1 /*
2 * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "code/codeCache.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "code/nmethod.hpp"
30 #include "code/pcDesc.hpp"
31 #include "code/scopeDesc.hpp"
32 #include "interpreter/bytecode.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "interpreter/oopMapCache.hpp"
35 #include "memory/allocation.inline.hpp"
36 #include "memory/oopFactory.hpp"
37 #include "memory/resourceArea.hpp"
38 #include "oops/method.hpp"
39 #include "oops/objArrayOop.inline.hpp"
40 #include "oops/oop.inline.hpp"
41 #include "oops/fieldStreams.hpp"
42 #include "oops/valueArrayKlass.hpp"
43 #include "oops/valueArrayOop.hpp"
44 #include "oops/valueKlass.hpp"
45 #include "oops/verifyOopClosure.hpp"
46 #include "prims/jvmtiThreadState.hpp"
47 #include "runtime/biasedLocking.hpp"
48 #include "runtime/compilationPolicy.hpp"
49 #include "runtime/deoptimization.hpp"
50 #include "runtime/interfaceSupport.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "runtime/signature.hpp"
53 #include "runtime/stubRoutines.hpp"
54 #include "runtime/thread.hpp"
55 #include "runtime/vframe.hpp"
56 #include "runtime/vframeArray.hpp"
57 #include "runtime/vframe_hp.hpp"
58 #include "utilities/events.hpp"
59 #include "utilities/xmlstream.hpp"
60
61 #if INCLUDE_JVMCI
62 #include "jvmci/jvmciRuntime.hpp"
63 #include "jvmci/jvmciJavaClasses.hpp"
64 #endif
65
66
67 bool DeoptimizationMarker::_is_active = false;
68
69 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
70 int caller_adjustment,
71 int caller_actual_parameters,
72 int number_of_frames,
73 intptr_t* frame_sizes,
74 address* frame_pcs,
75 BasicType return_type,
76 int exec_mode) {
77 _size_of_deoptimized_frame = size_of_deoptimized_frame;
78 _caller_adjustment = caller_adjustment;
79 _caller_actual_parameters = caller_actual_parameters;
80 _number_of_frames = number_of_frames;
81 _frame_sizes = frame_sizes;
82 _frame_pcs = frame_pcs;
83 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
84 _return_type = return_type;
85 _initial_info = 0;
86 // PD (x86 only)
87 _counter_temp = 0;
88 _unpack_kind = exec_mode;
89 _sender_sp_temp = 0;
90
91 _total_frame_sizes = size_of_frames();
92 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
93 }
94
95
96 Deoptimization::UnrollBlock::~UnrollBlock() {
97 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
98 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
99 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
100 }
101
102
103 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
104 assert(register_number < RegisterMap::reg_count, "checking register number");
105 return &_register_block[register_number * 2];
106 }
107
108
109
110 int Deoptimization::UnrollBlock::size_of_frames() const {
111 // Acount first for the adjustment of the initial frame
112 int result = _caller_adjustment;
113 for (int index = 0; index < number_of_frames(); index++) {
114 result += frame_sizes()[index];
115 }
116 return result;
117 }
118
119
120 void Deoptimization::UnrollBlock::print() {
121 ttyLocker ttyl;
122 tty->print_cr("UnrollBlock");
123 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
124 tty->print( " frame_sizes: ");
125 for (int index = 0; index < number_of_frames(); index++) {
126 tty->print(INTX_FORMAT " ", frame_sizes()[index]);
127 }
128 tty->cr();
129 }
130
131
132 // In order to make fetch_unroll_info work properly with escape
133 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
134 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
135 // of previously eliminated objects occurs in realloc_objects, which is
136 // called from the method fetch_unroll_info_helper below.
137 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode))
138 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
139 // but makes the entry a little slower. There is however a little dance we have to
140 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
141
142 // fetch_unroll_info() is called at the beginning of the deoptimization
143 // handler. Note this fact before we start generating temporary frames
144 // that can confuse an asynchronous stack walker. This counter is
145 // decremented at the end of unpack_frames().
146 if (TraceDeoptimization) {
147 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread));
148 }
149 thread->inc_in_deopt_handler();
150
151 return fetch_unroll_info_helper(thread, exec_mode);
152 JRT_END
153
154
155 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
156 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) {
157
158 // Note: there is a safepoint safety issue here. No matter whether we enter
159 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
160 // the vframeArray is created.
161 //
162
163 // Allocate our special deoptimization ResourceMark
164 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
165 assert(thread->deopt_mark() == NULL, "Pending deopt!");
166 thread->set_deopt_mark(dmark);
167
168 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
169 RegisterMap map(thread, true);
170 RegisterMap dummy_map(thread, false);
171 // Now get the deoptee with a valid map
172 frame deoptee = stub_frame.sender(&map);
173 // Set the deoptee nmethod
174 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!");
175 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
176 thread->set_deopt_compiled_method(cm);
177
178 if (VerifyStack) {
179 thread->validate_frame_layout();
180 }
181
182 // Create a growable array of VFrames where each VFrame represents an inlined
183 // Java frame. This storage is allocated with the usual system arena.
184 assert(deoptee.is_compiled_frame(), "Wrong frame type");
185 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
186 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
187 while (!vf->is_top()) {
188 assert(vf->is_compiled_frame(), "Wrong frame type");
189 chunk->push(compiledVFrame::cast(vf));
190 vf = vf->sender();
191 }
192 assert(vf->is_compiled_frame(), "Wrong frame type");
193 chunk->push(compiledVFrame::cast(vf));
194
195 bool realloc_failures = false;
196
197 #if defined(COMPILER2) || INCLUDE_JVMCI
198 // Reallocate the non-escaping objects and restore their fields. Then
199 // relock objects if synchronization on them was eliminated.
200 #ifndef INCLUDE_JVMCI
201 if (DoEscapeAnalysis || EliminateNestedLocks) {
202 if (EliminateAllocations) {
203 #endif // INCLUDE_JVMCI
204 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
205 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
206
207 // The flag return_oop() indicates call sites which return oop
208 // in compiled code. Such sites include java method calls,
209 // runtime calls (for example, used to allocate new objects/arrays
210 // on slow code path) and any other calls generated in compiled code.
211 // It is not guaranteed that we can get such information here only
212 // by analyzing bytecode in deoptimized frames. This is why this flag
213 // is set during method compilation (see Compile::Process_OopMap_Node()).
214 // If the previous frame was popped or if we are dispatching an exception,
215 // we don't have an oop result.
216 ScopeDesc* scope = chunk->at(0)->scope();
217 bool save_oop_result = scope->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt);
218 // In case of the return of multiple values, we must take care
219 // of all oop return values.
220 GrowableArray<Handle> return_oops;
221 ValueKlass* vk = NULL;
222 if (save_oop_result) {
223 if (scope->return_vt()) {
224 vk = ValueKlass::returned_value_type(map);
225 if (vk != NULL) {
226 bool success = vk->save_oop_results(map, return_oops);
227 assert(success, "found klass ptr being returned: saving oops can't fail");
228 save_oop_result = false;
229 } else {
230 vk = NULL;
231 }
232 }
233 }
234 if (save_oop_result) {
235 // Reallocation may trigger GC. If deoptimization happened on return from
236 // call which returns oop we need to save it since it is not in oopmap.
237 oop result = deoptee.saved_oop_result(&map);
238 assert(result == NULL || result->is_oop(), "must be oop");
239 return_oops.push(Handle(thread, result));
240 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
241 if (TraceDeoptimization) {
242 ttyLocker ttyl;
243 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
244 }
245 }
246 if (objects != NULL || vk != NULL) {
247 bool skip_internal = (cm != NULL) && !cm->is_compiled_by_jvmci();
248 JRT_BLOCK
249 if (vk != NULL) {
250 realloc_failures = realloc_value_type_result(vk, map, return_oops, THREAD);
251 }
252 if (objects != NULL) {
253 realloc_failures = realloc_failures || realloc_objects(thread, &deoptee, objects, THREAD);
254 reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, THREAD);
255 }
256 JRT_END
257 #ifndef PRODUCT
258 if (TraceDeoptimization) {
259 ttyLocker ttyl;
260 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
261 print_objects(objects, realloc_failures);
262 }
263 #endif
264 }
265 if (save_oop_result || vk != NULL) {
266 // Restore result.
267 assert(return_oops.length() == 1, "no value type");
268 deoptee.set_saved_oop_result(&map, return_oops.pop()());
269 }
270 #ifndef INCLUDE_JVMCI
271 }
272 if (EliminateLocks) {
273 #endif // INCLUDE_JVMCI
274 #ifndef PRODUCT
275 bool first = true;
276 #endif
277 for (int i = 0; i < chunk->length(); i++) {
278 compiledVFrame* cvf = chunk->at(i);
279 assert (cvf->scope() != NULL,"expect only compiled java frames");
280 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
281 if (monitors->is_nonempty()) {
282 relock_objects(monitors, thread, realloc_failures);
283 #ifndef PRODUCT
284 if (PrintDeoptimizationDetails) {
285 ttyLocker ttyl;
286 for (int j = 0; j < monitors->length(); j++) {
287 MonitorInfo* mi = monitors->at(j);
288 if (mi->eliminated()) {
289 if (first) {
290 first = false;
291 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
292 }
293 if (mi->owner_is_scalar_replaced()) {
294 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
295 tty->print_cr(" failed reallocation for klass %s", k->external_name());
296 } else {
297 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
298 }
299 }
300 }
301 }
302 #endif // !PRODUCT
303 }
304 }
305 #ifndef INCLUDE_JVMCI
306 }
307 }
308 #endif // INCLUDE_JVMCI
309 #endif // COMPILER2 || INCLUDE_JVMCI
310
311 ScopeDesc* trap_scope = chunk->at(0)->scope();
312 Handle exceptionObject;
313 if (trap_scope->rethrow_exception()) {
314 if (PrintDeoptimizationDetails) {
315 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());
316 }
317 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
318 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw");
319 ScopeValue* topOfStack = expressions->top();
320 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
321 guarantee(exceptionObject() != NULL, "exception oop can not be null");
322 }
323
324 // Ensure that no safepoint is taken after pointers have been stored
325 // in fields of rematerialized objects. If a safepoint occurs from here on
326 // out the java state residing in the vframeArray will be missed.
327 NoSafepointVerifier no_safepoint;
328
329 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
330 #if defined(COMPILER2) || INCLUDE_JVMCI
331 if (realloc_failures) {
332 pop_frames_failed_reallocs(thread, array);
333 }
334 #endif
335
336 assert(thread->vframe_array_head() == NULL, "Pending deopt!");
337 thread->set_vframe_array_head(array);
338
339 // Now that the vframeArray has been created if we have any deferred local writes
340 // added by jvmti then we can free up that structure as the data is now in the
341 // vframeArray
342
343 if (thread->deferred_locals() != NULL) {
344 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
345 int i = 0;
346 do {
347 // Because of inlining we could have multiple vframes for a single frame
348 // and several of the vframes could have deferred writes. Find them all.
349 if (list->at(i)->id() == array->original().id()) {
350 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
351 list->remove_at(i);
352 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
353 delete dlv;
354 } else {
355 i++;
356 }
357 } while ( i < list->length() );
358 if (list->length() == 0) {
359 thread->set_deferred_locals(NULL);
360 // free the list and elements back to C heap.
361 delete list;
362 }
363
364 }
365
366 #ifndef SHARK
367 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
368 CodeBlob* cb = stub_frame.cb();
369 // Verify we have the right vframeArray
370 assert(cb->frame_size() >= 0, "Unexpected frame size");
371 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
372
373 // If the deopt call site is a MethodHandle invoke call site we have
374 // to adjust the unpack_sp.
375 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
376 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
377 unpack_sp = deoptee.unextended_sp();
378
379 #ifdef ASSERT
380 assert(cb->is_deoptimization_stub() ||
381 cb->is_uncommon_trap_stub() ||
382 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
383 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
384 "unexpected code blob: %s", cb->name());
385 #endif
386 #else
387 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
388 #endif // !SHARK
389
390 // This is a guarantee instead of an assert because if vframe doesn't match
391 // we will unpack the wrong deoptimized frame and wind up in strange places
392 // where it will be very difficult to figure out what went wrong. Better
393 // to die an early death here than some very obscure death later when the
394 // trail is cold.
395 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
396 // in that it will fail to detect a problem when there is one. This needs
397 // more work in tiger timeframe.
398 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
399
400 int number_of_frames = array->frames();
401
402 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
403 // virtual activation, which is the reverse of the elements in the vframes array.
404 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
405 // +1 because we always have an interpreter return address for the final slot.
406 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
407 int popframe_extra_args = 0;
408 // Create an interpreter return address for the stub to use as its return
409 // address so the skeletal frames are perfectly walkable
410 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
411
412 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
413 // activation be put back on the expression stack of the caller for reexecution
414 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
415 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
416 }
417
418 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
419 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
420 // than simply use array->sender.pc(). This requires us to walk the current set of frames
421 //
422 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
423 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
424
425 // It's possible that the number of parameters at the call site is
426 // different than number of arguments in the callee when method
427 // handles are used. If the caller is interpreted get the real
428 // value so that the proper amount of space can be added to it's
429 // frame.
430 bool caller_was_method_handle = false;
431 if (deopt_sender.is_interpreted_frame()) {
432 methodHandle method = deopt_sender.interpreter_frame_method();
433 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
434 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
435 // Method handle invokes may involve fairly arbitrary chains of
436 // calls so it's impossible to know how much actual space the
437 // caller has for locals.
438 caller_was_method_handle = true;
439 }
440 }
441
442 //
443 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
444 // frame_sizes/frame_pcs[1] next oldest frame (int)
445 // frame_sizes/frame_pcs[n] youngest frame (int)
446 //
447 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
448 // owns the space for the return address to it's caller). Confusing ain't it.
449 //
450 // The vframe array can address vframes with indices running from
451 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
452 // When we create the skeletal frames we need the oldest frame to be in the zero slot
453 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
454 // so things look a little strange in this loop.
455 //
456 int callee_parameters = 0;
457 int callee_locals = 0;
458 for (int index = 0; index < array->frames(); index++ ) {
459 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
460 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
461 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
462 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
463 callee_locals,
464 index == 0,
465 popframe_extra_args);
466 // This pc doesn't have to be perfect just good enough to identify the frame
467 // as interpreted so the skeleton frame will be walkable
468 // The correct pc will be set when the skeleton frame is completely filled out
469 // The final pc we store in the loop is wrong and will be overwritten below
470 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
471
472 callee_parameters = array->element(index)->method()->size_of_parameters();
473 callee_locals = array->element(index)->method()->max_locals();
474 popframe_extra_args = 0;
475 }
476
477 // Compute whether the root vframe returns a float or double value.
478 BasicType return_type;
479 {
480 methodHandle method(thread, array->element(0)->method());
481 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
482 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
483 }
484
485 // Compute information for handling adapters and adjusting the frame size of the caller.
486 int caller_adjustment = 0;
487
488 // Compute the amount the oldest interpreter frame will have to adjust
489 // its caller's stack by. If the caller is a compiled frame then
490 // we pretend that the callee has no parameters so that the
491 // extension counts for the full amount of locals and not just
492 // locals-parms. This is because without a c2i adapter the parm
493 // area as created by the compiled frame will not be usable by
494 // the interpreter. (Depending on the calling convention there
495 // may not even be enough space).
496
497 // QQQ I'd rather see this pushed down into last_frame_adjust
498 // and have it take the sender (aka caller).
499
500 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
501 caller_adjustment = last_frame_adjust(0, callee_locals);
502 } else if (callee_locals > callee_parameters) {
503 // The caller frame may need extending to accommodate
504 // non-parameter locals of the first unpacked interpreted frame.
505 // Compute that adjustment.
506 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
507 }
508
509 // If the sender is deoptimized the we must retrieve the address of the handler
510 // since the frame will "magically" show the original pc before the deopt
511 // and we'd undo the deopt.
512
513 frame_pcs[0] = deopt_sender.raw_pc();
514
515 #ifndef SHARK
516 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
517 #endif // SHARK
518
519 #ifdef INCLUDE_JVMCI
520 if (exceptionObject() != NULL) {
521 thread->set_exception_oop(exceptionObject());
522 exec_mode = Unpack_exception;
523 }
524 #endif
525
526 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
527 assert(thread->has_pending_exception(), "should have thrown OOME");
528 thread->set_exception_oop(thread->pending_exception());
529 thread->clear_pending_exception();
530 exec_mode = Unpack_exception;
531 }
532
533 #if INCLUDE_JVMCI
534 if (thread->frames_to_pop_failed_realloc() > 0) {
535 thread->set_pending_monitorenter(false);
536 }
537 #endif
538
539 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
540 caller_adjustment * BytesPerWord,
541 caller_was_method_handle ? 0 : callee_parameters,
542 number_of_frames,
543 frame_sizes,
544 frame_pcs,
545 return_type,
546 exec_mode);
547 // On some platforms, we need a way to pass some platform dependent
548 // information to the unpacking code so the skeletal frames come out
549 // correct (initial fp value, unextended sp, ...)
550 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
551
552 if (array->frames() > 1) {
553 if (VerifyStack && TraceDeoptimization) {
554 ttyLocker ttyl;
555 tty->print_cr("Deoptimizing method containing inlining");
556 }
557 }
558
559 array->set_unroll_block(info);
560 return info;
561 }
562
563 // Called to cleanup deoptimization data structures in normal case
564 // after unpacking to stack and when stack overflow error occurs
565 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
566 vframeArray *array) {
567
568 // Get array if coming from exception
569 if (array == NULL) {
570 array = thread->vframe_array_head();
571 }
572 thread->set_vframe_array_head(NULL);
573
574 // Free the previous UnrollBlock
575 vframeArray* old_array = thread->vframe_array_last();
576 thread->set_vframe_array_last(array);
577
578 if (old_array != NULL) {
579 UnrollBlock* old_info = old_array->unroll_block();
580 old_array->set_unroll_block(NULL);
581 delete old_info;
582 delete old_array;
583 }
584
585 // Deallocate any resource creating in this routine and any ResourceObjs allocated
586 // inside the vframeArray (StackValueCollections)
587
588 delete thread->deopt_mark();
589 thread->set_deopt_mark(NULL);
590 thread->set_deopt_compiled_method(NULL);
591
592
593 if (JvmtiExport::can_pop_frame()) {
594 #ifndef CC_INTERP
595 // Regardless of whether we entered this routine with the pending
596 // popframe condition bit set, we should always clear it now
597 thread->clear_popframe_condition();
598 #else
599 // C++ interpreter will clear has_pending_popframe when it enters
600 // with method_resume. For deopt_resume2 we clear it now.
601 if (thread->popframe_forcing_deopt_reexecution())
602 thread->clear_popframe_condition();
603 #endif /* CC_INTERP */
604 }
605
606 // unpack_frames() is called at the end of the deoptimization handler
607 // and (in C2) at the end of the uncommon trap handler. Note this fact
608 // so that an asynchronous stack walker can work again. This counter is
609 // incremented at the beginning of fetch_unroll_info() and (in C2) at
610 // the beginning of uncommon_trap().
611 thread->dec_in_deopt_handler();
612 }
613
614 // Moved from cpu directories because none of the cpus has callee save values.
615 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
616 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
617
618 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
619 // the days we had adapter frames. When we deoptimize a situation where a
620 // compiled caller calls a compiled caller will have registers it expects
621 // to survive the call to the callee. If we deoptimize the callee the only
622 // way we can restore these registers is to have the oldest interpreter
623 // frame that we create restore these values. That is what this routine
624 // will accomplish.
625
626 // At the moment we have modified c2 to not have any callee save registers
627 // so this problem does not exist and this routine is just a place holder.
628
629 assert(f->is_interpreted_frame(), "must be interpreted");
630 }
631
632 // Return BasicType of value being returned
633 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
634
635 // We are already active int he special DeoptResourceMark any ResourceObj's we
636 // allocate will be freed at the end of the routine.
637
638 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
639 // but makes the entry a little slower. There is however a little dance we have to
640 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
641 ResetNoHandleMark rnhm; // No-op in release/product versions
642 HandleMark hm;
643
644 frame stub_frame = thread->last_frame();
645
646 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
647 // must point to the vframeArray for the unpack frame.
648 vframeArray* array = thread->vframe_array_head();
649
650 #ifndef PRODUCT
651 if (TraceDeoptimization) {
652 ttyLocker ttyl;
653 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d",
654 p2i(thread), p2i(array), exec_mode);
655 }
656 #endif
657 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
658 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode);
659
660 UnrollBlock* info = array->unroll_block();
661
662 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
663 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
664
665 BasicType bt = info->return_type();
666
667 // If we have an exception pending, claim that the return type is an oop
668 // so the deopt_blob does not overwrite the exception_oop.
669
670 if (exec_mode == Unpack_exception)
671 bt = T_OBJECT;
672
673 // Cleanup thread deopt data
674 cleanup_deopt_info(thread, array);
675
676 #ifndef PRODUCT
677 if (VerifyStack) {
678 ResourceMark res_mark;
679
680 thread->validate_frame_layout();
681
682 // Verify that the just-unpacked frames match the interpreter's
683 // notions of expression stack and locals
684 vframeArray* cur_array = thread->vframe_array_last();
685 RegisterMap rm(thread, false);
686 rm.set_include_argument_oops(false);
687 bool is_top_frame = true;
688 int callee_size_of_parameters = 0;
689 int callee_max_locals = 0;
690 for (int i = 0; i < cur_array->frames(); i++) {
691 vframeArrayElement* el = cur_array->element(i);
692 frame* iframe = el->iframe();
693 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
694
695 // Get the oop map for this bci
696 InterpreterOopMap mask;
697 int cur_invoke_parameter_size = 0;
698 bool try_next_mask = false;
699 int next_mask_expression_stack_size = -1;
700 int top_frame_expression_stack_adjustment = 0;
701 methodHandle mh(thread, iframe->interpreter_frame_method());
702 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
703 BytecodeStream str(mh);
704 str.set_start(iframe->interpreter_frame_bci());
705 int max_bci = mh->code_size();
706 // Get to the next bytecode if possible
707 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
708 // Check to see if we can grab the number of outgoing arguments
709 // at an uncommon trap for an invoke (where the compiler
710 // generates debug info before the invoke has executed)
711 Bytecodes::Code cur_code = str.next();
712 if (cur_code == Bytecodes::_invokevirtual ||
713 cur_code == Bytecodes::_invokespecial ||
714 cur_code == Bytecodes::_invokestatic ||
715 cur_code == Bytecodes::_invokeinterface ||
716 cur_code == Bytecodes::_invokedynamic) {
717 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
718 Symbol* signature = invoke.signature();
719 ArgumentSizeComputer asc(signature);
720 cur_invoke_parameter_size = asc.size();
721 if (invoke.has_receiver()) {
722 // Add in receiver
723 ++cur_invoke_parameter_size;
724 }
725 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
726 callee_size_of_parameters++;
727 }
728 }
729 if (str.bci() < max_bci) {
730 Bytecodes::Code bc = str.next();
731 if (bc >= 0) {
732 // The interpreter oop map generator reports results before
733 // the current bytecode has executed except in the case of
734 // calls. It seems to be hard to tell whether the compiler
735 // has emitted debug information matching the "state before"
736 // a given bytecode or the state after, so we try both
737 switch (cur_code) {
738 case Bytecodes::_invokevirtual:
739 case Bytecodes::_invokespecial:
740 case Bytecodes::_invokestatic:
741 case Bytecodes::_invokeinterface:
742 case Bytecodes::_invokedynamic:
743 case Bytecodes::_athrow:
744 break;
745 default: {
746 InterpreterOopMap next_mask;
747 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
748 next_mask_expression_stack_size = next_mask.expression_stack_size();
749 // Need to subtract off the size of the result type of
750 // the bytecode because this is not described in the
751 // debug info but returned to the interpreter in the TOS
752 // caching register
753 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
754 if (bytecode_result_type != T_ILLEGAL) {
755 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
756 }
757 assert(top_frame_expression_stack_adjustment >= 0, "");
758 try_next_mask = true;
759 break;
760 }
761 }
762 }
763 }
764
765 // Verify stack depth and oops in frame
766 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
767 if (!(
768 /* SPARC */
769 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
770 /* x86 */
771 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
772 (try_next_mask &&
773 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
774 top_frame_expression_stack_adjustment))) ||
775 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
776 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
777 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
778 )) {
779 ttyLocker ttyl;
780
781 // Print out some information that will help us debug the problem
782 tty->print_cr("Wrong number of expression stack elements during deoptimization");
783 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
784 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
785 iframe->interpreter_frame_expression_stack_size());
786 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
787 tty->print_cr(" try_next_mask = %d", try_next_mask);
788 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
789 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
790 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
791 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
792 tty->print_cr(" exec_mode = %d", exec_mode);
793 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
794 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
795 tty->print_cr(" Interpreted frames:");
796 for (int k = 0; k < cur_array->frames(); k++) {
797 vframeArrayElement* el = cur_array->element(k);
798 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
799 }
800 cur_array->print_on_2(tty);
801 guarantee(false, "wrong number of expression stack elements during deopt");
802 }
803 VerifyOopClosure verify;
804 iframe->oops_interpreted_do(&verify, &rm, false);
805 callee_size_of_parameters = mh->size_of_parameters();
806 callee_max_locals = mh->max_locals();
807 is_top_frame = false;
808 }
809 }
810 #endif /* !PRODUCT */
811
812
813 return bt;
814 JRT_END
815
816
817 int Deoptimization::deoptimize_dependents() {
818 Threads::deoptimized_wrt_marked_nmethods();
819 return 0;
820 }
821
822 Deoptimization::DeoptAction Deoptimization::_unloaded_action
823 = Deoptimization::Action_reinterpret;
824
825 #if defined(COMPILER2) || INCLUDE_JVMCI
826 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
827 Handle pending_exception(THREAD, thread->pending_exception());
828 const char* exception_file = thread->exception_file();
829 int exception_line = thread->exception_line();
830 thread->clear_pending_exception();
831
832 bool failures = false;
833
834 for (int i = 0; i < objects->length(); i++) {
835 assert(objects->at(i)->is_object(), "invalid debug information");
836 ObjectValue* sv = (ObjectValue*) objects->at(i);
837
838 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
839 oop obj = NULL;
840
841 if (k->is_instance_klass()) {
842 InstanceKlass* ik = InstanceKlass::cast(k);
843 obj = ik->allocate_instance(THREAD);
844 } else if (k->is_valueArray_klass()) {
845 ValueArrayKlass* ak = ValueArrayKlass::cast(k);
846 // Value type array must be zeroed because not all memory is reassigned
847 obj = ak->allocate(sv->field_size(), true, THREAD);
848 } else if (k->is_typeArray_klass()) {
849 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
850 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
851 int len = sv->field_size() / type2size[ak->element_type()];
852 obj = ak->allocate(len, THREAD);
853 } else if (k->is_objArray_klass()) {
854 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
855 obj = ak->allocate(sv->field_size(), THREAD);
856 }
857
858 if (obj == NULL) {
859 failures = true;
860 }
861
862 assert(sv->value().is_null(), "redundant reallocation");
863 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
864 CLEAR_PENDING_EXCEPTION;
865 sv->set_value(obj);
866 }
867
868 if (failures) {
869 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
870 } else if (pending_exception.not_null()) {
871 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
872 }
873
874 return failures;
875 }
876
877 // We're deoptimizing at the return of a call, value type fields are
878 // in registers. When we go back to the interpreter, it will expect a
879 // reference to a value type instance. Allocate and initialize it from
880 // the register values here.
881 bool Deoptimization::realloc_value_type_result(ValueKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) {
882 oop new_vt = vk->realloc_result(map, return_oops, THREAD);
883 if (new_vt == NULL) {
884 CLEAR_PENDING_EXCEPTION;
885 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true);
886 }
887 return_oops.clear();
888 return_oops.push(Handle(THREAD, new_vt));
889 return false;
890 }
891
892 // restore elements of an eliminated type array
893 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
894 int index = 0;
895 intptr_t val;
896
897 for (int i = 0; i < sv->field_size(); i++) {
898 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
899 switch(type) {
900 case T_LONG: case T_DOUBLE: {
901 assert(value->type() == T_INT, "Agreement.");
902 StackValue* low =
903 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
904 #ifdef _LP64
905 jlong res = (jlong)low->get_int();
906 #else
907 #ifdef SPARC
908 // For SPARC we have to swap high and low words.
909 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
910 #else
911 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
912 #endif //SPARC
913 #endif
914 obj->long_at_put(index, res);
915 break;
916 }
917
918 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
919 case T_INT: case T_FLOAT: { // 4 bytes.
920 assert(value->type() == T_INT, "Agreement.");
921 bool big_value = false;
922 if (i + 1 < sv->field_size() && type == T_INT) {
923 if (sv->field_at(i)->is_location()) {
924 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
925 if (type == Location::dbl || type == Location::lng) {
926 big_value = true;
927 }
928 } else if (sv->field_at(i)->is_constant_int()) {
929 ScopeValue* next_scope_field = sv->field_at(i + 1);
930 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
931 big_value = true;
932 }
933 }
934 }
935
936 if (big_value) {
937 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
938 #ifdef _LP64
939 jlong res = (jlong)low->get_int();
940 #else
941 #ifdef SPARC
942 // For SPARC we have to swap high and low words.
943 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
944 #else
945 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
946 #endif //SPARC
947 #endif
948 obj->int_at_put(index, (jint)*((jint*)&res));
949 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
950 } else {
951 val = value->get_int();
952 obj->int_at_put(index, (jint)*((jint*)&val));
953 }
954 break;
955 }
956
957 case T_SHORT:
958 assert(value->type() == T_INT, "Agreement.");
959 val = value->get_int();
960 obj->short_at_put(index, (jshort)*((jint*)&val));
961 break;
962
963 case T_CHAR:
964 assert(value->type() == T_INT, "Agreement.");
965 val = value->get_int();
966 obj->char_at_put(index, (jchar)*((jint*)&val));
967 break;
968
969 case T_BYTE:
970 assert(value->type() == T_INT, "Agreement.");
971 val = value->get_int();
972 obj->byte_at_put(index, (jbyte)*((jint*)&val));
973 break;
974
975 case T_BOOLEAN:
976 assert(value->type() == T_INT, "Agreement.");
977 val = value->get_int();
978 obj->bool_at_put(index, (jboolean)*((jint*)&val));
979 break;
980
981 default:
982 ShouldNotReachHere();
983 }
984 index++;
985 }
986 }
987
988
989 // restore fields of an eliminated object array
990 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
991 for (int i = 0; i < sv->field_size(); i++) {
992 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
993 assert(value->type() == T_OBJECT, "object element expected");
994 obj->obj_at_put(i, value->get_obj()());
995 }
996 }
997
998 class ReassignedField {
999 public:
1000 int _offset;
1001 BasicType _type;
1002 InstanceKlass* _klass;
1003 public:
1004 ReassignedField() {
1005 _offset = 0;
1006 _type = T_ILLEGAL;
1007 _klass = NULL;
1008 }
1009 };
1010
1011 int compare(ReassignedField* left, ReassignedField* right) {
1012 return left->_offset - right->_offset;
1013 }
1014
1015 // Restore fields of an eliminated instance object using the same field order
1016 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1017 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal, int base_offset, TRAPS) {
1018 if (klass->superklass() != NULL) {
1019 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal, 0, CHECK_0);
1020 }
1021
1022 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1023 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
1024 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
1025 ReassignedField field;
1026 field._offset = fs.offset();
1027 field._type = FieldType::basic_type(fs.signature());
1028 if (field._type == T_VALUETYPE) {
1029 // Resolve klass of flattened value type field
1030 SignatureStream ss(fs.signature(), false);
1031 Klass* vk = ss.as_klass(Handle(THREAD, klass->class_loader()), Handle(THREAD, klass->protection_domain()), SignatureStream::NCDFError, THREAD);
1032 guarantee(!HAS_PENDING_EXCEPTION, "Should not have any exceptions pending");
1033 assert(vk->is_value(), "must be a ValueKlass");
1034 field._klass = InstanceKlass::cast(vk);
1035 }
1036 fields->append(field);
1037 }
1038 }
1039 fields->sort(compare);
1040 for (int i = 0; i < fields->length(); i++) {
1041 intptr_t val;
1042 ScopeValue* scope_field = sv->field_at(svIndex);
1043 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1044 int offset = base_offset + fields->at(i)._offset;
1045 BasicType type = fields->at(i)._type;
1046 switch (type) {
1047 case T_OBJECT: case T_ARRAY:
1048 assert(value->type() == T_OBJECT, "Agreement.");
1049 obj->obj_field_put(offset, value->get_obj()());
1050 break;
1051
1052 case T_VALUETYPE: {
1053 // Recursively re-assign flattened value type fields
1054 InstanceKlass* vk = fields->at(i)._klass;
1055 assert(vk != NULL, "must be resolved");
1056 offset -= ValueKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header
1057 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0);
1058 continue; // Continue because we don't need to increment svIndex
1059 }
1060
1061 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1062 case T_INT: case T_FLOAT: { // 4 bytes.
1063 assert(value->type() == T_INT, "Agreement.");
1064 bool big_value = false;
1065 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1066 if (scope_field->is_location()) {
1067 Location::Type type = ((LocationValue*) scope_field)->location().type();
1068 if (type == Location::dbl || type == Location::lng) {
1069 big_value = true;
1070 }
1071 }
1072 if (scope_field->is_constant_int()) {
1073 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1074 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1075 big_value = true;
1076 }
1077 }
1078 }
1079
1080 if (big_value) {
1081 i++;
1082 assert(i < fields->length(), "second T_INT field needed");
1083 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1084 } else {
1085 val = value->get_int();
1086 obj->int_field_put(offset, (jint)*((jint*)&val));
1087 break;
1088 }
1089 }
1090 /* no break */
1091
1092 case T_LONG: case T_DOUBLE: {
1093 assert(value->type() == T_INT, "Agreement.");
1094 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1095 #ifdef _LP64
1096 jlong res = (jlong)low->get_int();
1097 #else
1098 #ifdef SPARC
1099 // For SPARC we have to swap high and low words.
1100 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1101 #else
1102 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1103 #endif //SPARC
1104 #endif
1105 obj->long_field_put(offset, res);
1106 break;
1107 }
1108
1109 case T_SHORT:
1110 assert(value->type() == T_INT, "Agreement.");
1111 val = value->get_int();
1112 obj->short_field_put(offset, (jshort)*((jint*)&val));
1113 break;
1114
1115 case T_CHAR:
1116 assert(value->type() == T_INT, "Agreement.");
1117 val = value->get_int();
1118 obj->char_field_put(offset, (jchar)*((jint*)&val));
1119 break;
1120
1121 case T_BYTE:
1122 assert(value->type() == T_INT, "Agreement.");
1123 val = value->get_int();
1124 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1125 break;
1126
1127 case T_BOOLEAN:
1128 assert(value->type() == T_INT, "Agreement.");
1129 val = value->get_int();
1130 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1131 break;
1132
1133 default:
1134 ShouldNotReachHere();
1135 }
1136 svIndex++;
1137 }
1138 return svIndex;
1139 }
1140
1141 // restore fields of an eliminated value type array
1142 void Deoptimization::reassign_value_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, valueArrayOop obj, ValueArrayKlass* vak, TRAPS) {
1143 ValueKlass* vk = vak->element_klass();
1144 assert(vk->flatten_array(), "should only be used for flattened value type arrays");
1145 // Adjust offset to omit oop header
1146 int base_offset = arrayOopDesc::base_offset_in_bytes(T_VALUETYPE) - ValueKlass::cast(vk)->first_field_offset();
1147 // Initialize all elements of the flattened value type array
1148 for (int i = 0; i < sv->field_size(); i++) {
1149 ScopeValue* val = sv->field_at(i);
1150 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper()));
1151 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, false /* skip_internal */, offset, CHECK);
1152 }
1153 }
1154
1155 // restore fields of all eliminated objects and arrays
1156 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) {
1157 for (int i = 0; i < objects->length(); i++) {
1158 ObjectValue* sv = (ObjectValue*) objects->at(i);
1159 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1160 Handle obj = sv->value();
1161 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1162 if (PrintDeoptimizationDetails) {
1163 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1164 }
1165 if (obj.is_null()) {
1166 continue;
1167 }
1168
1169 if (k->is_instance_klass()) {
1170 InstanceKlass* ik = InstanceKlass::cast(k);
1171 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK);
1172 } else if (k->is_valueArray_klass()) {
1173 ValueArrayKlass* vak = ValueArrayKlass::cast(k);
1174 reassign_value_array_elements(fr, reg_map, sv, (valueArrayOop) obj(), vak, CHECK);
1175 } else if (k->is_typeArray_klass()) {
1176 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1177 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1178 } else if (k->is_objArray_klass()) {
1179 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1180 }
1181 }
1182 }
1183
1184
1185 // relock objects for which synchronization was eliminated
1186 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
1187 for (int i = 0; i < monitors->length(); i++) {
1188 MonitorInfo* mon_info = monitors->at(i);
1189 if (mon_info->eliminated()) {
1190 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1191 if (!mon_info->owner_is_scalar_replaced()) {
1192 Handle obj(thread, mon_info->owner());
1193 markOop mark = obj->mark();
1194 if (UseBiasedLocking && mark->has_bias_pattern()) {
1195 // New allocated objects may have the mark set to anonymously biased.
1196 // Also the deoptimized method may called methods with synchronization
1197 // where the thread-local object is bias locked to the current thread.
1198 assert(mark->is_biased_anonymously() ||
1199 mark->biased_locker() == thread, "should be locked to current thread");
1200 // Reset mark word to unbiased prototype.
1201 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
1202 obj->set_mark(unbiased_prototype);
1203 }
1204 BasicLock* lock = mon_info->lock();
1205 ObjectSynchronizer::slow_enter(obj, lock, thread);
1206 assert(mon_info->owner()->is_locked(), "object must be locked now");
1207 }
1208 }
1209 }
1210 }
1211
1212
1213 #ifndef PRODUCT
1214 // print information about reallocated objects
1215 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1216 fieldDescriptor fd;
1217
1218 for (int i = 0; i < objects->length(); i++) {
1219 ObjectValue* sv = (ObjectValue*) objects->at(i);
1220 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1221 Handle obj = sv->value();
1222
1223 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1224 k->print_value();
1225 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1226 if (obj.is_null()) {
1227 tty->print(" allocation failed");
1228 } else {
1229 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1230 }
1231 tty->cr();
1232
1233 if (Verbose && !obj.is_null()) {
1234 k->oop_print_on(obj(), tty);
1235 }
1236 }
1237 }
1238 #endif
1239 #endif // COMPILER2 || INCLUDE_JVMCI
1240
1241 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1242 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1243
1244 #ifndef PRODUCT
1245 if (PrintDeoptimizationDetails) {
1246 ttyLocker ttyl;
1247 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1248 fr.print_on(tty);
1249 tty->print_cr(" Virtual frames (innermost first):");
1250 for (int index = 0; index < chunk->length(); index++) {
1251 compiledVFrame* vf = chunk->at(index);
1252 tty->print(" %2d - ", index);
1253 vf->print_value();
1254 int bci = chunk->at(index)->raw_bci();
1255 const char* code_name;
1256 if (bci == SynchronizationEntryBCI) {
1257 code_name = "sync entry";
1258 } else {
1259 Bytecodes::Code code = vf->method()->code_at(bci);
1260 code_name = Bytecodes::name(code);
1261 }
1262 tty->print(" - %s", code_name);
1263 tty->print_cr(" @ bci %d ", bci);
1264 if (Verbose) {
1265 vf->print();
1266 tty->cr();
1267 }
1268 }
1269 }
1270 #endif
1271
1272 // Register map for next frame (used for stack crawl). We capture
1273 // the state of the deopt'ing frame's caller. Thus if we need to
1274 // stuff a C2I adapter we can properly fill in the callee-save
1275 // register locations.
1276 frame caller = fr.sender(reg_map);
1277 int frame_size = caller.sp() - fr.sp();
1278
1279 frame sender = caller;
1280
1281 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1282 // the vframeArray containing the unpacking information is allocated in the C heap.
1283 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1284 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1285
1286 // Compare the vframeArray to the collected vframes
1287 assert(array->structural_compare(thread, chunk), "just checking");
1288
1289 #ifndef PRODUCT
1290 if (PrintDeoptimizationDetails) {
1291 ttyLocker ttyl;
1292 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1293 }
1294 #endif // PRODUCT
1295
1296 return array;
1297 }
1298
1299 #if defined(COMPILER2) || INCLUDE_JVMCI
1300 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1301 // Reallocation of some scalar replaced objects failed. Record
1302 // that we need to pop all the interpreter frames for the
1303 // deoptimized compiled frame.
1304 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1305 thread->set_frames_to_pop_failed_realloc(array->frames());
1306 // Unlock all monitors here otherwise the interpreter will see a
1307 // mix of locked and unlocked monitors (because of failed
1308 // reallocations of synchronized objects) and be confused.
1309 for (int i = 0; i < array->frames(); i++) {
1310 MonitorChunk* monitors = array->element(i)->monitors();
1311 if (monitors != NULL) {
1312 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1313 BasicObjectLock* src = monitors->at(j);
1314 if (src->obj() != NULL) {
1315 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread);
1316 }
1317 }
1318 array->element(i)->free_monitors(thread);
1319 #ifdef ASSERT
1320 array->element(i)->set_removed_monitors();
1321 #endif
1322 }
1323 }
1324 }
1325 #endif
1326
1327 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1328 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1329 Thread* thread = Thread::current();
1330 for (int i = 0; i < monitors->length(); i++) {
1331 MonitorInfo* mon_info = monitors->at(i);
1332 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1333 objects_to_revoke->append(Handle(thread, mon_info->owner()));
1334 }
1335 }
1336 }
1337
1338
1339 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1340 if (!UseBiasedLocking) {
1341 return;
1342 }
1343
1344 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1345
1346 // Unfortunately we don't have a RegisterMap available in most of
1347 // the places we want to call this routine so we need to walk the
1348 // stack again to update the register map.
1349 if (map == NULL || !map->update_map()) {
1350 StackFrameStream sfs(thread, true);
1351 bool found = false;
1352 while (!found && !sfs.is_done()) {
1353 frame* cur = sfs.current();
1354 sfs.next();
1355 found = cur->id() == fr.id();
1356 }
1357 assert(found, "frame to be deoptimized not found on target thread's stack");
1358 map = sfs.register_map();
1359 }
1360
1361 vframe* vf = vframe::new_vframe(&fr, map, thread);
1362 compiledVFrame* cvf = compiledVFrame::cast(vf);
1363 // Revoke monitors' biases in all scopes
1364 while (!cvf->is_top()) {
1365 collect_monitors(cvf, objects_to_revoke);
1366 cvf = compiledVFrame::cast(cvf->sender());
1367 }
1368 collect_monitors(cvf, objects_to_revoke);
1369
1370 if (SafepointSynchronize::is_at_safepoint()) {
1371 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1372 } else {
1373 BiasedLocking::revoke(objects_to_revoke);
1374 }
1375 }
1376
1377
1378 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1379 if (!UseBiasedLocking) {
1380 return;
1381 }
1382
1383 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1384 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1385 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1386 if (jt->has_last_Java_frame()) {
1387 StackFrameStream sfs(jt, true);
1388 while (!sfs.is_done()) {
1389 frame* cur = sfs.current();
1390 if (cb->contains(cur->pc())) {
1391 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1392 compiledVFrame* cvf = compiledVFrame::cast(vf);
1393 // Revoke monitors' biases in all scopes
1394 while (!cvf->is_top()) {
1395 collect_monitors(cvf, objects_to_revoke);
1396 cvf = compiledVFrame::cast(cvf->sender());
1397 }
1398 collect_monitors(cvf, objects_to_revoke);
1399 }
1400 sfs.next();
1401 }
1402 }
1403 }
1404 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1405 }
1406
1407
1408 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1409 assert(fr.can_be_deoptimized(), "checking frame type");
1410
1411 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1412
1413 if (LogCompilation && xtty != NULL) {
1414 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1415 assert(cm != NULL, "only compiled methods can deopt");
1416
1417 ttyLocker ttyl;
1418 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1419 cm->log_identity(xtty);
1420 xtty->end_head();
1421 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1422 xtty->begin_elem("jvms bci='%d'", sd->bci());
1423 xtty->method(sd->method());
1424 xtty->end_elem();
1425 if (sd->is_top()) break;
1426 }
1427 xtty->tail("deoptimized");
1428 }
1429
1430 // Patch the compiled method so that when execution returns to it we will
1431 // deopt the execution state and return to the interpreter.
1432 fr.deoptimize(thread);
1433 }
1434
1435 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1436 deoptimize(thread, fr, map, Reason_constraint);
1437 }
1438
1439 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) {
1440 // Deoptimize only if the frame comes from compile code.
1441 // Do not deoptimize the frame which is already patched
1442 // during the execution of the loops below.
1443 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1444 return;
1445 }
1446 ResourceMark rm;
1447 DeoptimizationMarker dm;
1448 if (UseBiasedLocking) {
1449 revoke_biases_of_monitors(thread, fr, map);
1450 }
1451 deoptimize_single_frame(thread, fr, reason);
1452
1453 }
1454
1455
1456 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1457 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1458 "can only deoptimize other thread at a safepoint");
1459 // Compute frame and register map based on thread and sp.
1460 RegisterMap reg_map(thread, UseBiasedLocking);
1461 frame fr = thread->last_frame();
1462 while (fr.id() != id) {
1463 fr = fr.sender(®_map);
1464 }
1465 deoptimize(thread, fr, ®_map, reason);
1466 }
1467
1468
1469 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1470 if (thread == Thread::current()) {
1471 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1472 } else {
1473 VM_DeoptimizeFrame deopt(thread, id, reason);
1474 VMThread::execute(&deopt);
1475 }
1476 }
1477
1478 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1479 deoptimize_frame(thread, id, Reason_constraint);
1480 }
1481
1482 // JVMTI PopFrame support
1483 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1484 {
1485 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1486 }
1487 JRT_END
1488
1489 MethodData*
1490 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1491 bool create_if_missing) {
1492 Thread* THREAD = thread;
1493 MethodData* mdo = m()->method_data();
1494 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1495 // Build an MDO. Ignore errors like OutOfMemory;
1496 // that simply means we won't have an MDO to update.
1497 Method::build_interpreter_method_data(m, THREAD);
1498 if (HAS_PENDING_EXCEPTION) {
1499 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1500 CLEAR_PENDING_EXCEPTION;
1501 }
1502 mdo = m()->method_data();
1503 }
1504 return mdo;
1505 }
1506
1507 #if defined(COMPILER2) || defined(SHARK) || INCLUDE_JVMCI
1508 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1509 // in case of an unresolved klass entry, load the class.
1510 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1511 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1512 return;
1513 }
1514
1515 if (!constant_pool->tag_at(index).is_symbol()) return;
1516
1517 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1518 Symbol* symbol = constant_pool->symbol_at(index);
1519
1520 // class name?
1521 if (symbol->byte_at(0) != '(') {
1522 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1523 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1524 return;
1525 }
1526
1527 // then it must be a signature!
1528 ResourceMark rm(THREAD);
1529 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1530 if (ss.is_object()) {
1531 Symbol* class_name = ss.as_symbol(CHECK);
1532 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1533 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1534 }
1535 }
1536 }
1537
1538
1539 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1540 EXCEPTION_MARK;
1541 load_class_by_index(constant_pool, index, THREAD);
1542 if (HAS_PENDING_EXCEPTION) {
1543 // Exception happened during classloading. We ignore the exception here, since it
1544 // is going to be rethrown since the current activation is going to be deoptimized and
1545 // the interpreter will re-execute the bytecode.
1546 CLEAR_PENDING_EXCEPTION;
1547 // Class loading called java code which may have caused a stack
1548 // overflow. If the exception was thrown right before the return
1549 // to the runtime the stack is no longer guarded. Reguard the
1550 // stack otherwise if we return to the uncommon trap blob and the
1551 // stack bang causes a stack overflow we crash.
1552 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1553 JavaThread* thread = (JavaThread*)THREAD;
1554 bool guard_pages_enabled = thread->stack_guards_enabled();
1555 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1556 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1557 }
1558 }
1559
1560 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1561 HandleMark hm;
1562
1563 // uncommon_trap() is called at the beginning of the uncommon trap
1564 // handler. Note this fact before we start generating temporary frames
1565 // that can confuse an asynchronous stack walker. This counter is
1566 // decremented at the end of unpack_frames().
1567 thread->inc_in_deopt_handler();
1568
1569 // We need to update the map if we have biased locking.
1570 #if INCLUDE_JVMCI
1571 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1572 RegisterMap reg_map(thread, true);
1573 #else
1574 RegisterMap reg_map(thread, UseBiasedLocking);
1575 #endif
1576 frame stub_frame = thread->last_frame();
1577 frame fr = stub_frame.sender(®_map);
1578 // Make sure the calling nmethod is not getting deoptimized and removed
1579 // before we are done with it.
1580 nmethodLocker nl(fr.pc());
1581
1582 // Log a message
1583 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1584 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1585
1586 {
1587 ResourceMark rm;
1588
1589 // Revoke biases of any monitors in the frame to ensure we can migrate them
1590 revoke_biases_of_monitors(thread, fr, ®_map);
1591
1592 DeoptReason reason = trap_request_reason(trap_request);
1593 DeoptAction action = trap_request_action(trap_request);
1594 #if INCLUDE_JVMCI
1595 int debug_id = trap_request_debug_id(trap_request);
1596 #endif
1597 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1598
1599 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1600 compiledVFrame* cvf = compiledVFrame::cast(vf);
1601
1602 CompiledMethod* nm = cvf->code();
1603
1604 ScopeDesc* trap_scope = cvf->scope();
1605
1606 if (TraceDeoptimization) {
1607 ttyLocker ttyl;
1608 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()
1609 #if INCLUDE_JVMCI
1610 , debug_id
1611 #endif
1612 );
1613 }
1614
1615 methodHandle trap_method = trap_scope->method();
1616 int trap_bci = trap_scope->bci();
1617 #if INCLUDE_JVMCI
1618 oop speculation = thread->pending_failed_speculation();
1619 if (nm->is_compiled_by_jvmci()) {
1620 if (speculation != NULL) {
1621 oop speculation_log = nm->as_nmethod()->speculation_log();
1622 if (speculation_log != NULL) {
1623 if (TraceDeoptimization || TraceUncollectedSpeculations) {
1624 if (HotSpotSpeculationLog::lastFailed(speculation_log) != NULL) {
1625 tty->print_cr("A speculation that was not collected by the compiler is being overwritten");
1626 }
1627 }
1628 if (TraceDeoptimization) {
1629 tty->print_cr("Saving speculation to speculation log");
1630 }
1631 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation);
1632 } else {
1633 if (TraceDeoptimization) {
1634 tty->print_cr("Speculation present but no speculation log");
1635 }
1636 }
1637 thread->set_pending_failed_speculation(NULL);
1638 } else {
1639 if (TraceDeoptimization) {
1640 tty->print_cr("No speculation");
1641 }
1642 }
1643 } else {
1644 assert(speculation == NULL, "There should not be a speculation for method compiled by non-JVMCI compilers");
1645 }
1646
1647 if (trap_bci == SynchronizationEntryBCI) {
1648 trap_bci = 0;
1649 thread->set_pending_monitorenter(true);
1650 }
1651
1652 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1653 thread->set_pending_transfer_to_interpreter(true);
1654 }
1655 #endif
1656
1657 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1658 // Record this event in the histogram.
1659 gather_statistics(reason, action, trap_bc);
1660
1661 // Ensure that we can record deopt. history:
1662 // Need MDO to record RTM code generation state.
1663 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1664
1665 methodHandle profiled_method;
1666 #if INCLUDE_JVMCI
1667 if (nm->is_compiled_by_jvmci()) {
1668 profiled_method = nm->method();
1669 } else {
1670 profiled_method = trap_method;
1671 }
1672 #else
1673 profiled_method = trap_method;
1674 #endif
1675
1676 MethodData* trap_mdo =
1677 get_method_data(thread, profiled_method, create_if_missing);
1678
1679 // Log a message
1680 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
1681 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1682 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
1683
1684 // Print a bunch of diagnostics, if requested.
1685 if (TraceDeoptimization || LogCompilation) {
1686 ResourceMark rm;
1687 ttyLocker ttyl;
1688 char buf[100];
1689 if (xtty != NULL) {
1690 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1691 os::current_thread_id(),
1692 format_trap_request(buf, sizeof(buf), trap_request));
1693 nm->log_identity(xtty);
1694 }
1695 Symbol* class_name = NULL;
1696 bool unresolved = false;
1697 if (unloaded_class_index >= 0) {
1698 constantPoolHandle constants (THREAD, trap_method->constants());
1699 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1700 class_name = constants->klass_name_at(unloaded_class_index);
1701 unresolved = true;
1702 if (xtty != NULL)
1703 xtty->print(" unresolved='1'");
1704 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1705 class_name = constants->symbol_at(unloaded_class_index);
1706 }
1707 if (xtty != NULL)
1708 xtty->name(class_name);
1709 }
1710 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
1711 // Dump the relevant MDO state.
1712 // This is the deopt count for the current reason, any previous
1713 // reasons or recompiles seen at this point.
1714 int dcnt = trap_mdo->trap_count(reason);
1715 if (dcnt != 0)
1716 xtty->print(" count='%d'", dcnt);
1717 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1718 int dos = (pdata == NULL)? 0: pdata->trap_state();
1719 if (dos != 0) {
1720 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1721 if (trap_state_is_recompiled(dos)) {
1722 int recnt2 = trap_mdo->overflow_recompile_count();
1723 if (recnt2 != 0)
1724 xtty->print(" recompiles2='%d'", recnt2);
1725 }
1726 }
1727 }
1728 if (xtty != NULL) {
1729 xtty->stamp();
1730 xtty->end_head();
1731 }
1732 if (TraceDeoptimization) { // make noise on the tty
1733 tty->print("Uncommon trap occurred in");
1734 nm->method()->print_short_name(tty);
1735 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
1736 #if INCLUDE_JVMCI
1737 if (nm->is_nmethod()) {
1738 oop installedCode = nm->as_nmethod()->jvmci_installed_code();
1739 if (installedCode != NULL) {
1740 oop installedCodeName = NULL;
1741 if (installedCode->is_a(InstalledCode::klass())) {
1742 installedCodeName = InstalledCode::name(installedCode);
1743 }
1744 if (installedCodeName != NULL) {
1745 tty->print(" (JVMCI: installedCodeName=%s) ", java_lang_String::as_utf8_string(installedCodeName));
1746 } else {
1747 tty->print(" (JVMCI: installed code has no name) ");
1748 }
1749 } else if (nm->is_compiled_by_jvmci()) {
1750 tty->print(" (JVMCI: no installed code) ");
1751 }
1752 }
1753 #endif
1754 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
1755 p2i(fr.pc()),
1756 os::current_thread_id(),
1757 trap_reason_name(reason),
1758 trap_action_name(action),
1759 unloaded_class_index
1760 #if INCLUDE_JVMCI
1761 , debug_id
1762 #endif
1763 );
1764 if (class_name != NULL) {
1765 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1766 class_name->print_symbol_on(tty);
1767 }
1768 tty->cr();
1769 }
1770 if (xtty != NULL) {
1771 // Log the precise location of the trap.
1772 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1773 xtty->begin_elem("jvms bci='%d'", sd->bci());
1774 xtty->method(sd->method());
1775 xtty->end_elem();
1776 if (sd->is_top()) break;
1777 }
1778 xtty->tail("uncommon_trap");
1779 }
1780 }
1781 // (End diagnostic printout.)
1782
1783 // Load class if necessary
1784 if (unloaded_class_index >= 0) {
1785 constantPoolHandle constants(THREAD, trap_method->constants());
1786 load_class_by_index(constants, unloaded_class_index);
1787 }
1788
1789 // Flush the nmethod if necessary and desirable.
1790 //
1791 // We need to avoid situations where we are re-flushing the nmethod
1792 // because of a hot deoptimization site. Repeated flushes at the same
1793 // point need to be detected by the compiler and avoided. If the compiler
1794 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1795 // module must take measures to avoid an infinite cycle of recompilation
1796 // and deoptimization. There are several such measures:
1797 //
1798 // 1. If a recompilation is ordered a second time at some site X
1799 // and for the same reason R, the action is adjusted to 'reinterpret',
1800 // to give the interpreter time to exercise the method more thoroughly.
1801 // If this happens, the method's overflow_recompile_count is incremented.
1802 //
1803 // 2. If the compiler fails to reduce the deoptimization rate, then
1804 // the method's overflow_recompile_count will begin to exceed the set
1805 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1806 // is adjusted to 'make_not_compilable', and the method is abandoned
1807 // to the interpreter. This is a performance hit for hot methods,
1808 // but is better than a disastrous infinite cycle of recompilations.
1809 // (Actually, only the method containing the site X is abandoned.)
1810 //
1811 // 3. In parallel with the previous measures, if the total number of
1812 // recompilations of a method exceeds the much larger set limit
1813 // PerMethodRecompilationCutoff, the method is abandoned.
1814 // This should only happen if the method is very large and has
1815 // many "lukewarm" deoptimizations. The code which enforces this
1816 // limit is elsewhere (class nmethod, class Method).
1817 //
1818 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1819 // to recompile at each bytecode independently of the per-BCI cutoff.
1820 //
1821 // The decision to update code is up to the compiler, and is encoded
1822 // in the Action_xxx code. If the compiler requests Action_none
1823 // no trap state is changed, no compiled code is changed, and the
1824 // computation suffers along in the interpreter.
1825 //
1826 // The other action codes specify various tactics for decompilation
1827 // and recompilation. Action_maybe_recompile is the loosest, and
1828 // allows the compiled code to stay around until enough traps are seen,
1829 // and until the compiler gets around to recompiling the trapping method.
1830 //
1831 // The other actions cause immediate removal of the present code.
1832
1833 // Traps caused by injected profile shouldn't pollute trap counts.
1834 bool injected_profile_trap = trap_method->has_injected_profile() &&
1835 (reason == Reason_intrinsic || reason == Reason_unreached);
1836
1837 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
1838 bool make_not_entrant = false;
1839 bool make_not_compilable = false;
1840 bool reprofile = false;
1841 switch (action) {
1842 case Action_none:
1843 // Keep the old code.
1844 update_trap_state = false;
1845 break;
1846 case Action_maybe_recompile:
1847 // Do not need to invalidate the present code, but we can
1848 // initiate another
1849 // Start compiler without (necessarily) invalidating the nmethod.
1850 // The system will tolerate the old code, but new code should be
1851 // generated when possible.
1852 break;
1853 case Action_reinterpret:
1854 // Go back into the interpreter for a while, and then consider
1855 // recompiling form scratch.
1856 make_not_entrant = true;
1857 // Reset invocation counter for outer most method.
1858 // This will allow the interpreter to exercise the bytecodes
1859 // for a while before recompiling.
1860 // By contrast, Action_make_not_entrant is immediate.
1861 //
1862 // Note that the compiler will track null_check, null_assert,
1863 // range_check, and class_check events and log them as if they
1864 // had been traps taken from compiled code. This will update
1865 // the MDO trap history so that the next compilation will
1866 // properly detect hot trap sites.
1867 reprofile = true;
1868 break;
1869 case Action_make_not_entrant:
1870 // Request immediate recompilation, and get rid of the old code.
1871 // Make them not entrant, so next time they are called they get
1872 // recompiled. Unloaded classes are loaded now so recompile before next
1873 // time they are called. Same for uninitialized. The interpreter will
1874 // link the missing class, if any.
1875 make_not_entrant = true;
1876 break;
1877 case Action_make_not_compilable:
1878 // Give up on compiling this method at all.
1879 make_not_entrant = true;
1880 make_not_compilable = true;
1881 break;
1882 default:
1883 ShouldNotReachHere();
1884 }
1885
1886 // Setting +ProfileTraps fixes the following, on all platforms:
1887 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1888 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1889 // recompile relies on a MethodData* to record heroic opt failures.
1890
1891 // Whether the interpreter is producing MDO data or not, we also need
1892 // to use the MDO to detect hot deoptimization points and control
1893 // aggressive optimization.
1894 bool inc_recompile_count = false;
1895 ProfileData* pdata = NULL;
1896 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) {
1897 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
1898 uint this_trap_count = 0;
1899 bool maybe_prior_trap = false;
1900 bool maybe_prior_recompile = false;
1901 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
1902 #if INCLUDE_JVMCI
1903 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
1904 #endif
1905 nm->method(),
1906 //outputs:
1907 this_trap_count,
1908 maybe_prior_trap,
1909 maybe_prior_recompile);
1910 // Because the interpreter also counts null, div0, range, and class
1911 // checks, these traps from compiled code are double-counted.
1912 // This is harmless; it just means that the PerXTrapLimit values
1913 // are in effect a little smaller than they look.
1914
1915 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1916 if (per_bc_reason != Reason_none) {
1917 // Now take action based on the partially known per-BCI history.
1918 if (maybe_prior_trap
1919 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1920 // If there are too many traps at this BCI, force a recompile.
1921 // This will allow the compiler to see the limit overflow, and
1922 // take corrective action, if possible. The compiler generally
1923 // does not use the exact PerBytecodeTrapLimit value, but instead
1924 // changes its tactics if it sees any traps at all. This provides
1925 // a little hysteresis, delaying a recompile until a trap happens
1926 // several times.
1927 //
1928 // Actually, since there is only one bit of counter per BCI,
1929 // the possible per-BCI counts are {0,1,(per-method count)}.
1930 // This produces accurate results if in fact there is only
1931 // one hot trap site, but begins to get fuzzy if there are
1932 // many sites. For example, if there are ten sites each
1933 // trapping two or more times, they each get the blame for
1934 // all of their traps.
1935 make_not_entrant = true;
1936 }
1937
1938 // Detect repeated recompilation at the same BCI, and enforce a limit.
1939 if (make_not_entrant && maybe_prior_recompile) {
1940 // More than one recompile at this point.
1941 inc_recompile_count = maybe_prior_trap;
1942 }
1943 } else {
1944 // For reasons which are not recorded per-bytecode, we simply
1945 // force recompiles unconditionally.
1946 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1947 make_not_entrant = true;
1948 }
1949
1950 // Go back to the compiler if there are too many traps in this method.
1951 if (this_trap_count >= per_method_trap_limit(reason)) {
1952 // If there are too many traps in this method, force a recompile.
1953 // This will allow the compiler to see the limit overflow, and
1954 // take corrective action, if possible.
1955 // (This condition is an unlikely backstop only, because the
1956 // PerBytecodeTrapLimit is more likely to take effect first,
1957 // if it is applicable.)
1958 make_not_entrant = true;
1959 }
1960
1961 // Here's more hysteresis: If there has been a recompile at
1962 // this trap point already, run the method in the interpreter
1963 // for a while to exercise it more thoroughly.
1964 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1965 reprofile = true;
1966 }
1967 }
1968
1969 // Take requested actions on the method:
1970
1971 // Recompile
1972 if (make_not_entrant) {
1973 if (!nm->make_not_entrant()) {
1974 return; // the call did not change nmethod's state
1975 }
1976
1977 if (pdata != NULL) {
1978 // Record the recompilation event, if any.
1979 int tstate0 = pdata->trap_state();
1980 int tstate1 = trap_state_set_recompiled(tstate0, true);
1981 if (tstate1 != tstate0)
1982 pdata->set_trap_state(tstate1);
1983 }
1984
1985 #if INCLUDE_RTM_OPT
1986 // Restart collecting RTM locking abort statistic if the method
1987 // is recompiled for a reason other than RTM state change.
1988 // Assume that in new recompiled code the statistic could be different,
1989 // for example, due to different inlining.
1990 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
1991 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
1992 trap_mdo->atomic_set_rtm_state(ProfileRTM);
1993 }
1994 #endif
1995 // For code aging we count traps separately here, using make_not_entrant()
1996 // as a guard against simultaneous deopts in multiple threads.
1997 if (reason == Reason_tenured && trap_mdo != NULL) {
1998 trap_mdo->inc_tenure_traps();
1999 }
2000 }
2001
2002 if (inc_recompile_count) {
2003 trap_mdo->inc_overflow_recompile_count();
2004 if ((uint)trap_mdo->overflow_recompile_count() >
2005 (uint)PerBytecodeRecompilationCutoff) {
2006 // Give up on the method containing the bad BCI.
2007 if (trap_method() == nm->method()) {
2008 make_not_compilable = true;
2009 } else {
2010 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
2011 // But give grace to the enclosing nm->method().
2012 }
2013 }
2014 }
2015
2016 // Reprofile
2017 if (reprofile) {
2018 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
2019 }
2020
2021 // Give up compiling
2022 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2023 assert(make_not_entrant, "consistent");
2024 nm->method()->set_not_compilable(CompLevel_full_optimization);
2025 }
2026
2027 } // Free marked resources
2028
2029 }
2030 JRT_END
2031
2032 ProfileData*
2033 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2034 int trap_bci,
2035 Deoptimization::DeoptReason reason,
2036 bool update_total_trap_count,
2037 #if INCLUDE_JVMCI
2038 bool is_osr,
2039 #endif
2040 Method* compiled_method,
2041 //outputs:
2042 uint& ret_this_trap_count,
2043 bool& ret_maybe_prior_trap,
2044 bool& ret_maybe_prior_recompile) {
2045 bool maybe_prior_trap = false;
2046 bool maybe_prior_recompile = false;
2047 uint this_trap_count = 0;
2048 if (update_total_trap_count) {
2049 uint idx = reason;
2050 #if INCLUDE_JVMCI
2051 if (is_osr) {
2052 idx += Reason_LIMIT;
2053 }
2054 #endif
2055 uint prior_trap_count = trap_mdo->trap_count(idx);
2056 this_trap_count = trap_mdo->inc_trap_count(idx);
2057
2058 // If the runtime cannot find a place to store trap history,
2059 // it is estimated based on the general condition of the method.
2060 // If the method has ever been recompiled, or has ever incurred
2061 // a trap with the present reason , then this BCI is assumed
2062 // (pessimistically) to be the culprit.
2063 maybe_prior_trap = (prior_trap_count != 0);
2064 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2065 }
2066 ProfileData* pdata = NULL;
2067
2068
2069 // For reasons which are recorded per bytecode, we check per-BCI data.
2070 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2071 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2072 if (per_bc_reason != Reason_none) {
2073 // Find the profile data for this BCI. If there isn't one,
2074 // try to allocate one from the MDO's set of spares.
2075 // This will let us detect a repeated trap at this point.
2076 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
2077
2078 if (pdata != NULL) {
2079 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2080 if (LogCompilation && xtty != NULL) {
2081 ttyLocker ttyl;
2082 // no more room for speculative traps in this MDO
2083 xtty->elem("speculative_traps_oom");
2084 }
2085 }
2086 // Query the trap state of this profile datum.
2087 int tstate0 = pdata->trap_state();
2088 if (!trap_state_has_reason(tstate0, per_bc_reason))
2089 maybe_prior_trap = false;
2090 if (!trap_state_is_recompiled(tstate0))
2091 maybe_prior_recompile = false;
2092
2093 // Update the trap state of this profile datum.
2094 int tstate1 = tstate0;
2095 // Record the reason.
2096 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2097 // Store the updated state on the MDO, for next time.
2098 if (tstate1 != tstate0)
2099 pdata->set_trap_state(tstate1);
2100 } else {
2101 if (LogCompilation && xtty != NULL) {
2102 ttyLocker ttyl;
2103 // Missing MDP? Leave a small complaint in the log.
2104 xtty->elem("missing_mdp bci='%d'", trap_bci);
2105 }
2106 }
2107 }
2108
2109 // Return results:
2110 ret_this_trap_count = this_trap_count;
2111 ret_maybe_prior_trap = maybe_prior_trap;
2112 ret_maybe_prior_recompile = maybe_prior_recompile;
2113 return pdata;
2114 }
2115
2116 void
2117 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2118 ResourceMark rm;
2119 // Ignored outputs:
2120 uint ignore_this_trap_count;
2121 bool ignore_maybe_prior_trap;
2122 bool ignore_maybe_prior_recompile;
2123 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2124 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2125 bool update_total_counts = JVMCI_ONLY(false) NOT_JVMCI(true);
2126 query_update_method_data(trap_mdo, trap_bci,
2127 (DeoptReason)reason,
2128 update_total_counts,
2129 #if INCLUDE_JVMCI
2130 false,
2131 #endif
2132 NULL,
2133 ignore_this_trap_count,
2134 ignore_maybe_prior_trap,
2135 ignore_maybe_prior_recompile);
2136 }
2137
2138 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2139 if (TraceDeoptimization) {
2140 tty->print("Uncommon trap ");
2141 }
2142 // Still in Java no safepoints
2143 {
2144 // This enters VM and may safepoint
2145 uncommon_trap_inner(thread, trap_request);
2146 }
2147 return fetch_unroll_info_helper(thread, exec_mode);
2148 }
2149
2150 // Local derived constants.
2151 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2152 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
2153 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2154
2155 //---------------------------trap_state_reason---------------------------------
2156 Deoptimization::DeoptReason
2157 Deoptimization::trap_state_reason(int trap_state) {
2158 // This assert provides the link between the width of DataLayout::trap_bits
2159 // and the encoding of "recorded" reasons. It ensures there are enough
2160 // bits to store all needed reasons in the per-BCI MDO profile.
2161 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2162 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2163 trap_state -= recompile_bit;
2164 if (trap_state == DS_REASON_MASK) {
2165 return Reason_many;
2166 } else {
2167 assert((int)Reason_none == 0, "state=0 => Reason_none");
2168 return (DeoptReason)trap_state;
2169 }
2170 }
2171 //-------------------------trap_state_has_reason-------------------------------
2172 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2173 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2174 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2175 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2176 trap_state -= recompile_bit;
2177 if (trap_state == DS_REASON_MASK) {
2178 return -1; // true, unspecifically (bottom of state lattice)
2179 } else if (trap_state == reason) {
2180 return 1; // true, definitely
2181 } else if (trap_state == 0) {
2182 return 0; // false, definitely (top of state lattice)
2183 } else {
2184 return 0; // false, definitely
2185 }
2186 }
2187 //-------------------------trap_state_add_reason-------------------------------
2188 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2189 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2190 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2191 trap_state -= recompile_bit;
2192 if (trap_state == DS_REASON_MASK) {
2193 return trap_state + recompile_bit; // already at state lattice bottom
2194 } else if (trap_state == reason) {
2195 return trap_state + recompile_bit; // the condition is already true
2196 } else if (trap_state == 0) {
2197 return reason + recompile_bit; // no condition has yet been true
2198 } else {
2199 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2200 }
2201 }
2202 //-----------------------trap_state_is_recompiled------------------------------
2203 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2204 return (trap_state & DS_RECOMPILE_BIT) != 0;
2205 }
2206 //-----------------------trap_state_set_recompiled-----------------------------
2207 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2208 if (z) return trap_state | DS_RECOMPILE_BIT;
2209 else return trap_state & ~DS_RECOMPILE_BIT;
2210 }
2211 //---------------------------format_trap_state---------------------------------
2212 // This is used for debugging and diagnostics, including LogFile output.
2213 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2214 int trap_state) {
2215 assert(buflen > 0, "sanity");
2216 DeoptReason reason = trap_state_reason(trap_state);
2217 bool recomp_flag = trap_state_is_recompiled(trap_state);
2218 // Re-encode the state from its decoded components.
2219 int decoded_state = 0;
2220 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2221 decoded_state = trap_state_add_reason(decoded_state, reason);
2222 if (recomp_flag)
2223 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2224 // If the state re-encodes properly, format it symbolically.
2225 // Because this routine is used for debugging and diagnostics,
2226 // be robust even if the state is a strange value.
2227 size_t len;
2228 if (decoded_state != trap_state) {
2229 // Random buggy state that doesn't decode??
2230 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2231 } else {
2232 len = jio_snprintf(buf, buflen, "%s%s",
2233 trap_reason_name(reason),
2234 recomp_flag ? " recompiled" : "");
2235 }
2236 return buf;
2237 }
2238
2239
2240 //--------------------------------statics--------------------------------------
2241 const char* Deoptimization::_trap_reason_name[] = {
2242 // Note: Keep this in sync. with enum DeoptReason.
2243 "none",
2244 "null_check",
2245 "null_assert" JVMCI_ONLY("_or_unreached0"),
2246 "range_check",
2247 "class_check",
2248 "array_check",
2249 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2250 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2251 "unloaded",
2252 "uninitialized",
2253 "unreached",
2254 "unhandled",
2255 "constraint",
2256 "div0_check",
2257 "age",
2258 "predicate",
2259 "loop_limit_check",
2260 "speculate_class_check",
2261 "speculate_null_check",
2262 "rtm_state_change",
2263 "unstable_if",
2264 "unstable_fused_if",
2265 #if INCLUDE_JVMCI
2266 "aliasing",
2267 "transfer_to_interpreter",
2268 "not_compiled_exception_handler",
2269 "unresolved",
2270 "jsr_mismatch",
2271 #endif
2272 "tenured"
2273 };
2274 const char* Deoptimization::_trap_action_name[] = {
2275 // Note: Keep this in sync. with enum DeoptAction.
2276 "none",
2277 "maybe_recompile",
2278 "reinterpret",
2279 "make_not_entrant",
2280 "make_not_compilable"
2281 };
2282
2283 const char* Deoptimization::trap_reason_name(int reason) {
2284 // Check that every reason has a name
2285 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2286
2287 if (reason == Reason_many) return "many";
2288 if ((uint)reason < Reason_LIMIT)
2289 return _trap_reason_name[reason];
2290 static char buf[20];
2291 sprintf(buf, "reason%d", reason);
2292 return buf;
2293 }
2294 const char* Deoptimization::trap_action_name(int action) {
2295 // Check that every action has a name
2296 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2297
2298 if ((uint)action < Action_LIMIT)
2299 return _trap_action_name[action];
2300 static char buf[20];
2301 sprintf(buf, "action%d", action);
2302 return buf;
2303 }
2304
2305 // This is used for debugging and diagnostics, including LogFile output.
2306 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2307 int trap_request) {
2308 jint unloaded_class_index = trap_request_index(trap_request);
2309 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2310 const char* action = trap_action_name(trap_request_action(trap_request));
2311 #if INCLUDE_JVMCI
2312 int debug_id = trap_request_debug_id(trap_request);
2313 #endif
2314 size_t len;
2315 if (unloaded_class_index < 0) {
2316 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2317 reason, action
2318 #if INCLUDE_JVMCI
2319 ,debug_id
2320 #endif
2321 );
2322 } else {
2323 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2324 reason, action, unloaded_class_index
2325 #if INCLUDE_JVMCI
2326 ,debug_id
2327 #endif
2328 );
2329 }
2330 return buf;
2331 }
2332
2333 juint Deoptimization::_deoptimization_hist
2334 [Deoptimization::Reason_LIMIT]
2335 [1 + Deoptimization::Action_LIMIT]
2336 [Deoptimization::BC_CASE_LIMIT]
2337 = {0};
2338
2339 enum {
2340 LSB_BITS = 8,
2341 LSB_MASK = right_n_bits(LSB_BITS)
2342 };
2343
2344 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2345 Bytecodes::Code bc) {
2346 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2347 assert(action >= 0 && action < Action_LIMIT, "oob");
2348 _deoptimization_hist[Reason_none][0][0] += 1; // total
2349 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2350 juint* cases = _deoptimization_hist[reason][1+action];
2351 juint* bc_counter_addr = NULL;
2352 juint bc_counter = 0;
2353 // Look for an unused counter, or an exact match to this BC.
2354 if (bc != Bytecodes::_illegal) {
2355 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2356 juint* counter_addr = &cases[bc_case];
2357 juint counter = *counter_addr;
2358 if ((counter == 0 && bc_counter_addr == NULL)
2359 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2360 // this counter is either free or is already devoted to this BC
2361 bc_counter_addr = counter_addr;
2362 bc_counter = counter | bc;
2363 }
2364 }
2365 }
2366 if (bc_counter_addr == NULL) {
2367 // Overflow, or no given bytecode.
2368 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2369 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2370 }
2371 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2372 }
2373
2374 jint Deoptimization::total_deoptimization_count() {
2375 return _deoptimization_hist[Reason_none][0][0];
2376 }
2377
2378 jint Deoptimization::deoptimization_count(DeoptReason reason) {
2379 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2380 return _deoptimization_hist[reason][0][0];
2381 }
2382
2383 void Deoptimization::print_statistics() {
2384 juint total = total_deoptimization_count();
2385 juint account = total;
2386 if (total != 0) {
2387 ttyLocker ttyl;
2388 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2389 tty->print_cr("Deoptimization traps recorded:");
2390 #define PRINT_STAT_LINE(name, r) \
2391 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2392 PRINT_STAT_LINE("total", total);
2393 // For each non-zero entry in the histogram, print the reason,
2394 // the action, and (if specifically known) the type of bytecode.
2395 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2396 for (int action = 0; action < Action_LIMIT; action++) {
2397 juint* cases = _deoptimization_hist[reason][1+action];
2398 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2399 juint counter = cases[bc_case];
2400 if (counter != 0) {
2401 char name[1*K];
2402 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2403 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2404 bc = Bytecodes::_illegal;
2405 sprintf(name, "%s/%s/%s",
2406 trap_reason_name(reason),
2407 trap_action_name(action),
2408 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2409 juint r = counter >> LSB_BITS;
2410 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2411 account -= r;
2412 }
2413 }
2414 }
2415 }
2416 if (account != 0) {
2417 PRINT_STAT_LINE("unaccounted", account);
2418 }
2419 #undef PRINT_STAT_LINE
2420 if (xtty != NULL) xtty->tail("statistics");
2421 }
2422 }
2423 #else // COMPILER2 || SHARK || INCLUDE_JVMCI
2424
2425
2426 // Stubs for C1 only system.
2427 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2428 return false;
2429 }
2430
2431 const char* Deoptimization::trap_reason_name(int reason) {
2432 return "unknown";
2433 }
2434
2435 void Deoptimization::print_statistics() {
2436 // no output
2437 }
2438
2439 void
2440 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2441 // no udpate
2442 }
2443
2444 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2445 return 0;
2446 }
2447
2448 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2449 Bytecodes::Code bc) {
2450 // no update
2451 }
2452
2453 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2454 int trap_state) {
2455 jio_snprintf(buf, buflen, "#%d", trap_state);
2456 return buf;
2457 }
2458
2459 #endif // COMPILER2 || SHARK || INCLUDE_JVMCI