rev 54936 : imported patch 8221734-v3
rev 54937 : imported patch 8221734-v4
1 /*
2 * Copyright (c) 1997, 2019, 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/symbolTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/codeCache.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "code/nmethod.hpp"
32 #include "code/pcDesc.hpp"
33 #include "code/scopeDesc.hpp"
34 #include "interpreter/bytecode.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "interpreter/oopMapCache.hpp"
37 #include "memory/allocation.inline.hpp"
38 #include "memory/oopFactory.hpp"
39 #include "memory/resourceArea.hpp"
40 #include "memory/universe.hpp"
41 #include "oops/constantPool.hpp"
42 #include "oops/method.hpp"
43 #include "oops/objArrayKlass.hpp"
44 #include "oops/objArrayOop.inline.hpp"
45 #include "oops/oop.inline.hpp"
46 #include "oops/fieldStreams.hpp"
47 #include "oops/typeArrayOop.inline.hpp"
48 #include "oops/verifyOopClosure.hpp"
49 #include "prims/jvmtiThreadState.hpp"
50 #include "runtime/biasedLocking.hpp"
51 #include "runtime/compilationPolicy.hpp"
52 #include "runtime/deoptimization.hpp"
53 #include "runtime/frame.inline.hpp"
54 #include "runtime/handles.inline.hpp"
55 #include "runtime/interfaceSupport.inline.hpp"
56 #include "runtime/safepointVerifiers.hpp"
57 #include "runtime/sharedRuntime.hpp"
58 #include "runtime/signature.hpp"
59 #include "runtime/stubRoutines.hpp"
60 #include "runtime/thread.hpp"
61 #include "runtime/threadSMR.hpp"
62 #include "runtime/vframe.hpp"
63 #include "runtime/vframeArray.hpp"
64 #include "runtime/vframe_hp.hpp"
65 #include "utilities/events.hpp"
66 #include "utilities/preserveException.hpp"
67 #include "utilities/xmlstream.hpp"
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, iframe->interpreter_frame_bci());
681 int max_bci = mh->code_size();
682 // Get to the next bytecode if possible
683 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
684 // Check to see if we can grab the number of outgoing arguments
685 // at an uncommon trap for an invoke (where the compiler
686 // generates debug info before the invoke has executed)
687 Bytecodes::Code cur_code = str.next();
688 if (Bytecodes::is_invoke(cur_code)) {
689 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
690 cur_invoke_parameter_size = invoke.size_of_parameters();
691 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
692 callee_size_of_parameters++;
693 }
694 }
695 if (str.bci() < max_bci) {
696 Bytecodes::Code next_code = str.next();
697 if (next_code >= 0) {
698 // The interpreter oop map generator reports results before
699 // the current bytecode has executed except in the case of
700 // calls. It seems to be hard to tell whether the compiler
701 // has emitted debug information matching the "state before"
702 // a given bytecode or the state after, so we try both
703 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) {
704 // Get expression stack size for the next bytecode
705 InterpreterOopMap next_mask;
706 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
707 next_mask_expression_stack_size = next_mask.expression_stack_size();
708 if (Bytecodes::is_invoke(next_code)) {
709 Bytecode_invoke invoke(mh, str.bci());
710 next_mask_expression_stack_size += invoke.size_of_parameters();
711 }
712 // Need to subtract off the size of the result type of
713 // the bytecode because this is not described in the
714 // debug info but returned to the interpreter in the TOS
715 // caching register
716 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
717 if (bytecode_result_type != T_ILLEGAL) {
718 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
719 }
720 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
721 try_next_mask = true;
722 }
723 }
724 }
725
726 // Verify stack depth and oops in frame
727 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
728 if (!(
729 /* SPARC */
730 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
731 /* x86 */
732 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
733 (try_next_mask &&
734 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
735 top_frame_expression_stack_adjustment))) ||
736 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
737 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
738 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
739 )) {
740 {
741 ttyLocker ttyl;
742
743 // Print out some information that will help us debug the problem
744 tty->print_cr("Wrong number of expression stack elements during deoptimization");
745 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
746 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
747 iframe->interpreter_frame_expression_stack_size());
748 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
749 tty->print_cr(" try_next_mask = %d", try_next_mask);
750 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
751 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
752 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
753 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
754 tty->print_cr(" exec_mode = %d", exec_mode);
755 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
756 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
757 tty->print_cr(" Interpreted frames:");
758 for (int k = 0; k < cur_array->frames(); k++) {
759 vframeArrayElement* el = cur_array->element(k);
760 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
761 }
762 cur_array->print_on_2(tty);
763 } // release tty lock before calling guarantee
764 guarantee(false, "wrong number of expression stack elements during deopt");
765 }
766 VerifyOopClosure verify;
767 iframe->oops_interpreted_do(&verify, &rm, false);
768 callee_size_of_parameters = mh->size_of_parameters();
769 callee_max_locals = mh->max_locals();
770 is_top_frame = false;
771 }
772 }
773 #endif /* !PRODUCT */
774
775
776 return bt;
777 JRT_END
778
779 class DeoptimizeMarkedTC : public ThreadClosure {
780 bool _in_handshake;
781 public:
782 DeoptimizeMarkedTC(bool in_handshake) : _in_handshake(in_handshake) {}
783 virtual void do_thread(Thread* thread) {
784 assert(thread->is_Java_thread(), "must be");
785 JavaThread* jt = (JavaThread*)thread;
786 jt->deoptimize_marked_methods(_in_handshake);
787 }
788 };
789
790 void Deoptimization::deoptimize_all_marked() {
791 ResourceMark rm;
792 DeoptimizationMarker dm;
793
794 if (SafepointSynchronize::is_at_safepoint()) {
795 DeoptimizeMarkedTC deopt(false);
796 // Make the dependent methods not entrant
797 CodeCache::make_marked_nmethods_not_entrant();
798 Threads::java_threads_do(&deopt);
799 } else {
800 // Make the dependent methods not entrant
801 {
802 MutexLocker mu(CodeCache_lock, Mutex::_no_safepoint_check_flag);
803 CodeCache::make_marked_nmethods_not_entrant();
804 }
805 DeoptimizeMarkedTC deopt(true);
806 Handshake::execute(&deopt);
807 }
808 }
809
810 Deoptimization::DeoptAction Deoptimization::_unloaded_action
811 = Deoptimization::Action_reinterpret;
812
813 #if COMPILER2_OR_JVMCI
814 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
815 Handle pending_exception(THREAD, thread->pending_exception());
816 const char* exception_file = thread->exception_file();
817 int exception_line = thread->exception_line();
818 thread->clear_pending_exception();
819
820 bool failures = false;
821
822 for (int i = 0; i < objects->length(); i++) {
823 assert(objects->at(i)->is_object(), "invalid debug information");
824 ObjectValue* sv = (ObjectValue*) objects->at(i);
825
826 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
827 oop obj = NULL;
828
829 if (k->is_instance_klass()) {
830 InstanceKlass* ik = InstanceKlass::cast(k);
831 obj = ik->allocate_instance(THREAD);
832 } else if (k->is_typeArray_klass()) {
833 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
834 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
835 int len = sv->field_size() / type2size[ak->element_type()];
836 obj = ak->allocate(len, THREAD);
837 } else if (k->is_objArray_klass()) {
838 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
839 obj = ak->allocate(sv->field_size(), THREAD);
840 }
841
842 if (obj == NULL) {
843 failures = true;
844 }
845
846 assert(sv->value().is_null(), "redundant reallocation");
847 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
848 CLEAR_PENDING_EXCEPTION;
849 sv->set_value(obj);
850 }
851
852 if (failures) {
853 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
854 } else if (pending_exception.not_null()) {
855 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
856 }
857
858 return failures;
859 }
860
861 // restore elements of an eliminated type array
862 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
863 int index = 0;
864 intptr_t val;
865
866 for (int i = 0; i < sv->field_size(); i++) {
867 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
868 switch(type) {
869 case T_LONG: case T_DOUBLE: {
870 assert(value->type() == T_INT, "Agreement.");
871 StackValue* low =
872 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
873 #ifdef _LP64
874 jlong res = (jlong)low->get_int();
875 #else
876 #ifdef SPARC
877 // For SPARC we have to swap high and low words.
878 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
879 #else
880 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
881 #endif //SPARC
882 #endif
883 obj->long_at_put(index, res);
884 break;
885 }
886
887 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
888 case T_INT: case T_FLOAT: { // 4 bytes.
889 assert(value->type() == T_INT, "Agreement.");
890 bool big_value = false;
891 if (i + 1 < sv->field_size() && type == T_INT) {
892 if (sv->field_at(i)->is_location()) {
893 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
894 if (type == Location::dbl || type == Location::lng) {
895 big_value = true;
896 }
897 } else if (sv->field_at(i)->is_constant_int()) {
898 ScopeValue* next_scope_field = sv->field_at(i + 1);
899 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
900 big_value = true;
901 }
902 }
903 }
904
905 if (big_value) {
906 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
907 #ifdef _LP64
908 jlong res = (jlong)low->get_int();
909 #else
910 #ifdef SPARC
911 // For SPARC we have to swap high and low words.
912 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
913 #else
914 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
915 #endif //SPARC
916 #endif
917 obj->int_at_put(index, (jint)*((jint*)&res));
918 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
919 } else {
920 val = value->get_int();
921 obj->int_at_put(index, (jint)*((jint*)&val));
922 }
923 break;
924 }
925
926 case T_SHORT:
927 assert(value->type() == T_INT, "Agreement.");
928 val = value->get_int();
929 obj->short_at_put(index, (jshort)*((jint*)&val));
930 break;
931
932 case T_CHAR:
933 assert(value->type() == T_INT, "Agreement.");
934 val = value->get_int();
935 obj->char_at_put(index, (jchar)*((jint*)&val));
936 break;
937
938 case T_BYTE:
939 assert(value->type() == T_INT, "Agreement.");
940 val = value->get_int();
941 obj->byte_at_put(index, (jbyte)*((jint*)&val));
942 break;
943
944 case T_BOOLEAN:
945 assert(value->type() == T_INT, "Agreement.");
946 val = value->get_int();
947 obj->bool_at_put(index, (jboolean)*((jint*)&val));
948 break;
949
950 default:
951 ShouldNotReachHere();
952 }
953 index++;
954 }
955 }
956
957
958 // restore fields of an eliminated object array
959 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
960 for (int i = 0; i < sv->field_size(); i++) {
961 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
962 assert(value->type() == T_OBJECT, "object element expected");
963 obj->obj_at_put(i, value->get_obj()());
964 }
965 }
966
967 class ReassignedField {
968 public:
969 int _offset;
970 BasicType _type;
971 public:
972 ReassignedField() {
973 _offset = 0;
974 _type = T_ILLEGAL;
975 }
976 };
977
978 int compare(ReassignedField* left, ReassignedField* right) {
979 return left->_offset - right->_offset;
980 }
981
982 // Restore fields of an eliminated instance object using the same field order
983 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
984 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
985 if (klass->superklass() != NULL) {
986 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal);
987 }
988
989 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
990 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
991 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
992 ReassignedField field;
993 field._offset = fs.offset();
994 field._type = FieldType::basic_type(fs.signature());
995 fields->append(field);
996 }
997 }
998 fields->sort(compare);
999 for (int i = 0; i < fields->length(); i++) {
1000 intptr_t val;
1001 ScopeValue* scope_field = sv->field_at(svIndex);
1002 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1003 int offset = fields->at(i)._offset;
1004 BasicType type = fields->at(i)._type;
1005 switch (type) {
1006 case T_OBJECT: case T_ARRAY:
1007 assert(value->type() == T_OBJECT, "Agreement.");
1008 obj->obj_field_put(offset, value->get_obj()());
1009 break;
1010
1011 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1012 case T_INT: case T_FLOAT: { // 4 bytes.
1013 assert(value->type() == T_INT, "Agreement.");
1014 bool big_value = false;
1015 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1016 if (scope_field->is_location()) {
1017 Location::Type type = ((LocationValue*) scope_field)->location().type();
1018 if (type == Location::dbl || type == Location::lng) {
1019 big_value = true;
1020 }
1021 }
1022 if (scope_field->is_constant_int()) {
1023 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1024 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1025 big_value = true;
1026 }
1027 }
1028 }
1029
1030 if (big_value) {
1031 i++;
1032 assert(i < fields->length(), "second T_INT field needed");
1033 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1034 } else {
1035 val = value->get_int();
1036 obj->int_field_put(offset, (jint)*((jint*)&val));
1037 break;
1038 }
1039 }
1040 /* no break */
1041
1042 case T_LONG: case T_DOUBLE: {
1043 assert(value->type() == T_INT, "Agreement.");
1044 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1045 #ifdef _LP64
1046 jlong res = (jlong)low->get_int();
1047 #else
1048 #ifdef SPARC
1049 // For SPARC we have to swap high and low words.
1050 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1051 #else
1052 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1053 #endif //SPARC
1054 #endif
1055 obj->long_field_put(offset, res);
1056 break;
1057 }
1058
1059 case T_SHORT:
1060 assert(value->type() == T_INT, "Agreement.");
1061 val = value->get_int();
1062 obj->short_field_put(offset, (jshort)*((jint*)&val));
1063 break;
1064
1065 case T_CHAR:
1066 assert(value->type() == T_INT, "Agreement.");
1067 val = value->get_int();
1068 obj->char_field_put(offset, (jchar)*((jint*)&val));
1069 break;
1070
1071 case T_BYTE:
1072 assert(value->type() == T_INT, "Agreement.");
1073 val = value->get_int();
1074 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1075 break;
1076
1077 case T_BOOLEAN:
1078 assert(value->type() == T_INT, "Agreement.");
1079 val = value->get_int();
1080 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1081 break;
1082
1083 default:
1084 ShouldNotReachHere();
1085 }
1086 svIndex++;
1087 }
1088 return svIndex;
1089 }
1090
1091 // restore fields of all eliminated objects and arrays
1092 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1093 for (int i = 0; i < objects->length(); i++) {
1094 ObjectValue* sv = (ObjectValue*) objects->at(i);
1095 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1096 Handle obj = sv->value();
1097 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1098 if (PrintDeoptimizationDetails) {
1099 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1100 }
1101 if (obj.is_null()) {
1102 continue;
1103 }
1104
1105 if (k->is_instance_klass()) {
1106 InstanceKlass* ik = InstanceKlass::cast(k);
1107 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1108 } else if (k->is_typeArray_klass()) {
1109 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1110 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1111 } else if (k->is_objArray_klass()) {
1112 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1113 }
1114 }
1115 }
1116
1117
1118 // relock objects for which synchronization was eliminated
1119 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
1120 for (int i = 0; i < monitors->length(); i++) {
1121 MonitorInfo* mon_info = monitors->at(i);
1122 if (mon_info->eliminated()) {
1123 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1124 if (!mon_info->owner_is_scalar_replaced()) {
1125 Handle obj(thread, mon_info->owner());
1126 markOop mark = obj->mark();
1127 if (UseBiasedLocking && mark->has_bias_pattern()) {
1128 // New allocated objects may have the mark set to anonymously biased.
1129 // Also the deoptimized method may called methods with synchronization
1130 // where the thread-local object is bias locked to the current thread.
1131 assert(mark->is_biased_anonymously() ||
1132 mark->biased_locker() == thread, "should be locked to current thread");
1133 // Reset mark word to unbiased prototype.
1134 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
1135 obj->set_mark(unbiased_prototype);
1136 }
1137 BasicLock* lock = mon_info->lock();
1138 ObjectSynchronizer::slow_enter(obj, lock, thread);
1139 assert(mon_info->owner()->is_locked(), "object must be locked now");
1140 }
1141 }
1142 }
1143 }
1144
1145
1146 #ifndef PRODUCT
1147 // print information about reallocated objects
1148 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1149 fieldDescriptor fd;
1150
1151 for (int i = 0; i < objects->length(); i++) {
1152 ObjectValue* sv = (ObjectValue*) objects->at(i);
1153 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1154 Handle obj = sv->value();
1155
1156 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1157 k->print_value();
1158 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1159 if (obj.is_null()) {
1160 tty->print(" allocation failed");
1161 } else {
1162 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1163 }
1164 tty->cr();
1165
1166 if (Verbose && !obj.is_null()) {
1167 k->oop_print_on(obj(), tty);
1168 }
1169 }
1170 }
1171 #endif
1172 #endif // COMPILER2_OR_JVMCI
1173
1174 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1175 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1176
1177 #ifndef PRODUCT
1178 if (PrintDeoptimizationDetails) {
1179 ttyLocker ttyl;
1180 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1181 fr.print_on(tty);
1182 tty->print_cr(" Virtual frames (innermost first):");
1183 for (int index = 0; index < chunk->length(); index++) {
1184 compiledVFrame* vf = chunk->at(index);
1185 tty->print(" %2d - ", index);
1186 vf->print_value();
1187 int bci = chunk->at(index)->raw_bci();
1188 const char* code_name;
1189 if (bci == SynchronizationEntryBCI) {
1190 code_name = "sync entry";
1191 } else {
1192 Bytecodes::Code code = vf->method()->code_at(bci);
1193 code_name = Bytecodes::name(code);
1194 }
1195 tty->print(" - %s", code_name);
1196 tty->print_cr(" @ bci %d ", bci);
1197 if (Verbose) {
1198 vf->print();
1199 tty->cr();
1200 }
1201 }
1202 }
1203 #endif
1204
1205 // Register map for next frame (used for stack crawl). We capture
1206 // the state of the deopt'ing frame's caller. Thus if we need to
1207 // stuff a C2I adapter we can properly fill in the callee-save
1208 // register locations.
1209 frame caller = fr.sender(reg_map);
1210 int frame_size = caller.sp() - fr.sp();
1211
1212 frame sender = caller;
1213
1214 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1215 // the vframeArray containing the unpacking information is allocated in the C heap.
1216 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1217 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1218
1219 // Compare the vframeArray to the collected vframes
1220 assert(array->structural_compare(thread, chunk), "just checking");
1221
1222 #ifndef PRODUCT
1223 if (PrintDeoptimizationDetails) {
1224 ttyLocker ttyl;
1225 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1226 }
1227 #endif // PRODUCT
1228
1229 return array;
1230 }
1231
1232 #if COMPILER2_OR_JVMCI
1233 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1234 // Reallocation of some scalar replaced objects failed. Record
1235 // that we need to pop all the interpreter frames for the
1236 // deoptimized compiled frame.
1237 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1238 thread->set_frames_to_pop_failed_realloc(array->frames());
1239 // Unlock all monitors here otherwise the interpreter will see a
1240 // mix of locked and unlocked monitors (because of failed
1241 // reallocations of synchronized objects) and be confused.
1242 for (int i = 0; i < array->frames(); i++) {
1243 MonitorChunk* monitors = array->element(i)->monitors();
1244 if (monitors != NULL) {
1245 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1246 BasicObjectLock* src = monitors->at(j);
1247 if (src->obj() != NULL) {
1248 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread);
1249 }
1250 }
1251 array->element(i)->free_monitors(thread);
1252 #ifdef ASSERT
1253 array->element(i)->set_removed_monitors();
1254 #endif
1255 }
1256 }
1257 }
1258 #endif
1259
1260 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1261 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1262 Thread* thread = Thread::current();
1263 for (int i = 0; i < monitors->length(); i++) {
1264 MonitorInfo* mon_info = monitors->at(i);
1265 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1266 objects_to_revoke->append(Handle(thread, mon_info->owner()));
1267 }
1268 }
1269 }
1270
1271 static void get_monitors_from_stack(GrowableArray<Handle>* objects_to_revoke, JavaThread* thread, frame fr, RegisterMap* map) {
1272 // Unfortunately we don't have a RegisterMap available in most of
1273 // the places we want to call this routine so we need to walk the
1274 // stack again to update the register map.
1275 if (map == NULL || !map->update_map()) {
1276 StackFrameStream sfs(thread, true);
1277 bool found = false;
1278 while (!found && !sfs.is_done()) {
1279 frame* cur = sfs.current();
1280 sfs.next();
1281 found = cur->id() == fr.id();
1282 }
1283 assert(found, "frame to be deoptimized not found on target thread's stack");
1284 map = sfs.register_map();
1285 }
1286
1287 vframe* vf = vframe::new_vframe(&fr, map, thread);
1288 compiledVFrame* cvf = compiledVFrame::cast(vf);
1289 // Revoke monitors' biases in all scopes
1290 while (!cvf->is_top()) {
1291 collect_monitors(cvf, objects_to_revoke);
1292 cvf = compiledVFrame::cast(cvf->sender());
1293 }
1294 collect_monitors(cvf, objects_to_revoke);
1295 }
1296
1297 void Deoptimization::revoke_safepoint(JavaThread* thread, frame fr, RegisterMap* map) {
1298 if (!UseBiasedLocking) {
1299 return;
1300 }
1301 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1302 get_monitors_from_stack(objects_to_revoke, thread, fr, map);
1303
1304 if (SafepointSynchronize::is_at_safepoint()) {
1305 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1306 } else {
1307 BiasedLocking::revoke(objects_to_revoke);
1308 }
1309 }
1310
1311 void Deoptimization::revoke_handshake(JavaThread* thread, frame fr, RegisterMap* map) {
1312 if (!UseBiasedLocking) {
1313 return;
1314 }
1315 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1316 get_monitors_from_stack(objects_to_revoke, thread, fr, map);
1317
1318 int len = objects_to_revoke->length();
1319 for (int i = 0; i < len; i++) {
1320 oop obj = (objects_to_revoke->at(i))();
1321 markOop mark = obj->mark();
1322 assert(!mark->has_bias_pattern() || mark->biased_locker() == thread, "Can't revoke");
1323 BiasedLocking::revoke_own_locks_in_handshake(objects_to_revoke->at(i), thread);
1324 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
1325 }
1326 }
1327
1328
1329 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1330 assert(fr.can_be_deoptimized(), "checking frame type");
1331
1332 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1333
1334 if (LogCompilation && xtty != NULL) {
1335 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1336 assert(cm != NULL, "only compiled methods can deopt");
1337
1338 ttyLocker ttyl;
1339 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1340 cm->log_identity(xtty);
1341 xtty->end_head();
1342 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1343 xtty->begin_elem("jvms bci='%d'", sd->bci());
1344 xtty->method(sd->method());
1345 xtty->end_elem();
1346 if (sd->is_top()) break;
1347 }
1348 xtty->tail("deoptimized");
1349 }
1350
1351 // Patch the compiled method so that when execution returns to it we will
1352 // deopt the execution state and return to the interpreter.
1353 fr.deoptimize(thread);
1354 }
1355
1356 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, bool in_handshake) {
1357 deopt_thread(in_handshake, thread, fr, map, Reason_constraint);
1358 }
1359
1360 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) {
1361 deopt_thread(false, thread, fr, map, reason);
1362 }
1363
1364 void Deoptimization::deopt_thread(bool in_handshake, JavaThread* thread,
1365 frame fr, RegisterMap *map, DeoptReason reason) {
1366 // Deoptimize only if the frame comes from compile code.
1367 // Do not deoptimize the frame which is already patched
1368 // during the execution of the loops below.
1369 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1370 return;
1371 }
1372 ResourceMark rm;
1373 DeoptimizationMarker dm;
1374 if (UseBiasedLocking) {
1375 if (in_handshake) {
1376 revoke_handshake(thread, fr, map);
1377 } else {
1378 revoke_safepoint(thread, fr, map);
1379 }
1380 }
1381 deoptimize_single_frame(thread, fr, reason);
1382
1383 }
1384
1385 #if INCLUDE_JVMCI
1386 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) {
1387 // there is no exception handler for this pc => deoptimize
1388 cm->make_not_entrant();
1389
1390 // Use Deoptimization::deoptimize for all of its side-effects:
1391 // revoking biases of monitors, gathering traps statistics, logging...
1392 // it also patches the return pc but we do not care about that
1393 // since we return a continuation to the deopt_blob below.
1394 JavaThread* thread = JavaThread::current();
1395 RegisterMap reg_map(thread, UseBiasedLocking);
1396 frame runtime_frame = thread->last_frame();
1397 frame caller_frame = runtime_frame.sender(®_map);
1398 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method");
1399 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler);
1400
1401 MethodData* trap_mdo = get_method_data(thread, cm->method(), true);
1402 if (trap_mdo != NULL) {
1403 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1404 }
1405
1406 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1407 }
1408 #endif
1409
1410 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1411 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1412 "can only deoptimize other thread at a safepoint");
1413 // Compute frame and register map based on thread and sp.
1414 RegisterMap reg_map(thread, UseBiasedLocking);
1415 frame fr = thread->last_frame();
1416 while (fr.id() != id) {
1417 fr = fr.sender(®_map);
1418 }
1419 deoptimize(thread, fr, ®_map, reason);
1420 }
1421
1422
1423 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1424 if (thread == Thread::current()) {
1425 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1426 } else {
1427 VM_DeoptimizeFrame deopt(thread, id, reason);
1428 VMThread::execute(&deopt);
1429 }
1430 }
1431
1432 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1433 deoptimize_frame(thread, id, Reason_constraint);
1434 }
1435
1436 // JVMTI PopFrame support
1437 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1438 {
1439 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1440 }
1441 JRT_END
1442
1443 MethodData*
1444 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1445 bool create_if_missing) {
1446 Thread* THREAD = thread;
1447 MethodData* mdo = m()->method_data();
1448 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1449 // Build an MDO. Ignore errors like OutOfMemory;
1450 // that simply means we won't have an MDO to update.
1451 Method::build_interpreter_method_data(m, THREAD);
1452 if (HAS_PENDING_EXCEPTION) {
1453 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1454 CLEAR_PENDING_EXCEPTION;
1455 }
1456 mdo = m()->method_data();
1457 }
1458 return mdo;
1459 }
1460
1461 #if COMPILER2_OR_JVMCI
1462 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1463 // in case of an unresolved klass entry, load the class.
1464 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1465 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1466 return;
1467 }
1468
1469 if (!constant_pool->tag_at(index).is_symbol()) return;
1470
1471 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1472 Symbol* symbol = constant_pool->symbol_at(index);
1473
1474 // class name?
1475 if (symbol->char_at(0) != '(') {
1476 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1477 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1478 return;
1479 }
1480
1481 // then it must be a signature!
1482 ResourceMark rm(THREAD);
1483 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1484 if (ss.is_object()) {
1485 Symbol* class_name = ss.as_symbol();
1486 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1487 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1488 }
1489 }
1490 }
1491
1492
1493 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1494 EXCEPTION_MARK;
1495 load_class_by_index(constant_pool, index, THREAD);
1496 if (HAS_PENDING_EXCEPTION) {
1497 // Exception happened during classloading. We ignore the exception here, since it
1498 // is going to be rethrown since the current activation is going to be deoptimized and
1499 // the interpreter will re-execute the bytecode.
1500 CLEAR_PENDING_EXCEPTION;
1501 // Class loading called java code which may have caused a stack
1502 // overflow. If the exception was thrown right before the return
1503 // to the runtime the stack is no longer guarded. Reguard the
1504 // stack otherwise if we return to the uncommon trap blob and the
1505 // stack bang causes a stack overflow we crash.
1506 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1507 JavaThread* thread = (JavaThread*)THREAD;
1508 bool guard_pages_enabled = thread->stack_guards_enabled();
1509 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1510 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1511 }
1512 }
1513
1514 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1515 HandleMark hm;
1516
1517 // uncommon_trap() is called at the beginning of the uncommon trap
1518 // handler. Note this fact before we start generating temporary frames
1519 // that can confuse an asynchronous stack walker. This counter is
1520 // decremented at the end of unpack_frames().
1521 thread->inc_in_deopt_handler();
1522
1523 // We need to update the map if we have biased locking.
1524 #if INCLUDE_JVMCI
1525 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1526 RegisterMap reg_map(thread, true);
1527 #else
1528 RegisterMap reg_map(thread, UseBiasedLocking);
1529 #endif
1530 frame stub_frame = thread->last_frame();
1531 frame fr = stub_frame.sender(®_map);
1532 // Make sure the calling nmethod is not getting deoptimized and removed
1533 // before we are done with it.
1534 nmethodLocker nl(fr.pc());
1535
1536 // Log a message
1537 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1538 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1539
1540 {
1541 ResourceMark rm;
1542
1543 // Revoke biases of any monitors in the frame to ensure we can migrate them
1544 revoke_biases_of_monitors(thread, fr, ®_map);
1545
1546 DeoptReason reason = trap_request_reason(trap_request);
1547 DeoptAction action = trap_request_action(trap_request);
1548 #if INCLUDE_JVMCI
1549 int debug_id = trap_request_debug_id(trap_request);
1550 #endif
1551 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1552
1553 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1554 compiledVFrame* cvf = compiledVFrame::cast(vf);
1555
1556 CompiledMethod* nm = cvf->code();
1557
1558 ScopeDesc* trap_scope = cvf->scope();
1559
1560 if (TraceDeoptimization) {
1561 ttyLocker ttyl;
1562 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()
1563 #if INCLUDE_JVMCI
1564 , debug_id
1565 #endif
1566 );
1567 }
1568
1569 methodHandle trap_method = trap_scope->method();
1570 int trap_bci = trap_scope->bci();
1571 #if INCLUDE_JVMCI
1572 jlong speculation = thread->pending_failed_speculation();
1573 if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods
1574 nm->as_nmethod()->update_speculation(thread);
1575 } else {
1576 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
1577 }
1578
1579 if (trap_bci == SynchronizationEntryBCI) {
1580 trap_bci = 0;
1581 thread->set_pending_monitorenter(true);
1582 }
1583
1584 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1585 thread->set_pending_transfer_to_interpreter(true);
1586 }
1587 #endif
1588
1589 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1590 // Record this event in the histogram.
1591 gather_statistics(reason, action, trap_bc);
1592
1593 // Ensure that we can record deopt. history:
1594 // Need MDO to record RTM code generation state.
1595 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1596
1597 methodHandle profiled_method;
1598 #if INCLUDE_JVMCI
1599 if (nm->is_compiled_by_jvmci()) {
1600 profiled_method = nm->method();
1601 } else {
1602 profiled_method = trap_method;
1603 }
1604 #else
1605 profiled_method = trap_method;
1606 #endif
1607
1608 MethodData* trap_mdo =
1609 get_method_data(thread, profiled_method, create_if_missing);
1610
1611 // Log a message
1612 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
1613 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1614 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
1615
1616 // Print a bunch of diagnostics, if requested.
1617 if (TraceDeoptimization || LogCompilation) {
1618 ResourceMark rm;
1619 ttyLocker ttyl;
1620 char buf[100];
1621 if (xtty != NULL) {
1622 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1623 os::current_thread_id(),
1624 format_trap_request(buf, sizeof(buf), trap_request));
1625 #if INCLUDE_JVMCI
1626 if (speculation != 0) {
1627 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
1628 }
1629 #endif
1630 nm->log_identity(xtty);
1631 }
1632 Symbol* class_name = NULL;
1633 bool unresolved = false;
1634 if (unloaded_class_index >= 0) {
1635 constantPoolHandle constants (THREAD, trap_method->constants());
1636 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1637 class_name = constants->klass_name_at(unloaded_class_index);
1638 unresolved = true;
1639 if (xtty != NULL)
1640 xtty->print(" unresolved='1'");
1641 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1642 class_name = constants->symbol_at(unloaded_class_index);
1643 }
1644 if (xtty != NULL)
1645 xtty->name(class_name);
1646 }
1647 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
1648 // Dump the relevant MDO state.
1649 // This is the deopt count for the current reason, any previous
1650 // reasons or recompiles seen at this point.
1651 int dcnt = trap_mdo->trap_count(reason);
1652 if (dcnt != 0)
1653 xtty->print(" count='%d'", dcnt);
1654 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1655 int dos = (pdata == NULL)? 0: pdata->trap_state();
1656 if (dos != 0) {
1657 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1658 if (trap_state_is_recompiled(dos)) {
1659 int recnt2 = trap_mdo->overflow_recompile_count();
1660 if (recnt2 != 0)
1661 xtty->print(" recompiles2='%d'", recnt2);
1662 }
1663 }
1664 }
1665 if (xtty != NULL) {
1666 xtty->stamp();
1667 xtty->end_head();
1668 }
1669 if (TraceDeoptimization) { // make noise on the tty
1670 tty->print("Uncommon trap occurred in");
1671 nm->method()->print_short_name(tty);
1672 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
1673 #if INCLUDE_JVMCI
1674 if (nm->is_nmethod()) {
1675 const char* installed_code_name = nm->as_nmethod()->jvmci_name();
1676 if (installed_code_name != NULL) {
1677 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name);
1678 }
1679 }
1680 #endif
1681 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
1682 p2i(fr.pc()),
1683 os::current_thread_id(),
1684 trap_reason_name(reason),
1685 trap_action_name(action),
1686 unloaded_class_index
1687 #if INCLUDE_JVMCI
1688 , debug_id
1689 #endif
1690 );
1691 if (class_name != NULL) {
1692 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1693 class_name->print_symbol_on(tty);
1694 }
1695 tty->cr();
1696 }
1697 if (xtty != NULL) {
1698 // Log the precise location of the trap.
1699 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1700 xtty->begin_elem("jvms bci='%d'", sd->bci());
1701 xtty->method(sd->method());
1702 xtty->end_elem();
1703 if (sd->is_top()) break;
1704 }
1705 xtty->tail("uncommon_trap");
1706 }
1707 }
1708 // (End diagnostic printout.)
1709
1710 // Load class if necessary
1711 if (unloaded_class_index >= 0) {
1712 constantPoolHandle constants(THREAD, trap_method->constants());
1713 load_class_by_index(constants, unloaded_class_index);
1714 }
1715
1716 // Flush the nmethod if necessary and desirable.
1717 //
1718 // We need to avoid situations where we are re-flushing the nmethod
1719 // because of a hot deoptimization site. Repeated flushes at the same
1720 // point need to be detected by the compiler and avoided. If the compiler
1721 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1722 // module must take measures to avoid an infinite cycle of recompilation
1723 // and deoptimization. There are several such measures:
1724 //
1725 // 1. If a recompilation is ordered a second time at some site X
1726 // and for the same reason R, the action is adjusted to 'reinterpret',
1727 // to give the interpreter time to exercise the method more thoroughly.
1728 // If this happens, the method's overflow_recompile_count is incremented.
1729 //
1730 // 2. If the compiler fails to reduce the deoptimization rate, then
1731 // the method's overflow_recompile_count will begin to exceed the set
1732 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1733 // is adjusted to 'make_not_compilable', and the method is abandoned
1734 // to the interpreter. This is a performance hit for hot methods,
1735 // but is better than a disastrous infinite cycle of recompilations.
1736 // (Actually, only the method containing the site X is abandoned.)
1737 //
1738 // 3. In parallel with the previous measures, if the total number of
1739 // recompilations of a method exceeds the much larger set limit
1740 // PerMethodRecompilationCutoff, the method is abandoned.
1741 // This should only happen if the method is very large and has
1742 // many "lukewarm" deoptimizations. The code which enforces this
1743 // limit is elsewhere (class nmethod, class Method).
1744 //
1745 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1746 // to recompile at each bytecode independently of the per-BCI cutoff.
1747 //
1748 // The decision to update code is up to the compiler, and is encoded
1749 // in the Action_xxx code. If the compiler requests Action_none
1750 // no trap state is changed, no compiled code is changed, and the
1751 // computation suffers along in the interpreter.
1752 //
1753 // The other action codes specify various tactics for decompilation
1754 // and recompilation. Action_maybe_recompile is the loosest, and
1755 // allows the compiled code to stay around until enough traps are seen,
1756 // and until the compiler gets around to recompiling the trapping method.
1757 //
1758 // The other actions cause immediate removal of the present code.
1759
1760 // Traps caused by injected profile shouldn't pollute trap counts.
1761 bool injected_profile_trap = trap_method->has_injected_profile() &&
1762 (reason == Reason_intrinsic || reason == Reason_unreached);
1763
1764 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
1765 bool make_not_entrant = false;
1766 bool make_not_compilable = false;
1767 bool reprofile = false;
1768 switch (action) {
1769 case Action_none:
1770 // Keep the old code.
1771 update_trap_state = false;
1772 break;
1773 case Action_maybe_recompile:
1774 // Do not need to invalidate the present code, but we can
1775 // initiate another
1776 // Start compiler without (necessarily) invalidating the nmethod.
1777 // The system will tolerate the old code, but new code should be
1778 // generated when possible.
1779 break;
1780 case Action_reinterpret:
1781 // Go back into the interpreter for a while, and then consider
1782 // recompiling form scratch.
1783 make_not_entrant = true;
1784 // Reset invocation counter for outer most method.
1785 // This will allow the interpreter to exercise the bytecodes
1786 // for a while before recompiling.
1787 // By contrast, Action_make_not_entrant is immediate.
1788 //
1789 // Note that the compiler will track null_check, null_assert,
1790 // range_check, and class_check events and log them as if they
1791 // had been traps taken from compiled code. This will update
1792 // the MDO trap history so that the next compilation will
1793 // properly detect hot trap sites.
1794 reprofile = true;
1795 break;
1796 case Action_make_not_entrant:
1797 // Request immediate recompilation, and get rid of the old code.
1798 // Make them not entrant, so next time they are called they get
1799 // recompiled. Unloaded classes are loaded now so recompile before next
1800 // time they are called. Same for uninitialized. The interpreter will
1801 // link the missing class, if any.
1802 make_not_entrant = true;
1803 break;
1804 case Action_make_not_compilable:
1805 // Give up on compiling this method at all.
1806 make_not_entrant = true;
1807 make_not_compilable = true;
1808 break;
1809 default:
1810 ShouldNotReachHere();
1811 }
1812
1813 // Setting +ProfileTraps fixes the following, on all platforms:
1814 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1815 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1816 // recompile relies on a MethodData* to record heroic opt failures.
1817
1818 // Whether the interpreter is producing MDO data or not, we also need
1819 // to use the MDO to detect hot deoptimization points and control
1820 // aggressive optimization.
1821 bool inc_recompile_count = false;
1822 ProfileData* pdata = NULL;
1823 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) {
1824 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
1825 uint this_trap_count = 0;
1826 bool maybe_prior_trap = false;
1827 bool maybe_prior_recompile = false;
1828 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
1829 #if INCLUDE_JVMCI
1830 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
1831 #endif
1832 nm->method(),
1833 //outputs:
1834 this_trap_count,
1835 maybe_prior_trap,
1836 maybe_prior_recompile);
1837 // Because the interpreter also counts null, div0, range, and class
1838 // checks, these traps from compiled code are double-counted.
1839 // This is harmless; it just means that the PerXTrapLimit values
1840 // are in effect a little smaller than they look.
1841
1842 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1843 if (per_bc_reason != Reason_none) {
1844 // Now take action based on the partially known per-BCI history.
1845 if (maybe_prior_trap
1846 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1847 // If there are too many traps at this BCI, force a recompile.
1848 // This will allow the compiler to see the limit overflow, and
1849 // take corrective action, if possible. The compiler generally
1850 // does not use the exact PerBytecodeTrapLimit value, but instead
1851 // changes its tactics if it sees any traps at all. This provides
1852 // a little hysteresis, delaying a recompile until a trap happens
1853 // several times.
1854 //
1855 // Actually, since there is only one bit of counter per BCI,
1856 // the possible per-BCI counts are {0,1,(per-method count)}.
1857 // This produces accurate results if in fact there is only
1858 // one hot trap site, but begins to get fuzzy if there are
1859 // many sites. For example, if there are ten sites each
1860 // trapping two or more times, they each get the blame for
1861 // all of their traps.
1862 make_not_entrant = true;
1863 }
1864
1865 // Detect repeated recompilation at the same BCI, and enforce a limit.
1866 if (make_not_entrant && maybe_prior_recompile) {
1867 // More than one recompile at this point.
1868 inc_recompile_count = maybe_prior_trap;
1869 }
1870 } else {
1871 // For reasons which are not recorded per-bytecode, we simply
1872 // force recompiles unconditionally.
1873 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1874 make_not_entrant = true;
1875 }
1876
1877 // Go back to the compiler if there are too many traps in this method.
1878 if (this_trap_count >= per_method_trap_limit(reason)) {
1879 // If there are too many traps in this method, force a recompile.
1880 // This will allow the compiler to see the limit overflow, and
1881 // take corrective action, if possible.
1882 // (This condition is an unlikely backstop only, because the
1883 // PerBytecodeTrapLimit is more likely to take effect first,
1884 // if it is applicable.)
1885 make_not_entrant = true;
1886 }
1887
1888 // Here's more hysteresis: If there has been a recompile at
1889 // this trap point already, run the method in the interpreter
1890 // for a while to exercise it more thoroughly.
1891 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1892 reprofile = true;
1893 }
1894 }
1895
1896 // Take requested actions on the method:
1897
1898 // Recompile
1899 if (make_not_entrant) {
1900 if (!nm->make_not_entrant()) {
1901 return; // the call did not change nmethod's state
1902 }
1903
1904 if (pdata != NULL) {
1905 // Record the recompilation event, if any.
1906 int tstate0 = pdata->trap_state();
1907 int tstate1 = trap_state_set_recompiled(tstate0, true);
1908 if (tstate1 != tstate0)
1909 pdata->set_trap_state(tstate1);
1910 }
1911
1912 #if INCLUDE_RTM_OPT
1913 // Restart collecting RTM locking abort statistic if the method
1914 // is recompiled for a reason other than RTM state change.
1915 // Assume that in new recompiled code the statistic could be different,
1916 // for example, due to different inlining.
1917 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
1918 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
1919 trap_mdo->atomic_set_rtm_state(ProfileRTM);
1920 }
1921 #endif
1922 // For code aging we count traps separately here, using make_not_entrant()
1923 // as a guard against simultaneous deopts in multiple threads.
1924 if (reason == Reason_tenured && trap_mdo != NULL) {
1925 trap_mdo->inc_tenure_traps();
1926 }
1927 }
1928
1929 if (inc_recompile_count) {
1930 trap_mdo->inc_overflow_recompile_count();
1931 if ((uint)trap_mdo->overflow_recompile_count() >
1932 (uint)PerBytecodeRecompilationCutoff) {
1933 // Give up on the method containing the bad BCI.
1934 if (trap_method() == nm->method()) {
1935 make_not_compilable = true;
1936 } else {
1937 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
1938 // But give grace to the enclosing nm->method().
1939 }
1940 }
1941 }
1942
1943 // Reprofile
1944 if (reprofile) {
1945 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1946 }
1947
1948 // Give up compiling
1949 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1950 assert(make_not_entrant, "consistent");
1951 nm->method()->set_not_compilable(CompLevel_full_optimization);
1952 }
1953
1954 } // Free marked resources
1955
1956 }
1957 JRT_END
1958
1959 ProfileData*
1960 Deoptimization::query_update_method_data(MethodData* trap_mdo,
1961 int trap_bci,
1962 Deoptimization::DeoptReason reason,
1963 bool update_total_trap_count,
1964 #if INCLUDE_JVMCI
1965 bool is_osr,
1966 #endif
1967 Method* compiled_method,
1968 //outputs:
1969 uint& ret_this_trap_count,
1970 bool& ret_maybe_prior_trap,
1971 bool& ret_maybe_prior_recompile) {
1972 bool maybe_prior_trap = false;
1973 bool maybe_prior_recompile = false;
1974 uint this_trap_count = 0;
1975 if (update_total_trap_count) {
1976 uint idx = reason;
1977 #if INCLUDE_JVMCI
1978 if (is_osr) {
1979 idx += Reason_LIMIT;
1980 }
1981 #endif
1982 uint prior_trap_count = trap_mdo->trap_count(idx);
1983 this_trap_count = trap_mdo->inc_trap_count(idx);
1984
1985 // If the runtime cannot find a place to store trap history,
1986 // it is estimated based on the general condition of the method.
1987 // If the method has ever been recompiled, or has ever incurred
1988 // a trap with the present reason , then this BCI is assumed
1989 // (pessimistically) to be the culprit.
1990 maybe_prior_trap = (prior_trap_count != 0);
1991 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1992 }
1993 ProfileData* pdata = NULL;
1994
1995
1996 // For reasons which are recorded per bytecode, we check per-BCI data.
1997 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1998 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
1999 if (per_bc_reason != Reason_none) {
2000 // Find the profile data for this BCI. If there isn't one,
2001 // try to allocate one from the MDO's set of spares.
2002 // This will let us detect a repeated trap at this point.
2003 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
2004
2005 if (pdata != NULL) {
2006 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2007 if (LogCompilation && xtty != NULL) {
2008 ttyLocker ttyl;
2009 // no more room for speculative traps in this MDO
2010 xtty->elem("speculative_traps_oom");
2011 }
2012 }
2013 // Query the trap state of this profile datum.
2014 int tstate0 = pdata->trap_state();
2015 if (!trap_state_has_reason(tstate0, per_bc_reason))
2016 maybe_prior_trap = false;
2017 if (!trap_state_is_recompiled(tstate0))
2018 maybe_prior_recompile = false;
2019
2020 // Update the trap state of this profile datum.
2021 int tstate1 = tstate0;
2022 // Record the reason.
2023 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2024 // Store the updated state on the MDO, for next time.
2025 if (tstate1 != tstate0)
2026 pdata->set_trap_state(tstate1);
2027 } else {
2028 if (LogCompilation && xtty != NULL) {
2029 ttyLocker ttyl;
2030 // Missing MDP? Leave a small complaint in the log.
2031 xtty->elem("missing_mdp bci='%d'", trap_bci);
2032 }
2033 }
2034 }
2035
2036 // Return results:
2037 ret_this_trap_count = this_trap_count;
2038 ret_maybe_prior_trap = maybe_prior_trap;
2039 ret_maybe_prior_recompile = maybe_prior_recompile;
2040 return pdata;
2041 }
2042
2043 void
2044 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2045 ResourceMark rm;
2046 // Ignored outputs:
2047 uint ignore_this_trap_count;
2048 bool ignore_maybe_prior_trap;
2049 bool ignore_maybe_prior_recompile;
2050 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2051 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2052 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2053 query_update_method_data(trap_mdo, trap_bci,
2054 (DeoptReason)reason,
2055 update_total_counts,
2056 #if INCLUDE_JVMCI
2057 false,
2058 #endif
2059 NULL,
2060 ignore_this_trap_count,
2061 ignore_maybe_prior_trap,
2062 ignore_maybe_prior_recompile);
2063 }
2064
2065 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2066 if (TraceDeoptimization) {
2067 tty->print("Uncommon trap ");
2068 }
2069 // Still in Java no safepoints
2070 {
2071 // This enters VM and may safepoint
2072 uncommon_trap_inner(thread, trap_request);
2073 }
2074 return fetch_unroll_info_helper(thread, exec_mode);
2075 }
2076
2077 // Local derived constants.
2078 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2079 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2080 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2081
2082 //---------------------------trap_state_reason---------------------------------
2083 Deoptimization::DeoptReason
2084 Deoptimization::trap_state_reason(int trap_state) {
2085 // This assert provides the link between the width of DataLayout::trap_bits
2086 // and the encoding of "recorded" reasons. It ensures there are enough
2087 // bits to store all needed reasons in the per-BCI MDO profile.
2088 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2089 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2090 trap_state -= recompile_bit;
2091 if (trap_state == DS_REASON_MASK) {
2092 return Reason_many;
2093 } else {
2094 assert((int)Reason_none == 0, "state=0 => Reason_none");
2095 return (DeoptReason)trap_state;
2096 }
2097 }
2098 //-------------------------trap_state_has_reason-------------------------------
2099 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2100 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2101 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2102 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2103 trap_state -= recompile_bit;
2104 if (trap_state == DS_REASON_MASK) {
2105 return -1; // true, unspecifically (bottom of state lattice)
2106 } else if (trap_state == reason) {
2107 return 1; // true, definitely
2108 } else if (trap_state == 0) {
2109 return 0; // false, definitely (top of state lattice)
2110 } else {
2111 return 0; // false, definitely
2112 }
2113 }
2114 //-------------------------trap_state_add_reason-------------------------------
2115 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2116 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2117 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2118 trap_state -= recompile_bit;
2119 if (trap_state == DS_REASON_MASK) {
2120 return trap_state + recompile_bit; // already at state lattice bottom
2121 } else if (trap_state == reason) {
2122 return trap_state + recompile_bit; // the condition is already true
2123 } else if (trap_state == 0) {
2124 return reason + recompile_bit; // no condition has yet been true
2125 } else {
2126 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2127 }
2128 }
2129 //-----------------------trap_state_is_recompiled------------------------------
2130 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2131 return (trap_state & DS_RECOMPILE_BIT) != 0;
2132 }
2133 //-----------------------trap_state_set_recompiled-----------------------------
2134 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2135 if (z) return trap_state | DS_RECOMPILE_BIT;
2136 else return trap_state & ~DS_RECOMPILE_BIT;
2137 }
2138 //---------------------------format_trap_state---------------------------------
2139 // This is used for debugging and diagnostics, including LogFile output.
2140 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2141 int trap_state) {
2142 assert(buflen > 0, "sanity");
2143 DeoptReason reason = trap_state_reason(trap_state);
2144 bool recomp_flag = trap_state_is_recompiled(trap_state);
2145 // Re-encode the state from its decoded components.
2146 int decoded_state = 0;
2147 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2148 decoded_state = trap_state_add_reason(decoded_state, reason);
2149 if (recomp_flag)
2150 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2151 // If the state re-encodes properly, format it symbolically.
2152 // Because this routine is used for debugging and diagnostics,
2153 // be robust even if the state is a strange value.
2154 size_t len;
2155 if (decoded_state != trap_state) {
2156 // Random buggy state that doesn't decode??
2157 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2158 } else {
2159 len = jio_snprintf(buf, buflen, "%s%s",
2160 trap_reason_name(reason),
2161 recomp_flag ? " recompiled" : "");
2162 }
2163 return buf;
2164 }
2165
2166
2167 //--------------------------------statics--------------------------------------
2168 const char* Deoptimization::_trap_reason_name[] = {
2169 // Note: Keep this in sync. with enum DeoptReason.
2170 "none",
2171 "null_check",
2172 "null_assert" JVMCI_ONLY("_or_unreached0"),
2173 "range_check",
2174 "class_check",
2175 "array_check",
2176 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2177 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2178 "profile_predicate",
2179 "unloaded",
2180 "uninitialized",
2181 "unreached",
2182 "unhandled",
2183 "constraint",
2184 "div0_check",
2185 "age",
2186 "predicate",
2187 "loop_limit_check",
2188 "speculate_class_check",
2189 "speculate_null_check",
2190 "speculate_null_assert",
2191 "rtm_state_change",
2192 "unstable_if",
2193 "unstable_fused_if",
2194 #if INCLUDE_JVMCI
2195 "aliasing",
2196 "transfer_to_interpreter",
2197 "not_compiled_exception_handler",
2198 "unresolved",
2199 "jsr_mismatch",
2200 #endif
2201 "tenured"
2202 };
2203 const char* Deoptimization::_trap_action_name[] = {
2204 // Note: Keep this in sync. with enum DeoptAction.
2205 "none",
2206 "maybe_recompile",
2207 "reinterpret",
2208 "make_not_entrant",
2209 "make_not_compilable"
2210 };
2211
2212 const char* Deoptimization::trap_reason_name(int reason) {
2213 // Check that every reason has a name
2214 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2215
2216 if (reason == Reason_many) return "many";
2217 if ((uint)reason < Reason_LIMIT)
2218 return _trap_reason_name[reason];
2219 static char buf[20];
2220 sprintf(buf, "reason%d", reason);
2221 return buf;
2222 }
2223 const char* Deoptimization::trap_action_name(int action) {
2224 // Check that every action has a name
2225 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2226
2227 if ((uint)action < Action_LIMIT)
2228 return _trap_action_name[action];
2229 static char buf[20];
2230 sprintf(buf, "action%d", action);
2231 return buf;
2232 }
2233
2234 // This is used for debugging and diagnostics, including LogFile output.
2235 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2236 int trap_request) {
2237 jint unloaded_class_index = trap_request_index(trap_request);
2238 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2239 const char* action = trap_action_name(trap_request_action(trap_request));
2240 #if INCLUDE_JVMCI
2241 int debug_id = trap_request_debug_id(trap_request);
2242 #endif
2243 size_t len;
2244 if (unloaded_class_index < 0) {
2245 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2246 reason, action
2247 #if INCLUDE_JVMCI
2248 ,debug_id
2249 #endif
2250 );
2251 } else {
2252 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2253 reason, action, unloaded_class_index
2254 #if INCLUDE_JVMCI
2255 ,debug_id
2256 #endif
2257 );
2258 }
2259 return buf;
2260 }
2261
2262 juint Deoptimization::_deoptimization_hist
2263 [Deoptimization::Reason_LIMIT]
2264 [1 + Deoptimization::Action_LIMIT]
2265 [Deoptimization::BC_CASE_LIMIT]
2266 = {0};
2267
2268 enum {
2269 LSB_BITS = 8,
2270 LSB_MASK = right_n_bits(LSB_BITS)
2271 };
2272
2273 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2274 Bytecodes::Code bc) {
2275 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2276 assert(action >= 0 && action < Action_LIMIT, "oob");
2277 _deoptimization_hist[Reason_none][0][0] += 1; // total
2278 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2279 juint* cases = _deoptimization_hist[reason][1+action];
2280 juint* bc_counter_addr = NULL;
2281 juint bc_counter = 0;
2282 // Look for an unused counter, or an exact match to this BC.
2283 if (bc != Bytecodes::_illegal) {
2284 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2285 juint* counter_addr = &cases[bc_case];
2286 juint counter = *counter_addr;
2287 if ((counter == 0 && bc_counter_addr == NULL)
2288 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2289 // this counter is either free or is already devoted to this BC
2290 bc_counter_addr = counter_addr;
2291 bc_counter = counter | bc;
2292 }
2293 }
2294 }
2295 if (bc_counter_addr == NULL) {
2296 // Overflow, or no given bytecode.
2297 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2298 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2299 }
2300 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2301 }
2302
2303 jint Deoptimization::total_deoptimization_count() {
2304 return _deoptimization_hist[Reason_none][0][0];
2305 }
2306
2307 void Deoptimization::print_statistics() {
2308 juint total = total_deoptimization_count();
2309 juint account = total;
2310 if (total != 0) {
2311 ttyLocker ttyl;
2312 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2313 tty->print_cr("Deoptimization traps recorded:");
2314 #define PRINT_STAT_LINE(name, r) \
2315 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2316 PRINT_STAT_LINE("total", total);
2317 // For each non-zero entry in the histogram, print the reason,
2318 // the action, and (if specifically known) the type of bytecode.
2319 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2320 for (int action = 0; action < Action_LIMIT; action++) {
2321 juint* cases = _deoptimization_hist[reason][1+action];
2322 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2323 juint counter = cases[bc_case];
2324 if (counter != 0) {
2325 char name[1*K];
2326 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2327 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2328 bc = Bytecodes::_illegal;
2329 sprintf(name, "%s/%s/%s",
2330 trap_reason_name(reason),
2331 trap_action_name(action),
2332 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2333 juint r = counter >> LSB_BITS;
2334 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2335 account -= r;
2336 }
2337 }
2338 }
2339 }
2340 if (account != 0) {
2341 PRINT_STAT_LINE("unaccounted", account);
2342 }
2343 #undef PRINT_STAT_LINE
2344 if (xtty != NULL) xtty->tail("statistics");
2345 }
2346 }
2347 #else // COMPILER2_OR_JVMCI
2348
2349
2350 // Stubs for C1 only system.
2351 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2352 return false;
2353 }
2354
2355 const char* Deoptimization::trap_reason_name(int reason) {
2356 return "unknown";
2357 }
2358
2359 void Deoptimization::print_statistics() {
2360 // no output
2361 }
2362
2363 void
2364 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2365 // no udpate
2366 }
2367
2368 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2369 return 0;
2370 }
2371
2372 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2373 Bytecodes::Code bc) {
2374 // no update
2375 }
2376
2377 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2378 int trap_state) {
2379 jio_snprintf(buf, buflen, "#%d", trap_state);
2380 return buf;
2381 }
2382
2383 #endif // COMPILER2_OR_JVMCI
--- EOF ---