1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "jvm.h"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/codeCache.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "code/nmethod.hpp"
31 #include "code/pcDesc.hpp"
32 #include "code/scopeDesc.hpp"
33 #include "interpreter/bytecode.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "interpreter/oopMapCache.hpp"
36 #include "memory/allocation.inline.hpp"
37 #include "memory/oopFactory.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "oops/method.hpp"
40 #include "oops/objArrayOop.inline.hpp"
41 #include "oops/oop.inline.hpp"
42 #include "oops/fieldStreams.hpp"
43 #include "oops/typeArrayOop.inline.hpp"
44 #include "oops/verifyOopClosure.hpp"
45 #include "prims/jvmtiThreadState.hpp"
46 #include "runtime/biasedLocking.hpp"
47 #include "runtime/compilationPolicy.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/frame.inline.hpp"
50 #include "runtime/interfaceSupport.inline.hpp"
51 #include "runtime/safepointVerifiers.hpp"
52 #include "runtime/sharedRuntime.hpp"
53 #include "runtime/signature.hpp"
54 #include "runtime/stubRoutines.hpp"
55 #include "runtime/thread.hpp"
56 #include "runtime/threadSMR.hpp"
57 #include "runtime/vframe.hpp"
58 #include "runtime/vframeArray.hpp"
59 #include "runtime/vframe_hp.hpp"
60 #include "utilities/events.hpp"
61 #include "utilities/preserveException.hpp"
62 #include "utilities/xmlstream.hpp"
63
64 #if INCLUDE_JVMCI
65 #include "jvmci/jvmciRuntime.hpp"
66 #include "jvmci/jvmciJavaClasses.hpp"
67 #endif
68
69
70 bool DeoptimizationMarker::_is_active = false;
71
72 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
73 int caller_adjustment,
74 int caller_actual_parameters,
75 int number_of_frames,
76 intptr_t* frame_sizes,
77 address* frame_pcs,
78 BasicType return_type,
79 int exec_mode) {
80 _size_of_deoptimized_frame = size_of_deoptimized_frame;
81 _caller_adjustment = caller_adjustment;
82 _caller_actual_parameters = caller_actual_parameters;
83 _number_of_frames = number_of_frames;
84 _frame_sizes = frame_sizes;
85 _frame_pcs = frame_pcs;
86 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
87 _return_type = return_type;
88 _initial_info = 0;
89 // PD (x86 only)
90 _counter_temp = 0;
91 _unpack_kind = exec_mode;
92 _sender_sp_temp = 0;
93
94 _total_frame_sizes = size_of_frames();
95 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
96 }
97
98
99 Deoptimization::UnrollBlock::~UnrollBlock() {
100 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
101 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
102 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
103 }
104
105
106 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
107 assert(register_number < RegisterMap::reg_count, "checking register number");
108 return &_register_block[register_number * 2];
109 }
110
111
112
113 int Deoptimization::UnrollBlock::size_of_frames() const {
114 // Acount first for the adjustment of the initial frame
115 int result = _caller_adjustment;
116 for (int index = 0; index < number_of_frames(); index++) {
117 result += frame_sizes()[index];
118 }
119 return result;
120 }
121
122
123 void Deoptimization::UnrollBlock::print() {
124 ttyLocker ttyl;
125 tty->print_cr("UnrollBlock");
126 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
127 tty->print( " frame_sizes: ");
128 for (int index = 0; index < number_of_frames(); index++) {
129 tty->print(INTX_FORMAT " ", frame_sizes()[index]);
130 }
131 tty->cr();
132 }
133
134
135 // In order to make fetch_unroll_info work properly with escape
136 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
137 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
138 // of previously eliminated objects occurs in realloc_objects, which is
139 // called from the method fetch_unroll_info_helper below.
140 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode))
141 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
142 // but makes the entry a little slower. There is however a little dance we have to
143 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
144
145 // fetch_unroll_info() is called at the beginning of the deoptimization
146 // handler. Note this fact before we start generating temporary frames
147 // that can confuse an asynchronous stack walker. This counter is
148 // decremented at the end of unpack_frames().
149 if (TraceDeoptimization) {
150 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread));
151 }
152 thread->inc_in_deopt_handler();
153
154 return fetch_unroll_info_helper(thread, exec_mode);
155 JRT_END
156
157
158 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
159 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) {
160
161 // Note: there is a safepoint safety issue here. No matter whether we enter
162 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
163 // the vframeArray is created.
164 //
165
166 // Allocate our special deoptimization ResourceMark
167 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
168 assert(thread->deopt_mark() == NULL, "Pending deopt!");
169 thread->set_deopt_mark(dmark);
170
171 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
172 RegisterMap map(thread, true);
173 RegisterMap dummy_map(thread, false);
174 // Now get the deoptee with a valid map
175 frame deoptee = stub_frame.sender(&map);
176 // Set the deoptee nmethod
177 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!");
178 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
179 thread->set_deopt_compiled_method(cm);
180
181 if (VerifyStack) {
182 thread->validate_frame_layout();
183 }
184
185 // Create a growable array of VFrames where each VFrame represents an inlined
186 // Java frame. This storage is allocated with the usual system arena.
187 assert(deoptee.is_compiled_frame(), "Wrong frame type");
188 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
189 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
190 while (!vf->is_top()) {
191 assert(vf->is_compiled_frame(), "Wrong frame type");
192 chunk->push(compiledVFrame::cast(vf));
193 vf = vf->sender();
194 }
195 assert(vf->is_compiled_frame(), "Wrong frame type");
196 chunk->push(compiledVFrame::cast(vf));
197
198 bool realloc_failures = false;
199
200 #if COMPILER2_OR_JVMCI
201 // Reallocate the non-escaping objects and restore their fields. Then
202 // relock objects if synchronization on them was eliminated.
203 #if !INCLUDE_JVMCI
204 if (DoEscapeAnalysis || EliminateNestedLocks) {
205 if (EliminateAllocations) {
206 #endif // INCLUDE_JVMCI
207 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
208 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
209
210 // The flag return_oop() indicates call sites which return oop
211 // in compiled code. Such sites include java method calls,
212 // runtime calls (for example, used to allocate new objects/arrays
213 // on slow code path) and any other calls generated in compiled code.
214 // It is not guaranteed that we can get such information here only
215 // by analyzing bytecode in deoptimized frames. This is why this flag
216 // is set during method compilation (see Compile::Process_OopMap_Node()).
217 // If the previous frame was popped or if we are dispatching an exception,
218 // we don't have an oop result.
219 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt);
220 Handle return_value;
221 if (save_oop_result) {
222 // Reallocation may trigger GC. If deoptimization happened on return from
223 // call which returns oop we need to save it since it is not in oopmap.
224 oop result = deoptee.saved_oop_result(&map);
225 assert(oopDesc::is_oop_or_null(result), "must be oop");
226 return_value = Handle(thread, result);
227 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
228 if (TraceDeoptimization) {
229 ttyLocker ttyl;
230 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
231 }
232 }
233 if (objects != NULL) {
234 JRT_BLOCK
235 realloc_failures = realloc_objects(thread, &deoptee, objects, THREAD);
236 JRT_END
237 bool skip_internal = (cm != NULL) && !cm->is_compiled_by_jvmci();
238 reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
239 #ifndef PRODUCT
240 if (TraceDeoptimization) {
241 ttyLocker ttyl;
242 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
243 print_objects(objects, realloc_failures);
244 }
245 #endif
246 }
247 if (save_oop_result) {
248 // Restore result.
249 deoptee.set_saved_oop_result(&map, return_value());
250 }
251 #if !INCLUDE_JVMCI
252 }
253 if (EliminateLocks) {
254 #endif // INCLUDE_JVMCI
255 #ifndef PRODUCT
256 bool first = true;
257 #endif
258 for (int i = 0; i < chunk->length(); i++) {
259 compiledVFrame* cvf = chunk->at(i);
260 assert (cvf->scope() != NULL,"expect only compiled java frames");
261 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
262 if (monitors->is_nonempty()) {
263 relock_objects(monitors, thread, realloc_failures);
264 #ifndef PRODUCT
265 if (PrintDeoptimizationDetails) {
266 ttyLocker ttyl;
267 for (int j = 0; j < monitors->length(); j++) {
268 MonitorInfo* mi = monitors->at(j);
269 if (mi->eliminated()) {
270 if (first) {
271 first = false;
272 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
273 }
274 if (mi->owner_is_scalar_replaced()) {
275 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
276 tty->print_cr(" failed reallocation for klass %s", k->external_name());
277 } else {
278 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
279 }
280 }
281 }
282 }
283 #endif // !PRODUCT
284 }
285 }
286 #if !INCLUDE_JVMCI
287 }
288 }
289 #endif // INCLUDE_JVMCI
290 #endif // COMPILER2_OR_JVMCI
291
292 ScopeDesc* trap_scope = chunk->at(0)->scope();
293 Handle exceptionObject;
294 if (trap_scope->rethrow_exception()) {
295 if (PrintDeoptimizationDetails) {
296 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());
297 }
298 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
299 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw");
300 ScopeValue* topOfStack = expressions->top();
301 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
302 guarantee(exceptionObject() != NULL, "exception oop can not be null");
303 }
304
305 // Ensure that no safepoint is taken after pointers have been stored
306 // in fields of rematerialized objects. If a safepoint occurs from here on
307 // out the java state residing in the vframeArray will be missed.
308 NoSafepointVerifier no_safepoint;
309
310 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
311 #if COMPILER2_OR_JVMCI
312 if (realloc_failures) {
313 pop_frames_failed_reallocs(thread, array);
314 }
315 #endif
316
317 assert(thread->vframe_array_head() == NULL, "Pending deopt!");
318 thread->set_vframe_array_head(array);
319
320 // Now that the vframeArray has been created if we have any deferred local writes
321 // added by jvmti then we can free up that structure as the data is now in the
322 // vframeArray
323
324 if (thread->deferred_locals() != NULL) {
325 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
326 int i = 0;
327 do {
328 // Because of inlining we could have multiple vframes for a single frame
329 // and several of the vframes could have deferred writes. Find them all.
330 if (list->at(i)->id() == array->original().id()) {
331 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
332 list->remove_at(i);
333 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
334 delete dlv;
335 } else {
336 i++;
337 }
338 } while ( i < list->length() );
339 if (list->length() == 0) {
340 thread->set_deferred_locals(NULL);
341 // free the list and elements back to C heap.
342 delete list;
343 }
344
345 }
346
347 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
348 CodeBlob* cb = stub_frame.cb();
349 // Verify we have the right vframeArray
350 assert(cb->frame_size() >= 0, "Unexpected frame size");
351 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
352
353 // If the deopt call site is a MethodHandle invoke call site we have
354 // to adjust the unpack_sp.
355 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
356 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
357 unpack_sp = deoptee.unextended_sp();
358
359 #ifdef ASSERT
360 assert(cb->is_deoptimization_stub() ||
361 cb->is_uncommon_trap_stub() ||
362 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
363 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
364 "unexpected code blob: %s", cb->name());
365 #endif
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 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
493
494 #if INCLUDE_JVMCI
495 if (exceptionObject() != NULL) {
496 thread->set_exception_oop(exceptionObject());
497 exec_mode = Unpack_exception;
498 }
499 #endif
500
501 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
502 assert(thread->has_pending_exception(), "should have thrown OOME");
503 thread->set_exception_oop(thread->pending_exception());
504 thread->clear_pending_exception();
505 exec_mode = Unpack_exception;
506 }
507
508 #if INCLUDE_JVMCI
509 if (thread->frames_to_pop_failed_realloc() > 0) {
510 thread->set_pending_monitorenter(false);
511 }
512 #endif
513
514 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
515 caller_adjustment * BytesPerWord,
516 caller_was_method_handle ? 0 : callee_parameters,
517 number_of_frames,
518 frame_sizes,
519 frame_pcs,
520 return_type,
521 exec_mode);
522 // On some platforms, we need a way to pass some platform dependent
523 // information to the unpacking code so the skeletal frames come out
524 // correct (initial fp value, unextended sp, ...)
525 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
526
527 if (array->frames() > 1) {
528 if (VerifyStack && TraceDeoptimization) {
529 ttyLocker ttyl;
530 tty->print_cr("Deoptimizing method containing inlining");
531 }
532 }
533
534 array->set_unroll_block(info);
535 return info;
536 }
537
538 // Called to cleanup deoptimization data structures in normal case
539 // after unpacking to stack and when stack overflow error occurs
540 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
541 vframeArray *array) {
542
543 // Get array if coming from exception
544 if (array == NULL) {
545 array = thread->vframe_array_head();
546 }
547 thread->set_vframe_array_head(NULL);
548
549 // Free the previous UnrollBlock
550 vframeArray* old_array = thread->vframe_array_last();
551 thread->set_vframe_array_last(array);
552
553 if (old_array != NULL) {
554 UnrollBlock* old_info = old_array->unroll_block();
555 old_array->set_unroll_block(NULL);
556 delete old_info;
557 delete old_array;
558 }
559
560 // Deallocate any resource creating in this routine and any ResourceObjs allocated
561 // inside the vframeArray (StackValueCollections)
562
563 delete thread->deopt_mark();
564 thread->set_deopt_mark(NULL);
565 thread->set_deopt_compiled_method(NULL);
566
567
568 if (JvmtiExport::can_pop_frame()) {
569 #ifndef CC_INTERP
570 // Regardless of whether we entered this routine with the pending
571 // popframe condition bit set, we should always clear it now
572 thread->clear_popframe_condition();
573 #else
574 // C++ interpreter will clear has_pending_popframe when it enters
575 // with method_resume. For deopt_resume2 we clear it now.
576 if (thread->popframe_forcing_deopt_reexecution())
577 thread->clear_popframe_condition();
578 #endif /* CC_INTERP */
579 }
580
581 // unpack_frames() is called at the end of the deoptimization handler
582 // and (in C2) at the end of the uncommon trap handler. Note this fact
583 // so that an asynchronous stack walker can work again. This counter is
584 // incremented at the beginning of fetch_unroll_info() and (in C2) at
585 // the beginning of uncommon_trap().
586 thread->dec_in_deopt_handler();
587 }
588
589 // Moved from cpu directories because none of the cpus has callee save values.
590 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
591 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
592
593 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
594 // the days we had adapter frames. When we deoptimize a situation where a
595 // compiled caller calls a compiled caller will have registers it expects
596 // to survive the call to the callee. If we deoptimize the callee the only
597 // way we can restore these registers is to have the oldest interpreter
598 // frame that we create restore these values. That is what this routine
599 // will accomplish.
600
601 // At the moment we have modified c2 to not have any callee save registers
602 // so this problem does not exist and this routine is just a place holder.
603
604 assert(f->is_interpreted_frame(), "must be interpreted");
605 }
606
607 // Return BasicType of value being returned
608 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
609
610 // We are already active in the special DeoptResourceMark any ResourceObj's we
611 // allocate will be freed at the end of the routine.
612
613 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
614 // but makes the entry a little slower. There is however a little dance we have to
615 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
616 ResetNoHandleMark rnhm; // No-op in release/product versions
617 HandleMark hm;
618
619 frame stub_frame = thread->last_frame();
620
621 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
622 // must point to the vframeArray for the unpack frame.
623 vframeArray* array = thread->vframe_array_head();
624
625 #ifndef PRODUCT
626 if (TraceDeoptimization) {
627 ttyLocker ttyl;
628 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d",
629 p2i(thread), p2i(array), exec_mode);
630 }
631 #endif
632 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
633 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode);
634
635 UnrollBlock* info = array->unroll_block();
636
637 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
638 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
639
640 BasicType bt = info->return_type();
641
642 // If we have an exception pending, claim that the return type is an oop
643 // so the deopt_blob does not overwrite the exception_oop.
644
645 if (exec_mode == Unpack_exception)
646 bt = T_OBJECT;
647
648 // Cleanup thread deopt data
649 cleanup_deopt_info(thread, array);
650
651 #ifndef PRODUCT
652 if (VerifyStack) {
653 ResourceMark res_mark;
654 // Clear pending exception to not break verification code (restored afterwards)
655 PRESERVE_EXCEPTION_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 (Bytecodes::is_invoke(cur_code)) {
690 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
691 cur_invoke_parameter_size = invoke.size_of_parameters();
692 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
693 callee_size_of_parameters++;
694 }
695 }
696 if (str.bci() < max_bci) {
697 Bytecodes::Code next_code = str.next();
698 if (next_code >= 0) {
699 // The interpreter oop map generator reports results before
700 // the current bytecode has executed except in the case of
701 // calls. It seems to be hard to tell whether the compiler
702 // has emitted debug information matching the "state before"
703 // a given bytecode or the state after, so we try both
704 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) {
705 // Get expression stack size for the next bytecode
706 InterpreterOopMap next_mask;
707 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
708 next_mask_expression_stack_size = next_mask.expression_stack_size();
709 if (Bytecodes::is_invoke(next_code)) {
710 Bytecode_invoke invoke(mh, str.bci());
711 next_mask_expression_stack_size += invoke.size_of_parameters();
712 }
713 // Need to subtract off the size of the result type of
714 // the bytecode because this is not described in the
715 // debug info but returned to the interpreter in the TOS
716 // caching register
717 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
718 if (bytecode_result_type != T_ILLEGAL) {
719 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
720 }
721 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
722 try_next_mask = true;
723 }
724 }
725 }
726
727 // Verify stack depth and oops in frame
728 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
729 if (!(
730 /* SPARC */
731 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
732 /* x86 */
733 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
734 (try_next_mask &&
735 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
736 top_frame_expression_stack_adjustment))) ||
737 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
738 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
739 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
740 )) {
741 {
742 ttyLocker ttyl;
743
744 // Print out some information that will help us debug the problem
745 tty->print_cr("Wrong number of expression stack elements during deoptimization");
746 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
747 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
748 iframe->interpreter_frame_expression_stack_size());
749 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
750 tty->print_cr(" try_next_mask = %d", try_next_mask);
751 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
752 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
753 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
754 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
755 tty->print_cr(" exec_mode = %d", exec_mode);
756 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
757 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
758 tty->print_cr(" Interpreted frames:");
759 for (int k = 0; k < cur_array->frames(); k++) {
760 vframeArrayElement* el = cur_array->element(k);
761 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
762 }
763 cur_array->print_on_2(tty);
764 } // release tty lock before calling guarantee
765 guarantee(false, "wrong number of expression stack elements during deopt");
766 }
767 VerifyOopClosure verify;
768 iframe->oops_interpreted_do(&verify, &rm, false);
769 callee_size_of_parameters = mh->size_of_parameters();
770 callee_max_locals = mh->max_locals();
771 is_top_frame = false;
772 }
773 }
774 #endif /* !PRODUCT */
775
776
777 return bt;
778 JRT_END
779
780
781 int Deoptimization::deoptimize_dependents() {
782 Threads::deoptimized_wrt_marked_nmethods();
783 return 0;
784 }
785
786 Deoptimization::DeoptAction Deoptimization::_unloaded_action
787 = Deoptimization::Action_reinterpret;
788
789 #if COMPILER2_OR_JVMCI
790 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
791 Handle pending_exception(THREAD, thread->pending_exception());
792 const char* exception_file = thread->exception_file();
793 int exception_line = thread->exception_line();
794 thread->clear_pending_exception();
795
796 bool failures = false;
797
798 for (int i = 0; i < objects->length(); i++) {
799 assert(objects->at(i)->is_object(), "invalid debug information");
800 ObjectValue* sv = (ObjectValue*) objects->at(i);
801
802 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
803 oop obj = NULL;
804
805 if (k->is_instance_klass()) {
806 InstanceKlass* ik = InstanceKlass::cast(k);
807 obj = ik->allocate_instance(THREAD);
808 } else if (k->is_typeArray_klass()) {
809 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
810 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
811 int len = sv->field_size() / type2size[ak->element_type()];
812 obj = ak->allocate(len, THREAD);
813 } else if (k->is_objArray_klass()) {
814 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
815 obj = ak->allocate(sv->field_size(), THREAD);
816 }
817
818 if (obj == NULL) {
819 failures = true;
820 }
821
822 assert(sv->value().is_null(), "redundant reallocation");
823 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
824 CLEAR_PENDING_EXCEPTION;
825 sv->set_value(obj);
826 }
827
828 if (failures) {
829 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
830 } else if (pending_exception.not_null()) {
831 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
832 }
833
834 return failures;
835 }
836
837 // restore elements of an eliminated type array
838 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
839 int index = 0;
840 intptr_t val;
841
842 for (int i = 0; i < sv->field_size(); i++) {
843 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
844 switch(type) {
845 case T_LONG: case T_DOUBLE: {
846 assert(value->type() == T_INT, "Agreement.");
847 StackValue* low =
848 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
849 #ifdef _LP64
850 jlong res = (jlong)low->get_int();
851 #else
852 #ifdef SPARC
853 // For SPARC we have to swap high and low words.
854 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
855 #else
856 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
857 #endif //SPARC
858 #endif
859 obj->long_at_put(index, res);
860 break;
861 }
862
863 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
864 case T_INT: case T_FLOAT: { // 4 bytes.
865 assert(value->type() == T_INT, "Agreement.");
866 bool big_value = false;
867 if (i + 1 < sv->field_size() && type == T_INT) {
868 if (sv->field_at(i)->is_location()) {
869 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
870 if (type == Location::dbl || type == Location::lng) {
871 big_value = true;
872 }
873 } else if (sv->field_at(i)->is_constant_int()) {
874 ScopeValue* next_scope_field = sv->field_at(i + 1);
875 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
876 big_value = true;
877 }
878 }
879 }
880
881 if (big_value) {
882 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
883 #ifdef _LP64
884 jlong res = (jlong)low->get_int();
885 #else
886 #ifdef SPARC
887 // For SPARC we have to swap high and low words.
888 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
889 #else
890 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
891 #endif //SPARC
892 #endif
893 obj->int_at_put(index, (jint)*((jint*)&res));
894 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
895 } else {
896 val = value->get_int();
897 obj->int_at_put(index, (jint)*((jint*)&val));
898 }
899 break;
900 }
901
902 case T_SHORT:
903 assert(value->type() == T_INT, "Agreement.");
904 val = value->get_int();
905 obj->short_at_put(index, (jshort)*((jint*)&val));
906 break;
907
908 case T_CHAR:
909 assert(value->type() == T_INT, "Agreement.");
910 val = value->get_int();
911 obj->char_at_put(index, (jchar)*((jint*)&val));
912 break;
913
914 case T_BYTE:
915 assert(value->type() == T_INT, "Agreement.");
916 val = value->get_int();
917 obj->byte_at_put(index, (jbyte)*((jint*)&val));
918 break;
919
920 case T_BOOLEAN:
921 assert(value->type() == T_INT, "Agreement.");
922 val = value->get_int();
923 obj->bool_at_put(index, (jboolean)*((jint*)&val));
924 break;
925
926 default:
927 ShouldNotReachHere();
928 }
929 index++;
930 }
931 }
932
933
934 // restore fields of an eliminated object array
935 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
936 for (int i = 0; i < sv->field_size(); i++) {
937 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
938 assert(value->type() == T_OBJECT, "object element expected");
939 obj->obj_at_put(i, value->get_obj()());
940 }
941 }
942
943 class ReassignedField {
944 public:
945 int _offset;
946 BasicType _type;
947 public:
948 ReassignedField() {
949 _offset = 0;
950 _type = T_ILLEGAL;
951 }
952 };
953
954 int compare(ReassignedField* left, ReassignedField* right) {
955 return left->_offset - right->_offset;
956 }
957
958 // Restore fields of an eliminated instance object using the same field order
959 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
960 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
961 if (klass->superklass() != NULL) {
962 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal);
963 }
964
965 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
966 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
967 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
968 ReassignedField field;
969 field._offset = fs.offset();
970 field._type = FieldType::basic_type(fs.signature());
971 fields->append(field);
972 }
973 }
974 fields->sort(compare);
975 for (int i = 0; i < fields->length(); i++) {
976 intptr_t val;
977 ScopeValue* scope_field = sv->field_at(svIndex);
978 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
979 int offset = fields->at(i)._offset;
980 BasicType type = fields->at(i)._type;
981 switch (type) {
982 case T_OBJECT: case T_ARRAY:
983 assert(value->type() == T_OBJECT, "Agreement.");
984 obj->obj_field_put(offset, value->get_obj()());
985 break;
986
987 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
988 case T_INT: case T_FLOAT: { // 4 bytes.
989 assert(value->type() == T_INT, "Agreement.");
990 bool big_value = false;
991 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
992 if (scope_field->is_location()) {
993 Location::Type type = ((LocationValue*) scope_field)->location().type();
994 if (type == Location::dbl || type == Location::lng) {
995 big_value = true;
996 }
997 }
998 if (scope_field->is_constant_int()) {
999 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1000 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1001 big_value = true;
1002 }
1003 }
1004 }
1005
1006 if (big_value) {
1007 i++;
1008 assert(i < fields->length(), "second T_INT field needed");
1009 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1010 } else {
1011 val = value->get_int();
1012 obj->int_field_put(offset, (jint)*((jint*)&val));
1013 break;
1014 }
1015 }
1016 /* no break */
1017
1018 case T_LONG: case T_DOUBLE: {
1019 assert(value->type() == T_INT, "Agreement.");
1020 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1021 #ifdef _LP64
1022 jlong res = (jlong)low->get_int();
1023 #else
1024 #ifdef SPARC
1025 // For SPARC we have to swap high and low words.
1026 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1027 #else
1028 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1029 #endif //SPARC
1030 #endif
1031 obj->long_field_put(offset, res);
1032 break;
1033 }
1034
1035 case T_SHORT:
1036 assert(value->type() == T_INT, "Agreement.");
1037 val = value->get_int();
1038 obj->short_field_put(offset, (jshort)*((jint*)&val));
1039 break;
1040
1041 case T_CHAR:
1042 assert(value->type() == T_INT, "Agreement.");
1043 val = value->get_int();
1044 obj->char_field_put(offset, (jchar)*((jint*)&val));
1045 break;
1046
1047 case T_BYTE:
1048 assert(value->type() == T_INT, "Agreement.");
1049 val = value->get_int();
1050 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1051 break;
1052
1053 case T_BOOLEAN:
1054 assert(value->type() == T_INT, "Agreement.");
1055 val = value->get_int();
1056 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1057 break;
1058
1059 default:
1060 ShouldNotReachHere();
1061 }
1062 svIndex++;
1063 }
1064 return svIndex;
1065 }
1066
1067 // restore fields of all eliminated objects and arrays
1068 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1069 for (int i = 0; i < objects->length(); i++) {
1070 ObjectValue* sv = (ObjectValue*) objects->at(i);
1071 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1072 Handle obj = sv->value();
1073 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1074 if (PrintDeoptimizationDetails) {
1075 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1076 }
1077 if (obj.is_null()) {
1078 continue;
1079 }
1080
1081 if (k->is_instance_klass()) {
1082 InstanceKlass* ik = InstanceKlass::cast(k);
1083 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1084 } else if (k->is_typeArray_klass()) {
1085 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1086 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1087 } else if (k->is_objArray_klass()) {
1088 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1089 }
1090 }
1091 }
1092
1093
1094 // relock objects for which synchronization was eliminated
1095 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
1096 for (int i = 0; i < monitors->length(); i++) {
1097 MonitorInfo* mon_info = monitors->at(i);
1098 if (mon_info->eliminated()) {
1099 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1100 if (!mon_info->owner_is_scalar_replaced()) {
1101 Handle obj(thread, mon_info->owner());
1102 markOop mark = obj->mark();
1103 if (UseBiasedLocking && mark->has_bias_pattern()) {
1104 // New allocated objects may have the mark set to anonymously biased.
1105 // Also the deoptimized method may called methods with synchronization
1106 // where the thread-local object is bias locked to the current thread.
1107 assert(mark->is_biased_anonymously() ||
1108 mark->biased_locker() == thread, "should be locked to current thread");
1109 // Reset mark word to unbiased prototype.
1110 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
1111 obj->set_mark(unbiased_prototype);
1112 }
1113 BasicLock* lock = mon_info->lock();
1114 ObjectSynchronizer::slow_enter(obj, lock, thread);
1115 assert(mon_info->owner()->is_locked(), "object must be locked now");
1116 }
1117 }
1118 }
1119 }
1120
1121
1122 #ifndef PRODUCT
1123 // print information about reallocated objects
1124 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1125 fieldDescriptor fd;
1126
1127 for (int i = 0; i < objects->length(); i++) {
1128 ObjectValue* sv = (ObjectValue*) objects->at(i);
1129 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1130 Handle obj = sv->value();
1131
1132 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1133 k->print_value();
1134 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1135 if (obj.is_null()) {
1136 tty->print(" allocation failed");
1137 } else {
1138 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1139 }
1140 tty->cr();
1141
1142 if (Verbose && !obj.is_null()) {
1143 k->oop_print_on(obj(), tty);
1144 }
1145 }
1146 }
1147 #endif
1148 #endif // COMPILER2_OR_JVMCI
1149
1150 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1151 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1152
1153 #ifndef PRODUCT
1154 if (PrintDeoptimizationDetails) {
1155 ttyLocker ttyl;
1156 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1157 fr.print_on(tty);
1158 tty->print_cr(" Virtual frames (innermost first):");
1159 for (int index = 0; index < chunk->length(); index++) {
1160 compiledVFrame* vf = chunk->at(index);
1161 tty->print(" %2d - ", index);
1162 vf->print_value();
1163 int bci = chunk->at(index)->raw_bci();
1164 const char* code_name;
1165 if (bci == SynchronizationEntryBCI) {
1166 code_name = "sync entry";
1167 } else {
1168 Bytecodes::Code code = vf->method()->code_at(bci);
1169 code_name = Bytecodes::name(code);
1170 }
1171 tty->print(" - %s", code_name);
1172 tty->print_cr(" @ bci %d ", bci);
1173 if (Verbose) {
1174 vf->print();
1175 tty->cr();
1176 }
1177 }
1178 }
1179 #endif
1180
1181 // Register map for next frame (used for stack crawl). We capture
1182 // the state of the deopt'ing frame's caller. Thus if we need to
1183 // stuff a C2I adapter we can properly fill in the callee-save
1184 // register locations.
1185 frame caller = fr.sender(reg_map);
1186 int frame_size = caller.sp() - fr.sp();
1187
1188 frame sender = caller;
1189
1190 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1191 // the vframeArray containing the unpacking information is allocated in the C heap.
1192 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1193 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1194
1195 // Compare the vframeArray to the collected vframes
1196 assert(array->structural_compare(thread, chunk), "just checking");
1197
1198 #ifndef PRODUCT
1199 if (PrintDeoptimizationDetails) {
1200 ttyLocker ttyl;
1201 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1202 }
1203 #endif // PRODUCT
1204
1205 return array;
1206 }
1207
1208 #if COMPILER2_OR_JVMCI
1209 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1210 // Reallocation of some scalar replaced objects failed. Record
1211 // that we need to pop all the interpreter frames for the
1212 // deoptimized compiled frame.
1213 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1214 thread->set_frames_to_pop_failed_realloc(array->frames());
1215 // Unlock all monitors here otherwise the interpreter will see a
1216 // mix of locked and unlocked monitors (because of failed
1217 // reallocations of synchronized objects) and be confused.
1218 for (int i = 0; i < array->frames(); i++) {
1219 MonitorChunk* monitors = array->element(i)->monitors();
1220 if (monitors != NULL) {
1221 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1222 BasicObjectLock* src = monitors->at(j);
1223 if (src->obj() != NULL) {
1224 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread);
1225 }
1226 }
1227 array->element(i)->free_monitors(thread);
1228 #ifdef ASSERT
1229 array->element(i)->set_removed_monitors();
1230 #endif
1231 }
1232 }
1233 }
1234 #endif
1235
1236 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1237 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1238 Thread* thread = Thread::current();
1239 for (int i = 0; i < monitors->length(); i++) {
1240 MonitorInfo* mon_info = monitors->at(i);
1241 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1242 objects_to_revoke->append(Handle(thread, mon_info->owner()));
1243 }
1244 }
1245 }
1246
1247
1248 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1249 if (!UseBiasedLocking) {
1250 return;
1251 }
1252
1253 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1254
1255 // Unfortunately we don't have a RegisterMap available in most of
1256 // the places we want to call this routine so we need to walk the
1257 // stack again to update the register map.
1258 if (map == NULL || !map->update_map()) {
1259 StackFrameStream sfs(thread, true);
1260 bool found = false;
1261 while (!found && !sfs.is_done()) {
1262 frame* cur = sfs.current();
1263 sfs.next();
1264 found = cur->id() == fr.id();
1265 }
1266 assert(found, "frame to be deoptimized not found on target thread's stack");
1267 map = sfs.register_map();
1268 }
1269
1270 vframe* vf = vframe::new_vframe(&fr, map, thread);
1271 compiledVFrame* cvf = compiledVFrame::cast(vf);
1272 // Revoke monitors' biases in all scopes
1273 while (!cvf->is_top()) {
1274 collect_monitors(cvf, objects_to_revoke);
1275 cvf = compiledVFrame::cast(cvf->sender());
1276 }
1277 collect_monitors(cvf, objects_to_revoke);
1278
1279 if (SafepointSynchronize::is_at_safepoint()) {
1280 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1281 } else {
1282 BiasedLocking::revoke(objects_to_revoke);
1283 }
1284 }
1285
1286
1287 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1288 if (!UseBiasedLocking) {
1289 return;
1290 }
1291
1292 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1293 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1294 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
1295 if (jt->has_last_Java_frame()) {
1296 StackFrameStream sfs(jt, true);
1297 while (!sfs.is_done()) {
1298 frame* cur = sfs.current();
1299 if (cb->contains(cur->pc())) {
1300 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1301 compiledVFrame* cvf = compiledVFrame::cast(vf);
1302 // Revoke monitors' biases in all scopes
1303 while (!cvf->is_top()) {
1304 collect_monitors(cvf, objects_to_revoke);
1305 cvf = compiledVFrame::cast(cvf->sender());
1306 }
1307 collect_monitors(cvf, objects_to_revoke);
1308 }
1309 sfs.next();
1310 }
1311 }
1312 }
1313 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1314 }
1315
1316
1317 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1318 assert(fr.can_be_deoptimized(), "checking frame type");
1319
1320 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1321
1322 if (LogCompilation && xtty != NULL) {
1323 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1324 assert(cm != NULL, "only compiled methods can deopt");
1325
1326 ttyLocker ttyl;
1327 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1328 cm->log_identity(xtty);
1329 xtty->end_head();
1330 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1331 xtty->begin_elem("jvms bci='%d'", sd->bci());
1332 xtty->method(sd->method());
1333 xtty->end_elem();
1334 if (sd->is_top()) break;
1335 }
1336 xtty->tail("deoptimized");
1337 }
1338
1339 // Patch the compiled method so that when execution returns to it we will
1340 // deopt the execution state and return to the interpreter.
1341 fr.deoptimize(thread);
1342 }
1343
1344 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1345 deoptimize(thread, fr, map, Reason_constraint);
1346 }
1347
1348 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) {
1349 // Deoptimize only if the frame comes from compile code.
1350 // Do not deoptimize the frame which is already patched
1351 // during the execution of the loops below.
1352 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1353 return;
1354 }
1355 ResourceMark rm;
1356 DeoptimizationMarker dm;
1357 if (UseBiasedLocking) {
1358 revoke_biases_of_monitors(thread, fr, map);
1359 }
1360 deoptimize_single_frame(thread, fr, reason);
1361
1362 }
1363
1364 #if INCLUDE_JVMCI
1365 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) {
1366 // there is no exception handler for this pc => deoptimize
1367 cm->make_not_entrant();
1368
1369 // Use Deoptimization::deoptimize for all of its side-effects:
1370 // revoking biases of monitors, gathering traps statistics, logging...
1371 // it also patches the return pc but we do not care about that
1372 // since we return a continuation to the deopt_blob below.
1373 JavaThread* thread = JavaThread::current();
1374 RegisterMap reg_map(thread, UseBiasedLocking);
1375 frame runtime_frame = thread->last_frame();
1376 frame caller_frame = runtime_frame.sender(®_map);
1377 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method");
1378 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler);
1379
1380 MethodData* trap_mdo = get_method_data(thread, cm->method(), true);
1381 if (trap_mdo != NULL) {
1382 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1383 }
1384
1385 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1386 }
1387 #endif
1388
1389 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1390 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1391 "can only deoptimize other thread at a safepoint");
1392 // Compute frame and register map based on thread and sp.
1393 RegisterMap reg_map(thread, UseBiasedLocking);
1394 frame fr = thread->last_frame();
1395 while (fr.id() != id) {
1396 fr = fr.sender(®_map);
1397 }
1398 deoptimize(thread, fr, ®_map, reason);
1399 }
1400
1401
1402 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1403 if (thread == Thread::current()) {
1404 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1405 } else {
1406 VM_DeoptimizeFrame deopt(thread, id, reason);
1407 VMThread::execute(&deopt);
1408 }
1409 }
1410
1411 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1412 deoptimize_frame(thread, id, Reason_constraint);
1413 }
1414
1415 // JVMTI PopFrame support
1416 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1417 {
1418 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1419 }
1420 JRT_END
1421
1422 MethodData*
1423 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1424 bool create_if_missing) {
1425 Thread* THREAD = thread;
1426 MethodData* mdo = m()->method_data();
1427 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1428 // Build an MDO. Ignore errors like OutOfMemory;
1429 // that simply means we won't have an MDO to update.
1430 Method::build_interpreter_method_data(m, THREAD);
1431 if (HAS_PENDING_EXCEPTION) {
1432 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1433 CLEAR_PENDING_EXCEPTION;
1434 }
1435 mdo = m()->method_data();
1436 }
1437 return mdo;
1438 }
1439
1440 #if COMPILER2_OR_JVMCI
1441 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1442 // in case of an unresolved klass entry, load the class.
1443 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1444 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1445 return;
1446 }
1447
1448 if (!constant_pool->tag_at(index).is_symbol()) return;
1449
1450 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1451 Symbol* symbol = constant_pool->symbol_at(index);
1452
1453 // class name?
1454 if (symbol->char_at(0) != '(') {
1455 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1456 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1457 return;
1458 }
1459
1460 // then it must be a signature!
1461 ResourceMark rm(THREAD);
1462 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1463 if (ss.is_object()) {
1464 Symbol* class_name = ss.as_symbol(CHECK);
1465 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1466 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1467 }
1468 }
1469 }
1470
1471
1472 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1473 EXCEPTION_MARK;
1474 load_class_by_index(constant_pool, index, THREAD);
1475 if (HAS_PENDING_EXCEPTION) {
1476 // Exception happened during classloading. We ignore the exception here, since it
1477 // is going to be rethrown since the current activation is going to be deoptimized and
1478 // the interpreter will re-execute the bytecode.
1479 CLEAR_PENDING_EXCEPTION;
1480 // Class loading called java code which may have caused a stack
1481 // overflow. If the exception was thrown right before the return
1482 // to the runtime the stack is no longer guarded. Reguard the
1483 // stack otherwise if we return to the uncommon trap blob and the
1484 // stack bang causes a stack overflow we crash.
1485 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1486 JavaThread* thread = (JavaThread*)THREAD;
1487 bool guard_pages_enabled = thread->stack_guards_enabled();
1488 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1489 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1490 }
1491 }
1492
1493 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1494 HandleMark hm;
1495
1496 // uncommon_trap() is called at the beginning of the uncommon trap
1497 // handler. Note this fact before we start generating temporary frames
1498 // that can confuse an asynchronous stack walker. This counter is
1499 // decremented at the end of unpack_frames().
1500 thread->inc_in_deopt_handler();
1501
1502 // We need to update the map if we have biased locking.
1503 #if INCLUDE_JVMCI
1504 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1505 RegisterMap reg_map(thread, true);
1506 #else
1507 RegisterMap reg_map(thread, UseBiasedLocking);
1508 #endif
1509 frame stub_frame = thread->last_frame();
1510 frame fr = stub_frame.sender(®_map);
1511 // Make sure the calling nmethod is not getting deoptimized and removed
1512 // before we are done with it.
1513 nmethodLocker nl(fr.pc());
1514
1515 // Log a message
1516 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1517 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1518
1519 {
1520 ResourceMark rm;
1521
1522 // Revoke biases of any monitors in the frame to ensure we can migrate them
1523 revoke_biases_of_monitors(thread, fr, ®_map);
1524
1525 DeoptReason reason = trap_request_reason(trap_request);
1526 DeoptAction action = trap_request_action(trap_request);
1527 #if INCLUDE_JVMCI
1528 int debug_id = trap_request_debug_id(trap_request);
1529 #endif
1530 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1531
1532 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1533 compiledVFrame* cvf = compiledVFrame::cast(vf);
1534
1535 CompiledMethod* nm = cvf->code();
1536
1537 ScopeDesc* trap_scope = cvf->scope();
1538
1539 if (TraceDeoptimization) {
1540 ttyLocker ttyl;
1541 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()
1542 #if INCLUDE_JVMCI
1543 , debug_id
1544 #endif
1545 );
1546 }
1547
1548 methodHandle trap_method = trap_scope->method();
1549 int trap_bci = trap_scope->bci();
1550 #if INCLUDE_JVMCI
1551 long speculation = thread->pending_failed_speculation();
1552 if (nm->is_compiled_by_jvmci()) {
1553 if (speculation != 0) {
1554 oop speculation_log = nm->as_nmethod()->speculation_log();
1555 if (speculation_log != NULL) {
1556 if (TraceDeoptimization || TraceUncollectedSpeculations) {
1557 if (HotSpotSpeculationLog::lastFailed(speculation_log) != 0) {
1558 tty->print_cr("A speculation that was not collected by the compiler is being overwritten");
1559 }
1560 }
1561 if (TraceDeoptimization) {
1562 tty->print_cr("Saving speculation to speculation log");
1563 }
1564 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation);
1565 } else {
1566 if (TraceDeoptimization) {
1567 tty->print_cr("Speculation present but no speculation log");
1568 }
1569 }
1570 thread->set_pending_failed_speculation(0);
1571 } else {
1572 if (TraceDeoptimization) {
1573 tty->print_cr("No speculation");
1574 }
1575 }
1576 } else {
1577 assert(speculation == 0, "There should not be a speculation for method compiled by non-JVMCI compilers");
1578 }
1579
1580 if (trap_bci == SynchronizationEntryBCI) {
1581 trap_bci = 0;
1582 thread->set_pending_monitorenter(true);
1583 }
1584
1585 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1586 thread->set_pending_transfer_to_interpreter(true);
1587 }
1588 #endif
1589
1590 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1591 // Record this event in the histogram.
1592 gather_statistics(reason, action, trap_bc);
1593
1594 // Ensure that we can record deopt. history:
1595 // Need MDO to record RTM code generation state.
1596 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1597
1598 methodHandle profiled_method;
1599 #if INCLUDE_JVMCI
1600 if (nm->is_compiled_by_jvmci()) {
1601 profiled_method = nm->method();
1602 } else {
1603 profiled_method = trap_method;
1604 }
1605 #else
1606 profiled_method = trap_method;
1607 #endif
1608
1609 MethodData* trap_mdo =
1610 get_method_data(thread, profiled_method, create_if_missing);
1611
1612 // Log a message
1613 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
1614 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1615 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
1616
1617 // Print a bunch of diagnostics, if requested.
1618 if (TraceDeoptimization || LogCompilation) {
1619 ResourceMark rm;
1620 ttyLocker ttyl;
1621 char buf[100];
1622 if (xtty != NULL) {
1623 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1624 os::current_thread_id(),
1625 format_trap_request(buf, sizeof(buf), trap_request));
1626 nm->log_identity(xtty);
1627 }
1628 Symbol* class_name = NULL;
1629 bool unresolved = false;
1630 if (unloaded_class_index >= 0) {
1631 constantPoolHandle constants (THREAD, trap_method->constants());
1632 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1633 class_name = constants->klass_name_at(unloaded_class_index);
1634 unresolved = true;
1635 if (xtty != NULL)
1636 xtty->print(" unresolved='1'");
1637 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1638 class_name = constants->symbol_at(unloaded_class_index);
1639 }
1640 if (xtty != NULL)
1641 xtty->name(class_name);
1642 }
1643 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
1644 // Dump the relevant MDO state.
1645 // This is the deopt count for the current reason, any previous
1646 // reasons or recompiles seen at this point.
1647 int dcnt = trap_mdo->trap_count(reason);
1648 if (dcnt != 0)
1649 xtty->print(" count='%d'", dcnt);
1650 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1651 int dos = (pdata == NULL)? 0: pdata->trap_state();
1652 if (dos != 0) {
1653 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1654 if (trap_state_is_recompiled(dos)) {
1655 int recnt2 = trap_mdo->overflow_recompile_count();
1656 if (recnt2 != 0)
1657 xtty->print(" recompiles2='%d'", recnt2);
1658 }
1659 }
1660 }
1661 if (xtty != NULL) {
1662 xtty->stamp();
1663 xtty->end_head();
1664 }
1665 if (TraceDeoptimization) { // make noise on the tty
1666 tty->print("Uncommon trap occurred in");
1667 nm->method()->print_short_name(tty);
1668 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
1669 #if INCLUDE_JVMCI
1670 if (nm->is_nmethod()) {
1671 char* installed_code_name = nm->as_nmethod()->jvmci_installed_code_name(buf, sizeof(buf));
1672 if (installed_code_name != NULL) {
1673 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name);
1674 }
1675 }
1676 #endif
1677 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
1678 p2i(fr.pc()),
1679 os::current_thread_id(),
1680 trap_reason_name(reason),
1681 trap_action_name(action),
1682 unloaded_class_index
1683 #if INCLUDE_JVMCI
1684 , debug_id
1685 #endif
1686 );
1687 if (class_name != NULL) {
1688 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1689 class_name->print_symbol_on(tty);
1690 }
1691 tty->cr();
1692 }
1693 if (xtty != NULL) {
1694 // Log the precise location of the trap.
1695 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1696 xtty->begin_elem("jvms bci='%d'", sd->bci());
1697 xtty->method(sd->method());
1698 xtty->end_elem();
1699 if (sd->is_top()) break;
1700 }
1701 xtty->tail("uncommon_trap");
1702 }
1703 }
1704 // (End diagnostic printout.)
1705
1706 // Load class if necessary
1707 if (unloaded_class_index >= 0) {
1708 constantPoolHandle constants(THREAD, trap_method->constants());
1709 load_class_by_index(constants, unloaded_class_index);
1710 }
1711
1712 // Flush the nmethod if necessary and desirable.
1713 //
1714 // We need to avoid situations where we are re-flushing the nmethod
1715 // because of a hot deoptimization site. Repeated flushes at the same
1716 // point need to be detected by the compiler and avoided. If the compiler
1717 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1718 // module must take measures to avoid an infinite cycle of recompilation
1719 // and deoptimization. There are several such measures:
1720 //
1721 // 1. If a recompilation is ordered a second time at some site X
1722 // and for the same reason R, the action is adjusted to 'reinterpret',
1723 // to give the interpreter time to exercise the method more thoroughly.
1724 // If this happens, the method's overflow_recompile_count is incremented.
1725 //
1726 // 2. If the compiler fails to reduce the deoptimization rate, then
1727 // the method's overflow_recompile_count will begin to exceed the set
1728 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1729 // is adjusted to 'make_not_compilable', and the method is abandoned
1730 // to the interpreter. This is a performance hit for hot methods,
1731 // but is better than a disastrous infinite cycle of recompilations.
1732 // (Actually, only the method containing the site X is abandoned.)
1733 //
1734 // 3. In parallel with the previous measures, if the total number of
1735 // recompilations of a method exceeds the much larger set limit
1736 // PerMethodRecompilationCutoff, the method is abandoned.
1737 // This should only happen if the method is very large and has
1738 // many "lukewarm" deoptimizations. The code which enforces this
1739 // limit is elsewhere (class nmethod, class Method).
1740 //
1741 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1742 // to recompile at each bytecode independently of the per-BCI cutoff.
1743 //
1744 // The decision to update code is up to the compiler, and is encoded
1745 // in the Action_xxx code. If the compiler requests Action_none
1746 // no trap state is changed, no compiled code is changed, and the
1747 // computation suffers along in the interpreter.
1748 //
1749 // The other action codes specify various tactics for decompilation
1750 // and recompilation. Action_maybe_recompile is the loosest, and
1751 // allows the compiled code to stay around until enough traps are seen,
1752 // and until the compiler gets around to recompiling the trapping method.
1753 //
1754 // The other actions cause immediate removal of the present code.
1755
1756 // Traps caused by injected profile shouldn't pollute trap counts.
1757 bool injected_profile_trap = trap_method->has_injected_profile() &&
1758 (reason == Reason_intrinsic || reason == Reason_unreached);
1759
1760 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
1761 bool make_not_entrant = false;
1762 bool make_not_compilable = false;
1763 bool reprofile = false;
1764 switch (action) {
1765 case Action_none:
1766 // Keep the old code.
1767 update_trap_state = false;
1768 break;
1769 case Action_maybe_recompile:
1770 // Do not need to invalidate the present code, but we can
1771 // initiate another
1772 // Start compiler without (necessarily) invalidating the nmethod.
1773 // The system will tolerate the old code, but new code should be
1774 // generated when possible.
1775 break;
1776 case Action_reinterpret:
1777 // Go back into the interpreter for a while, and then consider
1778 // recompiling form scratch.
1779 make_not_entrant = true;
1780 // Reset invocation counter for outer most method.
1781 // This will allow the interpreter to exercise the bytecodes
1782 // for a while before recompiling.
1783 // By contrast, Action_make_not_entrant is immediate.
1784 //
1785 // Note that the compiler will track null_check, null_assert,
1786 // range_check, and class_check events and log them as if they
1787 // had been traps taken from compiled code. This will update
1788 // the MDO trap history so that the next compilation will
1789 // properly detect hot trap sites.
1790 reprofile = true;
1791 break;
1792 case Action_make_not_entrant:
1793 // Request immediate recompilation, and get rid of the old code.
1794 // Make them not entrant, so next time they are called they get
1795 // recompiled. Unloaded classes are loaded now so recompile before next
1796 // time they are called. Same for uninitialized. The interpreter will
1797 // link the missing class, if any.
1798 make_not_entrant = true;
1799 break;
1800 case Action_make_not_compilable:
1801 // Give up on compiling this method at all.
1802 make_not_entrant = true;
1803 make_not_compilable = true;
1804 break;
1805 default:
1806 ShouldNotReachHere();
1807 }
1808
1809 // Setting +ProfileTraps fixes the following, on all platforms:
1810 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1811 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1812 // recompile relies on a MethodData* to record heroic opt failures.
1813
1814 // Whether the interpreter is producing MDO data or not, we also need
1815 // to use the MDO to detect hot deoptimization points and control
1816 // aggressive optimization.
1817 bool inc_recompile_count = false;
1818 ProfileData* pdata = NULL;
1819 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) {
1820 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
1821 uint this_trap_count = 0;
1822 bool maybe_prior_trap = false;
1823 bool maybe_prior_recompile = false;
1824 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
1825 #if INCLUDE_JVMCI
1826 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
1827 #endif
1828 nm->method(),
1829 //outputs:
1830 this_trap_count,
1831 maybe_prior_trap,
1832 maybe_prior_recompile);
1833 // Because the interpreter also counts null, div0, range, and class
1834 // checks, these traps from compiled code are double-counted.
1835 // This is harmless; it just means that the PerXTrapLimit values
1836 // are in effect a little smaller than they look.
1837
1838 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1839 if (per_bc_reason != Reason_none) {
1840 // Now take action based on the partially known per-BCI history.
1841 if (maybe_prior_trap
1842 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1843 // If there are too many traps at this BCI, force a recompile.
1844 // This will allow the compiler to see the limit overflow, and
1845 // take corrective action, if possible. The compiler generally
1846 // does not use the exact PerBytecodeTrapLimit value, but instead
1847 // changes its tactics if it sees any traps at all. This provides
1848 // a little hysteresis, delaying a recompile until a trap happens
1849 // several times.
1850 //
1851 // Actually, since there is only one bit of counter per BCI,
1852 // the possible per-BCI counts are {0,1,(per-method count)}.
1853 // This produces accurate results if in fact there is only
1854 // one hot trap site, but begins to get fuzzy if there are
1855 // many sites. For example, if there are ten sites each
1856 // trapping two or more times, they each get the blame for
1857 // all of their traps.
1858 make_not_entrant = true;
1859 }
1860
1861 // Detect repeated recompilation at the same BCI, and enforce a limit.
1862 if (make_not_entrant && maybe_prior_recompile) {
1863 // More than one recompile at this point.
1864 inc_recompile_count = maybe_prior_trap;
1865 }
1866 } else {
1867 // For reasons which are not recorded per-bytecode, we simply
1868 // force recompiles unconditionally.
1869 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1870 make_not_entrant = true;
1871 }
1872
1873 // Go back to the compiler if there are too many traps in this method.
1874 if (this_trap_count >= per_method_trap_limit(reason)) {
1875 // If there are too many traps in this method, force a recompile.
1876 // This will allow the compiler to see the limit overflow, and
1877 // take corrective action, if possible.
1878 // (This condition is an unlikely backstop only, because the
1879 // PerBytecodeTrapLimit is more likely to take effect first,
1880 // if it is applicable.)
1881 make_not_entrant = true;
1882 }
1883
1884 // Here's more hysteresis: If there has been a recompile at
1885 // this trap point already, run the method in the interpreter
1886 // for a while to exercise it more thoroughly.
1887 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1888 reprofile = true;
1889 }
1890 }
1891
1892 // Take requested actions on the method:
1893
1894 // Recompile
1895 if (make_not_entrant) {
1896 if (!nm->make_not_entrant()) {
1897 return; // the call did not change nmethod's state
1898 }
1899
1900 if (pdata != NULL) {
1901 // Record the recompilation event, if any.
1902 int tstate0 = pdata->trap_state();
1903 int tstate1 = trap_state_set_recompiled(tstate0, true);
1904 if (tstate1 != tstate0)
1905 pdata->set_trap_state(tstate1);
1906 }
1907
1908 #if INCLUDE_RTM_OPT
1909 // Restart collecting RTM locking abort statistic if the method
1910 // is recompiled for a reason other than RTM state change.
1911 // Assume that in new recompiled code the statistic could be different,
1912 // for example, due to different inlining.
1913 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
1914 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
1915 trap_mdo->atomic_set_rtm_state(ProfileRTM);
1916 }
1917 #endif
1918 // For code aging we count traps separately here, using make_not_entrant()
1919 // as a guard against simultaneous deopts in multiple threads.
1920 if (reason == Reason_tenured && trap_mdo != NULL) {
1921 trap_mdo->inc_tenure_traps();
1922 }
1923 }
1924
1925 if (inc_recompile_count) {
1926 trap_mdo->inc_overflow_recompile_count();
1927 if ((uint)trap_mdo->overflow_recompile_count() >
1928 (uint)PerBytecodeRecompilationCutoff) {
1929 // Give up on the method containing the bad BCI.
1930 if (trap_method() == nm->method()) {
1931 make_not_compilable = true;
1932 } else {
1933 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
1934 // But give grace to the enclosing nm->method().
1935 }
1936 }
1937 }
1938
1939 // Reprofile
1940 if (reprofile) {
1941 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1942 }
1943
1944 // Give up compiling
1945 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1946 assert(make_not_entrant, "consistent");
1947 nm->method()->set_not_compilable(CompLevel_full_optimization);
1948 }
1949
1950 } // Free marked resources
1951
1952 }
1953 JRT_END
1954
1955 ProfileData*
1956 Deoptimization::query_update_method_data(MethodData* trap_mdo,
1957 int trap_bci,
1958 Deoptimization::DeoptReason reason,
1959 bool update_total_trap_count,
1960 #if INCLUDE_JVMCI
1961 bool is_osr,
1962 #endif
1963 Method* compiled_method,
1964 //outputs:
1965 uint& ret_this_trap_count,
1966 bool& ret_maybe_prior_trap,
1967 bool& ret_maybe_prior_recompile) {
1968 bool maybe_prior_trap = false;
1969 bool maybe_prior_recompile = false;
1970 uint this_trap_count = 0;
1971 if (update_total_trap_count) {
1972 uint idx = reason;
1973 #if INCLUDE_JVMCI
1974 if (is_osr) {
1975 idx += Reason_LIMIT;
1976 }
1977 #endif
1978 uint prior_trap_count = trap_mdo->trap_count(idx);
1979 this_trap_count = trap_mdo->inc_trap_count(idx);
1980
1981 // If the runtime cannot find a place to store trap history,
1982 // it is estimated based on the general condition of the method.
1983 // If the method has ever been recompiled, or has ever incurred
1984 // a trap with the present reason , then this BCI is assumed
1985 // (pessimistically) to be the culprit.
1986 maybe_prior_trap = (prior_trap_count != 0);
1987 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1988 }
1989 ProfileData* pdata = NULL;
1990
1991
1992 // For reasons which are recorded per bytecode, we check per-BCI data.
1993 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1994 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
1995 if (per_bc_reason != Reason_none) {
1996 // Find the profile data for this BCI. If there isn't one,
1997 // try to allocate one from the MDO's set of spares.
1998 // This will let us detect a repeated trap at this point.
1999 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
2000
2001 if (pdata != NULL) {
2002 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2003 if (LogCompilation && xtty != NULL) {
2004 ttyLocker ttyl;
2005 // no more room for speculative traps in this MDO
2006 xtty->elem("speculative_traps_oom");
2007 }
2008 }
2009 // Query the trap state of this profile datum.
2010 int tstate0 = pdata->trap_state();
2011 if (!trap_state_has_reason(tstate0, per_bc_reason))
2012 maybe_prior_trap = false;
2013 if (!trap_state_is_recompiled(tstate0))
2014 maybe_prior_recompile = false;
2015
2016 // Update the trap state of this profile datum.
2017 int tstate1 = tstate0;
2018 // Record the reason.
2019 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2020 // Store the updated state on the MDO, for next time.
2021 if (tstate1 != tstate0)
2022 pdata->set_trap_state(tstate1);
2023 } else {
2024 if (LogCompilation && xtty != NULL) {
2025 ttyLocker ttyl;
2026 // Missing MDP? Leave a small complaint in the log.
2027 xtty->elem("missing_mdp bci='%d'", trap_bci);
2028 }
2029 }
2030 }
2031
2032 // Return results:
2033 ret_this_trap_count = this_trap_count;
2034 ret_maybe_prior_trap = maybe_prior_trap;
2035 ret_maybe_prior_recompile = maybe_prior_recompile;
2036 return pdata;
2037 }
2038
2039 void
2040 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2041 ResourceMark rm;
2042 // Ignored outputs:
2043 uint ignore_this_trap_count;
2044 bool ignore_maybe_prior_trap;
2045 bool ignore_maybe_prior_recompile;
2046 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2047 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2048 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2049 query_update_method_data(trap_mdo, trap_bci,
2050 (DeoptReason)reason,
2051 update_total_counts,
2052 #if INCLUDE_JVMCI
2053 false,
2054 #endif
2055 NULL,
2056 ignore_this_trap_count,
2057 ignore_maybe_prior_trap,
2058 ignore_maybe_prior_recompile);
2059 }
2060
2061 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2062 if (TraceDeoptimization) {
2063 tty->print("Uncommon trap ");
2064 }
2065 // Still in Java no safepoints
2066 {
2067 // This enters VM and may safepoint
2068 uncommon_trap_inner(thread, trap_request);
2069 }
2070 return fetch_unroll_info_helper(thread, exec_mode);
2071 }
2072
2073 // Local derived constants.
2074 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2075 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2076 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2077
2078 //---------------------------trap_state_reason---------------------------------
2079 Deoptimization::DeoptReason
2080 Deoptimization::trap_state_reason(int trap_state) {
2081 // This assert provides the link between the width of DataLayout::trap_bits
2082 // and the encoding of "recorded" reasons. It ensures there are enough
2083 // bits to store all needed reasons in the per-BCI MDO profile.
2084 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2085 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2086 trap_state -= recompile_bit;
2087 if (trap_state == DS_REASON_MASK) {
2088 return Reason_many;
2089 } else {
2090 assert((int)Reason_none == 0, "state=0 => Reason_none");
2091 return (DeoptReason)trap_state;
2092 }
2093 }
2094 //-------------------------trap_state_has_reason-------------------------------
2095 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2096 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2097 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2098 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2099 trap_state -= recompile_bit;
2100 if (trap_state == DS_REASON_MASK) {
2101 return -1; // true, unspecifically (bottom of state lattice)
2102 } else if (trap_state == reason) {
2103 return 1; // true, definitely
2104 } else if (trap_state == 0) {
2105 return 0; // false, definitely (top of state lattice)
2106 } else {
2107 return 0; // false, definitely
2108 }
2109 }
2110 //-------------------------trap_state_add_reason-------------------------------
2111 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2112 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2113 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2114 trap_state -= recompile_bit;
2115 if (trap_state == DS_REASON_MASK) {
2116 return trap_state + recompile_bit; // already at state lattice bottom
2117 } else if (trap_state == reason) {
2118 return trap_state + recompile_bit; // the condition is already true
2119 } else if (trap_state == 0) {
2120 return reason + recompile_bit; // no condition has yet been true
2121 } else {
2122 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2123 }
2124 }
2125 //-----------------------trap_state_is_recompiled------------------------------
2126 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2127 return (trap_state & DS_RECOMPILE_BIT) != 0;
2128 }
2129 //-----------------------trap_state_set_recompiled-----------------------------
2130 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2131 if (z) return trap_state | DS_RECOMPILE_BIT;
2132 else return trap_state & ~DS_RECOMPILE_BIT;
2133 }
2134 //---------------------------format_trap_state---------------------------------
2135 // This is used for debugging and diagnostics, including LogFile output.
2136 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2137 int trap_state) {
2138 assert(buflen > 0, "sanity");
2139 DeoptReason reason = trap_state_reason(trap_state);
2140 bool recomp_flag = trap_state_is_recompiled(trap_state);
2141 // Re-encode the state from its decoded components.
2142 int decoded_state = 0;
2143 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2144 decoded_state = trap_state_add_reason(decoded_state, reason);
2145 if (recomp_flag)
2146 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2147 // If the state re-encodes properly, format it symbolically.
2148 // Because this routine is used for debugging and diagnostics,
2149 // be robust even if the state is a strange value.
2150 size_t len;
2151 if (decoded_state != trap_state) {
2152 // Random buggy state that doesn't decode??
2153 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2154 } else {
2155 len = jio_snprintf(buf, buflen, "%s%s",
2156 trap_reason_name(reason),
2157 recomp_flag ? " recompiled" : "");
2158 }
2159 return buf;
2160 }
2161
2162
2163 //--------------------------------statics--------------------------------------
2164 const char* Deoptimization::_trap_reason_name[] = {
2165 // Note: Keep this in sync. with enum DeoptReason.
2166 "none",
2167 "null_check",
2168 "null_assert" JVMCI_ONLY("_or_unreached0"),
2169 "range_check",
2170 "class_check",
2171 "array_check",
2172 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2173 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2174 "profile_predicate",
2175 "unloaded",
2176 "uninitialized",
2177 "unreached",
2178 "unhandled",
2179 "constraint",
2180 "div0_check",
2181 "age",
2182 "predicate",
2183 "loop_limit_check",
2184 "speculate_class_check",
2185 "speculate_null_check",
2186 "speculate_null_assert",
2187 "rtm_state_change",
2188 "unstable_if",
2189 "unstable_fused_if",
2190 #if INCLUDE_JVMCI
2191 "aliasing",
2192 "transfer_to_interpreter",
2193 "not_compiled_exception_handler",
2194 "unresolved",
2195 "jsr_mismatch",
2196 #endif
2197 "tenured"
2198 };
2199 const char* Deoptimization::_trap_action_name[] = {
2200 // Note: Keep this in sync. with enum DeoptAction.
2201 "none",
2202 "maybe_recompile",
2203 "reinterpret",
2204 "make_not_entrant",
2205 "make_not_compilable"
2206 };
2207
2208 const char* Deoptimization::trap_reason_name(int reason) {
2209 // Check that every reason has a name
2210 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2211
2212 if (reason == Reason_many) return "many";
2213 if ((uint)reason < Reason_LIMIT)
2214 return _trap_reason_name[reason];
2215 static char buf[20];
2216 sprintf(buf, "reason%d", reason);
2217 return buf;
2218 }
2219 const char* Deoptimization::trap_action_name(int action) {
2220 // Check that every action has a name
2221 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2222
2223 if ((uint)action < Action_LIMIT)
2224 return _trap_action_name[action];
2225 static char buf[20];
2226 sprintf(buf, "action%d", action);
2227 return buf;
2228 }
2229
2230 // This is used for debugging and diagnostics, including LogFile output.
2231 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2232 int trap_request) {
2233 jint unloaded_class_index = trap_request_index(trap_request);
2234 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2235 const char* action = trap_action_name(trap_request_action(trap_request));
2236 #if INCLUDE_JVMCI
2237 int debug_id = trap_request_debug_id(trap_request);
2238 #endif
2239 size_t len;
2240 if (unloaded_class_index < 0) {
2241 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2242 reason, action
2243 #if INCLUDE_JVMCI
2244 ,debug_id
2245 #endif
2246 );
2247 } else {
2248 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2249 reason, action, unloaded_class_index
2250 #if INCLUDE_JVMCI
2251 ,debug_id
2252 #endif
2253 );
2254 }
2255 return buf;
2256 }
2257
2258 juint Deoptimization::_deoptimization_hist
2259 [Deoptimization::Reason_LIMIT]
2260 [1 + Deoptimization::Action_LIMIT]
2261 [Deoptimization::BC_CASE_LIMIT]
2262 = {0};
2263
2264 enum {
2265 LSB_BITS = 8,
2266 LSB_MASK = right_n_bits(LSB_BITS)
2267 };
2268
2269 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2270 Bytecodes::Code bc) {
2271 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2272 assert(action >= 0 && action < Action_LIMIT, "oob");
2273 _deoptimization_hist[Reason_none][0][0] += 1; // total
2274 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2275 juint* cases = _deoptimization_hist[reason][1+action];
2276 juint* bc_counter_addr = NULL;
2277 juint bc_counter = 0;
2278 // Look for an unused counter, or an exact match to this BC.
2279 if (bc != Bytecodes::_illegal) {
2280 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2281 juint* counter_addr = &cases[bc_case];
2282 juint counter = *counter_addr;
2283 if ((counter == 0 && bc_counter_addr == NULL)
2284 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2285 // this counter is either free or is already devoted to this BC
2286 bc_counter_addr = counter_addr;
2287 bc_counter = counter | bc;
2288 }
2289 }
2290 }
2291 if (bc_counter_addr == NULL) {
2292 // Overflow, or no given bytecode.
2293 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2294 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2295 }
2296 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2297 }
2298
2299 jint Deoptimization::total_deoptimization_count() {
2300 return _deoptimization_hist[Reason_none][0][0];
2301 }
2302
2303 jint Deoptimization::deoptimization_count(DeoptReason reason) {
2304 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2305 return _deoptimization_hist[reason][0][0];
2306 }
2307
2308 void Deoptimization::print_statistics() {
2309 juint total = total_deoptimization_count();
2310 juint account = total;
2311 if (total != 0) {
2312 ttyLocker ttyl;
2313 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2314 tty->print_cr("Deoptimization traps recorded:");
2315 #define PRINT_STAT_LINE(name, r) \
2316 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2317 PRINT_STAT_LINE("total", total);
2318 // For each non-zero entry in the histogram, print the reason,
2319 // the action, and (if specifically known) the type of bytecode.
2320 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2321 for (int action = 0; action < Action_LIMIT; action++) {
2322 juint* cases = _deoptimization_hist[reason][1+action];
2323 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2324 juint counter = cases[bc_case];
2325 if (counter != 0) {
2326 char name[1*K];
2327 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2328 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2329 bc = Bytecodes::_illegal;
2330 sprintf(name, "%s/%s/%s",
2331 trap_reason_name(reason),
2332 trap_action_name(action),
2333 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2334 juint r = counter >> LSB_BITS;
2335 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2336 account -= r;
2337 }
2338 }
2339 }
2340 }
2341 if (account != 0) {
2342 PRINT_STAT_LINE("unaccounted", account);
2343 }
2344 #undef PRINT_STAT_LINE
2345 if (xtty != NULL) xtty->tail("statistics");
2346 }
2347 }
2348 #else // COMPILER2_OR_JVMCI
2349
2350
2351 // Stubs for C1 only system.
2352 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2353 return false;
2354 }
2355
2356 const char* Deoptimization::trap_reason_name(int reason) {
2357 return "unknown";
2358 }
2359
2360 void Deoptimization::print_statistics() {
2361 // no output
2362 }
2363
2364 void
2365 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2366 // no udpate
2367 }
2368
2369 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2370 return 0;
2371 }
2372
2373 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2374 Bytecodes::Code bc) {
2375 // no update
2376 }
2377
2378 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2379 int trap_state) {
2380 jio_snprintf(buf, buflen, "#%d", trap_state);
2381 return buf;
2382 }
2383
2384 #endif // COMPILER2_OR_JVMCI