggy state that doesn't decode??
2150 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2151 } else {
2152 len = jio_snprintf(buf, buflen, "%s%s",
2153 trap_reason_name(reason),
2154 recomp_flag ? " recompiled" : "");
2155 }
2156 return buf;
2157 }
2158
2159
2160 //--------------------------------statics--------------------------------------
2161 const char* Deoptimization::_trap_reason_name[] = {
2162 // Note: Keep this in sync. with enum DeoptReason.
2163 "none",
2164 "null_check",
2165 "null_assert" JVMCI_ONLY("_or_unreached0"),
2166 "range_check",
2167 "class_check",
2168 "array_check",
2169 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2170 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2171 "unloaded",
2172 "uninitialized",
2173 "unreached",
2174 "unhandled",
2175 "constraint",
2176 "div0_check",
2177 "age",
2178 "predicate",
2179 "loop_limit_check",
2180 "speculate_class_check",
2181 "speculate_null_check",
2182 "rtm_state_change",
2183 "unstable_if",
2184 "unstable_fused_if",
2185 #if INCLUDE_JVMCI
2186 "aliasing",
2187 "transfer_to_interpreter",
2188 "not_compiled_exception_handler",
2189 "unresolved",
2190 "jsr_mismatch",
2191 #endif
2192 "tenured"
2193 };
2194 const char* Deoptimization::_trap_action_name[] = {
2195 // Note: Keep this in sync. with enum DeoptAction.
2196 "none",
2197 "maybe_recompile",
2198 "reinterpret",
2199 "make_not_entrant",
2200 "make_not_compilable"
2201 };
2202
2203 const char* Deoptimization::trap_reason_name(int reason) {
2204 // Check that every reason has a name
2205 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2206
2207 if (reason == Reason_many) return "many";
2208 if ((uint)reason < Reason_LIMIT)
2209 return _trap_reason_name[reason];
2210 static char buf[20];
2211 sprintf(buf, "reason%d", reason);
2212 return buf;
2213 }
2214 const char* Deoptimization::trap_action_name(int action) {
2215 // Check that every action has a name
2216 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2217
2218 if ((uint)action < Action_LIMIT)
2219 return _trap_action_name[action];
2220 static char buf[20];
2221 sprintf(buf, "action%d", action);
2222 return buf;
2223 }
2224
2225 // This is used for debugging and diagnostics, including LogFile output.
2226 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2227 int trap_request) {
2228 jint unloaded_class_index = trap_request_index(trap_request);
2229 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2230 const char* action = trap_action_name(trap_request_action(trap_request));
2231 #if INCLUDE_JVMCI
2232 int debug_id = trap_request_debug_id(trap_request);
2233 #endif
2234 size_t len;
2235 if (unloaded_class_index < 0) {
2236 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2237 reason, action
2238 #if INCLUDE_JVMCI
2239 ,debug_id
2240 #endif
2241 );
2242 } else {
2243 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2244 reason, action, unloaded_class_index
2245 #if INCLUDE_JVMCI
2246 ,debug_id
2247 #endif
2248 );
2249 }
2250 return buf;
2251 }
2252
2253 juint Deoptimization::_deoptimization_hist
2254 [Deoptimization::Reason_LIMIT]
2255 [1 + Deoptimization::Action_LIMIT]
2256 [Deoptimization::BC_CASE_LIMIT]
2257 = {0};
2258
2259 enum {
2260 LSB_BITS = 8,
2261 LSB_MASK = right_n_bits(LSB_BITS)
2262 };
2263
2264 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2265 Bytecodes::Code bc) {
2266 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2267 assert(action >= 0 && action < Action_LIMIT, "oob");
2268 _deoptimization_hist[Reason_none][0][0] += 1; // total
2269 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2270 juint* cases = _deoptimization_hist[reason][1+action];
2271 juint* bc_counter_addr = NULL;
2272 juint bc_counter = 0;
2273 // Look for an unused counter, or an exact match to this BC.
2274 if (bc != Bytecodes::_illegal) {
2275 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2276 juint* counter_addr = &cases[bc_case];
2277 juint counter = *counter_addr;
2278 if ((counter == 0 && bc_counter_addr == NULL)
2279 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2280 // this counter is either free or is already devoted to this BC
2281 bc_counter_addr = counter_addr;
2282 bc_counter = counter | bc;
2283 }
2284 }
2285 }
2286 if (bc_counter_addr == NULL) {
2287 // Overflow, or no given bytecode.
2288 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2289 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2290 }
2291 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2292 }
2293
2294 jint Deoptimization::total_deoptimization_count() {
2295 return _deoptimization_hist[Reason_none][0][0];
2296 }
2297
2298 jint Deoptimization::deoptimization_count(DeoptReason reason) {
2299 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2300 return _deoptimization_hist[reason][0][0];
2301 }
2302
2303 void Deoptimization::print_statistics() {
2304 juint total = total_deoptimization_count();
2305 juint account = total;
2306 if (total != 0) {
2307 ttyLocker ttyl;
2308 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2309 tty->print_cr("Deoptimization traps recorded:");
2310 #define PRINT_STAT_LINE(name, r) \
2311 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2312 PRINT_STAT_LINE("total", total);
2313 // For each non-zero entry in the histogram, print the reason,
2314 // the action, and (if specifically known) the type of bytecode.
2315 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2316 for (int action = 0; action < Action_LIMIT; action++) {
2317 juint* cases = _deoptimization_hist[reason][1+action];
2318 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2319 juint counter = cases[bc_case];
2320 if (counter != 0) {
2321 char name[1*K];
2322 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2323 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2324 bc = Bytecodes::_illegal;
2325 sprintf(name, "%s/%s/%s",
2326 trap_reason_name(reason),
2327 trap_action_name(action),
2328 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2329 juint r = counter >> LSB_BITS;
2330 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2331 account -= r;
2332 }
2333 }
2334 }
2335 }
2336 if (account != 0) {
2337 PRINT_STAT_LINE("unaccounted", account);
2338 }
2339 #undef PRINT_STAT_LINE
2340 if (xtty != NULL) xtty->tail("statistics");
2341 }
2342 }
2343 #else // COMPILER2 || SHARK || INCLUDE_JVMCI
2344
2345
2346 // Stubs for C1 only system.
2347 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2348 return false;
2349 }
2350
2351 const char* Deoptimization::trap_reason_name(int reason) {
2352 return "unknown";
2353 }
2354
2355 void Deoptimization::print_statistics() {
2356 // no output
2357 }
2358
2359 void
2360 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2361 // no udpate
2362 }
2363
2364 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2365 return 0;
2366 }
2367
2368 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2369 Bytecodes::Code bc) {
2370 // no update
2371 }
2372
2373 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2374 int trap_state) {
2375 jio_snprintf(buf, buflen, "#%d", trap_state);
2376 return buf;
2377 }
2378
2379 #endif // COMPILER2 || SHARK || INCLUDE_JVMCI