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 "jvm.h"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/codeCache.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "code/nmethod.hpp"
31 #include "code/pcDesc.hpp"
32 #include "code/scopeDesc.hpp"
33 #include "interpreter/bytecode.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "interpreter/oopMapCache.hpp"
36 #include "memory/allocation.inline.hpp"
37 #include "memory/oopFactory.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "oops/method.hpp"
40 #include "oops/objArrayOop.inline.hpp"
41 #include "oops/oop.inline.hpp"
42 #include "oops/fieldStreams.hpp"
43 #include "oops/typeArrayOop.inline.hpp"
44 #include "oops/verifyOopClosure.hpp"
45 #include "prims/jvmtiThreadState.hpp"
46 #include "runtime/biasedLocking.hpp"
47 #include "runtime/compilationPolicy.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/frame.inline.hpp"
50 #include "runtime/interfaceSupport.inline.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "runtime/signature.hpp"
53 #include "runtime/stubRoutines.hpp"
54 #include "runtime/thread.hpp"
55 #include "runtime/threadSMR.hpp"
56 #include "runtime/vframe.hpp"
57 #include "runtime/vframeArray.hpp"
58 #include "runtime/vframe_hp.hpp"
59 #include "utilities/events.hpp"
60 #include "utilities/preserveException.hpp"
61 #include "utilities/xmlstream.hpp"
62
63 #if INCLUDE_JVMCI
64 #include "jvmci/jvmciRuntime.hpp"
65 #include "jvmci/jvmciJavaClasses.hpp"
66 #endif
67
68
69 bool DeoptimizationMarker::_is_active = false;
70
71 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
72 int caller_adjustment,
73 int caller_actual_parameters,
74 int number_of_frames,
75 intptr_t* frame_sizes,
76 address* frame_pcs,
77 BasicType return_type,
78 int exec_mode) {
79 _size_of_deoptimized_frame = size_of_deoptimized_frame;
80 _caller_adjustment = caller_adjustment;
81 _caller_actual_parameters = caller_actual_parameters;
82 _number_of_frames = number_of_frames;
83 _frame_sizes = frame_sizes;
84 _frame_pcs = frame_pcs;
85 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
86 _return_type = return_type;
87 _initial_info = 0;
88 // PD (x86 only)
89 _counter_temp = 0;
90 _unpack_kind = exec_mode;
91 _sender_sp_temp = 0;
92
93 _total_frame_sizes = size_of_frames();
94 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
95 }
96
97
98 Deoptimization::UnrollBlock::~UnrollBlock() {
99 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
100 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
101 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
102 }
103
104
105 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
106 assert(register_number < RegisterMap::reg_count, "checking register number");
107 return &_register_block[register_number * 2];
108 }
109
110
111
112 int Deoptimization::UnrollBlock::size_of_frames() const {
113 // Acount first for the adjustment of the initial frame
114 int result = _caller_adjustment;
115 for (int index = 0; index < number_of_frames(); index++) {
116 result += frame_sizes()[index];
117 }
118 return result;
119 }
120
121
122 void Deoptimization::UnrollBlock::print() {
123 ttyLocker ttyl;
124 tty->print_cr("UnrollBlock");
125 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
126 tty->print( " frame_sizes: ");
127 for (int index = 0; index < number_of_frames(); index++) {
128 tty->print(INTX_FORMAT " ", frame_sizes()[index]);
129 }
130 tty->cr();
131 }
132
133
134 // In order to make fetch_unroll_info work properly with escape
135 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
136 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
137 // of previously eliminated objects occurs in realloc_objects, which is
138 // called from the method fetch_unroll_info_helper below.
139 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode))
140 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
141 // but makes the entry a little slower. There is however a little dance we have to
142 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
143
144 // fetch_unroll_info() is called at the beginning of the deoptimization
145 // handler. Note this fact before we start generating temporary frames
146 // that can confuse an asynchronous stack walker. This counter is
147 // decremented at the end of unpack_frames().
148 if (TraceDeoptimization) {
149 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread));
150 }
151 thread->inc_in_deopt_handler();
152
153 return fetch_unroll_info_helper(thread, exec_mode);
154 JRT_END
155
156
157 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
158 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) {
159
160 // Note: there is a safepoint safety issue here. No matter whether we enter
161 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
162 // the vframeArray is created.
163 //
164
165 // Allocate our special deoptimization ResourceMark
166 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
167 assert(thread->deopt_mark() == NULL, "Pending deopt!");
168 thread->set_deopt_mark(dmark);
169
170 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
171 RegisterMap map(thread, true);
172 RegisterMap dummy_map(thread, false);
173 // Now get the deoptee with a valid map
174 frame deoptee = stub_frame.sender(&map);
175 // Set the deoptee nmethod
176 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!");
177 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
178 thread->set_deopt_compiled_method(cm);
179
180 if (VerifyStack) {
181 thread->validate_frame_layout();
182 }
183
184 // Create a growable array of VFrames where each VFrame represents an inlined
185 // Java frame. This storage is allocated with the usual system arena.
186 assert(deoptee.is_compiled_frame(), "Wrong frame type");
187 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
188 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
189 while (!vf->is_top()) {
190 assert(vf->is_compiled_frame(), "Wrong frame type");
191 chunk->push(compiledVFrame::cast(vf));
192 vf = vf->sender();
193 }
194 assert(vf->is_compiled_frame(), "Wrong frame type");
195 chunk->push(compiledVFrame::cast(vf));
196
197 bool realloc_failures = false;
198
199 #if COMPILER2_OR_JVMCI
200 // Reallocate the non-escaping objects and restore their fields. Then
201 // relock objects if synchronization on them was eliminated.
202 #ifndef INCLUDE_JVMCI
203 if (DoEscapeAnalysis || EliminateNestedLocks) {
204 if (EliminateAllocations) {
205 #endif // INCLUDE_JVMCI
206 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
207 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
208
209 // The flag return_oop() indicates call sites which return oop
210 // in compiled code. Such sites include java method calls,
211 // runtime calls (for example, used to allocate new objects/arrays
212 // on slow code path) and any other calls generated in compiled code.
213 // It is not guaranteed that we can get such information here only
214 // by analyzing bytecode in deoptimized frames. This is why this flag
215 // is set during method compilation (see Compile::Process_OopMap_Node()).
216 // If the previous frame was popped or if we are dispatching an exception,
217 // we don't have an oop result.
218 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt);
219 Handle return_value;
220 if (save_oop_result) {
221 // Reallocation may trigger GC. If deoptimization happened on return from
222 // call which returns oop we need to save it since it is not in oopmap.
223 oop result = deoptee.saved_oop_result(&map);
224 assert(oopDesc::is_oop_or_null(result), "must be oop");
225 return_value = Handle(thread, result);
226 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
227 if (TraceDeoptimization) {
228 ttyLocker ttyl;
229 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
230 }
231 }
232 if (objects != NULL) {
233 JRT_BLOCK
234 realloc_failures = realloc_objects(thread, &deoptee, objects, THREAD);
235 JRT_END
236 bool skip_internal = (cm != NULL) && !cm->is_compiled_by_jvmci();
237 reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
238 #ifndef PRODUCT
239 if (TraceDeoptimization) {
240 ttyLocker ttyl;
241 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
242 print_objects(objects, realloc_failures);
243 }
244 #endif
245 }
246 if (save_oop_result) {
247 // Restore result.
248 deoptee.set_saved_oop_result(&map, return_value());
249 }
250 #ifndef INCLUDE_JVMCI
251 }
252 if (EliminateLocks) {
253 #endif // INCLUDE_JVMCI
254 #ifndef PRODUCT
255 bool first = true;
256 #endif
257 for (int i = 0; i < chunk->length(); i++) {
258 compiledVFrame* cvf = chunk->at(i);
259 assert (cvf->scope() != NULL,"expect only compiled java frames");
260 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
261 if (monitors->is_nonempty()) {
262 relock_objects(monitors, thread, realloc_failures);
263 #ifndef PRODUCT
264 if (PrintDeoptimizationDetails) {
265 ttyLocker ttyl;
266 for (int j = 0; j < monitors->length(); j++) {
267 MonitorInfo* mi = monitors->at(j);
268 if (mi->eliminated()) {
269 if (first) {
270 first = false;
271 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
272 }
273 if (mi->owner_is_scalar_replaced()) {
274 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
275 tty->print_cr(" failed reallocation for klass %s", k->external_name());
276 } else {
277 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
278 }
279 }
280 }
281 }
282 #endif // !PRODUCT
283 }
284 }
285 #ifndef INCLUDE_JVMCI
286 }
287 }
288 #endif // INCLUDE_JVMCI
289 #endif // COMPILER2_OR_JVMCI
290
291 ScopeDesc* trap_scope = chunk->at(0)->scope();
292 Handle exceptionObject;
293 if (trap_scope->rethrow_exception()) {
294 if (PrintDeoptimizationDetails) {
295 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());
296 }
297 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
298 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw");
299 ScopeValue* topOfStack = expressions->top();
300 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
301 guarantee(exceptionObject() != NULL, "exception oop can not be null");
302 }
303
304 // Ensure that no safepoint is taken after pointers have been stored
305 // in fields of rematerialized objects. If a safepoint occurs from here on
306 // out the java state residing in the vframeArray will be missed.
307 NoSafepointVerifier no_safepoint;
308
309 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
310 #if COMPILER2_OR_JVMCI
311 if (realloc_failures) {
312 pop_frames_failed_reallocs(thread, array);
313 }
314 #endif
315
316 assert(thread->vframe_array_head() == NULL, "Pending deopt!");
317 thread->set_vframe_array_head(array);
318
319 // Now that the vframeArray has been created if we have any deferred local writes
320 // added by jvmti then we can free up that structure as the data is now in the
321 // vframeArray
322
323 if (thread->deferred_locals() != NULL) {
324 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
325 int i = 0;
326 do {
327 // Because of inlining we could have multiple vframes for a single frame
328 // and several of the vframes could have deferred writes. Find them all.
329 if (list->at(i)->id() == array->original().id()) {
330 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
331 list->remove_at(i);
332 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
333 delete dlv;
334 } else {
335 i++;
336 }
337 } while ( i < list->length() );
338 if (list->length() == 0) {
339 thread->set_deferred_locals(NULL);
340 // free the list and elements back to C heap.
341 delete list;
342 }
343
344 }
345
346 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
347 CodeBlob* cb = stub_frame.cb();
348 // Verify we have the right vframeArray
349 assert(cb->frame_size() >= 0, "Unexpected frame size");
350 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
351
352 // If the deopt call site is a MethodHandle invoke call site we have
353 // to adjust the unpack_sp.
354 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
355 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
356 unpack_sp = deoptee.unextended_sp();
357
358 #ifdef ASSERT
359 assert(cb->is_deoptimization_stub() ||
360 cb->is_uncommon_trap_stub() ||
361 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
362 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
363 "unexpected code blob: %s", cb->name());
364 #endif
365
366 // This is a guarantee instead of an assert because if vframe doesn't match
367 // we will unpack the wrong deoptimized frame and wind up in strange places
368 // where it will be very difficult to figure out what went wrong. Better
369 // to die an early death here than some very obscure death later when the
370 // trail is cold.
371 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
372 // in that it will fail to detect a problem when there is one. This needs
373 // more work in tiger timeframe.
374 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
375
376 int number_of_frames = array->frames();
377
378 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
379 // virtual activation, which is the reverse of the elements in the vframes array.
380 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
381 // +1 because we always have an interpreter return address for the final slot.
382 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
383 int popframe_extra_args = 0;
384 // Create an interpreter return address for the stub to use as its return
385 // address so the skeletal frames are perfectly walkable
386 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
387
388 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
389 // activation be put back on the expression stack of the caller for reexecution
390 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
391 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
392 }
393
394 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
395 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
396 // than simply use array->sender.pc(). This requires us to walk the current set of frames
397 //
398 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
399 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
400
401 // It's possible that the number of parameters at the call site is
402 // different than number of arguments in the callee when method
403 // handles are used. If the caller is interpreted get the real
404 // value so that the proper amount of space can be added to it's
405 // frame.
406 bool caller_was_method_handle = false;
407 if (deopt_sender.is_interpreted_frame()) {
408 methodHandle method = deopt_sender.interpreter_frame_method();
409 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
410 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
411 // Method handle invokes may involve fairly arbitrary chains of
412 // calls so it's impossible to know how much actual space the
413 // caller has for locals.
414 caller_was_method_handle = true;
415 }
416 }
417
418 //
419 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
420 // frame_sizes/frame_pcs[1] next oldest frame (int)
421 // frame_sizes/frame_pcs[n] youngest frame (int)
422 //
423 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
424 // owns the space for the return address to it's caller). Confusing ain't it.
425 //
426 // The vframe array can address vframes with indices running from
427 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
428 // When we create the skeletal frames we need the oldest frame to be in the zero slot
429 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
430 // so things look a little strange in this loop.
431 //
432 int callee_parameters = 0;
433 int callee_locals = 0;
434 for (int index = 0; index < array->frames(); index++ ) {
435 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
436 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
437 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
438 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
439 callee_locals,
440 index == 0,
441 popframe_extra_args);
442 // This pc doesn't have to be perfect just good enough to identify the frame
443 // as interpreted so the skeleton frame will be walkable
444 // The correct pc will be set when the skeleton frame is completely filled out
445 // The final pc we store in the loop is wrong and will be overwritten below
446 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
447
448 callee_parameters = array->element(index)->method()->size_of_parameters();
449 callee_locals = array->element(index)->method()->max_locals();
450 popframe_extra_args = 0;
451 }
452
453 // Compute whether the root vframe returns a float or double value.
454 BasicType return_type;
455 {
456 methodHandle method(thread, array->element(0)->method());
457 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
458 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
459 }
460
461 // Compute information for handling adapters and adjusting the frame size of the caller.
462 int caller_adjustment = 0;
463
464 // Compute the amount the oldest interpreter frame will have to adjust
465 // its caller's stack by. If the caller is a compiled frame then
466 // we pretend that the callee has no parameters so that the
467 // extension counts for the full amount of locals and not just
468 // locals-parms. This is because without a c2i adapter the parm
469 // area as created by the compiled frame will not be usable by
470 // the interpreter. (Depending on the calling convention there
471 // may not even be enough space).
472
473 // QQQ I'd rather see this pushed down into last_frame_adjust
474 // and have it take the sender (aka caller).
475
476 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
477 caller_adjustment = last_frame_adjust(0, callee_locals);
478 } else if (callee_locals > callee_parameters) {
479 // The caller frame may need extending to accommodate
480 // non-parameter locals of the first unpacked interpreted frame.
481 // Compute that adjustment.
482 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
483 }
484
485 // If the sender is deoptimized the we must retrieve the address of the handler
486 // since the frame will "magically" show the original pc before the deopt
487 // and we'd undo the deopt.
488
489 frame_pcs[0] = deopt_sender.raw_pc();
490
491 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
492
493 #ifdef INCLUDE_JVMCI
494 if (exceptionObject() != NULL) {
495 thread->set_exception_oop(exceptionObject());
496 exec_mode = Unpack_exception;
497 }
498 #endif
499
500 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
501 assert(thread->has_pending_exception(), "should have thrown OOME");
502 thread->set_exception_oop(thread->pending_exception());
503 thread->clear_pending_exception();
504 exec_mode = Unpack_exception;
505 }
506
507 #if INCLUDE_JVMCI
508 if (thread->frames_to_pop_failed_realloc() > 0) {
509 thread->set_pending_monitorenter(false);
510 }
511 #endif
512
513 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
514 caller_adjustment * BytesPerWord,
515 caller_was_method_handle ? 0 : callee_parameters,
516 number_of_frames,
517 frame_sizes,
518 frame_pcs,
519 return_type,
520 exec_mode);
521 // On some platforms, we need a way to pass some platform dependent
522 // information to the unpacking code so the skeletal frames come out
523 // correct (initial fp value, unextended sp, ...)
524 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
525
526 if (array->frames() > 1) {
527 if (VerifyStack && TraceDeoptimization) {
528 ttyLocker ttyl;
529 tty->print_cr("Deoptimizing method containing inlining");
530 }
531 }
532
533 array->set_unroll_block(info);
534 return info;
535 }
536
537 // Called to cleanup deoptimization data structures in normal case
538 // after unpacking to stack and when stack overflow error occurs
539 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
540 vframeArray *array) {
541
542 // Get array if coming from exception
543 if (array == NULL) {
544 array = thread->vframe_array_head();
545 }
546 thread->set_vframe_array_head(NULL);
547
548 // Free the previous UnrollBlock
549 vframeArray* old_array = thread->vframe_array_last();
550 thread->set_vframe_array_last(array);
551
552 if (old_array != NULL) {
553 UnrollBlock* old_info = old_array->unroll_block();
554 old_array->set_unroll_block(NULL);
555 delete old_info;
556 delete old_array;
557 }
558
559 // Deallocate any resource creating in this routine and any ResourceObjs allocated
560 // inside the vframeArray (StackValueCollections)
561
562 delete thread->deopt_mark();
563 thread->set_deopt_mark(NULL);
564 thread->set_deopt_compiled_method(NULL);
565
566
567 if (JvmtiExport::can_pop_frame()) {
568 #ifndef CC_INTERP
569 // Regardless of whether we entered this routine with the pending
570 // popframe condition bit set, we should always clear it now
571 thread->clear_popframe_condition();
572 #else
573 // C++ interpreter will clear has_pending_popframe when it enters
574 // with method_resume. For deopt_resume2 we clear it now.
575 if (thread->popframe_forcing_deopt_reexecution())
576 thread->clear_popframe_condition();
577 #endif /* CC_INTERP */
578 }
579
580 // unpack_frames() is called at the end of the deoptimization handler
581 // and (in C2) at the end of the uncommon trap handler. Note this fact
582 // so that an asynchronous stack walker can work again. This counter is
583 // incremented at the beginning of fetch_unroll_info() and (in C2) at
584 // the beginning of uncommon_trap().
585 thread->dec_in_deopt_handler();
586 }
587
588 // Moved from cpu directories because none of the cpus has callee save values.
589 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
590 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
591
592 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
593 // the days we had adapter frames. When we deoptimize a situation where a
594 // compiled caller calls a compiled caller will have registers it expects
595 // to survive the call to the callee. If we deoptimize the callee the only
596 // way we can restore these registers is to have the oldest interpreter
597 // frame that we create restore these values. That is what this routine
598 // will accomplish.
599
600 // At the moment we have modified c2 to not have any callee save registers
601 // so this problem does not exist and this routine is just a place holder.
602
603 assert(f->is_interpreted_frame(), "must be interpreted");
604 }
605
606 // Return BasicType of value being returned
607 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
608
609 // We are already active in the special DeoptResourceMark any ResourceObj's we
610 // allocate will be freed at the end of the routine.
611
612 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
613 // but makes the entry a little slower. There is however a little dance we have to
614 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
615 ResetNoHandleMark rnhm; // No-op in release/product versions
616 HandleMark hm;
617
618 frame stub_frame = thread->last_frame();
619
620 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
621 // must point to the vframeArray for the unpack frame.
622 vframeArray* array = thread->vframe_array_head();
623
624 #ifndef PRODUCT
625 if (TraceDeoptimization) {
626 ttyLocker ttyl;
627 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d",
628 p2i(thread), p2i(array), exec_mode);
629 }
630 #endif
631 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
632 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode);
633
634 UnrollBlock* info = array->unroll_block();
635
636 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
637 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
638
639 BasicType bt = info->return_type();
640
641 // If we have an exception pending, claim that the return type is an oop
642 // so the deopt_blob does not overwrite the exception_oop.
643
644 if (exec_mode == Unpack_exception)
645 bt = T_OBJECT;
646
647 // Cleanup thread deopt data
648 cleanup_deopt_info(thread, array);
649
650 #ifndef PRODUCT
651 if (VerifyStack) {
652 ResourceMark res_mark;
653 // Clear pending exception to not break verification code (restored afterwards)
654 PRESERVE_EXCEPTION_MARK;
655
656 thread->validate_frame_layout();
657
658 // Verify that the just-unpacked frames match the interpreter's
659 // notions of expression stack and locals
660 vframeArray* cur_array = thread->vframe_array_last();
661 RegisterMap rm(thread, false);
662 rm.set_include_argument_oops(false);
663 bool is_top_frame = true;
664 int callee_size_of_parameters = 0;
665 int callee_max_locals = 0;
666 for (int i = 0; i < cur_array->frames(); i++) {
667 vframeArrayElement* el = cur_array->element(i);
668 frame* iframe = el->iframe();
669 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
670
671 // Get the oop map for this bci
672 InterpreterOopMap mask;
673 int cur_invoke_parameter_size = 0;
674 bool try_next_mask = false;
675 int next_mask_expression_stack_size = -1;
676 int top_frame_expression_stack_adjustment = 0;
677 methodHandle mh(thread, iframe->interpreter_frame_method());
678 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
679 BytecodeStream str(mh);
680 str.set_start(iframe->interpreter_frame_bci());
681 int max_bci = mh->code_size();
682 // Get to the next bytecode if possible
683 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
684 // Check to see if we can grab the number of outgoing arguments
685 // at an uncommon trap for an invoke (where the compiler
686 // generates debug info before the invoke has executed)
687 Bytecodes::Code cur_code = str.next();
688 if (Bytecodes::is_invoke(cur_code)) {
689 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
690 cur_invoke_parameter_size = invoke.size_of_parameters();
691 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
692 callee_size_of_parameters++;
693 }
694 }
695 if (str.bci() < max_bci) {
696 Bytecodes::Code next_code = str.next();
697 if (next_code >= 0) {
698 // The interpreter oop map generator reports results before
699 // the current bytecode has executed except in the case of
700 // calls. It seems to be hard to tell whether the compiler
701 // has emitted debug information matching the "state before"
702 // a given bytecode or the state after, so we try both
703 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) {
704 // Get expression stack size for the next bytecode
705 if (Bytecodes::is_invoke(next_code)) {
706 Bytecode_invoke invoke(mh, str.bci());
707 next_mask_expression_stack_size = invoke.size_of_parameters();
708 } else {
709 InterpreterOopMap next_mask;
710 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
711 next_mask_expression_stack_size = next_mask.expression_stack_size();
712 }
713 // Need to subtract off the size of the result type of
714 // the bytecode because this is not described in the
715 // debug info but returned to the interpreter in the TOS
716 // caching register
717 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
718 if (bytecode_result_type != T_ILLEGAL) {
719 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
720 }
721 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
722 try_next_mask = true;
723 }
724 }
725 }
726
727 // Verify stack depth and oops in frame
728 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
729 if (!(
730 /* SPARC */
731 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
732 /* x86 */
733 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
734 (try_next_mask &&
735 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
736 top_frame_expression_stack_adjustment))) ||
737 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
738 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
739 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
740 )) {
741 ttyLocker ttyl;
742
743 // Print out some information that will help us debug the problem
744 tty->print_cr("Wrong number of expression stack elements during deoptimization");
745 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
746 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
747 iframe->interpreter_frame_expression_stack_size());
748 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
749 tty->print_cr(" try_next_mask = %d", try_next_mask);
750 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
751 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
752 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
753 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
754 tty->print_cr(" exec_mode = %d", exec_mode);
755 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
756 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
757 tty->print_cr(" Interpreted frames:");
758 for (int k = 0; k < cur_array->frames(); k++) {
759 vframeArrayElement* el = cur_array->element(k);
760 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
761 }
762 cur_array->print_on_2(tty);
763 guarantee(false, "wrong number of expression stack elements during deopt");
764 }
765 VerifyOopClosure verify;
766 iframe->oops_interpreted_do(&verify, &rm, false);
767 callee_size_of_parameters = mh->size_of_parameters();
768 callee_max_locals = mh->max_locals();
769 is_top_frame = false;
770 }
771 }
772 #endif /* !PRODUCT */
773
774
775 return bt;
776 JRT_END
777
778
779 int Deoptimization::deoptimize_dependents() {
780 Threads::deoptimized_wrt_marked_nmethods();
781 return 0;
782 }
783
784 Deoptimization::DeoptAction Deoptimization::_unloaded_action
785 = Deoptimization::Action_reinterpret;
786
787 #if COMPILER2_OR_JVMCI
788 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
789 Handle pending_exception(THREAD, thread->pending_exception());
790 const char* exception_file = thread->exception_file();
791 int exception_line = thread->exception_line();
792 thread->clear_pending_exception();
793
794 bool failures = false;
795
796 for (int i = 0; i < objects->length(); i++) {
797 assert(objects->at(i)->is_object(), "invalid debug information");
798 ObjectValue* sv = (ObjectValue*) objects->at(i);
799
800 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
801 oop obj = NULL;
802
803 if (k->is_instance_klass()) {
804 InstanceKlass* ik = InstanceKlass::cast(k);
805 obj = ik->allocate_instance(THREAD);
806 } else if (k->is_typeArray_klass()) {
807 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
808 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
809 int len = sv->field_size() / type2size[ak->element_type()];
810 obj = ak->allocate(len, THREAD);
811 } else if (k->is_objArray_klass()) {
812 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
813 obj = ak->allocate(sv->field_size(), THREAD);
814 }
815
816 if (obj == NULL) {
817 failures = true;
818 }
819
820 assert(sv->value().is_null(), "redundant reallocation");
821 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
822 CLEAR_PENDING_EXCEPTION;
823 sv->set_value(obj);
824 }
825
826 if (failures) {
827 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
828 } else if (pending_exception.not_null()) {
829 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
830 }
831
832 return failures;
833 }
834
835 // restore elements of an eliminated type array
836 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
837 int index = 0;
838 intptr_t val;
839
840 for (int i = 0; i < sv->field_size(); i++) {
841 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
842 switch(type) {
843 case T_LONG: case T_DOUBLE: {
844 assert(value->type() == T_INT, "Agreement.");
845 StackValue* low =
846 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
847 #ifdef _LP64
848 jlong res = (jlong)low->get_int();
849 #else
850 #ifdef SPARC
851 // For SPARC we have to swap high and low words.
852 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
853 #else
854 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
855 #endif //SPARC
856 #endif
857 obj->long_at_put(index, res);
858 break;
859 }
860
861 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
862 case T_INT: case T_FLOAT: { // 4 bytes.
863 assert(value->type() == T_INT, "Agreement.");
864 bool big_value = false;
865 if (i + 1 < sv->field_size() && type == T_INT) {
866 if (sv->field_at(i)->is_location()) {
867 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
868 if (type == Location::dbl || type == Location::lng) {
869 big_value = true;
870 }
871 } else if (sv->field_at(i)->is_constant_int()) {
872 ScopeValue* next_scope_field = sv->field_at(i + 1);
873 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
874 big_value = true;
875 }
876 }
877 }
878
879 if (big_value) {
880 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
881 #ifdef _LP64
882 jlong res = (jlong)low->get_int();
883 #else
884 #ifdef SPARC
885 // For SPARC we have to swap high and low words.
886 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
887 #else
888 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
889 #endif //SPARC
890 #endif
891 obj->int_at_put(index, (jint)*((jint*)&res));
892 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
893 } else {
894 val = value->get_int();
895 obj->int_at_put(index, (jint)*((jint*)&val));
896 }
897 break;
898 }
899
900 case T_SHORT:
901 assert(value->type() == T_INT, "Agreement.");
902 val = value->get_int();
903 obj->short_at_put(index, (jshort)*((jint*)&val));
904 break;
905
906 case T_CHAR:
907 assert(value->type() == T_INT, "Agreement.");
908 val = value->get_int();
909 obj->char_at_put(index, (jchar)*((jint*)&val));
910 break;
911
912 case T_BYTE:
913 assert(value->type() == T_INT, "Agreement.");
914 val = value->get_int();
915 obj->byte_at_put(index, (jbyte)*((jint*)&val));
916 break;
917
918 case T_BOOLEAN:
919 assert(value->type() == T_INT, "Agreement.");
920 val = value->get_int();
921 obj->bool_at_put(index, (jboolean)*((jint*)&val));
922 break;
923
924 default:
925 ShouldNotReachHere();
926 }
927 index++;
928 }
929 }
930
931
932 // restore fields of an eliminated object array
933 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
934 for (int i = 0; i < sv->field_size(); i++) {
935 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
936 assert(value->type() == T_OBJECT, "object element expected");
937 obj->obj_at_put(i, value->get_obj()());
938 }
939 }
940
941 class ReassignedField {
942 public:
943 int _offset;
944 BasicType _type;
945 public:
946 ReassignedField() {
947 _offset = 0;
948 _type = T_ILLEGAL;
949 }
950 };
951
952 int compare(ReassignedField* left, ReassignedField* right) {
953 return left->_offset - right->_offset;
954 }
955
956 // Restore fields of an eliminated instance object using the same field order
957 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
958 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
959 if (klass->superklass() != NULL) {
960 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal);
961 }
962
963 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
964 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
965 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
966 ReassignedField field;
967 field._offset = fs.offset();
968 field._type = FieldType::basic_type(fs.signature());
969 fields->append(field);
970 }
971 }
972 fields->sort(compare);
973 for (int i = 0; i < fields->length(); i++) {
974 intptr_t val;
975 ScopeValue* scope_field = sv->field_at(svIndex);
976 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
977 int offset = fields->at(i)._offset;
978 BasicType type = fields->at(i)._type;
979 switch (type) {
980 case T_OBJECT: case T_ARRAY:
981 assert(value->type() == T_OBJECT, "Agreement.");
982 obj->obj_field_put(offset, value->get_obj()());
983 break;
984
985 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
986 case T_INT: case T_FLOAT: { // 4 bytes.
987 assert(value->type() == T_INT, "Agreement.");
988 bool big_value = false;
989 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
990 if (scope_field->is_location()) {
991 Location::Type type = ((LocationValue*) scope_field)->location().type();
992 if (type == Location::dbl || type == Location::lng) {
993 big_value = true;
994 }
995 }
996 if (scope_field->is_constant_int()) {
997 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
998 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
999 big_value = true;
1000 }
1001 }
1002 }
1003
1004 if (big_value) {
1005 i++;
1006 assert(i < fields->length(), "second T_INT field needed");
1007 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1008 } else {
1009 val = value->get_int();
1010 obj->int_field_put(offset, (jint)*((jint*)&val));
1011 break;
1012 }
1013 }
1014 /* no break */
1015
1016 case T_LONG: case T_DOUBLE: {
1017 assert(value->type() == T_INT, "Agreement.");
1018 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1019 #ifdef _LP64
1020 jlong res = (jlong)low->get_int();
1021 #else
1022 #ifdef SPARC
1023 // For SPARC we have to swap high and low words.
1024 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1025 #else
1026 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1027 #endif //SPARC
1028 #endif
1029 obj->long_field_put(offset, res);
1030 break;
1031 }
1032
1033 case T_SHORT:
1034 assert(value->type() == T_INT, "Agreement.");
1035 val = value->get_int();
1036 obj->short_field_put(offset, (jshort)*((jint*)&val));
1037 break;
1038
1039 case T_CHAR:
1040 assert(value->type() == T_INT, "Agreement.");
1041 val = value->get_int();
1042 obj->char_field_put(offset, (jchar)*((jint*)&val));
1043 break;
1044
1045 case T_BYTE:
1046 assert(value->type() == T_INT, "Agreement.");
1047 val = value->get_int();
1048 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1049 break;
1050
1051 case T_BOOLEAN:
1052 assert(value->type() == T_INT, "Agreement.");
1053 val = value->get_int();
1054 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1055 break;
1056
1057 default:
1058 ShouldNotReachHere();
1059 }
1060 svIndex++;
1061 }
1062 return svIndex;
1063 }
1064
1065 // restore fields of all eliminated objects and arrays
1066 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1067 for (int i = 0; i < objects->length(); i++) {
1068 ObjectValue* sv = (ObjectValue*) objects->at(i);
1069 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1070 Handle obj = sv->value();
1071 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1072 if (PrintDeoptimizationDetails) {
1073 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1074 }
1075 if (obj.is_null()) {
1076 continue;
1077 }
1078
1079 if (k->is_instance_klass()) {
1080 InstanceKlass* ik = InstanceKlass::cast(k);
1081 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1082 } else if (k->is_typeArray_klass()) {
1083 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1084 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1085 } else if (k->is_objArray_klass()) {
1086 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1087 }
1088 }
1089 }
1090
1091
1092 // relock objects for which synchronization was eliminated
1093 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
1094 for (int i = 0; i < monitors->length(); i++) {
1095 MonitorInfo* mon_info = monitors->at(i);
1096 if (mon_info->eliminated()) {
1097 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1098 if (!mon_info->owner_is_scalar_replaced()) {
1099 Handle obj(thread, mon_info->owner());
1100 markOop mark = obj->mark();
1101 if (UseBiasedLocking && mark->has_bias_pattern()) {
1102 // New allocated objects may have the mark set to anonymously biased.
1103 // Also the deoptimized method may called methods with synchronization
1104 // where the thread-local object is bias locked to the current thread.
1105 assert(mark->is_biased_anonymously() ||
1106 mark->biased_locker() == thread, "should be locked to current thread");
1107 // Reset mark word to unbiased prototype.
1108 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
1109 obj->set_mark(unbiased_prototype);
1110 }
1111 BasicLock* lock = mon_info->lock();
1112 ObjectSynchronizer::slow_enter(obj, lock, thread);
1113 assert(mon_info->owner()->is_locked(), "object must be locked now");
1114 }
1115 }
1116 }
1117 }
1118
1119
1120 #ifndef PRODUCT
1121 // print information about reallocated objects
1122 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1123 fieldDescriptor fd;
1124
1125 for (int i = 0; i < objects->length(); i++) {
1126 ObjectValue* sv = (ObjectValue*) objects->at(i);
1127 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1128 Handle obj = sv->value();
1129
1130 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1131 k->print_value();
1132 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1133 if (obj.is_null()) {
1134 tty->print(" allocation failed");
1135 } else {
1136 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1137 }
1138 tty->cr();
1139
1140 if (Verbose && !obj.is_null()) {
1141 k->oop_print_on(obj(), tty);
1142 }
1143 }
1144 }
1145 #endif
1146 #endif // COMPILER2_OR_JVMCI
1147
1148 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1149 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1150
1151 #ifndef PRODUCT
1152 if (PrintDeoptimizationDetails) {
1153 ttyLocker ttyl;
1154 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1155 fr.print_on(tty);
1156 tty->print_cr(" Virtual frames (innermost first):");
1157 for (int index = 0; index < chunk->length(); index++) {
1158 compiledVFrame* vf = chunk->at(index);
1159 tty->print(" %2d - ", index);
1160 vf->print_value();
1161 int bci = chunk->at(index)->raw_bci();
1162 const char* code_name;
1163 if (bci == SynchronizationEntryBCI) {
1164 code_name = "sync entry";
1165 } else {
1166 Bytecodes::Code code = vf->method()->code_at(bci);
1167 code_name = Bytecodes::name(code);
1168 }
1169 tty->print(" - %s", code_name);
1170 tty->print_cr(" @ bci %d ", bci);
1171 if (Verbose) {
1172 vf->print();
1173 tty->cr();
1174 }
1175 }
1176 }
1177 #endif
1178
1179 // Register map for next frame (used for stack crawl). We capture
1180 // the state of the deopt'ing frame's caller. Thus if we need to
1181 // stuff a C2I adapter we can properly fill in the callee-save
1182 // register locations.
1183 frame caller = fr.sender(reg_map);
1184 int frame_size = caller.sp() - fr.sp();
1185
1186 frame sender = caller;
1187
1188 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1189 // the vframeArray containing the unpacking information is allocated in the C heap.
1190 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1191 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1192
1193 // Compare the vframeArray to the collected vframes
1194 assert(array->structural_compare(thread, chunk), "just checking");
1195
1196 #ifndef PRODUCT
1197 if (PrintDeoptimizationDetails) {
1198 ttyLocker ttyl;
1199 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1200 }
1201 #endif // PRODUCT
1202
1203 return array;
1204 }
1205
1206 #if COMPILER2_OR_JVMCI
1207 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1208 // Reallocation of some scalar replaced objects failed. Record
1209 // that we need to pop all the interpreter frames for the
1210 // deoptimized compiled frame.
1211 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1212 thread->set_frames_to_pop_failed_realloc(array->frames());
1213 // Unlock all monitors here otherwise the interpreter will see a
1214 // mix of locked and unlocked monitors (because of failed
1215 // reallocations of synchronized objects) and be confused.
1216 for (int i = 0; i < array->frames(); i++) {
1217 MonitorChunk* monitors = array->element(i)->monitors();
1218 if (monitors != NULL) {
1219 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1220 BasicObjectLock* src = monitors->at(j);
1221 if (src->obj() != NULL) {
1222 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread);
1223 }
1224 }
1225 array->element(i)->free_monitors(thread);
1226 #ifdef ASSERT
1227 array->element(i)->set_removed_monitors();
1228 #endif
1229 }
1230 }
1231 }
1232 #endif
1233
1234 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1235 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1236 Thread* thread = Thread::current();
1237 for (int i = 0; i < monitors->length(); i++) {
1238 MonitorInfo* mon_info = monitors->at(i);
1239 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1240 objects_to_revoke->append(Handle(thread, mon_info->owner()));
1241 }
1242 }
1243 }
1244
1245
1246 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1247 if (!UseBiasedLocking) {
1248 return;
1249 }
1250
1251 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1252
1253 // Unfortunately we don't have a RegisterMap available in most of
1254 // the places we want to call this routine so we need to walk the
1255 // stack again to update the register map.
1256 if (map == NULL || !map->update_map()) {
1257 StackFrameStream sfs(thread, true);
1258 bool found = false;
1259 while (!found && !sfs.is_done()) {
1260 frame* cur = sfs.current();
1261 sfs.next();
1262 found = cur->id() == fr.id();
1263 }
1264 assert(found, "frame to be deoptimized not found on target thread's stack");
1265 map = sfs.register_map();
1266 }
1267
1268 vframe* vf = vframe::new_vframe(&fr, map, thread);
1269 compiledVFrame* cvf = compiledVFrame::cast(vf);
1270 // Revoke monitors' biases in all scopes
1271 while (!cvf->is_top()) {
1272 collect_monitors(cvf, objects_to_revoke);
1273 cvf = compiledVFrame::cast(cvf->sender());
1274 }
1275 collect_monitors(cvf, objects_to_revoke);
1276
1277 if (SafepointSynchronize::is_at_safepoint()) {
1278 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1279 } else {
1280 BiasedLocking::revoke(objects_to_revoke);
1281 }
1282 }
1283
1284
1285 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1286 if (!UseBiasedLocking) {
1287 return;
1288 }
1289
1290 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1291 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1292 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
1293 if (jt->has_last_Java_frame()) {
1294 StackFrameStream sfs(jt, true);
1295 while (!sfs.is_done()) {
1296 frame* cur = sfs.current();
1297 if (cb->contains(cur->pc())) {
1298 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1299 compiledVFrame* cvf = compiledVFrame::cast(vf);
1300 // Revoke monitors' biases in all scopes
1301 while (!cvf->is_top()) {
1302 collect_monitors(cvf, objects_to_revoke);
1303 cvf = compiledVFrame::cast(cvf->sender());
1304 }
1305 collect_monitors(cvf, objects_to_revoke);
1306 }
1307 sfs.next();
1308 }
1309 }
1310 }
1311 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1312 }
1313
1314
1315 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1316 assert(fr.can_be_deoptimized(), "checking frame type");
1317
1318 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1319
1320 if (LogCompilation && xtty != NULL) {
1321 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1322 assert(cm != NULL, "only compiled methods can deopt");
1323
1324 ttyLocker ttyl;
1325 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1326 cm->log_identity(xtty);
1327 xtty->end_head();
1328 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1329 xtty->begin_elem("jvms bci='%d'", sd->bci());
1330 xtty->method(sd->method());
1331 xtty->end_elem();
1332 if (sd->is_top()) break;
1333 }
1334 xtty->tail("deoptimized");
1335 }
1336
1337 // Patch the compiled method so that when execution returns to it we will
1338 // deopt the execution state and return to the interpreter.
1339 fr.deoptimize(thread);
1340 }
1341
1342 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1343 deoptimize(thread, fr, map, Reason_constraint);
1344 }
1345
1346 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) {
1347 // Deoptimize only if the frame comes from compile code.
1348 // Do not deoptimize the frame which is already patched
1349 // during the execution of the loops below.
1350 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1351 return;
1352 }
1353 ResourceMark rm;
1354 DeoptimizationMarker dm;
1355 if (UseBiasedLocking) {
1356 revoke_biases_of_monitors(thread, fr, map);
1357 }
1358 deoptimize_single_frame(thread, fr, reason);
1359
1360 }
1361
1362 #if INCLUDE_JVMCI
1363 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) {
1364 // there is no exception handler for this pc => deoptimize
1365 cm->make_not_entrant();
1366
1367 // Use Deoptimization::deoptimize for all of its side-effects:
1368 // revoking biases of monitors, gathering traps statistics, logging...
1369 // it also patches the return pc but we do not care about that
1370 // since we return a continuation to the deopt_blob below.
1371 JavaThread* thread = JavaThread::current();
1372 RegisterMap reg_map(thread, UseBiasedLocking);
1373 frame runtime_frame = thread->last_frame();
1374 frame caller_frame = runtime_frame.sender(®_map);
1375 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method");
1376 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler);
1377
1378 MethodData* trap_mdo = get_method_data(thread, cm->method(), true);
1379 if (trap_mdo != NULL) {
1380 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1381 }
1382
1383 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1384 }
1385 #endif
1386
1387 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1388 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1389 "can only deoptimize other thread at a safepoint");
1390 // Compute frame and register map based on thread and sp.
1391 RegisterMap reg_map(thread, UseBiasedLocking);
1392 frame fr = thread->last_frame();
1393 while (fr.id() != id) {
1394 fr = fr.sender(®_map);
1395 }
1396 deoptimize(thread, fr, ®_map, reason);
1397 }
1398
1399
1400 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1401 if (thread == Thread::current()) {
1402 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1403 } else {
1404 VM_DeoptimizeFrame deopt(thread, id, reason);
1405 VMThread::execute(&deopt);
1406 }
1407 }
1408
1409 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1410 deoptimize_frame(thread, id, Reason_constraint);
1411 }
1412
1413 // JVMTI PopFrame support
1414 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1415 {
1416 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1417 }
1418 JRT_END
1419
1420 MethodData*
1421 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1422 bool create_if_missing) {
1423 Thread* THREAD = thread;
1424 MethodData* mdo = m()->method_data();
1425 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1426 // Build an MDO. Ignore errors like OutOfMemory;
1427 // that simply means we won't have an MDO to update.
1428 Method::build_interpreter_method_data(m, THREAD);
1429 if (HAS_PENDING_EXCEPTION) {
1430 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1431 CLEAR_PENDING_EXCEPTION;
1432 }
1433 mdo = m()->method_data();
1434 }
1435 return mdo;
1436 }
1437
1438 #if COMPILER2_OR_JVMCI
1439 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1440 // in case of an unresolved klass entry, load the class.
1441 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1442 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1443 return;
1444 }
1445
1446 if (!constant_pool->tag_at(index).is_symbol()) return;
1447
1448 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1449 Symbol* symbol = constant_pool->symbol_at(index);
1450
1451 // class name?
1452 if (symbol->byte_at(0) != '(') {
1453 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1454 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1455 return;
1456 }
1457
1458 // then it must be a signature!
1459 ResourceMark rm(THREAD);
1460 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1461 if (ss.is_object()) {
1462 Symbol* class_name = ss.as_symbol(CHECK);
1463 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1464 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1465 }
1466 }
1467 }
1468
1469
1470 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1471 EXCEPTION_MARK;
1472 load_class_by_index(constant_pool, index, THREAD);
1473 if (HAS_PENDING_EXCEPTION) {
1474 // Exception happened during classloading. We ignore the exception here, since it
1475 // is going to be rethrown since the current activation is going to be deoptimized and
1476 // the interpreter will re-execute the bytecode.
1477 CLEAR_PENDING_EXCEPTION;
1478 // Class loading called java code which may have caused a stack
1479 // overflow. If the exception was thrown right before the return
1480 // to the runtime the stack is no longer guarded. Reguard the
1481 // stack otherwise if we return to the uncommon trap blob and the
1482 // stack bang causes a stack overflow we crash.
1483 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1484 JavaThread* thread = (JavaThread*)THREAD;
1485 bool guard_pages_enabled = thread->stack_guards_enabled();
1486 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1487 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1488 }
1489 }
1490
1491 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1492 HandleMark hm;
1493
1494 // uncommon_trap() is called at the beginning of the uncommon trap
1495 // handler. Note this fact before we start generating temporary frames
1496 // that can confuse an asynchronous stack walker. This counter is
1497 // decremented at the end of unpack_frames().
1498 thread->inc_in_deopt_handler();
1499
1500 // We need to update the map if we have biased locking.
1501 #if INCLUDE_JVMCI
1502 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1503 RegisterMap reg_map(thread, true);
1504 #else
1505 RegisterMap reg_map(thread, UseBiasedLocking);
1506 #endif
1507 frame stub_frame = thread->last_frame();
1508 frame fr = stub_frame.sender(®_map);
1509 // Make sure the calling nmethod is not getting deoptimized and removed
1510 // before we are done with it.
1511 nmethodLocker nl(fr.pc());
1512
1513 // Log a message
1514 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1515 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1516
1517 {
1518 ResourceMark rm;
1519
1520 // Revoke biases of any monitors in the frame to ensure we can migrate them
1521 revoke_biases_of_monitors(thread, fr, ®_map);
1522
1523 DeoptReason reason = trap_request_reason(trap_request);
1524 DeoptAction action = trap_request_action(trap_request);
1525 #if INCLUDE_JVMCI
1526 int debug_id = trap_request_debug_id(trap_request);
1527 #endif
1528 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1529
1530 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1531 compiledVFrame* cvf = compiledVFrame::cast(vf);
1532
1533 CompiledMethod* nm = cvf->code();
1534
1535 ScopeDesc* trap_scope = cvf->scope();
1536
1537 if (TraceDeoptimization) {
1538 ttyLocker ttyl;
1539 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()
1540 #if INCLUDE_JVMCI
1541 , debug_id
1542 #endif
1543 );
1544 }
1545
1546 methodHandle trap_method = trap_scope->method();
1547 int trap_bci = trap_scope->bci();
1548 #if INCLUDE_JVMCI
1549 oop speculation = thread->pending_failed_speculation();
1550 if (nm->is_compiled_by_jvmci()) {
1551 if (speculation != NULL) {
1552 oop speculation_log = nm->as_nmethod()->speculation_log();
1553 if (speculation_log != NULL) {
1554 if (TraceDeoptimization || TraceUncollectedSpeculations) {
1555 if (HotSpotSpeculationLog::lastFailed(speculation_log) != NULL) {
1556 tty->print_cr("A speculation that was not collected by the compiler is being overwritten");
1557 }
1558 }
1559 if (TraceDeoptimization) {
1560 tty->print_cr("Saving speculation to speculation log");
1561 }
1562 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation);
1563 } else {
1564 if (TraceDeoptimization) {
1565 tty->print_cr("Speculation present but no speculation log");
1566 }
1567 }
1568 thread->set_pending_failed_speculation(NULL);
1569 } else {
1570 if (TraceDeoptimization) {
1571 tty->print_cr("No speculation");
1572 }
1573 }
1574 } else {
1575 assert(speculation == NULL, "There should not be a speculation for method compiled by non-JVMCI compilers");
1576 }
1577
1578 if (trap_bci == SynchronizationEntryBCI) {
1579 trap_bci = 0;
1580 thread->set_pending_monitorenter(true);
1581 }
1582
1583 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1584 thread->set_pending_transfer_to_interpreter(true);
1585 }
1586 #endif
1587
1588 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1589 // Record this event in the histogram.
1590 gather_statistics(reason, action, trap_bc);
1591
1592 // Ensure that we can record deopt. history:
1593 // Need MDO to record RTM code generation state.
1594 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1595
1596 methodHandle profiled_method;
1597 #if INCLUDE_JVMCI
1598 if (nm->is_compiled_by_jvmci()) {
1599 profiled_method = nm->method();
1600 } else {
1601 profiled_method = trap_method;
1602 }
1603 #else
1604 profiled_method = trap_method;
1605 #endif
1606
1607 MethodData* trap_mdo =
1608 get_method_data(thread, profiled_method, create_if_missing);
1609
1610 // Log a message
1611 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
1612 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1613 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
1614
1615 // Print a bunch of diagnostics, if requested.
1616 if (TraceDeoptimization || LogCompilation) {
1617 ResourceMark rm;
1618 ttyLocker ttyl;
1619 char buf[100];
1620 if (xtty != NULL) {
1621 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1622 os::current_thread_id(),
1623 format_trap_request(buf, sizeof(buf), trap_request));
1624 nm->log_identity(xtty);
1625 }
1626 Symbol* class_name = NULL;
1627 bool unresolved = false;
1628 if (unloaded_class_index >= 0) {
1629 constantPoolHandle constants (THREAD, trap_method->constants());
1630 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1631 class_name = constants->klass_name_at(unloaded_class_index);
1632 unresolved = true;
1633 if (xtty != NULL)
1634 xtty->print(" unresolved='1'");
1635 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1636 class_name = constants->symbol_at(unloaded_class_index);
1637 }
1638 if (xtty != NULL)
1639 xtty->name(class_name);
1640 }
1641 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
1642 // Dump the relevant MDO state.
1643 // This is the deopt count for the current reason, any previous
1644 // reasons or recompiles seen at this point.
1645 int dcnt = trap_mdo->trap_count(reason);
1646 if (dcnt != 0)
1647 xtty->print(" count='%d'", dcnt);
1648 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1649 int dos = (pdata == NULL)? 0: pdata->trap_state();
1650 if (dos != 0) {
1651 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1652 if (trap_state_is_recompiled(dos)) {
1653 int recnt2 = trap_mdo->overflow_recompile_count();
1654 if (recnt2 != 0)
1655 xtty->print(" recompiles2='%d'", recnt2);
1656 }
1657 }
1658 }
1659 if (xtty != NULL) {
1660 xtty->stamp();
1661 xtty->end_head();
1662 }
1663 if (TraceDeoptimization) { // make noise on the tty
1664 tty->print("Uncommon trap occurred in");
1665 nm->method()->print_short_name(tty);
1666 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
1667 #if INCLUDE_JVMCI
1668 if (nm->is_nmethod()) {
1669 char* installed_code_name = nm->as_nmethod()->jvmci_installed_code_name(buf, sizeof(buf));
1670 if (installed_code_name != NULL) {
1671 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name);
1672 }
1673 }
1674 #endif
1675 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
1676 p2i(fr.pc()),
1677 os::current_thread_id(),
1678 trap_reason_name(reason),
1679 trap_action_name(action),
1680 unloaded_class_index
1681 #if INCLUDE_JVMCI
1682 , debug_id
1683 #endif
1684 );
1685 if (class_name != NULL) {
1686 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1687 class_name->print_symbol_on(tty);
1688 }
1689 tty->cr();
1690 }
1691 if (xtty != NULL) {
1692 // Log the precise location of the trap.
1693 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1694 xtty->begin_elem("jvms bci='%d'", sd->bci());
1695 xtty->method(sd->method());
1696 xtty->end_elem();
1697 if (sd->is_top()) break;
1698 }
1699 xtty->tail("uncommon_trap");
1700 }
1701 }
1702 // (End diagnostic printout.)
1703
1704 // Load class if necessary
1705 if (unloaded_class_index >= 0) {
1706 constantPoolHandle constants(THREAD, trap_method->constants());
1707 load_class_by_index(constants, unloaded_class_index);
1708 }
1709
1710 // Flush the nmethod if necessary and desirable.
1711 //
1712 // We need to avoid situations where we are re-flushing the nmethod
1713 // because of a hot deoptimization site. Repeated flushes at the same
1714 // point need to be detected by the compiler and avoided. If the compiler
1715 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1716 // module must take measures to avoid an infinite cycle of recompilation
1717 // and deoptimization. There are several such measures:
1718 //
1719 // 1. If a recompilation is ordered a second time at some site X
1720 // and for the same reason R, the action is adjusted to 'reinterpret',
1721 // to give the interpreter time to exercise the method more thoroughly.
1722 // If this happens, the method's overflow_recompile_count is incremented.
1723 //
1724 // 2. If the compiler fails to reduce the deoptimization rate, then
1725 // the method's overflow_recompile_count will begin to exceed the set
1726 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1727 // is adjusted to 'make_not_compilable', and the method is abandoned
1728 // to the interpreter. This is a performance hit for hot methods,
1729 // but is better than a disastrous infinite cycle of recompilations.
1730 // (Actually, only the method containing the site X is abandoned.)
1731 //
1732 // 3. In parallel with the previous measures, if the total number of
1733 // recompilations of a method exceeds the much larger set limit
1734 // PerMethodRecompilationCutoff, the method is abandoned.
1735 // This should only happen if the method is very large and has
1736 // many "lukewarm" deoptimizations. The code which enforces this
1737 // limit is elsewhere (class nmethod, class Method).
1738 //
1739 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1740 // to recompile at each bytecode independently of the per-BCI cutoff.
1741 //
1742 // The decision to update code is up to the compiler, and is encoded
1743 // in the Action_xxx code. If the compiler requests Action_none
1744 // no trap state is changed, no compiled code is changed, and the
1745 // computation suffers along in the interpreter.
1746 //
1747 // The other action codes specify various tactics for decompilation
1748 // and recompilation. Action_maybe_recompile is the loosest, and
1749 // allows the compiled code to stay around until enough traps are seen,
1750 // and until the compiler gets around to recompiling the trapping method.
1751 //
1752 // The other actions cause immediate removal of the present code.
1753
1754 // Traps caused by injected profile shouldn't pollute trap counts.
1755 bool injected_profile_trap = trap_method->has_injected_profile() &&
1756 (reason == Reason_intrinsic || reason == Reason_unreached);
1757
1758 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
1759 bool make_not_entrant = false;
1760 bool make_not_compilable = false;
1761 bool reprofile = false;
1762 switch (action) {
1763 case Action_none:
1764 // Keep the old code.
1765 update_trap_state = false;
1766 break;
1767 case Action_maybe_recompile:
1768 // Do not need to invalidate the present code, but we can
1769 // initiate another
1770 // Start compiler without (necessarily) invalidating the nmethod.
1771 // The system will tolerate the old code, but new code should be
1772 // generated when possible.
1773 break;
1774 case Action_reinterpret:
1775 // Go back into the interpreter for a while, and then consider
1776 // recompiling form scratch.
1777 make_not_entrant = true;
1778 // Reset invocation counter for outer most method.
1779 // This will allow the interpreter to exercise the bytecodes
1780 // for a while before recompiling.
1781 // By contrast, Action_make_not_entrant is immediate.
1782 //
1783 // Note that the compiler will track null_check, null_assert,
1784 // range_check, and class_check events and log them as if they
1785 // had been traps taken from compiled code. This will update
1786 // the MDO trap history so that the next compilation will
1787 // properly detect hot trap sites.
1788 reprofile = true;
1789 break;
1790 case Action_make_not_entrant:
1791 // Request immediate recompilation, and get rid of the old code.
1792 // Make them not entrant, so next time they are called they get
1793 // recompiled. Unloaded classes are loaded now so recompile before next
1794 // time they are called. Same for uninitialized. The interpreter will
1795 // link the missing class, if any.
1796 make_not_entrant = true;
1797 break;
1798 case Action_make_not_compilable:
1799 // Give up on compiling this method at all.
1800 make_not_entrant = true;
1801 make_not_compilable = true;
1802 break;
1803 default:
1804 ShouldNotReachHere();
1805 }
1806
1807 // Setting +ProfileTraps fixes the following, on all platforms:
1808 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1809 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1810 // recompile relies on a MethodData* to record heroic opt failures.
1811
1812 // Whether the interpreter is producing MDO data or not, we also need
1813 // to use the MDO to detect hot deoptimization points and control
1814 // aggressive optimization.
1815 bool inc_recompile_count = false;
1816 ProfileData* pdata = NULL;
1817 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) {
1818 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
1819 uint this_trap_count = 0;
1820 bool maybe_prior_trap = false;
1821 bool maybe_prior_recompile = false;
1822 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
1823 #if INCLUDE_JVMCI
1824 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
1825 #endif
1826 nm->method(),
1827 //outputs:
1828 this_trap_count,
1829 maybe_prior_trap,
1830 maybe_prior_recompile);
1831 // Because the interpreter also counts null, div0, range, and class
1832 // checks, these traps from compiled code are double-counted.
1833 // This is harmless; it just means that the PerXTrapLimit values
1834 // are in effect a little smaller than they look.
1835
1836 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1837 if (per_bc_reason != Reason_none) {
1838 // Now take action based on the partially known per-BCI history.
1839 if (maybe_prior_trap
1840 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1841 // If there are too many traps at this BCI, force a recompile.
1842 // This will allow the compiler to see the limit overflow, and
1843 // take corrective action, if possible. The compiler generally
1844 // does not use the exact PerBytecodeTrapLimit value, but instead
1845 // changes its tactics if it sees any traps at all. This provides
1846 // a little hysteresis, delaying a recompile until a trap happens
1847 // several times.
1848 //
1849 // Actually, since there is only one bit of counter per BCI,
1850 // the possible per-BCI counts are {0,1,(per-method count)}.
1851 // This produces accurate results if in fact there is only
1852 // one hot trap site, but begins to get fuzzy if there are
1853 // many sites. For example, if there are ten sites each
1854 // trapping two or more times, they each get the blame for
1855 // all of their traps.
1856 make_not_entrant = true;
1857 }
1858
1859 // Detect repeated recompilation at the same BCI, and enforce a limit.
1860 if (make_not_entrant && maybe_prior_recompile) {
1861 // More than one recompile at this point.
1862 inc_recompile_count = maybe_prior_trap;
1863 }
1864 } else {
1865 // For reasons which are not recorded per-bytecode, we simply
1866 // force recompiles unconditionally.
1867 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1868 make_not_entrant = true;
1869 }
1870
1871 // Go back to the compiler if there are too many traps in this method.
1872 if (this_trap_count >= per_method_trap_limit(reason)) {
1873 // If there are too many traps in this method, force a recompile.
1874 // This will allow the compiler to see the limit overflow, and
1875 // take corrective action, if possible.
1876 // (This condition is an unlikely backstop only, because the
1877 // PerBytecodeTrapLimit is more likely to take effect first,
1878 // if it is applicable.)
1879 make_not_entrant = true;
1880 }
1881
1882 // Here's more hysteresis: If there has been a recompile at
1883 // this trap point already, run the method in the interpreter
1884 // for a while to exercise it more thoroughly.
1885 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1886 reprofile = true;
1887 }
1888 }
1889
1890 // Take requested actions on the method:
1891
1892 // Recompile
1893 if (make_not_entrant) {
1894 if (!nm->make_not_entrant()) {
1895 return; // the call did not change nmethod's state
1896 }
1897
1898 if (pdata != NULL) {
1899 // Record the recompilation event, if any.
1900 int tstate0 = pdata->trap_state();
1901 int tstate1 = trap_state_set_recompiled(tstate0, true);
1902 if (tstate1 != tstate0)
1903 pdata->set_trap_state(tstate1);
1904 }
1905
1906 #if INCLUDE_RTM_OPT
1907 // Restart collecting RTM locking abort statistic if the method
1908 // is recompiled for a reason other than RTM state change.
1909 // Assume that in new recompiled code the statistic could be different,
1910 // for example, due to different inlining.
1911 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
1912 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
1913 trap_mdo->atomic_set_rtm_state(ProfileRTM);
1914 }
1915 #endif
1916 // For code aging we count traps separately here, using make_not_entrant()
1917 // as a guard against simultaneous deopts in multiple threads.
1918 if (reason == Reason_tenured && trap_mdo != NULL) {
1919 trap_mdo->inc_tenure_traps();
1920 }
1921 }
1922
1923 if (inc_recompile_count) {
1924 trap_mdo->inc_overflow_recompile_count();
1925 if ((uint)trap_mdo->overflow_recompile_count() >
1926 (uint)PerBytecodeRecompilationCutoff) {
1927 // Give up on the method containing the bad BCI.
1928 if (trap_method() == nm->method()) {
1929 make_not_compilable = true;
1930 } else {
1931 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
1932 // But give grace to the enclosing nm->method().
1933 }
1934 }
1935 }
1936
1937 // Reprofile
1938 if (reprofile) {
1939 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1940 }
1941
1942 // Give up compiling
1943 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1944 assert(make_not_entrant, "consistent");
1945 nm->method()->set_not_compilable(CompLevel_full_optimization);
1946 }
1947
1948 } // Free marked resources
1949
1950 }
1951 JRT_END
1952
1953 ProfileData*
1954 Deoptimization::query_update_method_data(MethodData* trap_mdo,
1955 int trap_bci,
1956 Deoptimization::DeoptReason reason,
1957 bool update_total_trap_count,
1958 #if INCLUDE_JVMCI
1959 bool is_osr,
1960 #endif
1961 Method* compiled_method,
1962 //outputs:
1963 uint& ret_this_trap_count,
1964 bool& ret_maybe_prior_trap,
1965 bool& ret_maybe_prior_recompile) {
1966 bool maybe_prior_trap = false;
1967 bool maybe_prior_recompile = false;
1968 uint this_trap_count = 0;
1969 if (update_total_trap_count) {
1970 uint idx = reason;
1971 #if INCLUDE_JVMCI
1972 if (is_osr) {
1973 idx += Reason_LIMIT;
1974 }
1975 #endif
1976 uint prior_trap_count = trap_mdo->trap_count(idx);
1977 this_trap_count = trap_mdo->inc_trap_count(idx);
1978
1979 // If the runtime cannot find a place to store trap history,
1980 // it is estimated based on the general condition of the method.
1981 // If the method has ever been recompiled, or has ever incurred
1982 // a trap with the present reason , then this BCI is assumed
1983 // (pessimistically) to be the culprit.
1984 maybe_prior_trap = (prior_trap_count != 0);
1985 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1986 }
1987 ProfileData* pdata = NULL;
1988
1989
1990 // For reasons which are recorded per bytecode, we check per-BCI data.
1991 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1992 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
1993 if (per_bc_reason != Reason_none) {
1994 // Find the profile data for this BCI. If there isn't one,
1995 // try to allocate one from the MDO's set of spares.
1996 // This will let us detect a repeated trap at this point.
1997 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
1998
1999 if (pdata != NULL) {
2000 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2001 if (LogCompilation && xtty != NULL) {
2002 ttyLocker ttyl;
2003 // no more room for speculative traps in this MDO
2004 xtty->elem("speculative_traps_oom");
2005 }
2006 }
2007 // Query the trap state of this profile datum.
2008 int tstate0 = pdata->trap_state();
2009 if (!trap_state_has_reason(tstate0, per_bc_reason))
2010 maybe_prior_trap = false;
2011 if (!trap_state_is_recompiled(tstate0))
2012 maybe_prior_recompile = false;
2013
2014 // Update the trap state of this profile datum.
2015 int tstate1 = tstate0;
2016 // Record the reason.
2017 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2018 // Store the updated state on the MDO, for next time.
2019 if (tstate1 != tstate0)
2020 pdata->set_trap_state(tstate1);
2021 } else {
2022 if (LogCompilation && xtty != NULL) {
2023 ttyLocker ttyl;
2024 // Missing MDP? Leave a small complaint in the log.
2025 xtty->elem("missing_mdp bci='%d'", trap_bci);
2026 }
2027 }
2028 }
2029
2030 // Return results:
2031 ret_this_trap_count = this_trap_count;
2032 ret_maybe_prior_trap = maybe_prior_trap;
2033 ret_maybe_prior_recompile = maybe_prior_recompile;
2034 return pdata;
2035 }
2036
2037 void
2038 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2039 ResourceMark rm;
2040 // Ignored outputs:
2041 uint ignore_this_trap_count;
2042 bool ignore_maybe_prior_trap;
2043 bool ignore_maybe_prior_recompile;
2044 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2045 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2046 bool update_total_counts = JVMCI_ONLY(false) NOT_JVMCI(true);
2047 query_update_method_data(trap_mdo, trap_bci,
2048 (DeoptReason)reason,
2049 update_total_counts,
2050 #if INCLUDE_JVMCI
2051 false,
2052 #endif
2053 NULL,
2054 ignore_this_trap_count,
2055 ignore_maybe_prior_trap,
2056 ignore_maybe_prior_recompile);
2057 }
2058
2059 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2060 if (TraceDeoptimization) {
2061 tty->print("Uncommon trap ");
2062 }
2063 // Still in Java no safepoints
2064 {
2065 // This enters VM and may safepoint
2066 uncommon_trap_inner(thread, trap_request);
2067 }
2068 return fetch_unroll_info_helper(thread, exec_mode);
2069 }
2070
2071 // Local derived constants.
2072 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2073 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
2074 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2075
2076 //---------------------------trap_state_reason---------------------------------
2077 Deoptimization::DeoptReason
2078 Deoptimization::trap_state_reason(int trap_state) {
2079 // This assert provides the link between the width of DataLayout::trap_bits
2080 // and the encoding of "recorded" reasons. It ensures there are enough
2081 // bits to store all needed reasons in the per-BCI MDO profile.
2082 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2083 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2084 trap_state -= recompile_bit;
2085 if (trap_state == DS_REASON_MASK) {
2086 return Reason_many;
2087 } else {
2088 assert((int)Reason_none == 0, "state=0 => Reason_none");
2089 return (DeoptReason)trap_state;
2090 }
2091 }
2092 //-------------------------trap_state_has_reason-------------------------------
2093 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2094 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2095 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2096 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2097 trap_state -= recompile_bit;
2098 if (trap_state == DS_REASON_MASK) {
2099 return -1; // true, unspecifically (bottom of state lattice)
2100 } else if (trap_state == reason) {
2101 return 1; // true, definitely
2102 } else if (trap_state == 0) {
2103 return 0; // false, definitely (top of state lattice)
2104 } else {
2105 return 0; // false, definitely
2106 }
2107 }
2108 //-------------------------trap_state_add_reason-------------------------------
2109 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2110 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2111 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2112 trap_state -= recompile_bit;
2113 if (trap_state == DS_REASON_MASK) {
2114 return trap_state + recompile_bit; // already at state lattice bottom
2115 } else if (trap_state == reason) {
2116 return trap_state + recompile_bit; // the condition is already true
2117 } else if (trap_state == 0) {
2118 return reason + recompile_bit; // no condition has yet been true
2119 } else {
2120 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2121 }
2122 }
2123 //-----------------------trap_state_is_recompiled------------------------------
2124 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2125 return (trap_state & DS_RECOMPILE_BIT) != 0;
2126 }
2127 //-----------------------trap_state_set_recompiled-----------------------------
2128 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2129 if (z) return trap_state | DS_RECOMPILE_BIT;
2130 else return trap_state & ~DS_RECOMPILE_BIT;
2131 }
2132 //---------------------------format_trap_state---------------------------------
2133 // This is used for debugging and diagnostics, including LogFile output.
2134 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2135 int trap_state) {
2136 assert(buflen > 0, "sanity");
2137 DeoptReason reason = trap_state_reason(trap_state);
2138 bool recomp_flag = trap_state_is_recompiled(trap_state);
2139 // Re-encode the state from its decoded components.
2140 int decoded_state = 0;
2141 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2142 decoded_state = trap_state_add_reason(decoded_state, reason);
2143 if (recomp_flag)
2144 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2145 // If the state re-encodes properly, format it symbolically.
2146 // Because this routine is used for debugging and diagnostics,
2147 // be robust even if the state is a strange value.
2148 size_t len;
2149 if (decoded_state != trap_state) {
2150 // Random buggy state that doesn't decode??
2151 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2152 } else {
2153 len = jio_snprintf(buf, buflen, "%s%s",
2154 trap_reason_name(reason),
2155 recomp_flag ? " recompiled" : "");
2156 }
2157 return buf;
2158 }
2159
2160
2161 //--------------------------------statics--------------------------------------
2162 const char* Deoptimization::_trap_reason_name[] = {
2163 // Note: Keep this in sync. with enum DeoptReason.
2164 "none",
2165 "null_check",
2166 "null_assert" JVMCI_ONLY("_or_unreached0"),
2167 "range_check",
2168 "class_check",
2169 "array_check",
2170 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2171 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2172 "unloaded",
2173 "uninitialized",
2174 "unreached",
2175 "unhandled",
2176 "constraint",
2177 "div0_check",
2178 "age",
2179 "predicate",
2180 "loop_limit_check",
2181 "speculate_class_check",
2182 "speculate_null_check",
2183 "speculate_null_assert",
2184 "rtm_state_change",
2185 "unstable_if",
2186 "unstable_fused_if",
2187 #if INCLUDE_JVMCI
2188 "aliasing",
2189 "transfer_to_interpreter",
2190 "not_compiled_exception_handler",
2191 "unresolved",
2192 "jsr_mismatch",
2193 #endif
2194 "tenured"
2195 };
2196 const char* Deoptimization::_trap_action_name[] = {
2197 // Note: Keep this in sync. with enum DeoptAction.
2198 "none",
2199 "maybe_recompile",
2200 "reinterpret",
2201 "make_not_entrant",
2202 "make_not_compilable"
2203 };
2204
2205 const char* Deoptimization::trap_reason_name(int reason) {
2206 // Check that every reason has a name
2207 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2208
2209 if (reason == Reason_many) return "many";
2210 if ((uint)reason < Reason_LIMIT)
2211 return _trap_reason_name[reason];
2212 static char buf[20];
2213 sprintf(buf, "reason%d", reason);
2214 return buf;
2215 }
2216 const char* Deoptimization::trap_action_name(int action) {
2217 // Check that every action has a name
2218 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2219
2220 if ((uint)action < Action_LIMIT)
2221 return _trap_action_name[action];
2222 static char buf[20];
2223 sprintf(buf, "action%d", action);
2224 return buf;
2225 }
2226
2227 // This is used for debugging and diagnostics, including LogFile output.
2228 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2229 int trap_request) {
2230 jint unloaded_class_index = trap_request_index(trap_request);
2231 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2232 const char* action = trap_action_name(trap_request_action(trap_request));
2233 #if INCLUDE_JVMCI
2234 int debug_id = trap_request_debug_id(trap_request);
2235 #endif
2236 size_t len;
2237 if (unloaded_class_index < 0) {
2238 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2239 reason, action
2240 #if INCLUDE_JVMCI
2241 ,debug_id
2242 #endif
2243 );
2244 } else {
2245 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2246 reason, action, unloaded_class_index
2247 #if INCLUDE_JVMCI
2248 ,debug_id
2249 #endif
2250 );
2251 }
2252 return buf;
2253 }
2254
2255 juint Deoptimization::_deoptimization_hist
2256 [Deoptimization::Reason_LIMIT]
2257 [1 + Deoptimization::Action_LIMIT]
2258 [Deoptimization::BC_CASE_LIMIT]
2259 = {0};
2260
2261 enum {
2262 LSB_BITS = 8,
2263 LSB_MASK = right_n_bits(LSB_BITS)
2264 };
2265
2266 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2267 Bytecodes::Code bc) {
2268 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2269 assert(action >= 0 && action < Action_LIMIT, "oob");
2270 _deoptimization_hist[Reason_none][0][0] += 1; // total
2271 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2272 juint* cases = _deoptimization_hist[reason][1+action];
2273 juint* bc_counter_addr = NULL;
2274 juint bc_counter = 0;
2275 // Look for an unused counter, or an exact match to this BC.
2276 if (bc != Bytecodes::_illegal) {
2277 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2278 juint* counter_addr = &cases[bc_case];
2279 juint counter = *counter_addr;
2280 if ((counter == 0 && bc_counter_addr == NULL)
2281 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2282 // this counter is either free or is already devoted to this BC
2283 bc_counter_addr = counter_addr;
2284 bc_counter = counter | bc;
2285 }
2286 }
2287 }
2288 if (bc_counter_addr == NULL) {
2289 // Overflow, or no given bytecode.
2290 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2291 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2292 }
2293 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2294 }
2295
2296 jint Deoptimization::total_deoptimization_count() {
2297 return _deoptimization_hist[Reason_none][0][0];
2298 }
2299
2300 jint Deoptimization::deoptimization_count(DeoptReason reason) {
2301 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2302 return _deoptimization_hist[reason][0][0];
2303 }
2304
2305 void Deoptimization::print_statistics() {
2306 juint total = total_deoptimization_count();
2307 juint account = total;
2308 if (total != 0) {
2309 ttyLocker ttyl;
2310 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2311 tty->print_cr("Deoptimization traps recorded:");
2312 #define PRINT_STAT_LINE(name, r) \
2313 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2314 PRINT_STAT_LINE("total", total);
2315 // For each non-zero entry in the histogram, print the reason,
2316 // the action, and (if specifically known) the type of bytecode.
2317 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2318 for (int action = 0; action < Action_LIMIT; action++) {
2319 juint* cases = _deoptimization_hist[reason][1+action];
2320 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2321 juint counter = cases[bc_case];
2322 if (counter != 0) {
2323 char name[1*K];
2324 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2325 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2326 bc = Bytecodes::_illegal;
2327 sprintf(name, "%s/%s/%s",
2328 trap_reason_name(reason),
2329 trap_action_name(action),
2330 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2331 juint r = counter >> LSB_BITS;
2332 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2333 account -= r;
2334 }
2335 }
2336 }
2337 }
2338 if (account != 0) {
2339 PRINT_STAT_LINE("unaccounted", account);
2340 }
2341 #undef PRINT_STAT_LINE
2342 if (xtty != NULL) xtty->tail("statistics");
2343 }
2344 }
2345 #else // COMPILER2_OR_JVMCI
2346
2347
2348 // Stubs for C1 only system.
2349 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2350 return false;
2351 }
2352
2353 const char* Deoptimization::trap_reason_name(int reason) {
2354 return "unknown";
2355 }
2356
2357 void Deoptimization::print_statistics() {
2358 // no output
2359 }
2360
2361 void
2362 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2363 // no udpate
2364 }
2365
2366 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2367 return 0;
2368 }
2369
2370 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2371 Bytecodes::Code bc) {
2372 // no update
2373 }
2374
2375 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2376 int trap_state) {
2377 jio_snprintf(buf, buflen, "#%d", trap_state);
2378 return buf;
2379 }
2380
2381 #endif // COMPILER2_OR_JVMCI