1 /*
2 * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "jvm.h"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/codeCache.hpp"
29 #include "code/debugInfoRec.hpp"
30 #include "code/nmethod.hpp"
31 #include "code/pcDesc.hpp"
32 #include "code/scopeDesc.hpp"
33 #include "interpreter/bytecode.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "interpreter/oopMapCache.hpp"
36 #include "memory/allocation.inline.hpp"
37 #include "memory/oopFactory.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "oops/method.hpp"
40 #include "oops/objArrayOop.inline.hpp"
41 #include "oops/oop.inline.hpp"
42 #include "oops/fieldStreams.hpp"
43 #include "oops/valueArrayKlass.hpp"
44 #include "oops/valueArrayOop.hpp"
45 #include "oops/valueKlass.hpp"
46 #include "oops/verifyOopClosure.hpp"
47 #include "prims/jvmtiThreadState.hpp"
48 #include "runtime/biasedLocking.hpp"
49 #include "runtime/compilationPolicy.hpp"
50 #include "runtime/deoptimization.hpp"
51 #include "runtime/interfaceSupport.hpp"
52 #include "runtime/sharedRuntime.hpp"
53 #include "runtime/signature.hpp"
54 #include "runtime/stubRoutines.hpp"
55 #include "runtime/thread.hpp"
56 #include "runtime/threadSMR.hpp"
57 #include "runtime/vframe.hpp"
58 #include "runtime/vframeArray.hpp"
59 #include "runtime/vframe_hp.hpp"
60 #include "utilities/events.hpp"
61 #include "utilities/xmlstream.hpp"
62
63 #if INCLUDE_JVMCI
64 #include "jvmci/jvmciRuntime.hpp"
65 #include "jvmci/jvmciJavaClasses.hpp"
66 #endif
67
68
69 bool DeoptimizationMarker::_is_active = false;
70
71 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
72 int caller_adjustment,
73 int caller_actual_parameters,
74 int number_of_frames,
75 intptr_t* frame_sizes,
76 address* frame_pcs,
77 BasicType return_type,
78 int exec_mode) {
79 _size_of_deoptimized_frame = size_of_deoptimized_frame;
80 _caller_adjustment = caller_adjustment;
81 _caller_actual_parameters = caller_actual_parameters;
82 _number_of_frames = number_of_frames;
83 _frame_sizes = frame_sizes;
84 _frame_pcs = frame_pcs;
85 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
86 _return_type = return_type;
87 _initial_info = 0;
88 // PD (x86 only)
89 _counter_temp = 0;
90 _unpack_kind = exec_mode;
91 _sender_sp_temp = 0;
92
93 _total_frame_sizes = size_of_frames();
94 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
95 }
96
97
98 Deoptimization::UnrollBlock::~UnrollBlock() {
99 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
100 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
101 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
102 }
103
104
105 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
106 assert(register_number < RegisterMap::reg_count, "checking register number");
107 return &_register_block[register_number * 2];
108 }
109
110
111
112 int Deoptimization::UnrollBlock::size_of_frames() const {
113 // Acount first for the adjustment of the initial frame
114 int result = _caller_adjustment;
115 for (int index = 0; index < number_of_frames(); index++) {
116 result += frame_sizes()[index];
117 }
118 return result;
119 }
120
121
122 void Deoptimization::UnrollBlock::print() {
123 ttyLocker ttyl;
124 tty->print_cr("UnrollBlock");
125 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
126 tty->print( " frame_sizes: ");
127 for (int index = 0; index < number_of_frames(); index++) {
128 tty->print(INTX_FORMAT " ", frame_sizes()[index]);
129 }
130 tty->cr();
131 }
132
133
134 // In order to make fetch_unroll_info work properly with escape
135 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
136 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
137 // of previously eliminated objects occurs in realloc_objects, which is
138 // called from the method fetch_unroll_info_helper below.
139 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode))
140 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
141 // but makes the entry a little slower. There is however a little dance we have to
142 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
143
144 // fetch_unroll_info() is called at the beginning of the deoptimization
145 // handler. Note this fact before we start generating temporary frames
146 // that can confuse an asynchronous stack walker. This counter is
147 // decremented at the end of unpack_frames().
148 if (TraceDeoptimization) {
149 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread));
150 }
151 thread->inc_in_deopt_handler();
152
153 return fetch_unroll_info_helper(thread, exec_mode);
154 JRT_END
155
156
157 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
158 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) {
159
160 // Note: there is a safepoint safety issue here. No matter whether we enter
161 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
162 // the vframeArray is created.
163 //
164
165 // Allocate our special deoptimization ResourceMark
166 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
167 assert(thread->deopt_mark() == NULL, "Pending deopt!");
168 thread->set_deopt_mark(dmark);
169
170 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
171 RegisterMap map(thread, true);
172 RegisterMap dummy_map(thread, false);
173 // Now get the deoptee with a valid map
174 frame deoptee = stub_frame.sender(&map);
175 // Set the deoptee nmethod
176 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!");
177 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
178 thread->set_deopt_compiled_method(cm);
179
180 if (VerifyStack) {
181 thread->validate_frame_layout();
182 }
183
184 // Create a growable array of VFrames where each VFrame represents an inlined
185 // Java frame. This storage is allocated with the usual system arena.
186 assert(deoptee.is_compiled_frame(), "Wrong frame type");
187 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
188 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
189 while (!vf->is_top()) {
190 assert(vf->is_compiled_frame(), "Wrong frame type");
191 chunk->push(compiledVFrame::cast(vf));
192 vf = vf->sender();
193 }
194 assert(vf->is_compiled_frame(), "Wrong frame type");
195 chunk->push(compiledVFrame::cast(vf));
196
197 bool realloc_failures = false;
198
199 #if COMPILER2_OR_JVMCI
200 // Reallocate the non-escaping objects and restore their fields. Then
201 // relock objects if synchronization on them was eliminated.
202 #ifndef INCLUDE_JVMCI
203 if (DoEscapeAnalysis || EliminateNestedLocks) {
204 if (EliminateAllocations) {
205 #endif // INCLUDE_JVMCI
206 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
207 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
208
209 // The flag return_oop() indicates call sites which return oop
210 // in compiled code. Such sites include java method calls,
211 // runtime calls (for example, used to allocate new objects/arrays
212 // on slow code path) and any other calls generated in compiled code.
213 // It is not guaranteed that we can get such information here only
214 // by analyzing bytecode in deoptimized frames. This is why this flag
215 // is set during method compilation (see Compile::Process_OopMap_Node()).
216 // If the previous frame was popped or if we are dispatching an exception,
217 // we don't have an oop result.
218 ScopeDesc* scope = chunk->at(0)->scope();
219 bool save_oop_result = scope->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt);
220 // In case of the return of multiple values, we must take care
221 // of all oop return values.
222 GrowableArray<Handle> return_oops;
223 ValueKlass* vk = NULL;
224 if (save_oop_result && scope->return_vt()) {
225 vk = ValueKlass::returned_value_klass(map);
226 if (vk != NULL) {
227 vk->save_oop_fields(map, return_oops);
228 save_oop_result = false;
229 }
230 }
231 if (save_oop_result) {
232 // Reallocation may trigger GC. If deoptimization happened on return from
233 // call which returns oop we need to save it since it is not in oopmap.
234 oop result = deoptee.saved_oop_result(&map);
235 assert(oopDesc::is_oop_or_null(result), "must be oop");
236 return_oops.push(Handle(thread, result));
237 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
238 if (TraceDeoptimization) {
239 ttyLocker ttyl;
240 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
241 }
242 }
243 if (objects != NULL || vk != NULL) {
244 bool skip_internal = (cm != NULL) && !cm->is_compiled_by_jvmci();
245 JRT_BLOCK
246 if (vk != NULL) {
247 realloc_failures = realloc_value_type_result(vk, map, return_oops, THREAD);
248 }
249 if (objects != NULL) {
250 realloc_failures = realloc_failures || realloc_objects(thread, &deoptee, objects, THREAD);
251 reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, THREAD);
252 }
253 JRT_END
254 #ifndef PRODUCT
255 if (TraceDeoptimization) {
256 ttyLocker ttyl;
257 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
258 print_objects(objects, realloc_failures);
259 }
260 #endif
261 }
262 if (save_oop_result || vk != NULL) {
263 // Restore result.
264 assert(return_oops.length() == 1, "no value type");
265 deoptee.set_saved_oop_result(&map, return_oops.pop()());
266 }
267 #ifndef INCLUDE_JVMCI
268 }
269 if (EliminateLocks) {
270 #endif // INCLUDE_JVMCI
271 #ifndef PRODUCT
272 bool first = true;
273 #endif
274 for (int i = 0; i < chunk->length(); i++) {
275 compiledVFrame* cvf = chunk->at(i);
276 assert (cvf->scope() != NULL,"expect only compiled java frames");
277 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
278 if (monitors->is_nonempty()) {
279 relock_objects(monitors, thread, realloc_failures);
280 #ifndef PRODUCT
281 if (PrintDeoptimizationDetails) {
282 ttyLocker ttyl;
283 for (int j = 0; j < monitors->length(); j++) {
284 MonitorInfo* mi = monitors->at(j);
285 if (mi->eliminated()) {
286 if (first) {
287 first = false;
288 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
289 }
290 if (mi->owner_is_scalar_replaced()) {
291 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
292 tty->print_cr(" failed reallocation for klass %s", k->external_name());
293 } else {
294 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
295 }
296 }
297 }
298 }
299 #endif // !PRODUCT
300 }
301 }
302 #ifndef INCLUDE_JVMCI
303 }
304 }
305 #endif // INCLUDE_JVMCI
306 #endif // COMPILER2_OR_JVMCI
307
308 ScopeDesc* trap_scope = chunk->at(0)->scope();
309 Handle exceptionObject;
310 if (trap_scope->rethrow_exception()) {
311 if (PrintDeoptimizationDetails) {
312 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());
313 }
314 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
315 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw");
316 ScopeValue* topOfStack = expressions->top();
317 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
318 guarantee(exceptionObject() != NULL, "exception oop can not be null");
319 }
320
321 // Ensure that no safepoint is taken after pointers have been stored
322 // in fields of rematerialized objects. If a safepoint occurs from here on
323 // out the java state residing in the vframeArray will be missed.
324 NoSafepointVerifier no_safepoint;
325
326 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
327 #if COMPILER2_OR_JVMCI
328 if (realloc_failures) {
329 pop_frames_failed_reallocs(thread, array);
330 }
331 #endif
332
333 assert(thread->vframe_array_head() == NULL, "Pending deopt!");
334 thread->set_vframe_array_head(array);
335
336 // Now that the vframeArray has been created if we have any deferred local writes
337 // added by jvmti then we can free up that structure as the data is now in the
338 // vframeArray
339
340 if (thread->deferred_locals() != NULL) {
341 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
342 int i = 0;
343 do {
344 // Because of inlining we could have multiple vframes for a single frame
345 // and several of the vframes could have deferred writes. Find them all.
346 if (list->at(i)->id() == array->original().id()) {
347 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
348 list->remove_at(i);
349 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
350 delete dlv;
351 } else {
352 i++;
353 }
354 } while ( i < list->length() );
355 if (list->length() == 0) {
356 thread->set_deferred_locals(NULL);
357 // free the list and elements back to C heap.
358 delete list;
359 }
360
361 }
362
363 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
364 CodeBlob* cb = stub_frame.cb();
365 // Verify we have the right vframeArray
366 assert(cb->frame_size() >= 0, "Unexpected frame size");
367 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
368
369 // If the deopt call site is a MethodHandle invoke call site we have
370 // to adjust the unpack_sp.
371 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
372 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
373 unpack_sp = deoptee.unextended_sp();
374
375 #ifdef ASSERT
376 assert(cb->is_deoptimization_stub() ||
377 cb->is_uncommon_trap_stub() ||
378 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
379 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
380 "unexpected code blob: %s", cb->name());
381 #endif
382
383 // This is a guarantee instead of an assert because if vframe doesn't match
384 // we will unpack the wrong deoptimized frame and wind up in strange places
385 // where it will be very difficult to figure out what went wrong. Better
386 // to die an early death here than some very obscure death later when the
387 // trail is cold.
388 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
389 // in that it will fail to detect a problem when there is one. This needs
390 // more work in tiger timeframe.
391 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
392
393 int number_of_frames = array->frames();
394
395 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
396 // virtual activation, which is the reverse of the elements in the vframes array.
397 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
398 // +1 because we always have an interpreter return address for the final slot.
399 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
400 int popframe_extra_args = 0;
401 // Create an interpreter return address for the stub to use as its return
402 // address so the skeletal frames are perfectly walkable
403 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
404
405 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
406 // activation be put back on the expression stack of the caller for reexecution
407 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
408 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
409 }
410
411 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
412 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
413 // than simply use array->sender.pc(). This requires us to walk the current set of frames
414 //
415 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
416 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
417
418 // It's possible that the number of parameters at the call site is
419 // different than number of arguments in the callee when method
420 // handles are used. If the caller is interpreted get the real
421 // value so that the proper amount of space can be added to it's
422 // frame.
423 bool caller_was_method_handle = false;
424 if (deopt_sender.is_interpreted_frame()) {
425 methodHandle method = deopt_sender.interpreter_frame_method();
426 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
427 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
428 // Method handle invokes may involve fairly arbitrary chains of
429 // calls so it's impossible to know how much actual space the
430 // caller has for locals.
431 caller_was_method_handle = true;
432 }
433 }
434
435 //
436 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
437 // frame_sizes/frame_pcs[1] next oldest frame (int)
438 // frame_sizes/frame_pcs[n] youngest frame (int)
439 //
440 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
441 // owns the space for the return address to it's caller). Confusing ain't it.
442 //
443 // The vframe array can address vframes with indices running from
444 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
445 // When we create the skeletal frames we need the oldest frame to be in the zero slot
446 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
447 // so things look a little strange in this loop.
448 //
449 int callee_parameters = 0;
450 int callee_locals = 0;
451 for (int index = 0; index < array->frames(); index++ ) {
452 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
453 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
454 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
455 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
456 callee_locals,
457 index == 0,
458 popframe_extra_args);
459 // This pc doesn't have to be perfect just good enough to identify the frame
460 // as interpreted so the skeleton frame will be walkable
461 // The correct pc will be set when the skeleton frame is completely filled out
462 // The final pc we store in the loop is wrong and will be overwritten below
463 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
464
465 callee_parameters = array->element(index)->method()->size_of_parameters();
466 callee_locals = array->element(index)->method()->max_locals();
467 popframe_extra_args = 0;
468 }
469
470 // Compute whether the root vframe returns a float or double value.
471 BasicType return_type;
472 {
473 methodHandle method(thread, array->element(0)->method());
474 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
475 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
476 }
477
478 // Compute information for handling adapters and adjusting the frame size of the caller.
479 int caller_adjustment = 0;
480
481 // Compute the amount the oldest interpreter frame will have to adjust
482 // its caller's stack by. If the caller is a compiled frame then
483 // we pretend that the callee has no parameters so that the
484 // extension counts for the full amount of locals and not just
485 // locals-parms. This is because without a c2i adapter the parm
486 // area as created by the compiled frame will not be usable by
487 // the interpreter. (Depending on the calling convention there
488 // may not even be enough space).
489
490 // QQQ I'd rather see this pushed down into last_frame_adjust
491 // and have it take the sender (aka caller).
492
493 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
494 caller_adjustment = last_frame_adjust(0, callee_locals);
495 } else if (callee_locals > callee_parameters) {
496 // The caller frame may need extending to accommodate
497 // non-parameter locals of the first unpacked interpreted frame.
498 // Compute that adjustment.
499 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
500 }
501
502 // If the sender is deoptimized the we must retrieve the address of the handler
503 // since the frame will "magically" show the original pc before the deopt
504 // and we'd undo the deopt.
505
506 frame_pcs[0] = deopt_sender.raw_pc();
507
508 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
509
510 #ifdef INCLUDE_JVMCI
511 if (exceptionObject() != NULL) {
512 thread->set_exception_oop(exceptionObject());
513 exec_mode = Unpack_exception;
514 }
515 #endif
516
517 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
518 assert(thread->has_pending_exception(), "should have thrown OOME");
519 thread->set_exception_oop(thread->pending_exception());
520 thread->clear_pending_exception();
521 exec_mode = Unpack_exception;
522 }
523
524 #if INCLUDE_JVMCI
525 if (thread->frames_to_pop_failed_realloc() > 0) {
526 thread->set_pending_monitorenter(false);
527 }
528 #endif
529
530 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
531 caller_adjustment * BytesPerWord,
532 caller_was_method_handle ? 0 : callee_parameters,
533 number_of_frames,
534 frame_sizes,
535 frame_pcs,
536 return_type,
537 exec_mode);
538 // On some platforms, we need a way to pass some platform dependent
539 // information to the unpacking code so the skeletal frames come out
540 // correct (initial fp value, unextended sp, ...)
541 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
542
543 if (array->frames() > 1) {
544 if (VerifyStack && TraceDeoptimization) {
545 ttyLocker ttyl;
546 tty->print_cr("Deoptimizing method containing inlining");
547 }
548 }
549
550 array->set_unroll_block(info);
551 return info;
552 }
553
554 // Called to cleanup deoptimization data structures in normal case
555 // after unpacking to stack and when stack overflow error occurs
556 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
557 vframeArray *array) {
558
559 // Get array if coming from exception
560 if (array == NULL) {
561 array = thread->vframe_array_head();
562 }
563 thread->set_vframe_array_head(NULL);
564
565 // Free the previous UnrollBlock
566 vframeArray* old_array = thread->vframe_array_last();
567 thread->set_vframe_array_last(array);
568
569 if (old_array != NULL) {
570 UnrollBlock* old_info = old_array->unroll_block();
571 old_array->set_unroll_block(NULL);
572 delete old_info;
573 delete old_array;
574 }
575
576 // Deallocate any resource creating in this routine and any ResourceObjs allocated
577 // inside the vframeArray (StackValueCollections)
578
579 delete thread->deopt_mark();
580 thread->set_deopt_mark(NULL);
581 thread->set_deopt_compiled_method(NULL);
582
583
584 if (JvmtiExport::can_pop_frame()) {
585 #ifndef CC_INTERP
586 // Regardless of whether we entered this routine with the pending
587 // popframe condition bit set, we should always clear it now
588 thread->clear_popframe_condition();
589 #else
590 // C++ interpreter will clear has_pending_popframe when it enters
591 // with method_resume. For deopt_resume2 we clear it now.
592 if (thread->popframe_forcing_deopt_reexecution())
593 thread->clear_popframe_condition();
594 #endif /* CC_INTERP */
595 }
596
597 // unpack_frames() is called at the end of the deoptimization handler
598 // and (in C2) at the end of the uncommon trap handler. Note this fact
599 // so that an asynchronous stack walker can work again. This counter is
600 // incremented at the beginning of fetch_unroll_info() and (in C2) at
601 // the beginning of uncommon_trap().
602 thread->dec_in_deopt_handler();
603 }
604
605 // Moved from cpu directories because none of the cpus has callee save values.
606 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
607 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
608
609 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
610 // the days we had adapter frames. When we deoptimize a situation where a
611 // compiled caller calls a compiled caller will have registers it expects
612 // to survive the call to the callee. If we deoptimize the callee the only
613 // way we can restore these registers is to have the oldest interpreter
614 // frame that we create restore these values. That is what this routine
615 // will accomplish.
616
617 // At the moment we have modified c2 to not have any callee save registers
618 // so this problem does not exist and this routine is just a place holder.
619
620 assert(f->is_interpreted_frame(), "must be interpreted");
621 }
622
623 // Return BasicType of value being returned
624 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
625
626 // We are already active int he special DeoptResourceMark any ResourceObj's we
627 // allocate will be freed at the end of the routine.
628
629 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
630 // but makes the entry a little slower. There is however a little dance we have to
631 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
632 ResetNoHandleMark rnhm; // No-op in release/product versions
633 HandleMark hm;
634
635 frame stub_frame = thread->last_frame();
636
637 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
638 // must point to the vframeArray for the unpack frame.
639 vframeArray* array = thread->vframe_array_head();
640
641 #ifndef PRODUCT
642 if (TraceDeoptimization) {
643 ttyLocker ttyl;
644 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d",
645 p2i(thread), p2i(array), exec_mode);
646 }
647 #endif
648 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
649 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode);
650
651 UnrollBlock* info = array->unroll_block();
652
653 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
654 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
655
656 BasicType bt = info->return_type();
657
658 // If we have an exception pending, claim that the return type is an oop
659 // so the deopt_blob does not overwrite the exception_oop.
660
661 if (exec_mode == Unpack_exception)
662 bt = T_OBJECT;
663
664 // Cleanup thread deopt data
665 cleanup_deopt_info(thread, array);
666
667 #ifndef PRODUCT
668 if (VerifyStack) {
669 ResourceMark res_mark;
670
671 thread->validate_frame_layout();
672
673 // Verify that the just-unpacked frames match the interpreter's
674 // notions of expression stack and locals
675 vframeArray* cur_array = thread->vframe_array_last();
676 RegisterMap rm(thread, false);
677 rm.set_include_argument_oops(false);
678 bool is_top_frame = true;
679 int callee_size_of_parameters = 0;
680 int callee_max_locals = 0;
681 for (int i = 0; i < cur_array->frames(); i++) {
682 vframeArrayElement* el = cur_array->element(i);
683 frame* iframe = el->iframe();
684 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
685
686 // Get the oop map for this bci
687 InterpreterOopMap mask;
688 int cur_invoke_parameter_size = 0;
689 bool try_next_mask = false;
690 int next_mask_expression_stack_size = -1;
691 int top_frame_expression_stack_adjustment = 0;
692 methodHandle mh(thread, iframe->interpreter_frame_method());
693 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
694 BytecodeStream str(mh);
695 str.set_start(iframe->interpreter_frame_bci());
696 int max_bci = mh->code_size();
697 // Get to the next bytecode if possible
698 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
699 // Check to see if we can grab the number of outgoing arguments
700 // at an uncommon trap for an invoke (where the compiler
701 // generates debug info before the invoke has executed)
702 Bytecodes::Code cur_code = str.next();
703 if (cur_code == Bytecodes::_invokevirtual ||
704 cur_code == Bytecodes::_invokespecial ||
705 cur_code == Bytecodes::_invokestatic ||
706 cur_code == Bytecodes::_invokeinterface ||
707 cur_code == Bytecodes::_invokedynamic) {
708 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
709 Symbol* signature = invoke.signature();
710 ArgumentSizeComputer asc(signature);
711 cur_invoke_parameter_size = asc.size();
712 if (invoke.has_receiver()) {
713 // Add in receiver
714 ++cur_invoke_parameter_size;
715 }
716 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
717 callee_size_of_parameters++;
718 }
719 }
720 if (str.bci() < max_bci) {
721 Bytecodes::Code bc = str.next();
722 if (bc >= 0) {
723 // The interpreter oop map generator reports results before
724 // the current bytecode has executed except in the case of
725 // calls. It seems to be hard to tell whether the compiler
726 // has emitted debug information matching the "state before"
727 // a given bytecode or the state after, so we try both
728 switch (cur_code) {
729 case Bytecodes::_invokevirtual:
730 case Bytecodes::_invokespecial:
731 case Bytecodes::_invokestatic:
732 case Bytecodes::_invokeinterface:
733 case Bytecodes::_invokedynamic:
734 case Bytecodes::_athrow:
735 break;
736 default: {
737 InterpreterOopMap next_mask;
738 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
739 next_mask_expression_stack_size = next_mask.expression_stack_size();
740 // Need to subtract off the size of the result type of
741 // the bytecode because this is not described in the
742 // debug info but returned to the interpreter in the TOS
743 // caching register
744 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
745 if (bytecode_result_type != T_ILLEGAL) {
746 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
747 }
748 assert(top_frame_expression_stack_adjustment >= 0, "");
749 try_next_mask = true;
750 break;
751 }
752 }
753 }
754 }
755
756 // Verify stack depth and oops in frame
757 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
758 if (!(
759 /* SPARC */
760 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
761 /* x86 */
762 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
763 (try_next_mask &&
764 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
765 top_frame_expression_stack_adjustment))) ||
766 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
767 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
768 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
769 )) {
770 ttyLocker ttyl;
771
772 // Print out some information that will help us debug the problem
773 tty->print_cr("Wrong number of expression stack elements during deoptimization");
774 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
775 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
776 iframe->interpreter_frame_expression_stack_size());
777 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
778 tty->print_cr(" try_next_mask = %d", try_next_mask);
779 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
780 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
781 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
782 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
783 tty->print_cr(" exec_mode = %d", exec_mode);
784 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
785 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
786 tty->print_cr(" Interpreted frames:");
787 for (int k = 0; k < cur_array->frames(); k++) {
788 vframeArrayElement* el = cur_array->element(k);
789 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
790 }
791 cur_array->print_on_2(tty);
792 guarantee(false, "wrong number of expression stack elements during deopt");
793 }
794 VerifyOopClosure verify;
795 iframe->oops_interpreted_do(&verify, &rm, false);
796 callee_size_of_parameters = mh->size_of_parameters();
797 callee_max_locals = mh->max_locals();
798 is_top_frame = false;
799 }
800 }
801 #endif /* !PRODUCT */
802
803
804 return bt;
805 JRT_END
806
807
808 int Deoptimization::deoptimize_dependents() {
809 Threads::deoptimized_wrt_marked_nmethods();
810 return 0;
811 }
812
813 Deoptimization::DeoptAction Deoptimization::_unloaded_action
814 = Deoptimization::Action_reinterpret;
815
816 #if COMPILER2_OR_JVMCI
817 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
818 Handle pending_exception(THREAD, thread->pending_exception());
819 const char* exception_file = thread->exception_file();
820 int exception_line = thread->exception_line();
821 thread->clear_pending_exception();
822
823 bool failures = false;
824
825 for (int i = 0; i < objects->length(); i++) {
826 assert(objects->at(i)->is_object(), "invalid debug information");
827 ObjectValue* sv = (ObjectValue*) objects->at(i);
828
829 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
830 oop obj = NULL;
831
832 if (k->is_instance_klass()) {
833 InstanceKlass* ik = InstanceKlass::cast(k);
834 obj = ik->allocate_instance(THREAD);
835 } else if (k->is_valueArray_klass()) {
836 ValueArrayKlass* ak = ValueArrayKlass::cast(k);
837 // Value type array must be zeroed because not all memory is reassigned
838 obj = ak->allocate(sv->field_size(), true, THREAD);
839 } else if (k->is_typeArray_klass()) {
840 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
841 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
842 int len = sv->field_size() / type2size[ak->element_type()];
843 obj = ak->allocate(len, THREAD);
844 } else if (k->is_objArray_klass()) {
845 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
846 obj = ak->allocate(sv->field_size(), THREAD);
847 }
848
849 if (obj == NULL) {
850 failures = true;
851 }
852
853 assert(sv->value().is_null(), "redundant reallocation");
854 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
855 CLEAR_PENDING_EXCEPTION;
856 sv->set_value(obj);
857 }
858
859 if (failures) {
860 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
861 } else if (pending_exception.not_null()) {
862 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
863 }
864
865 return failures;
866 }
867
868 // We're deoptimizing at the return of a call, value type fields are
869 // in registers. When we go back to the interpreter, it will expect a
870 // reference to a value type instance. Allocate and initialize it from
871 // the register values here.
872 bool Deoptimization::realloc_value_type_result(ValueKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) {
873 oop new_vt = vk->realloc_result(map, return_oops, false, THREAD);
874 if (new_vt == NULL) {
875 CLEAR_PENDING_EXCEPTION;
876 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true);
877 }
878 return_oops.clear();
879 return_oops.push(Handle(THREAD, new_vt));
880 return false;
881 }
882
883 // restore elements of an eliminated type array
884 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
885 int index = 0;
886 intptr_t val;
887
888 for (int i = 0; i < sv->field_size(); i++) {
889 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
890 switch(type) {
891 case T_LONG: case T_DOUBLE: {
892 assert(value->type() == T_INT, "Agreement.");
893 StackValue* low =
894 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
895 #ifdef _LP64
896 jlong res = (jlong)low->get_int();
897 #else
898 #ifdef SPARC
899 // For SPARC we have to swap high and low words.
900 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
901 #else
902 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
903 #endif //SPARC
904 #endif
905 obj->long_at_put(index, res);
906 break;
907 }
908
909 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
910 case T_INT: case T_FLOAT: { // 4 bytes.
911 assert(value->type() == T_INT, "Agreement.");
912 bool big_value = false;
913 if (i + 1 < sv->field_size() && type == T_INT) {
914 if (sv->field_at(i)->is_location()) {
915 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
916 if (type == Location::dbl || type == Location::lng) {
917 big_value = true;
918 }
919 } else if (sv->field_at(i)->is_constant_int()) {
920 ScopeValue* next_scope_field = sv->field_at(i + 1);
921 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
922 big_value = true;
923 }
924 }
925 }
926
927 if (big_value) {
928 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
929 #ifdef _LP64
930 jlong res = (jlong)low->get_int();
931 #else
932 #ifdef SPARC
933 // For SPARC we have to swap high and low words.
934 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
935 #else
936 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
937 #endif //SPARC
938 #endif
939 obj->int_at_put(index, (jint)*((jint*)&res));
940 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
941 } else {
942 val = value->get_int();
943 obj->int_at_put(index, (jint)*((jint*)&val));
944 }
945 break;
946 }
947
948 case T_SHORT:
949 assert(value->type() == T_INT, "Agreement.");
950 val = value->get_int();
951 obj->short_at_put(index, (jshort)*((jint*)&val));
952 break;
953
954 case T_CHAR:
955 assert(value->type() == T_INT, "Agreement.");
956 val = value->get_int();
957 obj->char_at_put(index, (jchar)*((jint*)&val));
958 break;
959
960 case T_BYTE:
961 assert(value->type() == T_INT, "Agreement.");
962 val = value->get_int();
963 obj->byte_at_put(index, (jbyte)*((jint*)&val));
964 break;
965
966 case T_BOOLEAN:
967 assert(value->type() == T_INT, "Agreement.");
968 val = value->get_int();
969 obj->bool_at_put(index, (jboolean)*((jint*)&val));
970 break;
971
972 default:
973 ShouldNotReachHere();
974 }
975 index++;
976 }
977 }
978
979
980 // restore fields of an eliminated object array
981 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
982 for (int i = 0; i < sv->field_size(); i++) {
983 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
984 assert(value->type() == T_OBJECT, "object element expected");
985 obj->obj_at_put(i, value->get_obj()());
986 }
987 }
988
989 class ReassignedField {
990 public:
991 int _offset;
992 BasicType _type;
993 InstanceKlass* _klass;
994 public:
995 ReassignedField() {
996 _offset = 0;
997 _type = T_ILLEGAL;
998 _klass = NULL;
999 }
1000 };
1001
1002 int compare(ReassignedField* left, ReassignedField* right) {
1003 return left->_offset - right->_offset;
1004 }
1005
1006 // Restore fields of an eliminated instance object using the same field order
1007 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1008 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal, int base_offset, TRAPS) {
1009 if (klass->superklass() != NULL) {
1010 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal, 0, CHECK_0);
1011 }
1012
1013 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1014 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
1015 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
1016 ReassignedField field;
1017 field._offset = fs.offset();
1018 field._type = FieldType::basic_type(fs.signature());
1019 if (field._type == T_VALUETYPE) {
1020 if (fs.is_flattened()) {
1021 // Resolve klass of flattened value type field
1022 SignatureStream ss(fs.signature(), false);
1023 Klass* vk = ss.as_klass(Handle(THREAD, klass->class_loader()), Handle(THREAD, klass->protection_domain()), SignatureStream::NCDFError, THREAD);
1024 guarantee(!HAS_PENDING_EXCEPTION, "Should not have any exceptions pending");
1025 assert(vk->is_value(), "must be a ValueKlass");
1026 field._klass = InstanceKlass::cast(vk);
1027 } else {
1028 // Non-flattened value type field
1029 field._type = T_VALUETYPEPTR;
1030 }
1031 }
1032 fields->append(field);
1033 }
1034 }
1035 fields->sort(compare);
1036 for (int i = 0; i < fields->length(); i++) {
1037 intptr_t val;
1038 ScopeValue* scope_field = sv->field_at(svIndex);
1039 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1040 int offset = base_offset + fields->at(i)._offset;
1041 BasicType type = fields->at(i)._type;
1042 switch (type) {
1043 case T_OBJECT:
1044 case T_VALUETYPEPTR:
1045 case T_ARRAY:
1046 assert(value->type() == T_OBJECT, "Agreement.");
1047 obj->obj_field_put(offset, value->get_obj()());
1048 break;
1049
1050 case T_VALUETYPE: {
1051 // Recursively re-assign flattened value type fields
1052 InstanceKlass* vk = fields->at(i)._klass;
1053 assert(vk != NULL, "must be resolved");
1054 offset -= ValueKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header
1055 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0);
1056 continue; // Continue because we don't need to increment svIndex
1057 }
1058
1059 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1060 case T_INT: case T_FLOAT: { // 4 bytes.
1061 assert(value->type() == T_INT, "Agreement.");
1062 bool big_value = false;
1063 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1064 if (scope_field->is_location()) {
1065 Location::Type type = ((LocationValue*) scope_field)->location().type();
1066 if (type == Location::dbl || type == Location::lng) {
1067 big_value = true;
1068 }
1069 }
1070 if (scope_field->is_constant_int()) {
1071 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1072 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1073 big_value = true;
1074 }
1075 }
1076 }
1077
1078 if (big_value) {
1079 i++;
1080 assert(i < fields->length(), "second T_INT field needed");
1081 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1082 } else {
1083 val = value->get_int();
1084 obj->int_field_put(offset, (jint)*((jint*)&val));
1085 break;
1086 }
1087 }
1088 /* no break */
1089
1090 case T_LONG: case T_DOUBLE: {
1091 assert(value->type() == T_INT, "Agreement.");
1092 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1093 #ifdef _LP64
1094 jlong res = (jlong)low->get_int();
1095 #else
1096 #ifdef SPARC
1097 // For SPARC we have to swap high and low words.
1098 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1099 #else
1100 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1101 #endif //SPARC
1102 #endif
1103 obj->long_field_put(offset, res);
1104 break;
1105 }
1106
1107 case T_SHORT:
1108 assert(value->type() == T_INT, "Agreement.");
1109 val = value->get_int();
1110 obj->short_field_put(offset, (jshort)*((jint*)&val));
1111 break;
1112
1113 case T_CHAR:
1114 assert(value->type() == T_INT, "Agreement.");
1115 val = value->get_int();
1116 obj->char_field_put(offset, (jchar)*((jint*)&val));
1117 break;
1118
1119 case T_BYTE:
1120 assert(value->type() == T_INT, "Agreement.");
1121 val = value->get_int();
1122 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1123 break;
1124
1125 case T_BOOLEAN:
1126 assert(value->type() == T_INT, "Agreement.");
1127 val = value->get_int();
1128 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1129 break;
1130
1131 default:
1132 ShouldNotReachHere();
1133 }
1134 svIndex++;
1135 }
1136 return svIndex;
1137 }
1138
1139 // restore fields of an eliminated value type array
1140 void Deoptimization::reassign_value_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, valueArrayOop obj, ValueArrayKlass* vak, TRAPS) {
1141 ValueKlass* vk = vak->element_klass();
1142 assert(vk->flatten_array(), "should only be used for flattened value type arrays");
1143 // Adjust offset to omit oop header
1144 int base_offset = arrayOopDesc::base_offset_in_bytes(T_VALUETYPE) - ValueKlass::cast(vk)->first_field_offset();
1145 // Initialize all elements of the flattened value type array
1146 for (int i = 0; i < sv->field_size(); i++) {
1147 ScopeValue* val = sv->field_at(i);
1148 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper()));
1149 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, false /* skip_internal */, offset, CHECK);
1150 }
1151 }
1152
1153 // restore fields of all eliminated objects and arrays
1154 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) {
1155 for (int i = 0; i < objects->length(); i++) {
1156 ObjectValue* sv = (ObjectValue*) objects->at(i);
1157 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1158 Handle obj = sv->value();
1159 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1160 if (PrintDeoptimizationDetails) {
1161 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1162 }
1163 if (obj.is_null()) {
1164 continue;
1165 }
1166
1167 if (k->is_instance_klass()) {
1168 InstanceKlass* ik = InstanceKlass::cast(k);
1169 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK);
1170 } else if (k->is_valueArray_klass()) {
1171 ValueArrayKlass* vak = ValueArrayKlass::cast(k);
1172 reassign_value_array_elements(fr, reg_map, sv, (valueArrayOop) obj(), vak, CHECK);
1173 } else if (k->is_typeArray_klass()) {
1174 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1175 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1176 } else if (k->is_objArray_klass()) {
1177 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1178 }
1179 }
1180 }
1181
1182
1183 // relock objects for which synchronization was eliminated
1184 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
1185 for (int i = 0; i < monitors->length(); i++) {
1186 MonitorInfo* mon_info = monitors->at(i);
1187 if (mon_info->eliminated()) {
1188 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1189 if (!mon_info->owner_is_scalar_replaced()) {
1190 Handle obj(thread, mon_info->owner());
1191 markOop mark = obj->mark();
1192 if (UseBiasedLocking && mark->has_bias_pattern()) {
1193 // New allocated objects may have the mark set to anonymously biased.
1194 // Also the deoptimized method may called methods with synchronization
1195 // where the thread-local object is bias locked to the current thread.
1196 assert(mark->is_biased_anonymously() ||
1197 mark->biased_locker() == thread, "should be locked to current thread");
1198 // Reset mark word to unbiased prototype.
1199 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
1200 obj->set_mark(unbiased_prototype);
1201 }
1202 BasicLock* lock = mon_info->lock();
1203 ObjectSynchronizer::slow_enter(obj, lock, thread);
1204 assert(mon_info->owner()->is_locked(), "object must be locked now");
1205 }
1206 }
1207 }
1208 }
1209
1210
1211 #ifndef PRODUCT
1212 // print information about reallocated objects
1213 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1214 fieldDescriptor fd;
1215
1216 for (int i = 0; i < objects->length(); i++) {
1217 ObjectValue* sv = (ObjectValue*) objects->at(i);
1218 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1219 Handle obj = sv->value();
1220
1221 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1222 k->print_value();
1223 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1224 if (obj.is_null()) {
1225 tty->print(" allocation failed");
1226 } else {
1227 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1228 }
1229 tty->cr();
1230
1231 if (Verbose && !obj.is_null()) {
1232 k->oop_print_on(obj(), tty);
1233 }
1234 }
1235 }
1236 #endif
1237 #endif // COMPILER2_OR_JVMCI
1238
1239 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1240 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1241
1242 #ifndef PRODUCT
1243 if (PrintDeoptimizationDetails) {
1244 ttyLocker ttyl;
1245 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1246 fr.print_on(tty);
1247 tty->print_cr(" Virtual frames (innermost first):");
1248 for (int index = 0; index < chunk->length(); index++) {
1249 compiledVFrame* vf = chunk->at(index);
1250 tty->print(" %2d - ", index);
1251 vf->print_value();
1252 int bci = chunk->at(index)->raw_bci();
1253 const char* code_name;
1254 if (bci == SynchronizationEntryBCI) {
1255 code_name = "sync entry";
1256 } else {
1257 Bytecodes::Code code = vf->method()->code_at(bci);
1258 code_name = Bytecodes::name(code);
1259 }
1260 tty->print(" - %s", code_name);
1261 tty->print_cr(" @ bci %d ", bci);
1262 if (Verbose) {
1263 vf->print();
1264 tty->cr();
1265 }
1266 }
1267 }
1268 #endif
1269
1270 // Register map for next frame (used for stack crawl). We capture
1271 // the state of the deopt'ing frame's caller. Thus if we need to
1272 // stuff a C2I adapter we can properly fill in the callee-save
1273 // register locations.
1274 frame caller = fr.sender(reg_map);
1275 int frame_size = caller.sp() - fr.sp();
1276
1277 frame sender = caller;
1278
1279 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1280 // the vframeArray containing the unpacking information is allocated in the C heap.
1281 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1282 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1283
1284 // Compare the vframeArray to the collected vframes
1285 assert(array->structural_compare(thread, chunk), "just checking");
1286
1287 #ifndef PRODUCT
1288 if (PrintDeoptimizationDetails) {
1289 ttyLocker ttyl;
1290 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1291 }
1292 #endif // PRODUCT
1293
1294 return array;
1295 }
1296
1297 #if COMPILER2_OR_JVMCI
1298 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1299 // Reallocation of some scalar replaced objects failed. Record
1300 // that we need to pop all the interpreter frames for the
1301 // deoptimized compiled frame.
1302 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1303 thread->set_frames_to_pop_failed_realloc(array->frames());
1304 // Unlock all monitors here otherwise the interpreter will see a
1305 // mix of locked and unlocked monitors (because of failed
1306 // reallocations of synchronized objects) and be confused.
1307 for (int i = 0; i < array->frames(); i++) {
1308 MonitorChunk* monitors = array->element(i)->monitors();
1309 if (monitors != NULL) {
1310 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1311 BasicObjectLock* src = monitors->at(j);
1312 if (src->obj() != NULL) {
1313 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread);
1314 }
1315 }
1316 array->element(i)->free_monitors(thread);
1317 #ifdef ASSERT
1318 array->element(i)->set_removed_monitors();
1319 #endif
1320 }
1321 }
1322 }
1323 #endif
1324
1325 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1326 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1327 Thread* thread = Thread::current();
1328 for (int i = 0; i < monitors->length(); i++) {
1329 MonitorInfo* mon_info = monitors->at(i);
1330 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1331 objects_to_revoke->append(Handle(thread, mon_info->owner()));
1332 }
1333 }
1334 }
1335
1336
1337 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1338 if (!UseBiasedLocking) {
1339 return;
1340 }
1341
1342 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1343
1344 // Unfortunately we don't have a RegisterMap available in most of
1345 // the places we want to call this routine so we need to walk the
1346 // stack again to update the register map.
1347 if (map == NULL || !map->update_map()) {
1348 StackFrameStream sfs(thread, true);
1349 bool found = false;
1350 while (!found && !sfs.is_done()) {
1351 frame* cur = sfs.current();
1352 sfs.next();
1353 found = cur->id() == fr.id();
1354 }
1355 assert(found, "frame to be deoptimized not found on target thread's stack");
1356 map = sfs.register_map();
1357 }
1358
1359 vframe* vf = vframe::new_vframe(&fr, map, thread);
1360 compiledVFrame* cvf = compiledVFrame::cast(vf);
1361 // Revoke monitors' biases in all scopes
1362 while (!cvf->is_top()) {
1363 collect_monitors(cvf, objects_to_revoke);
1364 cvf = compiledVFrame::cast(cvf->sender());
1365 }
1366 collect_monitors(cvf, objects_to_revoke);
1367
1368 if (SafepointSynchronize::is_at_safepoint()) {
1369 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1370 } else {
1371 BiasedLocking::revoke(objects_to_revoke);
1372 }
1373 }
1374
1375
1376 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1377 if (!UseBiasedLocking) {
1378 return;
1379 }
1380
1381 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1382 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1383 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
1384 if (jt->has_last_Java_frame()) {
1385 StackFrameStream sfs(jt, true);
1386 while (!sfs.is_done()) {
1387 frame* cur = sfs.current();
1388 if (cb->contains(cur->pc())) {
1389 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1390 compiledVFrame* cvf = compiledVFrame::cast(vf);
1391 // Revoke monitors' biases in all scopes
1392 while (!cvf->is_top()) {
1393 collect_monitors(cvf, objects_to_revoke);
1394 cvf = compiledVFrame::cast(cvf->sender());
1395 }
1396 collect_monitors(cvf, objects_to_revoke);
1397 }
1398 sfs.next();
1399 }
1400 }
1401 }
1402 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1403 }
1404
1405
1406 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1407 assert(fr.can_be_deoptimized(), "checking frame type");
1408
1409 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1410
1411 if (LogCompilation && xtty != NULL) {
1412 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1413 assert(cm != NULL, "only compiled methods can deopt");
1414
1415 ttyLocker ttyl;
1416 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1417 cm->log_identity(xtty);
1418 xtty->end_head();
1419 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1420 xtty->begin_elem("jvms bci='%d'", sd->bci());
1421 xtty->method(sd->method());
1422 xtty->end_elem();
1423 if (sd->is_top()) break;
1424 }
1425 xtty->tail("deoptimized");
1426 }
1427
1428 // Patch the compiled method so that when execution returns to it we will
1429 // deopt the execution state and return to the interpreter.
1430 fr.deoptimize(thread);
1431 }
1432
1433 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1434 deoptimize(thread, fr, map, Reason_constraint);
1435 }
1436
1437 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) {
1438 // Deoptimize only if the frame comes from compile code.
1439 // Do not deoptimize the frame which is already patched
1440 // during the execution of the loops below.
1441 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1442 return;
1443 }
1444 ResourceMark rm;
1445 DeoptimizationMarker dm;
1446 if (UseBiasedLocking) {
1447 revoke_biases_of_monitors(thread, fr, map);
1448 }
1449 deoptimize_single_frame(thread, fr, reason);
1450
1451 }
1452
1453 #if INCLUDE_JVMCI
1454 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) {
1455 // there is no exception handler for this pc => deoptimize
1456 cm->make_not_entrant();
1457
1458 // Use Deoptimization::deoptimize for all of its side-effects:
1459 // revoking biases of monitors, gathering traps statistics, logging...
1460 // it also patches the return pc but we do not care about that
1461 // since we return a continuation to the deopt_blob below.
1462 JavaThread* thread = JavaThread::current();
1463 RegisterMap reg_map(thread, UseBiasedLocking);
1464 frame runtime_frame = thread->last_frame();
1465 frame caller_frame = runtime_frame.sender(®_map);
1466 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method");
1467 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler);
1468
1469 MethodData* trap_mdo = get_method_data(thread, cm->method(), true);
1470 if (trap_mdo != NULL) {
1471 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1472 }
1473
1474 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1475 }
1476 #endif
1477
1478 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1479 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1480 "can only deoptimize other thread at a safepoint");
1481 // Compute frame and register map based on thread and sp.
1482 RegisterMap reg_map(thread, UseBiasedLocking);
1483 frame fr = thread->last_frame();
1484 while (fr.id() != id) {
1485 fr = fr.sender(®_map);
1486 }
1487 deoptimize(thread, fr, ®_map, reason);
1488 }
1489
1490
1491 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1492 if (thread == Thread::current()) {
1493 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1494 } else {
1495 VM_DeoptimizeFrame deopt(thread, id, reason);
1496 VMThread::execute(&deopt);
1497 }
1498 }
1499
1500 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1501 deoptimize_frame(thread, id, Reason_constraint);
1502 }
1503
1504 // JVMTI PopFrame support
1505 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1506 {
1507 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1508 }
1509 JRT_END
1510
1511 MethodData*
1512 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1513 bool create_if_missing) {
1514 Thread* THREAD = thread;
1515 MethodData* mdo = m()->method_data();
1516 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1517 // Build an MDO. Ignore errors like OutOfMemory;
1518 // that simply means we won't have an MDO to update.
1519 Method::build_interpreter_method_data(m, THREAD);
1520 if (HAS_PENDING_EXCEPTION) {
1521 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1522 CLEAR_PENDING_EXCEPTION;
1523 }
1524 mdo = m()->method_data();
1525 }
1526 return mdo;
1527 }
1528
1529 #if COMPILER2_OR_JVMCI
1530 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1531 // in case of an unresolved klass entry, load the class.
1532 if (constant_pool->tag_at(index).is_unresolved_klass() ||
1533 constant_pool->tag_at(index).is_unresolved_value_type()) {
1534 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1535 return;
1536 }
1537
1538 if (!constant_pool->tag_at(index).is_symbol()) return;
1539
1540 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1541 Symbol* symbol = constant_pool->symbol_at(index);
1542
1543 // class name?
1544 if (symbol->byte_at(0) != '(') {
1545 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1546 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1547 return;
1548 }
1549
1550 // then it must be a signature!
1551 ResourceMark rm(THREAD);
1552 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1553 if (ss.is_object()) {
1554 Symbol* class_name = ss.as_symbol(CHECK);
1555 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1556 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1557 }
1558 }
1559 }
1560
1561
1562 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1563 EXCEPTION_MARK;
1564 load_class_by_index(constant_pool, index, THREAD);
1565 if (HAS_PENDING_EXCEPTION) {
1566 // Exception happened during classloading. We ignore the exception here, since it
1567 // is going to be rethrown since the current activation is going to be deoptimized and
1568 // the interpreter will re-execute the bytecode.
1569 CLEAR_PENDING_EXCEPTION;
1570 // Class loading called java code which may have caused a stack
1571 // overflow. If the exception was thrown right before the return
1572 // to the runtime the stack is no longer guarded. Reguard the
1573 // stack otherwise if we return to the uncommon trap blob and the
1574 // stack bang causes a stack overflow we crash.
1575 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1576 JavaThread* thread = (JavaThread*)THREAD;
1577 bool guard_pages_enabled = thread->stack_guards_enabled();
1578 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1579 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1580 }
1581 }
1582
1583 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1584 HandleMark hm;
1585
1586 // uncommon_trap() is called at the beginning of the uncommon trap
1587 // handler. Note this fact before we start generating temporary frames
1588 // that can confuse an asynchronous stack walker. This counter is
1589 // decremented at the end of unpack_frames().
1590 thread->inc_in_deopt_handler();
1591
1592 // We need to update the map if we have biased locking.
1593 #if INCLUDE_JVMCI
1594 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1595 RegisterMap reg_map(thread, true);
1596 #else
1597 RegisterMap reg_map(thread, UseBiasedLocking);
1598 #endif
1599 frame stub_frame = thread->last_frame();
1600 frame fr = stub_frame.sender(®_map);
1601 // Make sure the calling nmethod is not getting deoptimized and removed
1602 // before we are done with it.
1603 nmethodLocker nl(fr.pc());
1604
1605 // Log a message
1606 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1607 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1608
1609 {
1610 ResourceMark rm;
1611
1612 // Revoke biases of any monitors in the frame to ensure we can migrate them
1613 revoke_biases_of_monitors(thread, fr, ®_map);
1614
1615 DeoptReason reason = trap_request_reason(trap_request);
1616 DeoptAction action = trap_request_action(trap_request);
1617 #if INCLUDE_JVMCI
1618 int debug_id = trap_request_debug_id(trap_request);
1619 #endif
1620 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1621
1622 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1623 compiledVFrame* cvf = compiledVFrame::cast(vf);
1624
1625 CompiledMethod* nm = cvf->code();
1626
1627 ScopeDesc* trap_scope = cvf->scope();
1628
1629 if (TraceDeoptimization) {
1630 ttyLocker ttyl;
1631 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()
1632 #if INCLUDE_JVMCI
1633 , debug_id
1634 #endif
1635 );
1636 }
1637
1638 methodHandle trap_method = trap_scope->method();
1639 int trap_bci = trap_scope->bci();
1640 #if INCLUDE_JVMCI
1641 oop speculation = thread->pending_failed_speculation();
1642 if (nm->is_compiled_by_jvmci()) {
1643 if (speculation != NULL) {
1644 oop speculation_log = nm->as_nmethod()->speculation_log();
1645 if (speculation_log != NULL) {
1646 if (TraceDeoptimization || TraceUncollectedSpeculations) {
1647 if (HotSpotSpeculationLog::lastFailed(speculation_log) != NULL) {
1648 tty->print_cr("A speculation that was not collected by the compiler is being overwritten");
1649 }
1650 }
1651 if (TraceDeoptimization) {
1652 tty->print_cr("Saving speculation to speculation log");
1653 }
1654 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation);
1655 } else {
1656 if (TraceDeoptimization) {
1657 tty->print_cr("Speculation present but no speculation log");
1658 }
1659 }
1660 thread->set_pending_failed_speculation(NULL);
1661 } else {
1662 if (TraceDeoptimization) {
1663 tty->print_cr("No speculation");
1664 }
1665 }
1666 } else {
1667 assert(speculation == NULL, "There should not be a speculation for method compiled by non-JVMCI compilers");
1668 }
1669
1670 if (trap_bci == SynchronizationEntryBCI) {
1671 trap_bci = 0;
1672 thread->set_pending_monitorenter(true);
1673 }
1674
1675 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1676 thread->set_pending_transfer_to_interpreter(true);
1677 }
1678 #endif
1679
1680 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1681 // Record this event in the histogram.
1682 gather_statistics(reason, action, trap_bc);
1683
1684 // Ensure that we can record deopt. history:
1685 // Need MDO to record RTM code generation state.
1686 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1687
1688 methodHandle profiled_method;
1689 #if INCLUDE_JVMCI
1690 if (nm->is_compiled_by_jvmci()) {
1691 profiled_method = nm->method();
1692 } else {
1693 profiled_method = trap_method;
1694 }
1695 #else
1696 profiled_method = trap_method;
1697 #endif
1698
1699 MethodData* trap_mdo =
1700 get_method_data(thread, profiled_method, create_if_missing);
1701
1702 // Log a message
1703 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
1704 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1705 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
1706
1707 // Print a bunch of diagnostics, if requested.
1708 if (TraceDeoptimization || LogCompilation) {
1709 ResourceMark rm;
1710 ttyLocker ttyl;
1711 char buf[100];
1712 if (xtty != NULL) {
1713 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1714 os::current_thread_id(),
1715 format_trap_request(buf, sizeof(buf), trap_request));
1716 nm->log_identity(xtty);
1717 }
1718 Symbol* class_name = NULL;
1719 bool unresolved = false;
1720 if (unloaded_class_index >= 0) {
1721 constantPoolHandle constants (THREAD, trap_method->constants());
1722 if (constants->tag_at(unloaded_class_index).is_unresolved_klass() ||
1723 constants->tag_at(unloaded_class_index).is_unresolved_value_type()) {
1724 class_name = constants->klass_name_at(unloaded_class_index);
1725 unresolved = true;
1726 if (xtty != NULL)
1727 xtty->print(" unresolved='1'");
1728 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1729 class_name = constants->symbol_at(unloaded_class_index);
1730 }
1731 if (xtty != NULL)
1732 xtty->name(class_name);
1733 }
1734 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
1735 // Dump the relevant MDO state.
1736 // This is the deopt count for the current reason, any previous
1737 // reasons or recompiles seen at this point.
1738 int dcnt = trap_mdo->trap_count(reason);
1739 if (dcnt != 0)
1740 xtty->print(" count='%d'", dcnt);
1741 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1742 int dos = (pdata == NULL)? 0: pdata->trap_state();
1743 if (dos != 0) {
1744 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1745 if (trap_state_is_recompiled(dos)) {
1746 int recnt2 = trap_mdo->overflow_recompile_count();
1747 if (recnt2 != 0)
1748 xtty->print(" recompiles2='%d'", recnt2);
1749 }
1750 }
1751 }
1752 if (xtty != NULL) {
1753 xtty->stamp();
1754 xtty->end_head();
1755 }
1756 if (TraceDeoptimization) { // make noise on the tty
1757 tty->print("Uncommon trap occurred in");
1758 nm->method()->print_short_name(tty);
1759 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
1760 #if INCLUDE_JVMCI
1761 if (nm->is_nmethod()) {
1762 char* installed_code_name = nm->as_nmethod()->jvmci_installed_code_name(buf, sizeof(buf));
1763 if (installed_code_name != NULL) {
1764 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name);
1765 }
1766 }
1767 #endif
1768 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
1769 p2i(fr.pc()),
1770 os::current_thread_id(),
1771 trap_reason_name(reason),
1772 trap_action_name(action),
1773 unloaded_class_index
1774 #if INCLUDE_JVMCI
1775 , debug_id
1776 #endif
1777 );
1778 if (class_name != NULL) {
1779 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1780 class_name->print_symbol_on(tty);
1781 }
1782 tty->cr();
1783 }
1784 if (xtty != NULL) {
1785 // Log the precise location of the trap.
1786 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1787 xtty->begin_elem("jvms bci='%d'", sd->bci());
1788 xtty->method(sd->method());
1789 xtty->end_elem();
1790 if (sd->is_top()) break;
1791 }
1792 xtty->tail("uncommon_trap");
1793 }
1794 }
1795 // (End diagnostic printout.)
1796
1797 // Load class if necessary
1798 if (unloaded_class_index >= 0) {
1799 constantPoolHandle constants(THREAD, trap_method->constants());
1800 load_class_by_index(constants, unloaded_class_index);
1801 }
1802
1803 // Flush the nmethod if necessary and desirable.
1804 //
1805 // We need to avoid situations where we are re-flushing the nmethod
1806 // because of a hot deoptimization site. Repeated flushes at the same
1807 // point need to be detected by the compiler and avoided. If the compiler
1808 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1809 // module must take measures to avoid an infinite cycle of recompilation
1810 // and deoptimization. There are several such measures:
1811 //
1812 // 1. If a recompilation is ordered a second time at some site X
1813 // and for the same reason R, the action is adjusted to 'reinterpret',
1814 // to give the interpreter time to exercise the method more thoroughly.
1815 // If this happens, the method's overflow_recompile_count is incremented.
1816 //
1817 // 2. If the compiler fails to reduce the deoptimization rate, then
1818 // the method's overflow_recompile_count will begin to exceed the set
1819 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1820 // is adjusted to 'make_not_compilable', and the method is abandoned
1821 // to the interpreter. This is a performance hit for hot methods,
1822 // but is better than a disastrous infinite cycle of recompilations.
1823 // (Actually, only the method containing the site X is abandoned.)
1824 //
1825 // 3. In parallel with the previous measures, if the total number of
1826 // recompilations of a method exceeds the much larger set limit
1827 // PerMethodRecompilationCutoff, the method is abandoned.
1828 // This should only happen if the method is very large and has
1829 // many "lukewarm" deoptimizations. The code which enforces this
1830 // limit is elsewhere (class nmethod, class Method).
1831 //
1832 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1833 // to recompile at each bytecode independently of the per-BCI cutoff.
1834 //
1835 // The decision to update code is up to the compiler, and is encoded
1836 // in the Action_xxx code. If the compiler requests Action_none
1837 // no trap state is changed, no compiled code is changed, and the
1838 // computation suffers along in the interpreter.
1839 //
1840 // The other action codes specify various tactics for decompilation
1841 // and recompilation. Action_maybe_recompile is the loosest, and
1842 // allows the compiled code to stay around until enough traps are seen,
1843 // and until the compiler gets around to recompiling the trapping method.
1844 //
1845 // The other actions cause immediate removal of the present code.
1846
1847 // Traps caused by injected profile shouldn't pollute trap counts.
1848 bool injected_profile_trap = trap_method->has_injected_profile() &&
1849 (reason == Reason_intrinsic || reason == Reason_unreached);
1850
1851 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
1852 bool make_not_entrant = false;
1853 bool make_not_compilable = false;
1854 bool reprofile = false;
1855 switch (action) {
1856 case Action_none:
1857 // Keep the old code.
1858 update_trap_state = false;
1859 break;
1860 case Action_maybe_recompile:
1861 // Do not need to invalidate the present code, but we can
1862 // initiate another
1863 // Start compiler without (necessarily) invalidating the nmethod.
1864 // The system will tolerate the old code, but new code should be
1865 // generated when possible.
1866 break;
1867 case Action_reinterpret:
1868 // Go back into the interpreter for a while, and then consider
1869 // recompiling form scratch.
1870 make_not_entrant = true;
1871 // Reset invocation counter for outer most method.
1872 // This will allow the interpreter to exercise the bytecodes
1873 // for a while before recompiling.
1874 // By contrast, Action_make_not_entrant is immediate.
1875 //
1876 // Note that the compiler will track null_check, null_assert,
1877 // range_check, and class_check events and log them as if they
1878 // had been traps taken from compiled code. This will update
1879 // the MDO trap history so that the next compilation will
1880 // properly detect hot trap sites.
1881 reprofile = true;
1882 break;
1883 case Action_make_not_entrant:
1884 // Request immediate recompilation, and get rid of the old code.
1885 // Make them not entrant, so next time they are called they get
1886 // recompiled. Unloaded classes are loaded now so recompile before next
1887 // time they are called. Same for uninitialized. The interpreter will
1888 // link the missing class, if any.
1889 make_not_entrant = true;
1890 break;
1891 case Action_make_not_compilable:
1892 // Give up on compiling this method at all.
1893 make_not_entrant = true;
1894 make_not_compilable = true;
1895 break;
1896 default:
1897 ShouldNotReachHere();
1898 }
1899
1900 // Setting +ProfileTraps fixes the following, on all platforms:
1901 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1902 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1903 // recompile relies on a MethodData* to record heroic opt failures.
1904
1905 // Whether the interpreter is producing MDO data or not, we also need
1906 // to use the MDO to detect hot deoptimization points and control
1907 // aggressive optimization.
1908 bool inc_recompile_count = false;
1909 ProfileData* pdata = NULL;
1910 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) {
1911 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
1912 uint this_trap_count = 0;
1913 bool maybe_prior_trap = false;
1914 bool maybe_prior_recompile = false;
1915 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
1916 #if INCLUDE_JVMCI
1917 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
1918 #endif
1919 nm->method(),
1920 //outputs:
1921 this_trap_count,
1922 maybe_prior_trap,
1923 maybe_prior_recompile);
1924 // Because the interpreter also counts null, div0, range, and class
1925 // checks, these traps from compiled code are double-counted.
1926 // This is harmless; it just means that the PerXTrapLimit values
1927 // are in effect a little smaller than they look.
1928
1929 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1930 if (per_bc_reason != Reason_none) {
1931 // Now take action based on the partially known per-BCI history.
1932 if (maybe_prior_trap
1933 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1934 // If there are too many traps at this BCI, force a recompile.
1935 // This will allow the compiler to see the limit overflow, and
1936 // take corrective action, if possible. The compiler generally
1937 // does not use the exact PerBytecodeTrapLimit value, but instead
1938 // changes its tactics if it sees any traps at all. This provides
1939 // a little hysteresis, delaying a recompile until a trap happens
1940 // several times.
1941 //
1942 // Actually, since there is only one bit of counter per BCI,
1943 // the possible per-BCI counts are {0,1,(per-method count)}.
1944 // This produces accurate results if in fact there is only
1945 // one hot trap site, but begins to get fuzzy if there are
1946 // many sites. For example, if there are ten sites each
1947 // trapping two or more times, they each get the blame for
1948 // all of their traps.
1949 make_not_entrant = true;
1950 }
1951
1952 // Detect repeated recompilation at the same BCI, and enforce a limit.
1953 if (make_not_entrant && maybe_prior_recompile) {
1954 // More than one recompile at this point.
1955 inc_recompile_count = maybe_prior_trap;
1956 }
1957 } else {
1958 // For reasons which are not recorded per-bytecode, we simply
1959 // force recompiles unconditionally.
1960 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1961 make_not_entrant = true;
1962 }
1963
1964 // Go back to the compiler if there are too many traps in this method.
1965 if (this_trap_count >= per_method_trap_limit(reason)) {
1966 // If there are too many traps in this method, force a recompile.
1967 // This will allow the compiler to see the limit overflow, and
1968 // take corrective action, if possible.
1969 // (This condition is an unlikely backstop only, because the
1970 // PerBytecodeTrapLimit is more likely to take effect first,
1971 // if it is applicable.)
1972 make_not_entrant = true;
1973 }
1974
1975 // Here's more hysteresis: If there has been a recompile at
1976 // this trap point already, run the method in the interpreter
1977 // for a while to exercise it more thoroughly.
1978 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1979 reprofile = true;
1980 }
1981 }
1982
1983 // Take requested actions on the method:
1984
1985 // Recompile
1986 if (make_not_entrant) {
1987 if (!nm->make_not_entrant()) {
1988 return; // the call did not change nmethod's state
1989 }
1990
1991 if (pdata != NULL) {
1992 // Record the recompilation event, if any.
1993 int tstate0 = pdata->trap_state();
1994 int tstate1 = trap_state_set_recompiled(tstate0, true);
1995 if (tstate1 != tstate0)
1996 pdata->set_trap_state(tstate1);
1997 }
1998
1999 #if INCLUDE_RTM_OPT
2000 // Restart collecting RTM locking abort statistic if the method
2001 // is recompiled for a reason other than RTM state change.
2002 // Assume that in new recompiled code the statistic could be different,
2003 // for example, due to different inlining.
2004 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
2005 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
2006 trap_mdo->atomic_set_rtm_state(ProfileRTM);
2007 }
2008 #endif
2009 // For code aging we count traps separately here, using make_not_entrant()
2010 // as a guard against simultaneous deopts in multiple threads.
2011 if (reason == Reason_tenured && trap_mdo != NULL) {
2012 trap_mdo->inc_tenure_traps();
2013 }
2014 }
2015
2016 if (inc_recompile_count) {
2017 trap_mdo->inc_overflow_recompile_count();
2018 if ((uint)trap_mdo->overflow_recompile_count() >
2019 (uint)PerBytecodeRecompilationCutoff) {
2020 // Give up on the method containing the bad BCI.
2021 if (trap_method() == nm->method()) {
2022 make_not_compilable = true;
2023 } else {
2024 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
2025 // But give grace to the enclosing nm->method().
2026 }
2027 }
2028 }
2029
2030 // Reprofile
2031 if (reprofile) {
2032 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
2033 }
2034
2035 // Give up compiling
2036 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2037 assert(make_not_entrant, "consistent");
2038 nm->method()->set_not_compilable(CompLevel_full_optimization);
2039 }
2040
2041 } // Free marked resources
2042
2043 }
2044 JRT_END
2045
2046 ProfileData*
2047 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2048 int trap_bci,
2049 Deoptimization::DeoptReason reason,
2050 bool update_total_trap_count,
2051 #if INCLUDE_JVMCI
2052 bool is_osr,
2053 #endif
2054 Method* compiled_method,
2055 //outputs:
2056 uint& ret_this_trap_count,
2057 bool& ret_maybe_prior_trap,
2058 bool& ret_maybe_prior_recompile) {
2059 bool maybe_prior_trap = false;
2060 bool maybe_prior_recompile = false;
2061 uint this_trap_count = 0;
2062 if (update_total_trap_count) {
2063 uint idx = reason;
2064 #if INCLUDE_JVMCI
2065 if (is_osr) {
2066 idx += Reason_LIMIT;
2067 }
2068 #endif
2069 uint prior_trap_count = trap_mdo->trap_count(idx);
2070 this_trap_count = trap_mdo->inc_trap_count(idx);
2071
2072 // If the runtime cannot find a place to store trap history,
2073 // it is estimated based on the general condition of the method.
2074 // If the method has ever been recompiled, or has ever incurred
2075 // a trap with the present reason , then this BCI is assumed
2076 // (pessimistically) to be the culprit.
2077 maybe_prior_trap = (prior_trap_count != 0);
2078 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2079 }
2080 ProfileData* pdata = NULL;
2081
2082
2083 // For reasons which are recorded per bytecode, we check per-BCI data.
2084 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2085 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2086 if (per_bc_reason != Reason_none) {
2087 // Find the profile data for this BCI. If there isn't one,
2088 // try to allocate one from the MDO's set of spares.
2089 // This will let us detect a repeated trap at this point.
2090 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
2091
2092 if (pdata != NULL) {
2093 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2094 if (LogCompilation && xtty != NULL) {
2095 ttyLocker ttyl;
2096 // no more room for speculative traps in this MDO
2097 xtty->elem("speculative_traps_oom");
2098 }
2099 }
2100 // Query the trap state of this profile datum.
2101 int tstate0 = pdata->trap_state();
2102 if (!trap_state_has_reason(tstate0, per_bc_reason))
2103 maybe_prior_trap = false;
2104 if (!trap_state_is_recompiled(tstate0))
2105 maybe_prior_recompile = false;
2106
2107 // Update the trap state of this profile datum.
2108 int tstate1 = tstate0;
2109 // Record the reason.
2110 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2111 // Store the updated state on the MDO, for next time.
2112 if (tstate1 != tstate0)
2113 pdata->set_trap_state(tstate1);
2114 } else {
2115 if (LogCompilation && xtty != NULL) {
2116 ttyLocker ttyl;
2117 // Missing MDP? Leave a small complaint in the log.
2118 xtty->elem("missing_mdp bci='%d'", trap_bci);
2119 }
2120 }
2121 }
2122
2123 // Return results:
2124 ret_this_trap_count = this_trap_count;
2125 ret_maybe_prior_trap = maybe_prior_trap;
2126 ret_maybe_prior_recompile = maybe_prior_recompile;
2127 return pdata;
2128 }
2129
2130 void
2131 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2132 ResourceMark rm;
2133 // Ignored outputs:
2134 uint ignore_this_trap_count;
2135 bool ignore_maybe_prior_trap;
2136 bool ignore_maybe_prior_recompile;
2137 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2138 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2139 bool update_total_counts = JVMCI_ONLY(false) NOT_JVMCI(true);
2140 query_update_method_data(trap_mdo, trap_bci,
2141 (DeoptReason)reason,
2142 update_total_counts,
2143 #if INCLUDE_JVMCI
2144 false,
2145 #endif
2146 NULL,
2147 ignore_this_trap_count,
2148 ignore_maybe_prior_trap,
2149 ignore_maybe_prior_recompile);
2150 }
2151
2152 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2153 if (TraceDeoptimization) {
2154 tty->print("Uncommon trap ");
2155 }
2156 // Still in Java no safepoints
2157 {
2158 // This enters VM and may safepoint
2159 uncommon_trap_inner(thread, trap_request);
2160 }
2161 return fetch_unroll_info_helper(thread, exec_mode);
2162 }
2163
2164 // Local derived constants.
2165 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2166 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
2167 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2168
2169 //---------------------------trap_state_reason---------------------------------
2170 Deoptimization::DeoptReason
2171 Deoptimization::trap_state_reason(int trap_state) {
2172 // This assert provides the link between the width of DataLayout::trap_bits
2173 // and the encoding of "recorded" reasons. It ensures there are enough
2174 // bits to store all needed reasons in the per-BCI MDO profile.
2175 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2176 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2177 trap_state -= recompile_bit;
2178 if (trap_state == DS_REASON_MASK) {
2179 return Reason_many;
2180 } else {
2181 assert((int)Reason_none == 0, "state=0 => Reason_none");
2182 return (DeoptReason)trap_state;
2183 }
2184 }
2185 //-------------------------trap_state_has_reason-------------------------------
2186 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2187 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2188 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2189 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2190 trap_state -= recompile_bit;
2191 if (trap_state == DS_REASON_MASK) {
2192 return -1; // true, unspecifically (bottom of state lattice)
2193 } else if (trap_state == reason) {
2194 return 1; // true, definitely
2195 } else if (trap_state == 0) {
2196 return 0; // false, definitely (top of state lattice)
2197 } else {
2198 return 0; // false, definitely
2199 }
2200 }
2201 //-------------------------trap_state_add_reason-------------------------------
2202 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2203 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2204 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2205 trap_state -= recompile_bit;
2206 if (trap_state == DS_REASON_MASK) {
2207 return trap_state + recompile_bit; // already at state lattice bottom
2208 } else if (trap_state == reason) {
2209 return trap_state + recompile_bit; // the condition is already true
2210 } else if (trap_state == 0) {
2211 return reason + recompile_bit; // no condition has yet been true
2212 } else {
2213 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2214 }
2215 }
2216 //-----------------------trap_state_is_recompiled------------------------------
2217 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2218 return (trap_state & DS_RECOMPILE_BIT) != 0;
2219 }
2220 //-----------------------trap_state_set_recompiled-----------------------------
2221 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2222 if (z) return trap_state | DS_RECOMPILE_BIT;
2223 else return trap_state & ~DS_RECOMPILE_BIT;
2224 }
2225 //---------------------------format_trap_state---------------------------------
2226 // This is used for debugging and diagnostics, including LogFile output.
2227 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2228 int trap_state) {
2229 assert(buflen > 0, "sanity");
2230 DeoptReason reason = trap_state_reason(trap_state);
2231 bool recomp_flag = trap_state_is_recompiled(trap_state);
2232 // Re-encode the state from its decoded components.
2233 int decoded_state = 0;
2234 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2235 decoded_state = trap_state_add_reason(decoded_state, reason);
2236 if (recomp_flag)
2237 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2238 // If the state re-encodes properly, format it symbolically.
2239 // Because this routine is used for debugging and diagnostics,
2240 // be robust even if the state is a strange value.
2241 size_t len;
2242 if (decoded_state != trap_state) {
2243 // Random buggy state that doesn't decode??
2244 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2245 } else {
2246 len = jio_snprintf(buf, buflen, "%s%s",
2247 trap_reason_name(reason),
2248 recomp_flag ? " recompiled" : "");
2249 }
2250 return buf;
2251 }
2252
2253
2254 //--------------------------------statics--------------------------------------
2255 const char* Deoptimization::_trap_reason_name[] = {
2256 // Note: Keep this in sync. with enum DeoptReason.
2257 "none",
2258 "null_check",
2259 "null_assert" JVMCI_ONLY("_or_unreached0"),
2260 "range_check",
2261 "class_check",
2262 "array_check",
2263 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2264 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2265 "unloaded",
2266 "uninitialized",
2267 "unreached",
2268 "unhandled",
2269 "constraint",
2270 "div0_check",
2271 "age",
2272 "predicate",
2273 "loop_limit_check",
2274 "speculate_class_check",
2275 "speculate_null_check",
2276 "speculate_null_assert",
2277 "rtm_state_change",
2278 "unstable_if",
2279 "unstable_fused_if",
2280 #if INCLUDE_JVMCI
2281 "aliasing",
2282 "transfer_to_interpreter",
2283 "not_compiled_exception_handler",
2284 "unresolved",
2285 "jsr_mismatch",
2286 #endif
2287 "tenured"
2288 };
2289 const char* Deoptimization::_trap_action_name[] = {
2290 // Note: Keep this in sync. with enum DeoptAction.
2291 "none",
2292 "maybe_recompile",
2293 "reinterpret",
2294 "make_not_entrant",
2295 "make_not_compilable"
2296 };
2297
2298 const char* Deoptimization::trap_reason_name(int reason) {
2299 // Check that every reason has a name
2300 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2301
2302 if (reason == Reason_many) return "many";
2303 if ((uint)reason < Reason_LIMIT)
2304 return _trap_reason_name[reason];
2305 static char buf[20];
2306 sprintf(buf, "reason%d", reason);
2307 return buf;
2308 }
2309 const char* Deoptimization::trap_action_name(int action) {
2310 // Check that every action has a name
2311 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2312
2313 if ((uint)action < Action_LIMIT)
2314 return _trap_action_name[action];
2315 static char buf[20];
2316 sprintf(buf, "action%d", action);
2317 return buf;
2318 }
2319
2320 // This is used for debugging and diagnostics, including LogFile output.
2321 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2322 int trap_request) {
2323 jint unloaded_class_index = trap_request_index(trap_request);
2324 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2325 const char* action = trap_action_name(trap_request_action(trap_request));
2326 #if INCLUDE_JVMCI
2327 int debug_id = trap_request_debug_id(trap_request);
2328 #endif
2329 size_t len;
2330 if (unloaded_class_index < 0) {
2331 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2332 reason, action
2333 #if INCLUDE_JVMCI
2334 ,debug_id
2335 #endif
2336 );
2337 } else {
2338 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2339 reason, action, unloaded_class_index
2340 #if INCLUDE_JVMCI
2341 ,debug_id
2342 #endif
2343 );
2344 }
2345 return buf;
2346 }
2347
2348 juint Deoptimization::_deoptimization_hist
2349 [Deoptimization::Reason_LIMIT]
2350 [1 + Deoptimization::Action_LIMIT]
2351 [Deoptimization::BC_CASE_LIMIT]
2352 = {0};
2353
2354 enum {
2355 LSB_BITS = 8,
2356 LSB_MASK = right_n_bits(LSB_BITS)
2357 };
2358
2359 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2360 Bytecodes::Code bc) {
2361 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2362 assert(action >= 0 && action < Action_LIMIT, "oob");
2363 _deoptimization_hist[Reason_none][0][0] += 1; // total
2364 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2365 juint* cases = _deoptimization_hist[reason][1+action];
2366 juint* bc_counter_addr = NULL;
2367 juint bc_counter = 0;
2368 // Look for an unused counter, or an exact match to this BC.
2369 if (bc != Bytecodes::_illegal) {
2370 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2371 juint* counter_addr = &cases[bc_case];
2372 juint counter = *counter_addr;
2373 if ((counter == 0 && bc_counter_addr == NULL)
2374 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2375 // this counter is either free or is already devoted to this BC
2376 bc_counter_addr = counter_addr;
2377 bc_counter = counter | bc;
2378 }
2379 }
2380 }
2381 if (bc_counter_addr == NULL) {
2382 // Overflow, or no given bytecode.
2383 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2384 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2385 }
2386 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2387 }
2388
2389 jint Deoptimization::total_deoptimization_count() {
2390 return _deoptimization_hist[Reason_none][0][0];
2391 }
2392
2393 jint Deoptimization::deoptimization_count(DeoptReason reason) {
2394 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2395 return _deoptimization_hist[reason][0][0];
2396 }
2397
2398 void Deoptimization::print_statistics() {
2399 juint total = total_deoptimization_count();
2400 juint account = total;
2401 if (total != 0) {
2402 ttyLocker ttyl;
2403 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2404 tty->print_cr("Deoptimization traps recorded:");
2405 #define PRINT_STAT_LINE(name, r) \
2406 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2407 PRINT_STAT_LINE("total", total);
2408 // For each non-zero entry in the histogram, print the reason,
2409 // the action, and (if specifically known) the type of bytecode.
2410 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2411 for (int action = 0; action < Action_LIMIT; action++) {
2412 juint* cases = _deoptimization_hist[reason][1+action];
2413 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2414 juint counter = cases[bc_case];
2415 if (counter != 0) {
2416 char name[1*K];
2417 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2418 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2419 bc = Bytecodes::_illegal;
2420 sprintf(name, "%s/%s/%s",
2421 trap_reason_name(reason),
2422 trap_action_name(action),
2423 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2424 juint r = counter >> LSB_BITS;
2425 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2426 account -= r;
2427 }
2428 }
2429 }
2430 }
2431 if (account != 0) {
2432 PRINT_STAT_LINE("unaccounted", account);
2433 }
2434 #undef PRINT_STAT_LINE
2435 if (xtty != NULL) xtty->tail("statistics");
2436 }
2437 }
2438 #else // COMPILER2_OR_JVMCI
2439
2440
2441 // Stubs for C1 only system.
2442 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2443 return false;
2444 }
2445
2446 const char* Deoptimization::trap_reason_name(int reason) {
2447 return "unknown";
2448 }
2449
2450 void Deoptimization::print_statistics() {
2451 // no output
2452 }
2453
2454 void
2455 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2456 // no udpate
2457 }
2458
2459 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2460 return 0;
2461 }
2462
2463 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2464 Bytecodes::Code bc) {
2465 // no update
2466 }
2467
2468 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2469 int trap_state) {
2470 jio_snprintf(buf, buflen, "#%d", trap_state);
2471 return buf;
2472 }
2473
2474 #endif // COMPILER2_OR_JVMCI