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