1 /*
2 * Copyright (c) 1997, 2017, 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/valueArrayKlass.hpp"
43 #include "oops/valueArrayOop.hpp"
44 #include "oops/valueKlass.hpp"
45 #include "oops/verifyOopClosure.hpp"
46 #include "prims/jvmtiThreadState.hpp"
47 #include "runtime/biasedLocking.hpp"
48 #include "runtime/compilationPolicy.hpp"
49 #include "runtime/deoptimization.hpp"
50 #include "runtime/interfaceSupport.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "runtime/signature.hpp"
53 #include "runtime/stubRoutines.hpp"
54 #include "runtime/thread.hpp"
55 #include "runtime/vframe.hpp"
56 #include "runtime/vframeArray.hpp"
57 #include "runtime/vframe_hp.hpp"
58 #include "utilities/events.hpp"
59 #include "utilities/xmlstream.hpp"
60
61 #if INCLUDE_JVMCI
62 #include "jvmci/jvmciRuntime.hpp"
63 #include "jvmci/jvmciJavaClasses.hpp"
64 #endif
65
66
67 bool DeoptimizationMarker::_is_active = false;
68
69 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
70 int caller_adjustment,
71 int caller_actual_parameters,
72 int number_of_frames,
73 intptr_t* frame_sizes,
74 address* frame_pcs,
75 BasicType return_type,
76 int exec_mode) {
77 _size_of_deoptimized_frame = size_of_deoptimized_frame;
78 _caller_adjustment = caller_adjustment;
79 _caller_actual_parameters = caller_actual_parameters;
80 _number_of_frames = number_of_frames;
81 _frame_sizes = frame_sizes;
82 _frame_pcs = frame_pcs;
83 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
84 _return_type = return_type;
85 _initial_info = 0;
86 // PD (x86 only)
87 _counter_temp = 0;
88 _unpack_kind = exec_mode;
89 _sender_sp_temp = 0;
90
91 _total_frame_sizes = size_of_frames();
92 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
93 }
94
95
96 Deoptimization::UnrollBlock::~UnrollBlock() {
97 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
98 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
99 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
100 }
101
102
103 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
104 assert(register_number < RegisterMap::reg_count, "checking register number");
105 return &_register_block[register_number * 2];
106 }
107
108
109
110 int Deoptimization::UnrollBlock::size_of_frames() const {
111 // Acount first for the adjustment of the initial frame
112 int result = _caller_adjustment;
113 for (int index = 0; index < number_of_frames(); index++) {
114 result += frame_sizes()[index];
115 }
116 return result;
117 }
118
119
120 void Deoptimization::UnrollBlock::print() {
121 ttyLocker ttyl;
122 tty->print_cr("UnrollBlock");
123 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
124 tty->print( " frame_sizes: ");
125 for (int index = 0; index < number_of_frames(); index++) {
126 tty->print(INTX_FORMAT " ", frame_sizes()[index]);
127 }
128 tty->cr();
129 }
130
131
132 // In order to make fetch_unroll_info work properly with escape
133 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
134 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
135 // of previously eliminated objects occurs in realloc_objects, which is
136 // called from the method fetch_unroll_info_helper below.
137 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode))
138 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
139 // but makes the entry a little slower. There is however a little dance we have to
140 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
141
142 // fetch_unroll_info() is called at the beginning of the deoptimization
143 // handler. Note this fact before we start generating temporary frames
144 // that can confuse an asynchronous stack walker. This counter is
145 // decremented at the end of unpack_frames().
146 if (TraceDeoptimization) {
147 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread));
148 }
149 thread->inc_in_deopt_handler();
150
151 return fetch_unroll_info_helper(thread, exec_mode);
152 JRT_END
153
154
155 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
156 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) {
157
158 // Note: there is a safepoint safety issue here. No matter whether we enter
159 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
160 // the vframeArray is created.
161 //
162
163 // Allocate our special deoptimization ResourceMark
164 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
165 assert(thread->deopt_mark() == NULL, "Pending deopt!");
166 thread->set_deopt_mark(dmark);
167
168 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
169 RegisterMap map(thread, true);
170 RegisterMap dummy_map(thread, false);
171 // Now get the deoptee with a valid map
172 frame deoptee = stub_frame.sender(&map);
173 // Set the deoptee nmethod
174 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!");
175 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
176 thread->set_deopt_compiled_method(cm);
177
178 if (VerifyStack) {
179 thread->validate_frame_layout();
180 }
181
182 // Create a growable array of VFrames where each VFrame represents an inlined
183 // Java frame. This storage is allocated with the usual system arena.
184 assert(deoptee.is_compiled_frame(), "Wrong frame type");
185 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
186 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
187 while (!vf->is_top()) {
188 assert(vf->is_compiled_frame(), "Wrong frame type");
189 chunk->push(compiledVFrame::cast(vf));
190 vf = vf->sender();
191 }
192 assert(vf->is_compiled_frame(), "Wrong frame type");
193 chunk->push(compiledVFrame::cast(vf));
194
195 bool realloc_failures = false;
196
197 #if defined(COMPILER2) || INCLUDE_JVMCI
198 // Reallocate the non-escaping objects and restore their fields. Then
199 // relock objects if synchronization on them was eliminated.
200 #ifndef INCLUDE_JVMCI
201 if (DoEscapeAnalysis || EliminateNestedLocks) {
202 if (EliminateAllocations) {
203 #endif // INCLUDE_JVMCI
204 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
205 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
206
207 // The flag return_oop() indicates call sites which return oop
208 // in compiled code. Such sites include java method calls,
209 // runtime calls (for example, used to allocate new objects/arrays
210 // on slow code path) and any other calls generated in compiled code.
211 // It is not guaranteed that we can get such information here only
212 // by analyzing bytecode in deoptimized frames. This is why this flag
213 // is set during method compilation (see Compile::Process_OopMap_Node()).
214 // If the previous frame was popped or if we are dispatching an exception,
215 // we don't have an oop result.
216 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt);
217 Handle return_value;
218 if (save_oop_result) {
219 // Reallocation may trigger GC. If deoptimization happened on return from
220 // call which returns oop we need to save it since it is not in oopmap.
221 oop result = deoptee.saved_oop_result(&map);
222 assert(result == NULL || result->is_oop(), "must be oop");
223 return_value = Handle(thread, result);
224 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
225 if (TraceDeoptimization) {
226 ttyLocker ttyl;
227 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
228 }
229 }
230 if (objects != NULL) {
231 bool skip_internal = (cm != NULL) && !cm->is_compiled_by_jvmci();
232 JRT_BLOCK
233 realloc_failures = realloc_objects(thread, &deoptee, objects, THREAD);
234 reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, THREAD);
235 JRT_END
236 #ifndef PRODUCT
237 if (TraceDeoptimization) {
238 ttyLocker ttyl;
239 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
240 print_objects(objects, realloc_failures);
241 }
242 #endif
243 }
244 if (save_oop_result) {
245 // Restore result.
246 deoptee.set_saved_oop_result(&map, return_value());
247 }
248 #ifndef INCLUDE_JVMCI
249 }
250 if (EliminateLocks) {
251 #endif // INCLUDE_JVMCI
252 #ifndef PRODUCT
253 bool first = true;
254 #endif
255 for (int i = 0; i < chunk->length(); i++) {
256 compiledVFrame* cvf = chunk->at(i);
257 assert (cvf->scope() != NULL,"expect only compiled java frames");
258 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
259 if (monitors->is_nonempty()) {
260 relock_objects(monitors, thread, realloc_failures);
261 #ifndef PRODUCT
262 if (PrintDeoptimizationDetails) {
263 ttyLocker ttyl;
264 for (int j = 0; j < monitors->length(); j++) {
265 MonitorInfo* mi = monitors->at(j);
266 if (mi->eliminated()) {
267 if (first) {
268 first = false;
269 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
270 }
271 if (mi->owner_is_scalar_replaced()) {
272 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
273 tty->print_cr(" failed reallocation for klass %s", k->external_name());
274 } else {
275 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
276 }
277 }
278 }
279 }
280 #endif // !PRODUCT
281 }
282 }
283 #ifndef INCLUDE_JVMCI
284 }
285 }
286 #endif // INCLUDE_JVMCI
287 #endif // COMPILER2 || INCLUDE_JVMCI
288
289 ScopeDesc* trap_scope = chunk->at(0)->scope();
290 Handle exceptionObject;
291 if (trap_scope->rethrow_exception()) {
292 if (PrintDeoptimizationDetails) {
293 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());
294 }
295 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
296 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw");
297 ScopeValue* topOfStack = expressions->top();
298 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
299 guarantee(exceptionObject() != NULL, "exception oop can not be null");
300 }
301
302 // Ensure that no safepoint is taken after pointers have been stored
303 // in fields of rematerialized objects. If a safepoint occurs from here on
304 // out the java state residing in the vframeArray will be missed.
305 NoSafepointVerifier no_safepoint;
306
307 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
308 #if defined(COMPILER2) || INCLUDE_JVMCI
309 if (realloc_failures) {
310 pop_frames_failed_reallocs(thread, array);
311 }
312 #endif
313
314 assert(thread->vframe_array_head() == NULL, "Pending deopt!");
315 thread->set_vframe_array_head(array);
316
317 // Now that the vframeArray has been created if we have any deferred local writes
318 // added by jvmti then we can free up that structure as the data is now in the
319 // vframeArray
320
321 if (thread->deferred_locals() != NULL) {
322 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
323 int i = 0;
324 do {
325 // Because of inlining we could have multiple vframes for a single frame
326 // and several of the vframes could have deferred writes. Find them all.
327 if (list->at(i)->id() == array->original().id()) {
328 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
329 list->remove_at(i);
330 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
331 delete dlv;
332 } else {
333 i++;
334 }
335 } while ( i < list->length() );
336 if (list->length() == 0) {
337 thread->set_deferred_locals(NULL);
338 // free the list and elements back to C heap.
339 delete list;
340 }
341
342 }
343
344 #ifndef SHARK
345 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
346 CodeBlob* cb = stub_frame.cb();
347 // Verify we have the right vframeArray
348 assert(cb->frame_size() >= 0, "Unexpected frame size");
349 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
350
351 // If the deopt call site is a MethodHandle invoke call site we have
352 // to adjust the unpack_sp.
353 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
354 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
355 unpack_sp = deoptee.unextended_sp();
356
357 #ifdef ASSERT
358 assert(cb->is_deoptimization_stub() ||
359 cb->is_uncommon_trap_stub() ||
360 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
361 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
362 "unexpected code blob: %s", cb->name());
363 #endif
364 #else
365 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
366 #endif // !SHARK
367
368 // This is a guarantee instead of an assert because if vframe doesn't match
369 // we will unpack the wrong deoptimized frame and wind up in strange places
370 // where it will be very difficult to figure out what went wrong. Better
371 // to die an early death here than some very obscure death later when the
372 // trail is cold.
373 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
374 // in that it will fail to detect a problem when there is one. This needs
375 // more work in tiger timeframe.
376 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
377
378 int number_of_frames = array->frames();
379
380 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
381 // virtual activation, which is the reverse of the elements in the vframes array.
382 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
383 // +1 because we always have an interpreter return address for the final slot.
384 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
385 int popframe_extra_args = 0;
386 // Create an interpreter return address for the stub to use as its return
387 // address so the skeletal frames are perfectly walkable
388 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
389
390 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
391 // activation be put back on the expression stack of the caller for reexecution
392 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
393 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
394 }
395
396 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
397 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
398 // than simply use array->sender.pc(). This requires us to walk the current set of frames
399 //
400 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
401 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
402
403 // It's possible that the number of parameters at the call site is
404 // different than number of arguments in the callee when method
405 // handles are used. If the caller is interpreted get the real
406 // value so that the proper amount of space can be added to it's
407 // frame.
408 bool caller_was_method_handle = false;
409 if (deopt_sender.is_interpreted_frame()) {
410 methodHandle method = deopt_sender.interpreter_frame_method();
411 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
412 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
413 // Method handle invokes may involve fairly arbitrary chains of
414 // calls so it's impossible to know how much actual space the
415 // caller has for locals.
416 caller_was_method_handle = true;
417 }
418 }
419
420 //
421 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
422 // frame_sizes/frame_pcs[1] next oldest frame (int)
423 // frame_sizes/frame_pcs[n] youngest frame (int)
424 //
425 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
426 // owns the space for the return address to it's caller). Confusing ain't it.
427 //
428 // The vframe array can address vframes with indices running from
429 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
430 // When we create the skeletal frames we need the oldest frame to be in the zero slot
431 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
432 // so things look a little strange in this loop.
433 //
434 int callee_parameters = 0;
435 int callee_locals = 0;
436 for (int index = 0; index < array->frames(); index++ ) {
437 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
438 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
439 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
440 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
441 callee_locals,
442 index == 0,
443 popframe_extra_args);
444 // This pc doesn't have to be perfect just good enough to identify the frame
445 // as interpreted so the skeleton frame will be walkable
446 // The correct pc will be set when the skeleton frame is completely filled out
447 // The final pc we store in the loop is wrong and will be overwritten below
448 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
449
450 callee_parameters = array->element(index)->method()->size_of_parameters();
451 callee_locals = array->element(index)->method()->max_locals();
452 popframe_extra_args = 0;
453 }
454
455 // Compute whether the root vframe returns a float or double value.
456 BasicType return_type;
457 {
458 methodHandle method(thread, array->element(0)->method());
459 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
460 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
461 }
462
463 // Compute information for handling adapters and adjusting the frame size of the caller.
464 int caller_adjustment = 0;
465
466 // Compute the amount the oldest interpreter frame will have to adjust
467 // its caller's stack by. If the caller is a compiled frame then
468 // we pretend that the callee has no parameters so that the
469 // extension counts for the full amount of locals and not just
470 // locals-parms. This is because without a c2i adapter the parm
471 // area as created by the compiled frame will not be usable by
472 // the interpreter. (Depending on the calling convention there
473 // may not even be enough space).
474
475 // QQQ I'd rather see this pushed down into last_frame_adjust
476 // and have it take the sender (aka caller).
477
478 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
479 caller_adjustment = last_frame_adjust(0, callee_locals);
480 } else if (callee_locals > callee_parameters) {
481 // The caller frame may need extending to accommodate
482 // non-parameter locals of the first unpacked interpreted frame.
483 // Compute that adjustment.
484 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
485 }
486
487 // If the sender is deoptimized the we must retrieve the address of the handler
488 // since the frame will "magically" show the original pc before the deopt
489 // and we'd undo the deopt.
490
491 frame_pcs[0] = deopt_sender.raw_pc();
492
493 #ifndef SHARK
494 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
495 #endif // SHARK
496
497 #ifdef INCLUDE_JVMCI
498 if (exceptionObject() != NULL) {
499 thread->set_exception_oop(exceptionObject());
500 exec_mode = Unpack_exception;
501 }
502 #endif
503
504 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
505 assert(thread->has_pending_exception(), "should have thrown OOME");
506 thread->set_exception_oop(thread->pending_exception());
507 thread->clear_pending_exception();
508 exec_mode = Unpack_exception;
509 }
510
511 #if INCLUDE_JVMCI
512 if (thread->frames_to_pop_failed_realloc() > 0) {
513 thread->set_pending_monitorenter(false);
514 }
515 #endif
516
517 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
518 caller_adjustment * BytesPerWord,
519 caller_was_method_handle ? 0 : callee_parameters,
520 number_of_frames,
521 frame_sizes,
522 frame_pcs,
523 return_type,
524 exec_mode);
525 // On some platforms, we need a way to pass some platform dependent
526 // information to the unpacking code so the skeletal frames come out
527 // correct (initial fp value, unextended sp, ...)
528 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
529
530 if (array->frames() > 1) {
531 if (VerifyStack && TraceDeoptimization) {
532 ttyLocker ttyl;
533 tty->print_cr("Deoptimizing method containing inlining");
534 }
535 }
536
537 array->set_unroll_block(info);
538 return info;
539 }
540
541 // Called to cleanup deoptimization data structures in normal case
542 // after unpacking to stack and when stack overflow error occurs
543 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
544 vframeArray *array) {
545
546 // Get array if coming from exception
547 if (array == NULL) {
548 array = thread->vframe_array_head();
549 }
550 thread->set_vframe_array_head(NULL);
551
552 // Free the previous UnrollBlock
553 vframeArray* old_array = thread->vframe_array_last();
554 thread->set_vframe_array_last(array);
555
556 if (old_array != NULL) {
557 UnrollBlock* old_info = old_array->unroll_block();
558 old_array->set_unroll_block(NULL);
559 delete old_info;
560 delete old_array;
561 }
562
563 // Deallocate any resource creating in this routine and any ResourceObjs allocated
564 // inside the vframeArray (StackValueCollections)
565
566 delete thread->deopt_mark();
567 thread->set_deopt_mark(NULL);
568 thread->set_deopt_compiled_method(NULL);
569
570
571 if (JvmtiExport::can_pop_frame()) {
572 #ifndef CC_INTERP
573 // Regardless of whether we entered this routine with the pending
574 // popframe condition bit set, we should always clear it now
575 thread->clear_popframe_condition();
576 #else
577 // C++ interpreter will clear has_pending_popframe when it enters
578 // with method_resume. For deopt_resume2 we clear it now.
579 if (thread->popframe_forcing_deopt_reexecution())
580 thread->clear_popframe_condition();
581 #endif /* CC_INTERP */
582 }
583
584 // unpack_frames() is called at the end of the deoptimization handler
585 // and (in C2) at the end of the uncommon trap handler. Note this fact
586 // so that an asynchronous stack walker can work again. This counter is
587 // incremented at the beginning of fetch_unroll_info() and (in C2) at
588 // the beginning of uncommon_trap().
589 thread->dec_in_deopt_handler();
590 }
591
592 // Moved from cpu directories because none of the cpus has callee save values.
593 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
594 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
595
596 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
597 // the days we had adapter frames. When we deoptimize a situation where a
598 // compiled caller calls a compiled caller will have registers it expects
599 // to survive the call to the callee. If we deoptimize the callee the only
600 // way we can restore these registers is to have the oldest interpreter
601 // frame that we create restore these values. That is what this routine
602 // will accomplish.
603
604 // At the moment we have modified c2 to not have any callee save registers
605 // so this problem does not exist and this routine is just a place holder.
606
607 assert(f->is_interpreted_frame(), "must be interpreted");
608 }
609
610 // Return BasicType of value being returned
611 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
612
613 // We are already active int he special DeoptResourceMark any ResourceObj's we
614 // allocate will be freed at the end of the routine.
615
616 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
617 // but makes the entry a little slower. There is however a little dance we have to
618 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
619 ResetNoHandleMark rnhm; // No-op in release/product versions
620 HandleMark hm;
621
622 frame stub_frame = thread->last_frame();
623
624 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
625 // must point to the vframeArray for the unpack frame.
626 vframeArray* array = thread->vframe_array_head();
627
628 #ifndef PRODUCT
629 if (TraceDeoptimization) {
630 ttyLocker ttyl;
631 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d",
632 p2i(thread), p2i(array), exec_mode);
633 }
634 #endif
635 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
636 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode);
637
638 UnrollBlock* info = array->unroll_block();
639
640 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
641 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
642
643 BasicType bt = info->return_type();
644
645 // If we have an exception pending, claim that the return type is an oop
646 // so the deopt_blob does not overwrite the exception_oop.
647
648 if (exec_mode == Unpack_exception)
649 bt = T_OBJECT;
650
651 // Cleanup thread deopt data
652 cleanup_deopt_info(thread, array);
653
654 #ifndef PRODUCT
655 if (VerifyStack) {
656 ResourceMark res_mark;
657
658 thread->validate_frame_layout();
659
660 // Verify that the just-unpacked frames match the interpreter's
661 // notions of expression stack and locals
662 vframeArray* cur_array = thread->vframe_array_last();
663 RegisterMap rm(thread, false);
664 rm.set_include_argument_oops(false);
665 bool is_top_frame = true;
666 int callee_size_of_parameters = 0;
667 int callee_max_locals = 0;
668 for (int i = 0; i < cur_array->frames(); i++) {
669 vframeArrayElement* el = cur_array->element(i);
670 frame* iframe = el->iframe();
671 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
672
673 // Get the oop map for this bci
674 InterpreterOopMap mask;
675 int cur_invoke_parameter_size = 0;
676 bool try_next_mask = false;
677 int next_mask_expression_stack_size = -1;
678 int top_frame_expression_stack_adjustment = 0;
679 methodHandle mh(thread, iframe->interpreter_frame_method());
680 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
681 BytecodeStream str(mh);
682 str.set_start(iframe->interpreter_frame_bci());
683 int max_bci = mh->code_size();
684 // Get to the next bytecode if possible
685 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
686 // Check to see if we can grab the number of outgoing arguments
687 // at an uncommon trap for an invoke (where the compiler
688 // generates debug info before the invoke has executed)
689 Bytecodes::Code cur_code = str.next();
690 if (cur_code == Bytecodes::_invokevirtual ||
691 cur_code == Bytecodes::_invokespecial ||
692 cur_code == Bytecodes::_invokestatic ||
693 cur_code == Bytecodes::_invokeinterface ||
694 cur_code == Bytecodes::_invokedynamic) {
695 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
696 Symbol* signature = invoke.signature();
697 ArgumentSizeComputer asc(signature);
698 cur_invoke_parameter_size = asc.size();
699 if (invoke.has_receiver()) {
700 // Add in receiver
701 ++cur_invoke_parameter_size;
702 }
703 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
704 callee_size_of_parameters++;
705 }
706 }
707 if (str.bci() < max_bci) {
708 Bytecodes::Code bc = str.next();
709 if (bc >= 0) {
710 // The interpreter oop map generator reports results before
711 // the current bytecode has executed except in the case of
712 // calls. It seems to be hard to tell whether the compiler
713 // has emitted debug information matching the "state before"
714 // a given bytecode or the state after, so we try both
715 switch (cur_code) {
716 case Bytecodes::_invokevirtual:
717 case Bytecodes::_invokespecial:
718 case Bytecodes::_invokestatic:
719 case Bytecodes::_invokeinterface:
720 case Bytecodes::_invokedynamic:
721 case Bytecodes::_athrow:
722 break;
723 default: {
724 InterpreterOopMap next_mask;
725 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
726 next_mask_expression_stack_size = next_mask.expression_stack_size();
727 // Need to subtract off the size of the result type of
728 // the bytecode because this is not described in the
729 // debug info but returned to the interpreter in the TOS
730 // caching register
731 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
732 if (bytecode_result_type != T_ILLEGAL) {
733 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
734 }
735 assert(top_frame_expression_stack_adjustment >= 0, "");
736 try_next_mask = true;
737 break;
738 }
739 }
740 }
741 }
742
743 // Verify stack depth and oops in frame
744 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
745 if (!(
746 /* SPARC */
747 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
748 /* x86 */
749 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
750 (try_next_mask &&
751 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
752 top_frame_expression_stack_adjustment))) ||
753 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
754 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
755 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
756 )) {
757 ttyLocker ttyl;
758
759 // Print out some information that will help us debug the problem
760 tty->print_cr("Wrong number of expression stack elements during deoptimization");
761 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
762 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
763 iframe->interpreter_frame_expression_stack_size());
764 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
765 tty->print_cr(" try_next_mask = %d", try_next_mask);
766 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
767 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
768 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
769 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
770 tty->print_cr(" exec_mode = %d", exec_mode);
771 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
772 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
773 tty->print_cr(" Interpreted frames:");
774 for (int k = 0; k < cur_array->frames(); k++) {
775 vframeArrayElement* el = cur_array->element(k);
776 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
777 }
778 cur_array->print_on_2(tty);
779 guarantee(false, "wrong number of expression stack elements during deopt");
780 }
781 VerifyOopClosure verify;
782 iframe->oops_interpreted_do(&verify, &rm, false);
783 callee_size_of_parameters = mh->size_of_parameters();
784 callee_max_locals = mh->max_locals();
785 is_top_frame = false;
786 }
787 }
788 #endif /* !PRODUCT */
789
790
791 return bt;
792 JRT_END
793
794
795 int Deoptimization::deoptimize_dependents() {
796 Threads::deoptimized_wrt_marked_nmethods();
797 return 0;
798 }
799
800 Deoptimization::DeoptAction Deoptimization::_unloaded_action
801 = Deoptimization::Action_reinterpret;
802
803 #if defined(COMPILER2) || INCLUDE_JVMCI
804 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
805 Handle pending_exception(THREAD, thread->pending_exception());
806 const char* exception_file = thread->exception_file();
807 int exception_line = thread->exception_line();
808 thread->clear_pending_exception();
809
810 bool failures = false;
811
812 for (int i = 0; i < objects->length(); i++) {
813 assert(objects->at(i)->is_object(), "invalid debug information");
814 ObjectValue* sv = (ObjectValue*) objects->at(i);
815
816 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
817 oop obj = NULL;
818
819 if (k->is_instance_klass()) {
820 InstanceKlass* ik = InstanceKlass::cast(k);
821 obj = ik->allocate_instance(THREAD);
822 } else if (k->is_valueArray_klass()) {
823 ValueArrayKlass* ak = ValueArrayKlass::cast(k);
824 // Value type array must be zeroed because not all memory is reassigned
825 obj = ak->allocate(sv->field_size(), true, THREAD);
826 } else if (k->is_typeArray_klass()) {
827 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
828 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
829 int len = sv->field_size() / type2size[ak->element_type()];
830 obj = ak->allocate(len, THREAD);
831 } else if (k->is_objArray_klass()) {
832 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
833 obj = ak->allocate(sv->field_size(), THREAD);
834 }
835
836 if (obj == NULL) {
837 failures = true;
838 }
839
840 assert(sv->value().is_null(), "redundant reallocation");
841 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
842 CLEAR_PENDING_EXCEPTION;
843 sv->set_value(obj);
844 }
845
846 if (failures) {
847 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
848 } else if (pending_exception.not_null()) {
849 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
850 }
851
852 return failures;
853 }
854
855 // restore elements of an eliminated type array
856 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
857 int index = 0;
858 intptr_t val;
859
860 for (int i = 0; i < sv->field_size(); i++) {
861 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
862 switch(type) {
863 case T_LONG: case T_DOUBLE: {
864 assert(value->type() == T_INT, "Agreement.");
865 StackValue* low =
866 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
867 #ifdef _LP64
868 jlong res = (jlong)low->get_int();
869 #else
870 #ifdef SPARC
871 // For SPARC we have to swap high and low words.
872 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
873 #else
874 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
875 #endif //SPARC
876 #endif
877 obj->long_at_put(index, res);
878 break;
879 }
880
881 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
882 case T_INT: case T_FLOAT: { // 4 bytes.
883 assert(value->type() == T_INT, "Agreement.");
884 bool big_value = false;
885 if (i + 1 < sv->field_size() && type == T_INT) {
886 if (sv->field_at(i)->is_location()) {
887 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
888 if (type == Location::dbl || type == Location::lng) {
889 big_value = true;
890 }
891 } else if (sv->field_at(i)->is_constant_int()) {
892 ScopeValue* next_scope_field = sv->field_at(i + 1);
893 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
894 big_value = true;
895 }
896 }
897 }
898
899 if (big_value) {
900 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
901 #ifdef _LP64
902 jlong res = (jlong)low->get_int();
903 #else
904 #ifdef SPARC
905 // For SPARC we have to swap high and low words.
906 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
907 #else
908 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
909 #endif //SPARC
910 #endif
911 obj->int_at_put(index, (jint)*((jint*)&res));
912 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
913 } else {
914 val = value->get_int();
915 obj->int_at_put(index, (jint)*((jint*)&val));
916 }
917 break;
918 }
919
920 case T_SHORT:
921 assert(value->type() == T_INT, "Agreement.");
922 val = value->get_int();
923 obj->short_at_put(index, (jshort)*((jint*)&val));
924 break;
925
926 case T_CHAR:
927 assert(value->type() == T_INT, "Agreement.");
928 val = value->get_int();
929 obj->char_at_put(index, (jchar)*((jint*)&val));
930 break;
931
932 case T_BYTE:
933 assert(value->type() == T_INT, "Agreement.");
934 val = value->get_int();
935 obj->byte_at_put(index, (jbyte)*((jint*)&val));
936 break;
937
938 case T_BOOLEAN:
939 assert(value->type() == T_INT, "Agreement.");
940 val = value->get_int();
941 obj->bool_at_put(index, (jboolean)*((jint*)&val));
942 break;
943
944 default:
945 ShouldNotReachHere();
946 }
947 index++;
948 }
949 }
950
951
952 // restore fields of an eliminated object array
953 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
954 for (int i = 0; i < sv->field_size(); i++) {
955 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
956 assert(value->type() == T_OBJECT, "object element expected");
957 obj->obj_at_put(i, value->get_obj()());
958 }
959 }
960
961 class ReassignedField {
962 public:
963 int _offset;
964 BasicType _type;
965 InstanceKlass* _klass;
966 public:
967 ReassignedField() {
968 _offset = 0;
969 _type = T_ILLEGAL;
970 _klass = NULL;
971 }
972 };
973
974 int compare(ReassignedField* left, ReassignedField* right) {
975 return left->_offset - right->_offset;
976 }
977
978 // Restore fields of an eliminated instance object using the same field order
979 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
980 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal, int base_offset, TRAPS) {
981 if (klass->superklass() != NULL) {
982 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal, 0, CHECK_0);
983 }
984
985 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
986 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
987 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
988 ReassignedField field;
989 field._offset = fs.offset();
990 field._type = FieldType::basic_type(fs.signature());
991 if (field._type == T_VALUETYPE) {
992 // Resolve klass of flattened value type field
993 SignatureStream ss(fs.signature(), false);
994 Klass* vk = ss.as_klass(Handle(THREAD, klass->class_loader()), Handle(THREAD, klass->protection_domain()), SignatureStream::NCDFError, THREAD);
995 guarantee(!HAS_PENDING_EXCEPTION, "Should not have any exceptions pending");
996 assert(vk->is_value(), "must be a ValueKlass");
997 field._klass = InstanceKlass::cast(vk);
998 }
999 fields->append(field);
1000 }
1001 }
1002 fields->sort(compare);
1003 for (int i = 0; i < fields->length(); i++) {
1004 intptr_t val;
1005 ScopeValue* scope_field = sv->field_at(svIndex);
1006 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1007 int offset = base_offset + fields->at(i)._offset;
1008 BasicType type = fields->at(i)._type;
1009 switch (type) {
1010 case T_OBJECT: case T_ARRAY:
1011 assert(value->type() == T_OBJECT, "Agreement.");
1012 obj->obj_field_put(offset, value->get_obj()());
1013 break;
1014
1015 case T_VALUETYPE: {
1016 // Recursively re-assign flattened value type fields
1017 InstanceKlass* vk = fields->at(i)._klass;
1018 assert(vk != NULL, "must be resolved");
1019 offset -= ValueKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header
1020 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0);
1021 continue; // Continue because we don't need to increment svIndex
1022 }
1023
1024 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1025 case T_INT: case T_FLOAT: { // 4 bytes.
1026 assert(value->type() == T_INT, "Agreement.");
1027 bool big_value = false;
1028 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1029 if (scope_field->is_location()) {
1030 Location::Type type = ((LocationValue*) scope_field)->location().type();
1031 if (type == Location::dbl || type == Location::lng) {
1032 big_value = true;
1033 }
1034 }
1035 if (scope_field->is_constant_int()) {
1036 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1037 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1038 big_value = true;
1039 }
1040 }
1041 }
1042
1043 if (big_value) {
1044 i++;
1045 assert(i < fields->length(), "second T_INT field needed");
1046 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1047 } else {
1048 val = value->get_int();
1049 obj->int_field_put(offset, (jint)*((jint*)&val));
1050 break;
1051 }
1052 }
1053 /* no break */
1054
1055 case T_LONG: case T_DOUBLE: {
1056 assert(value->type() == T_INT, "Agreement.");
1057 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1058 #ifdef _LP64
1059 jlong res = (jlong)low->get_int();
1060 #else
1061 #ifdef SPARC
1062 // For SPARC we have to swap high and low words.
1063 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1064 #else
1065 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1066 #endif //SPARC
1067 #endif
1068 obj->long_field_put(offset, res);
1069 break;
1070 }
1071
1072 case T_SHORT:
1073 assert(value->type() == T_INT, "Agreement.");
1074 val = value->get_int();
1075 obj->short_field_put(offset, (jshort)*((jint*)&val));
1076 break;
1077
1078 case T_CHAR:
1079 assert(value->type() == T_INT, "Agreement.");
1080 val = value->get_int();
1081 obj->char_field_put(offset, (jchar)*((jint*)&val));
1082 break;
1083
1084 case T_BYTE:
1085 assert(value->type() == T_INT, "Agreement.");
1086 val = value->get_int();
1087 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1088 break;
1089
1090 case T_BOOLEAN:
1091 assert(value->type() == T_INT, "Agreement.");
1092 val = value->get_int();
1093 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1094 break;
1095
1096 default:
1097 ShouldNotReachHere();
1098 }
1099 svIndex++;
1100 }
1101 return svIndex;
1102 }
1103
1104 // restore fields of an eliminated value type array
1105 void Deoptimization::reassign_value_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, valueArrayOop obj, ValueArrayKlass* vak, TRAPS) {
1106 ValueKlass* vk = vak->element_klass();
1107 assert(vk->flatten_array(), "should only be used for flattened value type arrays");
1108 // Adjust offset to omit oop header
1109 int base_offset = arrayOopDesc::base_offset_in_bytes(T_VALUETYPE) - ValueKlass::cast(vk)->first_field_offset();
1110 // Initialize all elements of the flattened value type array
1111 for (int i = 0; i < sv->field_size(); i++) {
1112 ScopeValue* val = sv->field_at(i);
1113 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper()));
1114 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, false /* skip_internal */, offset, CHECK);
1115 }
1116 }
1117
1118 // restore fields of all eliminated objects and arrays
1119 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) {
1120 for (int i = 0; i < objects->length(); i++) {
1121 ObjectValue* sv = (ObjectValue*) objects->at(i);
1122 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1123 Handle obj = sv->value();
1124 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1125 if (PrintDeoptimizationDetails) {
1126 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1127 }
1128 if (obj.is_null()) {
1129 continue;
1130 }
1131
1132 if (k->is_instance_klass()) {
1133 InstanceKlass* ik = InstanceKlass::cast(k);
1134 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK);
1135 } else if (k->is_valueArray_klass()) {
1136 ValueArrayKlass* vak = ValueArrayKlass::cast(k);
1137 reassign_value_array_elements(fr, reg_map, sv, (valueArrayOop) obj(), vak, CHECK);
1138 } else if (k->is_typeArray_klass()) {
1139 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1140 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1141 } else if (k->is_objArray_klass()) {
1142 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1143 }
1144 }
1145 }
1146
1147
1148 // relock objects for which synchronization was eliminated
1149 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
1150 for (int i = 0; i < monitors->length(); i++) {
1151 MonitorInfo* mon_info = monitors->at(i);
1152 if (mon_info->eliminated()) {
1153 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1154 if (!mon_info->owner_is_scalar_replaced()) {
1155 Handle obj(thread, mon_info->owner());
1156 markOop mark = obj->mark();
1157 if (UseBiasedLocking && mark->has_bias_pattern()) {
1158 // New allocated objects may have the mark set to anonymously biased.
1159 // Also the deoptimized method may called methods with synchronization
1160 // where the thread-local object is bias locked to the current thread.
1161 assert(mark->is_biased_anonymously() ||
1162 mark->biased_locker() == thread, "should be locked to current thread");
1163 // Reset mark word to unbiased prototype.
1164 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
1165 obj->set_mark(unbiased_prototype);
1166 }
1167 BasicLock* lock = mon_info->lock();
1168 ObjectSynchronizer::slow_enter(obj, lock, thread);
1169 assert(mon_info->owner()->is_locked(), "object must be locked now");
1170 }
1171 }
1172 }
1173 }
1174
1175
1176 #ifndef PRODUCT
1177 // print information about reallocated objects
1178 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1179 fieldDescriptor fd;
1180
1181 for (int i = 0; i < objects->length(); i++) {
1182 ObjectValue* sv = (ObjectValue*) objects->at(i);
1183 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1184 Handle obj = sv->value();
1185
1186 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1187 k->print_value();
1188 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1189 if (obj.is_null()) {
1190 tty->print(" allocation failed");
1191 } else {
1192 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1193 }
1194 tty->cr();
1195
1196 if (Verbose && !obj.is_null()) {
1197 k->oop_print_on(obj(), tty);
1198 }
1199 }
1200 }
1201 #endif
1202 #endif // COMPILER2 || INCLUDE_JVMCI
1203
1204 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1205 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1206
1207 #ifndef PRODUCT
1208 if (PrintDeoptimizationDetails) {
1209 ttyLocker ttyl;
1210 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1211 fr.print_on(tty);
1212 tty->print_cr(" Virtual frames (innermost first):");
1213 for (int index = 0; index < chunk->length(); index++) {
1214 compiledVFrame* vf = chunk->at(index);
1215 tty->print(" %2d - ", index);
1216 vf->print_value();
1217 int bci = chunk->at(index)->raw_bci();
1218 const char* code_name;
1219 if (bci == SynchronizationEntryBCI) {
1220 code_name = "sync entry";
1221 } else {
1222 Bytecodes::Code code = vf->method()->code_at(bci);
1223 code_name = Bytecodes::name(code);
1224 }
1225 tty->print(" - %s", code_name);
1226 tty->print_cr(" @ bci %d ", bci);
1227 if (Verbose) {
1228 vf->print();
1229 tty->cr();
1230 }
1231 }
1232 }
1233 #endif
1234
1235 // Register map for next frame (used for stack crawl). We capture
1236 // the state of the deopt'ing frame's caller. Thus if we need to
1237 // stuff a C2I adapter we can properly fill in the callee-save
1238 // register locations.
1239 frame caller = fr.sender(reg_map);
1240 int frame_size = caller.sp() - fr.sp();
1241
1242 frame sender = caller;
1243
1244 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1245 // the vframeArray containing the unpacking information is allocated in the C heap.
1246 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1247 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1248
1249 // Compare the vframeArray to the collected vframes
1250 assert(array->structural_compare(thread, chunk), "just checking");
1251
1252 #ifndef PRODUCT
1253 if (PrintDeoptimizationDetails) {
1254 ttyLocker ttyl;
1255 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1256 }
1257 #endif // PRODUCT
1258
1259 return array;
1260 }
1261
1262 #if defined(COMPILER2) || INCLUDE_JVMCI
1263 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1264 // Reallocation of some scalar replaced objects failed. Record
1265 // that we need to pop all the interpreter frames for the
1266 // deoptimized compiled frame.
1267 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1268 thread->set_frames_to_pop_failed_realloc(array->frames());
1269 // Unlock all monitors here otherwise the interpreter will see a
1270 // mix of locked and unlocked monitors (because of failed
1271 // reallocations of synchronized objects) and be confused.
1272 for (int i = 0; i < array->frames(); i++) {
1273 MonitorChunk* monitors = array->element(i)->monitors();
1274 if (monitors != NULL) {
1275 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1276 BasicObjectLock* src = monitors->at(j);
1277 if (src->obj() != NULL) {
1278 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread);
1279 }
1280 }
1281 array->element(i)->free_monitors(thread);
1282 #ifdef ASSERT
1283 array->element(i)->set_removed_monitors();
1284 #endif
1285 }
1286 }
1287 }
1288 #endif
1289
1290 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1291 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1292 Thread* thread = Thread::current();
1293 for (int i = 0; i < monitors->length(); i++) {
1294 MonitorInfo* mon_info = monitors->at(i);
1295 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1296 objects_to_revoke->append(Handle(thread, mon_info->owner()));
1297 }
1298 }
1299 }
1300
1301
1302 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1303 if (!UseBiasedLocking) {
1304 return;
1305 }
1306
1307 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1308
1309 // Unfortunately we don't have a RegisterMap available in most of
1310 // the places we want to call this routine so we need to walk the
1311 // stack again to update the register map.
1312 if (map == NULL || !map->update_map()) {
1313 StackFrameStream sfs(thread, true);
1314 bool found = false;
1315 while (!found && !sfs.is_done()) {
1316 frame* cur = sfs.current();
1317 sfs.next();
1318 found = cur->id() == fr.id();
1319 }
1320 assert(found, "frame to be deoptimized not found on target thread's stack");
1321 map = sfs.register_map();
1322 }
1323
1324 vframe* vf = vframe::new_vframe(&fr, map, thread);
1325 compiledVFrame* cvf = compiledVFrame::cast(vf);
1326 // Revoke monitors' biases in all scopes
1327 while (!cvf->is_top()) {
1328 collect_monitors(cvf, objects_to_revoke);
1329 cvf = compiledVFrame::cast(cvf->sender());
1330 }
1331 collect_monitors(cvf, objects_to_revoke);
1332
1333 if (SafepointSynchronize::is_at_safepoint()) {
1334 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1335 } else {
1336 BiasedLocking::revoke(objects_to_revoke);
1337 }
1338 }
1339
1340
1341 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1342 if (!UseBiasedLocking) {
1343 return;
1344 }
1345
1346 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1347 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1348 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1349 if (jt->has_last_Java_frame()) {
1350 StackFrameStream sfs(jt, true);
1351 while (!sfs.is_done()) {
1352 frame* cur = sfs.current();
1353 if (cb->contains(cur->pc())) {
1354 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1355 compiledVFrame* cvf = compiledVFrame::cast(vf);
1356 // Revoke monitors' biases in all scopes
1357 while (!cvf->is_top()) {
1358 collect_monitors(cvf, objects_to_revoke);
1359 cvf = compiledVFrame::cast(cvf->sender());
1360 }
1361 collect_monitors(cvf, objects_to_revoke);
1362 }
1363 sfs.next();
1364 }
1365 }
1366 }
1367 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1368 }
1369
1370
1371 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1372 assert(fr.can_be_deoptimized(), "checking frame type");
1373
1374 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1375
1376 if (LogCompilation && xtty != NULL) {
1377 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1378 assert(cm != NULL, "only compiled methods can deopt");
1379
1380 ttyLocker ttyl;
1381 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1382 cm->log_identity(xtty);
1383 xtty->end_head();
1384 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1385 xtty->begin_elem("jvms bci='%d'", sd->bci());
1386 xtty->method(sd->method());
1387 xtty->end_elem();
1388 if (sd->is_top()) break;
1389 }
1390 xtty->tail("deoptimized");
1391 }
1392
1393 // Patch the compiled method so that when execution returns to it we will
1394 // deopt the execution state and return to the interpreter.
1395 fr.deoptimize(thread);
1396 }
1397
1398 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1399 deoptimize(thread, fr, map, Reason_constraint);
1400 }
1401
1402 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) {
1403 // Deoptimize only if the frame comes from compile code.
1404 // Do not deoptimize the frame which is already patched
1405 // during the execution of the loops below.
1406 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1407 return;
1408 }
1409 ResourceMark rm;
1410 DeoptimizationMarker dm;
1411 if (UseBiasedLocking) {
1412 revoke_biases_of_monitors(thread, fr, map);
1413 }
1414 deoptimize_single_frame(thread, fr, reason);
1415
1416 }
1417
1418
1419 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1420 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1421 "can only deoptimize other thread at a safepoint");
1422 // Compute frame and register map based on thread and sp.
1423 RegisterMap reg_map(thread, UseBiasedLocking);
1424 frame fr = thread->last_frame();
1425 while (fr.id() != id) {
1426 fr = fr.sender(®_map);
1427 }
1428 deoptimize(thread, fr, ®_map, reason);
1429 }
1430
1431
1432 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1433 if (thread == Thread::current()) {
1434 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1435 } else {
1436 VM_DeoptimizeFrame deopt(thread, id, reason);
1437 VMThread::execute(&deopt);
1438 }
1439 }
1440
1441 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1442 deoptimize_frame(thread, id, Reason_constraint);
1443 }
1444
1445 // JVMTI PopFrame support
1446 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1447 {
1448 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1449 }
1450 JRT_END
1451
1452 MethodData*
1453 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1454 bool create_if_missing) {
1455 Thread* THREAD = thread;
1456 MethodData* mdo = m()->method_data();
1457 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1458 // Build an MDO. Ignore errors like OutOfMemory;
1459 // that simply means we won't have an MDO to update.
1460 Method::build_interpreter_method_data(m, THREAD);
1461 if (HAS_PENDING_EXCEPTION) {
1462 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1463 CLEAR_PENDING_EXCEPTION;
1464 }
1465 mdo = m()->method_data();
1466 }
1467 return mdo;
1468 }
1469
1470 #if defined(COMPILER2) || defined(SHARK) || INCLUDE_JVMCI
1471 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1472 // in case of an unresolved klass entry, load the class.
1473 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1474 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1475 return;
1476 }
1477
1478 if (!constant_pool->tag_at(index).is_symbol()) return;
1479
1480 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1481 Symbol* symbol = constant_pool->symbol_at(index);
1482
1483 // class name?
1484 if (symbol->byte_at(0) != '(') {
1485 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1486 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1487 return;
1488 }
1489
1490 // then it must be a signature!
1491 ResourceMark rm(THREAD);
1492 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1493 if (ss.is_object()) {
1494 Symbol* class_name = ss.as_symbol(CHECK);
1495 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1496 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1497 }
1498 }
1499 }
1500
1501
1502 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1503 EXCEPTION_MARK;
1504 load_class_by_index(constant_pool, index, THREAD);
1505 if (HAS_PENDING_EXCEPTION) {
1506 // Exception happened during classloading. We ignore the exception here, since it
1507 // is going to be rethrown since the current activation is going to be deoptimized and
1508 // the interpreter will re-execute the bytecode.
1509 CLEAR_PENDING_EXCEPTION;
1510 // Class loading called java code which may have caused a stack
1511 // overflow. If the exception was thrown right before the return
1512 // to the runtime the stack is no longer guarded. Reguard the
1513 // stack otherwise if we return to the uncommon trap blob and the
1514 // stack bang causes a stack overflow we crash.
1515 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1516 JavaThread* thread = (JavaThread*)THREAD;
1517 bool guard_pages_enabled = thread->stack_guards_enabled();
1518 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1519 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1520 }
1521 }
1522
1523 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1524 HandleMark hm;
1525
1526 // uncommon_trap() is called at the beginning of the uncommon trap
1527 // handler. Note this fact before we start generating temporary frames
1528 // that can confuse an asynchronous stack walker. This counter is
1529 // decremented at the end of unpack_frames().
1530 thread->inc_in_deopt_handler();
1531
1532 // We need to update the map if we have biased locking.
1533 #if INCLUDE_JVMCI
1534 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1535 RegisterMap reg_map(thread, true);
1536 #else
1537 RegisterMap reg_map(thread, UseBiasedLocking);
1538 #endif
1539 frame stub_frame = thread->last_frame();
1540 frame fr = stub_frame.sender(®_map);
1541 // Make sure the calling nmethod is not getting deoptimized and removed
1542 // before we are done with it.
1543 nmethodLocker nl(fr.pc());
1544
1545 // Log a message
1546 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1547 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1548
1549 {
1550 ResourceMark rm;
1551
1552 // Revoke biases of any monitors in the frame to ensure we can migrate them
1553 revoke_biases_of_monitors(thread, fr, ®_map);
1554
1555 DeoptReason reason = trap_request_reason(trap_request);
1556 DeoptAction action = trap_request_action(trap_request);
1557 #if INCLUDE_JVMCI
1558 int debug_id = trap_request_debug_id(trap_request);
1559 #endif
1560 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1561
1562 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1563 compiledVFrame* cvf = compiledVFrame::cast(vf);
1564
1565 CompiledMethod* nm = cvf->code();
1566
1567 ScopeDesc* trap_scope = cvf->scope();
1568
1569 if (TraceDeoptimization) {
1570 ttyLocker ttyl;
1571 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()
1572 #if INCLUDE_JVMCI
1573 , debug_id
1574 #endif
1575 );
1576 }
1577
1578 methodHandle trap_method = trap_scope->method();
1579 int trap_bci = trap_scope->bci();
1580 #if INCLUDE_JVMCI
1581 oop speculation = thread->pending_failed_speculation();
1582 if (nm->is_compiled_by_jvmci()) {
1583 if (speculation != NULL) {
1584 oop speculation_log = nm->as_nmethod()->speculation_log();
1585 if (speculation_log != NULL) {
1586 if (TraceDeoptimization || TraceUncollectedSpeculations) {
1587 if (HotSpotSpeculationLog::lastFailed(speculation_log) != NULL) {
1588 tty->print_cr("A speculation that was not collected by the compiler is being overwritten");
1589 }
1590 }
1591 if (TraceDeoptimization) {
1592 tty->print_cr("Saving speculation to speculation log");
1593 }
1594 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation);
1595 } else {
1596 if (TraceDeoptimization) {
1597 tty->print_cr("Speculation present but no speculation log");
1598 }
1599 }
1600 thread->set_pending_failed_speculation(NULL);
1601 } else {
1602 if (TraceDeoptimization) {
1603 tty->print_cr("No speculation");
1604 }
1605 }
1606 } else {
1607 assert(speculation == NULL, "There should not be a speculation for method compiled by non-JVMCI compilers");
1608 }
1609
1610 if (trap_bci == SynchronizationEntryBCI) {
1611 trap_bci = 0;
1612 thread->set_pending_monitorenter(true);
1613 }
1614
1615 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1616 thread->set_pending_transfer_to_interpreter(true);
1617 }
1618 #endif
1619
1620 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1621 // Record this event in the histogram.
1622 gather_statistics(reason, action, trap_bc);
1623
1624 // Ensure that we can record deopt. history:
1625 // Need MDO to record RTM code generation state.
1626 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1627
1628 methodHandle profiled_method;
1629 #if INCLUDE_JVMCI
1630 if (nm->is_compiled_by_jvmci()) {
1631 profiled_method = nm->method();
1632 } else {
1633 profiled_method = trap_method;
1634 }
1635 #else
1636 profiled_method = trap_method;
1637 #endif
1638
1639 MethodData* trap_mdo =
1640 get_method_data(thread, profiled_method, create_if_missing);
1641
1642 // Log a message
1643 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
1644 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1645 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
1646
1647 // Print a bunch of diagnostics, if requested.
1648 if (TraceDeoptimization || LogCompilation) {
1649 ResourceMark rm;
1650 ttyLocker ttyl;
1651 char buf[100];
1652 if (xtty != NULL) {
1653 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1654 os::current_thread_id(),
1655 format_trap_request(buf, sizeof(buf), trap_request));
1656 nm->log_identity(xtty);
1657 }
1658 Symbol* class_name = NULL;
1659 bool unresolved = false;
1660 if (unloaded_class_index >= 0) {
1661 constantPoolHandle constants (THREAD, trap_method->constants());
1662 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1663 class_name = constants->klass_name_at(unloaded_class_index);
1664 unresolved = true;
1665 if (xtty != NULL)
1666 xtty->print(" unresolved='1'");
1667 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1668 class_name = constants->symbol_at(unloaded_class_index);
1669 }
1670 if (xtty != NULL)
1671 xtty->name(class_name);
1672 }
1673 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
1674 // Dump the relevant MDO state.
1675 // This is the deopt count for the current reason, any previous
1676 // reasons or recompiles seen at this point.
1677 int dcnt = trap_mdo->trap_count(reason);
1678 if (dcnt != 0)
1679 xtty->print(" count='%d'", dcnt);
1680 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1681 int dos = (pdata == NULL)? 0: pdata->trap_state();
1682 if (dos != 0) {
1683 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1684 if (trap_state_is_recompiled(dos)) {
1685 int recnt2 = trap_mdo->overflow_recompile_count();
1686 if (recnt2 != 0)
1687 xtty->print(" recompiles2='%d'", recnt2);
1688 }
1689 }
1690 }
1691 if (xtty != NULL) {
1692 xtty->stamp();
1693 xtty->end_head();
1694 }
1695 if (TraceDeoptimization) { // make noise on the tty
1696 tty->print("Uncommon trap occurred in");
1697 nm->method()->print_short_name(tty);
1698 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
1699 #if INCLUDE_JVMCI
1700 if (nm->is_nmethod()) {
1701 oop installedCode = nm->as_nmethod()->jvmci_installed_code();
1702 if (installedCode != NULL) {
1703 oop installedCodeName = NULL;
1704 if (installedCode->is_a(InstalledCode::klass())) {
1705 installedCodeName = InstalledCode::name(installedCode);
1706 }
1707 if (installedCodeName != NULL) {
1708 tty->print(" (JVMCI: installedCodeName=%s) ", java_lang_String::as_utf8_string(installedCodeName));
1709 } else {
1710 tty->print(" (JVMCI: installed code has no name) ");
1711 }
1712 } else if (nm->is_compiled_by_jvmci()) {
1713 tty->print(" (JVMCI: no installed code) ");
1714 }
1715 }
1716 #endif
1717 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
1718 p2i(fr.pc()),
1719 os::current_thread_id(),
1720 trap_reason_name(reason),
1721 trap_action_name(action),
1722 unloaded_class_index
1723 #if INCLUDE_JVMCI
1724 , debug_id
1725 #endif
1726 );
1727 if (class_name != NULL) {
1728 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1729 class_name->print_symbol_on(tty);
1730 }
1731 tty->cr();
1732 }
1733 if (xtty != NULL) {
1734 // Log the precise location of the trap.
1735 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1736 xtty->begin_elem("jvms bci='%d'", sd->bci());
1737 xtty->method(sd->method());
1738 xtty->end_elem();
1739 if (sd->is_top()) break;
1740 }
1741 xtty->tail("uncommon_trap");
1742 }
1743 }
1744 // (End diagnostic printout.)
1745
1746 // Load class if necessary
1747 if (unloaded_class_index >= 0) {
1748 constantPoolHandle constants(THREAD, trap_method->constants());
1749 load_class_by_index(constants, unloaded_class_index);
1750 }
1751
1752 // Flush the nmethod if necessary and desirable.
1753 //
1754 // We need to avoid situations where we are re-flushing the nmethod
1755 // because of a hot deoptimization site. Repeated flushes at the same
1756 // point need to be detected by the compiler and avoided. If the compiler
1757 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1758 // module must take measures to avoid an infinite cycle of recompilation
1759 // and deoptimization. There are several such measures:
1760 //
1761 // 1. If a recompilation is ordered a second time at some site X
1762 // and for the same reason R, the action is adjusted to 'reinterpret',
1763 // to give the interpreter time to exercise the method more thoroughly.
1764 // If this happens, the method's overflow_recompile_count is incremented.
1765 //
1766 // 2. If the compiler fails to reduce the deoptimization rate, then
1767 // the method's overflow_recompile_count will begin to exceed the set
1768 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1769 // is adjusted to 'make_not_compilable', and the method is abandoned
1770 // to the interpreter. This is a performance hit for hot methods,
1771 // but is better than a disastrous infinite cycle of recompilations.
1772 // (Actually, only the method containing the site X is abandoned.)
1773 //
1774 // 3. In parallel with the previous measures, if the total number of
1775 // recompilations of a method exceeds the much larger set limit
1776 // PerMethodRecompilationCutoff, the method is abandoned.
1777 // This should only happen if the method is very large and has
1778 // many "lukewarm" deoptimizations. The code which enforces this
1779 // limit is elsewhere (class nmethod, class Method).
1780 //
1781 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1782 // to recompile at each bytecode independently of the per-BCI cutoff.
1783 //
1784 // The decision to update code is up to the compiler, and is encoded
1785 // in the Action_xxx code. If the compiler requests Action_none
1786 // no trap state is changed, no compiled code is changed, and the
1787 // computation suffers along in the interpreter.
1788 //
1789 // The other action codes specify various tactics for decompilation
1790 // and recompilation. Action_maybe_recompile is the loosest, and
1791 // allows the compiled code to stay around until enough traps are seen,
1792 // and until the compiler gets around to recompiling the trapping method.
1793 //
1794 // The other actions cause immediate removal of the present code.
1795
1796 // Traps caused by injected profile shouldn't pollute trap counts.
1797 bool injected_profile_trap = trap_method->has_injected_profile() &&
1798 (reason == Reason_intrinsic || reason == Reason_unreached);
1799
1800 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
1801 bool make_not_entrant = false;
1802 bool make_not_compilable = false;
1803 bool reprofile = false;
1804 switch (action) {
1805 case Action_none:
1806 // Keep the old code.
1807 update_trap_state = false;
1808 break;
1809 case Action_maybe_recompile:
1810 // Do not need to invalidate the present code, but we can
1811 // initiate another
1812 // Start compiler without (necessarily) invalidating the nmethod.
1813 // The system will tolerate the old code, but new code should be
1814 // generated when possible.
1815 break;
1816 case Action_reinterpret:
1817 // Go back into the interpreter for a while, and then consider
1818 // recompiling form scratch.
1819 make_not_entrant = true;
1820 // Reset invocation counter for outer most method.
1821 // This will allow the interpreter to exercise the bytecodes
1822 // for a while before recompiling.
1823 // By contrast, Action_make_not_entrant is immediate.
1824 //
1825 // Note that the compiler will track null_check, null_assert,
1826 // range_check, and class_check events and log them as if they
1827 // had been traps taken from compiled code. This will update
1828 // the MDO trap history so that the next compilation will
1829 // properly detect hot trap sites.
1830 reprofile = true;
1831 break;
1832 case Action_make_not_entrant:
1833 // Request immediate recompilation, and get rid of the old code.
1834 // Make them not entrant, so next time they are called they get
1835 // recompiled. Unloaded classes are loaded now so recompile before next
1836 // time they are called. Same for uninitialized. The interpreter will
1837 // link the missing class, if any.
1838 make_not_entrant = true;
1839 break;
1840 case Action_make_not_compilable:
1841 // Give up on compiling this method at all.
1842 make_not_entrant = true;
1843 make_not_compilable = true;
1844 break;
1845 default:
1846 ShouldNotReachHere();
1847 }
1848
1849 // Setting +ProfileTraps fixes the following, on all platforms:
1850 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1851 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1852 // recompile relies on a MethodData* to record heroic opt failures.
1853
1854 // Whether the interpreter is producing MDO data or not, we also need
1855 // to use the MDO to detect hot deoptimization points and control
1856 // aggressive optimization.
1857 bool inc_recompile_count = false;
1858 ProfileData* pdata = NULL;
1859 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) {
1860 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
1861 uint this_trap_count = 0;
1862 bool maybe_prior_trap = false;
1863 bool maybe_prior_recompile = false;
1864 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
1865 #if INCLUDE_JVMCI
1866 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
1867 #endif
1868 nm->method(),
1869 //outputs:
1870 this_trap_count,
1871 maybe_prior_trap,
1872 maybe_prior_recompile);
1873 // Because the interpreter also counts null, div0, range, and class
1874 // checks, these traps from compiled code are double-counted.
1875 // This is harmless; it just means that the PerXTrapLimit values
1876 // are in effect a little smaller than they look.
1877
1878 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1879 if (per_bc_reason != Reason_none) {
1880 // Now take action based on the partially known per-BCI history.
1881 if (maybe_prior_trap
1882 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1883 // If there are too many traps at this BCI, force a recompile.
1884 // This will allow the compiler to see the limit overflow, and
1885 // take corrective action, if possible. The compiler generally
1886 // does not use the exact PerBytecodeTrapLimit value, but instead
1887 // changes its tactics if it sees any traps at all. This provides
1888 // a little hysteresis, delaying a recompile until a trap happens
1889 // several times.
1890 //
1891 // Actually, since there is only one bit of counter per BCI,
1892 // the possible per-BCI counts are {0,1,(per-method count)}.
1893 // This produces accurate results if in fact there is only
1894 // one hot trap site, but begins to get fuzzy if there are
1895 // many sites. For example, if there are ten sites each
1896 // trapping two or more times, they each get the blame for
1897 // all of their traps.
1898 make_not_entrant = true;
1899 }
1900
1901 // Detect repeated recompilation at the same BCI, and enforce a limit.
1902 if (make_not_entrant && maybe_prior_recompile) {
1903 // More than one recompile at this point.
1904 inc_recompile_count = maybe_prior_trap;
1905 }
1906 } else {
1907 // For reasons which are not recorded per-bytecode, we simply
1908 // force recompiles unconditionally.
1909 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1910 make_not_entrant = true;
1911 }
1912
1913 // Go back to the compiler if there are too many traps in this method.
1914 if (this_trap_count >= per_method_trap_limit(reason)) {
1915 // If there are too many traps in this method, force a recompile.
1916 // This will allow the compiler to see the limit overflow, and
1917 // take corrective action, if possible.
1918 // (This condition is an unlikely backstop only, because the
1919 // PerBytecodeTrapLimit is more likely to take effect first,
1920 // if it is applicable.)
1921 make_not_entrant = true;
1922 }
1923
1924 // Here's more hysteresis: If there has been a recompile at
1925 // this trap point already, run the method in the interpreter
1926 // for a while to exercise it more thoroughly.
1927 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1928 reprofile = true;
1929 }
1930 }
1931
1932 // Take requested actions on the method:
1933
1934 // Recompile
1935 if (make_not_entrant) {
1936 if (!nm->make_not_entrant()) {
1937 return; // the call did not change nmethod's state
1938 }
1939
1940 if (pdata != NULL) {
1941 // Record the recompilation event, if any.
1942 int tstate0 = pdata->trap_state();
1943 int tstate1 = trap_state_set_recompiled(tstate0, true);
1944 if (tstate1 != tstate0)
1945 pdata->set_trap_state(tstate1);
1946 }
1947
1948 #if INCLUDE_RTM_OPT
1949 // Restart collecting RTM locking abort statistic if the method
1950 // is recompiled for a reason other than RTM state change.
1951 // Assume that in new recompiled code the statistic could be different,
1952 // for example, due to different inlining.
1953 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
1954 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
1955 trap_mdo->atomic_set_rtm_state(ProfileRTM);
1956 }
1957 #endif
1958 // For code aging we count traps separately here, using make_not_entrant()
1959 // as a guard against simultaneous deopts in multiple threads.
1960 if (reason == Reason_tenured && trap_mdo != NULL) {
1961 trap_mdo->inc_tenure_traps();
1962 }
1963 }
1964
1965 if (inc_recompile_count) {
1966 trap_mdo->inc_overflow_recompile_count();
1967 if ((uint)trap_mdo->overflow_recompile_count() >
1968 (uint)PerBytecodeRecompilationCutoff) {
1969 // Give up on the method containing the bad BCI.
1970 if (trap_method() == nm->method()) {
1971 make_not_compilable = true;
1972 } else {
1973 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
1974 // But give grace to the enclosing nm->method().
1975 }
1976 }
1977 }
1978
1979 // Reprofile
1980 if (reprofile) {
1981 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1982 }
1983
1984 // Give up compiling
1985 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1986 assert(make_not_entrant, "consistent");
1987 nm->method()->set_not_compilable(CompLevel_full_optimization);
1988 }
1989
1990 } // Free marked resources
1991
1992 }
1993 JRT_END
1994
1995 ProfileData*
1996 Deoptimization::query_update_method_data(MethodData* trap_mdo,
1997 int trap_bci,
1998 Deoptimization::DeoptReason reason,
1999 bool update_total_trap_count,
2000 #if INCLUDE_JVMCI
2001 bool is_osr,
2002 #endif
2003 Method* compiled_method,
2004 //outputs:
2005 uint& ret_this_trap_count,
2006 bool& ret_maybe_prior_trap,
2007 bool& ret_maybe_prior_recompile) {
2008 bool maybe_prior_trap = false;
2009 bool maybe_prior_recompile = false;
2010 uint this_trap_count = 0;
2011 if (update_total_trap_count) {
2012 uint idx = reason;
2013 #if INCLUDE_JVMCI
2014 if (is_osr) {
2015 idx += Reason_LIMIT;
2016 }
2017 #endif
2018 uint prior_trap_count = trap_mdo->trap_count(idx);
2019 this_trap_count = trap_mdo->inc_trap_count(idx);
2020
2021 // If the runtime cannot find a place to store trap history,
2022 // it is estimated based on the general condition of the method.
2023 // If the method has ever been recompiled, or has ever incurred
2024 // a trap with the present reason , then this BCI is assumed
2025 // (pessimistically) to be the culprit.
2026 maybe_prior_trap = (prior_trap_count != 0);
2027 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2028 }
2029 ProfileData* pdata = NULL;
2030
2031
2032 // For reasons which are recorded per bytecode, we check per-BCI data.
2033 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2034 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2035 if (per_bc_reason != Reason_none) {
2036 // Find the profile data for this BCI. If there isn't one,
2037 // try to allocate one from the MDO's set of spares.
2038 // This will let us detect a repeated trap at this point.
2039 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
2040
2041 if (pdata != NULL) {
2042 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2043 if (LogCompilation && xtty != NULL) {
2044 ttyLocker ttyl;
2045 // no more room for speculative traps in this MDO
2046 xtty->elem("speculative_traps_oom");
2047 }
2048 }
2049 // Query the trap state of this profile datum.
2050 int tstate0 = pdata->trap_state();
2051 if (!trap_state_has_reason(tstate0, per_bc_reason))
2052 maybe_prior_trap = false;
2053 if (!trap_state_is_recompiled(tstate0))
2054 maybe_prior_recompile = false;
2055
2056 // Update the trap state of this profile datum.
2057 int tstate1 = tstate0;
2058 // Record the reason.
2059 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2060 // Store the updated state on the MDO, for next time.
2061 if (tstate1 != tstate0)
2062 pdata->set_trap_state(tstate1);
2063 } else {
2064 if (LogCompilation && xtty != NULL) {
2065 ttyLocker ttyl;
2066 // Missing MDP? Leave a small complaint in the log.
2067 xtty->elem("missing_mdp bci='%d'", trap_bci);
2068 }
2069 }
2070 }
2071
2072 // Return results:
2073 ret_this_trap_count = this_trap_count;
2074 ret_maybe_prior_trap = maybe_prior_trap;
2075 ret_maybe_prior_recompile = maybe_prior_recompile;
2076 return pdata;
2077 }
2078
2079 void
2080 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2081 ResourceMark rm;
2082 // Ignored outputs:
2083 uint ignore_this_trap_count;
2084 bool ignore_maybe_prior_trap;
2085 bool ignore_maybe_prior_recompile;
2086 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2087 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2088 bool update_total_counts = JVMCI_ONLY(false) NOT_JVMCI(true);
2089 query_update_method_data(trap_mdo, trap_bci,
2090 (DeoptReason)reason,
2091 update_total_counts,
2092 #if INCLUDE_JVMCI
2093 false,
2094 #endif
2095 NULL,
2096 ignore_this_trap_count,
2097 ignore_maybe_prior_trap,
2098 ignore_maybe_prior_recompile);
2099 }
2100
2101 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2102 if (TraceDeoptimization) {
2103 tty->print("Uncommon trap ");
2104 }
2105 // Still in Java no safepoints
2106 {
2107 // This enters VM and may safepoint
2108 uncommon_trap_inner(thread, trap_request);
2109 }
2110 return fetch_unroll_info_helper(thread, exec_mode);
2111 }
2112
2113 // Local derived constants.
2114 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2115 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
2116 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2117
2118 //---------------------------trap_state_reason---------------------------------
2119 Deoptimization::DeoptReason
2120 Deoptimization::trap_state_reason(int trap_state) {
2121 // This assert provides the link between the width of DataLayout::trap_bits
2122 // and the encoding of "recorded" reasons. It ensures there are enough
2123 // bits to store all needed reasons in the per-BCI MDO profile.
2124 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2125 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2126 trap_state -= recompile_bit;
2127 if (trap_state == DS_REASON_MASK) {
2128 return Reason_many;
2129 } else {
2130 assert((int)Reason_none == 0, "state=0 => Reason_none");
2131 return (DeoptReason)trap_state;
2132 }
2133 }
2134 //-------------------------trap_state_has_reason-------------------------------
2135 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2136 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2137 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2138 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2139 trap_state -= recompile_bit;
2140 if (trap_state == DS_REASON_MASK) {
2141 return -1; // true, unspecifically (bottom of state lattice)
2142 } else if (trap_state == reason) {
2143 return 1; // true, definitely
2144 } else if (trap_state == 0) {
2145 return 0; // false, definitely (top of state lattice)
2146 } else {
2147 return 0; // false, definitely
2148 }
2149 }
2150 //-------------------------trap_state_add_reason-------------------------------
2151 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2152 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2153 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2154 trap_state -= recompile_bit;
2155 if (trap_state == DS_REASON_MASK) {
2156 return trap_state + recompile_bit; // already at state lattice bottom
2157 } else if (trap_state == reason) {
2158 return trap_state + recompile_bit; // the condition is already true
2159 } else if (trap_state == 0) {
2160 return reason + recompile_bit; // no condition has yet been true
2161 } else {
2162 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2163 }
2164 }
2165 //-----------------------trap_state_is_recompiled------------------------------
2166 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2167 return (trap_state & DS_RECOMPILE_BIT) != 0;
2168 }
2169 //-----------------------trap_state_set_recompiled-----------------------------
2170 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2171 if (z) return trap_state | DS_RECOMPILE_BIT;
2172 else return trap_state & ~DS_RECOMPILE_BIT;
2173 }
2174 //---------------------------format_trap_state---------------------------------
2175 // This is used for debugging and diagnostics, including LogFile output.
2176 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2177 int trap_state) {
2178 assert(buflen > 0, "sanity");
2179 DeoptReason reason = trap_state_reason(trap_state);
2180 bool recomp_flag = trap_state_is_recompiled(trap_state);
2181 // Re-encode the state from its decoded components.
2182 int decoded_state = 0;
2183 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2184 decoded_state = trap_state_add_reason(decoded_state, reason);
2185 if (recomp_flag)
2186 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2187 // If the state re-encodes properly, format it symbolically.
2188 // Because this routine is used for debugging and diagnostics,
2189 // be robust even if the state is a strange value.
2190 size_t len;
2191 if (decoded_state != trap_state) {
2192 // Random buggy state that doesn't decode??
2193 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2194 } else {
2195 len = jio_snprintf(buf, buflen, "%s%s",
2196 trap_reason_name(reason),
2197 recomp_flag ? " recompiled" : "");
2198 }
2199 return buf;
2200 }
2201
2202
2203 //--------------------------------statics--------------------------------------
2204 const char* Deoptimization::_trap_reason_name[] = {
2205 // Note: Keep this in sync. with enum DeoptReason.
2206 "none",
2207 "null_check",
2208 "null_assert" JVMCI_ONLY("_or_unreached0"),
2209 "range_check",
2210 "class_check",
2211 "array_check",
2212 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2213 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2214 "unloaded",
2215 "uninitialized",
2216 "unreached",
2217 "unhandled",
2218 "constraint",
2219 "div0_check",
2220 "age",
2221 "predicate",
2222 "loop_limit_check",
2223 "speculate_class_check",
2224 "speculate_null_check",
2225 "rtm_state_change",
2226 "unstable_if",
2227 "unstable_fused_if",
2228 #if INCLUDE_JVMCI
2229 "aliasing",
2230 "transfer_to_interpreter",
2231 "not_compiled_exception_handler",
2232 "unresolved",
2233 "jsr_mismatch",
2234 #endif
2235 "tenured"
2236 };
2237 const char* Deoptimization::_trap_action_name[] = {
2238 // Note: Keep this in sync. with enum DeoptAction.
2239 "none",
2240 "maybe_recompile",
2241 "reinterpret",
2242 "make_not_entrant",
2243 "make_not_compilable"
2244 };
2245
2246 const char* Deoptimization::trap_reason_name(int reason) {
2247 // Check that every reason has a name
2248 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2249
2250 if (reason == Reason_many) return "many";
2251 if ((uint)reason < Reason_LIMIT)
2252 return _trap_reason_name[reason];
2253 static char buf[20];
2254 sprintf(buf, "reason%d", reason);
2255 return buf;
2256 }
2257 const char* Deoptimization::trap_action_name(int action) {
2258 // Check that every action has a name
2259 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2260
2261 if ((uint)action < Action_LIMIT)
2262 return _trap_action_name[action];
2263 static char buf[20];
2264 sprintf(buf, "action%d", action);
2265 return buf;
2266 }
2267
2268 // This is used for debugging and diagnostics, including LogFile output.
2269 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2270 int trap_request) {
2271 jint unloaded_class_index = trap_request_index(trap_request);
2272 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2273 const char* action = trap_action_name(trap_request_action(trap_request));
2274 #if INCLUDE_JVMCI
2275 int debug_id = trap_request_debug_id(trap_request);
2276 #endif
2277 size_t len;
2278 if (unloaded_class_index < 0) {
2279 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2280 reason, action
2281 #if INCLUDE_JVMCI
2282 ,debug_id
2283 #endif
2284 );
2285 } else {
2286 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2287 reason, action, unloaded_class_index
2288 #if INCLUDE_JVMCI
2289 ,debug_id
2290 #endif
2291 );
2292 }
2293 return buf;
2294 }
2295
2296 juint Deoptimization::_deoptimization_hist
2297 [Deoptimization::Reason_LIMIT]
2298 [1 + Deoptimization::Action_LIMIT]
2299 [Deoptimization::BC_CASE_LIMIT]
2300 = {0};
2301
2302 enum {
2303 LSB_BITS = 8,
2304 LSB_MASK = right_n_bits(LSB_BITS)
2305 };
2306
2307 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2308 Bytecodes::Code bc) {
2309 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2310 assert(action >= 0 && action < Action_LIMIT, "oob");
2311 _deoptimization_hist[Reason_none][0][0] += 1; // total
2312 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2313 juint* cases = _deoptimization_hist[reason][1+action];
2314 juint* bc_counter_addr = NULL;
2315 juint bc_counter = 0;
2316 // Look for an unused counter, or an exact match to this BC.
2317 if (bc != Bytecodes::_illegal) {
2318 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2319 juint* counter_addr = &cases[bc_case];
2320 juint counter = *counter_addr;
2321 if ((counter == 0 && bc_counter_addr == NULL)
2322 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2323 // this counter is either free or is already devoted to this BC
2324 bc_counter_addr = counter_addr;
2325 bc_counter = counter | bc;
2326 }
2327 }
2328 }
2329 if (bc_counter_addr == NULL) {
2330 // Overflow, or no given bytecode.
2331 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2332 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2333 }
2334 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2335 }
2336
2337 jint Deoptimization::total_deoptimization_count() {
2338 return _deoptimization_hist[Reason_none][0][0];
2339 }
2340
2341 jint Deoptimization::deoptimization_count(DeoptReason reason) {
2342 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2343 return _deoptimization_hist[reason][0][0];
2344 }
2345
2346 void Deoptimization::print_statistics() {
2347 juint total = total_deoptimization_count();
2348 juint account = total;
2349 if (total != 0) {
2350 ttyLocker ttyl;
2351 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2352 tty->print_cr("Deoptimization traps recorded:");
2353 #define PRINT_STAT_LINE(name, r) \
2354 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2355 PRINT_STAT_LINE("total", total);
2356 // For each non-zero entry in the histogram, print the reason,
2357 // the action, and (if specifically known) the type of bytecode.
2358 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2359 for (int action = 0; action < Action_LIMIT; action++) {
2360 juint* cases = _deoptimization_hist[reason][1+action];
2361 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2362 juint counter = cases[bc_case];
2363 if (counter != 0) {
2364 char name[1*K];
2365 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2366 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2367 bc = Bytecodes::_illegal;
2368 sprintf(name, "%s/%s/%s",
2369 trap_reason_name(reason),
2370 trap_action_name(action),
2371 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2372 juint r = counter >> LSB_BITS;
2373 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2374 account -= r;
2375 }
2376 }
2377 }
2378 }
2379 if (account != 0) {
2380 PRINT_STAT_LINE("unaccounted", account);
2381 }
2382 #undef PRINT_STAT_LINE
2383 if (xtty != NULL) xtty->tail("statistics");
2384 }
2385 }
2386 #else // COMPILER2 || SHARK || INCLUDE_JVMCI
2387
2388
2389 // Stubs for C1 only system.
2390 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2391 return false;
2392 }
2393
2394 const char* Deoptimization::trap_reason_name(int reason) {
2395 return "unknown";
2396 }
2397
2398 void Deoptimization::print_statistics() {
2399 // no output
2400 }
2401
2402 void
2403 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2404 // no udpate
2405 }
2406
2407 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2408 return 0;
2409 }
2410
2411 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2412 Bytecodes::Code bc) {
2413 // no update
2414 }
2415
2416 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2417 int trap_state) {
2418 jio_snprintf(buf, buflen, "#%d", trap_state);
2419 return buf;
2420 }
2421
2422 #endif // COMPILER2 || SHARK || INCLUDE_JVMCI