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